text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
Interactive Robot Transition Repair With SMT
Jarrett Holtz, Arjun Guha, Joydeep Biswas
arXiv:1802.01706v1 [cs.RO] 5 Feb 2018
University of Massachusetts Amherst
Abstract
Complex robot behaviors are often structured as
state machines, where states encapsulate actions
and a transition function switches between states.
Since transitions depend on physical parameters,
when the environment changes, a roboticist has to
painstakingly readjust the parameters to work in
the new environment. We present interactive SMTbased Robot Transition Repair (SRTR): instead of
manually adjusting parameters, we ask the roboticist to identify a few instances where the robot is
in a wrong state and what the right state should be.
A lightweight automated analysis of the transition
function’s source code then 1) identifies adjustable
parameters, 2) converts the transition function into
a system of logical constraints, and 3) formulates
the constraints and user-supplied corrections as a
MaxSMT problem that yields new parameter values. Our evaluation shows that SRTR is effective
on real robots and in simulation. We show that
SRTR finds new parameters 1) quickly, 2) with
only a few corrections, and 3) that the parameters
generalize to new scenarios. We also show that a
simple state machine corrected by SRTR can outperform a more complex, expert-tuned state machine in the real world.
1
Introduction
Complex robot control software is typically structured as a
state machine, where each state encapsulates a feedback controller. Even if each state is correct, the transitions between
states depend on parameters that are hard to get right, even
for experienced roboticists. It is very common for parameter
values to work in simulation but fail in the real world, to work
in one physical environment but fail in another, or to work on
one robot but fail on another. For example, Figure 1 shows
the trajectory of a robotic soccer player as it tries to kick a
moving ball. A very small change to its parameter values determines whether or not it succeeds.
Even a simple robot may have a large parameter space,
which makes exhaustive-search impractical. Moreover, robot
performance is usually non-convex with respect to parameter
Start
Go To
Intercept
Catch
Kick
End
(a) Attacker state machine
(b) Execution traces
Figure 1: A robot soccer attacker a) state machine, with b) successful
(blue) and unsuccessful (red) traces to Intercept a ball (orange) and
Kick at the goal. The green box isolates the error: the successful
trace transitions to Kick, the unsuccessful trace remains in Intercept.
values, which makes general optimization techniques susceptible to local minima. For some cases, there exist calibration
procedures to adjust parameters automatically (e.g., [Holtz
and Biswas, 2017]), but these are not general procedures.
Therefore, roboticists usually resort to adjusting parameters
manually—a tedious task that can result in poor performance.
We have the following key insights: a roboticist debugging
a robot can frequently identify when something goes wrong,
and what it should have done instead. When the robot control
software is structured as a state machine, this corresponds to
identifying when the robot is in the wrong state and what the
correct state should be. This is a partial specification of expected behavior: the roboticist does not need to enumerate a
complete sequence of states, the parameters to adjust, how to
adjust them, or even identify all errors.
We present SMT-based Robot Transition Repair (SRTR),
an approach to adjusting the parameters of robot state machines that leverages program repair techniques that have
gained traction in programming languages and software engineering research. Specifically, we define the grammar of
a repairable robot transition function as a subset of generalpurpose imperative programming languages. During execution, we log execution traces of the transition function. After
execution, the roboticist examines the execution traces and
corrects a handful of transitions. SRTR then converts the
set of corrections and the transition function into a logical
formula for a MaxSMT solver. A solution to this MaxSMT
problem minimizes changes to the parameters, while maxi-
mizing the number of satisfied corrections.
Our experimental results demonstrate that SRTR 1) is
more computationally tractable than exhaustive parameter
search; 2) is scalable; 3) generalizes to unseen situations; and
4) when applied to simple robot soccer attacker, can outperform a more complicated, expert-tuned attacker that won the
lower bracket finals at anonymized competition.
2
Related Work
Finite state machines (FSMs) can specify robot behavior and
enable robust autonomy. RoboChart [Miyazawa et al., 2017]
uses FSMs to verify liveness, timeliness, and other properties of robots. The HAMQ-INT algorithm [Bai and Russell,
2017] uses reinforcement learning to discover internal transitions and adapt to unexpected scenarios. SRTR also leverages FSMs, but focuses on correcting transition errors that
occur when the robot is deployed in a new environment.
Robot behavior can also be repaired by dynamic synthesis of new control structures, such as automatic synthesis of
new FSMs [Wong et al., 2014], or synthesis of code from a
context-free motion grammar with parameters derived from
human-inspired control [Dantam et al., 2013]. When automated synthesis is intractable, a user-generated specification
in a domain-specific language can be used to synthesize a
plan [Nedunuri et al., 2014]. SRTR assumes that the FSM
does not need new states or transitions, but that failures are
due to incorrect triggering of the transition function arising
from incorrect environment-dependent parameter values.
Robot behaviors rely on environment-dependent parameters for robust and accurate execution. If a precise model of
the dependency between parameters and behaviors is available, it may be possible to design a calibration procedure that
executes a specific sequence of actions and to recover correct
parameter values [Yang et al., 2014; Holtz and Biswas, 2017].
If a calibration procedure cannot be designed, but the effect
of parameters is well-understood, it may be possible to optimize for the parameters using a functional model [Cano et
al., 2016]. SRTR addresses the problem of repair even when
the dependence of the behaviors on parameters is not explicitly modeled. It automatically discovers such dependencies
by analysis of the source code of the transition function.
Human input can help overcome the limitations of autonomous algorithms [Kamar, 2016; Redacted, 2018]. Learning from demonstration (LfD) [Argall et al., 2009] and inverse reinforcement learning (IRL) [Abbeel and Ng, 2011]
allow robots to learn new behaviors from human demonstrations. LfD can also overcome model errors by correcting
portions of the state space [Meriçli et al., 2012]. These approaches require demonstrations in the full high-dimensional
state space of the robot, which can be tedious for users to
provide. When human demonstration does not specify why
an action was applied to a state, it can be hard to generalize
to a new situation. SRTR generalizes to new scenarios since
it infers dependencies from the source code of the transition
function. Furthermore, it requires only a partial specification
of correct behavior.
Domain-specific languages (DSLs) allow users to specify high-level behavior using abstractions such as Instruction
Graphs [Mericli et al., 2014]. Unlike DSLs that provide a
complete specification of robot behavior, SRTR only requires
sparse corrections that partially-specify expected behavior.
Automated program repair is an active area of research in
software engineering and programming languages. For example, DirectFix [Mechtaev et al., 2015] formulates program
repair as a MaxSMT problem and deems a program repaired
when all tests pass. Physical robots don’t have deterministic test cases and even user-provided corrections can be contradictory. SRTR uses MaxSMT to minimize changes and
maximize the number of satisfied corrections. There are several other domains that are not amenable to unit-testing. For
example, Tortoise [Weiss et al., 2017] allows the user to fix
system configuration errors using the Unix shell and propagates fixes to a system configuration specification. Whereas
Tortoise requires the user to completely fix the system, SRTR
only requires a partial specification.
Programming by example can synthesize programs from a
small number of examples; the approach has been very effective in domains such as synthesizing string transformations
on spreadsheet data [Gulwani, 2011]. RobustFill [Devlin et
al., 2017] generalizes the approach to also support noisy data.
Program templates can also be used to to synthesize task and
motion plans that satisfy user plan outlines [Nedunuri et al.,
2014]. Unlike these approaches, SRTR does not synthesis
new program structure and focuses on adjusting parameters.
For generalizability, SRTR minimizes the magnitude of adjustments and uses MaxSMT to discard noisy corrections.
3
Background
We use a real-world example to motivation SRTR: a robot
soccer attacker (also called a striker). The attacker must
1) go to the ball if the ball is stopped, 2) intercept the ball
if it is moving away from the attacker, 3) catch the ball if
it is moving toward the attacker, and 4) kick the ball at the
goal once the attacker has control of the ball. Each of these
sub-behaviors is a distinct, self-contained feedback controller
(e.g., ball interception [Biswas et al., 2014] or omnidirectional time optimal control [Kalmár-Nagy et al., 2004]). At
each time-step, the attacker 1) switches to a new controller
if necessary and 2) invokes the current controller to produce
new outputs. We represent the attacker as a robot state machine (RSM), where each state represents a controller. Figure 1a shows the structure of the attacker RSM.
In this paper, we assume that the output of each feedback
controller is nominally correct—while there might be minor
performance degradation of each controller when environmental factors change, we assume that they are convergent,
and hence will eventually result in the correct result. However, environmental factors also affect the transition function,
which transfers execution from one controller to another. For
example, the friction coefficient between the ball and the carpet, and the robot’s motor efficiency affect when the attacker
transitions from Intercept to Kick; and the mass of the ball affects when the attacker transitions from Kick to Done. These
factors vary from one environment to another. Since transition functions do not have any self-correcting mechanisms,
robots are prone to behaving incorrectly when their parame-
ters are incorrect for the given environment.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
inputs ballLoc : Vector2D , ballVel : Vector2D ,
robotLoc : Vector2D , robotVel : Vector2D ,
robotAng : Float , robotAngVel : Float ,
targetAng : Float ;
variables kickTime : Float ;
parameters kickedThresh , timeoutThresh , aimErrorThresh ,
angleVelThresh , distanceThresh , relativeYThresh ,
relativeVelThresh , alignThresh , offsetThresh ,
catchVelThresh , ivThresh ;
if (s == "START ") {
s = "GOTO"
} else if (s == "KICK ") {
robotHeading = [cos( robotAng ), sin( robotAng )];
if (|| ballVel · robotHeading || > kickedThresh ) {
s = "END"
} else if ( kickTime > timeoutThresh ) {
s = "START"
} else {
s = "KICK" }
} else {
ballToRobot = ballLoc - robotLoc ;
aimError = AngleMod ( targetAng - robotAng );
ballVelTarget = Rotate (ballVel , targetAng );
robotVelTarget = Rotate (robotVel , targetAng );
ballToRobotVel = robotVelTarget - ballVelTarget ;
robotYAxis = [-sin( robotAng }, cos( robotAng )];
ballToRobotY = robotYAxis · ballToRobot ;
ballIntercept = ballLoc +
,→ || ballVel || * UnitVector ( ballToRobot );
interceptOffset = || ballIntercept - robotLoc ||;
interceptVel = ballVel *
,→ ballVel · UnitVector ( ballToRobot );
if ( aimError < aimErrorThresh &&
robotAngVel < angleVelThresh &&
|| ballToRobot || < distanceThresh &&
[0, 1] · ballToRobotVel < relativeYThresh &&
|| ballToRobotVel || < relativeVelThresh &&
|| ballToRobotY || < alignThresh ) {
s = "KICK"
} else if ( interceptOffset < offsetThresh &&
|| ballVel || > catchVelThresh &&
|| interceptVel || / || ballVel || > ivThresh ) {
s = "CATCH"
} else if (|| ballVel || > catchVelThresh ) {
s = " INTERCEPT "
} else {
s = "GOTO" } }
Figure 2: Transition function of an attacker.
3.1
Robot State Machines
A robot state machine (RSM) is a discrete-time Mealy
machine that is extended with continuous inputs, outputs,
and program variables. Formally, an RSM is a 9-tuple
hS, S0 , SF , V, V0 , Y, U, T, Gi, where S is the finite set of
states, S0 ∈ S is the start state, SF ∈ S is the end state,
V ∈ Rm is the set of program variables values, V0 ∈ Rm
are the initial values of the program variables, Y ∈ Rn are
the continuous inputs, U ∈ Rl are the continuous actuation
outputs, T : S × Y × V → S is the transition function, and
G : S × Y × V → U × V is the emission function. At
each time step t, the RSM first uses the transition function to
select a state and then the emission function to run the controller associated with that state. The transition function can
only update the current state, whereas the emission function
can update program variables and produce outputs.
Figure 2 shows the transition function for the attacker, written in C-style pseudocode. The function depends on the inputs and variables (declared on lines 1–5) and uses them to
Unary Operators
op 1 ::= - | sin | cos | · · ·
Expressions
e ::= k
| s
| xv
| xy
| xp
| op 1 (e)
| e1 op 2 e2
Constants
State
Variables
Inputs
Parameters
Transition Functions
T ::= { m1 ; · · · ; mn }
Binary Operators
op 2 ::= + | - | * | > | · · ·
Statements
m ::= s := e
| xv := e
| if (e) m1 else m2
| { e1 ; · · · ; en }
Parameter Maps
n
P ::= hx1p 7→ k1 · · · xn
p 7→ k i
Traces
R ::= [τ1 · · · τn ]
Corrections
ct ::= s : s ∈ S
Trace Elements
m
n
i=1
j=1
τt ::= {in = h ∀ xiy 7→ ki i, vars = h ∀ xjv 7→ k0j i, state = st 7→ ks00 }
Figure 3: Syntax of transition functions, traces, and corrections.
perform geometric calculations. The function has 11 different
parameters (lines 6–9) that determine when certain transitions
occur. The figure does not specify the parameters’ values, but
they are global constants. Finally, the transition function uses
a few helpers that we elide: UnitVector scales a vector by
its L2-norm, AngleMod bounds an angle to the range [−π, π),
and Rotate rotates a vector by an angle.
3.2
Transition Errors
Figure 1b shows two traces of the attacker. In the blue trace,
the attacker intercepts the moving ball and kicks it at the goal.
But, in the red trace, the attacker fails to kick: it remains stuck
in the Intercept state and never transitions to Kick. Over the
course of several trials (e.g., a robot soccer game), we may
find that the attacker only occasionally fails to kick. When
this occurs, it is usually the case that the high-level structure
of the transition function is correct, but that the values of the
parameters need to be adjusted. Unfortunately, since there are
11 real-valued parameters, the search space is large.
To efficiently search for new parameter values, we need to
reason about the structure of the transition function. To do so,
the next section describes how we systematically convert it
to a formula in propositional logic, extended with arithmetic
operators. This formula encodes the structure of the transition
function along with constraints from the user corrections, and
allows an SMT solver to efficiently find new parameters.
4
Interactive Robot Transition Repair
The SRTR algorithm (lines 32–36 of Figure 4) has four inputs: 1) the transition function, 2) a map from parameters
to their values, 3) an execution trace, and 4) a set of userprovided corrections. The result of SRTR is a corrected parameter map that maximizes the number of corrections satisfied and minimizes the changes to the input parameter map.
(The tradeoff between these objectives is a hyperparameter.)
SRTR has three major steps. 1) It parses the transition
function code and converts it to an abstract syntax tree for repair (Section 4.1). 2) It uses a lightweight program analysis
to identify parameters that cannot be repaired, and for each
user-provided correction, it partially evaluates the transition
function for the inputs and variable values at the time of correction, yielding residual transition functions (Section 4.2).
1
1
n
e
n
1 def Peval(T ,x 7→ k · · · x 7→ k ):
2
3
4
5
6 def Rep(T):
7 // Produces the parameters of T that do not need to be partially evaluated.
8
9 def Unrep(T):
10 // Produces the parameters of T that must be partially evaluated.
11
12 def Residual(T ,τt ,P ):
13
l
m
i=1
j=1
m
{in = h ∀ xiy 7→ ki i, vars = h ∀ xjv 7→ k0j i, state = st 7→ ks00 } := τt
0
l
Peval(T , ∀ xiy
i=1
i
, ∀ xiv
j=1
0i
, st 7→ ks00 )
14
T :=
15
0
{x1p , · · · , xn
p } = Unrep(T )
0
0
T := Peval(T ,x1 7→ P (x1 ), · · · , xn 7→ P (xn ))
return T 0
16
7→ k
7→ k
17
18
19 def CorrectOne(T ,τt ,P ,ct ):
20
21
22
T 0 := Residual(T ,τt )
{x1p , · · · , xm
p } := Rep(T )
1
0m
m
return ∃δ 1 , · · · , δ m : ct = T 0 (s1 , x01
p + δ , · · · , xp + δ )
23
1
n
24 def CorrectAll(T , P , R, {ct , · · · , ct }):
25
26
27
28
29
30
31
{x1p , · · · , xm
p } = Rep(T )
Φ = true
for i ∈ [1 · · · n]:
∃δ 1 , · · · , δ m : φi = CorrectOne(T ,R[t],P ,cit )
Φ = Φ ∧ (wi = H Y (wi = 0 ∧ φi ))
return ∃δ 1 , · · · , δ m , w1 , · · · , wn : Φ
1
n
32 def SRTR(T , P , R,{ct , · · · , ct }):
1
n
33 assert(CorrectAll(T ,P ,R,{ct , · · · , ct }))
34
35
36
cos(θy ) < xp
cos(xp )
θp + π
sin(θp + π)
// 1) Substitute xi with ki in T . 2) Simplify all operations and conditionals
// with immediate arguments until simplification is no longer possible.
return T 0 // partially evaluated T
j
minimize(Σi=1 wi + Σm
j=1 kδ k)
1
m
{xp , · · · , xp } = Rep(T )
m
m
return [x1p 7→ P (x1p ) + δ 1 , · · · , xm
p 7→ P (xp ) + δ ]
Figure 4: The core SRTR algorithm.
3) Finally, it uses the residual transition functions to formulate an optimization problem for an off-the-shelf MaxSMT
solver [Bjørner et al., 2015]. The solution to this problem is
an adjustment to the parameter values (Section 4.3).
4.1
The Language Of Transition Functions
To abstract away language-specific details of our repair procedure, we first parse the source code and convert it into an
idealized imperative language that only has features essential
for writing transition functions (Figure 3). The language has
a special identifier s that refers to the current state and other
identifiers have a subscript to indicate whether they are inputs
(xy ), parameters (xp ), or ordinary variables (xv ). The transition function can read any identifier, but it can only write to
variables and the state. Figure 3 has a list of operators that
often appear in transition functions, such as arithmetic and
trigonometric functions, but the list is not exhaustive. Finally, a transition function (T ) is a block of statements and
a parameter map (P ) assigns values to the parameters in T .
An execution trace is a sequence of trace elements (τt ). A
trace element records the values of sensor inputs, ordinary
variables, and the state at the start of time-step t. In contrast,
a user-provided correction (ct ) specifies the expected state at
the end of time-step t. To repair parameters, SRTR estab-
τ.vars
τ.in
e0
P
θy = 0
1 < xp
−1
θp + π
0
xp 7→ π
xp 7→ π
xp 7→ π
xp 7→ π
Table 1: Examples of residuals and repairable parameters.
lishes constraints that relate the observed initial state, the expected output state, and the other identifiers.
4.2
Residual Transition Functions
SRTR needs to translate the transition function into a logical
formula for the SMT solver. However, solvers do not have decision procedures for non-linear arithmetic and trigonometry,
both of which commonly occur in transition functions.
Partial Evaluation [Jones et al., 1993] is a technique that
specializes programs to work on specific inputs. For example,
if we partially evaluate the program cos(x) + y for x 7→ 0,
we get the program 1 + y. In general, partial evaluation first
substitutes identifiers with concrete values and then simplifies
the program as much as possible. SRTR uses a canonical
partial evaluator for transition functions, which we call Peval
( Figure 4, line 1). The arguments of Peval are a transition
function and a mapping from identifiers to constants.
The Residual function specializes a transition function for
a particular trace element. We call its result a residual transition function. Residual first partially evaluates the transition
function with respect to all the sensor inputs and variable values recorded in the trace element (line 14). This step eliminates most functions that the solver cannot represent, thus the
only identifiers that remain are the parameters and the state.
For example, the first row of Table 1 shows an expression
that uses a trigonometric function. If this expression occurs
in the transition function, its residual after partial evaluation
is a simple inequality that cannot be further simplified.
However, it is possible for the transition function to use a
parameter in a context that the solver cannot represent. For
example, the second row of Table 1 applies a trigonometric
function to a parameter and the first step of Residual simply
yields the original expression. (i.e., Peval(cos(xp )) = cos(xp )).
To address this, Residual calculates the set of parameters that
cannot be repaired (line 15) and partially evaluates once again
with respect to their values in the parameter map (line 16).
Therefore, the final result in this example is −1.
The Residual function uses a pair of auxiliary functions
Rep and Unrep to calculate the set of parameters that can, and
cannot be repaired. These functions perform a lightweight
analysis by mutual recursion over the transition function.
Mutual recursion is necessary because a parameter may not
be repairable even if its immediate context is repairable. For
example, the third row of Table 1 shows an expression with
a repairable parameter thus its residual is θp + π. However,
the fourth row embeds the same expression in a context that
makes it unrepairable, thus the residual is partially evaluated
using the parameter map, yielding 0.
4.3
Transition Repair as a MaxSMT Problem
This section presents how SRTR formulates a MaxSMT
problem in three steps. 1) It translates each correction into
an independent formula. A solution to this formula (if one
exists) corresponds to parameter adjustments that satisfy the
correction. 2) It combines the formulas from the previous
step into a single formula with independent weights for each
sub-formula. A solution to this formula corresponds to parameter adjustments that satisfy a subset of the sub-formulas.
Any sub-formulas that could not be satisfied have a positive
weight. 3) Finally, we formulate a MaxSMT problem that
minimizes the magnitude of adjustments and the weight of
violated sub-formulas.
The CorrectOne function transforms a single correction
into a formula. This function 1) calculates the residual transition function (Figure 4, line 20), 2) gets the repairable parameters (line 21), and 3) produces a formula (line 22) with
a variable δ i that corresponds to an adjustment to the ith repairable parameter. This formula constrains the residual transition function’s output state to be equal to the correction.
Therefore, if the correction differs from the post-state of the
transition function at run-time, the solution to the formula
will have a non-zero adjustment for at least one of the repairable parameters. (i.e., δ i 6= 0 for some i.)
The CorrectAll function supports multiple corrections and
uses CorrectOne as a subroutine. The function iteratively
builds a conjunctive formula Φ, where each clause has a distinct weight wi and two mutually exclusive cases: Either 1)
wi = 0 thus the clause has no cost and the adjustments to the
parameters satisfy the ith correction (line 29) or 2) wi = H
and the ith correction is violated. Thus H ∈ R+ is a hyperparameter that induces a tradeoff between satisfying more
corrections vs. minimizing the magnitude of the adjustments:
large values of H allow more corrections to be satisfied, and
small values of H minimize the magnitude of the adjustments. The final formula has m real-valued variables δ i for
adjustments to the corresponding m repairable parameters xip ,
and n discrete variables wj that represent the cost of violating
formula φj corresponding to correction cjt .
Finally, the SRTR function uses CorrectAll as a subroutine
and invokes the MaxSMT solver. This function 1) adds the
assertion returned by CorrectAll to the solver, 2) directs the
solver to minimize the sum of weights and the sum of the
magnitude of parameter changes (line 34), and 3) returns a
map from repairable parameter names to their new values.
5
Evaluation
We evaluate SRTR in four ways. 1) We compare SRTR to
an exhaustive search; 2) We show how the number of corrections affect RSM performance and that SRTR requires only a
small number of corrections to perform well; 3) Using three
RSMs, we show that SRTR does not over-fit and performs
well in new scenarios; and 4) We use SRTR to improve the
performance of a real-world robot.
We present three RSMs in this section and measure their
success rates as follows. The attacker (Figure 1a), which is
our running example, fills the main offensive role in robot
soccer. Its success rate is the fraction of the test scenarios
where it successfully kicks the ball into the goal. The deflector (Figure 5a) plays a supporting role in robot soccer, performing one-touch passing [Bruce et al., 2008]. Its success
Start
Setup
Wait
Kick
End
(a) Deflector RSM
S1 Left
S1 Forward
End
Start
S2 Left
S2 Forward
S1 Right
S2 Right
(b) Docker RSM
Figure 5: RSMs used for experiments
Method
Success Rate (%) CPU Time
Initial Parameters
44
—
Exhaustive Search
89 1,300 hr
SRTR
89
10 ms
Table 2: Success rate and CPU time compared to exhaustive search.
rate is the fraction of the test scenarios where it successfully
deflects the ball. The docker (Figure 5b) drives a differential drive robot to line up and dock with a charging station.
Its success rate is the fraction of the test scenarios where it
successfully docks with the charging station.
5.1
Comparison To Exhaustive Search
Using the attacker, we compare SRTR to an exhaustive
search to show that 1) SRTR is dramatically faster and 2)
the adjustments found by SRTR are as good as those found
by exhaustive search. To limit the cost of exhaustive search,
the experiment only repairs the six parameters that affect transitions into the Kick state; we bound the search space by the
physical limits of the parameters; and we discretize the resulting hypercube in parameter space. We evaluate each parameter set using 13 simulated positions and manually specify if
the position should transition to the Kick state.
We evaluate the initial parameter values, the SRTRadjusted parameters, and the parameters found by exhaustive search on 20,000 randomly generated scenarios. Table 2
reports the success rate and running time of each approach.
SRTR and exhaustive search achieve a comparable success
rate. However, SRTR completes in 10 ms whereas exhaustive search takes 1,300 CPU hours (using 100 cores).
5.2
Scalability
Using the attacker, we now show how the number of corrections affects SRTR solver time and success rate. Starting with
the same set of initial parameters used in Section 5.1, we first
create a training dataset by simulating the attacker in 40 randomly generated scenarios. For each scenario, we provide
one correction, thus we have a training set of 40 corrections.
Each trial applies SRTR to a subset of these corrections and
evaluates the success rate using 150 random test scenarios.
We repeat this procedure for each N ∈ [1, 40] with 50 randomly chosen subsets of the training set (i.e., 300,000 total
trials). Figure 6a shows how the number of corrections affects the success rate. It is possible for a single informative
correction to dramatically increase the success rate, or for a
particularly under-informative correction to have little effect.
Therefore we also report the mean success rate and show the
99% confidence interval in gray. Success rate increases with
the number of corrections, and with 23 corrections, the attacker reaches a peak success rate of 87%. Figure 6b shows
0.04
RSM
Solver Time (s)
Success Rate (%)
70
60
0.03
Attacker
Deflector
Docker
0.02
SRTR
Corrections
12
2
5
3
9
3
Params
80
Success Rates (%)
Baseline Expert SRTR
57,600
42
44
89
16,776
1
65
80
5,000
0
0
100
Tests
Table 3: Success rates for baseline, expert, and SRTR parameters.
0.01
50
0.00
0
10
20
30
Number of Corrections
(a) Success Rate
40
0
10
20
30
40
Number of Corrections
(b) Solver Time
Figure 6: Scaling of SRTR with number of corrections. We report
the mean as a line, the 99% confidence interval in grey, inliers in
blue, and outliers in red. Darker points represent more occurrences.
that the solver time increases linearly with the number of corrections, and remains less than 0.04s with 40 corrections.
5.3
Generalizability
Since SRTR uses a handful of user-provided corrections to
adjust parameters, there is a risk that it may overfit and underperform in new scenarios. We use three RSMs to show
that 1) SRTR-adjusted parameters generalize to new scenarios and that 2) SRTR outperforms a domain-expert who has
30 minutes to manually adjust parameters.
Table 3 summarizes the results of this experiment. We evaluate the success rate of the Attacker, Deflector, and Docker
RSMs on a test dataset with several thousand test scenarios
each. The baseline parameters that we use for these RSMs
have a low success rate. We give a domain expert complete
access to the RSM code (i.e., the transition and emission functions), and subsequently our simulator for 30 mins. In that
time, the expert is able to dramatically increase the success
rate of the Deflector, but has minimal impact on the success
rate of the Attacker and the Docker. Finally, we apply SRTR
using a handful of corrections and the baseline parameters.
The SRTR-adjusted parameters perform significantly better
than the baseline and domain-expert parameters.
The heat maps in Figure 7 illustrates how parameters found
by the domain-expert, nd by SRTR generalize to novel scenarios with the Attacker. In both heat maps, the goal is the
green bar and the initial position of the Attacker is at the origin. Each coordinate corresponds to an initial position of the
ball and for each position we set the ball’s initial velocity to
12 uniformly distributed angles. With the expert-adjusted parameters, the Attacker performs well when the ball starts in its
immediate vicinity, but performs poorly otherwise. However,
with SRTR-adjusted parameters, the Attacker is able to catch
or intercept the ball from most positions on the field. For this
result, we required only two corrections and the cross-marks
in the figure show the initial position of the ball for both corrections. Therefore, although SRTR only adjusted parameters to account for these two corrections, the result generalized to many other positions on the field.
5.4
Case Study: SRTR In The Real World
To evaluate SRTR in the real world, we follow the same procedure that experts use (summarized in Table 4): we develop
the Attacker in a simulator, we adjust parameters until it performs well in simulation, and then we find that it performs
(a) Expert-adjusted parameters. (b) SRTR-adjusted parameters.
Figure 7: Attacker success rate with respect to different initial ball
positions. The corrections are marked with a cross.
poorly in the real world. To evaluate the success rate of the
attacker in the real world, we start the ball from 18 positions on the field and repeat each position five times with the
same velocity (i.e., 90 trials). The parameters from simulation have a 25% success rate. Using the execution logs of
this experiment, we apply SRTR with three corrections. The
adjusted parameters increase the success rate to 73%. In practice, an expert would iteratively adjust parameters, so we apply SRTR again with 2 more corrections, which increases the
success rate to 86%. Finally, our group has an attacker that we
tested and optimized extensively for anonymized competition,
where it was part of a team that won the lower bracket. This
Competition Attacker has additional states to handle special
casesthat do not occur in our tests. On our tests, the Competition Attacker’s success rate is 76%. Therefore, with two iterations of SRTR, the simpler Attacker actually outperforms
the Competition Attacker in typical, real-world scenarios.
6
Conclusion
This paper presents SMT-based Robot Transition Repair
(SRTR), which is a semi-automatic white-box approach for
adjusting the transition parameters of Robot State Machines.
We demonstrate that SRTR: 1) increases success rate for multiple behaviors; 2) finds new parameters quickly using a small
number of annotations; 3) produces solutions which generalize well to novel situations; and 4) improves performance in
a real world robot soccer application.
Trial
Competition Attacker
Parameters from Simulation
Real World SRTR Tuning 1
Real World SRTR Tuning 2
Success Rate (%)
75
24
73
85
Table 4: Attacker success rates on a real robot.
References
[Abbeel and Ng, 2011] Pieter Abbeel and Andrew Y Ng. Inverse reinforcement learning. In Encyclopedia of machine
learning, pages 554–558. Springer, 2011.
[Argall et al., 2009] Brenna D. Argall, Sonia Chernova,
Manuela Veloso, and Brett Browning. A Survey of Robot
Learning From Demonstration. Robotics and Autonomous
Systems, pages 469 – 483, 2009.
[Bai and Russell, 2017] Aijun Bai and Stuart Russell. Efficient Reinforcement Learning with Hierarchies of Machines by Leveraging Internal Transitions. In IEEE International Conference on Robotics and Automation, 2017.
[Biswas et al., 2014] Biswas, Mendoza, Zhu, Choi, Klee,
and Veloso. Opponent-driven planning and execution for
pass, attack, and defense in a multi-robot soccer team.
In International Conference on Autonomous Agents and
Multi-Agent Systems, pages 493–500, 2014.
[Bjørner et al., 2015] Nikolaj Bjørner, Anh-Dung Phan, and
Lars Fleckenstein. νZ-an optimizing SMT solver. In International Conference on Tools and Algorithms for the
Construction and Analysis of Systems, pages 194–199.
Springer, 2015.
[Bruce et al., 2008] James Bruce, Stefan Zickler, Mike Licitra, and Manuela Veloso. CMDragons: Dynamic Passing and Strategy on a Champion Robot Soccer Team. In
IEEE International Conference on Robotics and Automation, pages 4074–4079, 2008.
[Cano et al., 2016] José Cano, Alejandro Bordallo, Vijay
Nagarajan, Subramanian Ramamoorthy, and Sethu Vijayakumar. Automatic configuration of ros applications
for near-optimal performance. In IEEE/RSJ International
Conference on Intelligent Robots and Systems, pages
2217–2223, 2016.
[Dantam et al., 2013] Neil Dantam, Ayonga Hereid, Aaron
Ames, and Mike Stilman. Correct Software Synthesis for
Stable Speed-Controlled Robotic Walking. In Robotics:
Science and Systems, 2013.
[Devlin et al., 2017] Jacob Devlin, Jonathan Uesato, Surya
Bhupatiraju, Rishabh Singh, Abdelrahman Mohammad,
and Pushmeet Kohli. RobustFill: Neural program learning under noisy I/O. In ICML, 2017.
[Gulwani, 2011] Sumit Gulwani. Automating string processing in spreadsheets using input-output examples. In ACM
SIGPLAN Notices, pages 317–330. ACM, 2011.
[Holtz and Biswas, 2017] Jarrett Holtz and Joydeep Biswas.
Automatic Extrinsic Calibration of Depth Sensors with
Ambiguous Environments and Restricted Motion. In
IEEE/RSJ International Conference on Intelligent Robots
and Systems, pages 2235–2240, 2017.
[Jones et al., 1993] Neil D. Jones, Carsten K. Gomad, and
Peter Sestoft. Partial Evaluation and Automatic Program
Generation. Prentice-Hall, Inc., 1993.
[Kalmár-Nagy et al., 2004] Tamás Kalmár-Nagy, Raffaello
D’Andrea, and Pritam Ganguly. Near-Optimal Dynamic
Trajectory Generation and Control of an Omnidirectional
Vehicle. Robotics and Autonomous Systems, pages 47–64,
2004.
[Kamar, 2016] Ece Kamar. Directions in Hybrid Intelligence: Complementing AI Systems with Human Intelligence. In IEEE International Conference on Robotics and
Automation, 2016.
[Mechtaev et al., 2015] Sergey Mechtaev, Jooyong Yi, and
Abhik Roychoudhury. DirectFix: Looking for Simple Program Repairs. In IEEE/ACM International Conference on
Software Engineering, pages 448–458, 2015.
[Meriçli et al., 2012] Çetin Meriçli, Manuela Veloso, and
H Levent Akın. Multi-resolution corrective demonstration
for efficient task execution and refinement. International
Journal of Social Robotics, pages 423–435, 2012.
[Mericli et al., 2014] Cetin Mericli, Steven D. Klee, Jack Paparian, and Manuela Veloso. An Interactive Approach for
Situated Task Specification Through Verbal Instructions.
In Autonomous Agents and Multi-agent Systems, pages
1069–1076, 2014.
[Miyazawa et al., 2017] Alvaro Miyazawa, Pedro Ribeiro,
Wei Li, Ana Cavalcanti, and Jon Timmis. Automatic property checking of robotic applications. In IEEE/RSJ International Conference on Intelligent Robots and Systems,
pages 3869–3876, 2017.
[Nedunuri et al., 2014] Srinivas Nedunuri, Sailesh Prabhu,
Mark Moll, Swarat Chaudhuri, and Lydia E. Kavraki.
SMT-based synthesis of integrated task and motion plans
from plan outlines. In IEEE International Conference on
Robotics and Automation, pages 655–662, 2014.
[Redacted, 2018] Redacted. Human-in-the-loop slam. In
Redacted, 2018.
[Weiss et al., 2017] Aaron Weiss, Arjun Guha, and Yuriy
Brun. Tortoise: Interactive System Configuration Repair. In IEEE/ACM International Conference on Automated Software Engineering, pages 625–636, 2017.
[Wong et al., 2014] Kai Weng Wong, Rädiger Ehlers, and
Hadas Kress-Gazit. Correct High-level Robot Behavior
in Environments with Unexpected Events. In Robotics:
Science and Systems, 2014.
[Yang et al., 2014] Xiangdong Yang, Liao Wu, Jinquan Li,
and Ken Chen. A Minimal Kinematic Model for Serial
Robot Calibration Using POE Formula. Robotics and
Computer-Integrated Manufacturing, pages 326 – 334,
2014.
| 6cs.PL
|
arXiv:1504.03021v1 [physics.comp-ph] 12 Apr 2015
Elastic properties of the degenerate f.c.c.
crystal of polydisperse soft dimers at zero
temperature
J. W. Narojczyk 1 and K. W. Wojciechowski 2
Institute of Molecular Physics, Polish Academy of Science, Smoluchowskiego 17,
PL-60-179 Poznań, Poland
Abstract
Elastic properties of soft, three-dimensional dimers, interacting through site-site
n-inverse-power potential, are determined by computer simulations at zero temperature. The degenerate crystal of dimers exhibiting (Gaussian) size distribution
of atomic √
diameters - i.e. size polydispersity - is studied at the molecular number
density 1/ 2; the distance between centers of atoms forming dimers is considered
as a length unit. It is shown that, at the fixed number density of the dimers, increasing polydispersity causes, typically, an increase of pressure, elastic constants
and Poisson’s ratio; the latter is positive in most direction. A direction is found,
however, in which the size polydispersity causes substantial decrease of Poisson’s
ratio, down to negative values for large n. Thus, the system is partially auxetic for
large polydispersity and large n.
Keywords: auxetic materials, elastic constants, polydispersity, modeling and
simulation, mechanical properties, negative Poisson’s ratio
PACS: 07.05.Tp, 61.43.-j, 62.20.de, 62.20.dj
1
Introduction
Among the characteristics of real materials, elastic properties are the ones that
are crucial from the point of view of various applications. Recently, increasing
interest is observed in the area of materials with unusual elastic properties an example are materials with negative Poisson’s ratio [1–7]. When stretched
(or resp. compressed) such materials increase (resp. decrease) their size not
1
2
[email protected]
[email protected]
Preprint submitted to Elsevier
14 April 2015
only in the direction of the stretch (resp. compression) but also in one or
more directions transverse to it [7]. Materials exhibiting such counterintuitive
behavior are known as anti-rubber [8], dilational materials [9] or auxetics [2].
In this paper, the last name will be used for its brevity.
As for isotropic systems the Poisson’s ratio is direction independent, any
isotropic material can be either auxetic or not. For anisotropic systems, which
are the subject of the present paper, the situation is more complex. Apart
from the alternative that the Poisson’s ratio is in all directions negative or
not, a third possibility exists: in some directions it is negative and in other - it
is not. In the literature related to the field, various names have been proposed
to describe different situations [4, 10–12]. Here we will adopt the scheme introduced in Ref. [12], in which - as in the isotropic case - the names auxetics
and nonauxetics correspond, respectively, to negative or nonnegative Poisson’s
ratio in all directions. Materials with Poisson’s ratios negative in some directions and nonnegative in the other ones will be referred to as partially auxetic
materials or just partial auxetics.
The very existence of auxetics and partial auxetics encourages one to study
influence of various micro-level mechanisms on the elastic properties of condensed matter [13–44]. An important problem in this context is the influence of disorder on the Poisson’s ratio. For some two-dimensional (2D) models which are elastically isotropic (at least for small deformations) and for
some anisotropic 2D models, it has been found that various kinds of disorder
(such as vacancies, interstitials, aperiodicity or polydispersity) introduced in
crystalline structures, increase Poisson’s ratio [45–50]. Studies of some threedimensional (3D) molecular models [51] revealed similar effect caused by vacancies, what might suggest a general conclusion that (at least in simple molecular systems) disorder weakens or eliminates auxeticity.
Recently, however, it has been found that in one of the simplest 3D systems soft spheres, forming the f.c.c. phase, there exists a direction in which one of
the form of disorder (polydispersity) decreases Poisson’s ratio, which reaches
negative values down to -1 [52]. This observation might find practical applications, e.g. in production of partial auxetics of latex meso- or macroscopic
spheres for which size polydispersity is easy to control. Unfortunately, the
mentioned lowest Poisson’s ratios were observed only for the softest spheres
with n = 6, i.e. when the f.c.c. structure may be unstable with respect to the
b.c.c. structure [53].
Taking into account that elastic properties of many-body systems strongly depend on details of the intermolecular interaction potential and, in particular,
on the molecular shape, it is interesting to study other 3D systems searching
for stable structures for which increasing polydispersity can lead to more negative Poisson’s ratios in some directions. As it is meaningful to start with the
2
simplest molecular shapes, the natural candidate for a first 3D molecule to
investigate is a polydisperse dimer.
The aim of this study is to determine the influence of disorder introduced
to a solid of soft dimers by the size dispersion of their atoms, on the elastic
properties of the model. The study
√ concerns the zero temperature, T = 0, and
the number density N/V = 1/ 2, where V is the volume and N is the number
of dimers. At this density atoms of dimers can be arranged in a perfect f.c.c.
lattice of lattice constant equal to the distance of atomic centers in the dimer.
The dimers form a degenerate crystalline (DC) phase - an analogue of that
observed in 2D [54]. The phase is characterized by non-periodic positions and
orientations of molecules, whereas the atoms forming molecules are arranged in
a perfect f.c.c. lattice (when the atoms are identical) or (for slightly different
sizes of atoms) in a lattice close to the f.c.c. one. The DC phase is known
to be thermodynamically stable and effectively cubic from the point of view
of elastic constants for hard dimers [55] and is expected to preserve those
properties for soft dimers with short ranged interactions which are studied in
the present paper.
The soft dimer system can be seen as a rough model of some man-made
nano-, meso- or macro-particle systems, e.g. systems of latex particles [56, 57].
Recently, an increasing interest is observed in systems of soft particles [58],
especially in the context of physics of colloids. This is due to the fact that
the latter have a broad range of possible applications. Systems of soft polydisperse particles seem to be more realistic models of colloids than those built
of identical particles. Polydispersity is also important for nanoparticle crystals [59–61], where the speed of crystallization and resulting crystal quality is
strongly related to size dispersion of particles.
Studies of elastic properties of models of polydisperse systems can be thought
of as a first step on a path to better understanding mechanisms and phenomena
that govern the elastic properties of real systems. Such a knowledge is not only
important from the point of view of fundamental research but also useful from
the material engineering point of view - it can help to engineer materials with
desired elastic properties.
2
Description of the studied model
The model studied in this work consists of N diatomic molecules (dimers) in
periodic boundary conditions. The dimers are rigid i.e. the distance between
centers of the atoms forming each one is constant and equal to σ (which
constitutes a unit of length) for each dimer. Initially the atoms form the perfect
f.c.c. lattice and the molecules form a degenerate crystalline (DC) phase. The
3
atoms interact via purely repulsive potential of the form:
uij (dij , rij ) = u0
dij
rij
!n
,
(1)
where dij = (di + dj )/2, di , dj are the diameters of the interacting spheres
(atoms), rij is the distance between their centers and u0 is the energy unit.
The exponent n is further referred to as the hardness parameter and its inverse, 1/n, as the softness parameter. (These names come from the observation
that when n → ∞ the interaction potential described by Eq. (1) tends to the
hard-body potential, which is widely used in computer simulations and condensed matter theory [34].) Atomic interactions are assumed to be short-range
ones, i.e. only the nearest neighbors (shearing a face of their Voronoi polyhedra) interact. Since the molecules are considered as rigid, interactions between
atoms forming a molecule are neglected.
2.1 Atomic size polydispersity
In the studied system, atomic diameters were generated randomly according
to the Gauss distribution function with a given standard deviation δ
1q 2
hd i − hdi2 .
δ=
σ
(2)
In above equation, d is the atomic diameter and σ = hdi. Further, in this
work, δ is treated as the measure of disorder in the system and is referred to
as the polydispersity parameter. Generated values of atomic diameters were
a)
b)
Fig. 1. Example of studied structures with atomic size polydispersity. The structure
shown consists of 432 molecules and has the polydispersity parameter δ equal to 3%.
In (a) the atoms are drawn in the shades of gray, indicating their size relative to
other atoms. The ones drawn in white are the smallest and those drawn in black are
the largest. The gray cube indicates the periodic box. In (b) connections between
atoms in the same structure as in (a) are shown.
4
randomly distributed among atoms in the structure. In the Fig. 1 examples of
studied structures are presented. It can be seen that even in the presence of
atomic size polydispersity atoms form nearly perfect f.c.c. lattice.
3
Basic formulae
In the following, it will be assumed that the studied system of soft dimers
exhibits effective regular symmetry. For the latter symmetry, when the external pressure p is zero, elastic properties can be described by just three elastic
constants Cxxxx, Cxxyy and Cxyxy [62]. However, when p 6= 0, it is more convenient [22, 45] to use other elastic constants, Bijkl , obtained by expanding the
free enthalpy (Gibbs free energy), G, with respect to the components of the
Lagrange strain tensor, ε:
∆G/Vref = ∆(F + pV )/Vref =
1X
Bijkl εij εkl .
2 ijkl
(3)
F in the above equation is the (Helmholtz) free energy and Vref is the volume of
the deformation-free (reference) system. The elastic constants Bijkl are simply
related with the constants Cijkl [63, 64]
Bijkl = Cijkl − p (δik δjl + δil δjk − δij δkl ) ,
(4)
where δij is the Kronecker delta, equal to 1 for i = j and zero otherwise.
In the case of cubic symmetry, considered in this work, only the following
elastic constants are of interest (the Voigt notation is used below [63]):
B11 = C11 − p ,
B12 = C12 + p ,
B66 = C66 − p ,
(5)
(6)
(7)
and the free enthalpy expansion (3) takes the form:
1
∆G/Vref = B11 ε2xx + ε2yy + ε2zz
2
+ B12 (εxx εyy + εyy εzz + εzz εxx ) + 2B66 ε2xy + ε2yz + ε2zx .
(8)
For cubic systems, instead of constants defined by Eq. (5)-(7), the below elastic
moduli are also often used:
5
1
(B11 + 2B12 ) ,
3
µ1 = B66 ,
1
µ2 = (B11 − B12 ) ,
2
(9)
B=
(10)
(11)
where B is the bulk modulus, and the µi are the shear moduli. In the following,
both sets of the quantities will be used while presenting and discussing the
results.
The Poisson’s ratio is defined as the negative ratio of transverse to longitudinal
strain when only the longitudinal component of the stress tensor is infinitesimally changed. For anisotropic systems it depends, in general, on both the
longitudinal and the transverse direction. For cubic symmetry, the Poisson’s
ratio along two high-symmetry longitudinal directions ([100] and [111]) does
not depend on the choice of the transverse direction [10] and can be expressed
by the elastic constants, Bij , in the form [65]:
B12
1 3B − 2µ2
=
,
B11 + B12
2 3B + µ2
B11 + 2B12 − 2B66
1 3B − 2µ1
ν[111] =
=
.
2 (B11 + 2B12 + B66 )
2 3B + µ1
ν[100] =
(12)
(13)
In this paper, the Poisson’s ratio along the longitudinal direction [110] and
two transverse directions [11̄0] and [001] was also computed:
2
2
B11
+ B11 B12 − 2B12
− 2B11 B66
2
2
B11 + B11 B12 − 2B12 + 2B11 B66
3B(3µ2 − µ1 ) − 4µ1 µ2
,
=
3B (3µ2 + µ1 ) + 4µ1 µ2
(14)
4B12 B66
2
+ B11 B12 − 2B12
+ 2B11 B66
6Bµ1 − 4µ1 µ2
=
.
3B (3µ2 + µ1 ) + 4µ1 µ2
(15)
ν[110][11̄0] =
ν[110][001] =
4
2
B11
Numerical determination of the elastic properties
In order to obtain the elastic constants, a sample was considered which, at
equilibrium, is described by three vectors a0 , b0 and c0 parallel to the unit
cell edges of the f.c.c. lattice. Components of those vectors form columns of a
reference box matrix [66, 67] H. Applying a uniform deformation that transforms the sample into a parallelepiped described by the box matrix h one can
6
write the Lagrange strain tensor as [66, 67]:
ε = H′
−1
h′ h H−1 − I /2 ,
(16)
where h′ is the transposed matrix h and I is the unit matrix. Differentiating
the free energy F with respect to the strain tensor components, one obtains
all the elastic constants and pressure. Is is worth to note, that in the case of
zero temperature (T = 0◦ K), the free energy is just equal to the potential
energy of the system.
After determining equilibrium positions and orientations of all the molecules
in the reference state for a given δ, the following deformations were applied:
hαβ = (1 + ξα ) Hαβ δαβ ,
(17)
for which new equilibrium positions and orientations were determined. In (17)
α, β indicate directions, and ξα is a small real number. The relations below
define the elastic constants and pressure:
∂2F
∂ξα ∂ξβ
∂F
∂ξα
ξα =0
∂2F
∂ξα2
ξα =0
ξα =ξβ =0, α6=β
= −p ,
(18)
= B11 ,
(19)
= B12 − p .
(20)
The B66 constant was obtained by applying the deformation of the form:
hαα = Hαα ,
hαβ = Hαβ + ζαβ Hαα ,
(21)
for which new equilibrium positions and orientations were determined. In (21)
ζαβ = ζβα is a small real number. For Hxx = Hyy = Hzz this leads to:
∂2F
2
∂ζαβ
5
= 6 (2B66 + p) .
(22)
ζαβ =0
Results
In the Figs. 2 (a) and 2 (b) values of the pressure and the bulk modulus
are shown, respectively. Is is clearly visible that increasing polydispersity and
7
1025
1025
δ
δ
δ
δ
δ
δ
1020
=
=
=
=
=
=
0
0.001
0.003
0.007
0.01
0.03
δ
δ
δ
δ
δ
δ
1020
=
=
=
=
=
=
0
0.001
0.003
0.007
0.01
0.03
1015
p
B
1015
1010
1010
105
105
100
100
0.001
0.01
0.001
0.1
0.01
a)
1/n
0.1
b)
1/n
Fig. 2. The values of (a) pressure and (b) bulk modulus B for polydisperse dimers,
plotted against the softness parameter (1/n). The thick solid line represents values
for the system with δ = 0.
increasing n leads to increase of those quantities as it was earlier observed in
other systems [47–50, 52].
The observed behavior can be qualitatively understood by noticing that dense
packing of identical molecules in the hard body limit requires less space than
dense packing of polydisperse molecules of the same average size [46]. As all
the considered systems are studied at the same volume, their densities relative
to close packing grow with increasing polydispersity. Increase of the relative
density implies an increase of the pressure and the bulk modulus. Obviously,
the latter quantities show the largest increase for the largest n, when the
interaction potential is the steepest one.
In Fig. 3 the elastic constants Bij are shown. It can be seen that all elastic
constants also increase with increasing polydispersity and increasing vaule of
n. Presented behavior of elastic constants is also similar to that observed for
other 2D and 3D systems [47–50, 52] and can be qualitatively explained by
the close packing argument used above.
1025
1025
δ
δ
δ
δ
δ
δ
1020
=
=
=
=
=
=
0
0.001
0.003
0.007
0.01
0.03
1020
1015
=
=
=
=
=
=
0
0.001
0.003
0.007
0.01
0.03
105
1010
105
100
0.01
1/n
0.1
1010
100
0.001
a)
0
0.001
0.003
0.007
0.01
0.03
105
100
0.001
=
=
=
=
=
=
1015
B66
1010
δ
δ
δ
δ
δ
δ
1020
1015
B12
B11
1025
δ
δ
δ
δ
δ
δ
0.01
1/n
0.1
0.001
b)
0.01
1/n
0.1
c)
Fig. 3. The values of elastic constants (a) B11 , (b) B12 and (c) B66 . plotted against
the softness parameter (1/n). The thick solid line represents values for the system
with δ = 0.
8
In Fig. 4 plots of µi /B ratios are presented. As in the case of earlier studies [47–
50, 52] for both 2D and 3D, this ratios seem to tend to 0 for any nonzero
polydispersity, in the hard interactions limit (n → ∞).
1
1
δ
δ
δ
δ
δ
δ
0.8
=
=
=
=
=
=
0
0.001
0.003
0.007
0.01
0.03
0.8
=
=
=
=
=
=
0
0.001
0.003
0.007
0.01
0.03
0.6
µ2 /B
0.6
µ1 /B
δ
δ
δ
δ
δ
δ
0.4
0.4
0.2
0.2
0
0
0.001
0.01
1/n
0.001
0.1
a)
0.01
1/n
0.1
b)
Fig. 4. The values of µ/B ratio for (a) µ1 and (b) µ2 , plotted against the softness
parameter (1/n). Thick solid line represents values measured for the system with
δ = 0.
According to the formulae (12) and (13) the mentioned above asymptotics of
µi /B implies that the system behaves like rubber (νrubber = 0.5) when a small
stress is applied along the directions [100] and [111] in the limit n → ∞. This
is shown in Fig. 5, where the plots of Poisson’s ratio are presented. It can be
seen there that, typically, presence of any nonzero polydispersity causes an
increase of the Poisson’s ratio in most of directions corresponding to the high
symmetry axes. It is worth to notice that in the cubic system studied, the
Poisson’s ratio ν[110][001] exceeds the maximum value 1/2 allowed for isotropic
materials. One should also stress that a pair of exceptional directions (see
Fig. 5(d)) exists in the system for which the Poisson’s ratio decreases with
increasing polydispersity. This occurs when the direction of the deformation
is [110] and the response of the system is measured in the direction [11̄0]. For
this pair of directions, the Poisson’s ratio reaches negative values both for small
and large n limits! The decrease for large n is, however, much more significant.
This new behavior of Poisson’s ratio is opposite to what was discovered in
the soft polydisperse spheres system [52], where the most negative values of
Poisson’s ratio were observed for small n only.
The observed dependence of ν[110][11̄0] , surprising in the context of results
obtained for various 2D isotropic systems including the DC phase of 2D
dimers [49], is of interest from the point of view of production of partial auxetics. As shown in Fig. 5(d), solid structure of dimers obtained by ‘gluing’
together pairs of polydisperse soft spheres arranged in the f.c.c. lattice will
show the Poisson’s ratio as low as ν[110][11̄0] = −0.8.
9
0.4
0.4
0.3
0.3
ν[111]
0.5
ν[100]
0.5
0.2
0.2
0.1
δ
δ
δ
δ
δ
δ
=
=
=
=
=
=
0
0.001
0.003
0.007
0.01
0.03
δ
δ
δ
δ
δ
δ
0.1
=
=
=
=
=
=
0
0.001
0.003
0.007
0.01
0.03
0
0
0.001
0.01
a)
1/n
2
δ
δ
δ
δ
δ
δ
1.5
0.001
0.1
=
=
=
=
=
=
0.01
0.1
b)
1/n
0
0.001
0.003
0.007
0.01
0.03
0.4
0.2
ν[110][11̄0]
ν[110][001]
0
1
0.5
−0.2
−0.4
δ
δ
δ
δ
δ
δ
−0.6
−0.8
0
−1
0.001
0.01
1/n
0.1
c)
0.001
0.01
1/n
=
=
=
=
=
=
0
0.001
0.003
0.007
0.01
0.03
0.1
d)
Fig. 5. Poisson’s ratio plotted against the softness parameter (1/n). Plots
(a), (b) show the Poisson’s ratio measured perpendicularly to the direction of the
high-symmetry axis (a) [100] and (b) [111]. Plots (c), (d) show Poisson’s ratio perpendicularly to the direction [110] in the direction (c) [001] and (d) [11̄0], respectively.
6
Conclusions
The elastic properties of polydisperse soft dimer system in three dimensions
were studied by computer simulations in the static (zero temperature) case.
Disorder was introduced to the DC f.c.c. structure in the form of atomic size
dispersion with polydispersity parameter δ. The influence of polydispersity on
elastic properties of 3D DC dimers system was determined. It was found that
increasing polydispersity and hardness of the interaction potential causes a
significant increase of elastic constants and pressure.
The simulations have shown that Poisson’s ratio of the dimer system measured
in [100] and [111] directions increases to its maximum positive value, 1/2, in
the presence of any nonzero polydispersity and sufficiently hard interaction
10
potential. An increase of Poisson’s ratio was observed also in other directions.
Such a behavior was earlier observed in the case of 2D [47–50] and 3D [52]
systems and seems to be a typical result.
It is worth to stress, however, that for the longitudinal direction [110] and
transverse direction [11̄0] the Poisson’s ratio decreases reaching negative values
along with increasing polydispersity value both for large and small n. This
unusual phenomenon is of interest from the point of view of manufacturing
stable auxetics of cubic symmetry.
The questions whether the effect of auxeticity enhancement by polydispersity
is characteristic for the f.c.c. lattice only or if it can be observed for elastic
symmetries other than the f.c.c. one, e.g. for packing of spheres in the h.c.p.
lattice or stacking fault lattices, will be a subject of a separate paper. Studies
of elastic properties of other molecular shapes are in progres.
Acknowledgements
This work was partially supported by the grants NN202 070 32/1512 and
NN202 261438 MNiSzW. We are grateful to Dr. Mikolaj Kowalik for preparing
the samples of the DC phase of the dimers, used in the simulations. Most of the
simulations were carried out at the Poznań Supercomputing and Networking
Center (PCSS).
References
[1] R. S. Lakes. Foam structures with a negative Poisson’s ratio. Science,
235:1038–1040, 1987.
[2] K. E. Evans. Auxetic polymers: a new range of materials. Endeavour,
15:170–174, 1991.
[3] R. S. Lakes. Advances in negative Poisson’s ratio materials. Advanced
Materials, 5:293–296, 1993.
[4] R. H. Baughman, J. M. Shacklette, A. A. Zakhidov, and S. Stafstrom.
Negative Poisson’s ratios as a common feature of cubic metals. Nature,
392:362–365, 1998.
[5] K. E. Evans and K. L. Alderson. Auxetic materials: the positive side of
being negative. Engineering Science and Education Journal, 9:148–154,
2000.
[6] R. H. Baughman. Avoiding the shrink. Nature, 425:667, 2003.
[7] C. Remillat, F. Scarpa, and K. W. Wojciechowski. Preface. Physica
Status Solidi (b), 246:2007–2009, 2009. See also references therein and
other papers in that issue.
11
[8] R. S. Lakes, web page:. http://silver.neep.wisc.edu/∼lakes/Poisson.html.
[9] G. Milton. Composite materials with Poisson’s ratios close to -1. J. Mech.
Phys. Solids, 40:1105–1137, 1992.
[10] K. W. Wojciechowski. Poisson’s ratio of anisotropic systems. Comp.
Meth. Sci. Technol., 11:73–79, 2005.
[11] T. C. T. Ting and D. M. Barnett. Negative Poisson’s ratios in anisotropic
linear elastic media. Journal of Applied Mechanics - Transactions of the
ASME, 72:929–931, 2005.
[12] A. C. Branka, D. M. Heyes, and K. W. Wojciechowski. Auxeticity of
cubic materials. Physica Status Solidi (b), 246:2063–2071, 2009.
[13] R. F. Almgren. An isotropic three-dimensional structure with Poisson’s
ratio =-1. Journal of Elasticity, 15:427–430, 1985.
[14] K. W. Wojciechowski. Constant thermodynamic tension Monte Carlo
studies of elastic properties of a two-dimensional system of hard cyclic
hexamers. Molecular Physics, 61:1247–1258, 1987.
[15] R. J. Bathurst and L. Rothenburg. Note on a random isotropic granular
material with negative Poisson’s ratio. International Journal of Engineering Science, 26:373–383, 1988.
[16] K. W. Wojciechowski. Two-dimensional isotropic model with a negative
Poisson ratio. Physics Letters A, 137:60–64, 1989.
[17] K. W. Wojciechowski and A. C. Brańka. Negative Poisson ratio in a
two-dimensional isotropic solid. Phys. Rev. A, 40:7222–7225, 1989.
[18] K. E. Evans, M. A. Nkansah, I. J. Hutchinson, and S. C. Rogers. Molecular network design. Nature, 353:124, 1991.
[19] L. Rothenburg, A. A. Berlin, and R. J. Bathurst. Microstructure of
isotropic materials with negative Poisson’s ratio. Nature, 354:470–472,
1991.
[20] G. Wei. J. Chem. Phys., 96:3226, 1992.
[21] D. H. Boal, U. Seifert, and J. C. Shillcock. Negative Poisson ratio in
two-dimensional networks under tension. Phys. Rev. E, 48:4274–4283,
1993.
[22] K. W. Wojciechowski. Negative Poisson ratios at negative pressures.
Molecular Physics Reports, 10:129–136, 1995.
[23] D. Prall and R. S. Lakes. Properties of a chiral honeycomb with a Poisson’s ratio of -1. Int. J. Mech. Sci., 39:305, 1997.
[24] G. Y. Wei and S. F. Edwards. Phys. Rev. E, 58:6173, 1998.
[25] R. H. Baughman, S. O. Dantas, S. Stafstrom, A. A. Zakhidov, T. B.
Mitchell, and D. H. E. Dubin. Negative Poisson’s ratios for extreme
states of matter. Science, 288:2018–2022, 2000.
[26] J. N. Grima and K. E. Evans. Auxetic behavior from rotating squares.
J. Mater. Sci. Lett., 19:1563–1565, 2000.
[27] Y. Ishibashi and M. Iwata. A microscopic model of a negative Poisson’s
ratio in some crystals. J. Phys. Soc. Japan, 69:2702–2703, 2000.
[28] H. Kimizuka, H. Kaburaki, and Y. Kogure. Mechanism for negative
Poisson ratios over the a-b transition of cristobalite, sio2: A molecular12
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
dynamics study. Phys. Rev. Lett., 84:5548–5551, 2000.
C. W. Smith, J. N. Grima, and K. E. Evans. A novel mechanism for generating auxetic behaviour in reticulated foams: Missing rib foam model.
Acta Materialia, 48:4349–4356, 2000.
M. Bowick, A. Cacciuto, G. Thorleifsson, and Travesset. Universal negative Poisson ratio of self-avoiding fixed-connectivity membranes. Phys.
Rev. Lett., 87:148103, 2001.
A. A. Vasiliev, S. V. Dmitriev, Y. Ishibashi, and T. Shigenari. Elastic
properties of a two-dimensional model of crystals containing particles
with rotational degrees of freedom. Phys. Rev. E, 65:094101, 2002.
K. W. Wojciechowski. Non-chiral, molecular model of negative Poisson’s
ratio in two dimensions. Journal of Physics A: Math Gen., 36:11765–
11778, 2003.
X. Xing and L. Radzihovsky. Universal elasticity and fluctuations of
nematic gels. Phys. Rev. Lett., 90:168301, 2003.
K. W. Wojciechowski, K. V. Tretiakov, and M. Kowalik. Elastic properties of dense solid phases of hard cyclic pentamers and heptamers in two
dimensions. Physical Review E, 67:036121, 2003.
P. V. Pikhitsa. Regular network of contacting cylinders with implications
for materials with negative Poisson ratios. Phys. Rev. Lett., 93:155595,
2004.
J. N. Grima, V. Zammit, R. Gatt, A. Alderson, and K. E. Evans. Auxetic behaviour from rotating semi-rigid units. Physica Status Solidi (b),
244:866–882, 2007.
D. Attard and J. N. Grima. Auxetic behaviour from rotating rhombi.
Physica Status Solidi (b), 245:2395–2404, 2008.
J. N. Grima, R. Gatt, and P. S. Farrugia. On the properties of auxetic
meta-tetrachiral structures. Physica Status Solidi (b), 245:511–520, 2008.
L. J. Hall, V. R. Coluci, D. S. Galvao, M. E. Koslov, M. Zhang, S. O.
Dantas, and R. H. Baughman. Sign change of Poisson’s ratio for carbon
nanotube sheets. Science, 320:504–507, 2008.
M. C. Rechtsman, F. H. Stillinger, and S. Torquato. Negative Poisson’s
ratio materials via isotropic interactions. Phys. Rev. Lett., 101:085501,
2008.
R. Lakes and K. W. Wojciechowski. Negative compressibility, negative
Poisson’s ratio, and stability. Physica Status Solidi b, 245:545–551, 2008.
D. Attard, E. Manicaro, and J. N. Grima. On rotating rigid parallelograms and their potential for exhibiting auxetic behaviour. Physica Status
Solidi (b), 246:2033–2044, 2009.
D. Attard, E. Manicaro, R. Gatt, and J. N. Grima. On the properties of
auxetic rotating stretching squares. Physica Status Solidi (b), 246:2045–
2054, 2009.
P. V. Pikhitsa, M. Choi, H. J. Kim, and S. H. Ahn. Auxetic lattice of
multipods. Physica Status Solidi (b), 246:2098–2101, 2009.
K. W. Wojciechowski and K. V. Tretiakov. Computer simulation of elastic
13
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[59]
properties of solids under pressure. Computer Physics Communications,
121–122:528–530, 1999.
K. V. Tretiakov and K. W. Wojciechowski. Monte Carlo simulations of
two-dimensional hard body systems with extreme values of the Poisson’s
ratio. Physica Status Solidi (b), 242:730–741, 2005.
J. W. Narojczyk and K. W. Wojciechowski. Elastic properties of twodimensional soft polydisperse trimers at zero temperature. Physica Status
Solidi (b), 244:943–954, 2007.
J. W. Narojczyk and K. W. Wojciechowski. Elastic properties of twodimensional soft discs of various diameters at zero temperature. Journal
of Non-Crystalline Solids, 352:4292–4298, 2006.
J. W. Narojczyk and K. W. Wojciechowski. Elasticity of periodic and
aperiodic structures of polydisperse dimers in two dimensions at zero
temperature. Physica Status Solidi (b), 245:2463–2468, 2008.
J. W. Narojczyk, A. Alderson, A.R. Imre, F. Scarpa, and K. W. Wojciechowski. Negative Poisson’s ratio behavior in the planar model of
asymmetric trimers at zero temperature. Journal of Non-Crystalline
Solids, 354:4242–4248, 2008.
M. Kowalik and K. W. Wojciechowski. Poisson’s ratio of degenerate crystalline phases of three-dimensional hard dimers and hard cyclic trimers.
Physica Status Solidi (b), 242:626–631, 2005.
J. W. Narojczyk and K. W. Wojciechowski. Elastic properties of the fcc
crystals of soft spheres with size dispersion at zero temperature. Physica
Status Solidi (b), 245:606–613, 2008.
W. G. Hoover, D. A. Young, and R. Groover. Statistical mechanics of
phase diagrams. I. Inverse power potentials and the close-packed to bodycentered cubic transition. J. Chem. Phys., 56:2207–2210, 1971.
K. W. Wojciechowski, D. Frenkel, and A. C. Brańka. Nonperiodic solid
phase in a two-dimensional hard-dimer system. Phys. Rev. Let, 67:3168–
3171, 1991.
M. Kowalik. Computer simulations of the Poisson’s ratio for selected hard
core models. 2005. PhD Thesis (in Polish); see also references therein.
J. P. Hoogenboom, A. K. van Langen-Suurling, J. Romijn, and A. van
Blaaderen. Epitaxial growth of a colloidal hard-sphere hcp crystal and
the effects of epitaxial mismatch on crystal structure. Physical Review E,
69:051602, 2004.
V. N. Manoharan and D. J. Pine. Building materials by packing shperes.
MRS Bulletin, 29:91–95, 2004.
D. M. Heyes and A. C. Branka. Physical properties of soft repulsive
particle fluids. Physical Chemistry and Chemical Physics, 9:5570–5575,
2007.
A. M. Kalsin, M. Fialkowski, M. Paszewski, S. K. Smoukov, K. J. M.
Bishop, and B. A. Grzybowski. Electrostatic self-assembly of binary
nanoparticle crystals with a diamond-like lattice. Science, 312:420–424,
2006.
14
[60] A. M. Kalsin and B. A. Grzybowski. Controlling the growth of ionic
nanoparticle supracrystals. Nano Letters, 7:1018–1021, 2007.
[61] A. O. Pinchuk, A. M. Kalsin, B. Kowalczyk, G. C. Schatz, and B. A. Grzybowski. Modeling of electrodynamic interactions between metal nanoparticles aggregated by electrostatic interactions into closely-packed clusters.
Journal of Physical Chemistry C, 111:11816–11822, 2007.
[62] L. D. Landau and E. M. Lifshitz. Theory of Elasticity. Pergamon Press,
London, 1986.
[63] D. C. Wallace. Thermodynamics of Crystals. Wiley, New York, 1972.
[64] K. V. Tretiakov and K. W. Wojciechowski. Poisson’s ratio of the fcc hard
sphere crystal at high densities. J. Chem. Phys, 123:74509, 2005.
[65] M. Kowalik and K. W. Wojciechowski. Poisson’s ratio of orientationally
disordered hard dumbbell crystal in three dimensions. Journal of NonCrystalline Solids, 352:4269–4278, 2006.
[66] M. Parrinello and A. Rahman. Polymorphic transitions in single crystals:
A new molecular dynamics method. J. Appl. Phys., 52:7182–7190, 1981.
[67] M. Parrinello and A. Rahman. Strain fluctuations and elastic constants.
Journal of Chemical Physics., 76:2662–2666, 1982.
15
| 5cs.CE
|
1
Variation Evolving for Optimal Control
Computation, A Compact Way
Sheng ZHANG and Kai-Feng HE
(2017.08)
Abstract: A compact version of the Variation Evolving Method (VEM) is developed for the optimal control computation. It follows the
idea that originates from the continuous-time dynamics stability theory in the control field. The optimal solution is analogized to the
equilibrium point of a dynamic system and is anticipated to be obtained in an asymptotically evolving way. With the introduction of a
virtual dimension——the variation time, the Evolution Partial Differential Equation (EPDE), which describes the variation motion
towards the optimal solution, is deduced from the Optimal Control Problem (OCP), and the equivalent optimality conditions with no
employment of costates are established. In particular, it is found that theoretically the analytic feedback optimal control law does not
exist for general OCPs because the optimal control is related to the future state. Since the derived EPDE is suitable to be solved with the
semi-discrete method in the field of PDE numerical calculation, the resulting Initial-value Problems (IVPs) may be solved with mature
Ordinary Differential Equation (ODE) numerical integration methods.
Key words: Optimal control, dynamics stability, variation evolution, initial-value problem, costate-free optimality condition.
I. INTRODUCTION
Optimal control theory aims to determine the inputs to a dynamic system that optimize a specified performance index while
satisfying constraints on the motion of the system. It is closely related to engineering and has been widely studied [1]. Because of
the complexity, Optimal Control Problems (OCPs) are usually solved with numerical methods. Various numerical methods are
developed and generally they are divided into two classes, namely, the direct methods and the indirect methods [2]. The direct
methods discretize the control or/and state variables to obtain the Nonlinear Programming (NLP) problem, for example, the
widely-used direct shooting method [2] and the classic collocation method [3]. These methods are easy to apply, whereas the
results obtained are usually suboptimal [4], and the optimal may be infinitely approached. The indirect methods transform the OCP
to a Boundary-value Problem (BVP) through the optimality conditions. Typical methods of this type include the well-known
indirect shooting method [2] and the novel symplectic method [5]. Although be more precise, the indirect methods often suffer
from the significant numerical difficulty due to the ill-conditioning of the Hamiltonian dynamics, that is, the stability of costates
dynamics is adverse to that of the states dynamics [6]. The recent development, representatively the Pseudo-spectral (PS) method
[7], blends the two types of methods, as it unifies the NLP and the BVP in a dualization view [8]. Such methods inherit the
advantages of both types and blur their difference.
Theories in the control field often enlighten strategies for the optimal control computation, for example, the non-linear variable
transformation to reduce the variables [9]. Recently, a new Variation Evolving Method (VEM), which is enlightened by the states
The authors are with the Computational Aerodynamics Institution, China Aerodynamics Research and Development Center, Mianyang, 621000, China. (e-mail:
[email protected]).
2
evolution within the stable continuous-time dynamic system, is proposed for the optimal control computation [10]. The VEM also
synthesizes the direct and indirect methods, but from a new standpoint. The Partial Differential Equation (PDE), which describes
the evolution of variables towards the extremal solution, is derived from the viewpoint of variation motion in typical OCPs. Using
the well-known semi-discrete method in the field of PDE numerical calculation [11], the PDEs are transformed to the
finite-dimensional Initial-value Problems (IVPs) to be solved, with the mature Ordinary Differential Equation (ODE) integration
methods. Because the extremum is guaranteed be the equilibrium point of the deduced dynamic system, the optimal solution will
be gradually approached. However, in the work of Ref. [10], besides the states and the controls, the costates are also introduced,
which increases the complexity of the computation. In this paper, a compact version of the VEM that uses only the original
variables is developed. The corresponding variation dynamic evolution equations, which may be re-presented in the PDE
formulation, are derived.
Throughout the paper, our work is built upon the assumption that the solution for the optimization problem exists. We do not
describe the existing conditions for the purpose of brevity. Relevant researches such as the Filippov-Cesari theorem are
documented in [12]. In the following, first preliminaries that state the inspiration of the VEM are presented. Then the foundational
VEM bred under this idea is demonstrated for the unconstrained calculus-of-variations problem. Next the compact VEM to solve
the OCPs is established. During this course, the costate-free optimality conditions are derived, and it is proved equivalent to the
traditional conditions. Later illustrative examples are solved to verify the effectiveness of the method. Besides, comparison
between the evolution equation and that in Ref. [10] is presented at the end.
II. PRELIMINARIES
The VEM is a newly developed method for the optimal solutions. For a better understanding, its motivations are recalled. For a
continuous-time autonomous dynamic system like
x = f ( x )
where x ∈ \ n is the state, x =
(1)
dx
is its time derivative, and f : \ n → \ n is a vector function. Suppose that x̂ is a
dt
asymptotically stable equilibrium point of system (1) that satisfies f ( xˆ ) = 0 , then from any initial condition x (t ) t = 0 = x0 within
the stability domain D that contains x̂ , the state x will tend to x̂ over time t [13]. According to the Lyapunov theory, there is a
continuously differentiable function V : D → \ such that
i) V ( xˆ ) = 0 and V ( x ) > 0 in D / { xˆ } .
ii) V ( x ) ≤ 0
For example, maybe f ( x ) satisfies ( x − xˆ )T f ( x ) ≤ 0 , and then a feasible Lyapunov function can be constructed as
V=
1
( x − xˆ )T ( x − xˆ )
2
(2)
where the superscript “ T ” denotes the transpose operator. The dynamics given by f ( x ) determines that V ≤ 0 and x will
converge to the equilibrium x̂ . Fig. 1 sketches the trajectory of some state in the stable dynamic system and the corresponding
Lyapunov function value. No matter what the initial condition x0 is, as long as it falls into the stability domain D , the state x
will approaches the equilibrium x̂ gradually.
3
Fig. 1. The sketch for the state trajectory and the Lyapunov function value profile.
In the system dynamics theory, from the stable dynamics of state x , we may construct a monotonously decreasing function
V ( x ) , which will achieve its minimum when x reaches x̂ . Inspired by it, now we consider its inverse problem, that is, from a
performance index function to derive the dynamics that minimize this performance index. Consider the parameter optimization
problem with performance index
J = h(θ )
(3)
where θ is the optimization parameter vector and h : \ n → \ is a scalar function. To find the optimal value θ̂ that minimizes J ,
we make the analogy to the Lyapunov function and differentiate J , i.e., function h here, with respect to a virtual time τ , which is
used to describe the derived dynamics.
dJ dh
∂h dθ
=
= ( )T
dτ dτ
∂θ dτ
To guarantee that J decreases with respect to τ , i.e.,
(4)
δJ
≤ 0 , we may set
δτ
dθ
∂h
= −Kh
dτ
∂θ
(5)
where K h is a positive-definite matrix. Under this dynamics, h will decrease until it reaches a extremum, and θ will approaches
θ̂ , the equilibrium point of system (5), which satisfies
∂h
∂θ
= 0 . This equilibrium condition is exactly the first-order optimality
θ =θˆ
condition for the optimization problem (3).
A bolder idea further arises hereafter. If the optimization parameter θ can approach its optimal under the dynamics given by Eq.
(5), we can imagine a variable x (t ) might also evolve to the optimal solution to minimize some performance index within certain
dynamics. Fig. 2 illustrates the idea of the VEM in solving the OCP, and we formally introduce the virtual variation time τ , a new
dimension orthogonal to the normal time t , to describe the variation evolution process. Through the variation motion, the initial
guess of variable will evolve to the optimal solution. We will detail the implementation of the idea in the following.
4
Fig. 2. The illustration of the variable evolution along the variation time
τ
in the VEM.
III. THE FOUNDATIONAL VARIATION EVOLVING METHOD
Calculus-of-variations problems may be regarded as OCPs with integrator dynamics. To start with, the foundational VEM,
which was first demonstrated in Ref. [10], is again presented for the unconstrained calculus-of-variations problem defined as
Problem 1: For the following functional depending on variable vector y (t ) ∈ \ n
J = ∫ F ( y (t ), y (t ), t ) dt
tf
(6)
t0
where t ∈ \ is the time. The elements of y belong to C 2 [t0 , t f ] , which denotes the set of variables with continuous second-order
derivatives (indicated by the superscript). The function F : \ n × \ n × \ → \ and its first-order and second-order partial
dy
derivatives are continuous with respect to y , its time derivative y =
and t . t0 and t f are the fixed initial and terminal time,
dt
and the boundary conditions are prescribed as y (t0 ) = y0 and y(t f ) = y f . Find the extremal solution ŷ that minimizes J , i.e.
yˆ = arg min( J )
(7)
Follow the idea of dynamics evolution to reduce some performance index, we anticipate that any initial guess of y (t ) , whose
elements belong to C 2 [t0 , t f ] , will evolve to the minimum along the variation dimension. Like the decrease of a Lyapunov
function, if J in Eq. (6) decreases with respect to the variation time τ , i.e.,
δJ
≤ 0 , we may finally obtain the optimal solution.
δτ
Differentiating (6) with respect to τ (even τ does not explicitly exist) produces
t
δJ
=
δτ ∫t
0
f
δ y ⎞
⎛ Tδy
dt
+ Fy T
⎜ Fy
δτ
δτ ⎟⎠
⎝
t
δy
δy
= Fy
− Fy T
+
δτ t
δτ t ∫t
T
0
f
where the column vectors Fy =
0
f
T
⎛⎡
d
⎤ δy⎞
⎜ ⎢ Fy − ( Fy ) ⎥
⎟ dt
⎜⎣
dt
⎦ δτ ⎟⎠
⎝
(8)
∂F
∂F
and Fy =
are the shorthand notations of partial derivatives. From an initial guess y (t )
∂y
∂y
that satisfies the boundary conditions at t0 and t f , then by enforcing
δJ
≤ 0 , we may set that
δτ
δy
d
⎛
⎞
= − K ⎜ Fy − ( Fy ) ⎟ ，t ∈ (t0 , t f )
δτ
dt
⎝
⎠
(9)
5
where K is a n × n dimensional positive-definite matrix. The variation dynamic evolution equation (9) describes the variation
motion of y (t ) starting from y (t ) , and it is proved that the motion is directed to the extremum [10]. It drives the performance
index J to decrease until
δJ
δJ
= 0, this determines the optimal conditions, namely, the Euler-Lagrange
= 0 , and when
δτ
δτ
equation [14][15]
Fy −
d
( Fy ) = 0
dt
(10)
The variation dynamic evolution equation (9) may be considered from the view of PDE formulation, by replacing the variation
operation “ δ ” and the differential operator “ d ” with the partial differential operator “ ∂ ” as
∂y
∂
⎛
⎞
= − K ⎜ Fy − ( Fy ) ⎟
∂τ
∂t
⎝
⎠
(11)
For this PDE, its right function only depends on the time t . Thus it is suitable to be solved with the semi-discrete method in the
field of PDE numerical calculation. With the discretization along the normal time dimension, Eq. (11) is transformed to be IVPs
with finite states. Note that the resulting IVP is defined with respect to the variation time τ , not the normal time t . In the previous
work [10], a demonstrative example is solved to verify the result.
IV. THE COMPACT VARIATION EVOLVING METHOD
A. Problem definition
In this paper, we consider the following class of OCP that is defined as
Problem 2：Consider performance index of Bolza form
J = ϕ ( x (t f ), t f ) + ∫ L ( x (t ), u(t ), t ) dt
tf
t0
(12)
subject to the dynamic equation
x = f ( x,u, t )
(13)
where t ∈ \ is the time. x ∈ \ n is the state vector and its elements belong to C 2 [t0 , t f ] . u ∈ \ m is the control vector and its
elements belong to C1[t0 , t f ] . The function L : \ n × \ m × \ → \ and its first-order partial derivatives are continuous with
respect to x , u and t . The function ϕ : \ m × \ → \ and its first-order and second-order partial derivatives are continuous with
respect to x and t . The vector function f : \ n × \ m × \ → \ n and its first-order partial derivatives are continuous and Lipschitz
in x , u and t . The initial time t0 is fixed and the terminal time t f is free. The initial boundary conditions are prescribed as
x (t0 ) = x0
(14)
and the terminal states are free. Find the optimal solution ( xˆ , uˆ ) that minimizes J , i.e.
( xˆ , uˆ ) = arg min( J )
(15)
B. Derivation of variation dynamic evolution equations
In Ref. [10], the problem is addressed by constructing an equivalent unconstrained functional problem that has the same
extremum. This operation is practical but it introduces the costate vector, which has the same dimension as the state vector. To
avoid the resulting complexity, here we will address Problem 2 directly. Consider the problem within the feasible solution domain
6
D o , in which any solution satisfies Eqs. (13) and (14). First we transform the Bolza performance index to the equivalent Lagrange
type, i.e.
J =∫
tf
t0
(ϕ + ϕ
t
x
T
)
f ( x,u, t ) + L( x , u, t ) dt
(16)
where ϕt and ϕ x are the partial derivatives notated as before. Similarly we differentiate Eq. (16) with respect to the variation time
τ to obtain
δtf
t
δJ
= (ϕt + ϕ x T f + L)
+∫
t
t δτ
δτ
f
0
f
δu⎞
⎛
T
T
T
T δx
dt
+ (ϕ x T fu + Lu T )
⎜ (ϕtx + f ϕ xx + ϕ x f x + Lx )
δτ
δτ ⎟⎠
⎝
(17)
where ϕtx and ϕ xx are second-order partial derivatives in the form of (column) vector and matrix, and f x and f u are the Jacobi
matrixes. Different from the calculus-of-variations problems,
δx
δu
and
are related because the profiles of x are determined by
δτ
δτ
u . For the solutions in D o , they need to satisfies the following variation equation
δ x
δx
δu
= fx
+ fu
δτ
δτ
δτ
with the initial condition
(18)
δx
= 0 . Note that f x and f u are time-dependent matrixes linearized at the feasible solution x (t ) and
δτ t
0
u(t ) . Eq. (18) is linear with respect to the variables
δx
δu
and
, and has a zero initial value. Thus it satisfies the superposition
δτ
δτ
principle, and its solution may be explicitly expressed.
Lemma 1 [16]: For the linear system
x = A(t ) x + B (t )u
(19)
with the initial value x (t0 ) = 0 , where x ∈ \ n is the state vector, u ∈ \ m is the control vector, A(t ) and B (t ) are the right
dimensional coefficient matrixes, the solution of system (19) is
t
x (t ) = ∫ H (t , s ) u( s ) d s
t0
(20)
where is the n × m dimensional H (t , s) is the impulse response function that satisfies
⎧Φ (t , s) B ( s )
H (t , s) = ⎨
⎩0
t≥s
s <t
(21)
and Φ (t , s) is the n × n dimensional state transition matrix for the system from time point s to time point t .
According to Lemma 1, it may be derived that Eq. (18) has the solution
t
δx
δu
= ∫ H o (t , s )
(s) d s
t
δτ
δτ
(22)
0
where H o (t , s) is the impulse response function corresponding to the specific f x (t ) and f u (t ) . Substitute Eq. (22) into Eq. (17),
there is
7
δtf
t ⎧
δJ
δu
δ u⎫
⎛ t
⎞
( s ) d s ⎟ + (φ x T f u + Lu T ) ⎬ dt
= (ϕt + ϕ x T f + L)
+ ∫ ⎨ ϕtx T + f Tϕ xx + ϕ x T f x + Lx T ⎜ ∫ H o (t , s )
t
t
t
δτ
δτ
δτ
δτ ⎭
⎝
⎠
⎩
δtf
t
t ⎧
t ⎛
δu ⎫
δu⎞
( s ) ⎬ d s d t + ∫ ⎜ (ϕ x T f u + Lu T )
dt
= (ϕt + ϕ x T f + L)
+ ∫ ∫ ⎨ ϕtx T + f Tϕ xx + ϕ x T f x + Lx T H o (t , s )
t
t ⎩
t
t δτ
δτ
δτ ⎠⎟
⎭
⎝
f
0
f
(
)
(
f
0
f
0
0
)
(23)
f
0
By exchanging the order in the double integral, we may derive the following transformation as
δu ⎫
(s)⎬ d s d t
δτ
⎭
δu
T
T
T
T
( s) d s
ϕtx (t ) + f (t )ϕ xx (t ) + ϕ x (t ) f x (t ) + Lx (t ) H o (t , s) d t
δτ
∫t ∫t ⎨⎩(ϕtx
=∫
tf
t
0
0
tf
t0
⎧
(∫
tf
s
(
T
)
(t ) + f T (t )ϕ xx (t ) + ϕ x T (t ) f x (t ) + Lx T (t ) H o (t , s)
(24)
)
)
To make the result clear, the change of variable symbol t → σ and s → t , without changing the final result of Eq. (24), produce
tf
tf
δu
T
T
T
T
∫t0 ∫s ϕtx (t ) + f (t )ϕ xx (t ) + ϕ x (t ) f x (t ) + Lx (t ) H o (t , s) d t δτ (s) d s
(25)
tf
tf
δu
= ∫ ∫ ϕtx T (σ ) + f T (σ )ϕ xx (σ ) + ϕ x T (σ ) f x (σ ) + Lx T (σ ) H o (σ , t ) d σ
(t ) d t
(
(
t0
(
)
)
(
t
) δτ
)
Thus, Eq. (23) may be reformulated as
δtf
δJ
= (ϕt + ϕ x T f + L)
δτ
+∫
tf
t0
Now to achieve
{( ∫
tf
tf
t
(ϕ
tx
T
δτ
)
)
(σ ) + f (σ )ϕ xx (σ ) + ϕ x (σ ) f x (σ ) + Lx (σ ) H o (σ , t ) d σ + (ϕ x
T
T
T
T
δJ
≤ 0 , we may derive the variation dynamic evolution equation for the control u as
δτ
t
δu
= − K Lu + f u Tϕ x + ∫ H o T (σ , t ) Lx (σ ) + ϕtx (σ ) + ϕ xx T (σ ) f (σ ) + f x (σ )T ϕ x (σ ) d σ
t
δτ
{
(
(
f
}
δu
(t ) d t
fu + Lu )
δτ
T
)
)}
(26)
(27)
and for the terminal time as
δtf
δτ
(
= − kt f L + ϕt + ϕ x T f
)
(28)
tf
where K is the m × m dimensional positive-definite matrix and kt f is a positive constant. Presume that we already have a feasible
initial solution x (t ) and u (t ) , then Eqs. (22), (27) and (28) drive it to the one that satisfies
Lu + fu Tϕ x +
(∫
tf
t
(
)
)
H o T (σ , t ) Lx (σ ) + ϕtx (σ ) + ϕ xx T (σ ) f (σ ) + f x (σ )T ϕ x (σ ) d σ = 0
L(t f ) + φt (t f ) + ϕ x T (t f ) f (t f ) = 0
(29)
(30)
C. Equivalence to the classic optimality conditions
Actually, Eqs. (29) and (30) are the first-order optimality conditions for Problem 2 without the employment of costates. We will
show that they are equivalent to the traditional ones with costates [17]
λ + H x = λ + Lx + f x T λ = 0
H u = Lu + f u T λ = 0
(31)
(32)
and transversality conditions
λ(t f ) = ϕ x (t f )
(33)
H (t f ) + ϕt f = 0
(34)
8
where H = L + λT f is the Hamiltonian and λ is the costate vector.
Theorem 1: For Problem 2, the optimal conditions given by Eqs. (29) and (30) are equivalent to the optimality conditions given by
(31)-(34).
Proof: With Eq. (21), Eq. (29) may be reformulated as
Lu + fu Tϕ x + f u T
(∫
tf
t
(
)
)
Φo T (σ , t ) Lx (σ ) + ϕtx (σ ) + ϕ xx T (σ ) f (σ ) + f x (σ )T ϕ x (σ ) d σ = 0
where Φo (σ , t ) is the state transition matrix related with H o (σ , t ) . Define a quantity γ (t ) as
(
tf
(35)
)
γ (t ) = ϕ x (t ) + ∫ Φo T (σ , t ) Lx (σ ) + ϕtx (σ ) + ϕ xx T (σ ) f (σ ) + f x (σ )T ϕ x (σ ) d σ
t
(36)
Then Eq. (35) is simplified as
Lu + fu T γ = 0
(37)
γ (t f ) = φ x (t f )
(38)
Obviously, from Eq. (36), when t = t f , there is
Differentiate γ (t ) with respect to t . In the process, we will use the Leibniz rule [18]
d
dt
(∫
a (t )
b (t )
)
h(σ , t ) d σ = h ( a(t ), t )
a (t )
d
d
a(t ) − h ( b(t ), t ) b(t ) + ∫ ht (σ , t ) d σ
(t )
b
dt
dt
(39)
and the property of Φo (σ , t ) [16]
∂Φo (σ , t )
= −Φo (σ , t ) f x (t )
∂t
(40)
Φo (t , t ) = 1
(41)
where 1 is the n × n dimensional identity matrix. Then we have
(
)
(
)
tf
d
γ (t ) = ϕtx + ϕ xx T f − Lx + ϕtx + ϕ xx T f + f x Tϕ x − f x T ∫ Φo T (σ , t ) Lx (σ ) + ϕtx (σ ) + ϕ xx T (σ ) f (σ ) + f x (σ )T ϕ x (σ ) d σ
t
dt
(
tf
(
)
= − Lx − f x T ϕ x (t ) + ∫ Φo T (σ , t ) Lx (σ ) + ϕtx (σ ) + ϕ xx T (σ ) f (σ ) + f x (σ )T ϕ x (σ ) d σ
t
)
(42)
= − Lx − f x T γ (t )
From Eqs. (42) and (38), it is found that γ (t ) conforms to the same dynamics and boundary conditions as the costates λ(t ) . Thus,
we can conclude that γ (t ) = λ(t ) . Then Eq. (37) and Eq. (32) are identical. This means that Eq. (29) is equivalent to Eqs. (31), (32)
and (33).
On the other hand, with Eq. (33), it is readily to show that Eqs. (30) and (34) are same.
■
By investigating the optimality condition (29), it is found that the optimal control are related to the future state, and this means
the optimal feedback control law in the analytic form does not exists for general OCPs. However, optimal control aided with
numerical computation is still possible. After proving the equivalence of optimality conditions, now the variables evolving
direction using the VEM is easy to determine.
Theorem 2: Solving the IVP with respect to τ , defined by the variation dynamic evolution equations (22), (27) and (28), from a
feasible initial solution, when τ → +∞ , ( x, u) will satisfy the optimality conditions of Problem 2.
9
Proof: From a feasible initial solution, any evolution under the dynamics of Eq. (22) guarantees the feasibility of solution.
Substitute Eqs. (27) and (28) into Eq. (26), we have
τ → +∞ due to the asymptotical approach. When
δJ
δJ
≤ 0 . The functional J will decrease until
= 0 , which occurs when
δτ
δτ
δJ
=0, this determines the optimal conditions, namely, Eqs. (29) and (30).
δτ
■
Presume that we already have a feasible initial solution x (t ) and u (t ) that satisfy Eqs. (13) and (14), then theoretically the
infinite-dimensional variation dynamic evolution equations (22), (27) and (28) may be used to obtain the optimal solution that
minimizes Eq. (12). The corn during the computation is to solve H o (t , s ) which may be evaluated upon the current solution of
x (t ) and u(t ) .
D. Formulation of Evolution PDE
Similarly, we may use the partial differential operator “ ∂ ” and the differential operator “ d ” to reformulate the variation
dynamic equations to get the following Evolution PDE (EPDE) and Evolution Differential Equation (EDE) as
⎡
∂ ⎡ x⎤ ⎢
=⎢
∂τ ⎣⎢ u ⎦⎥ ⎢
− K Lu + f u Tϕ x +
⎢⎣
{
(∫
tf
t
∂u
(s) d s
∂τ
⎤
⎥
⎥
H o T (σ , t ) Lx (σ ) + ϕtx (σ ) + ϕ xx T (σ ) f (σ ) + f x (σ )T ϕ x (σ ) d σ ⎥
⎦⎥
t
∫t
H o (t , s )
0
(
dt f
dτ
)
(
= −kt f L + ϕt + ϕ x T f
)
)}
(43)
(44)
tf
Put into this perspective, the definite conditions are the initial guess of t f , i.e., t f
τ =0
= tf , and
⎡ x (t ,τ ) ⎤
⎡ x (t ) ⎤
=⎢
⎢ u(t ,τ ) ⎥
⎥
⎣
⎦ τ = 0 ⎣ u (t ) ⎦
(45)
where x (t ) and u (t ) are the initial feasible solution. Recall Fig. 2, Eqs. (43) and (44) realize the anticipated variable evolution
along the variation time τ . The initial conditions of x (t ,τ ) and u(t ,τ ) at τ = 0 belong to the feasible solution domain and their
value at τ = +∞ may be the optimal solution of the OCP. The right part of the EPDE (43) is also only a vector function of time t .
Thus we may apply the semi-discrete method to discretize it along the normal time dimension and further use ODE integration
methods to get the numerical solution.
This paper considers the OCPs defined with free terminal time, yet the results obtained are also applicable to the simpler case
with fixed terminal time. For those OCPs, the evolution equations regarding state variable x and control variable u are same,
while the equation regarding the terminal time t f is not necessary anymore.
V. ILLUSTRATIVE EXAMPLES
First a linear example taken from Xie [19] is considered.
Example 1: Consider the following dynamic system
x = Ax + bu
⎡x ⎤
⎡0 1 ⎤
⎡0⎤
where x = ⎢ 1 ⎥ , A = ⎢
, b = ⎢ ⎥ . Find the solution that minimizes the performance index
⎥
⎣0 0⎦
⎣1 ⎦
⎣ x2 ⎦
10
J=
(
)
1
1 tf
x (t f )T Fx (t f ) + ∫ x T Qx + Ru 2 dt
2
2 t0
⎡1⎤
with the initial boundary conditions x (t0 ) = ⎢ ⎥ , where the initial time t0 = 0 and the terminal time t f = 3 are fixed. The
⎣1⎦
⎡1 0 ⎤
⎡2 1⎤
1
weighted matrixes are F = ⎢
⎥ , Q = ⎢ 1 4 ⎥ and R = 2 .
0
2
⎣
⎦
⎣
⎦
In solving this example using the VEM, the EPDE derived is
t A ( t − s ) ∂u
⎡
⎤
∫t0 e b ∂τ (s) d s
⎥
∂ ⎡ x⎤ ⎢
⎥
=⎢
⎢
⎥
∂τ ⎣ u ⎦ ⎢
⎧
⎞ ⎫⎥
T
T ⎛ tf
T
A (σ −t ) T
−
+
+
+
+
+
K
Ru
e
u
σ
σ
σ
b
Fx
b
(
Q
FA
A
F
)
x
(
)
Fb
(
)
d
⎨
⎜
⎟⎬
⎢⎣
⎝ ∫t
⎠ ⎭⎥⎦
⎩
(
)(
)
with the one-dimensional matrix K = 2 × 10−2 . The definite conditions of the EPDE, i.e., the initial guess of the states x (t ) and the
control u (t ) , were obtained by numerical integration with control input u (t ) = 0 . Using the semi-discrete method, the time horizon
[t0 , t f ] was discretized uniformly, with 61 points. Thus, a dynamic system with 183 states was obtained and the OCP was
transformed to an IVP. The ODE integrator “ode45” in Matlab, with default relative error tolerance 1×10-3 and default absolute
error tolerance 1×10-6, was employed to solve the IVP. Even for this simple example, there is no analytic solution. For comparison,
we computed the optimal solution with GPOPS-II [20], a Radau PS method based OCP solver.
Figs. 3, 4 and 5 show the evolving process of x1 (t ) , x2 (t ) and u (t ) solutions to the optimal, respectively. At τ = 300s, the
numerical solutions are indistinguishable from the optimal, and this shows the effectiveness of the VEM. Fig. 6 plots the profile of
performance index value against the variation time. It declines rapidly at first and almost reaches the minimum when τ = 50s. Then
it keeps approaching the minimum monotonously.
4
3.5
The optimal solution
Numerical solutions with VEM
τ = 0s
3
x1(t)
2.5
τ = 0.4s
2
τ = 1.0s
1.5
τ = 20.7s
1
0.5
0
0
τ = 1.8s
τ = 300s
0.5
1
1.5
2
2.5
3
t
Fig. 3 The evolution of numerical solutions of x1 to the optimal solution.
11
The optimal solution
Numerical solutions with VEM
1.2
1
τ = 0s
x2(t)
0.8
τ = 0.4s
0.6
0.4
τ = 1.0s
0.2
τ = 1.8s
0
-0.2
τ = 20.7s
-0.4
τ = 300s
0
0.5
1
1.5
2
2.5
3
t
Fig. 4 The evolution of numerical solutions of
1
0
-1
x2
to the optimal solution.
τ = 0s
τ = 0.4s
τ = 1.8s
u (t)
τ = 20.7s
-2
-3
τ = 70.9s
-4
τ = 300s
The optimal solution
Numerical solutions with VEM
-5
-6
-0.5
0
0.5
1
1.5
2
2.5
3
t
Fig. 5 The evolution of numerical solutions of
u
to the optimal solution.
50
Minimum of J
Funcitonal value of numerical solution
40
40
20
J
J
30
20
0
0
10
0
0
50
100
150
τ (s)
200
2
4 6
τ (s)
250
300
Fig. 6 The approach to the minimum of performance index.
Now we consider a nonlinear example with free terminal time t f , the homing missile problem adapted from Hull [21].
Example 2: Consider the problem of a constant speed missile intercepting a constant speed target moving in a straight line. The
dynamic equations are
12
x = f ( x , u )
⎡ VT sin(θT ) − VM sin(θ M ) ⎤
⎡ x ⎤
⎢V cos(θ ) − V cos(θ ) ⎥
T
M
M ⎥
. x and y are the relative abscissa and ordinate, respectively. θ M is the
where x = ⎢⎢ y ⎥⎥ , and f = ⎢ T
⎢
⎥
u
⎢⎣θ M ⎥⎦
⎢
⎥
VM
⎣
⎦
azimuth of missile. θT =30 deg is the azimuth of the target. VM =1000 m/s is the constant speed of missile. VT =500 m/s is the
constant speed of the target. u is the missile normal acceleration. To intercept the target and penalize too large normal acceleration,
the performance index to be minimized is defined as
J=
⎡1× 10−2
⎢
where the weighted matrixes are F = ⎢ 0
⎢ 0
⎢⎣
1
1 tf
x (t f )T Fx (t f ) + ∫ Ru 2 dt
2
2 t0
0
2 × 10−2
0
0⎤
⎥
0 ⎥ and R = 5 × 10−4 . The initial boundary conditions are
0 ⎥⎥
⎦
⎡ x ⎤
⎡10000m ⎤
⎢ y ⎥
= ⎢⎢ 5000m ⎥⎥ and the terminal time t f is free.
⎢ ⎥
⎢⎣θ M ⎥⎦
⎢⎣ 0 deg ⎥⎦
t0 = 0
Before the computation, the states and the control variables were scaled to improve the numerical efficiency. In the specific form
of the EPDE (43) and the EDE (44), the parameters K and kt f were set to be 1.5 × 10−6 and 1× 10−4 , respectively. The definite
⎡ x (t ,τ ) ⎤
⎢
⎥
conditions, i.e., ⎢u (t ,τ ) ⎥
, were obtained by numerical integration at time horizon [0, 25]s with control input u (t ) = 0 . We also
⎢ t f (τ ) ⎥
⎣
⎦ τ =0
discretized the time horizon [t0 , t f ] uniformly, with 51 points. Thus, a large IVP with 205 states (including the terminal time) is
obtained. We still employed “ode45” in Matlab for the numerical integration. In the integrator setting, the default relative error
tolerance and the absolute error tolerance are 1×10-3 and 1×10-6, respectively. For comparison, the optimal solution is again
computed with GPOPS-II.
Fig. 7 gives the states curve in the x y relative coordinate plane, showing that the numerical results approach the optimal
solution over time. For the optimal solution, the missile will intercept the target with a fairly small position error. The control
solutions are plotted in Fig. 8, and the asymptotical approach of the numerical results are demonstrated. In Fig. 9, the terminal time
profile against the variation time τ is plotted. The result of t f oscillates sharply at first and then gradually approaches to the
optimal interception time, and it only changes slightly after τ = 40s. At τ = 300s, we compute that t f = 23.52s from the VEM,
same to the result from GPOPS-II.
13
12000
The optimal solution
Numerical solutions with VEM
10000
τ = 0s
y (m)
8000
τ = 0.6s
6000
4000
Initial relative coordintates
τ = 1.2s
τ = 300s
2000
τ = 2.1s
τ = 3.1s
0
-5000
Optimal terminal relative coordintates
0
5000
10000
x (m)
Fig. 7 The evolution of numerical solutions in the
xy
relative coordinate plane to the optimal solution.
70
The optimal solution
Numerical solutions with VEM
τ = 300s
60
50
u (t)
40
τ = 3.1s
τ = 2.1s
30
τ = 1.2s
20
10
τ = 0.6s
0
-10
0
τ = 0s
5
10
15
20
25
t
Fig. 8. The evolution of numerical solutions of
u
to the optimal solution.
25
Minimum decline time
Evolving profile of tf
24.8
24.6
25
tf (s)
24.2
f
t (s)
24.4
24
24.5
24
23.5
-2 0 2 4 6
23.8
τ (s)
23.6
23.4
0
50
100
150
τ (s)
200
250
300
Fig. 9 The evolution profile of t f to the optimal interception time.
14
VI. FURTHER COMMENTS
Through the substitution of operator, the variation dynamic evolution equations may be reformulated as PDE. Discussions
between the EPDE (43) from the compact VEM and the Augmented Evolution PDE (AEPDE) derived in [10], which was called
the ZS equation originally, is worthwhile. For convenience, the AEPDE is again presented here.
⎛
⎡⎛
∂λ ⎞ ⎤
⎡ ⎛ ∂x
⎞ ⎤⎞
⎜
⎢⎜ H x + ∂t ⎟ ⎥
⎢ ⎜ ∂t − f ⎟ ⎥ ⎟
⎠⎥
⎝
⎠ ⎥⎟
⎜
⎢⎝
∂y
∂ ⎢
= −2 K ⎜ H yy ⎢ ⎛
∂x ⎞ ⎥ − ⎢⎛ ∂λ
⎞⎥ ⎟
⎜
∂τ
⎢ ⎜ f − ⎟ ⎥ ∂t ⎢⎜ + H x ⎟ ⎥ ⎟
∂t ⎠ ⎥
⎠⎥ ⎟
⎜
⎢⎝
⎢⎝ ∂t
⎜
⎢⎣
⎥
⎢
⎥⎦ ⎟⎠
H
0
⎣
u
⎦
⎝
(46)
⎡ x⎤
where y = ⎢⎢ λ ⎥⎥ and H = L + λT f is the Hamiltonian.
⎢⎣ u ⎥⎦
Advantages of the EPDE over the AEPDE are listed as follows. i) The costates are not included and the evaluation on the
second-order derivatives of H is avoided, which will relieve the computation burden. ii) The solution of the EPDE is more
capable to reach a minimum, while the solution of AEPDE may halt at a saddle extremum. iii) Generally the EPDE requires the
integration, and the differentiation, as displayed in the AEPDE, may be avoided. This is advantageous to reduce the numerical error
in seeking solutions.
Regarding the disadvantages, i) currently the EPDE may only address OCPs with free terminal states. When the terminal states
are constrained, which is common in the OCP formulation, it is not directly applicable. One may penalize the terminal boundary
conditions in the performance index. However, this is not a satisfactory solution. ii) Moreover, it requires the initial solution to be
feasible, which is also inflexible.
About the computation, both the right parts of the PDEs are only the functions of time t . This makes them suitable to be solved
with the semi-discrete method in the field of PDE numerical calculation. Then they may be solved with the mature ODE integration
methods. During the numerical calculation, techniques such as the Legendre-Gauss (LG), Legendre-Gauss-Radau (LGR) and
Legendre-Gauss-Lobatto (LGL) discretization, which have stronger approximation capacity, may further improve the accuracy
and efficiency. The common deficiency that both equations face is the incapability to address the general state- and
control-constrained OCPs. These are interesting topics for future studies. Return to the principle, since complex numerical
computations are avoided, and the integration may be achieved with the simple analog circuit, the equations may provide a
promising way of optimal control in the engineering.
VII. CONCLUSION
A compact version of the Variation Evolving Method (VEM) for the optimal control computation is developed. It introduces no
extra variables in transforming the Optimal Control Problems (OCPs) to the Initial-value Problems (IVPs). The optimality
conditions that use no information of the costates are derived and they are proved to be equivalent to the traditional optimality
conditions. One important conclusion drawn by analyzing the optimality conditions is that generally the analytic optimal feedback
control law does not exists because the control is related to the future state. With the mature Ordinary Differential Equation (ODE)
integration methods, the solution of the resulting IVP may reach the minimum of the OCP. In particular, under the frame of IVPs
upon continuous-time dynamics, daunting task of searching reasonable step size and annoying oscillation phenomenon around the
extremum, as occurs in the discrete numerical method, are eliminated. However, currently the method can only solve the OCP with
free terminal states, and this restricts its application. Further studies will be carried out to address this issue.
15
REFERENCES
[1] H. J. Pesch and M. Plail, “The maximum principle of optimal control: A history of ingenious ideas and missed opportunities,” Control &
Cybernetics, vol. 38, no. 4, pp. 973-995, 2009.
[2] J. T. Betts, “Survey of numerical methods for trajectory optimization,” J. Guid. Control Dynam., vol. 21, no. 2, pp. 193-206, 1998.
[3] C. Hargraves and W. Paris, “Direct trajectory optimization using nonlinear programming and collocation,” J. Guid. Control Dynam., vol. 10,
no. 4, pp. 338-342, 1987.
[4] O. V. Stryk and R. Bulirsch, “Direct and indirect methods for trajectory optimization,” Ann. Oper. Res., vol. 37, no. 1, pp. 357-373, 1992.
[5] H. J. Peng, Q. Gao, Z. G. Wu, and W. X. Zhong, “Symplectic approaches for solving two-point boundary-value problems,” J. Guid. Control
Dynam., vol. 35, no. 2, pp. 653-658, 2012.
[6] A. V. Rao, “A survey of numerical methods for optimal control,” in Proc. AAS/AIAA Astrodynam. Specialist Conf., Pittsburgh, PA, 2009, AAS
Paper 09-334.
[7] D. Garg, M. A. Patterson, W. W. Hager, A. V. Rao, et al, A Unified framework for the numerical solution of optimal control problems using
pseudospectral methods,” Automatica, vol. 46, no. 11, pp. 1843-1851, 2010.
[8] I. M. Ross and F. Fahroo, “A perspective on methods for trajectory optimization,” in Proc. AIAA/AAS Astrodynam. Conf., Monterey, CA, 2002,
AIAA Paper No. 2002-4727.
[9] I. M. Ross and F. Fahroo, “Pseudospectral methods for optimal motion planning of differentially flat systems,” IEEE Trans. Autom. Control,
vol. 49, no. 8, pp. 1410-1413, 2004.
[10] S. Zhang, E. M. Yong, W. Q. Qian, and K. F. He, “A Variation Evolving Method for Optimal Control,” arXiv: 1703.10263 [cs.SY]
[11] H. X. Zhang and M. Y. Shen. Computational Fluid Dynamics — Fundamentals and Applications of Finite Difference Methods. Beijing, China:
National Defense Industry Press, 2003, pp. 76-78.
[12] R. F. Hartl, S. P. Sethi, and R. G. Vickson, “A survey of the maximum principles for optimal control problems with state constraint,” SIAM
Rev., vol. 37, no. 2, pp. 181-218, 1995.
[13] H. K. Khalil, Nonlinear Systems. New Jersey, USA: Prentice Hall, 2002, pp. 111-181.
[14] D. G. Wu, Variation Method. Beijing, China: Higher Education Press, 1987, pp. 12-68.
[15] K. W. Cassel, Variational Methods with Applications in Science and Engineering. New York: Cambridge University Press, 2013, pp. 28-81
[16] D.Z. Zhen. Linear System Theory. Beijing, China: Tsinghua University Press, 2002, pp. 85-134.
[17] A. E. Bryson and Y. C. Ho, Applied Optimal Control: Optimization, Estimation, and Control. Washington, DC, USA: Hemisphere, 1975, pp.
42-125.
[18] H. Wang, J. S. Luo, and J. M. Zhu. Advanced Mathematics. Changsha, China: National University of Defense Technology Press, 2000, pp.
203-210.
[19] X. S. Xie, Optimal Control Theory and Application. Beijing, China: Tsinghua Univ. Press, 1986, pp. 167-215.
[20] M. A. Patterson and A. V. Rao, “GPOPS-II: AMATLAB software for solving multiple-phase optimal control problems using hp-adaptive
Gaussian quadrature collocation methods and sparse nonlinear programming,” ACM Trans. Math. Software, vol. 41, no. 1, pp. 1-37, 2014.
[21] D. G. Hull, Optimal Control Theory for Applications, New York, USA: Springer, 2003, pp. 89-71.
| 3cs.SY
|
arXiv:1611.02466v1 [math.AC] 8 Nov 2016
Serre Dimension of Monoid Algebras
Manoj K. Keshari and Husney Parvez Sarwar
Department of Mathematics, IIT Bombay, Powai, Mumbai - 400076, India;
[email protected]; [email protected]
Abstract
Let R be a commutative Noetherian ring of dimension d, M a commutative cancellative torsionfree monoid of rank r and P a finitely generated projective R[M ]-module of rank t.
(1) Assume M is Φ-simplicial seminormal. (i) If M ∈ C(Φ), then Serre dim R[M ] ≤ d. (ii) If
r ≤ 3, then Serre dim R[int(M )] ≤ d.
(2) If M ⊂ Z2+ is a normal monoid of rank 2, then Serre dim R[M ] ≤ d.
(3) Assume M is c-divisible, d = 1 and t ≥ 3. Then P ∼
= ∧t P ⊕R[M ]t−1 .
(4) Assume R is a uni-branched affine algebra over an algebraically closed field and d = 1.
∼ ∧t P ⊕R[M ]t−1 .
Then P =
1
Introduction
Throughout rings are commutative Noetherian with 1; projective modules are finitely generated and of
constant rank; monoids are commutative cancellative torsion-free; Z+ denote the additive monoid of
non-negative integers.
Let A be a ring and P a projective A-module. An element p ∈ P is called unimodular, if there
exists φ ∈ Hom (P, A) such that φ(p) = 1. We say Serre dimension of A (denoted as Serre dim A)
is ≤ t, if every projective A-module of rank ≥ t + 1 has a unimodular element. Serre dimension of
A measures the surjective stabilization of the Grothendieck group K0 (A). Serre’s problem on the
freeness of projective k[X1 , . . . , Xn ]-modules, k a field, is equivalent to Serre dim k[X1 , . . . , Xn ] = 0.
After the solution of Serre’s problem by Quillen [16] and Suslin [21], many people worked on surjective stabilization of polynomial extension of a ring. Serre [20] proved Serre dim A ≤ dim A, Plumstead
[14] proved Serre dim A[X] ≤ dim A, Bhatwadekar-Roy [4] proved Serre dim A[X1 , . . . , Xn ] ≤ dim A
and Bhatwadekar-Lindel-Rao [3] proved Serre dim A[X1 , . . . , Xn , Y1±1 , . . . , Ym±1 ] ≤ dim A.
Anderson conjectured an analogue of Quillen-Suslin theorem for monoid algebras over a field which
was answered by Gubeladze [8] (see 1.1) as follows.
Theorem 1.1 Let k be a field and M a monoid. Then M is seminormal if and only if all projective
k[M ]-modules are free.
Gubeladze [11] asked the following
Question 1.2 Let M ⊂ Zr+ be a monoid of rank r with M ⊂ Zr+ an integral extension. Let R be a
ring of dimension d. Is Serre dim R[M ] ≤ d?
1
We answer Question 1.2 for some class of monoids. Recall that a finitely generated monoid M of
rank r is called Φ-simplicial if M can be embedded in Zr+ and the extension M ⊂ Zr+ is integral (see
[10]). A Φ-simplicial monoid is commutative, cancellative and torsion-free.
Definition 1.3 Let C(Φ) denote the class of seminormal Φ-simplicial monoids M ⊂ Zr+ of rank r
such that if Zr+ = {ts11 . . . tsrr | si ≥ 0}, then for 1 ≤ m ≤ r, Mm = M ∩ {ts11 . . . tsmm | si ≥ 0} satisfies
the following properties: Given a positive integer c, there exist integers ci > c for i = 1, . . . , m − 1
such that for any ring R, the automorphism η ∈ AutR[tm ] (R[t1 , . . . , tm ]) defined by η(ti ) = ti + tcmi for
i = 1, . . . , m − 1, restricts to an R-automorphism of R[Mm ]. It is easy to see that Mm ∈ C(Φ) and
rank Mm = m for 1 ≤ m ≤ r.
The following result (3.4, 3.8) answers Question 1.2 for monoids in C(Φ).
Theorem 1.4 Let M ⊂ Zr+ be a seminormal Φ-simplicial monoid of rank r and R a ring of dimension
d.
(1) If M ∈ C(Φ), then Serre dim R[M ] ≤ d.
(2) If r ≤ 3, then Serre dim R[int(M )] ≤ d, where int(M ) = int(R+ M ) ∩ Z3+ and int(R+ M ) is
the interior of the cone R+ M ⊂ R3 with respect to Euclidean topology.
The following result (3.6) follows from (1.4(1)). When R is a field, this result is due to Anderson
[1].
Theorem 1.5 Let R be a ring of dimension d and M ⊂ Z2+ a normal monoid of rank 2. Then Serre
dim R[M ] ≤ d.
The next result answers Question 1.2 partially for 1-dimensional rings (see 3.13, 3.16). The proof
uses the techniques of Kang [12], Roy [17] and Gubeladze’s [9]. Let us recall two definitions. (i) A
monoid M is called c-divisible, where c > 1 is an integer, if cX = m has a solution in M for all m ∈ M .
All c-divisible monoids are seminormal. (ii) Let R be a ring, R the integral closure of R and C the
conductor ideal of R ⊂ R. Then R is called uni-branched if for any p ∈ Spec R containing C, there is
a unique q ∈ Spec R such that q ∩ R = p.
Theorem 1.6 Let R be a ring of dimension 1, M a monoid and P a projective R[M ]-module of rank
r.
(i) If M is c-divisible and r ≥ 3, then P ∼
= ∧r P ⊕R[M ]r−1 .
(ii) If R is a uni-branched affine algebra over an algebraically closed field, then P ∼
= ∧r P ⊕R[M ]r−1 .
If R is a 1-dimensional anodal ring with finite seminormalization, then (1.6(ii)) is due to Sarwar
([18], Theorem 1.2). Note that if k is an algebraically closed field of characteristic 2, then node
k[X, Y ]/(X 2 − Y 2 − Y 3 ) is not anodal but is uni-branched, by Kang ([12], Example 2).
At the end, we give some applications to minimum number of generators of projective modules.
2
2
Preliminaries
Let A be a ring and Q an A-module. We say p ∈ Q is unimodular if the order ideal OQ (p) =
{φ(p) | φ ∈ Hom (Q, A)} equals A. The set of all unimodular elements in Q is denoted by Um(Q). We
write En (A) for the group generated by set of all n × n elementary matrices over A and Umn (A) for
Um(An ). We denote by Aut A (Q), the group of all A-automorphisms of Q.
For an ideal J of A, we denote by E(A ⊕ Q, J), the subgroup of Aut A (A ⊕ Q) generated by all the
aφ
and Γq = 1q id0Q with a ∈ J, φ ∈ Q∗ and q ∈ Q. Further, we shall
automorphisms ∆aφ = 10 id
Q
write E(A ⊕ Q) for E(A ⊕ Q, A). We denote by Um(A ⊕ Q, J) the set of all (a, q) ∈ Um(A ⊕ Q) with
a ∈ 1 + J and q ∈ JQ.
We state some results of Lindel [13] for later use.
Proposition 2.1 (Lindel [13], 1.1) Let A be a ring and Q an A-module. Let Qs be free of rank r for
some s ∈ A. Then there exist p1 , . . . , pr ∈ Q, φ1 , . . . , φr ∈ Q∗ and t ≥ 1 such that following holds:
2
(i) 0 :A s′ A = 0 :A s′ A, where s′ = st .
Pr
Pr
′
′ ∗
(ii) s Q ⊂ F and s Q ⊂ G, where F = i=1 Api ⊂ Q and G = i=1 Aφi ⊂ Q∗ .
(iii) the matrix (φi (pj ))1≤i,j≤r = diagonal (s′ , . . . , s′ ). We say F and G are s′ -dual submodules of
Q and Q∗ respectively.
Proposition 2.2 (Lindel [13], 1.2, 1.3) Let A be a ring and Q an A-module. Assume Qs is free of
rank r for some s ∈ A. Let F and G be s-dual submodules of Q and Q∗ respectively. Then
(i) for p ∈ Q, there exists q ∈ F such that ht (OQ (p + sq)As ) ≥ r.
(ii) If Q is projective A-module and p ∈ Um(Q/sQ), then there exists q ∈ F such that ht (OQ (p +
sq)) ≥ r.
Proposition 2.3 (Lindel [13], 1.6) Let Q be a module over a positively graded ring A = ⊕i≥0 Ai and
Qs be free for some s ∈ R = A0 . Let T ⊆ A be a multiplicatively closed set of homogeneous elements.
Let p ∈ Q be such that pT (1+sR) ∈ Um(QT (1+sR) ) and s ∈ rad(OQ (p) + A+ ), where A+ = ⊕i≥1 Ai .
Then there exists p′ ∈ p + sA+ Q such that p′T ∈ Um(QT ).
Proposition 2.4 (Lindel [13], 1.8) Under the assumptions of (2.3), let p ∈ Q be such that OQ (p) +
sA+ = A and A/OQ (p) is an integral extension of R/(R ∩ OQ (p)). Then there exists p′ ∈ Um(Q)
with p′ − p ∈ sA+ Q.
The following result is due to Amit Roy ([17], Proposition 3.4).
Proposition 2.5 Let A, B be two rings with f : A → B a ring homomorphism. Let s ∈ A be nonzerodivisor such that f (s) is a non-zerodivisor in B. Assume that we have the following cartesian
square.
A
As
f
fs
3
/B
/ Bf (s)
Further assume that SLr (Bf (s) ) = Er (Bf (s) ) for some r > 0. Let P and Q be two projective A-modules
of rank r such that (i) ∧r P ∼
= ∧r Q, (ii) Ps and Qs are free over As , (iii) P ⊗ A B ∼
= Q ⊗ A B and
∼
Q ⊗ A B has a unimodular element. Then P = Q.
Definition 2.6 (see [10], Section 6) Let R be a ring and M a Φ-simplicial monoid of rank r. Fix an
integral extension M ֒→ Zr+ . Let {t1 , . . . , tr } be a free basis of Zr+ . Then M can be thought of as a
monoid consisting of monomials in t1 , . . . , tr .
For x = ta1 1 . . . tar r and y = tb11 . . . tbrr in Zr+ , define x is lower than y if ai < bi for some i and
aj = bj for j > i. In particular, ti is lower than tj if and only if i < j.
For f ∈ R[M ], define the highest member H(f ) of f as am, where f = am + a1 m1 + . . . + ak mk
with m, mi ∈ M , a ∈ R \ {0}, ai ∈ R and each mi is strictly lower than m for 1 ≤ i ≤ k.
An element f ∈ R[Zr+ ] is called monic if H(f ) = atsr , where a ∈ R is a unit and s > 0. An element
f ∈ R[M ] is said to be monic if f is monic in R[Zr+ ] via the embedding R[M ] ֒→ R[Zr+ ].
sr−1
Define M0 to be the submonoid {ts11 . . . tr−1
| si ≥ 0} ∩ M of M . Clearly M0 is finitely generated
r−1
as M is finitely generated. Also M0 ֒→ Z+ is integral. Hence M0 is Φ-simplicial. Further, if M is
seminormal, then M0 is seminormal.
Grade R[M ] as R[M ] = R[M0 ] ⊕ A1 ⊕ A2 ⊕ . . ., where Ai is the R[M0 ]-module generated by the
sr−1 i
tr ∈ M . For an ideal I in R[M ], define its leading coefficient ideal λ(I) as
monomials ts11 . . . tr−1
{a ∈ R | ∃f ∈ I with H(f ) = am for some m ∈ M }.
Lemma 2.7 ([10], Lemma 6.5) Let R be a ring and M ⊂ Zr+ a Φ-simplicial monoid. If I ⊆ R[M ] is
an ideal, then ht (λ(I)) ≥ ht (I), where λ(I) is defined in (2.6).
3
Main Theorem
This section contains main results stated in the introduction. We also give some examples of monoids
in C(Φ).
3.1
Over C(Φ) class of monoids
Lemma 3.1 Let R be a ring and M ⊂ Zr+ a monoid in C(Φ) of rank r. Let f ∈ R[M ] ⊂ R[Zr+ ] =
R[t1 , . . . , tr ] with H(f ) = uts11 . . . tsrr for some unit u ∈ R. Then there exist η ∈ AutR (R[M ]) such
that η(f ) is a monic polynomial in tr .
Proof Since M ∈ C(Φ), we can choose positive integers c1 , . . . , cr−1 such that the automorphism
η ∈ AutR[tr ] R[t1 . . . , tr ] defined by η(ti ) = ti + tcri for i = 1, . . . , r − 1, restricts to an automorphism
of R[M ] and such that η(f ) is a monic polynomial in tr .
Lemma 3.2 Let R be a ring of dimension d and M ⊂ Zr+ a monoid in C(Φ) of rank r. Let P be
a projective R[M ]-module of rank > d. Write R[M ] = R[M0 ] ⊕ A1 ⊕ A2 . . ., as defined in (2.6) and
4
A+ = A1 ⊕ A2 ⊕ . . . an ideal of R[M ]. Assume that Ps is free for some s ∈ R and P/sA+ P has a
unimodular element. Then the natural map Um(P ) → Um(P/sA+ P ) is surjective. In particular, P
has a unimodular element.
Proof Write A = R[M ]. Since every unimodular element of P/sA+ P can be lifted to a unimodular
element of P1+sA+ , if s is nilpotent, then elements of 1+sA+ are units in A and we are done. Therefore,
assume that s is not nilpotent.
Let p ∈ P be such that p ∈ Um(P/sA+ P ). Then OP (p) + sA+ = A. Hence OP (p) contains an
element of 1 + sA+ . Choose g ∈ A+ such that 1 + sg ∈ OP (p). Applying (2.2) with sg in place of s, we
get q ∈ F ⊂ P such that ht (OP (p + sgq)) > d. Since p + sgq is a lift of p, replacing p by p + sgq, we
may assume that ht (OP (p)) > d. By (2.7), we get ht (λ(OP (p))) ≥ ht (OP (p)) > d. Since λ(OP (p))
is an ideal of R, we get 1 ∈ λ(OP (p)). Hence there exists f ∈ OP (p) such that the coefficient of H(f )
(highest member of f ) is a unit.
Suppose H(f ) = uts11 . . . tsrr with u a unit in R. Since M ∈ C(Φ), by (3.1), there exists α ∈
Aut R (R[M ]) such that α(f ) is monic in tr . Thus we may assume that OP (p) contains a monic
polynomial in tr . Hence A/OP (p) is an integral extension of R[M0 ]/(OP (p) ∩ R[M0 ]) and p ∈
Um(P/sA+ P ). By (2.4), there exists p′ ∈ Um(P ) such that p′ − p ∈ sA+ P . This means p′ ∈ Um(P )
is a lift of p. This proves the result.
Remark 3.3 In (3.2), we do not need the monoid M to be seminormal.
The next result proves (1.4(1)).
Theorem 3.4 Let R be a ring of dimension d and M a monoid in C(Φ) of rank r. If P is a projective
R[M ]-module of rank r′ ≥ d + 1, then P has a unimodular element. In other words, Serre dim
R[M ] ≤ d.
Proof We can assume that the ring is reduced with connected spectrum. If d = 0, then R is a field.
Since M is seminormal, projective R[M ]-modules are free, by (1.1). If r = 0, then M = 0 and we are
done by Serre [20]. Assume d, r ≥ 1 and use induction on d and r simultaneously.
If S is the set of all non-zerodivisor of R, then dim S −1 R = 0 and so S −1 P is free S −1 R[M ]-module
(d = 0 case). Choose s ∈ S such that Ps is free. Consider the ring R[M ]/(sR[M ]) = (R/sR)[M ].
Since dim R/sR = d − 1, by induction on d, Um(P/sP ) is non-empty.
Write R[M ] = R[M0 ] ⊕ A1 ⊕ A2 . . ., as defined in (2.6) and A+ = A1 ⊕ A2 ⊕ . . . an ideal of R[M ].
Note that M0 ∈ C(Φ) and rank M0 = r −1. Since R[M ]/A+ = R[M0 ], by induction on r, Um(P/A+ P )
is non-empty. Write A = R[M ] and consider the following fiber product diagram
A/(sA ∩ A+ )
/ A/sA
A/A+
/ A/(s, A+ )
5
If B = R/sR, then A/(s, A+ ) = B[M0 ]. Let u ∈ Um(P/A+ P ) and v ∈ Um(P/sP ). Let u and v
denote the images of u and v in P/(s, A+ )P . Write P/(s, A+ )P = B[M0 ] ⊕ P0 , where P0 is some
projective B[M0 ]-module of rank = r′ − 1. Note that dim(B) = d − 1 and u, v are two unimodular
elements in B[M0 ] ⊕ P0 .
Case 1. Assume rank(P0 ) ≥ max {2, d}. Then by ([6], Theorem 4.5), there exists σ ∈ E(B[M0 ] ⊕
P0 ) such that σ(u) = v. Lift σ to an element σ1 ∈ E(P/A+ P ) and write σ1 (u) = u1 ∈ Um(P/A+ P ).
Then images of u1 and v are same in P/(s, A+ )P . Patching u1 and v over P/(s, A+ )P in the above
fiber product diagram, we get an element p ∈ Um(P/(sA ∩ A+ )P ).
Note sA ∩ A+ = sA+ . We have Ps is free and P/sA+ P has a unimodular element. Use (3.2), to
conclude that P has a unimodular element.
Case 2. Now we consider the remaining case, namely d = 1 and rank(P ) = 2. Since B = R/sR
is 0 dimensional, projective modules over B[M0 ] and B[M ] are free, by (1.1). In particular, P/sP
and P/(s, A+ )P are free modules of rank 2 over the rings B[M ] and B[M0 ] respectively. Consider the
same fiber product diagram as above.
Since any two unimodular elements in Um2 (B[M0 ]) are connected by an element of GL2 (B[M0 ]).
Further B[M0 ] is a subring of B[M ] = A/sA. Hence the natural map GL2 (B[M ]) → GL2 (B[M0 ]) is
surjective. Hence any automorphism of P/(s, A+ )P can be lifted to an automorphism of P/sP . By
same argument as above, patching unimodular elements of P/sP and P/A+ P , we get a unimodular
element in P/(sA ∩ A+ )P . Since sA ∩ A+ = sA+ and P/sA+ P has a unimodular element, by (3.2),
P has a unimodular element. This completes the proof.
Example 3.5 (1) If M is a Φ-simplicial normal monoid of rank 2, then M ∈ C(Φ). To see this, by
([10], Lemma 1.3), M ∼
= (α1 , α2 ) ∩ Z2+ , where α1 = (a, b) and α2 = (0, c) and (α1 , α2 ) is the group
generated by α1 and α2 . It is easy to see that M ∼
= ((1, a1 ), (0, a2 )) ∩ Z2+ , where gcd(b, c) = g and
a1 = b/g, a2 = c/g. Hence M ∈ C(Φ).
(2) If M ⊂ Z2+ is a finitely generated rank 2 normal monoid, then it is easy to see that M is
Φ-simplicial. Hence M ∈ C(Φ) by (1).
(3) If M is a rank 3 normal quasi-truncated or truncated monoid (see [10], Definition 5.1), then
M ∈ C(Φ). To see this, by ([10], Lemma 6.6), M satisfies properties of (1.3). Further, M0 is a
Φ-simplicial normal monoid of rank 2. By (1), M0 ∈ C(Φ).
Corollary 3.6 Let R be a ring of dimension d and M ⊂ Z2+ a normal monoid of rank 2. Then Serre
dim R[M ] ≤ d.
Proof If M is finitely generated, then result follows from (3.5(2)) and (3.4).
If M is not finitely generated, then write M as a filtered union of finitely generated submonoids,
say M = ∪λ∈I Mλ . Since M is normal, the integral closure M λ of Mλ is contained in M . Hence
M = ∪λ∈I M λ . By ([5], Proposition 2.22), M λ is finitely generated. If P is a projective R[M ]-module,
then P is defined over R[M λ ] for some λ ∈ I as P is finitely generated. Now the result follows from
(3.5(2)) and (3.4).
6
The following result follows from (3.5(3)) and (3.4).
Corollary 3.7 Let R be a ring of dimension d and M a truncated or normal quasi-truncated monoid
of rank ≤ 3. Then Serre dim R[M ] ≤ d.
Now we prove (1.4(2)).
Proposition 3.8 Let R be a ring of dimension d and M a Φ-simplicial seminormal monoid of rank
≤ 3. Then Serre dim R[int(M )] ≤ d.
Proof Recall that int(M ) = int(R+ M ) ∩ Z3+ . Let P be a projective R[int(M )]-module of rank
≥ d + 1. Since M is seminormal, by ([5], Proposition 2.40), int(M ) = int(M ), where M is the
normalization of M . Since normalization of a finitely generated monoid is finitely generated (see [5],
Proposition 2.22), M is a Φ-simplicial normal monoid. By ([10], Theorem 3.1), int(M ) = int(M ) is a
filtered union of truncated (normal) monoids (see [10], Definition 2.2). Since P is finitely generated, we
get P is defined over R[N ], where N ⊂ int(M ) is a truncated monoid. By (3.7), Serre dim R[N ] ≤ d.
Hence P has a unimodular element. Therefore Serre dim R[int(M )] ≤ d.
Assumptions: In the following examples, R is a ring of dimension d, Monoid operations are written
multiplicatively and K(M ) denotes the group of fractions of monoid M .
P
Example 3.9 For n > 0, consider the monoid M ⊂ Zr+ generated by {ti11 ti22 . . . tirr | ij = n}. Then
M is a Φ-simplicial normal monoid. For integers ci = nki + 1, ki > 0 and i = 1, . . . , r − 1, consider
η ∈ Aut R[tr ] (R[t1 , . . . , tr ]) defined by ti 7→ ti + tcri for i = 1, . . . , r − 1.
ir−1 ir
c
A typical monomial in the expansion of η(ti11 . . . tr−1
tr ) = (t1 + tcr1 )i1 . . . (tr−1 + trr−1 )ir−1 tirr will
ir−1 −lr−1 cr−1 lr−1 ir
ir−1 −lr−1 l1 +...+lr−1 +ir n(k1 l1 +...+kr−1 lr−1 )
look like (t1i1 −l1 trc1 l1 ) . . . (tr−1
)tr = (t1i1 −l1 . . . tr−1
tr
)tr
tr
−1
which belong to M . So η(R[M ]) ⊂ R[M ]. Similarly, η (R[M ]) ⊂ R[M ]. Hence η restricts to an
R-automorphism of R[M ]. Therefore η satisfies the property of (1.3) for M . It is easy to see that
Mm = M ∩ {ts11 . . . tsmm |si ≥ 0} for 1 ≤ m ≤ r − 1 also satisfy this property. Hence M ∈ C(Φ). By
(3.4), Serre dim R[M ] ≤ d.
Example 3.10 Let M be a Φ-simplicial monoid generated by monomials t21 , t22 , t23 , t1 t3 , t2 t3 . For
integers cj = 2kj − 1 with kj > 1, consider the automorphism η ∈ Aut R[t3 ] (R[t1 , t2 , t3 ]) defined by
c
tj 7→ tj + t3j for j = 1, 2. Then it is easy to see that η restricts to an automorphism of R[M ].
We claim that M is seminormal but not normal. For this, let
z = (t23 )−1 (t1 t3 )(t2 t3 ) = t1 t2 ∈ K(M ) \ M, but z2 ∈ M,
showing that M is not normal. For seminormality, let
z = (t21 )α1 (t22 )α2 (t23 )α3 (t1 t3 )α4 (t2 t3 )α5 ∈ K(M ) with αi ∈ Z and z2 , z3 ∈ M.
7
We may assume that 0 ≤ α4 , α5 ≤ 1. Now z 2 ∈ M ⇒ α1 , α2 ≥ 0 and 2α3 + α4 + α5 ≥ 0. If α3 < 0,
1 +1 2α2 +1 3
then α4 = α5 = 1 and α3 = −1. In this case, z 3 = (t2α
t2
) ∈
/ M , a contradiction. Therefore
1
α3 ≥ 0 and z ∈ M . Hence M is seminormal. It is easy to see that M ∈ C(Φ). By (3.4), Serre dim
R[t21 , t22 , t23 , t1 t3 , t2 t3 ] ≤ d.
Remark 3.11 (1) Let R be a ring and P a projective R-module of rank ≥ 2. Let R be the
seminormalization of R. It follows from arguments in Bhatwadekar ([2], Lemma 3.1) that P ⊗ R R has
a unimodular element if and only if P has a unimodular element.
(2) Assume R is a ring of dimension d and M ∈ C(Φ). Let M be the seminormalization of M . If
M is in C(Φ), then Serre dim R[M ] ≤ max{1, d}, using ([2] and 3.4).
(3) Let (R, m, K) be a regular local ring of dimension d containing a field k such that either char
k = 0 or char k = p and tr-deg K/Fp ≥ 1. Let M be a seminormal monoid. Then, using Popescu
([15], Theorem 1) and Swan ([23], Theorem 1.2), we get Serre dim R[M ] = 0. If M is not seminormal,
then Serre dim R[M ] = 1 using ([11], [2] and [23]).
Example 3.12 For a monoid M , M denotes the seminormalization of M .
1. Let M ⊂ Z2+ be a Φ-simplicial monoid generated by tn1 , t1 t2 , tn2 , where n ∈ N. We claim that
M is normal. To see this, let z = ti1 tj2 = (tn1 )p (t1 t2 )q (tn2 )r ∈ K(M ) with p, q, r ∈ Z such that
z t ∈ M for some t > 0. Then i, j ≥ 0. We need to show that z ∈ M . We may assume that
0 ≤ q < n. Since i, j ≥ 0, we get p, r ≥ 0. Thus z ∈ M and M is normal. Hence, by (3.6), Serre
dim R[tn1 , t1 t2 , tn2 ] ≤ d.
2. The monoid M ⊂ Z2+ generated by t21 , t1 t22 , t22 is seminormal but not normal. For this, let
z = (t1 t22 )(t22 )−1 = t1 ∈ K(M ) \ M . Then z 2 ∈ M showing that M is not normal. For
seminormality, let z = (t21 )α (t1 t22 )β (t22 )γ ∈ K(M ) with α, β, γ ∈ Z be such that z 2 , z 3 ∈ M . We
may assume 0 ≤ β ≤ 1. If β = 0, then α, γ ≥ 0 and hence z ∈ M . If β = 1, then z 2 ∈ M implies
α ≥ 0 and γ + 1 ≥ 0. If γ = −1, then z 3 = (t1 )6α+3 ∈
/ M , a contradiction. Hence γ ≥ 0, proving
that z ∈ M and M is seminormal. It is easy to see that M ∈ C(Φ). Therefore, by (3.4), Serre
dim R[t21 , t1 t22 , t22 ] ≤ d.
3. Let M be a monoid generated by (t21 , t1 tj2 , t22 ), where j ≥ 3. Then M is not seminormal. For
2(j−2)
this, if z = (t1 tj2 )(t22 )−1 = t1 tj−2
∈ K(M ) \ M , then z 2 = t21 t2
and z 3 = (t21 )(t1 tj2 )(t2j−6
)
2
2
are in M , showing that M is not seminormal.
If j = 3, then observe that t1 t2 belongs to M . Since the monoid generated by t21 , t1 t2 , t22 is
normal, we get that M is generated by t21 , t1 t2 , t22 . Hence Serre dim R[M ] ≤ d by (1) above.
Observe that if j is odd, then M = (t21 , t1 t2 , t22 ) and if j is even, then M = (t21 , t1 t22 , t22 ). So Serre
dim R[M ] ≤ d by (1, 2) above.
In both cases, applying (3.11(1)), we get Serre dim R[M ] ≤ max {1, d}.
8
4. Let M be a monoid generated by (t31 , t1 t22 , t32 ) Then M is not seminormal. For this, let z =
2
3
2
3
6 3
(t1 t22 )2 t−3
2 ∈ K(M ) \ M . Then z = t1 (t1 t2 ) ∈ M and z = t1 t2 ∈ M . Hence seminormalization
of M is M = (t31 , t21 t2 , t1 t22 , t32 ). By (3.9), Serre dim R[M ] ≤ d. Therefore, applying (3.11(1)), we
get Serre dim R[M ] ≤ max {1, d}.
3.2
Monoid algebras over 1-dimensional rings
The following result proves (1.6(i)).
Theorem 3.13 Let R be a ring of dimension 1 and M a c-divisible monoid. If P is a projective
R[M ]-module of rank r ≥ 3, then P ∼
= ∧r P ⊕R[M ]r−1 .
Proof If R is normal, then we are done by Swan [23]. Assume R is not normal.
Case 1. Assume R has finite normalization. Let R be the normalization of R and C the conductor
ideal of the extension R ⊂ R. Then ht C = 1. Hence R/C and R/C are zero dimensional rings.
Consider the following fiber product diagram
R[M ]
/ R[M ]
(R/C)[M ]
/ (R/C)[M ]
∼ P ′ ⊗ R[M ].
If P ′ = ∧r P ⊕R[M ]r−1 , then by Swan [23], P ⊗ R[M ] ∼
= ∧r (P ⊗ R[M ])⊕R[M ]r−1 =
By Gubeladze [8], P/CP and P ′ /CP ′ are free (R/C)[M ]-modules. Further, SLr ((R/C)[M ]) =
Er ((R/C)[M ]) for r ≥ 3, by Gubeladze [9]. Now using standard arguments of fiber product diagram,
we get P ∼
= P ′.
Case 2. Now R need not have finite normalization. We may assume R is a reduced ring with
connected spectrum. Let S be the set of all non-zerodivisors of R. By [8], S −1 P is a free S −1 R[M ]module. Choose s ∈ S such that Ps is a free Rs [M ]-module.
Now we follow the arguments of Roy ([17],Theorem 4.1). Let R̂ denote the s-adic completion R.
Then R̂red has a finite normalization. Consider the following fiber product diagram
R[M ]
/ R̂[M ]
Rs [M ]
/ R̂s [M ]
Since R̂s is a zero dimensional ring, by [9], SLr (R̂s [M ]) = Er (R̂s [M ]) for r ≥ 3. If P ′ = ∧r P ⊕R[M ]r−1 ,
then Ps and Ps′ are free Rs [M ]-modules and by Case 1, P ⊗ R̂[M ] ∼
= P ′ ⊗ R̂[M ]. By (2.5), P ∼
= P ′.
This completes the proof.
The following result is due to Kang ([12], Lemma 7.1 and Remark).
9
Lemma 3.14 Let R be a 1-dimensional uni-branched affine algebra over an algebraically closed field,
R the normalization of R and C the conductor ideal of the extension R ⊂ R. Then R/C = R/C +
√
a1 R/C + · · · + am R/C, where ai ∈ C the radical ideal of C in R.
Lemma 3.15 Let R be a 1-dimensional ring, R the normalization of R and C the conductor ideal of
√
the extension R ⊂ R. Assume R/C = R/C + a1 R/C + · · · + am R/C, where ai ∈ C the radical ideal
of C in R. Let M be a monoid and write A = R/C.
(i) If σ ∈ SLn (A[M ]), then σ = σ1 σ2 , where σ1 ∈ SLn ((R/C)[M ]) and σ2 ∈ En (A[M ]).
(ii) If P is a projective R[M ]-module of rank r, then P ∼
= ∧r P ⊕R[M ]r−1 .
Proof (i) Let σ = (bij ) ∈ SLn (A[M ]). Write bij = (bij )0 + (bij )1 a1 + · · · + (bij )m am , where (bij )l ∈
√
(R/C)[M ]. If α = ((bij )0 ), then det(σ) = 1 = det(α) + c, where c ∈ ( C/C)[M ]. Since c ∈ (R/C)[M ]
is nilpotent, det(α) is a unit in (R/C)[M ]. Let β = diagonal (1/(1 − c), 1, . . . , 1) ∈ GLn ((R/C)[M ])
and σ1 = αβ ∈ SLn ((R/C)[M ]).
Note that σ1−1 σ = β −1 α−1 σ = β −1 1/(1 − c) ασ, where α = ((bij )0 ), (bij )0 are minors of (bij )0 .
σ2 := σ1−1 σ =
1
0
..
.
0
0
1
1−c
..
.
0
···
···
···
···
0
0
..
.
1
1−c
1 + c11
c21
..
.
cn1
c12
1 + c22
..
.
cn2
···
···
···
···
c1n
c2n
..
.
1 + cnn
,
√
where cij ∈ ( C/C)[M ].
Note that σ2 ∈ SLn (A[M ]) and σ2 = Id modulo the nilpotent ideal of A[M ]. Hence σ2 ∈ En (A[M ]).
Thus we get σ = σ1 σ2 with the desired properties.
(ii) Follow the proof of (3.13) and use (3.15(i)) to get the result.
Now we prove (1.6(ii)) which follows from (3.14) and (3.15).
Theorem 3.16 Let R be a 1-dimensional uni-branched affine algebra over an algebraically closed field
and M a monoid. If P is a projective R[M ]-module of rank r, then P ∼
= ∧r P ⊕R[M ]r−1 .
4
Applications
Let R be a ring of dimension d and Q a finitely generated R-module. Let µ(Q) denote the minimum
number of generators of Q. By Forster [7] and Swan [22], µ(Q) ≤ max{µ(Qp ) + dim(R/p)|p ∈
Spec (R), Qp 6= 0}. In particular, if P is a projective R-module of rank r, then µ(P ) ≤ r + d.
The following result is well known.
Theorem 4.1 Let A be a ring such that Serre dim A ≤ d. Assume Am is cancellative for m ≥ d + 1.
If P is a projective A-module of rank r ≥ d + 1, then µ(P ) ≤ r + d.
10
Proof Assume µ(P ) = n > r + d. Consider a surjection φ : An →
→ P with Q = ker(φ). Then
An ∼
= P ⊕Q. Since Q is a projective A-module of rank ≥ d + 1, Q has a unimodular element q. Since
→ P . Since n − 1 > d, An−1 is cancellative. Hence
φ(q) = 0, φ induces a surjection φ : An /qAn →
n−1 ∼ n
A
= A /qA and P is generated by n − 1 elements, a contradiction.
The following result is immediate from (4.1, 3.4, 3.6 and [6]).
Corollary 4.2 Let R be a ring of dimension d, M a monoid and P a projective R[M ]-module of rank
r > d. Then:
(i) If M ∈ C(Φ), then µ(P ) ≤ r + d.
(ii) If M ⊂ Z2+ is a normal monoid of rank 2, then µ(P ) ≤ r + d.
Let M be a c-divisible monoid, R a ring of dimension d and n ≥ max{2, d + 1}. Then Schaubhüser
[19] proved that En+1 (R[M ]) acts transitively on Umn+1 (R[M ]). Using Schaubhüser’s result and
arguments of Dhorajia-Keshari ([6], Theorem 4.4), we get that if P is a projective R[M ]-module
of rank n, then E(R[M ]⊕P ) acts transitively on Um(R[M ]⊕P ). Therefore the following result is
immediate from (4.1 and 3.13).
Corollary 4.3 Let R be a ring of dimension 1, M a c-divisible monoid and P a projective R[M ]module of rank r ≥ 3. Then µ(P ) ≤ r + 1.
Acknowledgement. We would like to thank the referee for his/her critical remark. The second
author would like to thank C.S.I.R., India for their fellowship.
References
[1] D.F. Anderson, Projective modules over subrings of k[X, Y] generated by monomials, Pacific J. Math. 79
(1978) 5-17.
[2] S.M. Bhatwadekar, Inversion of monic polynomials and existence of unimodular elements (II), Math. Z.
200 (1989) 233-238.
[3] S.M. Bhatwadekar, H. Lindel and R.A. Rao, The Bass-Murthy question: Serre dimension of Laurent
polynomial extensions, Invent. Math. 81 (1985) 189-203.
[4] S.M. Bhatwadekar and A. Roy, Some theorems about projective modules over polynomial rings, J. Algebra
86 (1984) 150-158.
[5] W. Bruns and J. Gubeladze, Polytopes, Rings and K-Theory, Springer Monographs in Mathematics, 2009.
[6] A.M. Dhorajia and M.K. Keshari, A note on cancellation of projective modules, J. Pure and Applied
Algebra 216 (2012) 126-129.
[7] Otto Forster, Über die Anzahl der Erzeugenden eines Ideals in einem Noetherschen Ring, Math. Z. 84
(1964) 80-87.
[8] J. Gubeladze, Anderson’s conjecture and the maximal class of monoid over which projective modules are
free, Math. USSR-Sb. 63 (1988), 165-188.
11
[9] J. Gubeladze, Classical algebraic K-theory of monoid algebras, Lect. Notes Math. 1437 (1990), Springer,
36-94.
[10] J. Gubeladze, The elementary action on unimodular rows over a monoid ring, J. Algebra 148 (1992)
135-161.
[11] J. Gubeladze, K-Theory of affine toric varieties, Homology, Homotopy and Appl. 1 (1999) 135-145.
[12] M.C. Kang, Projective modules over some polynomial rings, J. Algebra 59 (1979) 65-76.
[13] H. Lindel, Unimodular elements in projective modules, J. Algebra 172 (1995) no-2, 301-319.
[14] B. Plumstead, The conjectures of Eisenbud and Evans, Amer. J. Math. 105 (1983) 1417-1433.
[15] D. Popescu. On a question of Quillen, Bull. Math. Soc. Sci. Math. Roumanie (N.S.) 45 (93) no. 3-4,
(2002) 209-212.
[16] D. Quillen. Projective modules over polynomial rings, Invent. Math. 36 (1976), 167-171.
[17] A. Roy, Application of patching diagrams to some questions about projective modules, J. Pure Appl.
Algebra 24 (1982), no. 3, 313-319.
[18] H.P. Sarwar, Some results about projective modules over monoid algebras, to appear in Communications
in Algebra.
[19] G. Schabhüser, Cancellation properties of projective modules over monoid rings, Universitt Münster,
Mathematisches Institut, Münster, (1991) iv+86 pp.
[20] J.P. Serre, Sur les modules projectifs, Sem. Dubreil-Pisot 14 (1960-61) 1-16.
[21] A.A. Suslin, Projective modules over polynomial rings are free, Sov. Math. Dokl. 17 (1976), 1160-1164.
[22] R.G. Swan, The number of generators of a module, Math. Z. 102 (1967), 318-322.
[23] R.G Swan, Gubeladze proof of Anderson’s conjecture, Contemp. Math 124 (1992), 215-250.
12
| 0math.AC
|
Paradigm and Paradox in Topology Control of Power Grids
Shuai Wang & John Baillieul
arXiv:1803.05852v2 [cs.SY] 16 Mar 2018
Abstract
Corrective Transmission Switching can be used by the grid operator to relieve line overloading and voltage violations,
improve system reliability, and reduce system losses. Power grid optimization by means of line switching is typically formulated
as a mixed integer programming problem (MIP). Such problems are known to be computationally intractable, and accordingly,
a number of heuristic approaches to grid topology reconfiguration have been proposed in the power systems literature. By
means of some low order examples (3-bus systems), it is shown that within a reasonably large class of “greedy” heuristics,
none can be found that perform better than the others across all grid topologies. Despite this cautionary tale, statistical evidence
based on a large number of simulations using using IEEE 118- bus systems indicates that among three heuristics, a globally
greedy heuristic is the most computationally intensive, but has the best chance of reducing generation costs while enforcing
N-1 connectivity. It is argued that, among all iterative methods, the locally optimal switches at each stage have a better chance
in not only approximating a global optimal solution but also greatly limiting the number of lines that are switched.
I. I NTRODUCTION
Power grid control by means of transmission line switching has been the focus of research since the 1980s [4], [5] in the
hope of taking advantage of its fast time constants in changing the system state to reduce losses and improve grid security.
After the adoption of energy deregulation in the 1990s, a good deal of effort in the study of power markets has been focused
on the co-optimization of generation unit commitment and transmission line switching when congestion occurs.
A standard and widely adopted approach to modeling topology reconfiguration uses binary variables to indicate the status
of candidate lines (in or out of service) in a mixed integer programming (MIP) formulation [8]. To address the complexity
caused by the NP-hardness of MIP, most of recent work focuses on developing fast heuristic algorithms. References [9],
[10], [11], [12], [13] show the effectiveness of heuristics co-optimizing the generation and the network topology through
simulations on the IEEE 118-bus system and the the WECC 179-bus system, and the running time records of such heuristics
have been consistently set and then shattered. At the same time, the quest for firmer theoretical foundations of the heuristic
approaches continues. References [6] and [7], for instance, show that in a voltage-controlled or current-controlled DC circuit,
the sign of the overall I 2 R congestion change is always predictable. Unfortunately, the magnitude of change can only be
determined by solving Kirchhoff equations.
Some NP-complete or NP-hard problems may admit efficient heuristics in practice as long as their inputs meet certain
criteria. For example, there is a pseudo-polynomial time algorithm using dynamic programming for the knapsack problem if
the inputs are of relatively small magnitude [2], and the Fincke and Pohst algorithm [3] is suitable for solving integer
programming with pseudo continuous inputs, etc. There are, however, no such algorithms for general MIP problems,
indicating a strong form of NP-hardness for MIP.
The present paper examines fundamental challenges associated with heuristics for topology reconfiguration of power grids.
The paper is organized as follows. We begin the discussion in the next section by explicitly showing the NP-hardness of
topology control with a simple proof. In Section III, we explore a list of paradoxical behaviors associated with greedy
heuristics for topology control, and describe examples showing that many plausible heuristic approaches fail to achieve
acceptable grid operation. In particular, we define the concepts of commutativity which is the ease of creating an optimal
sequence of operations out of optimal sub-sequences, and monotonicity which is the possibility of continued step-wise
improvement on the effects of previous operations, and consistency which is the ease of obtaining better operation by
increasing the computational complexity. Simulation results in Section IV indicate that the shortcomings of heuristics pointed
out in Section III are statistically insignificant across a wide range of power grids that are frequently studied in the power
systems literature. The section ends with a theoretical justification in supporting the value of greedy heuristic in topology
control. Concluding remarks are contained in Section V.
II. T HE NP- HARDNESS OF T OPOLOGY C ONTROL
Mathematically, the DC model of power flow is equivalent to a current driven network, where power injections are
equivalent to current sources; power flowing through lines is equivalent to current through edges, etc. See Table 1.
Reference [7] approaches the topology control problem by combining graph theory and Kirchhoff laws. The work in [7]
shows that disconnecting a resistance edge in a current driven network always increases the total I 2 R congestion of the
Shuai Wang is with the Division of Systems Engineering and John Baillieul is with the Departments of Mechanical Engineering; Electrical and
Computer Engineering, and the Division of Systems Engineering at Boston University, Boston, MA 02115. Corresponding author is John Baillieul (Email:
[email protected]).
The authors gratefully acknowledge support of NSF EFRI grant number 1038230.
1
network. In other words, the new circuit network obtained after line switching has fewer edges but has higher overall I 2 R
congestion.
network potential injection
admittance
equation
circuit voltage V current I conductance G I = GV
grid
phase θ power P susceptance B P = Bθ
Table 1: The equivalence between a current driven circuit and a transmission grid.
By the equivalence between the current-controlled circuits and DC power flow models of transmission grids, we know
the total real f 2 /B (squared line flow over line susceptance) of the transmission network must be increased after switching
off a transmission line. Defining f 2 /B as the squared line flow limit over line susceptance, we find that the total f 2 /B
capacity must strictly decrease after disconnecting a line. This means that the total f 2 /B stability margin, the difference
between the total f 2 /B capacities and actual real f 2 /B, must be decreased after each step switching out lines. This also
suggests a good rule of thumb of line switching in transmission networks: do not switch out lines in a congested network
with relatively low f 2 /B stability margin on those uncongested lines. But what if we are dealing with a congested network
with high capacity margin on uncongested lines?
Topology control is typically formulated as a mixed integer programming (MIP) problem [8]. Although the common
wisdom is that many MIP problems are NP-hard and some of them don’t even have pseudo-polynomial time solutions, an
explicit proof of the NP-hardness of topology control, to the best of our knowledge, is somehow lacking in the literature.
Thus, we start the discussion by answering this question by proving the following.
Proposition 1: Given the capacity of each line and the net power injection/withdrawal of each bus in an arbitrary DC
model of a power network, the topology control problem is NP-complete.
Proof: First we know that the feasibility problem of topology control is NP because we can verify in polynomial time
whether an instance is a feasible solution. The verification involves two steps. The first is to calculte the new line flows by
solving the updated OPF problem. Next, we check whether line congestions still exist. Clearly, both steps can be done in
polynomial time.
The second part of the proof involves a reduction from an arbitrary instance of an NP complete problem to an instance
of the topology control problem.
Here, we use the well known subset sum problem for the reduction. The subset sub problem is this: given a set of numbers,
is there a non-empty subset whose sum is zero? For instance, given the set {−1, −2, −3, 4, 8}, the answer is yes because
the subset {−1, −3, 4} sums to zero. The problem is known to be NP-complete [1]. We take an instance of the subset sum
problem and reduce it to a topology control instance that has a feasible solution if and only if the subset sum problem has
a non-empty subset whose sum is zero.
Let X = {x1 , ..., xm , y1 , ..., yn } (xi > 0, i = 1, ..., m and yj < 0, j = 1, ..., n) be an instance of the subset sum problem.
We then can reduce it to the topology control problem shown in Fig. 1.
Bus 1
x1
Bus 2
. ..
~
xm
Main Source
-y1
Line L1
. ..
Line L2
-yn
Bus 3
Bus 4
Load = 2MW
Fig. 1.
An instance of topology control problem.
In Fig. 1, we assume that the susceptances of vertical lines are {x1 , ..., xm } (xi > 0, i = 1, ..., m) and {−y1 , ..., −yn }
(yj < 0, j = 1, ..., n), respectively. All other lines are assumed to be of same susceptance. In addition, the line capacities
of the lowest two lines {L1 , L2 } are both 1 M W , and we assume the line capacities of other lines are all large enough. To
2
satisfy the 2 M W demand of the lower bus, we must balance the susceptance of the left group of lines {x1 , ..., xm } with
that of the right group of lines {−y1 , ..., −yn } . Then we see that the topology control instance has a feasible solution if
and only if the subset sum problem {x1 , ..., xm , y1 , ..., yn } has a non-empty subset whose sum is zero.
III. PARADOX OF T OPOLOGY C ONTROL
Proposition 1 is not surprising as most integer programming is NP-hard. The NP-hardness naturally leads researchers
to think about possible heuristic methods. In recent years there has been a significant interest in developing heuristics to
co-optimize network topology and generation [10], [11], [12], [13]. For example, the work reported in [11] finds topology
improvements by iteratively solving the linear programming (LP) formulation of the DC optimal power flow (DCOPF) by
means of a heuristic algorithm that disconnects the one unprofitable line per iteration while enforcing a relaxed form of
N-1 reliability. Such a heuristic approach is one way of avoiding the severe computational complexity of the line switching
problems, and makes the real time topology control possible. The NP-hardness of the problem, however, makes optimality
not guaranteed by any heuristic, and we are led to a set of interesting paradoxical phenomena.
We focus our attention on several iterative greedy heuristics in which the main idea is to disconnect/connect the line(s)
that can reduce the cost most in each iteration. We show paradoxical behaviors associated with the heuristics by using some
simple 3-bus power system examples.
A. Non-commutativity
The first paradoxical behavior is that the most beneficial single line may not be included in the most beneficial pair of
lines or the most beneficial set of n (n > 2) lines. We call this the non-commutativity property of the topology control
problem. This problem can be easily illustrated as follows.
Bus 1
~
$10/MW
Line L1
|Flow|≤10MW
Bus 2
Line L3
|Flow|≤1MW
Line L2
|Flow|≤10MW
Line L4
|Flow|≤1MW
Line L5
|Flow|≤10MW
Load = 15MW
Fig. 2.
~
Bus 3
$80/MW
The example of non-commutativity problem.
In Fig. 2, there are two generators with the cheap one on bus 1 and the expensive one on bus 3 and a 15MW demand on
bus 3. We assume that all transmission lines have the same susceptance, and the line capacities range from 1MW to 10MW
as marked in the figure. The original cost before applying topology control is $710 with the cheap generator supplying
7MW and the expensive generator supplying 8MW.
If we remove the most beneficial single line per iteration, then we will remove the line L1 and the cost will be $500
with the cheap generator supplying 10MW and the expensive generator supplying 5MW. The algorithm will stop here as
we can’t further reduce the cost by removing another line. However, if we are allowed to remove the most beneficial pair
of lines per iteration, then we will keep L1 connected and instead remove the line L3 and L4 together and the cost will be
$150 which actually equals the cost of unconstrained optimal power flow.
B. Non-monotonicity
The second paradoxical behavior is that one line which was removed earlier may be reconnected later because keeping it
switched off is no longer beneficial. We call this the non-monotonicity property of the topology control problem.
This problem is illustrated in Fig. 3 which shows a power system with the cheapest generator on bus 2, a slightly more
expensive generator on bus 1, and the most expensive generator on bus 3. We assume that the susceptance of the horizontal
3
Bus 1
~
$20/MW
Line L1 B1 =∞
|Flow|≤100MW
Line L4 B
|Flow|≤10MW
Line L5 B
|Flow|≤10MW
Line L3 B
|Flow|≤2MW
Line L6 B
|Flow|≤10MW
~
$10/MW
Bus 2
Line L2 B
|Flow|≤1MW
Fig. 3.
~
Load = 100MW
Bus 3
$80/MW
The example of non-monotonicity.
transmission line is large enough and all other lines are of the same but limited susceptance. There is a demand of 100MW
on bus 3 and the line capacities range from 1MW to 100MW as marked in the figure. The original cost before applying
topology control is $7650 with the cheapest generator supplying 5MW and the most expensive generator supplying 95MW.
Suppose we can either remove or connect the most beneficial single line per iteration, then we will remove the line L1
first, line L2 second, and line L3 third. In the fourth iteration, however, we find that having L1 out of service is no longer
beneficial and we can reduce the cost by reconnecting L1. After the reconnecting, the cost would be $5900 with the cheapest
generator supplying 30MW and the most expensive generator supplying 70MW. The algorithm will stop here as we can’t
further reduce the cost by removing or reconnecting another line.
C. Non-consistency
The third example illustrated in in Fig. 4 shows that an algorithm that only allows removal of one line per iteration
may outperform another algorithm that allows to either remove or reconnect one line per iteration. We call this the nonconsistency property of the topology control problem. This shows that the increased computational complexity sometimes
is not compensated by lower generation cost.
~
Line L4 B
|Flow|≤10MW
Line L5 B
|Flow|≤10MW
Line L3 B
|Flow|≤3MW
Line L6 B
|Flow|≤15.5MW
Line L2 B
|Flow|≤1MW
Line L7 B
|Flow|≤35MW
Load = 100MW
Fig. 4.
~
Line L1 B1 =∞
|Flow|≤100MW
~
Bus 1
$20/MW
$10/MW
Bus 2
Bus 3
$80/MW
The first example of non-consistency problem.
Fig. 4 shows a power system with the cheapest generator on bus 2, and the mid-cost generator on bus 1, and the most
expensive generator on bus 3. We assume that the susceptance of the horizontal transmission line is large enough and all
other lines are of the same but limited susceptance. There is a demand of 100MW on bus 3 and the line capacities range
from 1MW to 100MW as marked in the figure. The OPF with all lines in service is $7580 with the cheapest generator
supplying 6MW and the most expensive generator supplying 94MW.
4
Algorithm (1) (only allowing removal of one line per iteration) will remove line L1 first, line L2 second, line L3 third,
line L5 fourth, and line L6 last. The final cost will be $4950 and with the cheapest generator supplying 35MW, and the
mid-cost generator supplying 10MW, and the most expensive generator supplying 55MW.
Algorithm (2) (allowing to either remove or reconnect one line per iteration) will remove line L1 first, line L2 second,
and line L3 third. In the fourth iteration, line L1 will be reconnected and the algorithm stops. The cost will be $5200 with
the cheapest generator supplying 40MW and the most expensive generator supplying 60MW.
As we can see algorithm (1) outperforms algorithm (2).
The fourth example illustrated in Fig. 5 shows another type of non-consistency property: one algorithm that only allows
removal of one line per iteration may outperform another algorithm that allows removal of at most 2 lines line per iteration.
Line L6
B=5S (R=0.2Ω)
|Flow|≤10MW
~
Bus 1
$10/MW
Line L1
B=8S
Line L2
B=1.1S
Line L3
B=1.1S
Line L4
B=5S
Line L5
B=5S
Line L7
B=10S
|Flow|≤10MW
Fig. 5.
~
Load = 100MW
Bus 2
Bus 3
$80/MW
The second example of non-consistency problem.
Fig. 5 shows a power system with the cheap generator on the upper bus and the expensive generator on the lower bus.
We assume that the susceptances of L1 through L7 are 8S, 1.1S, 1.1S, 5S, 5S, 5S and 10S, respectively. There is a demand
of 100MW on the lower bus and the line capacities of L6 and L7 are both 10MW. The line capacities of other lines are
assumed to be sufficiently large. The OPF with all lines in service is $6775 with the cheap generator supplying 17.5MW
and the expensive generator supplying 82.5MW.
Algorithm (1) (only allowing removal of one line per iteration) will remove line L1 first, line L2 second, and line L3
third. The cost will be $6600 with the cheap generator supplying 20MW and the expensive generator supplying 80MW. The
algorithm will stop here as we can’t further reduce the cost by removing another line.
Algorithm (3) (allowing to remove at most two lines per iteration) will remove line L4 and L5 in the first iteration, with
the resulting cost being $6607 with the cheap generator supplying 19.9MW and the expensive generator supplying 80.1MW.
The algorithm will stop here as we can’t further reduce the cost by removing another line or another pair of lines.
As we can see algorithm (1) outperforms algorithm (3).
Mathematically, the Algorithms (1), (2) and (3) described above are of the same level of performance, i.e. no one dominates
another. For a general binary programming problem, only when the search space of one heuristic explicitly contains that
of the other can it declare dominancy. Thus the paradoxical phenomenon, “no ‘guaranteed’ extra return for increased
computational complexity”, can be observed in all types of non-brute-force greedy heuristics. This phenomenon, which
makes the binary programming problem extremely challenging, is visualized in Fig. 6.
IV. PARADIGM OF T OPOLOGY C ONTROL
Section III explicitly discounts the value of increased computational effort of heuristics in reducing the generation cost for
specific grid scenarios. It is natural to ask if the next best thing is true: can the increased computational effort “statistically”
5
Relative performance
(dominancy level)
Brute force
method
Computational complexity
of heuristic
Fig. 6.
Visualization of the paradoxical phenomenon: no extra return for increased computational complexity of heuristics in topology control.
reduce generation costs for a large set of scenarios? The result of our simulations suggests the answer is yes.
A. Test Network Overview
The heuristics were tested on the IEEE 118-bus test system. This test system represents a portion of the Midwestern
US Power System as of December, 1962. The version of the test system employed is available at [14]. The generator cost
information used in the IEEE 118 network is extracted from [15]. The test system consists of 118 buses, 54 generators, and
194 lines, all of which are assumed to be connected in the initial topology.
B. Heuristics Overview
Three heuristics were tested and compared in this section. All of them sequentially disconnect transmission lines until no
further generation costs can be reduced.
To make this a fair comparison, all heuristics only open closed switches and do not close open switches since the Line
Profits Heuristic [11] (described below) cannot compute potential improvements from reconnecting disconnected lines.
1) Random Heuristic: The algorithm is specified in Fig. 7. This heuristic sequentially solves the DCOPFs. In each
iteration, if the DCOPF is feasible, a set of switchable candidate lines is then created. The cardinality of the switchable
set is reduced in each iteration after randomly disconnecting a switchable unprofitable line. In addition, the switchable set
enforces a relaxed form of the “N-1” reliability requirement: at no time is a load or generator bus connected by less than
TWO lines. The heuristic stops when no switchable unprofitable lines can be found.
Step 0
Solve the initial OPF and store solutions.
Initialize the switchable line set.
Step 1
Randomly select a candidate line from the
switchable set and change its status.
Reconnect the disconnected
line and remove it from the
switchable set.
No
Step 2
Solve the updated DCOPF (with one new
line removed) problem.
Store solution and remove
the disconnected line from
the switchable set.
Step 3
Improved
Solution?
Yes
Step 4
Stopping criteria?
Yes
Step 5
Store solution.
Fig. 7.
Flow chart describing general algorithm structure of the Random Heuristic.
No
6
2) Line Profit Heuristic: This heuristic is proposed in [11], and is named after the metric used in the switching criteria.
Its algorithm structure is almost the same as Fig. 7 except for the line switching step (Step 1 in Fig. 7). In the line switching
step, the line profits, fl (πnl − πml ), are computed where fl is the power flow on transmission line l, and πnl − πml are the
locational marginal price (LMP) difference between the two ending buses of line l. The convention is that the flow direction
of line l is from bus m and to bus n. The most unprofitable line (with lowest line profit fl (πnl − πml )), rather than a random
line, is selected as a candidate for opening. Please refer [11] for more details.
3) Standard Greedy Heuristic: The algorithm is specified in Fig. 8. Unlike the Line Profit Heuristic that first sorts the
candidate lines based on their line profits fl (πnl − πml ) and then selects the first line whose removal leads to savings in
the updated DCOPF, the Standard Greedy Heuristic redoes the updated DCOPF for each allowable line removal and then
selects the single line to be removed that leads to maximum savings in each iteration.
Step 0
Solve the initial OPF and store solutions.
Initialize the switchable line set.
Step 1
Calculate the savings for each switchable
line by redoing the DCOPF for each
switchable line.
Step 2
Select the line of maximum savings from
the switchable set and change its status.
Reconnect the disconnected
line and remove it from the
switchable set.
No
Step 3
Solve the updated DCOPF (with one new
line removed) problem.
Store solution and remove
the disconnected line from
the switchable set.
Step 4
Improved
Solution?
Yes
Step 5
Stopping criteria?
No
Yes
Step 6
Store solution.
Fig. 8.
Flow chart describing general algorithm structure of the Random Heuristic.
Clearly, optimality is not guaranteed with any of the heuristics, since the transmission topology is not co-optimized with
power dispatch.
C. Simulation Results and Analysis
To generate the test scenarios for the heuristics, a fixed load is maintained and we perform a Monte Carlo simulation
where the generator costs are randomly varied. The sample size used for the Monte Carlo simulation is 50. The percent
savings rather than the dollar value are the focus of the paper.
Two benchmark cases are used to evaluate the savings of the heuristic: Case one considers the initial topology and line
constraints whose DCOPF sets an upper bound of the total generation costs, and Case two considers the unconstrained case
whose DCOPF provides a lower bound of the generation costs. The cost difference between the two benchmark cases is
called the maximum attainable savings (MAS).
The three heuristics are implemented in MATLAB, using MATPOWER [16] as the DCOPF solver. The performance
comparison among the three heuristic is detailed in Table 2.
heuristic
Random
Line Profit
Greedy
saving/MAS lines disconnected mean effort
0.517 ± 0.092 31.22 ± 6.37
0.193
0.631 ± 0.049 14.53 ± 4.12
0.317
0.672 ± 0.051 10.18 ± 3.98
1
Table 2: Performance comparison.
7
Strictly speaking, the three heuristics are said to be of the same level computational complexity as they need to redo
the DCOPF for each switchable line in the worst case. For example, the Random Heuristic or the Line Profit Heuristic
may, though with very low probability, not be able to find a switchable unprofitable line until after redoing the DCOPF
for each allowable line removal. In general, the Standard Greedy Heuristic, however, costs much more computational effort
than the Random Heuristic and the Line Profit Heuristic especially for a large scale networks. The benefit of the additional
computational effort is the increased switching quality, i.e. higher cost reduction achieved at each stage. In general, the
switching savings in each iteration is the best in Standard Greedy Heuristic, the worst in the Random Heuristic and better
in Line Profit Heuristic. Unlike the paradoxical behaviors for specific scenarios, the better switching quality accumulated in
each iteration does lead to more average saving for a set of test scenarios as shown in Table 2, which may outweigh the
extra computational cost.
It is natural to expect one heuristic may save more if it disconnects more lines. What seems counterintuitive is that,
on average, the Standard Greedy Heuristic not only attains maximum savings but also disconnects the fewest lines. This,
however, can be well explained by the work in [7] and [17].
Theorem 1: [7] For an arbitrary current-controlled circuit (a circuit network that is comprised purely of current sources
and resistors), the total I 2 R losses of the network must increase after removing an arbitrary resistance link.
Theorem 2: [17] The increase of total I 2 R loss, ∆P , resulting from the removal of a resistor link with endpoint pair
0
{m, n} from a current-controlled circuit is given by |∆P | = |Imn Vmn |, where Imn denotes the current flowing on the link
0
before its removal, and Vmn denotes the voltage difference between node pair {m, n} after its removal.
Recalling the equivalence between the current-controlled circuits and DC power flow models of transmission grids shown
in Table 1, Theorem 1 tells us that the new transmission network obtained after line switching is of fewer lines but higher
total real f 2 /B (squared line flow over line susceptance) congestion. Theorem 2 quantifies the increase of the I 2 R losses of
a circuit due to a disconnected resistor by using the product of two electrical parameters associated with that disconnected
resistor. In the DC power flow models of transmission grids, it is equivalent to say that the increase of total real f 2 /B of a
transmission grid due to one line removal is given by the product of the power flowing through that transmission line flow
before its removal and the phase difference between its ending buses after its removal.
One profitable switching that can greatly reduce the generation cost usually significantly redistributes the power flowing
across the network. Such effect is unlikely to be obtained by switching a transmission line with relatively small pre-removal
line flow and post-removal phase difference between its ending buses. Since the Standard Greedy Heuristic always executes
the line removal leading to maximum savings in each iteration, it is reasonable to expect that the average change of of total
real f 2 /B in each iteration of the Standard Greedy Heuristic is usually larger than that of the Random Heuristic and Line
Profit Heuristic. The combination of the above result together with the ever decreasing total f 2 /B capacity due to fewer
lines in service explain the superior performance of the Standard Greedy Heuristic: not only better approximating the global
optimal solution but also better enforcing network connectivity.
V. C ONCLUSION
In spite of the seemingly promising simulation results of the heuristic approaches pursued by the research community
and a certain level of proven predictability [6], [7], [17] of topology control, satisfactory grid-scale solutions are everelusive. The cautionary tales associated with our non-commutativity, non-monotonicity, and non-consistency paradoxes leads
us to wonder whether other paradox – possibly more problematic – may associated with as yet untried heuristics. While
optimality for NP-hard problems is not obtained, our simulation results on a set of iterative heuristics nevertheless offer
some justification that, with similar searching space, a greedy heuristic has a better chance in reducing generation cost
and enforcing connectivity. Future work will be focused on further simplifying the problem by developing fast partition
algorithms for the decomposition approach [18]. Adaptability of the decomposition with respect to various types (traditional
& renewable) of generators will also be the subject of future work.
R EFERENCES
[1] Murty, K.G. and Kabadi, S.N., (1987). Some NP-complete problems in quadratic and nonlinear programming. Mathematical programming, 39(2),
pp.117-129.
[2] Kellerer, H., Pferschy, U. and Pisinger, D., 2004. Introduction to NP-Completeness of knapsack problems (pp. 483-493). Springer Berlin Heidelberg.
[3] Fincke, U. and Pohst, M., 1985. Improved methods for calculating vectors of short length in a lattice, including a complexity analysis. Mathematics
of computation, 44(170), pp.463-471.
[4] Koglin, H.J. and Muller, H., 1980, June. Overload reduction through corrective switching actions. In Int. Conf. Power Sys. Monitor. & Contr Vol. 24,
No. 26.
[5] Van Amerongen, R. and Van Meeteren, H.P., 1980. Security control by real power rescheduling network switching and load shedding. Proc. 1980
CTGRE Report 32-02.
[6] Baillieul, J., Zhang, B., and Wang, S. (2015). The Kirchhoff-Braess paradox and its implications for smart microgrids. In Decision and Control (CDC),
2015 IEEE 54th Annual Conference on, pp. 6556-6563. IEEE.
[7] Wang, S., and Baillieul, J. (2016). Kirchhoff-Braess phenomena in DC electric networks. In Decision and Control (CDC), 2016 IEEE 55th Conference
on, pp. 3286-3293. IEEE.
[8] Chen, C.S. and Cho, M.Y., 1993. Energy loss reduction by critical switches. IEEE Transactions on Power Delivery, 8(3), pp.1246-1253.
8
[9] Hedman, K.W., ONeill, R.P., Fisher, E.B. and Oren, S.S., 2009. Optimal transmission switching with contingency analysis. IEEE Transactions on
Power Systems, 24(3), pp.1577-1586.
[10] Fisher, E. B., O’Neill, R. P., and Ferris, M. C. (2008). Optimal transmission switching. IEEE Transactions on Power Systems, 23(3):13461355
[11] Ruiz, P. A., Foster, J. M., Rudkevich, A., and Caramanis, M. C. (2011). On fast transmission topology control heuristics. In 2011 IEEE Power and
Energy Society General Meeting, pp. 18. IEEE.
[12] Fuller, J. D., Ramasra, R., and Cha, A. (2012). Fast heuristics for transmission-line switching. IEEE Transactions on Power Systems, 27(3):1377-1386
[13] Soroush, M. and Fuller, J. D. (2014). Accuracies of optimal transmission switching heuristics based on dcopf and acopf. IEEE Transactions on Power
Systems, 29(2):924932.
[14] https://www2.ee.washington.edu/research/pstca/pf118/pg tca118bus.htm
[15] Blumsack, S. A., 2006, Network topologies and transmission investment under electric-industry restructuring, Ph.D. dissertation, Eng. Public Pol.,
Carnegie Mellon Univ., Pittsburgh, PA.
[16] Zimmerman, R.D., Murillo-Snchez, C.E. and Thomas, R.J., 2011. MATPOWER: Steady-state operations, planning, and analysis tools for power
systems research and education. IEEE Transactions on power systems, 26(1), pp.12-19.
[17] Wang, S. and Baillieul, J., 2017. A Novel Decomposition for Control of DC Circuits and Grid Models with Heterogeneous Energy Sources, arXiv
preprint arXiv:1709.07569, accepted to the American Control Conference (ACC), 2018.
[18] Wang, S. and Baillieul, J., 2018. Power Grid Decomposition Based on Vertex Cut Sets and Its Applications to Topology Control and Power Trading,
arXiv preprint arXiv:1803.05860.
| 3cs.SY
|
Understanding Negations in Information Processing:
Learning from Replicating Human Behavior
Nicolas Pröllochsa,∗, Stefan Feuerriegela , Dirk Neumanna
arXiv:1704.05356v1 [cs.AI] 18 Apr 2017
a
Chair for Information Systems Research, University of Freiburg, Platz der Alten Synagoge,
79098 Freiburg, Germany
Abstract
Information systems experience an ever-growing volume of unstructured data, particularly in the form of textual materials. This represents a rich source of information from which one can create value for people, organizations and businesses.
For instance, recommender systems can benefit from automatically understanding preferences based on user reviews or social media. However, it is difficult for
computer programs to correctly infer meaning from narrative content. One major
challenge is negations that invert the interpretation of words and sentences. As
a remedy, this paper proposes a novel learning strategy to detect negations: we
apply reinforcement learning to find a policy that replicates the human perception
of negations based on an exogenous response, such as a user rating for reviews.
Our method yields several benefits, as it eliminates the former need for expensive
and subjective manual labeling in an intermediate stage. Moreover, the inferred
policy can be used to derive statistical inferences and implications regarding how
humans process and act on negations.
Keywords:
Unstructured data, information processing, decision-making, natural language
∗
Corresponding author. Mail: [email protected]; Tel: +49 761 203 2395;
Fax: +49 761 203 2416.
Email addresses: [email protected] (Nicolas Pröllochs),
[email protected] (Stefan Feuerriegel),
[email protected] (Dirk Neumann)
processing, reinforcement learning, negations
1. Introduction
When making decisions in their daily lives, humans base their reasoning on
information, while pondering expected outcomes and the importance of individual
arguments. At the same time, they are continuously confronted with novel information of potentially additional value [1]. Psychological theories suggest that,
when processing information, humans constantly categorize and filter for relevant
tid bits [1, 2]. The outcome of this filtering then drives decision-making , which
in turn affects interactions with information technology, personal relationships,
businesses or whole organizations. Information Systems (IS) research [3] thus
strives for insights into how humans interpret and react to information in order
“to understand and improve the ways people create value with information” [4,
p. 20].
Information is increasingly encoded not only in quantitative figures, but also
in qualitative formats, such as textual materials [5]. Common examples from the
digital age include blog entries, posts on social media platforms, user-generated
reviews or negotiations in electronic commerce. These materials predominantly
encompass feedback, comments or reviews, and thus immediately impact the
decision-making processes of individuals and organizations [6, 7]. Rather than
a manual text analysis, a computerized evaluation is generally preferable, as it
can process massive volumes of documents, often in realtime. Recent advances
in information technology render it possible to automatically investigate the influence of qualitative information from narrative language and word-of-mouth –
especially in order to gain an understanding of its content [8]. This, in turn, opens
up novel opportunities for computerized decision support, e. g. question answering
and information retrieval (e. g. [9]). Consequently, understanding decision-making
2
and providing decision support both increasingly rely upon computerized natural
language processing.
The exact interpretation of language is largely impracticable with computer
programs at present. Among the numerous difficulties is the particularly daunting challenge of analyzing negations [10, 11], since their context-dependent nature
hampers bag-of-words approaches for natural language. The latter method considers only word frequencies without looking at the order of words from the beginning
to the end of a document. Such a careful consideration is, however, necessary as
negations occur in various forms; they reverse the meanings of individual words,
but also of phrases or even whole sentences [12]. Thus, one must handle negations
properly in order to achieve an accurate comprehension of texts. The importance
of profound language understanding is demonstrated by the following exemplary
applications:
Recommender Systems. Recommender systems support users by predicting
their rating or preference for a product or service. An increasing number of usergenerated reviews constitutes a compelling source of information [13]. Hence,
recommender systems must accurately classify positive and negative content in
order to interpret the intended opinion contained within reviews [10, 14, 11] or
their credibility [15].
Financial News. Investors in financial markets predominantly base their decision on news when choosing whether to exercise stock ownership. In addition
to quantitative numbers, qualitative information found in news, such as tone
and sentiment, strongly influences stock prices (e. g. [16, 17]). For example,
companies often frame negative news using positive words [18]; therefore, empirical research, investors and automated traders demand the precise processing of
negations.
3
Question Answering Systems. Question answering systems support users with
insights drawn from immense bodies of data. For instance, IBM’s Watson processes millions of medical documents in order to discover potential diseases and
recommends treatments based on symptoms. Similarly, one can automatically
determine software requirements from descriptions. For such applications of information retrieval, it is necessary to distinguish between certainty and beliefs
in natural language by considering negations [19, 20, 9].
Negotiations. Negotiations usually consists of a seesaw of offers and counteroffers, of which most are usually rejected until one is finally accepted. Even
in the digital area, negotiations are based on textual arguments [21] and, in
order for systems to automatically decode outcomes, it is necessary to examine
language correctly [22, 23]. Similarly, this holds for cases in which language helps
to predict deception in human communication [24, 25].
Despite the fact that language offers a rich source of information, its processing
and the underlying decision-making are still subject to ongoing research activities. This includes negation processing, which affects virtually every context or
domain, since neglecting negations can lead to erroneous implications or false interpretations. In the following, we refer to the part of a document whose meaning
is reversed as the negation scope. Identifying negation scopes is difficult as these
are latent, unobservable and – even among experts – highly subjective [12]. In
addition, many machine learning algorithms struggle with this type of problem as
it is virtually impossible to encode with a fixed-length vector while preserving its
order and context [8].
This paper develops a novel method, based on reinforcement learning, for
detecting, understanding and interpreting negations in natural language. This
approach exhibits several favorable features that overcome shortcomings found
in prior works. Among them, reinforcement learning is well suited to learning
4
tasks of varying lengths; that is, it can process sentences of arbitrary complexity
while preserving context and order of information. Furthermore, our approach
eliminates the need for manual labeling of individual words and thus avoids the
detrimental influence of subjectivity and misinterpretation. On the contrary, our
model is solely trained on an exogenous response variable at document level. We
refer to this score as the gold standard, not only because it represents a common
term in text mining, but also to stress that it can reflect any human behavior of
interest. As a result of its learning process, our method adapts to domain-specific
cues or particularities of the given prose.
Our contribution goes beyond the pure algorithmic benefits, since we envision
the goal of understanding human information processing. For this purpose, our
approach essentially learns to replicate human decision-making in order to gain
insights into the workings of the human mind when processing narrative content.
In fact, the reinforcement learning approach is trained based on past decisions.
While the model receives feedback as to how well it matches the human decision, it
does not receive explicit information regarding how to improve accuracy. Rather,
the learner iteratively processes information in textual materials and experiments
with different variants of negation processing in a trial-and-error manner that
imitates human behavior.
Reinforcement learning can considerably advance our understanding of decisionmaking. Indeed, learning itself has long been conceptualized as the process of creating new associations between stimuli, actions and outcomes, which then guide
decision-making in the presence of similar stimuli [26]. We thus show how to make
the knowledge of these associations explicit: we propose a framework by which
to study negation scopes that were previously assumed to be latent and unobservable. Contrary to this presumption, we manage to measure their objective
perception. Therefore, we exploit the action-value function inside the reinforce-
5
ment learning model in order draw statistical inferences and derive conclusions
regarding how the human mind processes and acts upon negations. As such, our
approach presents an alternative or supplement to experiments, such as those of
a psychological or neuro-physiological nature (along the lines of NeuroIS).
This paper is structured as follows. Section 2 provides an overview of related
works that investigate human information processing of textual materials, while
also explaining the motivation behind our research objective of understanding
negations in natural language. Subsequently, Section 3 explains how we adapt
reinforcement learning to improve existing methods of negation scope detection.
In Section 4, we demonstrate our novel approach with applications from recommender systems and finance in order to contribute to existing knowledge of
information processing. Finally, Section 5 discusses implications for IS research,
practice and management.
2. Background
This section presents background on natural language processing. First, we
discuss recent advances in computational intelligence and then outline challenges
that arise when working with narrative content. We conclude by briefly reviewing
previous works on the handling of negation in natural language.
2.1. Human Information Processing
Advances in computational intelligence have revolutionized our understanding
of learning processes in the human brain. As a result, research has yielded precise theories regarding the reception of information and function of human memory [26]. For instance, statistical models provide insights into human memory
formation and the dynamics of memory updating, while also validating these theories by replicating experiments with statistical computations [27]. Reinforcement
learning, especially, has gained considerable traction as it mines real experiences
6
with the help of trial-and-error learning to understand decision-making [26]. Accordingly, existing studies find that the brain naturally reduces the dimensionality
of real-world problems to only those dimensions that are relevant for predicting
the outcome [26]. Along these lines, a recent review argues for jointly combining
both perception and learning in order to draw statistical inferences regarding information processing [28]. While the previous reference materials predominantly
address visual perception, the focus of this paper is rather on natural language.
Behavioral theories suggest that human decision-makers seek as much information as possible in order to make an informed decision [29]. In the case of natural
language, researchers have devised advanced methods to study the influence of
textual information on the resulting decision. On the one hand, it is common to
extract specific facts or features from the content and relate these to a decision
variable [14]. On the other hand, information diffusion is also frequently studied
by measuring the overall tone of documents. This latter approach comprises a
variety of different aspects of perception, including negative language, sentiment
and emotions [11, 30, 17].
In the case of natural language, a variety of textual sources have served as
research subjects for studying word-of-mouth communication and information diffusion. For instance, the dissemination of information and sentiment has been empirically tested in social networks [31], revealing that emotionally-charged tweets
are retweeted more often and faster [30]. Similarly, measuring the response to information allows one to test behavioral theories, such as attribution theories or the
negativity bias, by distinguishing between the reaction to positive and negative
content [32, 30].
2.2. Natural Language Processing
Addressing the above research questions on information processing requires
accurate models for understanding and interpreting natural language. However,
7
the majority of such methods only count the occurrences of words (or combinations), resulting in so-called bag-of-words methods. By doing so, these techniques
have a tendency to ignore information relating to the order of words and their
context [8], such as inverted meanings through negations.
Neglecting negations can substantially impair accuracy when studying human
information processing; for example, it is common “to see the framing of negative
news using positive words” [18]. To avoid false attributions, one must identify
and predict negated text fragments precisely, since information is otherwise likely
to be classified erroneously. This holds true not only for negations in information
retrieval [19, 20], but especially when studying sentiment [10, 33]; even simple
heuristics can yield substantial improvements in such cases [34].
2.3. Negation Processing
Previous methods for detecting, handling and interpreting negations can be
grouped into different categories (cf. [35, 36, 20]).
Rule-based approaches are among the most common due to their ease of implementation and solid out-of-the-box performance. In addition, rules have been
found to work effectively across different domains and rarely need fine-tuning [37].
They identify negations based on pre-defined lists of negating cues and then hypothesize a language model which assumes a specific interpretation by the audience. For example, some rules invert the meaning of all words in a sentence,
while others suppose a forward influence of negation cues and thus invert only a
fixed number of subsequent words [38]. Furthermore, a rule-based approach can
also incorporate syntactic information in order to imitate subject and object [39].
However, rules cannot effectively cope with implicit expressions or particular,
domain-specific characteristics.
Machine learning approaches can partially overcome previous shortcomings [20],
such as the difficulty of recognizing implicit negations. Common examples of such
8
methods include generative probabilistic models in the form of Hidden Markov
models and conditional random fields (e. g. [12]). These methods can adapt to
domain-specific language, but require more computational resources and rely upon
ex ante transition probabilities. Although approaches based on unsupervised
learning avoid the need for any labels, practical applications reveal inferior performance compared to supervised approaches [35]. The latter usually depend
on manual labels at a granular level, which are not only costly but suffer from
subjective interpretations [12].
3. Method Development
This section posits the importance of developing a novel method for learning
negation scopes in textual materials. After first formulating a problem statement,
we introduce our approach, which is based on reinforcement learning.
3.1. Rationale and Intuition of Proposed Methodology
Negation scope detection in related research predominantly relies on rule-based
algorithms. Rule-based approaches entail several drawbacks, as the list of negations must be pre-defined and the selection criterion according to which rule a
rule is chosen is usually random or determined via cross validation. Rules aim to
reflect the “ground truth” but fail at actually learning this.
For those seeking to incorporate a learning strategy, a viable alternative exists
in the form of generative probabilistic models (e. g. Hidden Markov models or
conditional random fields [20]). These process narrative language word-by-word
and move between hidden states representing negated and non-negated parts. On
the one hand, unsupervised learning can estimate the models without annotations,
but yields less accurate results overall [35]. On the other hand, supervised learning
offers better performance, but this approach requires a training set with manual
labels for each word (see Figure 1) which are supposed to approximate the latent
9
negation scopes. Such labeling requires extensive manual work and is highly
subjective, thus yielding only fair performance [35, 36]. As a further drawback,
many approaches from supervised machine learning are simply infeasible as they
usually require an input vector of a fixed, pre-defined length without considering
its order. This circumstance thus necessitates a tailored method for dealing with
negations in narrative materials.
Rules
Apply static rule
Corpus
Negation
Scopes
Insert
Information
processing
Insert
Information
processing
Insert
Information
processing
Unsupervised learning
Estimate
Training set
Model
Apply
Negation
scopes
Corpus
Supervised learning
Training set
Label
Word
annotations
Estimate
Model
Apply
Negation
scopes
Corpus
Our method
Corpus
Find optimal policy for negation scopes
Information
processing
Interpret policy
Figure 1: Process chart compares the different stages of labeling, training and applying rules in
order to evaluate information processing.
In contrast to these suboptimal methods, we propose a novel approach to
determining latent negation scopes based on reinforcement learning. It works well
with learning tasks of arbitrary length [40] and is adaptable to domain-specific
features and particularities. Since it relies only upon a gold standard at document
level, it represents a more objective strategy. However, such an approach has
10
been largely overlooked in previous works on natural language processing (see
Section 2).
Reinforcement learning aims at learning a suitable policy directly through
trial-and-error experience. It updates its knowledge episodically and learns the
policy from past experience using only limited feedback in the form of a reward.
This reward indicates the current performance of the classifier, but does not necessarily specify how to improve the policy. In addition, this type of learning can
also handle highly complex sentences and is thus well suited to the given task.
3.2. Learning Task for Negation Detection
Understanding negations and their influence on language is – as previously
mentioned – a non-trivial computational problem, since the underlying learning
task suffers from several undesirable features:
(1) Even though sentences follow grammatical rules, they can be nested up to
arbitrary complexity and thus become arbitrarily long.
(2) Words have a meaning based on their context, which is implicitly established
by their order. By merely rearranging the word order, one can produce a
completely different meaning. This constitutes a dependency according to
which the meaning of words depends, in part, on all other words and their
order in the same document.
(3) Negation scopes affect individual words; however, we lack annotations on a
word-by-word basis. Instead, we only observe a gold standard for the whole
document upon which we must reconstruct negation scopes for each individual
word within the document.
Based on these challenging features, we can formalize the problem, resulting in
the following learning task. Both negation scopes and sentences are of different
11
length Nd depending on the specific document d. This length can theoretically
range from one to infinity. Each word wd,i in document d with i ∈ {1, . . . , Nd }
thus represents an individual classification task, which also depends on all other
words in that document, i. e.
f : (wd,i , [wd,1 , wd,2 , . . . , wd,Nd ]) 7→ {Negated, ¬Negated},
|
{z
}
(1)
Ordered sequence of words
where [wd,1 , wd,2 , . . .] is an ordered list of variable length providing context information.
Each document comes with a single label yd , i. e. the gold standard, which
reflects the response of human decision-making to the text processing. In order
to estimate f , we minimize the expected error (or any other loss-like function)
via the gold standard and the result of a text processing function. The latter
function maps the words as a predictor onto the gold standard. Examples of
text processing functions are functions that measure the accuracy of information
retrieval or sentiment based on the presence of polarity words.
3.3. Reinforcement Learning
Reinforcement learning constructs a suitable policy for negation classification
through trial-and-error experience. That is, it mimics human-like learning and
thus appears well suited to natural language processing. In the following section,
we introduce its key elements and tailor the method to our problem statement.
The overall goal is to train an agent based on a recurrent sequence of interactions. After observing the current state, the agent decides upon an action. Based
on the result of the action, the agent receives immediate feedback via a reward.
It is important to note that the agent aims only to maximize the rewards, but it
never requires pairs of input and the true output (i. e. words and a flag indicating
whether they are negated). This forms a setting in which the agent learns the
latent negation scopes.
12
More formally, the model consists of a finite set of environmental states S
and a finite set of actions A. The agent models the decision-maker by iteratively
interacting with an environment over a sequence of discrete steps and seeks to
maximize the reward over time. Here, the environment is a synonym for the
states and the transition rules between them. At each iteration i, the decisionmaking agent observes a state si ∈ S. Based on the current state si , the agent
picks an action ai ∈ A(si ), where A(si ) ⊆ A is the subset of available actions in
the given state si . Subsequently, the agent receives feedback related to its decision
in the form of a numerical reward ri+1 , after which it moves then moving to the
next state si+1 . The entire process is depicted in Figure 2.
Agent
State
si
Reward
ri
Action
ai
ri+1
si+1
Text Processing
Environment
Figure 2: Interaction between agent and environment in reinforcement learning [40].
In order to build up knowledge, reinforcement learning updates a state-action
function Q(si , ai ), which specifies the expected reward for each possible action ai
in state si . This knowledge can then be used to infer the optimal behavior, i. e.
the policy that maximizes the expected reward from any state. Specifically, the
optimal policy π ∗ (si , ai ) chooses in state si the actions ai that maximizes Q(si , ai ).
Several algorithms have been devised to learn an optimal policy π ∗ , among
which is an approach known as Q-learning [40, 41]. This methods seeks an optimal policy without an explicit model of the environment. In other words, it knows
neither explicitly the reward function nor the state transition function [42]. Instead, it iteratively updates its action-value Q(st , at ) based on past experience [41].
In our case, we use a variant with eligibility traces named Watkin’s Q(λ) due to
13
better convergence; see [40] for details.
3.4. Learning Negation Processing
In this section, we outline how we adapt reinforcement learning to our attempt
to simulate human negation processing. In each iteration, the agent observes the
current state si = (wi , ai−1 ) that we engineer as the combination of the i-th word
wi in a document and the previous action ai−1 . This specification establishes
a recurrent architecture whereby the previous negation can pass on to the next
word.1 At the same time, this allows for nested negations, as a word can first
introduce a negation scope and another subsequent negation can potentially revert
it based on ai−1 = Negated to follow a non-negating action again. In our case, we
incorporate the actual words into the states, while other variants are also possible,
such as using part-of-speech tags or word stems instead. The latter variants work
similarly; however, our tests suggest a lower out-of-sample performance.2
After observing the current state, the agent chooses an action at from of two
possibilities: (1) it can set the current word to negated or (2) it can mark it
as not negated. Hence, we obtain the following set of possible actions A =
{Negated, ¬Negated}. Based on the selected action, the agent receives a reward,
ri which updates the knowledge in the state-action function Q(si , ai ). This stateaction function is then used to infer the best possible action ai in each state si ,
i. e. the optimal policy π ∗ (si , ai ).
Our approach relies upon a text processing function that measures the correlation between a given gold standard at document level (e. g. the author’s rating in
movie reviews) and the content of a document. We later show possible extensions
1
Such a design is common in partially observable Markov decision processes (POMDP for
short) that feature a similar relaxation into so-called belief states [43].
2
By definition, the use of n-grams is not necessary, as the context is implicitly modeled by
the ordered sequence of states and actions.
14
(see Section 5.1), but for now demonstrate only how the tone (or sentiment) in
a document works as a predictor of its exogenous assessment. Examples of such
predicted variables are movie ratings in the case of reviews or stock market returns in the case of financial disclosures. Even though more advanced approaches
from machine learning are possible, we prefer – for reasons of clarity – an approach is based on pre-defined lists of positive and negative terms, Lpos and Lneg .
We then measure the tone Sd in document d as the difference between positively
and negatively opinionated terms divided by the overall number of terms in that
document [11], i. e.
Sd =
|{wi | wi ∈ Lpos }| − |{wi | wi ∈ Lneg }|
.
Nd
(2)
If a term is negated by the policy, the polarity of the corresponding term is
inverted, i. e. positively opinionated terms are counted as negative and vice versa.
Our list of opinionated words originates from the Harvard IV General Inquirer
dictionary which contains 1915 opinionated entries with a positive connotation
and 2291 entries marked as negative.
Let us now demonstrate the learning process via an example, where the agent
processes word-by-word the first document, which consists of “this is a good product”, with gold standard +1 (i. e. positive content). The agent might then, at
random, decide to explore the environment by negating the word “good ”. Upon
reaching the last word, it receives feedback in the form of a zero as there is no
improvement from having negation scopes compared to having none. Thus, the
agent will discard this action in the future. It then processes the second document:
(“this isn’t a good product” with gold standard −1), where it negates all words following “isn’t”. As this inversion now better reflects the gold standard, the agent
receives a positive reward and will apply this rule in the future. Ultimately, the
agent is also able to learn a suitable policy for nested negations, e. g. for the third
document “this product isn’t good but fantastic” with gold standard +1. Based
15
on the current policy, it negates all words following “isn’t” but receives a negative
reward as there is an inferior resulting correlation compared to not incorporating
negations. However, through further exploration, the agent learns that it is beneficial to terminate the negation scope after “but.” Thus, the agent will invert
all words subsequent to “isn’t” and terminate the negation scope subsequent to
“but” if (and potentially only if) the previous state is negated. Table 1 illustrates
an exemplary resulting state-action function for this learning process.
State=(wi , ai−1 )
ai = Negated
ai = ¬Negated
π ∗ (si , ai )
(this, ¬Negated)
3
6
¬Negated
(product, ¬Negated)
1
3
¬Negated
(isn’t, ¬Negated)
5
1
Negated
(good, Negated)
3
1
Negated
(but, Negated)
2
4
¬Negated
(fantastic, ¬Negated)
2
3
¬Negated
Table 1: Exemplary table for the state-action function Q(si , ai ) with recurrent state architecture and actions A = {Negated, ¬Negated}. Each cell contains the expected reward for the
corresponding state-action pairs. The last column shows optimal policy π ∗ (si , ai ) for all states
and actions.
We now specify a reward ri such that it incentivizes the outcome of the text
processing function to match the gold standard. When processing a document,
we cannot actually compute the reward (as not all negations are clear) until we
have processed all words. Therefore, we set the reward before the last word to
almost zero, i. e. ri ≈ 0 for all i = 1, . . . , Nd −1. Upon reaching the final word, the
agent compares the text processing function without any negation to the current
policy π ∗ . The former is defined by the absolute difference between gold standard
yd and tone Sd0 , whereas the latter is defined by the absolute difference between
gold standard yd and the adjusted tone using the current policy Sdπ . Then the
difference between the text processing functions returns the terminal reward rNd .
16
This results in the reward
0,
if ai = Negated and i < Nd ,
ri = c,
if ai = ¬Negated and i < Nd ,
|yd − Sd0 | − |yd − Sdπ | , if i = Nd ,
(3)
with constant c = 0.005 that adds a small reward for default (i. e. non-negating)
actions to avoid overfitting.
At the beginning, we initialize the action-value function Q(s, a), i. e. the current knowledge of the agent, to zero for all states and actions.3 The agent then
successively observes a sequence of words in which it can select between exploring
new actions or taking the current optimal one. This choice is made by ε-greedy
selection according to which the agent explores the environment by selecting a
random action with probability ε or, alternatively, exploits the current knowledge
with probability 1 − ε. In the latter case, the agent chooses the action with the
highest estimated reward for the given policy.4
3.5. Inferences, Understanding and Hypothesis Testing
Our approach features several beneficial characteristics that make inferences
and statistical testing easy. In contrast to black-box approaches in machine learning, we can use the state-action function Q(s, a) to infer rules regarding how the
content is processed because the function reflects the ground truth. For instance,
this state-action function specifically determines which cues introduce explicit or
3
This also controls our default action when encountering unknown states or words in the
out-of-sample dataset. In such cases, the non-negated action is preferred.
4
First, we perform 4000 iterations with a higher exploration rate as given by the following
parameters: exploration ε = 0.1 %, discount factor γ = 0 % and learning rate α = 0.5 %. In a
second phase, we run 1000 iterations for fine-tuning with exploration ε = 0.01 %, discount factor
γ = 0 % and learning rate α = 0.1 %. See [40, 41] for detailed explanations.
17
implicit negations. Additionally, we can gain a metric of confidence about the
rules by comparing the largest reward to all other rewards in a specific state. A
larger discrepancy expresses higher confidence with regard to a certain action.
Applying the policy to out-of-sample documents benchmarks its performance
in comparison to the absence of negation handling. Furthermore, we can study,
for instance, which cues prompt this policy to introduce negation scopes, as well
as their position, size or other characteristics as a basis for statistical testing.
4. Evaluating Negation Processing
This section evaluates our method for replicating human negation processing.
First, we show how policy learning can help to yield a more accurate interpretation
of movie reviews. We then detail the role of negation cues and compare explicit
versus implicit negations. In the next step, we validate the robustness of our
results and introduce a second application scenario which addresses the relevance
of accurate negation handling in financial disclosures.
4.1. Case Study: Recommender System
Recommender systems can benefit greatly from user-generated reviews, which
represent a rich source of information. We thus demonstrate our method using
a common dataset [44] of 5006 movie reviews from the Internet Movie Database
archive (IMDb), each annotated with an overall rating at document level.5 It is
widely accepted that measuring the tone of movie reviews is particularly difficult
because positive movie reviews often mention some unpleasant scenes, while negative reviews, conversely, often detail certain pleasant scenes [45]. Thus, this corpus
5
We
use
the
scaled
dataset
available
from
www.cs.cornell.edu/people/pabo/
movie-review-data/. All reviews are written by four different authors and preprocessed, e. g.
by removing explicit rating indicators [44].
18
appears particularly suitable for a case study since it allows one to examine the
importance of human-like negation processing beyond simple rules. Accordingly,
we use 10-fold cross validation to verify the predictive accuracy.
4.2. Policy Learning for Negation Processing
Shown below are the results from policy learning for negation processing. Table 2 summarizes the main results and we explicate these findings in more depth.
As part of a benchmark, we study the proportion of variance of the gold standard
that is explained by the tone when leaving negations untreated. We observe an
in-sample R2 of 0.0870 and a R2 of 0.0867 in the out-of-sample set. We then
compare this to our approach of policy learning. We perform 5000 learning iterations, thereby yielding significant improvements: the in-sample R2 increases by
158.94 %, leading to an overall R2 of 0.2233. Similarly, we see a rise by 58.94 % to
a R2 of 0.1378 in the out-of-sample set. A better handling of negations thus contributes to more accurate text processing (we later perform additional robustness
checks; see Section 4.5).
R2
(no negation handling)
R2
(with negation policy)
In-sample set
0.0870
0.2233
158.94
Out-of-sample set
0.0867
0.1378
58.94
Improvement
(in %)
Table 2: Comparison of gold standard variance explained by the tone. Figures are reported
for both the in-sample and out-of-sample sets using 10-fold cross validation after 5000 learning
iterations.
We now provide descriptive statistics of negation scopes in order to gain further
insights (Table 3). For this purpose, we apply the learned policy to the out-ofsample documents and record its effects. In the first place, the policy negates a
large share (38.11 %) of opinionated words, i. e. words that convey a positive or
negative polarity. On average, each document contains 75.18 separate negation
19
scopes of different size and extent, all of which invert opinion words. For example,
the length of the corresponding negation scopes, i. e. sequences that are uniformly
negated, is unevenly distributed and ranges from 1 to 18 words, whereas the
average length of each scope is 1.74 words. 75.46 % of all negation scopes consist
of only a single word, while 24.53 % encompass two or more words.
Minimum length of negation scopes
1
Maximum length of negation scopes
18
Mean length of negation scopes
1.74
Share of negation scopes with ≤ 1 word
75.46 %
Share of negation scopes with ≥ 2 word
24.53 %
Share of negated polarity words
38.11 %
Mean number of negation scopes per document
75.18
Table 3: Descriptive statistics on the out-of-sample set after applying the in-sample policy.
4.3. Negation Cues
Negation scopes are typically initiated by specific cues that invert the meaning
of surrounding words. These negation cues can be grouped in two categories. On
the one hand, a negation cue can be explicit, such as not in the sentence, “This
is not a terrible movie”. On the other hand, negations can also flip the meaning
of sentences implicitly, e. g. “The actor did a great job in his last movie; it was
the first and last time”.
Given this understanding, we investigate individual effects of explicit and implicit negations on text reception. For this purpose, we group the words that
initiate a negation scope according to their part-of-speech tag and depict in Figure 3 the resulting share of negation cues by word class. Here, the last bar relies
on a list of explicit negations as proposed by [34]. We thus find evidence that a
major share of negations (4 out of 8, i. e. 50.00 %) are evoked by explicit cues,
while the remainder originate from implicit negations.
20
Share of Negated Words in %
80.00
60.00
50.00
40.00
20.00
9.50
17.59
14.73
0.00
0.00
Adjective
Adverb
Conjunction
Implicit Negations
5.76
Noun
Verb
Negation
Explicit Negations
Figure 3: Negation cues per word class based on policy learned after 5000 iterations.
We now provide additional descriptions of the appearance of implicit negations. For instance, we frequently observe words such as fairly or hopefully as
part of implicit negation cues, which often transform a positive statement into a
negative one, e. g. “Hopefully, the movie is better next time”. Contrary to our
prior expectations, conjunctions seem unlikely to initiate a negation scope, but
are often accountable for double negations. Here, words such as but and nor frequently revert the meaning of negated words, i. e. terminate the negation scope,
as in the sentence, “The movie is not great but absolutely unmissable”.
Table 4 provides statistics for all explicit negation cues based on [34].6 Their
frequency in the documents differs considerably, while some cues also involve a
larger negation scope than others. For example, the negation word not negates
1.60 subsequent words on average, whereas this figure stands at 1.92 for the term
without. Based on the Q-value, we assess their strength, i. e. a larger value
6
Here, we divide the corpus into two subsets: an in-sample set of 4005 reviews which we
use to learn the agent, and (b) an out-of-sample set with the remaining 1001 documents to test
implications. We exclude the use of cross validation since we desire a single model with which
can perform statistical analyses.
21
indicates a higher reward from negating. We can also gain confidence in negations
by comparing the gap between the highest and second highest Q-value of each
word. Interestingly, several words that were previously considered negation cues
do not negate surrounding words in our case, namely, barely, less, hardly and
rarely.
Word
Negating Action
Q-Value
Confidence (Difference
to Second Best Policy)
Occurrences
Mean Length
of Negation Scope
not
3
0.0700
0.0456
1941
1.60
no
3
0.0696
0.0491
699
1.71
never
3
0.0680
0.0367
398
1.67
without
3
0.0779
0.0419
249
1.92
barely
7
–
–
–
–
less
7
–
–
–
–
hardly
7
–
–
–
–
rarely
7
–
–
–
–
Table 4: Explicit negation words from [34] according to policy learned after 5000 iterations.
This provides evidence that static negation lists are generally inadequate in
mimicking human perception. Even though explicit negations can be recognized
with predefined lists of cues, implicit ones are often hidden and difficult to identify
algorithmically. As a remedy to this shortcoming, our approach is capable of
learning both kinds of negations and can handle them accordingly.
4.4. Behavioral Implications of Negation Processing
Policy learning is also a valuable tool for analyzing behavioral implications.
In this section, we demonstrate potential policy learning applications that allow
for the testing of certain hypotheses regarding human information processing of
natural language.
As an example, our method allows one to test the hypothesis whether negations
appear evenly throughout different parts of narrative content. Such a test is not
tractable for rules or supervised learning with intermediate labeling, since these
22
introduce a subjective choice of negation cues, rules or labels; however, our method
infers a negation policy model from an exogenous response variable. Hence, we
can evaluate where authors place negations when composing reviews, i. e. do they
generally introduce negative aspects in the beginning or rather at the end? We
thus compare the frequency of negations (as a proxy for negativity) across different
parts of documents in the corpus. Let µ1 denote the mean of negated words in
the first half and µ2 in the second half of a document or sentence, respectively.
We can then test a null hypothesis H0 : µ1 = µ2 to infer behavioral implications.
As a result, we find that the second half of an out-of-sample document contains
1.38 % more negations than its first half on average. This difference is statistically
significant at even the 0.1 % significance level when performing a two-sided Welch
t-test. It also coincides with psychological research according to which senders
of information are more likely to place negative content at the end [46]; however,
we can provide evidence outside of a laboratory setting by utilizing human information behavior in a real-life environment. Furthermore, the share of negated
words also varies across different segments of sentences. However, at this level,
the effect tends in the opposite direction as the first half of a sentence in the
out-of-sample contains 0.09 % more negations than the second half. The latter is
also statistically significant at the 0.1 % level.
4.5. Robustness Checks
We investigate the convergence of the reinforcement learning process to a stationary policy. Accordingly, Figure 4 visualizes the proportion of variance of the
gold standard that is explained by the tone for the first 5000 learning iterations.
Here, the horizontal lines denote the explained variance in the benchmark setting
(i. e. no negation handling). Both the in-sample and out-of-sample R2 improve
relatively quickly and outperform the benchmark considerably. In the end, we
use our above policy based on 4000 iterations, since the next 1000 iterations con23
sistently show fluctuations below 0.05 % in terms of in-sample R2 . This pattern
indicates a fairly stationary outcome.
0.20
R2
0.15
Benchmark (no negation handling)
0.10
0.05
Explained variance (in−sample)
0.00
Explained variance (out−of−sample)
1000
2000
3000
Iteration
0
4000
5000
Figure 4: The fluctuating series show the converging explained variance of user ratings based
on tone (i. e. sentiment) across different learning iterations using 10-fold cross validation, while
the uniform lines shows it after smoothing. Here, the dark gray series corresponds to the insample set, whereas the light gray series corresponds to the out-of-sample set. The horizontal
line denotes the R2 in the benchmark setting (without handling negations).
Next, we compare the performance of our reinforcement learning approach to
common rules proposed in the literature [38, 37], which essentially try to imitate
the grammatical structure of a sentence. For this purpose, the negation rules
search for the occurrence of specific cues based on pre-defined lists and then
invert the meaning of a fixed number of surrounding words. Hence, we apply the
individual rules to each document and again compare the out-of-sample R2 ; see
Table 5 for results. Negating a fixed window of the next 4 words achieves the
highest fit among all rules similar to [47]. This rule exceeds the benchmark with
no negation handling by 10.84 %. Most importantly, our approach works even
more accurately, and dominates all of the rules, outperforming them by at least
43.39 %.
24
Approach
Correlation
Benchmark: no negation handling
0.0867
Negating all subsequent words
0.0840
Negating the whole sentence
0.0661
Negating a fixed window of 1 word
0.0918
Negating a fixed window of 2 words
0.0948
Negating a fixed window of 3 words
0.0953
Negating a fixed window of 4 words
0.0962
Negating a fixed window of 5 words
0.0961
Our approach (based on reinforcement learning)
0.1378
Table 5: Table compares the out-of-sample R2 (using 10-fold cross validation) of our approach
and different rules from previous literature.
Finally, we evaluated further setups and methods for handling negations. We
first tested alternative action sets for reinforcement learning that not only negate
single words but also whole phrases, including backward negations. However,
this configuration leads to inferior R2 values on both the in-sample and out-ofsample set. We also explored other dictionaries of opinionated words and the
performance of generative probabilistic models. Here, we find similar results for
alternative dictionaries but inferior results for generative probabilistic models. All
results confirm our findings.7
4.6. Comparison with Negation Processing in Financial News
Our second case study investigates information processing in financial markets by analyzing how qualitative content in financial disclosures influences stock
prices. For this purpose, we use 14,463 regulated ad hoc announcements8 from
European companies, all of which are written in English. These entail several
advantages; for example, ad hoc announcements must be authorized by company
7
8
Available on request.
Kindly provided by the Deutsche Gesellschaft für Ad-Hoc-Publizität (DGAP).
25
executives, their content is largely quality-checked by federal authorities and previous evidence finds a strong relationship between content and subsequent stock
market reaction [48]. As our gold standard, we calculate the daily abnormal return of the corresponding stock [49, 50]. We measure the tone in these disclosures
with the help of a finance-specific dictionary, the Loughran and McDonald dictionary [18]. This dictionary contains 354 entries with positive polarity, as well as
2350 entries marked as negative.
As detailed below, we derive a policy for negation processing and briefly introduce our main findings. Again, our reinforcement learning approach improves
the link between tone and market response. Our benchmark without negation
handling yields an out-of-sample R2 of 0.0042, while our method increases this by
37.02 %, resulting in an out-of-sample R2 for 10-fold cross validation of 0.0057.
Both the absolute R2 and its improvements are – as expected – higher for movie
reviews; this is a domain-specific disparity since “very few control variables predict
next-day returns” in efficient markets [51].
Next, we apply the learned policy to the out-of-sample documents and record
its effects. Interestingly, we find that negations are less frequent in financial news.
On average, each document contains 5.30 separate negation scopes that invert
10.88 % of all opinion words. The length of the corresponding negation scopes
is shorter, ranging from 1 to 14 words with an average length of 1.60 words.
Similarly, to the results for the movie reviews, we find that the end of a document
is more likely to contain negations than the beginning. On average, the second
half of an out-of-sample announcement contains 6.45 % more negation that its
first half. It is noteworthy that this difference is significant at the 10 % level using
a two-sided Welch t-test. This might suggest that authors of financial disclosures
utilize negations as a tool to convey negative information through positive words.
Overall, the results show that negations are domain-specific and depend on the
26
particularities of the chosen prose. Additionally, the comparison strongly affirms
that negation handling enhances the understanding of information processing for
natural language of an arbitrary domain.
5. Discussion
In the following sections, we discuss the implications of our research, as our
method not only improves text comprehension, but also suggests a new approach
to understanding decision-making in the social and behavioral sciences. Furthermore, our research is highly relevant for practitioners when extending information
systems with interfaces for natural language.
5.1. Extensibility
Our method of negation learning is not limited to the study of tone or sentiment; on the contrary, one can easily adapt it to all applications of natural
language processing which utilize a gold standard and where negations play an
important role. To accomplish this, one replaces the function calculating the sentiment with a corresponding counterpart that maps words onto a gold standard
for the given application. Our only requirement is that this function takes into
consideration – in some way – whether each word is negated or not.
To better illustrate this concept, we briefly describe how this works using
two examples. We first consider a medical question-answering system into which
users enter their symptoms and, in return, are provided a list of potential illnesses. The system bases its answers on a collection of medical reports and one
measures its performance by counting the number of correctly retrieved answers
relative to the given input. For example, the system should return “fever” for
input “flu” when the corpus contains “a flu causes fever”. Reinforcement learning can improve accuracy in the presence of negations; i. e. it learns that diseases
can also be unrelated to symptoms, as in the statement, “A flu does not result
27
in high blood pressure”. As a second example, we assume an information system
for negotiations that proposes offers to customers, who then reply in natural language. Subsequently, the information system automatically determines whether
a customer’s response was positive or negative based on its content. A naı̈ve
bag-of-words model considers only specific cues (such as accept without context),
whereas our approach can even learn to correctly classify cases with negations,
e. g. by appending the prefix “not ” to words that are negated [33].
We now generalize the reward function in order to search for an optimal negation policy for the above applications. For each document d with gold standard
yd , we calculate the predictive performance perf 0d that should forecast the gold
standard negation handling, as well as perf πd using the current policy π. The agent
then gains a reward
0,
if ai = Negated and i < Nd ,
ri = c,
if ai = ¬Negated and i < Nd ,
yd − perf 0d − |yd − perf πd | , if i = Nd
(4)
with a suitable constant c. The first two cases add a small reward c for default
(i. e. non-negating) actions to avoid overfitting, while the last case rewards how
much better the current policy approximates the gold standard compared to no
treatment of negations. This definition thus extends reinforcement learning to
seek optimal negation processing across almost arbitrary applications of natural
language processing.
5.2. Limitations
The current research faces a number of limitations, which can provide investigative possibilities for further works as follows: first and foremost, our method
exhibits shortcomings when language is intricate, such as when piece of text refers
28
to content that is located in an entirely different part of the document. Sometimes one even requires additional background knowledge to correctly interpret
the content, as in the statement, “The movie was not at all different from the last
one”. This complexity poses challenges to natural language processing – not only
for our method but also for those discussed in the related work. In addition, we
predominantly focus on implementations where negations have a forward-looking
scope, i. e. a negation cue affects subsequent words but not words that precede it.
Therefore, we have also tested variants with a backward-looking analysis as part of
our robustness checks; however, this offers opportunities for additional variations
with advanced actions which could, for instance, invert the meaning of the full
sentence or the subsequent object in order to further improve accuracy. Finally,
further effort is necessary to develop an unsupervised variant which eliminates the
need for a gold standard.
5.3. Implications for IS Research
The unique and enduring purpose of IS research as an academic field is to
understand and improve the ways in which people, organizations and businesses
create value with information [3, 4]. Hence, the design and implementation of systems to provide “the right information to the right person at the right time was
the raison d’être of early IS research” [32]. In the past, “information” predominantly referred to structured data, while companies nowadays also exploit unstructured data and especially textual materials. This development has found its
way into IS research, which thus focuses on how textual information is processed.
Among the earliest references to this area of inquiry is an article in Management
of Information Systems Quarterly from 1982 that explicitly addresses information
processing [52].
The field of information processing has gained great traction with advances
in neuroscience and NeuroIS [53]. By acquiring neuro-physiological data, schol29
ars can gather information on how the human brain reacts to external stimuli.
For this purpose, one measures (neuro-)physiological parameters (e. g. heart rate
and skin conductance) to study the information processing of human agents [54].
This makes it possible to measure informational and cognitive overload in users
in the course of their interaction with information systems [54] and text-based
information [55]. However, NeuroIS remains very costly and the methods exhibit
many weaknesses [56]. For example, interpreting data from functional magnetic
resonance imaging (fMRI) is hampered by the complexity and non-localizable
activities of the brain.
Our computational intelligence method promises to fill the gap in existing
approaches to understanding negations. It is analogous to revealed preferences
estimations in economics, where the choices of individuals reveal the individuals’
latent utility function, since we utilize text documents that are tagged by the
users with a rating or gold standard. In addition, applying computational intelligence offers the potential to automatically unveil negations in texts – without the
need to manually label individual words. This entails several advantages as the
understanding of language is highly subjective and, in contrast, we derive the (latent) negation model that best fits the data. The results can thus also contribute
to linguistic and psychological models of negation usage and representation [57].
Overall, reinforcement learning manifests immense potential for future IS research
involving the study of information processing in depth.
5.4. Implications for Practitioners
Previous IS research argues that “a basic issue in the design of expert systems
is how to equip them with representational and computational capabilities” [58].
As a remedy, this paper presents a tool to practitioners in order to improve the
automated processing of natural language in their information systems. As such,
our methodology can enhance the accuracy of decision support based on textual
30
data. It does not necessarily require changes in the derivation of the original
algorithms; instead, our methodology can be built on top of routines and thus
enables a seamless integration into an existing tool chain.
Practitioners can benefit from negation handling when assessing the semantic
orientation of written materials. For example, in the case of recommender systems
and opinion mining, texts provide decision support by tracking the public mood
in order to measure brand perception or judge the launch of a new product based
on blog posts, comments, reviews or tweets. Based on our case study, we see a
significant improvement of up to 58.94 % in explained variance by adjusting for
negated text units.
A better understanding of human language can spark business innovations
in multiple areas. For instance, our approach facilitates the interactive control
of information systems through natural language, such as in question-answering
systems. With the advent of cognitive computing, the accurate processing of
natural language will gain even more in importance [59]. Ultimately, the relevance
of our methodology goes beyond these examples and comprises almost all textbased applications of individuals, organizations and businesses.
6. Conclusion
Information is at the heart of all decision-making that affects humans, businesses and organizations. Consequently, understanding the formation of decisions
represents a compelling research topic and yet knowledge gaps become visible
when it comes to information processing with regard to natural language. Negations, for example, are a frequently utilized linguistic tool for expressing disapproval or framing negative content with positive words; however, existing methods
struggle to accurately recognize and interpret negations in written text. Moreover, these methods are often tethered to large volumes of manually labeled data,
31
which introduce an additional source of subjectivity and noise.
In order to address these shortcomings, this paper develops a novel approach
based on reinforcement learning, which has the advantage of being human-like
and thus capable of learning to replicate human decision-making. As a result, our
evaluation shows superior performance in predicting negation scopes, while this
method also reveals an unbiased approach to identifying negation scopes based on
an exogenous response variable collected at document level. It thereby sheds light
on the “ground truth” of negation scopes, which would have otherwise been latent
and unobservable. In addition, reinforcement learning allows for hypothesis testing in order to pinpoint how humans process and act on negations. For instance,
this paper demonstrates that negations are unequally distributed across document
segments, showing that the second half of movie reviews and financial news items
contain significantly more negations than the first half. Our approach serves as an
intriguing alternative or supplement to experimental research, as it unleashes computational intelligence for the purpose of performing behavioral research, thereby
fostering unprecedented insights into human information processing.
References
[1] D. LaBerge, S. Samuels, Toward a Theory of Automatic Information Processing in Reading, Cognitive Psychology 6 (1974) 293–323.
[2] W. Schneider, R. M. Shiffrin, Controlled and Automatic Human Information Processing: Detection, Search, and Attention, Psychological Review 84
(1977) 1–66.
[3] R. O. Briggs, Special Section: Cognitive Perspectives on Information Systems, Journal of Management Information Systems 31 (2015) 3–5.
[4] J. F. Nunamaker, R. O. Briggs, Toward A Broader Vision for Information
32
Systems, ACM Transactions on Management Information Systems 2 (2011)
1–12.
[5] M. C. Lacity, M. A. Janson, Understanding Qualitative Data: A Framework
of Text Analysis Methods, Journal of Management Information Systems 11
(1994) 137–155.
[6] M. Chau, J. Xu, Business Intelligence in Blogs: Understanding Consumer
Interactions and Communities, MIS Quarterly 36 (2012) 1189–1216.
[7] S. Vodanovich, D. Sundaram, M. Myers, Digital Natives and Ubiquitous
Information Systems, Information Systems Research 21 (2010) 711–723.
[8] J. Hirschberg, C. D. Manning, Advances in Natural Language Processing,
Science 349 (2015) 261–266.
[9] R. E. Vlas, W. N. Robinson, Two Rule-Based Natural Language Strategies
for Requirements Discovery and Classification in Open Source Software Development Projects, Journal of Management Information Systems 28 (2012)
11–38.
[10] N. P. Cruz, M. Taboada, R. Mitkov, A Machine–Learning Approach to
Negation and Speculation Detection for Sentiment Analysis, Journal of the
Association for Information Science and Technology In Press (2015).
[11] B. Pang, L. Lee, Opinion Mining and Sentiment Analysis, Foundations and
Trends in Information Retrieval 2 (2008) 1–135.
[12] I. G. Councill, R. McDonald, L. Velikovich, What’s Great and What’s Not:
Learning to Classify the Scope of Negation for Improved Sentiment Analysis,
in: Proceedings of the Workshop on Negation and Speculation in Natural
33
Language Processing, Association for Computational Linguistics, Stroudsburg, PA, USA, 2010, pp. 51–59.
[13] N. Archak, A. Ghose, P. G. Ipeirotis, Deriving the Pricing Power of Product
Features by Mining Consumer Reviews, Management Science 57 (2011) 1485–
1509.
[14] T. T. Thet, J.-C. Na, C. S. G. Khoo, Aspect-Based Sentiment Analysis
of Movie Reviews on Discussion Boards, Journal of Information Science 36
(2010) 823–848.
[15] M. L. Jensen, J. M. Averbeck, Z. Zhang, K. B. Wright, Credibility of Anonymous Online Product Reviews: A Language Expectancy Perspective, Journal
of Management Information Systems 30 (2013) 293–324.
[16] E. Henry, Are Investors Influenced by How Earnings Press Releases are
Written?, Journal of Business Communication 45 (2008) 363–407.
[17] P. C. Tetlock, Giving Content to Investor Sentiment: The Role of Media in
the Stock Market, Journal of Finance 62 (2007) 1139–1168.
[18] T. Loughran, B. McDonald, When Is a Liability Not a Liability? Textual
Analysis, Dictionaries, and 10-Ks, Journal of Finance 66 (2011) 35–65.
[19] N. P. Cruz Dı́az, Maña López, Manuel J., J. M. Vázquez, V. P. Álvarez,
A Machine-Learning Approach to Negation and Speculation Detection in
Clinical Texts, Journal of the American Society for Information Science and
Technology 63 (2012) 1398–1410.
[20] L. Rokach, R. Romano, O. Maimon, Negation Recognition in Medical Narrative Reports, Information Retrieval 11 (2008) 499–538.
34
[21] N. A. Johnson, R. B. Cooper, Understanding the Influence of Instant Messaging on Ending Concessions During Negotiations, Journal of Management
Information Systems 31 (2015) 311–342.
[22] H. Lai, W.-J. Lin, G. E. Kersten, Effects of Language Familiarity on eNegotiation: Use of Native vs. Nonnative Language, in: Proceedings of the
42nd Hawaii International Conference on System Sciences (HICSS), IEEE,
2002, pp. 1–9.
[23] D. P. Twitchell, M. L. Jensen, D. C. Derrick, J. K. Burgoon, J. F. Nunamaker, Negotiation Outcome Classification Using Language Features, Group
Decision and Negotiation 22 (2013) 135–151.
[24] C. M. Fuller, D. P. Biros, J. Burgoon, J. Nunamaker, An Examination
and Validation of Linguistic Constructs for Studying High-Stakes Deception,
Group Decision and Negotiation 22 (2013) 117–134.
[25] L. Zhou, J. K. Burgoon, D. P. Twitchell, T. Qin, J. F. Nunamaker, A
Comparison of Classification Methods for Predicting Deception in ComputerMediated Communication, Journal of Management Information Systems 20
(2004) 139–166.
[26] Y. Niv, R. Daniel, A. Geana, S. J. Gershman, Y. C. Leong, A. Radulescu,
R. C. Wilson, Reinforcement Learning in Multidimensional Environments
Relies on Attention Mechanisms, Journal of Neuroscience 35 (2015) 8145–
8157.
[27] S. J. Gershman, A. Radulescu, K. A. Norman, Y. Niv, O. Sporns, Statistical Computations Underlying the Dynamics of Memory Updating, PLoS
Computational Biology 10 (2014).
35
[28] J. Fiser, P. Berkes, G. Orbán, M. Lengyel, Statistically Optimal Perception
and Learning: From Behavior to Neural Representations, Trends in Cognitive
Sciences 14 (2010) 119–130.
[29] T. D. Wilson, Models in Information Behaviour Research, Journal of Documentation 55 (1999) 249–270.
[30] S. Stieglitz, L. Dang-Xuan, Emotions and Information Diffusion in Social Media: Sentiment of Microblogs and Sharing Behavior, Journal of Management
Information Systems 29 (2013) 217–248.
[31] D. N. Trung, T. T. Nguyen, J. J. Jung, D. Choi, Understanding Effect of
Sentiment Content Toward Information Diffusion Pattern in Online Social
Networks: A Case Study on TweetScope, in: P. C. Vinh, V. Alagar, E. Vassev, A. Khare (Eds.), Context-Aware Systems and Applications, volume 128
of Lecture Notes for Computer Sciences, Social Informatics and Telecommunications Engineering, Springer, Cham, Switzerland, 2014, pp. 349–358.
[32] R. Aggarwal, R. Gopal, R. Sankaranarayanan, P. V. Singh, Blog, Blogger,
and the Firm: Can Negative Employee Posts Lead to Positive Outcomes?,
Information Systems Research 23 (2012) 306–322.
[33] M. Wiegand, A. Balahur, B. Roth, D. Klakow, A. Montoyo, A Survey on the
Role of Negation in Sentiment Analysis, in: Proceedings of the Workshop on
Negation and Speculation in Natural Language Processing, Association for
Computational Linguistics, Stroudsburg, PA, USA, 2010, pp. 60–68.
[34] L. Jia, C. Yu, W. Meng, The Effect of Negation on Sentiment Analysis and
Retrieval Effectiveness, in: D. Cheung (Ed.), Proceeding of the 18th ACM
Conference on Information and Knowledge Management (CIKM ’09), ACM,
New York, NY, 2009, pp. 1827–1830.
36
[35] N. Pröllochs, S. Feuerriegel, D. Neumann, Enhancing Sentiment Analysis of
Financial News by Detecting Negation Scopes, in: 48th Hawaii International
Conference on System Sciences (HICSS), 2015, pp. 959–968.
[36] N. Pröllochs, S. Feuerriegel, D. Neumann, Detecting Negation Scopes for
Financial News Sentiment Using Reinforcement Learning, in: 49th Hawaii
International Conference on System Sciences (HICSS), 2016, pp. 1164–1173.
[37] M. Taboada, J. Brooke, M. Tofiloski, K. Voll, M. Stede, Lexicon-Based
Methods for Sentiment Analysis, Computational Linguistics 37 (2011) 267–
307.
[38] A. Hogenboom, P. van Iterson, B. Heerschop, F. Frasincar, U. Kaymak, Determining Negation Scope and Strength in Sentiment Analysis, in: IEEE
International Conference on Systems, Man, and Cybernetics, 2011, pp. 2589–
2594.
[39] S. Padmaja, S. Fatima, S. Bandu, Evaluating Sentiment Analysis Methods and Identifying Scope of Negation in Newspaper Articles, International
Journal of Advanced Research in Artificial Intelligence 3 (2014) 1–6.
[40] R. S. Sutton, A. G. Barto, Reinforcement Learning: An Introduction, Adaptive Computation and Machine Learning, MIT Press, Cambridge, MA, 1998.
[41] C. J. C. H. Watkins, P. Dayan, Q-Learning, Machine Learning 8 (1992)
279–292.
[42] J. Hu, M. P. Wellman, Nash Q-Learning for General-Sum Stochastic Games,
Journal of Machine Learning Research 4 (2003) 1039–1069.
[43] L. P. Kaelbling, M. L. Littman, A. W. Moore, Reinforcement Learning: A
Survey, Journal of Artificial Intelligence Research 4 (1996) 237–285.
37
[44] B. Pang, L. Lee, Seeing Stars: Exploiting Class Relationships for Sentiment
Categorization with Respect to Rating Scales, in: Proceedings of the 43rd
Annual Meeting on Association for Computational Linguistics (ACL ’05),
2005, pp. 115–124.
[45] P. D. Turney, Thumbs Up or Thumbs Down? Semantic Orientation Applied
to Unsupervised Classification of Reviews, Association for Computational
Linguistics, 2002.
[46] A. M. Legg, K. Sweeny, Do You Want the Good News or the Bad News First?
The Nature and Consequences of News Order Preferences, Personality and
Social Psychology Bulletin 40 (2014) 279–288.
[47] M. Dadvar, C. Hauff, F. de Jong, Scope of Negation Detection in Sentiment
Analysis, in: Proceedings of the Dutch-Belgian Information Retrieval Workshop, University of Amsterdam, Amsterdam, Netherlands, 2011, pp. 16–20.
[48] J. Muntermann, A. Guettler, Intraday Stock Price Effects of Ad Hoc Disclosures: The German Case, Journal of International Financial Markets,
Institutions and Money 17 (2007) 1–24.
[49] Y. Konchitchki, D. E. O’Leary, Event Study Methodologies in Information
Systems Research, International Journal of Accounting Information Systems
12 (2011) 99–115.
[50] A. C. MacKinlay, Event Studies in Economics and Finance, Journal of
Economic Literature 35 (1997) 13–39.
[51] P. C. Tetlock, M. Saar-Tsechansky, S. Macskassy, More Than Words: Quantifying Language to Measure Firms’ Fundamentals, Journal of Finance 63
(2008) 1437–1467.
38
[52] D. Robey, W. Taggart, Human Information Processing in Information and
Decision Support Systems, MIS Quarterly 6 (1982) 61.
[53] A. Dimoka, P. A. Pavlou, F. D. Davis, NeuroIS: The Potential of Cognitive Neuroscience for Information Systems Research: Research Commentary,
Information Systems Research 22 (2011) 687–702.
[54] R. Riedl, F. D. Davis, A. R. Hevner, Towards a NeuroIS Research Methodology: Intensifying the Discussion on Methods, Tools, and Measurement,
Journal of the Association for Information Systems 15 (2014) 1–35.
[55] R. K. Minas, R. F. Potter, A. R. Dennis, V. Bartelt, S. Bae, Putting on the
Thinking Cap: Using NeuroIS to Understand Information Processing Biases
in Virtual Teams, Journal of Management Information Systems 30 (2014)
49–82.
[56] A. Dimoka, R. D. Banker, I. Benbasat, F. D. Davis, A. R. Dennis, D. Gefen,
A. Gupta, A. Ischebeck, P. H. Kenning, P. A. Pavlou, et al., On the Use of
Neurophysiological Tools in IS Research: Developing a Research Agenda for
NeuroIS, MIS Quarterly 36 (2012) 679–702.
[57] S. Khemlani, I. Orenes, P. N. Johnson-Laird, Negation: A Theory of Its
Meaning, Representation, and Use,
Journal of Cognitive Psychology 24
(2012) 541–559.
[58] W.-R. Zhang, POOL: A Semantic Model for Approximate Reasoning and
Its Application in Decision Support, Journal of Management Information
Systems 3 (1987) 65–78.
[59] D. S. Modha, R. Ananthanarayanan, S. K. Esser, A. Ndirango, A. J. Sherbondy, R. Singh, Cognitive Computing, Communications of the ACM 54
(2011) 62.
39
| 2cs.AI
|
1
Deterministic Annealing Optimization for
Witsenhausen’s and Related Decentralized
Stochastic Control Problems
arXiv:1607.02893v1 [cs.SY] 11 Jul 2016
Mustafa Said Mehmetoglu, Student Member, IEEE, Emrah Akyol, Member, IEEE,
and Kenneth Rose, Fellow, IEEE
Abstract—This note studies the global optimization of controller mappings in discrete-time stochastic control problems
including Witsenhausen’s celebrated 1968 counter-example. We
propose a generally applicable non-convex numerical optimization method based on the concept of deterministic annealing
– which is derived from information theoretic principles and
was successfully employed in several problems including vector
quantization, classification and regression. We present comparative numerical results for two test problems that show strict
superiority of the proposed method over prior approaches in
literature.
Index Terms—Decentralized control, Optimization methods,
Numerical simulation, Physical models.
I. I NTRODUCTION
Decentralized control systems have multiple controllers
designed to collaboratively achieve a global objective while
taking actions based on local observations. One of the most
studied structures, termed “linear quadratic Gaussian” (LQG),
involves linear dynamics, quadratic cost functions and Gaussian variables. Since in the case of centralized LQG problems,
the optimal mappings are linear, it was naturally conjectured
that linear control mappings remain optimal in decentralized
settings. However, Witsenhausen proposed an example of a
decentralized LQG control problem, commonly referred to
as Witsenhausen’s counter-example (WCE), for which he
provided a simple non-linear control strategy that outperforms
all linear strategies [1]. The problem has been viewed as a
benchmark in stochastic networked control, see, e.g., [2] for
a detailed treatment.
Decentralized control systems such as WCE arise in many
practical applications, and numerous variations on WCE have
been studied in the literature (see, e.g., [3]–[8]). In general,
linear control strategies are not optimal for LQG systems,
except when the system admits some specific information
structure (see, e.g., [9], [10]). It is well-understood that if
the information structure in a decentralized control problem
Mustafa S. Mehmetoglu and Kenneth Rose are with the Department of
Electrical and Computer Engineering, University of California, Santa Barbara,
CA, 93106, USA (e-mail: [email protected], [email protected])
Emrah Akyol is with the Coordinated Science Laboratory, University of Illinois at Urbana-Champaign, Urbana, IL, 61801, USA (email:
[email protected])
This work is supported by the NSF under grants CCF-1118075 and CCF1016861. The material in this paper was presented in part at the IEEE
International Symposium on Information Theory (ISIT), Honolulu, HI, USA,
June 2014
is nonclassical, as in the case of WCE, non-linear strategies
may widely outperform optimal linear strategies. Finding the
optimal mappings for such problems is usually a difficult task
unless they admit an explicit (and often as simple as linear)
solution [3].
Recent research efforts have focused on developing efficient
numerical methods for decentralized control problems [11]–
[14], specifically for WCE [15]–[18]. Some of the existing
methods rely on simplifying properties of WCE, such as
monotonicity, and are therefore not easily generalizable [11],
[12], [19]. Moreover, methods that require analytical derivation
for each particular setting are not fully automated [14], [19].
In this work, building on our prior work [20], we propose
an optimization method, based on the concept of deterministic
annealing (DA), for a class of decentralized stochastic control
problems. DA has been successfully used in various problems
in control theory including dynamic coverage control problems
[21]–[23] and cluster analysis in control systems [24], in
addition to other problems involving non-convex optimization
such as vector quantization [25], regression [26], zero-delay
source-channel coding [27], and more (see review in [28]).
There are many important advantages of the proposed method
compared to prior work, including ability to avoid poor local
minima and independence from initialization.
We demonstrate the DA-based method on two specific
problems. We first analyze the numerically ”over-mined” WCE
problem. We then study a more involved variation, introduced
in [7], which includes an additional noisy channel over which
the two controllers communicate. The second controller, therefore, has access to some side information which is controlled
by the first controller. We refer to this setting as the “side
channel problem” motivated by the class of ”decoder side
information” problems in communications and information
theory [29]. It has been demonstrated in [7] that non-linear
strategies may outperform the best linear strategies, however,
the question of how to approach the optimal solution remains
open.
Having a powerful optimization method at hand, we analyze the structure of experimentally obtained mappings. For
instance, Wu and Verdú have shown [30] that the optimal
solution of WCE must have real analytic left inverse, thus,
a piecewise linear function cannot be optimal. Our numerical
results demonstrate that the “steps” in obtained mappings show
small deviations from linear, experimentally confirming this
theoretical finding.
2
II. P ROBLEM D EFINITION
A. Derivation
A. Notation
Let R, E(·) and P(·) denote the set of real numbers, the
expectation and probability operators, respectively. We represent random variables and their realizations with uppercase
and lowercase letters (e.g., X and x), respectively. Let Xij
denote the set Xi , . . . , Xj . The probability density function of
the random variable X is fX (x). The Gaussian density with
mean µ and standard deviation σ is denoted as N (µ, σ 2 ). We
use natural logarithms which, in general, may be complex, and
the integrals are, in general, Lebesgue integrals.
B. General Problem Definition
Formally, we consider a discrete-time stochastic control
problem with nonclassical information pattern involving n
controllers, and assume that the order of control actions is
fixed in advance, i.e., the system is sequential [31]. Following
the problem definition in [31], let (Ω, B, P) be a probability
space, where Ω denotes the random quantities involved in the
system such as initial input, and (Ui , Σi ), for i = 1, . . . , n,
are measurable spaces with Ui denoting the set of control
actions. Controller mappings (functions) are denoted by gi :
Ω×U1 ×U2 . . . Ui−1 → Ui , for i = 1, . . . , n. For convenience,
we denote the input set of gi by Xi = Ω × U1 × U2 . . . Ui−1 .
The system is then defined by the following set of equations
ui = gi (xi ), i = 1, . . . , n.
(1)
Let f be a real-valued and bounded measurable function of
ω, un1 on (Ω, B), i.e., f is a random variable. The problem
objective is to find the set of functions g1n that minimize the
value of the cost function J:
J = E{f (ω, un1 )}.
(2)
III. P ROPOSED M ETHOD
A quick overview of the proposed method at the high
level is as follows: We introduce controlled randomization
into the optimization process by randomizing the controller
mappings, and impose a constraint on the level of randomness
(measured by the Shannon entropy) while minimizing the
expected cost of the system. The resultant Lagrangian functional can be viewed as the “free energy” of a corresponding
physical system, wherein the Lagrangian parameter is the
“temperature”. The optimization is equivalent to an annealing
process that starts by minimizing the cost (free energy) at a
high temperature, which effectively maximizes the entropy.
The minimum cost is then tracked at successively lower
temperatures as the system typically undergoes a sequence of
phase transitions through which the complexity of the solution
(controller mappings) grows. As the temperature approaches
zero, hard (nonrandom) mappings are obtained. DA avoids
poor local minima through randomization of mappings, slowly
tracks the minimum cost, and can achieve the global minimum
under certain conditions on the type (and continuity) of
phase transitions. However, there is no general guarantee of
convergence to the globally optimal solution.
Let us denote the space of controller input xi by Rxi , for i =
1, . . . , n. Assume there exists a partition of Rxi into Mi > 0
disjoint regions denoted by Ri,mi (mi = 1, . . . , Mi ):
M
[i
Ri,mi = Rxi .
(3)
mi =1
Note that each value of xi belongs to exactly one of the
partition regions, referred to as a deterministic (non-random)
partition.
We begin our formulation by imposing a piecewise structure
on the controller mappings. Consider the structured mapping
gi , for i = 1, . . . , n, written as
gi (xi ) = gi,mi (xi ) for xi ∈ Ri,mi .
(4)
Each gi,mi (xi ) is a parametric function referred to as “local
model”. Effectively, each of the mappings gi is defined with
a structure determined by two components: a space partition
where regions are denoted by Ri,mi and a parametric local
model per partition cell, i.e., gi,mi (xi ) for Ri,mi . The number
of local models (partition regions) for mapping gi is Mi .
The local models can take any prescribed form such as
linear, quadratic or Gaussian and we let Λ(gi,mi ) denote the
parameter set for local model gi,mi .
The crucial idea in DA is to introduce controlled randomization into the problem formulation. We replace the deterministic
partition of space by a random partition, i.e., we associate
every input point (xi ) with partition regions in probability. To
this end, we introduce random variables Mi , for i = 1, . . . , n,
whose realization is the partition index mi . We define the
association probabilities as conditional distribution on the
partition index given the input:
pi (mi |xi ) = P{xi ∈ Ri,mi } = P{gi (xi ) = gi,mi (xi )}, (5)
for i = 1, . . . , n. Consequently, the mappings are now random,
in the sense that the output of controller gi for an input xi is
given in probability as
gi (xi ) = gi,mi (xi ) with probability pi (mi |xi ).
(6)
By construction, we have that given Xi , Mi is independent of
the random variables M1i−1 .
The expectation in (2) is now taken over Xin and Min . Let
us rewrite it for a fixed value of i as follows:
Z X
Mi
J=
E{f (ω, un1 )|mi , xi }pi (mi |xi )fxi (xi )dxi (7)
xi m =1
i
where E{f (ω, un1 )|mi , xi } can be viewed as the cost of associating xi with local model gi,mi . Assuming fixed local model
parameters, optimizing (7) with respect to pi (mi |xi ) would
clearly produce deterministic mappings, since the minimum is
achieved by setting pi (mi |xi ) = 1 for the pair {mi , xi } for
which E{f (ω, un1 )|mi , xi } is minimum. Therefore, the ultimate objective of obtaining optimum deterministic controllers
is preserved as the random encoders share the same global
minimum as deterministic ones. However, direct optimization
of the cost with respect to pi (mi |xi ) results in poor local
3
minima. Instead, we minimize (2) at prescribed levels of
randomness, which we measure by the Shannon entropy. The
joint entropy of the system can be written
H(X1n , M1n ) = H(X1 ) +
n
X
i=2
+H(M1 |X1 ) +
H(Xi |M1i−1 , X1i−1 )
n
X
H(Mi |Xi , M1i−1 , X1i−1 ).
n
X
H(Mi |Xi ).
i=2
(8)
It is easy to see from conditional independence arguments
that the conditional entropies in the second term in the
righthand side of (8) can be simplified to H(Xi |X1i−1 ), and
those in the last term to H(Mi |Xi ). Thus the first two terms
of (8) are fixed and determined by the problem statement (the
joint distribution of X1n ). We therefore discard the first two
fixed terms of (8), rearrange the remaining terms, to obtain a
conveniently compact measure of randomness defined as
H,
i=1
(9)
The conditional entropy H(Mi |Xi ) is given by
Z X
Mi
p(mi |xi ) log p(mi |xi )fXi (xi )dxi .
H(Mi |Xi ) = −
xi m =1
i
Accordingly, we construct the Lagrangian
F = J − TH
(10)
(11)
as the objective function to be minimized, where J is given
in (2), H is given in (9) and T is the Lagrange multiplier
associated with the entropy constraint. The notational choices
are to emphasize the analogy to statistical physics, where
F can be viewed as the Helmholtz free energy, J as the
thermodynamic energy, H as the entropy, and T as the
temperature, of a corresponding physical system (see [28] for
a detailed treatment of the statistical physics analogy).
B. Optimization Method
The practical method consists of gradually reducing the
temperature T while tracking the minimum of the free energy F . The central iteration at each temperature consists
of optimizing the local model parameters and association
probabilities. Initially, for T → ∞, the minimum F is obtained
by maximizing H, which is achieved by uniform association
probabilities as can be seen from (9). Consequently, for each
controller, the local models are all optimized for the same
distribution over the input space, and are therefore all identical,
i.e., there is effectively a single “distinct” local model. As
the temperature is decreased, the system will undergo “phase
transitions” where the current solution no longer represents a
minimum (the old minimum typically becomes a saddle point)
and there exists a better solution with increased number of
distinct local models. Since it is computationally more efficient
to keep only distinct models, we initialize with a single model
(Mi = 1) and trigger phase transitions by duplicating existing
models and slightly perturb them at each temperature. The
minimization of F will either produce the existing solution
(the duplicated and perturbed local models seek to merge
at their initial values), or a better solution with increased
number of distinct models, depending on whether a “critical
temperature” has been reached. Accordingly, in we combine
identical models. At the limit T → 0, we perform zero
temperature iteration (equivalent to gradient descent) by fully
assigning source points to the model that makes the smallest
contribution to J, thus obtaining the desired deterministic
mappings. (This is referred to as ”quenching” in the physical
analogy.)
Remark 1: Our method is derived without recourse to discretization. Although practical simulations involve sampling
of the continuous space during numerical computations of
integrals, this is in contrast to methods that are entirely
formulated in discrete settings.
Remark 2: Critical temperatures can be derived analytically
if, for the problem considered, phase transitions are of “continuous” nature, in the sense that tracked minimum becomes
a saddle point at the exact critical temperature. The condition
for saddle point can be obtained using variational calculus,
see [28] for phase transition analysis in DA. Our experiments
indicate that, at least for the test cases considered in this paper,
phase transitions are not continuous. While pre-calculating the
critical temperature may enable a numerical speed up of the
annealing process, it is not necessary to implementing the
practical algorithm. Hence, the derivation and characteristics
of phase transitions are kept outside the scope of this paper.
The optimal pi (mi |xi ) that minimize (11) can be derived
in closed form. Plugging (7) and (10) in (11), straightforward
derivation gives the optimal pi (mi |xi ) as
n
e−E{f (ω,u1 )|mi ,xi }/T
pi (mi |xi ) = P −E{f (ω,un )|m ,x }/T
i i
1
e
(12)
mi
Optimization of parameters in Λ(gmi ) can be done using
any standard method. Typically, a variant of gradient descent
is used when closed form expressions cannot be obtained.
IV. A PPLICATIONS OF THE P ROPOSED DA M ETHOD
A. Witsenhausen’s Counterexample
1) Problem Description: Let X0 and W be Gaussian ran2
dom variables with distributions N (0, σX
) and N (0, 1), re0
spectively. WCE is a 2-stage control problem with controllers
g1 : R → R and g2 : R → R, defined by the following
equations:
U1 = g1 (X0 ),
U2 = g2 (X1 + W ),
X1 = X0 + U1 , X2 = X1 − U2 .
(13)
The schematic representation is given in Figure 1a. The
objective is to minimize the cost
J = E{k 2 U12 + X22 }.
(14)
For convenience, we define f1 (X0 ) = g1 (X0 ) + X0 .
Some properties of f1 (·) are known, including the property
of symmetry about the origin (thus, positive half is enough to
describe a given solution) [1]. Witsenhausen has provided the
4
TABLE I
R ESULTS FOR WCE
Solution
Optimal linear Solution
1-step, Witsenhausen [1]
2-step, [11]
Sloped 2.5 - step, [15]
Sloped 3.5 - step, [19]
Sloped 3.5 - step, [17]
Sloped 4 - step, [16]
Sloped 5 - step, our result
(a)
(b)
Fig. 1. Settings used for testing the proposed method (a) Witsenhausen’s
counter-example (b) Side channel problem.
following solution that outperforms the optimal linear solution
for a given set of problem parameters (k = 0.2, σX0 = 5):
f1 (x0 ) = σX0 sgn(x0 )
(15)
where sgn(·) is the signum function. Since there is a single
“step” in the positive half of real line, this solution is referred
to as a “1-step” solution. Improved solutions that appeared in
literature utilize 2.5, 3, 3.5 and 4-step functions (an x.5 step
function has a step that straddles the origin). Moreover, the
latter solutions made improvements by using slightly sloped
steps rather than constant ones.
Although in standard application of DA-based method we
randomize all controllers, for computational efficiency, we
restrict the randomization to only g1 as
g1 (x0 ) = g1,m1 (x0 ) with probability p1 (m1 |x0 )
(16)
and numerically compute (update) g2 by using the fact that
optimal g2 given g1 is
g2 (Y2 ) = E{X1 |Y2 }
(17)
where Y2 = X1 + W .
For this particular problem, we use linear local models given
by
g1,m1 (x0 ) = a1,m1 x0 + b1,m1 .
(18)
while noting that optimal g1 must have analytic left inverse and
hence cannot be piecewise linear [30]. Nevertheless, the minimal cost can be approached arbitrarily closely by piecewise
linear functions [30]. Thus, for numerical algorithms, linear
models are sufficient.
2) Results for WCE: Preliminary results on the application
to WCE appeared in [20]. We first provide results for the standard benchmark case where k = 0.2, σX0 = 5 that was used in
many papers in literature. The annealing process is illustrated
in Figure 2, where evolution of the mapping can be seen (only
the positive half is shown thanks to the symmetry property).
The obtained mapping is referred to as a “sloped 5-step”
solution. At high temperature, there is only one local model,
thus, the function is 1-step. As the temperature is lowered, the
solution undergoes phase transitions, revealing more steps for
the mapping function. In this work we calculated the solution
with 5 steps. Although more steps possibly exist, improvement
to cost is numerically insignificant with additional steps. Some
Cost
0.96
0.404253
0.190
0.1701
0.1673132
0.1670790
0.16692462
0.16692291
earlier results from the literature are given in Table I, where
it can be seen that our method produced the minimum cost
achieved to date.
Another benchmark case, k = 0.63, was suggested in [5] as
potentially being more relevant for confirming the high gains
of optimal non-linear mappings. Our resulting mapping for
k = 0.63, σX0 = 5 is given in Figure 3a, which is a 6-step
solution. Our numerical results suggest that the gain over linear
solution is smaller compared to the standard benchmark case
above: J = 0.844 for the solution in Figure 3a whereas cost
associated with the optimal linear mapping is J = 0.961.
These numerical results illustrate an important theoretical
result as well. In [30] authors proved that the optimal f1 must
have analytic left inverse and therefore cannot be piecewise
linear, which was believed to be the case due to numerical
results (see, e.g., [19]). Our numerical results indicate that
steps are in fact non-linear, as shown in Figure 3b. The steps
become non-linear during the final stages of the algorithm as
multiple local models appear to form a single step. To the best
of our knowledge, this is the first numerical result illustrating
non-linearity of the steps.
B. Side Channel Problem
1) Problem Description: Let X0 be a Gaussian random
2
variable with distribution N (0, σX
), and W1 , W2 be inde0
pendent Gaussian random variables, both with a distribution
N (0, 1). The system is defined by the following equations:
U1 = g1 (X0 ), U2 = g2 (X0 ), U3 = g3 (X1 + W1 , U2 + W2 ),
X1 = X0 + U1 ,
X2 = X1 − U3 .
(19)
The problem is to optimize the cost function
J = E{k 2 U12 + X22 }
(20)
for given σX0 and positive parameter k, subject to a power
constraint on U2 :
bSN R = E{U22 }
(21)
where bSN R is the specified power level. We again define
f1 (X0 ) = g1 (X0 ) + X0 .
This problem setting is illustrated in Figure 1b and was introduced in [7]. It can be seen as a generalization of WCE with
an additional communication channel between the controllers,
i.e., a non-linear function of input X0 is communicated by
g2 to the controller denoted by g3 : R2 → R. The non-linear
mappings analyzed in [7], which widely outperform the best
linear solution in a large range of bSN R , are such that both f1
and g2 are staircase functions of x0 .
5
T = 5.000, D = 0.3510, H = 0.693
T = 0.275, D = 0.1898, H = 0.121
T = 0.153, D = 0.1729, H = 0.068
T = 0.021, D = 0.1671, H = 0.009
T = 0.005, D = 0.1670, H = 0.002
30
30
30
30
30
15
15
15
15
15
f 1 (x 0 )
0
0
0
10
20
30
0
0
10
20
30
0
0
10
20
30
0
0
10
20
30
0
10
20
30
Fig. 2. Evolving graph of f1 (x0 ) in WCE during various phases of the annealing process. We note that mapping is actually random during algorithm run.
Here, for demonstration, we fully associate every x0 with g1,m1 for which p1 (m1 |x0 ) is largest.
f1 (x0 )
30
0.04
20
0.02
10
0
0
0
10
f1 (x0 )-line
20
-0.02
30
0
2
(a)
4
6
(b)
Fig. 3. Numerical result for WCE in the case of k = 0.63, σX0 = 5. (a)
6-step solution. (b) The deviation of the first step in f1 (x0 ) from a straight
line between the end points of the step.
TABLE II
C OST C OMPARISON TABLE FOR S IDE C HANNEL P ROBLEM
bSN R
0
2.6
4.7
5.9
9.0
Linear Cost
0.960
0.696
0.432
0.344
0.203
J M ( [7])
0.185
0.149
0.101
0.081
0.052
J∗
0.167
0.079
0.040
0.026
0.012
V. A DVANTAGES OF P ROPOSED M ETHOD
(J M −J ∗)/J M
0.10
0.47
0.60
0.68
0.77
2) Results for Side Channel Problem: The original problem
is to minimize (20) subject to the constraint in (21). We follow
the standard approach in optimization theory and convert this
constrained problem to unconstrained Lagrangian formulation:
J = E{k 2 U12 + X22 + λU22 }
stage. Intuitively, as the second controller has access to better
side information (i.e. at higher SNR), the estimation error is
decreased and as observed in Figure 4, f1 (x0 ) approaches x0 .
The relative improvement in cost, given in Table II, increases
with SNR, which is consistent with the above observation.
The mappings for the side channel, g2 , are irregular and
the overall shape varies with SNR. This observation, together
with the above for f1 , suggests that the mappings f1 and g2 are
not scale invariant. The discontinuities in f1 and g2 coincide
as expected, as the discontinuities in side information signal
those in f1 to g3 .
Note: For numerical results in this paper, MATLAB codes
can be found in [32].
(22)
where λ is chosen to satisfy the power constraint (21) with
equality. In the experiments, we used the standard benchmark
parameters that were used for the original WCE, that is,
k = 0.2 and σX0 = 5. We have varied λ to obtain results
at different bSN R .
In Table II we compare the cost of our solutions (denoted
by J ∗ ) to the ones given in [7] (denoted by J M ), and the best
linear mappings. Significant cost reductions can be observed.
The relative improvement over the solution of [7] is listed in
the last column.
Remark 3: When bSN R = 0, the problem degenerates to
WCE, thus the cost is 0.1669, the best known to date.
We present several mappings obtained by our method in
Figure 4. Some interesting features of these mappings are
observed. The mappings f1 are staircase functions with constant steps similar to the ones obtained for the original WCE
problem, however, the steps get smaller and increase in number
as the side channel SNR increases; that is, f1 (x0 ) approaches
x0 . Note that the control cost term in (20), E{k 2 U12 }, achieves
its minimum when g1 = 0, i.e., f1 (x0 ) = x0 . This is,
however, not optimal due to the estimation error at the second
There are several improvements of the method proposed
here over existing methods in literature.
1) It is derived in the original, continuous domain, without
discretization. The continuous space is sampled during
numerical computation of integrals only. This is in
contrast with many prior methods such as those in [12],
[16], [17] that are entirely formulated in a discrete
setting.
2) Our method is based on DA, a powerful non-convex
optimization framework. DA has been successfully used
as a remedy to the problem of poor local minima in
non-convex optimization problems [28], and is shown to
outperform competing methods such as “noisy channel
relaxation” (NCR) [27]. Although NCR performs well
for the simple setting of WCE [16], it is susceptible to
get trapped in local minima in more involved settings.
3) From its DA foundation, our method directly inherits
notably useful properties including reduced sensitivity
to initialization. The authors in [15] had to experiment
with a large number of initial weight vectors to obtain
the result included in Table I.
4) We do not make any assumptions about the controller
mappings. Methods presented in [11], [12], [19] benefit
from the monotonicity of optimal mapping in WCE. The
results in Figure 4 demonstrate that monotonicity may
not hold for optimal mappings in the general setting.
5) The method is fully automated and does not require
analytical derivations or manual interventions during
algorithm run. This is in contrast to the method presented in [19] which requires analytical work during the
procedure.
6) Method is applicable to a broad class of stochastic
control problems. Many prior methods require nontrivial work in order to generalize, for instance, [17]
6
bsnr = 2.6
20
cost = 0.079
f1
0
-20
-15
cost = 0.040
0
-5
0
5
10
15
-20
-15
cost = 0.012
f1
g2
-10
bsnr = 9.0
20
f1
g2
-10
bsnr = 4.8
20
0
-5
x0
0
5
x0
10
15
-20
-15
g2
-10
-5
0
5
10
15
x0
Fig. 4. Some of the mappings we obtained for the side channel variation problem. The first controller is plotted at various SNR levels.
proposes a method that is generalizable, but it requires
conversion to a potential game problem.
VI. C ONCLUSIONS
In this paper, we proposed an optimization method for
distributed control problems, whose solutions are known to be
non-linear, and demonstrated its effectiveness on two problems
from the literature. The first problem is the celebrated benchmark problem known as Witsenhausen’s counter-example, for
which our approach obtained the best known cost value. As
a second test case we focused on the side channel setting
introduced in [7], where it is motivated as a two stage
noise cancellation problem. The mappings obtained are highly
nontrivial, offer considerably improved performance, and raise
interesting questions about the functional properties of optimal
mappings in decentralized control, which are the focus of
ongoing research.
R EFERENCES
[1] H. Witsenhausen, “A counterexample in stochastic optimum control,”
SIAM Journal on Control, vol. 6, no. 1, pp. 131–147, 1968.
[2] S. Yüksel and T. Başar, Stochastic Networked Control Systems: Stabilization and Optimization under Information Constraints. Springer,
2013.
[3] T. Başar, “Variations on the theme of the Witsenhausen counterexample,”
in 47th IEEE Conference on Decision and Control Proceedings (CDC).
IEEE, 2008, pp. 1614–1619.
[4] G. Gnecco, M. Sanguineti, and M. Gaggero, “Suboptimal solutions to
team optimization problems with stochastic information structure,” SIAM
Journal on Optimization, vol. 22, no. 1, pp. 212–243, 2012.
[5] P. Grover, S. Y. Park, and A. Sahai, “Approximately optimal solutions to
the finite-dimensional Witsenhausen counterexample,” Automatic Control, IEEE Transactions on, vol. 58, no. 9, pp. 2189–2204, 2013.
[6] Y. Ho and K. Chu, “Information structure in dynamic multi-person
control problems,” Automatica, vol. 10, no. 4, pp. 341 – 351, 1974.
[7] N. Martins, “Witsenhausen’s counter example holds in the presence of
side information,” in Decision and Control, 2006 45th IEEE Conference
on, 2006, pp. 1111–1116.
[8] N. Saldi, S. Yüksel, and T. Linder, “Finite Model Approximations and
Asymptotic Optimality of Quantized Policies in Decentralized Stochastic
Control,” ArXiv e-prints, 2015.
[9] Y. Ho and K. Chu, “Team decision theory and information structures in
optimal control problems–Part I,” Automatic Control, IEEE Transactions
on, vol. 17, no. 1, pp. 15–22, 1972.
[10] R. Radner, “Team decision problems,” The Annals of Mathematical
Statistics, vol. 33, no. 3, pp. 857–881, 09 1962.
[11] M. Deng and Y. Ho, “An ordinal optimization approach to optimal
control problems,” Automatica, vol. 35, no. 2, pp. 331 – 338, 1999.
[12] Y.-C. Ho and J. Lee, “Granular optimization: An approach to function
optimization,” in Decision and Control, 2000. Proceedings of the 39th
IEEE Conference on, vol. 1, 2000, pp. 103–111 vol.1.
[13] L. Jiang, D. Shah, J. Shin, and J. Walrand, “Distributed random access
algorithm: Scheduling and congestion control,” Information Theory,
IEEE Transactions on, vol. 56, no. 12, pp. 6182–6207, 2010.
[14] A. Kulkarni and T. Coleman, “An optimizer’s approach to stochastic
control problems with nonclassical information structures,” Automatic
Control, IEEE Transactions on, vol. PP, no. 99, pp. 1–1, 2014.
[15] M. Baglietto, T. Parisini, and R. Zoppoli, “Numerical solutions to the
Witsenhausen counterexample by approximating networks,” Automatic
Control, IEEE Transactions on, vol. 46, no. 9, pp. 1471–1477, 2001.
[16] J. Karlsson, A. Gattami, T. Oechtering, and M. Skoglund, “Iterative
source-channel coding approach to Witsenhausen’s counterexample,” in
American Control Conference (ACC), 2011. IEEE, 2011, pp. 5348–
5353.
[17] N. Li, J. Marden, and J. Shamma, “Learning approaches to the Witsenhausen counterexample from a view of potential games,” in Decision
and Control, 2009 held jointly with the 2009 28th Chinese Control
Conference. CDC/CCC 2009. Proceedings of the 48th IEEE Conference
on. IEEE, 2009, pp. 157–162.
[18] W. McEneaney and S. Han, “Optimization formulation and monotonic
solution method for the Witsenhausen problem,” Automatica, vol. 55,
pp. 55 – 65, 2015.
[19] J. Lee, E. Lau, and Y.-C. Ho, “The Witsenhausen counterexample:
a hierarchical search approach for nonconvex optimization problems,”
Automatic Control, IEEE Transactions on, vol. 46, no. 3, pp. 382–397,
2001.
[20] M. Mehmetoglu, E. Akyol, and K. Rose, “A deterministic annealing
approach to Witsenhausen’s counterexample,” in Information Theory
(ISIT), 2014 IEEE International Symposium on, June 2014, pp. 3032–
3036.
[21] P. Sharma, S. Salapaka, and C. Beck, “Entropy-based framework for
dynamic coverage and clustering problems,” Automatic Control, IEEE
Transactions on, vol. 57, no. 1, pp. 135–150, 2012.
[22] Y. Xu, S. Salapaka, and C. Beck, “Clustering and coverage control for
systems with acceleration-driven dynamics,” Automatic Control, IEEE
Transactions on, vol. 59, no. 5, pp. 1342–1347, 2014.
[23] A. Kwok and S. Martinez, “A distributed deterministic annealing algorithm for limited-range sensor coverage,” Control Systems Technology,
IEEE Transactions on, vol. 19, no. 4, pp. 792–804, 2011.
[24] M. Morozkov, O. Granichin, Z. Volkovich, and X. Zhang, “Fast algorithm for finding true number of clusters. Applications to control
systems,” in Control and Decision Conference (CCDC), 2012 24th
Chinese, 2012, pp. 2001–2006.
[25] K. Rose, E. Gurewitz, and G. Fox, “Vector quantization by deterministic
annealing,” Information Theory, IEEE Transactions on, vol. 38, no. 4,
pp. 1249–1257, 1992.
[26] A. Rao, D. Miller, K. Rose, and A. Gersho, “A deterministic annealing
approach for parsimonious design of piecewise regression models,”
Pattern Analysis and Machine Intelligence, IEEE Transactions on,
vol. 21, no. 2, pp. 159–173, 1999.
[27] M. Mehmetoglu, E. Akyol, and K. Rose, “Deterministic annealingbased optimization for zero-delay source-channel coding in networks,”
Communications, IEEE Transactions on, vol. 63, no. 12, pp. 5089–5100,
2015.
[28] K. Rose, “Deterministic annealing for clustering, compression, classification, regression, and related optimization problems,” Proceedings of
the IEEE, vol. 86, no. 11, pp. 2210–2239, 1998.
[29] A. El Gamal and Y. Kim, Network Information Theory. Cambridge
University Press, 2011.
[30] Y. Wu and S. Verdu, “Witsenhausen’s counterexample: A view from
optimal transport theory,” in Decision and Control and European Control
Conference (CDC-ECC), 2011 50th IEEE Conference on, 2011, pp.
5732–5737.
[31] H. Witsenhausen, “A standard form for sequential stochastic control,”
Mathematical systems theory, vol. 7, no. 1, pp. 5–11, 1973.
[32] http://lab.scl.ece.ucsb.edu/html/witsen.html.
| 3cs.SY
|
Non-Termination Inference of Logic Programs
arXiv:cs/0406041v1 [cs.PL] 22 Jun 2004
Etienne Payet and Fred Mesnard
IREMIA, Université de La Réunion, France
We present a static analysis technique for non-termination inference of logic programs. Our
framework relies on an extension of the subsumption test, where some specific argument
positions can be instantiated while others are generalized. We give syntactic criteria to
statically identify such argument positions from the text of a program. Atomic left looping
queries are generated bottom-up from selected subsets of the binary unfoldings of the program of interest. We propose a set of correct algorithms for automating the approach. Then,
non-termination inference is tailored to attempt proofs of optimality of left termination conditions computed by a termination inference tool. An experimental evaluation is reported.
When termination and non-termination analysis produce complementary results for a logic
procedure, then with respect to the leftmost selection rule and the language used to describe
sets of atomic queries, each analysis is optimal and together, they induce a characterization
of the operational behavior of the logic procedure.
Keywords: languages, verification, logic programming, static analysis, non-termination analysis, optimal termination condition
1
Introduction
Since the work of N. Lindenstrauss on TermiLog [20, 12], several automatic tools for termination
checking (e.g. TALP [3]) or termination inference (e.g. cTI [25, 26] or TerminWeb [17]) are now
available to the logic programmer. As the halting problem is undecidable for logic programs, such
analyzers compute sufficient termination conditions implying left termination. In most works,
only universal left termination is considered and termination conditions rely on a language for
describing classes of atomic queries. The search tree associated to any (concrete) query satisfying
a termination condition is guaranteed to be finite. When terms are abstracted using the term-size
norm, the termination conditions are (disjunctions of) conjunctions of conditions of the form “the
i-th argument is ground”. Let us call this language Lterm .
In this report, which is based on an earlier conference paper [27], we present the first approach
to non-termination inference tailored to attempt proofs of optimality of termination conditions at
verification time for pure logic programs. The aim is to ensure the existence, for each class of
atomic queries not covered by a termination condition, of one query from this class which leads to
an infinite search tree when such a query is proved using any standard Prolog engine. We shall first
present an analysis which computes classes of left looping queries, where any atomic query from
such a class is guaranteed to lead to at least one infinite derivation under the usual left-to-right
selection rule. Intuitively, we begin by computing looping queries from recursive binary clauses
of the form p(. . .) ← p(. . .). Then we try to add binary clauses of the form q(. . .) ← p(. . .) to
increase the set of looping queries. Finally by combining the result of non-termination inference
with termination inference, for each predicate, we compute the set of modes for which the overall
verification system has no information.
The main contributions of this work are:
• A new application of binary unfoldings to left loop inference. [16] introduced the binary
1
unfoldings of a logic program P as a goal independent technique to transform P into a
possibly infinite set of binary clauses, which preserves the termination property [7] while
abstracting the standard operational semantics. We present a correct algorithm to construct
left looping classes of atomic goals, where such classes are computed bottom-up from selected
subsets of the binary unfoldings of the analyzed program.
• A correct algorithm which, when combined with termination inference [23], may detect
optimal left termination conditions expressed in Lterm for logic programs. When termination
and non-termination analysis produce complementary results for a logic procedure, then with
respect to the leftmost selection rule and the language used to describe sets of atomic queries,
each analysis is optimal and together, they induce a characterization of the operational
behavior of the logic procedure.
• A report on the experimental evaluation we conduct. We have fully implemented termination
and non-termination inference for logic programs. We have run the couple of analyzers on
a set of classical logic programs, the sizes of which range from 2 to 177 clauses. The results
of this experiment should help the reader to appreciate the value of the approach.
We organize the paper as follows: Section 2 presents the notations. In Section 3 we study
loop inference for binary programs. We offer a full set of correct algorithms for non-termination
inference in Section 4 and optimality proofs of termination conditions in Section 5. Finally, in
Section 6, we discuss related works. The detailed proofs of the results can be found in Appendix B,
at the end of the article.
2
2.1
Preliminaries
Functions
Let E and F be two sets. Then, f : E → F denotes that f is a partial function from E to F and
f : E F denotes that f is a function from E to F . The domain of a partial function f from E
to F is denoted by Dom(f ) and is defined as: Dom(f ) = {e | e ∈ E, f (e) exists}. Thus, if f is a
function from E to F , then Dom(f ) = E. Finally, if f : E → F is a partial function and E ′ is a
set, then f |E ′ is the function from Dom(f ) ∩ E ′ to F such that for each e ∈ Dom(f ) ∩ E ′ , f |E ′
maps e to f (e).
2.2
Logic Programming
We strictly adhere to the notations, definitions, and results presented in [1].
N denotes the set of non-negative integers and for any n ∈ N , [1, n] denotes the set {1, . . . , n}.
If n = 0 then [1, n] = ∅.
From now on, we fix a language L of programs. We assume that L contains an infinite number
of constant symbols. The set of relation symbols of L is Π, and we assume that each relation
symbol p has a unique arity, denoted arity(p). T UL (resp. T BL ) denotes the set of all (ground
and non ground) terms of L (resp. atoms of L). A query is a finite sequence of atoms A1 , . . . , An
(where n ≥ 0). When n = 1, we say that the query is atomic. Throughout this article, the
variables of L are denoted by X, Y, Z, . . . , the constant symbols by a, b, . . . , the function symbols
by f, g, h, . . . , the relation symbols by p, q, r, . . . , the atoms by A, B, . . . and the queries by Q, Q′ ,
. . . or by A, B, . . .
Let t be a term. Then V ar(t) denotes the set of variables occurring in t. This notation is
extended to atoms, queries and clauses. Let θ := {X1 /t1 , . . . , Xn /tn } be a substitution. We
denote by Dom(θ) the set of variables {X1 , . . . , Xn } and by Ran(θ) the set of variables appearing
in t1 , . . . , tn . We define V ar(θ) = Dom(θ) ∪ Ran(θ). Given a set of variables V , θ|V denotes the
substitution obtained from θ by restricting its domain to V .
2
Let t be a term and θ be a substitution. Then, the term tθ is called an instance of t. If θ is a
renaming (i.e. a substitution that is a 1-1 and onto mapping from its domain to itself), then tθ is
called a variant of t. Finally, t is called more general than t′ if t′ is an instance of t.
A logic program is a finite set of definite clauses. In program examples, we use the ISO-Prolog
syntax. Let P be a logic program. Then ΠP denotes the set of relation symbols appearing in P .
In this paper, we only focus on left derivations i.e. we only consider the leftmost selection rule.
Consider a non-empty query B, C and a clause c. Let H ← B be a variant of c variable disjoint
θ
with B, C and assume that B and H unify. Let θ be an mgu of B and H. Then B, C =⇒(B, C)θ
c
is a left derivation step with H ← B as its input clause. If the substitution θ or the clause c is
irrelevant, we drop a reference to it.
θ2
θ1
· · · of left derivation steps is called a
Q1 =⇒
Let Q0 be a query. A maximal sequence Q0 =⇒
c1
c2
left derivation of P ∪ {Q0 } if c1 , c2 , . . . are clauses of P and if the standardization apart condition
holds, i.e. each input clause used is variable disjoint from the initial query Q0 and from the mgu’s
and input clauses used at earlier steps. A finite left derivation may end up either with the empty
query (then it is a successful left derivation) or with a non-empty query (then it is a failed left
derivation). We say Q0 left loops with respect to (w.r.t.) P if there exists an infinite left derivation
+
of P ∪ {Q0 }. We write Q =⇒ Q′ if there exists a finite non-empty prefix ending at Q′ of a left
P
derivation of P ∪ {Q}.
2.3
The Binary Unfoldings of a Logic Program
Let us present the main ideas about the binary unfoldings [16] of a logic program, borrowed from
[7]. This technique transforms a logic program P into a possibly infinite set of binary clauses.
Intuitively, each generated binary clause H ← B (where B is either an atom or the atom true
which denotes the empty query) specifies that, with respect to the original program P , a call to
H (or any of its instances) necessarily leads to a call to B (or its corresponding instance).
More precisely, let Q be an atomic query. Then A is a call in a left derivation of P ∪ {Q} if
+
Q =⇒ A, B. We denote by calls P (Q) the set of calls which occur in the left derivations of P ∪ {Q}.
P
The specialization of the goal independent semantics for call patterns for the left-to-right selection
rule is given as the fixpoint of an operator TPβ over the domain of binary clauses, viewed modulo
renaming. In the definition below, id denotes the set of all binary clauses of the form true ← true
or p(X1 , . . . , Xn ) ← p(X1 , . . . , Xn ) for any p ∈ ΠP , where arity(p) = n.
c := H ← B1 , . . . , Bm ∈ P, i ∈ [1, m],
hHj ← trueii−1
∈
X
renamed
with
fresh
variables,
j=1
β
TP (X) = (H ← B)θ Hi ← B ∈ X ∪ id renamed with fresh variables,
i < m ⇒ B 6= true
θ = mgu(hB1 , . . . , Bi i, hH1 , . . . , Hi i)
We define its powers as usual. It can be shown that the least fixpoint of this monotonic operator
always exists and we set bin unf (P ) := lfp(TPβ ). Then the calls that occur in the left derivations of
P ∪{Q} can be characterized as follows: calls P (Q) = {Bθ|H ← B ∈ bin unf (P ), θ = mgu(Q, H)}.
This last property was one of the main initial motivations of the proposed abstract semantics, enabling logic programs optimizations. Similarly, bin unf (P ) gives a goal independent representation
of the success patterns of P .
But we can extract more information from the binary unfoldings of a program P : universal
left termination of an atomic query Q with respect to P is identical to universal termination of
Q with respect to bin unf (P ). Note that the selection rule is irrelevant for a binary program and
an atomic query, as each subsequent query has at most one atom. The following result lies at the
heart of Codish’s approach to termination:
Theorem 2.1 [7] Let P be a program and Q an atomic query. Then Q left loops with respect to
P iff Q loops with respect to bin unf (P ).
3
Notice that bin unf (P ) is a possibly infinite set of binary clauses. For this reason, in the algorithms
of Section 4, we compute only the first max iterations of TPβ where max is a parameter of the
analysis. As an immediate consequence of Theorem 2.1, assume that we detect that Q loops with
respect to a subset of the binary clauses of TPβ ↑ i, with i ∈ N . Then Q loops with respect to
bin unf (P ) hence Q left loops with respect to P .
Example 2.2 Consider the following program P (see [21], p. 56–58):
p(X,Z) :- p(Y,Z),q(X,Y).
p(X,X).
q(a,b).
The binary unfoldings of P are:
TPβ
TPβ
TPβ
TPβ
TPβ
TPβ
↑0
↑1
↑2
↑3
↑4
↑5
=
=
=
=
=
=
∅
{p(X, Z) ← p(Y, Z), p(X, X) ← true, q(a, b) ← true} ∪ TPβ ↑ 0
{p(a, b) ← true, p(X, Y ) ← q(X, Y )} ∪ TPβ ↑ 1
{p(X, b) ← q(X, a), p(X, Z) ← q(Y, Z)} ∪ TPβ ↑ 2
{p(X, b) ← q(Y, a)} ∪ TPβ ↑ 3
TPβ ↑ 4 = bin unf (P )
Let Q := p(X, b). Note that Q loops w.r.t. TPβ ↑ 1, hence it loops w.r.t. bin unf (P ). So Q left
loops w.r.t. P .
3
Loop Inference Using Filters
In this paper, we propose a mechanism that, given a logic program P , generates at verification
time classes of atomic queries that left loop w.r.t. P . Our approach is completely based on
the binary unfoldings of P and relies on Theorem 2.1. It consists in computing a finite subset
BinProg of bin unf (P ) and then in inferring a set of atomic queries that loop w.r.t. BinProg. By
Theorem 2.1, these queries left loop w.r.t. P .
Hence, we reduce the problem of inferring looping atomic queries w.r.t. a logic program to
that of inferring looping atomic queries w.r.t. a binary program. This is why in the sequel, our
definitions, results and discussions mainly concentrate on binary programs only.
The central point of our method is the subsumption test, as the following lifting lemma,
specialized for the leftmost selection rule, holds:
Lemma 3.1 (One Step Lifting, [1]) Let Q =⇒ Q1 be a left derivation step, Q′ be a query that is
c
more general than Q and c′ be a variant of c variable disjoint with Q′ . Then, there exists a query
Q′1 that is more general than Q1 and such that Q′ =⇒ Q′1 with input clause c′ .
c
From this result, we derive:
Corollary 3.2 Let c := H ← B be a binary clause. If B is more general than H then H loops
w.r.t. {c}.
Corollary 3.3 Let c := H ← B be a clause from a binary program BinProg . If B loops w.r.t.
BinProg then H loops w.r.t. BinProg .
These corollaries provide two sufficient conditions that can be used to design an incremental
bottom-up mechanism that infers looping atomic queries. Given a binary program BinProg , it
suffices to build the set Q of atomic queries consisting of the heads of the clauses whose body is
more general than the head. By Corollary 3.2, the elements of Q loop w.r.t. BinProg. Then, by
Corollary 3.3, the head of the clauses whose body is more general than an element of Q can safely
been added to Q while retaining the property that every query in Q loops w.r.t. BinProg.
Notice that using this technique, we may not detect some looping queries. In [15], the authors
show that there is no algorithm that, when given a right-linear binary recursive clause (i.e. a
4
binary clause p(· · · ) ← p(· · · ) such that all variables occur at most once in the body) and given
an atomic query, always decides in a finite number of steps whether or not the resolution stops. In
the case of a linear atomic query (i.e. an atomic query such that all variables occur at most once)
however, the halting problem of derivations w.r.t. one binary clause is decidable [33, 13, 14].
It can be argued that the condition provided by Corollary 3.2 is rather weak because it fails
at inferring looping queries in some simple cases. This is illustrated by the following example.
Example 3.4 Let c be the clause p(X) ← p(f (X)). We have the infinite derivation:
p(X) =⇒ p(f (X)) =⇒ p(f (f (X))) =⇒ p(f (f (f (X)))) · · ·
c
c
c
But, since the body of c is not more general than its head, Corollary 3.2 does not allow to infer
that p(X) loops w.r.t. {c}.
In this section, we distinguish a special kind of argument positions that are “neutral” for
derivation. Our goal is to extend the relation “is more general than” by, roughly, disregarding the
predicate arguments whose position has been identified as neutral. Doing so, we aim at inferring
more looping queries.
Intuitively, a set of predicate argument positions ∆ is “Derivation Neutral” (DN for short) for
a binary clause c when the following holds. Let Q be an atomic query and Q′ be a query obtained
by replacing by any terms the predicate arguments in Q whose position is in ∆. If Q =⇒ Q1 then
c
Q′ =⇒ Q′1 where Q′1 is more general than Q1 up to the arguments whose position is in ∆.
c
Example 3.5 (Example 3.4 continued) The predicate p has only one argument position, so let
us consider ∆ := hp 7→ {1}i which distinguishes position 1 for predicate p. For any derivation
step p(s) =⇒ p(s1 ) if we replace s by any term t then there exists a derivation step p(t) =⇒ p(t1 ).
c
c
Notice that p(t1 ) is more general than p(s1 ) up to the argument of p. So, by the intuition described
above, ∆ is DN for c. Consequently, as in c the body p(f (X)) is more general than the head p(X)
up to the argument of p which is neutral, by an extended version of Corollary 3.2 there exists an
infinite derivation of {c} ∪ {p(X)}.
Let us give some more concrete examples of DN positions.
Example 3.6 The second argument position of the relation symbol append in the program APPEND:
append([],Ys,Ys).
append([X|Xs],Ys,[X|Zs]) :- append(Xs,Ys,Zs).
% C1
% C2
is DN for C2. Notice that a very common programming technique called accumulator passing
(see for instance e.g. [28], p. 21–25) always produces DN positions. A classical example of the
accumulator passing technique is the following program REVERSE.
reverse(L,R) :- rev(L,[],R).
rev([],R,R).
rev([X|Xs],R0,R) :- rev(Xs,[X|R0],R).
% C1
% C2
% C3
Concerning termination, we may ignore the second and the third argument of rev in the recursive
clause C3 while unfolding a query with this clause. Only the first argument can stop the unfolding.
But we can be even more precise. Instead of only identifying positions that can be totaly
disregarded as in the above examples, we can try to identify positions where we can place any
terms for which a given condition holds.
5
Example 3.7 Consider the clause c := p(f (X)) ← p(f (f (X))). If we mean by a DN position
a position where we can place any terms, then the argument position of p is not DN for c. This
is because, for example, we have the derivation step p(X) =⇒ p(f (f (X1 ))) but if we replace X by
c
g(X) then there is no derivation step of {c} ∪ {p(g(X))}. However, if we mean by a DN position
a position where we can place any instances of f (X), then the argument position of p is DN for
c.
In the sequel of the section, we define more precisely DN positions as positions where we can
place any terms satisfying certain conditions identified by “filters”. We use filters to present an
extension of the relation “is more general than” and we propose an extended version of Corollary 3.2. We offer two syntactic conditions of increasing power for easily identifying DN positions
from mere inspection of the text of a logic program. The practical impact of such filters will be
tackled in Section 5.
3.1
Filters
Let us first introduce the notion of a filter. We use filters in order to distinguish atoms, some
arguments of which satisfy a given condition. A condition upon atom arguments, i.e. terms, can
be defined as a function in the following way.
Definition 3.8 (Term-condition) A term-condition is a function from the set of terms T UL to
{true, false}.
Example 3.9 The following functions are term-conditions.
ftrue : T UL
t
{true, false}
7→ true
f1 : T UL
{true, false}
t
f2 : T UL
t
7→
true iff t is an instance of [X|Y ]
{true, false}
7→
true iff t unifies with h(a, X)
Notice that a term-condition might give distinct results for two terms which are equal modulo
renaming. For instance f2 (X) = false and f2 (Y ) = true. However, in Definition 3.12 below, we
will only consider variant independent term-conditions.
Definition 3.10 (Variant Independent Term-Condition) A term-condition f is variant independent if, for every term t, f (t) = true implies that f (t′ ) = true for every variant t′ of t.
Example 3.11 (Example 3.9 continued) ftrue and f1 are variant independent while f2 is not.
We restrict the class of term-conditions to that of variant independent ones because we want
to extend the relation “is more general than” so that if an atom A is linked to an atom B by
the extended relation, then every variant of A is also linked to B (see Proposition 3.16 below).
This will be essential to establish the forthcoming main Proposition 3.20 which is an extension of
Corollary 3.2. Now we can define what we exactly mean by a filter.
Definition 3.12 (Filter) A filter, denoted by ∆, is a function from Π such that: for each p ∈ Π,
∆(p) is a partial function from [1, arity(p)] to the set of variant independent term-conditions.
Example 3.13 (Example 3.9 continued) Let p be a relation symbol whose arity equals 3. The
filter ∆ which maps p to the function h1 7→ ftrue , 2 →
7 f1 i and any q ∈ Π \ {p} to hi is noted
∆ := h p 7→ h1 7→ ftrue , 2 7→ f1 i i.
6
3.2
Extension of the Relation “Is More General Than”
Given a filter ∆, the relation “is more general than” can be extended in the following way: an atom
A := p(· · · ) is ∆-more general than B := p(· · · ) if the “is more general than” requirement holds
for those arguments of A whose position is not in the domain of ∆(p) while the other arguments
satisfy their associated term-condition.
Definition 3.14 (∆-more general) Let ∆ be a filter and A and B be two atoms.
• Let η be a substitution. Then A is ∆-more general than B for η if:
A = p(s1 , . . . , sn )
B = p(t1 , . . . , tn )
∀i ∈ [1, n] \ Dom(∆(p)), ti = si η
∀i ∈ Dom(∆(p)), ∆(p)(i)(si ) = true .
• A is ∆-more general than B if there exists a substitution η s.t. A is ∆-more general than B
for η.
An atomic query Q is ∆-more general than an atomic query Q′ if either Q and Q′ are both empty
or Q contains the atom A, Q′ contains the atom B and A is ∆-more general than B.
Example 3.15 (Example 3.13 continued) Let
A := p( b , X , h(a, X) )
B := p( a , [a|b] , X
)
C := p( a , [a|b] , h(Y, b) ) .
Then, A is not ∆-more general than B and C because, for instance, its second argument X is
not an instance of [X|Y ] as required by f1 . On the other hand, B is ∆-more general than A for
the substitution {X/h(a, X)} and B is ∆-more general than C for the substitution {X/h(Y, b)}.
Finally, C is not ∆-more general than A because h(Y, b) is not more general than h(a, X) and C
is not ∆-more general than B because h(Y, b) is not more general than X.
As in a filter the term-conditions are variant independent, we get the following proposition.
Proposition 3.16 Let ∆ be a filter and A and B be two atoms. If A is ∆-more general than B
then every variant of A is ∆-more general than B.
The next proposition states an intuitive result:
Proposition 3.17 Let ∆ be a filter and A and B be two atoms. Then A is ∆-more general than
B if and only if there exists a substitution η such that V ar(η) ⊆ V ar(A, B) and A is ∆-more
general than B for η.
3.3
Derivation Neutral Filters: Operational Definition
In the sequel of this paper, we focus on “derivation neutral” filters. The name “derivation neutral”
stems from the fact that in any derivation of an atomic query Q, the arguments of Q whose position
is distinguished by such a filter can be safely replaced by any terms satisfying the associated termcondition. Such a replacement does not modify the derivation process.
Definition 3.18 (Derivation Neutral) Let ∆ be a filter and c be a binary clause. We say that ∆
is DN for c if for each derivation step Q =⇒ Q1 where Q is an atomic query, for each Q′ that is
c
∆-more general than Q and for each variant c′ of c variable disjoint with Q′ , there exists a query
Q′1 that is ∆-more general than Q1 and such that Q′ =⇒ Q′1 with input clause c′ . This definition
c
is extended to binary programs: ∆ is DN for P if it is DN for each clause of P .
7
Example 3.19 The following examples illustrate the previous definition.
• Let us reconsider the program APPEND from Example 3.6 with the term-condition ftrue defined
in Example 3.9 and the filter ∆ := happend 7→ h2 7→ ftrue ii. ∆ is DN for C2. However, ∆
is not DN for APPEND because it is not DN for C1.
• Consider the following clause:
merge([X|Xs],[Y|Ys],[X|Zs]) :- merge(Xs,[Y|Ys],Zs).
The filter hmerge 7→ h2 7→ f1 ii, where the term-condition f1 is defined in Example 3.9, is
DN for this clause.
In the next subsection, we present some syntactic criteria for identifying correct DN filters. For
proving that the above filters are indeed DN, we will just check that they actually fulfill these
syntactic criteria that are sufficient conditions.
Derivation neutral filters lead to the following extended version of Corollary 3.2 (take ∆ such
that for any p, ∆(p) is a function whose domain is empty):
Proposition 3.20 Let c := H ← B be a binary clause and ∆ be a filter that is DN for c. If B is
∆-more general than H then H loops w.r.t. {c}.
We point out that the above results remain valid when the program under consideration is
restricted to its set of clauses used in the derivation steps. For instance, although the filter ∆ of
Example 3.19 is not DN for APPEND, it will help us to construct queries which loop w.r.t. C2. Such
queries also loop w.r.t. APPEND.
Notice that lifting lemmas are used in the literature to prove completeness of SLD-resolution.
As Definition 3.18 corresponds to an extended version of the One Step Lifting Lemma 3.1, it may
be worth to investigate its consequences from the model theoretic point of view.
First of all, a filter may be used to “expand” atoms by replacing every argument whose position
is distinguished by any term that satisfies the associated term-condition.
Definition 3.21 Let ∆ be a filter and A be an atom. The expansion of A w.r.t. ∆, denoted A↑∆ ,
is the set defined as
def
A↑∆ = {A} ∪ {B ∈ T BL | B is ∆-more general than A for ǫ}
where ǫ denotes the empty substitution.
Notice that in this definition, we do not necessary have the inclusion
{A} ⊆ {B ∈ T BL | B is ∆-more general than A for ǫ} .
For instance, suppose that A := p(f (X)) and that ∆ maps p to the function h1 7→ f i where f is
the term-condition mapping any term t to true iff t is an instance of g(X). Then
{B ∈ T BL | B is ∆-more general than A} = {p(t) | t is an instance of g(X)}
with A 6∈ {p(t) | t is an instance of g(X)}.
Term interpretations in the context of logic programming were first introduced in [6] and further
investigated in [11] and then in [22]. A term interpretation for L is identified with a (possibly
empty) subset of the term base T BL . So, as for atoms, a term interpretation can be expanded by
a filter.
Definition 3.22 Let ∆ be a filter and I be a term interpretation for L. Then I↑∆ is the term
interpretation for L defined as:
def [
I↑∆ =
A↑∆ .
A∈I
8
For any logic program P , we denote by C(P ) its least term model.
Theorem 3.23 Let P be a binary program and ∆ be a DN filter for P . Then C(P )↑∆ = C(P ).
Proof. The inclusion C(P ) ⊆ C(P )↑∆ is straightforward so let us concentrate on the other one
i.e. C(P )↑∆ ⊆ C(P ). Let A′ ∈ C(P )↑∆ . Then there exists A ∈ C(P ) such that A′ ∈ A↑∆ . A well
known result states:
C(P ) = {B ∈ T BL | there exists a successful derivation of P ∪ {B}}
(1)
Consequently, there exists a successful derivation ξ of P ∪{A}. Therefore, by successively applying
Definition 3.18 to each step of ξ, one construct a successful derivation of A′ . So by (1) A′ ∈ C(P ).
3.4
Some Particular DN Filters
In this section, we provide two sufficient syntactic conditions for identifying DN filters.
3.4.1
DN Sets of Positions
The first instance we consider corresponds to filters, the associated term-conditions of which are
all equal to ftrue (see Example 3.9). Within such a context, as the term-conditions are fixed, each
filter ∆ is uniquely determined by the domains of the partial functions ∆(p) for p ∈ Π. Hence the
following definition.
Definition 3.24 (Set of Positions) A set of positions, denoted by τ , is a function from Π to 2N
such that: for each p ∈ Π, τ (p) is a subset of [1, arity(p)].
Example 3.25 Let append and append3 be two relation symbols. Assume that arity(append ) = 3
and arity(append3 ) = 4. Then τ := h append 7→ {2}, append3 7→ {2, 3, 4} i is a set of positions.
Not surprisingly, the filter that is generated by a set of positions is defined as follows.
Definition 3.26 (Associated Filter) Let τ be a set of positions and ftrue be the term-condition
defined in Example 3.9. The filter ∆[τ ] defined as:
for each p ∈ Π, ∆[τ ](p) is the function from τ (p) to {ftrue }
is called the filter associated to τ .
Example 3.27 (Example 3.25 continued) The filter associated to τ is
∆[τ ] :=
happend 7→ h2 7→ ftrue i, append3 7→ h2 7→ ftrue , 3 7→ ftrue , 4 7→ ftrue ii.
Now we define a particular kind of sets of positions. These are named after “DN” because, as
stated by Theorem 3.30 below, they generate DN filters.
Definition 3.28 (DN Set of Positions) Let τ be a set of positions. We say that τ is DN for a
binary clause p(s1 , . . . , sn ) ← q(t1 , . . . , tm ) if:
si is a variable
si occurs only once in p(s1 , . . . , sn )
∀i ∈ τ (p),
∀j ∈ [1, m], si ∈ V ar(tj ) ⇒ j ∈ τ (q) .
A set of positions is DN for a binary program P if it is DN for each clause of P .
9
The intuition of Definition 3.28 is the following. If for instance we have a clause c :=
p(X, Y, f (Z)) ← p(g(Y, Z), X, Z) then in the first two positions of p we can put any terms and get
a derivation step w.r.t. c because the first two arguments of the head of c are variables that appear
exactly once in the head. Moreover, X and Y of the head reappear in the body but again only in
the first two positions of p. So, if we have a derivation step p(s1 , s2 , s3 ) =⇒ p(t1 , t2 , t3 ), we can rec
place s1 and s2 by any terms s′1 and s′2 and get another derivation step p(s′1 , s′2 , s3 ) =⇒ p(t′1 , t′2 , t′3 )
where t′3 is the same as t3 up to variable names.
c
Example 3.29 (Example 3.25 continued) τ is DN for the program:
append([X|Xs],Ys,[X|Zs]) :- append(Xs,Ys,Zs).
append3(Xs,Ys,Zs,Ts) :- append(Xs,Ys,Us).
which is a subset of the binary unfoldings of the program APPEND3:
append([],Ys,Ys).
append([X|Xs],Ys,[X|Zs]) :- append(Xs,Ys,Zs).
append3(Xs,Ys,Zs,Ts) :- append(Xs,Ys,Us), append(Us,Zs,Ts).
DN sets of positions generate DN filters.
Theorem 3.30 Let τ be a DN set of positions for a binary program P . Then ∆[τ ] is DN for P .
Proof. As we will see in Section 3.4.2, this theorem is a particular case of Theorem 3.39.
Notice that the set of DN sets of positions of any binary program P is not empty because, by
Definition 3.28, τ0 := hp 7→ ∅ | p ∈ Πi is DN for P . Moreover, an atom A is ∆[τ0 ]-more general
than an atom B iff A is more general than B.
3.4.2
DN Sets of Positions with Associated Terms
Now we consider another instance of Definition 3.18. As we will see, it is more general than the
previous one. It corresponds to filters whose associated term-conditions have all the form “is an
instance of t” where t is a term that uniquely determines the term-condition. Notice that such
term-conditions are variant independent, so it makes sense to consider such filters. Hence the
following definition.
Definition 3.31 (Sets of Positions with Associated Terms) A set of positions with associated
terms, denoted by τ + , is a function from Π such that: for each p ∈ Π, τ + (p) is a partial function
from [1, arity(p)] to T UL .
Example 3.32 Let p and q be two relation symbols whose arity is 2. Then
τ + := h p 7→ h2 7→ Xi, q 7→ h2 7→ g(X)i i
is a set of positions with associated terms.
The filter that is generated by a set of positions with associated terms is defined as follows.
Definition 3.33 (Associated Filter) Let τ + be a set of positions with associated terms. The filter
associated to τ + , denoted by ∆[τ + ], is defined as: for each p ∈ Π, ∆[τ + ](p) is the function
Dom(τ + (p))
i
The set of term-conditions
(
T UL {true, false}
7→
t
7→ true iff t is an instance of τ + (p)(i)
10
Example 3.34 (Example 3.32 continued) The filter associated to τ + is
∆[τ + ] :=
where
f1 : T UL
t
f2 : T UL
t
h p 7→ h2 7→ f1 i, q 7→ h2 7→ f2 i i
{true, false}
7→
true iff t is an instance of X
{true, false}
7→
true iff t is an instance of g(X)
As for sets of positions, we define a special kind of sets of positions with associated terms.
Definition 3.35 (DN Sets of Positions with Associated Terms) Let τ + be a set of positions with
associated terms. We say that τ + is DN for a binary clause p(s1 , . . . , sn ) ← q(t1 , . . . , tm ) if these
conditions hold:
• (DN1) ∀i ∈ Dom(τ + (p)), ∀j ∈ [1, n] \ {i}: V ar(si ) ∩ V ar(sj ) = ∅,
• (DN2) ∀hi 7→ ui i ∈ τ + (p): si is more general than ui ,
• (DN3) ∀hj 7→ uj i ∈ τ + (q): tj is an instance of uj ,
• (DN4) ∀i ∈ Dom(τ + (p)), ∀j 6∈ Dom(τ + (q)): V ar(si ) ∩ V ar(tj ) = ∅.
A set of positions with associated terms is DN for a binary program P if it is DN for each clause
of P .
This definition says that any si where i is in the domain of τ + (p) (i.e. position i is distinguished
by τ + ): (DN1) does not share its variables with the other arguments of the head, (DN2) is more
general than the term ui that i is mapped to by τ + (p), (DN4) distributes its variables to some tj
such that j is in the domain of τ + (q) (i.e. position j is distinguished by τ + ). Moreover, (DN3)
says that any tj , where j is distinguished by τ + , is such that tj is an instance of the term uj that
j is mapped to by τ + (q).
Example 3.36 (Example 3.32 continued) τ + is DN for the following program:
p(f(X),Y) :- q(X,g(X)).
q(a,g(X)) :- q(a,g(b)).
The preceding notion is closed under renaming:
Proposition 3.37 Let c be a binary clause and τ + be a set of positions with associated terms that
is DN for c. Then τ + is DN for every variant of c.
Notice that a set of positions is a particular set of positions with associated terms in the
following sense.
Proposition 3.38 Let τ be a set of positions and X be a variable. Let τ + be the set of positions
with associated terms defined as: for each p ∈ Π, τ + (p) := ( τ (p) {X} ). Then, the following
holds.
1. An atom A is ∆[τ ]-more general than an atom B iff A is ∆[τ + ]-more general than B.
2. For any binary clause c, τ is DN for c iff τ + is DN for c.
Proof. A proof follows from these remarks.
11
• Item 1 is a direct consequence of the definition of “∆-more general” (see Definition 3.14)
and the definition of the filter associated to a set of positions (see Definition 3.26) and to a
set of positions with associated terms (see Definition 3.33).
• Item 2 is a direct consequence of the definition of DN sets of positions (see Definition 3.28)
and DN sets of positions with associated terms (see Definition 3.35).
The sets of positions with associated terms of Definition 3.35 were named after “DN” because
of the following result.
Theorem 3.39 Let P be a binary program and τ + be a set of positions with associated terms that
is DN for P . Then ∆[τ + ] is DN for P .
As in the case of sets of positions, the set of DN sets of positions with associated terms of any
binary program P is not empty because, by Definition 3.35, τ0+ := hp 7→ hi | p ∈ Πi is DN for
P . Moreover, an atom A is ∆[τ0+ ]-more general than an atom B iff A is more general than B.
Finally, in Appendix A, we give an incremental algorithm (see Section 4.2) that computes a DN
set of positions with associated terms. Its correctness proof is also presented.
3.5
Examples
This section presents some examples where we use filters obtained from DN sets of positions and
DN sets of positions with associated terms to infer looping queries. As the filters we use in each
case are not “empty” (i.e. are not obtained from τ0 or τ0+ ), we are able to compute more looping
queries than using the classical subsumption test.
Example 3.40 Consider the program APPEND that we introduced in Example 3.6. Every infinite
derivation w.r.t. APPEND starting from an atomic query only uses the non-unit clause C2. Therefore, as we aim at inferring looping atomic queries w.r.t. APPEND, we only focus on C2 in the
sequel of this example.
As in C2 the body, which is append (Xs, Ys, Zs), is more general than the head, which is
append ([X |Xs], Ys, [X |Zs]), by Corollary 3.2 we have that the query append ([X |Xs], Ys, [X |Zs])
loops w.r.t. {C2}. Consequently, by the One Step Lifting Lemma 3.1, each query that is more
general than append ([X |Xs], Ys, [X |Zs]) also loops w.r.t. {C2}.
But we can be more precise than that. According to Definition 3.28, τ := h append 7→ {2} i
is a DN set of positions for {C2}. The filter associated to τ (see Definition 3.26) is ∆[τ ] :=
h append 7→ h2 7→ ftrue i i. By Theorem 3.30, ∆[τ ] is a DN filter for {C2}. Consequently, by
Definition 3.18, each query that is ∆[τ ]-more general than append ([X |Xs], Ys, [X |Zs]) loops w.r.t.
{C2}. This means that
t2 is any term and
append (t1 , t2 , t3 ) ∈ TB L
t1 , t3 is more general than [X |Xs], [X |Zs]
is a set of atomic queries that loop w.r.t. {C2}, hence w.r.t. APPEND. This set includes the ’welltyped’ query append (As, [ ], Bs).
Example 3.41 Consider the program REVERSE that was introduced in Example 3.6. As in the
example above, in order to infer looping atomic queries w.r.t. REVERSE, we only focus on the
non-unit clauses C1 and C3 in the sequel of this example. More precisely, we process the relation
symbols of the program in a bottom-up way, so we start the study with clause C3 and end it with
clause C1.
According to Definition 3.28, τ := h rev 7→ {2, 3} i is a DN set of positions for {C3}. The
filter associated to τ (see Definition 3.26) is ∆[τ ] := h rev 7→ h2 7→ ftrue , 3 7→ ftrue i i. By
Theorem 3.30, ∆[τ ] is DN for {C3}. As rev (Xs, [X |R0 ], R) (the body of C3) is ∆[τ ]-more general
than rev ([X |Xs], R0 , R) (the head of C3), by Proposition 3.20 we get that rev ([X |Xs], R0 , R) loops
w.r.t. {C3}. Notice that, unlike the example above, here we do not get this result from Corollary 3.2
12
as rev (Xs, [X |R0 ], R) is not more general than rev ([X |Xs], R0 , R). Finally, as ∆[τ ] is DN for
{C3}, by Definition 3.18 we get that each query that is ∆[τ ]-more general than rev ([X |Xs], R0 , R)
loops w.r.t. {C3}, hence w.r.t. REVERSE. This means that
t2 and t3 are any terms and
Q := rev (t1 , t2 , t3 ) ∈ TB L
t1 is more general than [X |Xs]
is a set of atomic queries that loop w.r.t. REVERSE. This set includes the ’well-typed’ query
rev (As, [ ], [ ]).
Now, consider clause C1. As rev (L, [], R) (its body) is an element of Q, then rev (L, [], R) loops
w.r.t. {C3}, hence w.r.t. {C1, C3}. Consequently, by Corollary 3.3, reverse(L, R) (the head of C1)
loops w.r.t. {C1, C3}. Let τ ′ := h rev 7→ {2, 3}, reverse 7→ {2} i. By Definition 3.28, τ ′ is DN for
{C1, C3}, so ∆[τ ′ ] is DN for {C1, C3}. Consequently, each query that is ∆[τ ′ ]-more general than
reverse(L, R) also loops w.r.t. {C1, C3} hence w.r.t. REVERSE. This means that
reverse(X, t) ∈ TB L | X is a variable and t is any term
is a set of atomic queries that loop w.r.t. REVERSE. This set includes the ’well-typed’ query
reverse(As, [ ]).
Example 3.42 Consider the two recursive clauses of the program MERGE where we have removed
the inequalities:
merge([X|Xs],[Y|Ys],[X|Zs]) :- merge(Xs,[Y|Ys],Zs). % C3
merge([X|Xs],[Y|Ys],[Y|Zs]) :- merge([X|Xs],Ys,Zs). % C4
Every set of positions τ that is DN for {C3} is such that τ (merge) = ∅ because each argument
of the head of C3 is not a variable (see Definition 3.28). Hence, using Proposition 3.20 with
a filter obtained from a DN set of positions leads to the same results as using Corollary 3.2:
as merge(Xs, [Y |Ys], Zs) is more general than merge([X |Xs], [Y |Ys], [X |Zs]), by Corollary 3.2
merge([X |Xs], [Y |Ys], [X |Zs]) loops w.r.t. {C3}. So, by the One Step Lifting Lemma 3.1, each
query that is more general than merge([X |Xs], [Y |Ys], [X |Zs]) also loops w.r.t. {C3}, hence w.r.t.
MERGE.
But we can be more precise than that. According to Definition 3.35, τ + := hmerge 7→ h2 7→
[Y |Ys]i i is a set of positions with associated terms that is DN for {C3}. Hence, by Theorem 3.39,
the associated filter ∆[τ + ] (see Definition 3.33) is DN for {C3}. So, by Definition 3.18, each query
that is ∆[τ + ]-more general than merge([X |Xs], [Y |Ys], [X |Zs]) loops w.r.t. {C3}. This means that
t is any instance of [Y |Ys] and
merge(t1 , t2 , t3 ) ∈ TB L 2
t1 , t3 is more general than [X |Xs], [X |Zs]
is a set of atomic queries that loop w.r.t. MERGE. Notice that this set includes the ’well-typed’ query
merge(As, [0], Bs). Finally, let us turn to clause C4. Reasoning exactly as above with the set of
positions with associated terms hmerge 7→ h1 7→ [X |Xs]i i which is DN for {C4}, we conclude that:
t1 is any instance of [X |Xs] and
merge(t1 , t2 , t3 ) ∈ TB L
t2 , t3 is more general than [Y |Ys], [Y |Zs]
is a set of atomic queries that loop w.r.t. MERGE. Notice that this set includes the ’well-typed’ query
merge([0], As, Bs).
4
Algorithms
We have designed a set of correct algorithms for full automation of non-termination analysis of
logic programs. These algorithms are given in Appendix A with their correctness proofs. In this
section, we present the intuitions and conceptual definitions underlying our approach.
13
4.1
Loop Dictionaries
Our technique is based on a data structure called dictionary which is a set of pairs (BinSeq, τ + )
where BinSeq is a finite ordered sequence of binary clauses and τ + is a set of positions with
associated terms. In the sequel, we use the list notation of Prolog and a special kind of dictionaries
that we define as follows.
Definition 4.1 (Looping Pair, Loop Dictionary) A pair (BinSeq, τ + ), where the list BinSeq is a
finite ordered sequence of binary clauses and τ + is a set of positions with associated terms, is a
looping pair if τ + is DN for BinSeq and:
• either BinSeq = [H ← B] and B is ∆[τ + ]-more general than H,
• or BinSeq = [H ← B, H1 ← B1 | BinSeq 1 ] and there exists a set of positions with associated
terms τ1+ such that ([H1 ← B1 | BinSeq 1 ], τ1+ ) is a looping pair and B is ∆[τ1+ ]-more general
than H1 .
A loop dictionary is a finite set of looping pairs.
Example 4.2 The pair (BinSeq := [H1 ← B1 , H2 ← B2 , H3 ← B3 ], τ1+ ) where
H1 ← B1 := r(X) ← q(X, f (f (X)))
H2 ← B2 := q(X, f (Y )) ← p(f (X), a)
H3 ← B3 := p(f (g(X)), a) ← p(X, a)
and τ1+ := hp 7→ h2 7→ ai, q 7→ h2 7→ f (X)ii is a looping pair:
• Let τ3+ := hp 7→ h2 7→ aii. Then τ3+ is a DN set of positions with associated terms for
[H3 ← B3 ]. Moreover, B3 is ∆(τ3+ )-more general than H3 . Consequently, ([H3 ← B3 ], τ3+ )
is a looping pair.
• Notice that B2 is ∆(τ3+ )-more general than H3 . Now, let τ2+ := τ1+ . Then τ2+ is DN for
[H2 ← B2 , H3 ← B3 ]. So, ([H2 ← B2 , H3 ← B3 ], τ2+ ) is a looping pair.
• Finally, notice that B1 is ∆(τ2+ )-more general than H2 . As τ1+ is DN for BinSeq, we conclude
that (BinSeq, τ1+ ) is a looping pair.
A looping pair immediately provides an atomic looping query. It suffices to take the head of
the first clause of the binary program of the pair:
Proposition 4.3 Let ([H ← B|BinSeq], τ + ) be a looping pair. Then H loops with respect to
[H ← B|BinSeq].
Proof. By induction on the length of BinSeq using Proposition 3.20, Corollary 3.3 and Theorem 3.39. So, a looping pair denotes a proof outline for establishing that H left loops. Moreover,
looping pairs can be built incrementally in a simple way as described below.
4.2
Computing a Loop Dictionary
Given a logic program P and a positive integer max , the function infer loop dict from Appendix A first computes TPβ ↑ max (the first max iterations of the operator TPβ ), which is a
finite subset of bin unf (P ). Then, using the clauses of TPβ ↑ max , it incrementally builds a loop
dictionary Dict as follows.
At start, Dict is set to ∅. Then, for each clause H ← B in TPβ ↑ max , the following actions
are performed.
• infer loop dict tries to extract from H ← B the most simple form of a looping pair: it
computes a set of positions with associated terms τ + that is DN for H ← B, then it tests if
B is ∆[τ + ]-more general than H. If so, the looping pair ([H ← B], τ + ) is added to Dict .
14
• infer loop dict tries to combine H ← B to some looping pairs that have already been
added to Dict in order to build other looping pairs. For each ([H1 ← B1 |BinSeq 1 ], τ1+ ) in
Dict, if B is ∆[τ1+ ]-more general than H1 , then a set of positions with associated terms
τ + that is DN for [H ← B, H1 ← B1 |BinSeq 1 ] is computed and the looping pair ([H ←
B, H1 ← B1 |BinSeq 1 ], τ + ) is added to Dict .
Notice that in the second step above, we compute τ + that is DN for [H ← B, H1 ← B1 |BinSeq 1 ].
As we already hold τ1+ that is DN for [H1 ← B1 |BinSeq 1 ], it is more interesting, for efficiency
reasons, to compute τ + from τ1+ instead of starting from the ground. Indeed, starting from τ1+ ,
one uses the information stored in τ1+ about the program [H1 ← B1 |BinSeq 1 ], which may speed
up the computation substantially. This is why we have designed a function dna that takes two
arguments as input, a binary program BinProg and a set of positions with associated terms τ + . It
computes a set of positions with associated terms that is DN for BinProg and that refines τ + . On
+
the other hand, the function unit loop calls dna with τmax
which is the initial set of positions with
+
+
associated terms defined as follows: Dom(τmax (p)) = [1, arity(p)] for each p ∈ Π and τmax
(p)(i) is
a variable for each i ∈ [1, arity(p)].
Example 4.4 Consider the program APPEND3
append3(Xs,Ys,Zs,Us) :- append(Xs,Ys,Vs), append(Vs,Zs,Us).
β
augmented with the APPEND program. The set TAPPEND3
↑ 2 includes:
append([A|B],C,[A|D]) :- append(B,C,D).
append3(A,B,C,D) :- append(A,B,E).
append3([],A,B,C) :- append(A,B,C).
% BC1
% BC2
% BC3
From clause BC1 we get the looping pair (BinSeq 1 , τ1+ ) where
BinSeq 1 = append ([X1 |X2 ], X3 , [X1 |X4 ]) ← append (X2 , X3 , X4 )
and τ1+ (append ) = h2 7→ X3 i. From this pair and the clause BC2, we get the looping pair
(BinSeq 2 , τ2+ ) where:
BinSeq 2 =
append3 (X1 , X2 , X3 , X4 ) ← append (X1 , X2 , X5 ),
append ([X1 |X2 ], X3 , [X1 |X4 ]) ← append (X2 , X3 , X4 )
and τ2+ (append ) = h2 7→ X3 i and τ2+ (append3 ) = h2 7→ X2 , 3 7→ X3 , 4 7→ X4 i.
Finally, from (BinSeq 1 , τ1+ ) and BC3, we get the looping pair (BinSeq 3 , τ3+ ) where:
BinSeq 3 =
append3 ([], X1 , X2 , X3 ) ← append (X1 , X2 , X3 ),
append ([X1 |X2 ], X3 , [X1 |X4 ]) ← append (X2 , X3 , X4 )
and τ3+ (append ) = h2 7→ X3 i and τ3+ (append3 ) = h3 7→ X2 i.
Example 4.5 Consider the program PERMUTE:
delete(X,[X|Xs],Xs).
delete(Y,[X|Xs],[X|Ys]) :- delete(Y,Xs,Ys).
permute([],[]).
permute([X|Xs],[Y|Ys]) :- delete(Y,[X|Xs],Zs), permute(Zs,Ys).
β
The set TPERMUTE
↑ 1 includes:
15
delete(B,[C|D],[C|E]) :- delete(B,D,E).
permute([B|C],[D|E]) :- delete(D,[B|C],F).
% BC1
% BC2
From clause BC1 we get the looping pair (BinSeq 1 , τ1+ ) where
BinSeq 1 = delete(X1 , [X2 |X3 ], [X2 |X4 ]) ← delete(X1 , X3 , X4 )
and τ1+ (delete) = h1 7→ X1 i. From this pair and BC2, we get the looping pair (BinSeq 2 , τ2+ ) where:
BinSeq 2 =
permute([X1 |X2 ], [X3 |X4 ]) ← delete(X3 , [X1 |X2 ], X5 ),
delete(X1 , [X2 |X3 ], [X2 |X4 ]) ← delete(X1 , X3 , X4 )
and τ2+ (delete) = h1 7→ X1 i and τ2+ (permute) = h2 7→ [X3 |X4 ]i.
4.3
Looping Conditions
One of the main purposes of this article is the inference of classes of atomic queries that left loop
w.r.t. a given logic program. Classes of atomic queries we consider are defined by pairs (A, τ + )
where A is an atom and τ + is a set of positions with associated terms. Such a pair denotes the
set of queries A↑τ + , the definition of which is similar to that of the expansion of an atom, see
Definition 3.21.
Definition 4.6 Let A be an atom and τ + be a set of positions with associated terms. Then A↑τ +
denotes the class of atomic queries defined as:
def
A↑τ + = {A} ∪ {B ∈ T BL | B is ∆[τ + ]-more general than A} .
Once each element of A↑τ + left loops w.r.t. a logic program, we get what we call a looping
condition for that program:
Definition 4.7 (Looping Condition) Let P be a logic program. A looping condition for P is a
pair (A, τ + ) such that each element of A↑τ + left loops w.r.t. P .
The function infer loop cond takes as arguments a logic program P and a non-negative
integer max . Calling infer loop dict(P, max ), it first computes a loop dictionary Dict . Then,
it computes from Dict looping conditions for P as follows. The function extracts the pair (H, τ + )
from each element ([H ← B|BinSeq], τ + ) of Dict . By Proposition 4.3, H loops w.r.t. [H ←
B|BinSeq]. As τ + , hence ∆[τ + ], is DN for [H ← B|BinSeq], by Definition 3.18 each element of
H↑τ + loops w.r.t. [H ← B|BinSeq]. Finally, as [H ← B|BinSeq] ⊆ TPβ ↑ max ⊆ bin unf (P ), by
Theorem 2.1, each element of H↑τ + left loops w.r.t. P .
Example 4.8 (Example 4.4 continued) From each looping pair we have infered, we get the following information.
• (append ([X1 |X2 ], X3 , [X1 |X4 ]), τ1+ ) is a looping condition. So, each query append (t1 , t2 , t3 ),
where [X1 |X2 ] = t1 η and [X1 |X4 ] = t3 η for a substitution η and t2 is an instance of X3
(because τ1+ (append )(2) = X3 ), left loops w.r.t. APPEND3. In other words, each query
append (t1 , t2 , t3 ), where [X1 |X2 ] = t1 η and [X1 |X4 ] = t3 η for a substitution η and t2 is
any term, left loops w.r.t. APPEND3.
• (append3 (X1 , X2 , X3 , X4 ), τ2+ ) is a looping condition. As we have τ2+ (append3 )(2) = X2 ,
τ2+ (append3 )(3) = X3 and τ2+ (append3 )(4) = X4 , this means that each query of form
append3 (x1 , t2 , t3 , t4 ), where t2 , t3 and t4 are any terms, left loops w.r.t. APPEND3.
• (append3 ([], X1 , X2 , X3 ), τ3+ ) is a looping condition. So, as τ3+ (append3 )(3) = X2 , this
means that each query of form append3 ([], X1 , t, X3 ), where t is any term, left loops w.r.t.
APPEND3.
16
Example 4.9 (Example 4.5 continued) From each looping pair we have infered, we get the following information.
• (delete(X1 , [X2 |X3 ], [X2 |X4 ]), τ1+ ) is a looping condition. As τ1+ (delete)(1) = X1 , this means
that each query of form delete(t1 , t2 , t3 ), where t1 is any term and [X2 |X3 ] = t2 η and
[X2 |X4 ] = t3 η for a substitution η, left loops w.r.t. PERMUTE.
• (permute([X1 |X2 ], [X3 |X4 ]), τ2+ ) is a looping condition. As τ2+ (permute)(2) = [X3 |X4 ], this
means that each query of form permute(t1 , t2 ), where t1 is more general than [X1 |X2 ] and
t2 is any instance of [X3 |X4 ], left loops w.r.t. PERMUTE.
5
An Application: Proving Optimality of Termination Conditions
[26] presents a tool for inferring termination conditions that are expressed as multi-modes, i.e.
as disjunctions of conjunctions of propositions of form “the i-th argument is ground”. In this
section, we describe an algorithm that attempts proofs of optimality of such conditions using the
algorithms for non-termination inference of the previous section.
5.1
Optimal Terminating Multi-modes
Let P be a logic program and p ∈ ΠP be a relation symbol, with arity(p) = n. First, we describe
the language we use for abstracting sets of atomic queries:
Definition 5.1 (Mode) A mode mp for p is a subset of [1, n], and denotes the following set of
atomic goals: [mp ] = {p(t1 , . . . , tn ) ∈ TB L | ∀i ∈ mp Var (ti ) = ∅}. The set of all modes for p,
i.e. 2[1,n] , is denoted modes(p).
Note that if mp = ∅ then [mp ] = {p(t1 , . . . , tn ) ∈ TB L }. Since a logic procedure may have
multiple uses, we generalize:
Definition 5.2 (Multi-mode) A multi-mode Mp for p is a finite set of modes for p and denotes
the following set of atomic queries: [Mp ] = ∪m∈Mp [m].
Note that if Mp = ∅, then [Mp ] = ∅. Now we can define what we mean by terminating and
looping multi-modes:
Definition 5.3 (Terminating mode, terminating multi-mode) A terminating mode mp for p is a
mode for p such that any query in [mp ] left terminates w.r.t. P . A terminating multi-mode TM p
for p is a finite set of terminating modes for p.
Definition 5.4 (Looping mode, looping multi-mode) A looping mode mp for p is a mode for p
such that there exists a query in [mp ] which left loops w.r.t. P . A looping multi-mode LM p for p
is a finite set of looping modes for p.
As left termination is instantiation-closed, any mode that is “below” (less general than) a
terminating mode is also a terminating mode. Similarly, as left looping is generalization-closed,
any mode that is “above” (more general than) a looping mode is also a looping mode. Let us be
more precise:
Definition 5.5 (Less general, more general) Let Mp be a multi-mode for the relation symbol p.
We set:
less general (Mp ) =
more general (Mp ) =
{m ∈ modes(p) | ∃m′ ∈ Mp [m] ⊆ [m′ ]}
{m ∈ modes(p) | ∃m′ ∈ Mp [m′ ] ⊆ [m]}
17
looping modes(L, p):
in: L: a finite set of looping conditions
p: a predicate symbol
out: a looping multi-mode for p
1: LM p := ∅
2: for each (p(t1 , . . . , tn ), τ + ) ∈ L do
3:
mp := Dom(τ + (p)) ∪ {i ∈ [1, n] | Var (ti ) = ∅}
4:
LM p := LM p ∪ {mp }
5: return LM p
Figure 1:
We are now equipped to present a definition of optimality for terminating multi-modes:
Definition 5.6 (Optimal terminating multi-mode) A terminating multi-mode TM p for p is optimal if there exists a looping multi-mode LM p verifying:
modes(p) = less general (TM p ) ∪ more general (LM p )
Otherwise stated, given a terminating multi-mode TM p , if each mode which is not less general
than a mode of TM p is a looping mode, then TM p characterizes the operational behavior of p
w.r.t. left termination and our language for defining sets of queries.
Example 5.7 Consider the program APPEND. A well-known terminating multi-mode is the set
TM append = {{1}, {3}}. Indeed, any query of the form append(t,Ys,Zs) or append(Xs,Ys,t),
where t is a ground term ( i.e. such that Var(t) = ∅), left terminates. We have:
less general (TM append ) = {{1}, {3}, {1, 2}, {1, 3}, {2, 3}, {1, 2, 3}}
On the other hand, append(Xs,[],Zs) left loops. Hence LM append = {{2}} is a looping condition
and more general (LM append ) = {∅, {2}}.
Since modes(append ) = less general (TM append ) ∪ more general (LM append ), we conclude that
the terminating multi-mode TM append is optimal.
5.2
Algorithms
Suppose we hold a finite set L of looping conditions for P . Then, each element (p(t1 , . . . , tn ), τ + )
of L provides an obvious looping mode for p: it suffices to take {i ∈ [1, n] | Var(ti ) = ∅}. But
actually, we can extract more information from L. Let p(t′1 , . . . , t′n ) be an atom such that:
• for each hi 7→ ui i ∈ τ + (p), t′i is a ground instance of ui ,
• for each i in [1, n] \ Dom(τ + (p)), t′i = ti .
Then, p(t′1 , . . . , t′n ) belongs to p(t1 , . . . , tn )↑τ + , hence it left loops w.r.t. P . Consequently, we
have that Dom(τ + (p)) ∪ {i ∈ [1, n] | Var (ti ) = ∅} is a looping mode for p. The function
looping modes of Fig. 1 is an application of these remarks.
Now we have the essential material for the design of a tool that attempts proofs of optimality
of left terminating multi-modes computed by a termination inference tool as e.g. cTI [26] or
TerminWeb [17]. For each pair (p, ∅) in the set the function optimal tc of Fig. 2 returns, we can
conclude that the corresponding TM p is the optimal terminating multi-mode which characterizes
the operational behavior of p with respect to Lterm .
18
optimal tc(P , max , {TM p }p∈ΠP ):
in: P : a logic program
max : a non-negative integer
{TM p }p∈ΠP : a finite set of terminating multi-modes
out: a finite set of pairs (p, Mp ) such that p ∈ ΠP and
Mp is a multi-mode for p with no information w.r.t. its left behaviour
note: if for each p ∈ Πp , Mp = ∅, then {TM p }p∈ΠP is optimal
1:
2:
3:
4:
5:
6:
7:
Res := ∅
L := infer loop cond(P, max )
for each p ∈ ΠP do
LM p := looping modes(L, p)
Mp := modes(p) \ (less general(TM p ) ∪ more general(LM p ))
Res := Res ∪ {(p, Mp )}
return Res
Figure 2:
Example 5.8 (Example 4.8 continued) We apply our algorithm to the program APPEND3 of Example 4.4. We get that
L := { (append ([X1 |X2 ], X3 , [X1 |X4 ]),
(append3 (X1 , X2 , X3 , X4 ),
(append3 ([], X1 , X2 , X3 ),
τ1+ ),
τ2+ ),
τ3+ )
}
is a finite set of looping conditions for APPEND3 (see Example 4.8) with
Dom(τ1+ (append ))
=
{2}
Dom(τ2+ (append3 )) =
{2, 3, 4}
Dom(τ3+ (append3 ))
{3}
=
So, for append we have:
LM append :=
more general(LM append ) =
looping modes(L, append ) = {{2}}
{∅, {2}}
TM append
less general(TM append )
=
=
{{1}, {3}}
{{1}, {3}, {1, 2}, {1, 3}, {2, 3}, {1, 2, 3}}
Mappend
=
{}
For append3 , we get:
• the looping mode {2, 3, 4} from (append3 (X1 , X2 , X3 , X4 ), τ2+ ) and
• the looping mode mp := {1, 3} from (append3 ([], X1 , X2 , X3 ), τ3+ ) (notice that 3 ∈ mp because
Dom(τ3+ (append3 )) = {3} and 1 ∈ mp because of constant [] which is the first argument of
append3 ([], X1 , X2 , X3 ).)
19
So, we have:
LM append3 :=
more general(LM append3 ) =
looping modes(L, append3 ) = {{2, 3, 4}, {1, 3}}
{∅, {1}, {2}, {3}, {4}, {1, 3}, {2, 3}, {2, 4},
{3, 4}, {2, 3, 4}}
TM append3
less general(TM append3 )
Mappend3
=
=
{{1, 2}, {1, 4}}
{{1, 2}, {1, 4}, {1, 2, 3}, {1, 2, 4}, {1, 3, 4},
=
{1, 2, 3, 4}}
{}
Hence in both cases, we have characterized the left behaviour of the predicates by using two complementary tools.
5.3
An Experimental Evaluation
We have implemented1 the algorithms presented in Sections 4 and 5.2. The binary unfoldings
algorithm is derived from the one described in [7], where we added time stamps to precisely
control what is computed at each iteration. Looping modes are computed starting from the leaves
of the call graph then moving up to its roots. The cTI termination inference tool2 is detailed in
[26, 24]. Here is the configuration we used for this experiment: Intel 686, 2.4GHz, 512Mb, Linux
2.4, SICStus Prolog 3.10.1, 24.8 MLips. Timings in seconds are average over 10 runs.
First we have applied them on some small programs from standard benchmarks of the termination analysis literature [30, 2, 9] (predefined predicates were erased). The column opt? of Table
1 indicates whether the result of cTI (see [26]) is proved optimal (X) or not (? ). The column
max gives the least non-negative integer implying optimality or the least non-negative integer n
where it seems we get the most precise information from non-termination inference (i.e. for n and
n + 1, the analyser delivers the same results). Then timings in seconds (t[s]) appear, followed by
a pointer to a comment to the notes below.
Notes:
1. The predicate fold/3 is defined by:
fold(X,[],X).
fold(X,[Y|Ys],Z) :- op2(X,Y,V), fold(V,Ys,Z).
When the predicate op2/3 is defined by the fact op2(A,B,C), the result of cTI is optimal.
When the predicate op2/3 is defined by the fact op2(a,b,c), no looping mode is found and
the result of cTI is indeed sub-optimal as the query fold(X,Y,Z) terminates.
2. Termination proofs for mergesort require the list-size norm, while cTI applies the term-size
norm.
3. The result of cTI is not optimal. The analyzed program:
p(A,B) :- q(A,C),p(C,B).
p(A,A).
q(a,b).
has finite binary unfoldings because there is no function symbol. Hence its termination is
decidable (see [7]). This could be easily detected at analyze time. We notice that no looping
mode is found. But as any constant is mapped to 0 by the term-size norm, the modes
modes(p) remain undecided for cTI while they all terminate.
1 Available
2 Available
from http://www.univ-reunion.fr/~gcc
from http://www.cs.unipr.it/cTI
20
Table 1: Some De Schreye’s, Apt’s, and Plümer’s programs.
program
permute
duplicate
sum
merge
dis-con
reverse
append
list
fold
lte
map
member
mergesort
mergesort ap
naive rev
ordered
overlap
permutation
quicksort
select
subset
sum peano
pl2.3.1
pl3.5.6
pl4.4.6a
pl4.5.2
pl4.5.3a
pl5.2.2
pl7.2.9
pl7.6.2a
pl7.6.2b
pl7.6.2c
pl8.3.1a
pl8.4.1
pl8.4.2
top-level predicate
permute(X,Y)
duplicate(X,Y)
sum(X,Y,Z)
merge(X,Y,Z)
dis(X)
reverse(X,Y,Z)
append(X,Y,Z)
list(X)
fold(X,Y,Z)
goal
map(X,Y)
member(X,Y)
mergesort(X,Y)
mergesort ap(X,Y,Z)
naive rev(X,Y)
ordered(X)
overlap(X,Y)
permutation(X,Y)
quicksort(X,Y)
select(X,Y,Z)
subset(X,Y)
sum(X,Y,Z)
p(X,Y)
p(X)
perm(X,Y)
s(X,Y)
p(X)
turing(X,Y,Z,T)
mult(X,Y,Z)
reach(X,Y,Z)
reach(X,Y,Z,T)
reach(X,Y,Z,T)
minsort(X,Y)
even(X)
e(X,Y)
cTI
term-cond
X
X ∨Y
X ∨Y ∨Z
(X ∧ Y ) ∨ Z
X
X
X ∨Z
X
Y
1
X ∨Y
Y
0
Z
X
X
X ∧Y
X
X
Y ∨Z
X ∧Y
Y ∨Z
0
X
X
0
0
0
X ∧Y
0
0
Z∧T
X
X
X
21
t[s]
0.01
0.01
0.01
0.02
0.02
0.02
0.01
0.01
0.01
0.01
0.01
0.01
0.04
0.08
0.02
0.01
0.01
0.03
0.05
0.01
0.01
0.01
0.01
0.01
0.02
0.03
0.01
0.08
0.02
0.02
0.02
0.02
0.03
0.02
0.05
opt?
X
X
X
X
X
X
X
X
?
X
X
X
?
?
X
X
X
X
X
X
X
X
?
X
X
X
X
?
X
?
?
?
X
X
X
Optimal
max
t[s]
1
0.01
1
0.01
1
0.01
1
0.01
2
0.01
1
0.01
1
0.01
1
0.01
2
0.01
1
0.01
2
0.01
1
0.01
2
0.01
2
0.02
1
0.01
1
0.01
2
0.01
1
0.01
1
0.01
1
0.01
2
0.01
1
0.01
1
0.01
2
0.01
1
0.01
1
0.01
1
0.01
2
0.03
4
0.03
1
0.01
1
0.01
2
0.02
2
0.02
2
0.01
3
0.04
cf.
note 1
note 2
note 3
note 4
note 5
note 6
Table 2: Some middle-sized programs.
program
name
ann
bid
boyer
browse
credit
peephole
plan
qplan
rdtok
read
warplan
clauses
177
50
136
30
57
134
29
148
55
88
101
cTI
Q%
48
100
84
53
100
88
100
61
44
52
32
t[s]
1.00
0.14
0.30
0.26
0.11
1.08
0.11
1.13
0.65
1.72
0.49
max=1
Opt%
t[s]
46
0.14
55
0.02
80
0.03
46
0.03
91
0.02
23
0.06
68
0.02
50
0.11
44
0.11
39
0.04
37
0.07
Optimal
max=2
Opt%
t[s]
68
1.34
90
0.08
96
0.22
80
0.18
95
0.11
70
3.62
81
0.09
79
1.60
88
40.2
47
0.80
83
0.99
max=3
Opt%
t[s]
74
32.4
95
0.50
100
3.66
100
6.05
100
4.46
70
406
81
0.37
81
1911
?
> 3600
47
10.9
91
21.5
4. The analyzed program (from [30], p. 64) simulates a Turing machine. The result of cTI is
optimal.
5. With respect to the program:
mult(0,A,0).
mult(s(A),B,C) :- mult(A,B,D),add(D,B,C).
add(0,A,A).
add(s(A),B,s(C)) :- add(A,B,C).
the query mult(s(s(0)),A,B) is automatically detected as looping, although mult(0,A,B)
and mult(s(0),A,B) do terminate.
6. These three programs propose various definitions of the reachability relation between two
nodes in a list of edges. For the first and the third definition, cTI is indeed optimal. For the
second one, cTI is not optimal.
Next, we have applied the couple of analyzers to some middle-sized Prolog programs, see Table
2. Again, predefined predicates were all erased, while they are usually taken into account for cTI
which of course improves the analysis. In other words, we only consider the logic programming
skeleton of each program. The first two columns give the name of the file and its size (number of
clauses). The fourth column indicates the running time (in seconds) of the termination analysis,
while the third column is the ratio of predicates for which a non-false termination condition is
computed over the total number of predicates defined in the program. For instance, cTI is able
to show that there is at least one terminating mode for 48% of the predicates defined in the
program ann. We ran the non-termination analyzer with 1 ≤ max ≤ 3 iterations. For each value
of max, we give the running time (in seconds) and the ratio of predicates for which looping modes
complement terminating modes. For example, with respect to the program ann, for max = 2 we
get the full complete mode termination behavior of 68% of all the defined predicates.
We note that when we increase max, we obtain better results but the running times also
increase, which is fairly obvious. For max = 3, we get good to optimal results but the binary
unfoldings approach reveals its potentially explosive nature: we aborted the analysis of rdtok
after one hour of computation.
In conclusion, from such a naive implementation, we were rather surprised by the quality of the
combined analysis. Adopting some more clever implementation schemes, for instance computing
the binary unfoldings in a demand driven fashion, could be investigated to improve the running
times.
22
6
Related Works
Some extensions of the Lifting Theorem with respect to infinite derivations are presented in [18],
where the authors study numerous properties of finite failure. The non-ground finite failure set
of a logic program is defined as the set of possibly non-ground atoms which admit a fair finitely
failed SLD-tree w.r.t. the program. This denotation is shown correct in the following sense. If two
programs have the same non-ground finite failure set, then any ground or non-ground goal which
finitely fails w.r.t. one program also finitely fails w.r.t. the other. Such a property is false when we
consider the standard ground finite failure set. The proof of correctness of the non-ground finite
failure semantics relies on the following result. First, a derivation is called non-perpetual if it is
a fair infinite derivation and there exists a finite depth from which unfolding does not instantiate
the original goal any more. Then the authors define the definite answer goal of a non-perpetual
derivation as the maximal instantiation of the original goal. A crucial lemma states that any
instance of the definite answer goal admits a similar non-perpetual derivation. Compared to our
work, note that we do not need fairness as an hypothesis for our results. On the other hand,
investigating the relationships between non-ground arguments of the definite answer and neutral
arguments is an interesting problem.
In [35], the authors present a dynamic approach to characterize (in the form of a necessary
and sufficient condition) termination of general logic programs. Their technique employs some
key dynamic features of an infinite generalized SLDNF-derivation, such as repetition of selected
subgoals and recursive increase in term size.
Loop checking in logic programming is also a subject related to our work. In this area, [5] sets
up some solid foundations. A loop check is a device to prune derivations when it seems appropriate.
A loop checker is defined as sound if no solution is lost. It is complete if all infinite derivations
are pruned. A complete loop check may also prune finite derivations. The authors show that even
for function-free programs (also known as Datalog programs), sound and complete loop checks are
out of reach. Completeness is shown only for some restricted classes of function-free programs.
We now review loop checking in more details. To our best knowledge, among all existing
loop checking mechanisms only OS-check [32], EVA-check [34] and VAF-check [36] are suitable for
logic programs with function symbols. They rely on a structural characteristic of infinite SLDderivations, namely, the growth of the size of some generated subgoals. This is what the following
theorem states.
Theorem 6.1 Consider an infinite SLD-derivation ξ where the leftmost selection rule is used.
Then there are infinitely many queries Qi1 , Qi2 , . . . (with i1 < i2 < . . . ) in ξ such that for
any j ≥ 1, the selected atom Aij of Qij is an ancestor of the selected atom Aij+1 of Qij+1 and
size(Aij+1 ) ≥ size(Aij ).
Here, size is a given function that maps an atom to its size which is defined in terms of the number
of symbols appearing in the atom. As this theorem does not provide any sufficient condition to
detect infinite SLD-derivations, the three loop checking mechanisms mentioned above may detect
finite derivations as infinite. However, these mechanisms are complete w.r.t. the leftmost selection
rule i.e. they detect all infinite loops when the leftmost selection rule is used.
OS-check (for OverSize loop check) was first introduced by Shalin [31, 32] and was then formalized by Bol [4]. It is based on a function size that can have one of the three following definitions:
for any atoms A and B, either size(A) = size(B), either size(A) (resp. size(B)) is the count of
symbols appearing in A (resp. B), either size(A) ≤ size(B) if for each i, the count of symbols of
the i-th argument of A is smaller than or equal to that of the i-th argument of B. OS-check says
that an SLD-derivation may be infinite if it generates an atomic subgoal A that is oversized, i.e.
that has ancestor subgoals which have the same predicate symbol as A and whose size is smaller
than or equal to that of A.
EVA-check (for Extented Variant Atoms loop check) was introduced by Shen [34]. It is based
on the notion of generalized variants (if Gi and Gj (i < j) are two goals in an SLD-derivation, an
atom A in Gj is a generalized variant of an atom A′ in Gi if A is a variant of A′ except for some
arguments whose size increases from A′ to A via a set of recursive clauses.) EVA-check says that
23
an SLD-derivation may be infinite if it generates an atomic subgoal A that is a generalized variant
of some of its ancestor A′ . Here the size function that is used applies to predicate arguments,
i.e. to terms, and it is fixed: it is defined as the the count of symbols that appear in the terms.
EVA-check is more reliable than OS-check because it is less likely to mis-identify infinite loops
[34]. This is mainly due to the fact that, unlike OS-check, EVA-check refers to the informative
internal structure of subgoals.
VAF-check (for Variant Atoms loop check for logic programs with Functions) was proposed by
Shen et al. [36]. It is based on the notion of expanded variants. An atom A is an expanded variant
of an atom A′ if, after variable renaming, A becomes B that is the same as A′ except that there
may be some terms at certain positions in A′ , each A′ [i] . . . [k] of which grows in B into a function
B[i] . . . [k] = f (. . . , A′ [i] . . . [k], . . . ) (here, we use A′ [i] . . . [k] (resp. B[i] . . . [k]) to refer to the k-th
argument of . . . of the i-th argument of A′ (resp. B)). VAF-check says that an SLD-derivation
may be infinite if it generates an atomic subgoal A that is an expanded variant of some of its
ancestor A′ . VAF-check is as reliable as and more efficient than EVA-check [36].
The main difference with our work is that we want to infer atomic queries which are guaranteed
to be left looping. Hence, we consider sufficient conditions for looping, in contrast to the above
mentioned methods which consider necessary conditions. Our technique returns a set of queries
for which it has pinpointed one infinite derivation. Hence, we are not interested in soundness
as we do not care of finite derivations, nor in completeness, as the existence of just one infinite
derivation suffices. Of course, using the ∆-subsumption test as a loop checker leads to a device that
is neither sound (since ∆-subsumption is a particular case of subsumption) nor complete (since
the ∆-subsumption test provides a sufficient but not necessary condition). This is illustrated by
the following example.
Example 6.2 Let c := p(X, X) ← p(f (X), f (X)). As the arguments of the head of c have one
common variable X, every set of positions with associated terms τ + that is DN for {c} is such
that the domain of τ + (p) is empty (see (DN1) in Definition 3.35). Notice that from the head
p(X, X) of c we get
p(X, X) =⇒ p(f (X), f (X)) =⇒ · · · =⇒ p(f n (X), f n (X)) =⇒ · · ·
c
c
c
c
As the arguments of p grow from step to step, there cannot be any query in the derivation that is
∆[τ + ]-more general than one of its ancestors. Consequently, we can not conclude that p(X, X)
left loops w.r.t {c}.
On the other hand, using loop checking approaches to infer classes of atomic left looping queries
is not satisfactory because, as we said above, non-looping queries may be mis-identified as looping.
Example 6.3 We cannot replace, in Corollary 3.2, the subsumption test by the expanded variant
test used in VAF-check because, for instance, in the clause c := p(a) ← p(f (a)), we have: p(f (a))
is an expanded variant of p(a), but p(a) does not loop w.r.t. c.
Finally, [10] is also related to our study. In this paper, the authors describe an algorithm for
detecting non-terminating queries to clauses of the type p(· · · ) ← p(· · · ). The algorithm is able to
check if such a given clause has no non-terminating queries or has a query which either loops or fails
due to occur check. Moreover, given a linear atomic goal (i.e. a goal where all variable occurs at
most once), the algorithm is able to check if the goal loops or not w.r.t. the clause. The technique
of the algorithm is based on directed weighted graphs [14] and on a necessary and sufficient
condition for the existence of non-terminating queries to clauses of the type p(· · · ) ← p(· · · ). This
condition is proved in [8] and is expressed in terms of rational trees.
7
Conclusion
We have presented a extension of the subsumption test which allows to disregard some arguments,
termed neutral arguments, while checking for subsumption. We have proposed two syntactic
24
criteria for statically identifying neutral arguments. From these results, in the second part of this
report we have described algorithms for automating non-termination analysis of logic programs,
together with correctness proofs. Finally, we have applied these techniques to check the optimality
of termination conditions for logic programs.
This paper leaves numerous questions open. For instance, it might be interesting to try to
generalize this approach to constraint logic programming [19]. Can we obtain higher level proofs
compared to those we give? Can we propose more abstract criteria for identifying neutral arguments? A first step in this direction is presented in [29]. Also, our work aims at inferring classes
of atomic left looping queries, using a bottom-up point of view. Experimental data show that it
may sometimes lead to prohibitive time/space costs. How can we generate only the useful binary
clauses without fully computing the iterations of this TP -like operator? Or can we adapt our
algorithms towards a more efficient correct top-down approach for checking non-termination?
Acknowledgments
We thank Ulrich Neumerkel for numerous discussions on this topic, Roberto Bagnara and anonymous referees for interesting suggestions.
References
[1] K. R. Apt. From Logic Programming to Prolog. Prentice Hall, 1997.
[2] K. R. Apt and D. Pedreschi. Modular termination proofs for logic and pure Prolog programs.
In G. Levi, editor, Advances in Logic Programming Theory, pages 183–229. Oxford University
Press, 1994.
[3] T. Arts and H. Zantema. Termination of logic programs using semantic unification.
In Logic Program Synthesis and Transformation, volume 1048 of Lecture Notes in Computer Science. Springer-Verlag, Berlin, 1996. TALP can be used online at the address:
http://bibiserv.techfak.uni.biekefeld.de/talp/.
[4] R. N. Bol. Loop checking in partial deduction. Journal of Logic Programming, 16:25–46,
1993.
[5] R. N. Bol, K. R. Apt, and J. W. Klop. An analysis of loop checking mechanisms for logic
programs. Theoretical Computer Science, 86:35–79, 1991.
[6] K. L. Clark. Predicate logic as a computational formalism. Technical Report Doc 79/59,
Logic Programming Group, Imperial College, London, 1979.
[7] M. Codish and C. Taboch. A semantic basis for the termination analysis of logic programs.
Journal of Logic Programming, 41(1):103–123, 1999.
[8] D. De Schreye, M. Bruynooghe, and K. Verschaetse. On the existence of nonterminating
queries for a restricted class of Prolog-clauses. Artificial Intelligence, 41:237–248, 1989.
[9] D. De Schreye and S. Decorte. Termination of logic programs : the never-ending story. Journal
of Logic Programming, 19-20:199–260, 1994.
[10] D. De Schreye, K. Verschaetse, and M. Bruynooghe. A practical technique for detecting nonterminating queries for a restricted class of Horn clauses, using directed, weighted graphs. In
Proc. of ICLP’90, pages 649–663. The MIT Press, 1990.
[11] P. Deransart and G. Ferrand. Programmation en logique avec négation: présentation
formelle. Technical Report 87/3, Laboratoire d’Informatique, Département de Mathématiques
et d’Informatique, Université d’Orleans, 1987.
25
[12] N. Dershowitz, N. Lindenstrauss, Y. Sagiv, and A. Serebrenik. A general framework
for automatic termination analysis of logic programs. Applicable Algebra in Engineering,Communication and Computing, 12(1/2):117–156, 2001.
[13] P. Devienne. Weighted graphs, a tool for expressing the behavious of recursive rules in logic
programming. pages 397–404. OHMSHA Ltd. Tokyo and Springer-Verlag, 1988. Proc. of the
Inter. Conf. on Fifth Generation Computer Systems 88, Tokyo, Japan.
[14] P. Devienne. Weighted graphs: A tool for studying the halting problem and time complexity
in term rewriting systems and logic programming. Theoretical Computer Science, 75(2):157–
215, 1990.
[15] P. Devienne, P.Lebègue, and J-C. Routier. Halting problem of one binary Horn clause is
undecidable. In LNCS, volume 665, pages 48–57. Springer-Verlag, 1993. Proc. of STACS’93.
[16] M. Gabbrielli and R. Giacobazzi. Goal independency and call patterns in the analysis of
logic programs. In Proceedings of the ACM Symposium on applied computing, pages 394–399.
ACM Press, 1994.
[17] S. Genaim and M. Codish. Inferring termination condition for logic programs using backwards
analysis. In Proceedings of Logic for Programming, Artificial intelligence and Reasoning,
Lecture Notes in Computer Science. Springer-Verlag, Berlin, 2001. TerminWeb can be used
online from http://www.cs.bgu.ac.il/~codish.
[18] R. Gori and G. Levi. Finite failure is and-compositional. Journal of Logic and Computation,
7(6):753–776, 1997.
[19] J. Jaffar and J. L. Lassez. Constraint logic programming. In Proc. of the ACM Symposium
on Principles of Programming Languages, pages 111–119. ACM Press, 1987.
[20] N. Lindenstrauss. TermiLog: a system for checking termination of queries to logic programs,
1997. http://www.cs.huji.ac.il/~naomil.
[21] J. W. Lloyd. Foundations of Logic Programming. Springer-Verlag, 1987.
[22] M. Martelli M. Falaschi, G. Levi and C. Palamidessi. A model-theoretic reconstruction of the
operational semantics of logic programs. Information and Computation, 102(1):86–113, 1993.
[23] F. Mesnard. Inferring left-terminating classes of queries for constraint logic programs by
means of approximations. In M. J. Maher, editor, Proc. of the 1996 Joint Intl. Conf. and
Symp. on Logic Programming, pages 7–21. MIT Press, 1996.
[24] F. Mesnard and R. Bagnara. cTI: a constraint-based termination inference tool for ISOProlog. Theory and Practice of Logic Programming, 2004. To appear.
[25] F. Mesnard and U. Neumerkel. cTI: a tool for inferring termination conditions of ISO-Prolog,
april 2000. http://www.complang.tuwien.ac.at/cti.
[26] F. Mesnard and U. Neumerkel. Applying static analysis techniques for inferring termination
conditions of logic programs. In P. Cousot, editor, Static Analysis Symposium, volume 2126
of Lecture Notes in Computer Science, pages 93–110. Springer-Verlag, Berlin, 2001.
[27] F. Mesnard, E. Payet, and U. Neumerkel. Detecting optimal termination conditions of logic
programs. In M. Hermenegildo and G. Puebla, editors, Proc. of the 9th International Symposium on Static Analysis, volume 2477 of Lecture Notes in Computer Science, pages 509–525.
Springer-Verlag, Berlin, 2002.
[28] R. O’Keefe. The Craft of Prolog. MIT Press, 1990.
26
[29] E. Payet and F. Mesnard. Non-termination inference of logic programs. In R. Giacobazzi,
editor, Proc. of the 11th International Symposium on Static Analysis, volume 3148 of Lecture
Notes in Computer Science. Springer-Verlag, Berlin, 2004.
[30] L. Plümer. Terminations proofs for logic programs. Number 446 in LNAI. Springer-Verlag,
Berlin, 1990.
[31] D. Sahlin. The mixtus approach to automatic partial evaluation of full Prolog. In S. Debray and M. Hermenegildo, editors, Proc. of the 1990 North American Conference on Logic
Programming, pages 377–398. MIT Press, Cambridge, MA, 1990.
[32] D. Sahlin. Mixtus: an automatic partial evaluator for full Prolog. New Generation Computing,
12(1):7–51, 1993.
[33] M. Schmidt-Schauss. Implication of clauses is undecidable. Theoretical Computer Science,
59:287–296, 1988.
[34] Y-D. Shen. An extended variant of atoms loop check for positive logic programs. New
Generation Computing, 15(2):187–204, 1997.
[35] Y-D. Shen, J-H. You, L-Y. Yuan, S. Shen, and Q. Yang. A dynamic approach to characterizing
termination of general logic programs. ACM Transactions on Computational Logic, 4(4):417–
434, 2003.
[36] Y-D. Shen, L-Y. Yuan, and J-H. You. Loops checks for logic programs with functions. Theoretical Computer Science, 266(1-2):441–461, 2001.
27
A
Algorithms
First, we define a pre-order relation over sets of positions with associated terms. Such a relation
is useful for the design of the algorithms that we present in the sequel of this section.
+
Definition A.1 (4 and τmax
)
• τ1+ 4 τ2+ if for each p ∈ Π, Dom(τ1+ (p)) ⊆ Dom(τ2+ (p)) and for each i ∈ Dom(τ1+ (p)),
τ2+ (p)(i) is more general than τ1+ (p)(i).
+
+
• τmax
denotes a set of positions with associated terms s.t. Dom(τmax
(p)) = [1, arity(p)] for
+
each p ∈ Π and τmax (p)(i) is a variable for each i ∈ [1, arity(p)].
A.1
DN Sets of Positions with Associated Terms for Binary Programs
We present below an algorithm for computing DN sets of positions with associated terms.
dna BinProg, τ1+ :
in: BinProg: a finite set of binary clauses
τ1+ : a set of positions with associated terms
out: τ2+ s.t. τ2+ 4 τ1+ and τ2+ is DN for BinProg
1:
2:
3:
4:
5:
6:
7:
τ2+ := τ1+
τ2+ := satisfy DN1(BinProg, τ2+ )
τ2+ := satisfy DN2(BinProg, τ2+ )
τ2+ := satisfy DN3(BinProg, τ2+ )
while satisfy DN4(BinProg, τ2+ ) 6= τ2+ do
τ2+ := satisfy DN4(BinProg, τ2+ )
return τ2+
The algorithm dna calls four auxiliary functions that correspond to conditions (DN1), (DN2)
(DN3) and (DN4) in the definition of a DN set of positions with associated terms (see Definition 3.35). These functions are detailed below.
After τ2+ := satisfy DN1(BinProg , τ2+ ) at line 2 of dna, τ2+ satisfies item (DN1) of Definition 3.35.
satisfy DN1 BinProg, τ1+ :
1: τ2+ := τ1+
2: for each p(s1 , . . . , sn ) ← B ∈ BinProg do
3:
E := {i ∈ [1, n] | Var (si ) ∩ Var ({sj | j 6= i}) = ∅}
4:
τ2+ (p) := τ2+ (p)|(Dom(τ2+ (p)) ∩ E)
5: return τ2+
After τ2+ := satisfy DN2(BinProg , τ2+ ) at line 3 of dna, τ2+ satisfies item (DN2) of Definition 3.35. Notice that less general at line 5 of satisfy DN2 is a function that returns the less
general term of two given terms; if none of the given terms is less general than the other, then
this function returns undefined.
28
satisfy DN2 BinProg , τ1+ :
1: τ2+ := τ1+
2: for each p(s1 , . . . , sn ) ← B ∈ BinProg do
3:
F := ∅
4:
for each i ∈ Dom(τ2+ (p)) do
5:
ui := less general(si , τ2+ (p)(i))
6:
if ui = undefined then F := F ∪ {i}
7:
else τ2+ (p)(i) := ui
8:
τ2+ (p) := τ2+ (p)|(Dom(τ2+ (p)) \ F )
9: return τ2+
After τ2+ := satisfy DN3(BinProg , τ2+ ) at line 4 of dna, τ2+ satisfies item (DN3) of Definition 3.35. The function satisfy DN3 is detailed below.
satisfy DN3 BinProg, τ1+ :
1: τ2+ := τ1+
2: for each H ← q(t1 , . . . , tm ) ∈ BinProg do
3:
F := ∅
4:
for each i ∈ Dom(τ2+ (q)) do
5:
if τ2+ (q)(i) is not more general than ti then F := F ∪ {i}
6:
τ2+ (q) := τ2+ (q)|(Dom(τ2+ (q)) \ F )
7: return τ2+
Finally, the function satisfy DN4 is defined as follows. After line 6 of dna, the set of positions
with associated terms τ2+ satisfies item (DN4) of Definition 3.35.
satisfy DN4 BinProg, τ1+ :
1: τ2+ := τ1+
2: for each p(s1 , . . . , sn ) ← q(t1 , . . . , tm ) ∈ BinProg do
3:
F := ∅
4:
for each i ∈ Dom(τ2+ (p)) do
5:
for each j ∈ [1, m] \ Dom(τ2+ (q)) do
6:
if Var(si ) ∩ Var (tj ) 6= ∅ then F := F ∪ {i}
+
7:
τ2 (p) := τ2+ (p)|(Dom(τ2+ (p)) \ F )
8: return τ2+
Proposition A.2 (Correctness of dna) Let BinProg be a binary program and τ1+ be a set of
positions with associated terms.
1. satisfy DN1(BinProg, τ1+ ), . . . , satisfy DN4(BinProg, τ1+ ) terminate;
2. satisfy DN1(BinProg, τ1+ ) 4 τ1+ , . . . , satisfy DN4(BinProg , τ1+ ) 4 τ1+ ;
3. dna(BinProg, τ1+ ) terminates;
4. dna(BinProg , τ1+ ) 4 τ1+ and dna(BinProg, τ1+ ) is a set of positions with associated terms that
is DN for BinProg .
Proof. We have:
1. satisfy DN1(BinProg, τ1+ ) terminates because, as BinProg is a finite set of binary clauses,
the loop at lines 2–4 terminates. Concerning satisfy DN2–4, the inner loops terminate since
for each p ∈ Π, Dom(τ2+ (p)) ⊆ [1, arity(p)] and the outer loop terminates as BinProg is a
finite set of binary clauses.
29
• satisfy DN1(BinProg , τ1+ ) 4 τ1+ :
Line 1, we start from τ1+ . Line 4, we have for each relation symbol p from the heads of
the clauses of BinProg :
2.
Dom(τ2+ (p)) ⊆ Dom(τ1+ (p)) and ∀i ∈ Dom(τ2+ (p)), τ2+ (p)(i) = τ1+ (p)(i) .
Hence, when we reach line 5, we have: satisfy DN1(BinProg, τ1+ ) 4 τ1+ .
• satisfy DN2(BinProg , τ1+ ) 4 τ1+ :
Line 1, we start from τ1+ . Then, for each relation symbol p from the heads of the
clauses of BinProg and for each i ∈ Dom(τ2+ (p)), either τ2+ (p)(i) is set to a less general
term than τ2+ (p)(i) (line 7) or i is removed from the domain of τ2+ (p) (lines 6 and 8).
Therefore, when we reach line 9, we have: satisfy DN2(BinProg , τ1+ ) 4 τ1+ .
• satisfy DN3(BinProg , τ1+ ) 4 τ1+ :
Line 1, we start from τ1+ . Line 6, we have for each relation symbol q from the bodies
of the clauses of BinProg:
Dom(τ2+ (q)) ⊆ Dom(τ1+ (q)) and ∀i ∈ Dom(τ2+ (q)), τ2+ (q)(i) = τ1+ (q)(i) .
Hence, when we reach line 7, we have: satisfy DN3(BinProg, τ1+ ) 4 τ1+ .
• satisfy DN4(BinProg , τ1+ ) 4 τ1+ :
Line 1, we start from τ1+ . Line 7, we have for each relation symbol p from the heads of
the clauses of BinProg :
Dom(τ2+ (p)) ⊆ Dom(τ1+ (p)) and ∀i ∈ Dom(τ2+ (p)), τ2+ (p)(i) = τ1+ (p)(i) .
(2)
Hence, when we reach line 8, we have: satisfy DN4(BinProg, τ1+ ) 4 τ1+ .
3. Each call to satisfy DN1, . . . , satisfy DN4 terminate. Moreover, concerning function
satisfy DN4, we mentioned above that (2) holds. As ⊂ is a well-founded relation over
the set of sets, the loop at lines 5–6 terminates.
4. Line 1, we start from τ1+ . Then satisfy DN1, . . . , satisfy DN4 weaken τ1+ with respect to
Definition 3.35. When we reach the fixpoint, the property holds.
A.2
Loop Dictionaries
A.2.1
Proof of Proposition 4.3, page 14
We proceed by induction on the length n of BinSeq.
• Basis. If n = 0, then, as ([H ← B], τ + ) is a looping pair, B is ∆[τ + ]-more general than H
and τ + is DN for H ← B, i.e. ∆[τ + ] is DN for H ← B by Theorem 3.39. Consequently, by
Proposition 3.20, H loops w.r.t. [H ← B].
• Induction. Suppose that for an n ≥ 0, each looping pair ([H ← B|BinSeq], τ + ) with BinSeq
of length n is such that H loops w.r.t. [H ← B|BinSeq].
If BinSeq is of length n + 1, it has form [H1 ← B1 |BinSeq 1 ] with BinSeq 1 of length n.
Moreover, as ([H ← B|BinSeq], τ + ) is a looping pair, there exists a set of positions with
associated terms τ1+ such that ([H1 ← B1 |BinSeq 1 ], τ1+ ) is a looping pair and B is ∆[τ1+ ]more general than H1 . So, by the induction hypothesis, H1 loops w.r.t. [H1 ← B1 |BinSeq 1 ]
i.e. H1 loops w.r.t. BinSeq. As B is ∆[τ1+ ]-more general than H1 and ∆[τ1+ ] is DN for
BinSeq, by Definition 3.18 B loops w.r.t. BinSeq. Therefore, by Corollary 3.3, H loops
w.r.t. [H ← B|BinSeq].
30
A.2.2
Computing a Loop Dictionary
The top-level function for inferring loop dictionaries from a logic program is the following. It uses
the auxiliary functions unit loop and loops from dict described below.
infer loop dict(P , max ):
in: P : a logic program
max : a non-negative integer
out: a loop dictionary, each element (BinSeq, τ + ) of which
is such that BinSeq ⊆ TPβ ↑ max
1: Dict := ∅
2: for each H ← B ∈ TPβ ↑ max do
3:
Dict := unit loop(H ← B, Dict )
4:
Dict := loops from dict(H ← B, Dict )
5: return Dict
The function unit loop is used to extract from a binary clause the most simple form of a
looping pair:
unit loop(H ← B, Dict):
in: H ← B: a binary clause
Dict: a loop dictionary
out: Dict′ : a loop dictionary, every element (BinSeq, τ + ) of which is
such that either (BinSeq, τ + ) ∈ Dict or BinSeq = [H ← B]
1:
2:
3:
4:
5:
Dict′ := Dict
+
τ + := dna([H ← B], τmax
)
+
if B is ∆[τ ]-more general than H then
Dict ′ := Dict ′ ∪ {([H ← B], τ + )}
return Dict ′
Termination of unit loop relies on that of dna. Partial correctness is deduced from the next
theorem.
Theorem A.3 (Partial correctness of unit loop) Let H ← B be a binary clause and Dict be a
loop dictionary. Then unit loop(H ← B, Dict) is a loop dictionary, every element (BinSeq, τ + )
of which is such that either (BinSeq, τ + ) ∈ Dict or BinSeq = [H ← B].
Proof. Let τ + be the set of positions with associated terms computed at line 2. If B is not
∆[τ + ]-more general than H then, at line 5 of unit loop, we have Dict ′ = Dict, so the theorem
holds.
Now suppose that B is ∆[τ + ]-more general than H. Then, at line 5 we have Dict ′ := Dict ∪
{([H ← B], τ + )} where Dict is a loop dictionary, τ + is DN for H ← B and B is ∆[τ + ]-more
general than H. So at line 5 Dict ′ is a loop dictionary, every element (BinSeq, τ + ) of which is
such that either (BinSeq, τ + ) ∈ Dict or BinSeq = [H ← B].
The function loops from dict is used to combine a binary clause to some looping pairs that
have already been infered in order to get some more looping pairs.
31
loops from dict(H ← B, Dict ):
in: H ← B: a binary clause
Dict: a loop dictionary
out: Dict ′ : a loop dictionary, every element (BinSeq, τ + ) of which is
such that (BinSeq, τ + ) ∈ Dict or BinSeq = [H ← B|BinSeq 1 ]
for some (BinSeq 1 , τ1+ ) in Dict
1: Dict ′ := Dict
2: for each [H1 ← B1 |BinSeq 1 ], τ1+ ∈ Dict do
3:
if B is ∆[τ1+ ]-more general than H1 then
4:
τ + := dna([H ← B, H1 ← B1 |BinSeq 1 ], τ1+ )
5:
Dict ′ := Dict ′ ∪ {([H ← B, H1 ← B1 |BinSeq 1 ], τ + )}
6: return Dict ′
Termination of loops from dict follows from finiteness of Dict (because Dict is a loop dictionary) and termination of dna. Partial correctness follows from the result below.
Theorem A.4 (Partial correctness of loops from dict) Let H ← B be a binary clause and Dict
be a loop dictionary. Then loops from dict(H ← B, Dict) is a loop dictionary, every element
(BinSeq, τ + ) of which is such that (BinSeq, τ + ) ∈ Dict or BinSeq = [H ← B|BinSeq 1 ] for some
(BinSeq 1 , τ1+ ) in Dict .
Proof. Upon initialization at line 1, Dict ′ is a loop dictionary. Suppose that before an iteration of
the loop at line 2, Dict ′ is a loop dictionary. Let ([H1 ← B1 |BinSeq 1 ], τ1+ ) ∈ Dict.
If the condition at line 3 is false, then Dict ′ remains unchanged, so after the iteration Dict ′
is still a loop dictionary. Otherwise, the pair ([H ← B, H1 ← B1 |BinSeq 1 ], τ + ) is added to
Dict ′ at line 5. Notice that this pair is a looping one because τ + defined at line 4 is DN for
[H ← B, H1 ← B1 |BinSeq 1 ] and ([H1 ← B1 |BinSeq 1 ], τ1+ ) is a looping pair and B is ∆[τ1+ ]-more
general than H1 . Therefore, after the iteration, Dict ′ is a loop dictionary. Finally, notice that as
Dict is a finite set, the loop at line 2 terminates. Hence, at line 6 Dict ′ is a finite set of looping
pairs i.e. Dict ′ is a loop dictionary.
Moreover, at line 1, each element of Dict ′ belongs to Dict . Then, during the loop, pairs of form
([H ← B|BinSeq 1 ], τ + ) are added to Dict ′ where BinSeq 1 is such that there exists (BinSeq 1 , τ1+ ) ∈
Dict . Consequently, at line 6 each element (BinSeq, τ + ) of Dict ′ is such that either (BinSeq, τ + ) ∈
Dict or BinSeq = [H ← B|BinSeq 1 ] for some (BinSeq 1 , τ1+ ) in Dict .
Finally, here is the correctness proof of the function infer loop dict.
Theorem A.5 (Correctness of infer loop dict) Let P be a logic program and max be a nonnegative integer. Then, infer loop dict(P, max ) terminates and returns a loop dictionary, every
element (BinSeq, τ + ) of which is such that BinSeq ⊆ TPβ ↑ max .
Proof. At line 1, Dict is initialized to ∅ which is a loop dictionary. Suppose that before an
iteration of the loop at line 2, Dict is a loop dictionary. Then at lines 3 and 4 unit loop and
loops from dict fullfil their specifications. Hence, the call to these functions terminates and after
the iteration Dict is still a loop dictionary. Finally, as TPβ ↑ max is a finite set, the loop at line 2
terminates and at line 5 Dict is a loop dictionary.
Moreover, at line 1 each element (BinSeq, τ + ) of Dict is such that BinSeq ⊆ TPβ ↑ max . Then,
during the loop, unit loop and loops from dict are called with clauses from TPβ ↑ max . So, by
Theorem A.3 and Theorem A.4, after the iteration each element (BinSeq, τ + ) of Dict is such that
BinSeq ⊆ TPβ ↑ max .
A.3
Looping Conditions
The following function computes a finite set of looping conditions for any given logic program.
32
infer loop cond(P , max ):
in: P : a logic program
max : a non-negative integer
out: a finite set of looping conditions for P
1:
2:
3:
4:
5:
L := ∅
Dict := infer loop dict(P, max )
for each ([H ← B|BinSeq], τ + ) ∈ Dict do
L := L ∪ {(H, τ + )}
return L
A call to infer loop cond(P , max ) terminates for any logic program P and any non-negative
integer max because infer loop dict(P, max ) at line 2 terminates and the loop at line 3 has a
finite number of iterations (because, by correctness of infer loop dict, Dict is finite.) Partial
correctness of infer loop cond follows from the next theorem.
Theorem A.6 (Partial correctness of infer loop cond) Let P be a logic program and max be a
non-negative integer. Then infer loop cond(P, max ) is a finite set of looping conditions for P .
Proof. By correctness of infer loop dict, Dict is a loop dictionary.
Let ([H ← B|BinSeq], τ + ) ∈ Dict . Then ([H ← B|BinSeq], τ + ) is a looping pair. Consequently, by Proposition 4.3, H loops w.r.t. [H ← B|BinSeq]. As τ + , hence ∆[τ + ], is DN for
[H ← B|BinSeq], by Definition 3.18 every atom that is ∆[τ + ]-more general than H loops w.r.t.
[H ← B|BinSeq].
As ([H ← B|BinSeq], τ + ) ∈ Dict , by Theorem A.5 we have
[H ← B|BinSeq] ⊆ TPβ ↑ max ⊆ bin unf (P ) .
So, by Theorem 2.1, H left loops w.r.t. P and every atom that is ∆[τ + ]-more general than H left
loops w.r.t. P . So, (H, τ + ) is a looping condition for P . Consequently, at line 5, L is a finite set
of looping conditions for P because, as Dict is finite, the loop at line 3 iterates a finite number of
times.
B
Proofs
B.1
Two Useful Lemmas
Lemma B.1 Let c := H ← B be a binary clause. Then, for every variant cγ of c such that
Var (cγ) ∩ Var (H) = ∅, we have H =⇒ Bγ ′ where γ ′ := γ|Var(B) \ Var (H).
c
Proof. Let µ := {xγ/x|x ∈ Var (H)}. By Claim B.2 below, µ is an mgu of Hγ and H. Hence, as
µ
Var (cγ) ∩ Var(H) = ∅, we have the left derivation step H =⇒ Bγµ where cγ is the input clause
c
used.
µ
If Var(B) = ∅, then Bγµ = Bγ ′ , so we have H =⇒ Bγ ′ i.e. H =⇒ Bγ ′ .
c
If Var(B) 6= ∅, take a variable x ∈ Var(B):
• if x ∈ Var(H), then x(γµ) = (xγ)µ = x by definition of µ,
• if x 6∈ Var(H), then x(γµ) = (xγ)µ = xγ by definition of µ.
µ
Hence, Bγµ = Bγ ′ , so we have H =⇒ Bγ ′ i.e. H =⇒ Bγ ′ .
c
c
Claim B.2 µ is an mgu of Hγ and H.
33
c
Proof. Let p(s1 , . . . , sn ) := H. The set of unifiers of Hγ and H is the same as that of E1 :=
{s1 γ = s1 , . . . , sn γ = sn }. Let E2 := {xγ = x | x ∈ Var (H)}. Notice that, as γ is a renaming,
if x, y ∈ Var (H) then x 6= y ⇒ xγ 6= yγ. Moreover, for each x ∈ Var(H), xγ 6= x because
Var (cγ) ∩ Var (H) = ∅. So, E2 is solved. Consequently, by Lemma 2.15 page 32 of [1], µ is an
mgu of E2 . Notice that, by Claim B.3 below, the set of unifiers of E1 is that of E2 . So µ is an
mgu of E1 i.e. µ is an mgu of Hγ and H.
Claim B.3 E1 and E2 have the same set of unifiers.
Proof. Let θ be a unifier of E1 . Let x ∈ Var (H) and let i ∈ [1, n] such that x ∈ Var (si ). Then
si γθ = si θ, so, if xk is an occurrence of x in si , we have xk γθ = xk θ i.e. (xk γ)θ = xk θ. As x
denotes any variable of H, we conclude that θ is a unifier of E2 . Conversely, let θ be a unifier of
E2 . Then, for each i ∈ [1, n], (si γ)θ = si θ by definition of E2 . Hence, θ is a unifier of E1 .
Lemma B.4 Let c := H ← B be a binary clause, cγ be a variant of c such that Var (cγ) ∩
Var (H) = ∅ and γ ′ := γ|Var (B)\Var (H). Then, there exists a renaming γ ′′ such that Bγ ′ = Bγ ′′ .
Proof. Let A := {x | x ∈ Ran(γ ′ ) and x 6∈ Dom(γ ′ )} and B := {x | x ∈ Dom(γ ′ ) and x 6∈
Ran(γ ′ )}. Notice that Ran(γ ′ ) and Dom(γ ′ ) have the same number of elements, so A and B have
the same number of elements. Let σ be a 1-1 and onto mapping from A to B. Then, γ ′′ := γ ′ ∪ σ
is a well-defined substitution, is such that Dom(γ ′′ ) = Ran(γ ′′ ), is 1-1 and is onto. Consequently,
γ ′′ is a renaming.
Now, let us prove that Bγ ′ = Bγ ′′ . If Var (B) = ∅, then the result is straightforward.
Otherwise, let x ∈ Var(B).
• If x ∈ Var (H) then, as Dom(γ ′ ) ⊆ Var (B) \ Var (H), we have x 6∈ Dom(γ ′ ), so xγ ′ = x.
Moreover, xγ ′′ = x(γ ′ ∪σ) = xσ = x because Dom(σ) ⊆ Ran(γ ′ ) and Ran(γ ′ )∩Var (H) = ∅.
Consequently, we have xγ ′ = xγ ′′ .
• If x 6∈ Var (H) and x ∈ Dom(γ ′ ) then xγ ′′ = x(γ ′ ∪ σ) = xγ ′ .
• If x 6∈ Var (H) and x 6∈ Dom(γ ′ ) then xγ ′ = x.
Now, suppose that x ∈ Dom(σ). Then, as Dom(σ) ⊆ Ran(γ ′ ) ⊆ Ran(γ), we have x ∈
Ran(γ). But, as γ is a renaming, Ran(γ) = Dom(γ), so we have x ∈ Dom(γ). As
x ∈ Var (B), as x 6∈ Var(H) and as γ ′ := γ|Var (B) \ Var (H), we have x ∈ Dom(γ ′ ).
Contradiction.
Consequently, x 6∈ Dom(σ), so xσ = x and xγ ′′ = x(γ ′ ∪ σ) = xσ = x. Finally, we have
proved that xγ ′ = xγ ′′ .
B.2
Proof of Corollary 3.2, page 4
By Lemma B.1 and Lemma B.4, we have H =⇒ Bγ ′′ where γ ′′ is a renaming. As by hypothesis
c
B is more general than H, then Bγ ′′ is more general than H. Therefore, by the One Step Lifting
Lemma 3.1, H loops w.r.t. {c}.
B.3
Proof of Corollary 3.3, page 4
By Lemma B.1 and Lemma B.4, we have H =⇒ Bγ ′′ where γ ′′ is a renaming. As Bγ ′′ is more
c
general than B and as B loops w.r.t. P , then, by the One Step Lifting Lemma 3.1, Bγ ′′ loops
w.r.t. P , so H loops w.r.t. P .
34
B.4
Proof of Proposition 3.16, page 7
If A is ∆-more general than B, we have, for a substitution η:
A = p(s1 , . . . , sn )
B = p(t1 , . . . , tn )
∀i ∈ [1, n] \ Dom(∆(p)), ti = si η
∀i ∈ Dom(∆(p)), ∆(p)(i)(si ) = true .
Let A′ be a variant of A. Then, there exists a renaming γ such that A′ = Aγ. As for each
i ∈ Dom(∆(p)), ∆(p)(i) is a variant independent term-condition, we have:
′
A = p(s1 γ, . . . , sn γ)
B = p(t1 , . . . , tn )
∀i ∈ [1, n] \ Dom(∆(p)), ti = si η = (si γ)(γ −1 η)
∀i ∈ Dom(∆(p)), ∆(p)(i)(si γ) = true .
Consequently, A′ is ∆-more general than B for γ −1 η, i.e. A′ is ∆-more general than B.
B.5
Proof of Proposition 3.17, page 7
⇐ By definition.
⇒ Let p(s1 , . . . , sn ) := A and p(t1 , . . . , tn ) := B. As A is ∆-more general than B, there exists a
substitution σ such that A is ∆-more general than B for σ. Notice that A is also ∆-more
general than B for the substitution obtained by restricting the domain of σ to the variables
appearing in the positions of A not distinguished by ∆. More precisely, let
η := σ|V ar({si | i ∈ [1, n] \ Dom(∆(p))}) .
Then,
Dom(η) ⊆ V ar(A)
(3)
and A is ∆-more general than B for η.
Now, let x ∈ Dom(η). Then, there exists i ∈ [1, n] \ Dom(∆(p)) such that x ∈ V ar(si ).
As A is ∆-more general than B for η and i ∈ [1, n] \ Dom(∆(p)), we have ti = si η. So, as
x ∈ V ar(si ), xη is a subterm of ti . Consequently, V ar(xη) ⊆ V ar(ti ), so V ar(xη) ⊆ V ar(B).
So, we have proved that for each x ∈ Dom(η), V ar(xη) ⊆ V ar(B), i.e. we have proved that
Ran(η) ⊆ V ar(B) .
(4)
Finally, (3) and (4) imply that Dom(η) ∪ Ran(η) ⊆ V ar(A, B) i.e. that
V ar(η) ⊆ V ar(A, B) .
B.6
Proof of Proposition 3.20, page 8
By Lemma B.1 and Lemma B.4, we have H =⇒ Bγ ′′ where γ ′′ is a renaming. As by hypothesis B
c
is ∆-more general than H, then by Proposition 3.16 Bγ ′′ is ∆-more general than H. Therefore,
as ∆ is DN for c, by Definition 3.18, H loops w.r.t. {c}.
B.7
Proof of Proposition 3.37, page 11
Let c := p(s1 , . . . , sn ) ← q(t1 , . . . , tm ) and c′ := p(s′1 , . . . , s′n ) ← q(t′1 , . . . , t′m ) be a variant of c.
Then, there exists a renaming γ such that c′ = cγ.
35
(DN1) Let i ∈ Dom(τ + (p)). Suppose that there exists j 6= i such that V ar(s′i ) ∩ V ar(s′j ) 6= ∅
and let us derive a contradiction.
Let x′ ∈ V ar(s′i ) ∩ V ar(s′j ). As s′j = sj γ, there exists x ∈ V ar(sj ) such that x′ = xγ.
For such an x, as j 6= i and as V ar(si ) ∩ V ar(sj ) = ∅ (because τ + is DN for c), we
have x 6∈ V ar(si ). So, as γ is a 1-1 and onto mapping from its domain to itself, we have
xγ 6∈ V ar(si γ)3 , i.e. x′ 6∈ V ar(s′i ). Contradiction!
Consequently, V ar(s′i ) ∩ V ar(s′j ) = ∅.
(DN2) Let hi 7→ ui i ∈ τ + (p). As si is more general than ui (because τ + is DN for c) and as s′i
is a variant of si , s′i is more general than ui .
(DN3) Let hj 7→ uj i ∈ τ + (q). As tj is an instance of uj (because τ + is DN for c) and as t′j is a
variant of tj , t′j is an instance of uj .
(DN4) Let i ∈ Dom(τ + (p)). Suppose there exists j 6∈ Dom(τ + (q)) such that V ar(s′i )∩V ar(t′j ) 6=
∅. Let us derive a contradiction.
Let x′ ∈ V ar(s′i ) ∩ V ar(t′j ). As t′j = tj γ and x′ ∈ V ar(t′j ), there exists x ∈ V ar(tj ) such that
x′ = xγ. For such an x, as the elements of V ar(si ) only occur in those tk s.t. k ∈ Dom(τ + (q))
(because τ + is DN for c) and as x ∈ V ar(tj ) with j 6∈ Dom(τ + (q)), we have x 6∈ V ar(si ).
So, as γ is a 1-1 and onto mapping from its domain to itself, we have xγ 6∈ V ar(si γ) (see
footnote 3), i.e. x′ 6= V ar(s′i ). Contradiction! So, for each j 6∈ Dom(τ + (q)), we have
V ar(s′i ) ∩ V ar(t′j ) = ∅.
Finally, we have established that τ + is DN for c′ .
B.8
DN Sets of Positions with Associated Terms Generate DN Filters
In this section, we give a proof of Theorem 3.39, page 12.
B.8.1
Context
All the results of this section are parametric to the following context:
• P is a binary program and τ + is a set of positions with associated terms that is DN for P ,
θ
• Q =⇒ Q1 is a left derivation step where
c
– c ∈ P,
– Q := p(t1 , . . . , tn ),
– c1 := p(s1 , . . . , sn ) ← B is the input clause used (consequently, c1 is a variant of c that
is variable disjoint from Q),
• Q′ := p(t′1 , . . . , t′n ) is ∆[τ + ]-more general than Q i.e., by Proposition 3.17, there exists a
substitution η such that V ar(η) ⊆ V ar(Q, Q′ ) and Q′ is ∆[τ + ]-more general than Q for η.
3 Because if xγ ∈ V ar(s γ), then either x ∈ V ar(s ), either xγ ∈ V ar(s ) and (xγ)γ = xγ. The former case is
i
i
i
impossible because we said that x 6∈ V ar(si ). The latter case is impossible too because (xγ)γ = xγ implies that
xγ 6∈ Dom(γ) i.e. x 6∈ Dom(γ) (because γ is a 1-1 and onto mapping from its domain to itself); so, x = xγ i.e., as
xγ ∈ V ar(si ), x ∈ V ar(si ).
36
B.8.2
Technical Definitions and Lemmas
Definition B.5 (Technical Definition) Let c′1 := p(s′1 , . . . , s′n ) ← B ′ be a binary clause such that
• V ar(c′1 ) ∩ V ar(Q, Q′ ) = ∅ and
• c1 = c′1 γ for some renaming γ satisfying V ar(γ) ⊆ V ar(c1 , c′1 ).
As c′1 is a variant of c1 and c1 is a variant of c, then c′1 is a variant of c. Moreover, as τ +
is DN for c, by Proposition 3.37, τ + is DN for c′1 . So, by (DN2) in Definition 3.35, for each
hi 7→ ui i ∈ τ + (p) there exists a substitution δi such that ui = s′i δi .
Moreover, as p(t′1 , . . . , t′n ) is ∆[τ + ]-more general than p(t1 , . . . , tn ), for each hi 7→ ui i ∈ τ + (p),
′
ti is an instance of ui . So, there exists a substitution δi′ such that t′i = ui δi′ .
For each i ∈ Dom(τ + (p)), we set
def
σi = (δi δi′ )|V ar(s′i ) .
Moreover, we set:
def
[
σ =
σi .
i∈Dom(τ + (p))
Lemma B.6 The set σ of Definition B.5 is a well-defined substitution.
Proof. Notice that, as τ + is DN for c′1 , by (DN1) in Definition 3.35, we have
∀i ∈ Dom(τ + (p)), ∀j ∈ [1, n] \ {i}, V ar(s′i ) ∩ V ar(s′j ) = ∅ .
Consequently,
∀i, j ∈ Dom(τ + (p)), i 6= j ⇒ Dom(σi ) ∩ Dom(σj ) = ∅ .
Moreover, for each i ∈ Dom(τ + (p)), σi is a well-defined substitution. So, σ is a well-defined
substitution.
Lemma B.7 (Technical Lemma) Let c′1 := p(s′1 , . . . , s′n ) ← B ′ be a binary clause such that
• V ar(c′1 ) ∩ V ar(Q, Q′ ) = ∅ and
• c1 = c′1 γ for some renaming γ satisfying V ar(γ) ⊆ V ar(c1 , c′1 ).
Let σ be the substitution of Definition B.5.
p(t′1 , . . . , t′n ) and p(s′1 , . . . , s′n ).
Then, the substitution σηγθ is a unifier of
Proof. The result follows from the following facts.
• For each hi 7→ ui i ∈ τ + (p), we have:
s′i σ = s′i σi = s′i δi δi′ = (s′i δi )δi′ = ui δi′ = t′i
and t′i σ = t′i because Dom(σ) ⊆ V ar(c′1 ) and V ar(Q′ ) ∩ V ar(c′1 ) = ∅. So, s′i σ = t′i σ and
s′i σηγθ = t′i σηγθ.
• For each i ∈ [1, n] \ Dom(τ + (p)), we have:
s′i ηγθ = (s′i η)γθ = s′i γθ = (s′i γ)θ = si θ
and
t′i ηγθ = (t′i η)γθ = ti γθ = (ti γ)θ = ti θ
θ
and si θ = ti θ because θ is a unifier of p(s1 , . . . , sn ) and p(t1 , . . . , tn ) (because Q =⇒ Q1 with
c
c1 as input clause used). So,
s′i ηγθ = t′i ηγθ
37
(5)
• For each i ∈ [1, n] \ Dom(τ + (p)), we also have:
– s′i σ = s′i because Dom(σ) = V ar {s′j | j ∈ Dom(τ + (p))} and, by (DN1) in Defini
tion 3.35, V ar {s′j | j ∈ Dom(τ + (p))} ∩ V ar(s′i ) = ∅;
– t′i σ = t′i because Dom(σ) ⊆ V ar(c′1 ) and V ar(Q′ ) ∩ V ar(c′1 ) = ∅.
Therefore, because of (5), s′i σηγθ = t′i σηγθ.
B.8.3
∆-Propagation
Now we extend, in the case of left derivations with atomic queries and binary clauses, the following
Propagation Lemma that is proved by Apt in [1] p. 54–56.
Lemma B.8 (Propagation) Let G, G1 , G′ and G′1 be some queries such that
G =⇒ G1
G′ =⇒ G′1
and
c
c
and
• G is an instance of G′
• in G and G′ atoms in the same positions are selected.
Then, G1 is an instance of G′1 .
First we establish the following result.
θ′
Lemma B.9 Suppose there exists a left derivation step of form Q′ =⇒ Q′1 where the input clause
c
is c′1 such that V ar(Q) ∩ V ar(c′1 ) = ∅. Then, Q′1 is ∆[τ + ]-more general than Q1 .
Proof. Notice that we have
V ar(Q) ∩ V ar(c1 ) = V ar(Q, Q′ ) ∩ V ar(c′1 ) = ∅ .
Moreover, as c1 is a variant of c′1 , there exists a renaming γ such that
V ar(γ) ⊆ V ar(c1 , c′1 ) and
c1 = c′1 γ .
Let c′1 := p(s′1 , . . . , s′n ) ← B ′ . Then,
Q1 = Bθ
and
Q′1 = B ′ θ′ .
τ + is DN for c and c′1 is a variant of c. So, by Proposition 3.37, τ + is DN for c′1 . Let σ be the
substitution of Definition B.5.
′
Let q(v1′ , . . . , vm
) := B ′ . As B = B ′ γ, B has form q(v1 , . . . , vm ).
• For each hj 7→ uj i ∈ τ + (q), vj′ is an instance of uj (because τ + is DN for c′1 and (DN3) in
Definition 3.35.)
• For each j ∈ [1, m] \ Dom(τ + (q)) we have:
vj′ σηγθ = (vj′ σ)ηγθ = vj′ ηγθ
because, by (DN4) in Definition 3.35
V ar(vj′ ) ∩ V ar {s′i | i ∈ Dom(τ + (p))} = ∅
with Dom(σ) = V ar {s′i | i ∈ Dom(τ + (p))} . Moreover,
vj′ ηγθ = (vj′ η)γθ = vj′ γθ
because V ar(η) ⊆ V ar(Q, Q′ ) and V ar(c′1 ) ∩ V ar(Q, Q′ ) = ∅. Finally,
vj′ γθ = (vj′ γ)θ = vj θ
because B = B ′ γ.
38
Consequently, we have proved that
′
q(v1′ , . . . , vm
) is ∆[τ + ]-more general than q(v1 , . . . , vm )θ for σηγθ
i.e. that B ′ is ∆[τ + ]-more general than Bθ for σηγθ i.e. that
B ′ is ∆[τ + ]-more general than Q1 for σηγθ .
(6)
But, by the Technical Lemma B.7, σηγθ is a unifier of p(s′1 , . . . , s′n ) and p(t′1 , . . . , t′n ). As θ′ is an
θ′
mgu of p(s′1 , . . . , s′n ) and p(t′1 , . . . , t′n ) (because Q′ =⇒ Q′1 with c′1 as input clause), there exists δ
c
such that σηγθ = θ′ δ. Therefore, we conclude from (6) that B ′ is ∆[τ + ]-more general than Q1
for θ′ δ which implies that B ′ θ′ is ∆[τ + ]-more general than Q1 for δ i.e. that Q′1 is ∆[τ + ]-more
general than Q1 for δ. Finally, we have proved that Q′1 is ∆[τ + ]-more general than Q1 .
Using the Propagation Lemma B.8, the preceding result can be extended as follows.
θ′
Proposition B.10 (∆-Propagation) Suppose there exists a left derivation step Q′ =⇒ Q′1 . Then
c
Q′1 is ∆[τ + ]-more general than Q1 .
θ′
Proof. Let c′1 be the input clause used in Q′ =⇒ Q′1 . Take a variant Q′′ of Q such that
c
V ar(Q′′ ) ∩ V ar(c′1 ) = ∅
and a variant c′′1 of c such that
V ar(c′′1 ) ∩ V ar(Q′′ ) = ∅ .
Then, the left resolvent Q′′1 of Q′′ and c exists with the input clause c′′1 . So, for some θ′′ , we have
θ ′′
Q′′ =⇒ Q′′1 with input clause c′′1 . Consequently, we have:
c
θ
Q =⇒ Q1
c
θ ′′
and Q′′ =⇒ Q′′1 .
c
Q and Q′′ are instances of each other because Q′′ is a variant of Q. So, by the Propagation
Lemma B.8 used twice, Q′′1 is an instance of Q1 and Q1 is an instance of Q′′1 . So,
Q′′1 is a variant of Q1 .
But we also have
θ ′′
Q′′ =⇒ Q′′1
c
′
θ′
and Q′ =⇒ Q′1
c
+
with Q that is ∆[τ ]-more general than Q′′ (because Q′′ is a variant
with input clauses
and
′
of Q and Q is ∆[τ ]-more general than Q) and V ar(Q′′ ) ∩ V ar(c′1 ) = ∅. So, by Lemma B.9,
c′′1
+
c′1 ,
(7)
Q′1 is ∆[τ + ]-more general than Q′′1 .
(8)
Finally, from (7) and (8) we have: Q′1 is ∆[τ + ]-more general than Q1 .
B.8.4
Epilogue
Theorem 3.39 is a direct consequence of the following result.
Proposition B.11 (One Step ∆-Lifting) Let c′ be a variant of c variable disjoint with Q′ . Then,
θ′
there exist θ′ and a query Q′1 that is ∆[τ + ]-more general than Q1 such that Q′ =⇒ Q′1 with input
c
clause c′ .
39
Proof. Let c′1 := p(s′1 , . . . , s′n ) ← B ′ be a variant of c1 . Then there exists a renaming γ such that
V ar(γ) ⊆ V ar(c1 , c′1 ) and c1 = c′1 γ. Suppose also that
V ar(c′1 ) ∩ V ar(Q, Q′ ) = ∅ .
By the Technical Lemma B.7, p(s′1 , . . . , s′n ) and p(t′1 , . . . , t′n ) unify. Moreover, as V ar(c′1 ) ∩
V ar(Q′ ) = ∅, p(s′1 , . . . , s′n ) and p(t′1 , . . . , t′n ) are variable disjoint. Notice that the following claim
holds.
Claim B.12 Suppose that the atoms A and H are variable disjoint and unify. Then, A also
unifies with any variant H ′ of H variable disjoint with A.
Proof. For some γ such that Dom(γ) ⊆ V ar(H ′ ), we have H = H ′ γ. Let θ be a unifier of A and
H. Then, Aγθ = Aθ = Hθ = H ′ γθ, so A and H ′ unify.
Therefore, as c′ is a variant of c′1 and c′ is variable disjoint with Q′ , p(t′1 , . . . , t′n ) and the head
of c′ unify. As they also are variable disjoint, we have
θ′
Q′ =⇒ Q′1
c
for some θ′ and Q′1 where c′ is the input clause used. Moreover, by the ∆-Propagation Proposition B.10, Q′1 is ∆[τ + ]-more general than Q1 .
40
| 6cs.PL
|
Sensing-Constrained LQG Control
arXiv:1709.08826v1 [math.OC] 26 Sep 2017
Vasileios Tzoumas,1,2 Luca Carlone,2 George J. Pappas,1 Ali Jadbabaie2
Abstract—Linear-Quadratic-Gaussian (LQG) control is concerned with the design of an optimal controller and estimator
for linear Gaussian systems with imperfect state information.
Standard LQG assumes the set of sensor measurements, to be
fed to the estimator, to be given. However, in many problems,
arising in networked systems and robotics, one may not be able to
use all the available sensors, due to power or payload constraints,
or may be interested in using the smallest subset of sensors
that guarantees the attainment of a desired control goal. In this
paper, we introduce the sensing-constrained LQG control problem,
in which one has to jointly design sensing, estimation, and control,
under given constraints on the resources spent for sensing.
We focus on the realistic case in which the sensing strategy has to
be selected among a finite set of possible sensing modalities. While
the computation of the optimal sensing strategy is intractable,
we present the first scalable algorithm that computes a nearoptimal sensing strategy with provable sub-optimality guarantees.
To this end, we show that a separation principle holds, which
allows the design of sensing, estimation, and control policies in
isolation. We conclude the paper by discussing two applications
of sensing-constrained LQG control, namely, sensing-constrained
formation control and resource-constrained robot navigation.
I. I NTRODUCTION
Traditional approaches to control of systems with partially
observable state assume the choice of sensors used to observe
the system is given. The choice of sensors usually results from
a preliminary design phase in which an expert designer selects
a suitable sensor suite that accommodates estimation requirements (e.g., observability, desired estimation error) and system
constraints (e.g., size, cost). Modern control applications, from
large networked systems to miniaturized robotics systems,
pose serious limitations to the applicability of this traditional
paradigm. In large-scale networked systems (e.g., smart grids
or robot swarms), in which new nodes are continuously added
and removed from the network, a manual re-design of the
sensors becomes cumbersome and expensive, and it is simply
not scalable. In miniaturized robot systems, while the set of
onboard sensors is fixed, it may be desirable to selectively
activate only a subset of the sensors during different phases of
operation, in order to minimize power consumption. In both
application scenarios, one usually has access to a (possibly
large) list of potential sensors, but, due to resource constraints
(e.g., cost, power), can only utilize a subset of them. Moreover,
1 The authors are with the Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA 19104 USA (email:
{pappagsg, vtzoumas}@seas.upenn.edu).
2 The authors are with the Institute for Data, Systems and Society, and
the Laboratory for Information and Decision Systems, Massachusetts Institute of Technology, Cambridge, MA 02139 USA (email: {jadbabai,
lcarlone, vtzoumas}@mit.edu).
This work was supported in part by TerraSwarm, one of six centers
of STARnet, a Semiconductor Research Corporation program sponsored by
MARCO and DARPA, and in part by AFOSR Complex Networks Program.
the need for online and large-scale sensor selection demands
for automated approaches that efficiently select a subset of
sensors to maximize system performance.
Motivated by these applications, in this paper we consider
the problem of jointly designing control, estimation, and
sensor selection for a system with partially observable state.
Related work. One body of related work is control over
band-limited communication channels, which investigates the
trade-offs between communication constraints (e.g., data rate,
quantization, delays) and control performance (e.g., stability)
in networked control systems. Early work provides results
on the impact of quantization [1], finite data rates [2], [3],
and separation principles for LQG design with communication constraints [4]; more recent work focuses on privacy
constraints [5]. We refer the reader to the surveys [6]–[8].
A second set of related work is sensor selection and scheduling, in which one has to select a (possibly time-varying) set of
sensors in order to monitor a phenomenon of interest. Related
literature includes approaches based on randomized sensor
selection [9], dual volume sampling [10], [11], convex relaxations [12], [13], and submodularity [14]–[16]. The third set
of related works is information-constrained (or informationregularized) LQG control [17], [18]. Shafieepoorfard and Raginsky [17] study rationally inattentive control laws for LQG
control and discuss their effectiveness in stabilizing the system.
Tanaka and Mitter [18] consider the co-design of sensing,
control, and estimation, propose to augment the standard LQG
cost with an information-theoretic regularizer, and derive an
elegant solution based on semidefinite programming. The main
difference between our proposal and [18] is that we consider
the case in which the choice of sensors, rather than being
arbitrary, is restricted to a finite set of available sensors.
Contributions. We extend the Linear-Quadratic-Gaussian
(LQG) control to the case in which, besides designing an optimal controller and estimator, one has to select a set of sensors
to be used to observe the system state. In particular, we formulate the sensing-constrained (finite-horizon) LQG problem
as the joint design of an optimal control and estimation policy,
as well as the selection of a subset of k out of N available
sensors, that minimize the LQG objective, which quantifies
tracking performance and control effort. We first leverage a
separation principle to show that the design of sensing, control,
and estimation, can be performed independently. While the
computation of the optimal sensing strategy is combinatorial
in nature, a key contribution of this paper is to provide the
first scalable algorithm that computes a near-optimal sensing
strategy with provable sub-optimality guarantees. We demonstrate the effectiveness of the proposed algorithm in numerical
experiments, and motivate the importance of the sensing-
constrained LQG problem, by considering two application
scenarios, namely, sensing-constrained formation control and
resource-constrained robot navigation.
Notation. Lowercase letters denote vectors and scalars, and
uppercase letters denote matrices. We use calligraphic fonts to
denote sets. The identity matrix of size n is denoted with In
(dimension is omitted when clear from the context). For a
matrix M and a vector v of appropriate dimension, we define
kvk2M , v T M v. For matrices M1 , M2 , . . . , Mk , we define
diag (M1 , M2 , . . . , Mk ) as the block diagonal matrix with
diagonal blocks the M1 , M2 , . . . , Mk .
II. S ENSING -C ONSTRAINED LQG C ONTROL
In this section we formalize the sensing-constrained LQG
control problem considered in this paper. We start by introducing the notions of system, sensors, and control policies.
a) System: We consider a standard discrete-time (possibly time-varying) linear system with additive Gaussian noise:
xt+1 = At xt + Bt ut + wt ,
t = 1, 2, . . . , T,
(1)
where xt ∈ Rnt represents the state of the system at time t,
ut ∈ Rmt represents the control action, wt represents the
process noise, and T is a finite time horizon. In addition,
we consider the system’s initial condition x1 to be a Gaussian
random variable with covariance Σ1|0 , and wt to be a Gaussian
random variable with mean zero and covariance Wt , such that
wt is independent of x1 and wt′ for all t′ = 1, 2, . . . , T , t′ 6= t.
b) Sensors: We consider the case where we have a
(potentially large) set of available sensors, which take noisy
linear observations of the system’s state. In particular, let V be
a set of indices such that each index i ∈ V uniquely identifies
a sensor that can be used to observe the state of the system.
We consider sensors of the form
yi,t = Ci,t xt + vi,t ,
i ∈ V,
(2)
where yi,t ∈ Rpi,t represents the measurement of sensor i at
time t, and vi,t represents the measurement noise of sensor i.
We assume vi,t to be a Gaussian random variable with mean
zero and positive definite covariance Vi,t , such that vi,t is
independent of x1 , and of wt′ for any t′ 6= t, and independent
of vi′ ,t′ for all t′ 6= t, and any i′ ∈ V, i′ 6= i.
In this paper we are interested in the case in which we
cannot use all the available sensors, and as a result, we need
to select a convenient subset of sensors in V to maximize our
control performance (formalized in Problem 1 below).
Definition 1 (Active sensor set and measurement model).
Given a set of available sensors V, we say that S ⊂ V is an
active sensor set if we can observe the measurements from each
sensor i ∈ S for all t = 1, 2, . . . , T . Given an active sensor
set S = {i1 , i2 . . . , i|S| }, we define the following quantities
yt (S)
Ct (S)
Vt (S)
,
,
,
[yiT1 ,t , yiT2 ,t , . . . , yiT|S| ,t ]T ,
[CiT1 ,t , CiT2 ,t , . . . , CiT|S| ,t ]T ,
diag[Vi1 ,t , Vi2 ,t , . . . , Vi|S| ,t ]
(3)
which lead to the definition of the measurement model:
yt (S) = Ct (S)xt + vt (S)
(4)
where vt (S) is a zero-mean Gaussian noise with covariance Vt (S). Despite the availability of a possibly large set
of sensors V, our observer will only have access to the
measurements produced by the active sensors.
The following paragraph formalizes how the choice of the
active sensors affects the control policies.
c) Control policies: We consider control policies ut for
all t = 1, 2, . . . , T that are only informed by the measurements
collected by the active sensors:
ut = ut (S) = ut (y1 (S), y2 (S), . . . , yt (S)),
t = 1, 2, . . . , T.
Such policies are called admissible.
In this paper, we want to find a small set of active sensors S,
and admissible controllers u1 (S), u2 (S), . . . , uT (S), to solve
the following sensing-constrained LQG control problem.
Problem 1 (Sensing-constrained LQG control). Find a sensor set S ⊂ V of cardinality at most k to be active across all
times t = 1, 2, . . . , T , and control policies u1:T (S) , {u1 (S),
u2 (S), . . . , uT (S)}, that minimize the LQG cost function:
min
S ⊆ V, |S|≤ k,
u1:T (S)
T
X
t=1
E kxt+1 (S)k2Qt +kut (S)k2Rt ,
(5)
where the state-cost matrices Q1 , Q2 , . . . , QT are positive
semi-definite, the control-cost matrices R1 , R2 , . . . , RT are
positive definite, and the expectation is taken with respect to
the initial condition x1 , the process noises w1 , w2 , . . . , wT ,
and the measurement noises v1 (S), v2 (S), . . . , vT (S).
Problem 1 generalizes the imperfect state-information LQG
control problem from the case where all sensors in V are
active, and only optimal control policies are to be found [19,
Chapter 5], to the case where only a few sensors in V can
be active, and both optimal sensors and control policies are
to be found jointly. While we already noticed that admissible
control policies depend on the active sensor set S, it is worth
noticing that this in turn implies that the state evolution also
depends on S; for this reason we write xt+1 (S) in eq. (5).
The intertwining between control and sensing calls for a joint
design strategy. In the following section we focus on the design
of a jointly optimal control and sensing solution to Problem 1.
III. J OINT S ENSING
AND
C ONTROL D ESIGN
In this section we first present a separation principle that decouples sensing, estimation, and control, and allows designing
them in cascade (Section III-A). We then present a scalable
algorithm for sensing and control design (Section III-B).
A. Separability of Optimal Sensing and Control Design
We characterize the jointly optimal control and sensing
solutions to Problem 1, and prove that they can be found in
two separate steps, where first the sensing design is computed,
and second the corresponding optimal control design is found.
Algorithm 1 Joint Sensing and Control design for Problem 1.
Input: Time horizon T , available sensor set V, covariance
matrix Σ1|0 of initial condition x1 ; for all t = 1, 2, . . . , T ,
system matrix At , input matrix Bt , LQG cost matrices Qt
and Rt , process noise covariance matrix Wt ; and for all
sensors i ∈ V, measurement matrix Ci,t , and measurement
noise covariance matrix Vi,t .
b and control matrices K1 , . . . , KT .
Output: Active sensors S,
b
1: S is returned by Algorithm 2 that finds a (possibly approximate) solution to the optimization problem in eq. (6);
2: K1 , . . . , KT are computed using the recursion in eq. (8).
Theorem 1 (Separability of optimal sensing and control design). Let the sensor set S ⋆ and the controllers u⋆1 , u⋆2 , . . . , u⋆T
be a solution to the sensing-constrained LQG Problem 1. Then,
S ⋆ and u⋆1 , u⋆2 , . . . , u⋆T can be computed in cascade as follows:
S ⋆ ∈ arg min
T
X
S⊆V,|S|≤k t=1
u⋆t = Kt x̂t,S ⋆ ,
tr[Θt Σt|t (S)],
t = 1, . . . , T
(6)
(7)
where x̂t (S) is the Kalman estimator of the state xt ,
i.e., x̂t (S) , E(xt |y1 (S), y2 (S), . . . , yt (S)), and Σt|t (S)
is x̂t (S)’s error covariance, i.e., Σt|t (S) , E[(x̂t (S) −
xt )(x̂t (S) − xt )T ] [19, Appendix E]. In addition, the matrices
Θt and Kt are independent of the selected sensor set S, and
they are computed as follows: the matrices Θt and Kt are the
solution of the backward Riccati recursion
Algorithm 2 Sensing design for Problem 1.
Input: Time horizon T , available sensor set V, covariance
matrix Σ1|0 of system’s initial condition x1 , and for
any time t = 1, 2, . . . , T , any sensor i ∈ V, process
noise covariance matrix Wt , measurement matrix Ci,t , and
measurement noise covariance matrix Vi,t .
b
Output: Sensor set S.
1: Compute Θ1 , Θ2 , . . . , ΘT using recursion in eq. (8);
b ← ∅; i ← 0;
2: S
3: while i < k do
4:
for all a ∈ V \ Sb do
5:
Sba ← Sb ∪ {a}; Σ1|0 (Sba ) ← Σ1|0 ;
6:
for all t = 1, . . . , T do
7:
Σt|t (Sba ) ←
8:
[Σt|t−1 (Sba )−1 + Ct (Sba )T Vt (Sba )−1 Ct (Sba )]−1 ;
9:
Σt+1|t (Sba ) ← At Σt|t (Sba )AT
t + Wt ;
10:
end for P
T
11:
costa ← t=1 tr[Θt Σt|t (Sba )];
12:
end for
13:
ai ← arg mina∈V\S costa ;
14:
Sb ← Sb ∪ {ai }; i ← i + 1;
15: end while
vide more insight on the cost function in (6), we rewrite it as:
T
X
tr[Θt Σt|t (S)] =
t=1
T
X
t=1
=
T
X
t=1
St
Nt
Mt
Kt
Θt
= Qt + Nt+1 ,
−1
= AT
+ Bt Rt−1 BtT )−1 At ,
t (St
T
= Bt St Bt + Rt ,
= −Mt−1 BtT St At ,
= KtT Mt Kt ,
(8)
with boundary condition NT +1 = 0.
Remark 1 (Certainty equivalence principle). The control
gain matrices K1 , K2 , . . . , KT are the same as the ones that
make the controllers (K1 x1 , K1 x2 , . . . , KT xT ) optimal for
the perfect state-information version of Problem 1, where the
state xt is known to the controllers [19, Chapter 4].
Theorem 1 decouples the design of the sensing from the
controller design. Moreover, it suggests that once an optimal
sensor set S ⋆ is found, then the optimal controllers are equal
to Kt x̂t (S), which correspond to the standard LQG control
policy. This should not come as a surprise, since for a given
sensing strategy, Problem 1 reduces to standard LQG control.
We conclude this section with a remark providing a more
intuitive interpretation of the sensor design step in eq. (6).
Remark 2 (Control-aware sensor design). In order to pro-
E tr{[xt − x̂t (S)]T Θt [xt − x̂t (S)]}
E kKt xt − Kt x̂t (S)k2Mt ,
(9)
where
in the first line we used
the fact that Σt|t (S) =
E (xt − x̂t (S))(xt − x̂t (S))T , and in the second line we
substituted the definition of Θt = KtT Mt Kt from eq. (8).
From eq. (9), it is clear that each term tr[Θt Σt|t (S)]
captures the expected control mismatch between the imperfect
state-information controller ut (S) = Kt x̂t (S) (which is only
aware of the measurements from the active sensors) and the
perfect state-information controller Kt xt . This is an important
distinction from the existing sensor selection literature. In particular, while standard sensor selection attempts to minimize
the estimation covariance, for instance by minimizing
T
X
t=1
tr[Σt|t (S)] ,
T
X
t=1
E kxt − x̂t (S)k22 ,
(10)
the proposed LQG cost formulation attempts to minimize the
estimation error of only the informative states to the perfect
state-information controller: for example, the contribution of
all xt − x̂t (S) in the null space of Kt to the total control
mismatch in eq. (9) is zero. Hence, in contrast to minimizing
the cost function in eq. (10), minimizing the cost function in
eq. (9) results to a control-aware sensing design.
B. Scalable Near-optimal Sensing and Control Design
This section proposes a practical design algorithm for
Problem 1. The pseudo-code of the algorithm is presented in
Algorithm 1. Algorithm 1 follows the result of Theorem 1,
and jointly designs sensing and control by first computing an
active sensor set (line 1 in Algorithm 1) and then computing
the control policy (line 2 in Algorithm 1). We discuss each
step of the design process in the rest of this section.
1) Near-optimal Sensing design: The optimal sensor design
can be computed by solving the optimization problem in
eq. (6). The problem is combinatorial in nature, since it
requires to select a subset of elements of cardinality k out
of all the available sensors that induces the smallest cost.
In this section we propose a greedy algorithm, whose
pseudo-code is given in Algorithm 2, that computes a (possibly
approximate) solution to the problem in eq. (6). Our interest
towards this greedy algorithm is motivated by the fact that
it is scalable (in Section IV we show that its complexity is
linear in the number of available sensors) and is provably
close to the optimal solution of the problem in eq. (6)
(we provide suboptimality bounds in Section IV).
Algorithm 2 computes the matrices Θt (t = 1, 2, . . . , T )
which appear in the cost function in eq. (6) (line 1).
Note that these matrices are independent on the choice of
sensors. The set of active sensors Sb is initialized to the
empty set (line 2). The “while loop” in line 3 will be
executed k times and at each time a sensor is greedily
b In particular, the
added to the set of active sensors S.
“for loop” in lines 4-12 computes the estimation covariance
resulting by adding a sensor to the current active sensor
set and the corresponding cost (line 11). Finally, the sensor
inducing the smallest cost is selected (line 13) and added
to the current set of active sensors (line 14).
2) Control policy design: The optimal control design is
computed as in eq. (7), where the control policy matrices
K1 , K2 , . . . , KT are obtained from the recursion in eq. (8).
In the following section we characterize the approximation
and running-time performance of Algorithm 1.
IV. P ERFORMANCE G UARANTEES FOR J OINT S ENSING
AND C ONTROL D ESIGN
We prove that Algorithm 1 is the first scalable algorithm
for the joint sensing and control design Problem 1, and that it
achieves a value for the LQG cost function in eq. (5) that is
finitely close to the optimal. We start by introducing the notion
of supermodularity ratio (Section IV-A), which will enable to
bound the sub-optimality gap of Algorithm 1 (Section IV-B).
A. Supermodularity ratio of monotone functions
We define the supermodularity ratio of monotone functions.
We start with the notions of monotonicity and supermodularity.
Definition 2 (Monotonicity). Consider any finite ground
set V. The set function f : 2V 7→ R is non-increasing if and
only if for any A ⊆ A′ ⊆ V, f (A) ≥ f (A′ ).
Definition 3 (Supermodularity [20, Proposition 2.1]). Consider any finite ground set V. The set function f : 2V 7→ R is
supermodular if and only if for any A ⊆ A′ ⊆ V and x ∈ V,
f (A) − f (A ∪ {x}) ≥ f (A′ ) − f (A′ ∪ {x}).
In words, a set function f is supermodular if and only if it
satisfies the following intuitive diminishing returns property:
for any x ∈ V, the marginal drop f (A) − f (A ∪ {x})
diminishes as A grows; equivalently, for any A ⊆ V and
x ∈ V, the marginal drop f (A)−f (A∪{x}) is non-increasing.
Definition 4 (Supermodularity ratio). Consider any finite
ground set V, and a non-increasing set function f : 2V 7→ R.
We define the supermodularity ratio of f as
γf =
min
A⊆V,x,x′ ∈V\A
f (A) − f (A ∪ {x})
.
f (A ∪ {x′ }) − f [(A ∪ {x′ }) ∪ {x}]
In words, the supermodularity ratio of a monotone set
function f measures how far f is from being supermodular.
In particular, per the Definition 4 of supermodularity ratio, the
supermodularity ratio γf takes values in [0, 1], and
• γf = 1 if and only if f is supermodular, since if γf = 1,
then Definition 4 implies f (A) − f (A ∪ {x}) ≥ f (A ∪
{x′ }) − f [(A ∪ {x′ }) ∪ {x}], i.e., the drop f (A) − f (A ∪
{x}) is non-increasing as new elements are added in A.
• γf < 1 if and only if f is approximately supermodular, in
the sense that if γf < 1, then Definition 4 implies f (A)−
f (A ∪ {x}) ≥ γf {f (A ∪ {x′ }) − f [(A ∪ {x′ }) ∪ {x}]},
i.e., the drop f (A) − f (A ∪ {x}) is approximately nonincreasing as new elements are added in A; specifically,
the supermodularity ratio γf captures how much ones
needs to discount the drop f (A ∪ {x′ }) − f [(A ∪ {x′ }) ∪
{x}], such that f (A) − f (A ∪ {x}) remains greater then,
or equal to, f (A ∪ {x′ }) − f [(A ∪ {x′ }) ∪ {x}].
We next use the notion of supermodularity ratio Definition 4
to quantify the sub-optimality gap of Algorithm 1.
B. Performance Analysis for Algorithm 1
We quantify Algorithm 1’s running time, as well as, Algorithm 1’s approximation performance, using the notion of
supermodularity ratio introduced in Section IV-A. We conclude the section by showing that for appropriate LQG cost
matrices Q1 , Q2 , . . . , QT and R1 , R2 , . . . , RT , Algorithm 1
achieves near-optimal approximate performance.
Theorem 2 (Performance of Algorithm 1). For any active
sensor set S ⊆ V, and admissible control policies u1:T (S) ,
{u1 (S), u2 (S), . . . , uT (S)}, let h[S, u1:T (S)] be Problem 1’s
cost function, i.e.,
PT
h[S, u1:T (S)] , t=1 E(kxt+1 (S)k2Qt +kut(S)k2Rt );
Further define the following set-valued function and scalar:
g(S) , minu1:T (S) h[S, u1:T (S)],
g ⋆ , minS⊆V,|S|≤k, h[S, u1:T (S)].
u1:T (S)
The following results hold true:
(11)
1) (Approximation quality) Algorithm 1 returns an active
sensor set Sb ⊂ V of cardinality k, and gain matrices K1 ,
b u1:T (S)]
b attained by
K2 , . . . , KT , such that the cost h[S,
b
the sensor set S and the corresponding control policies
b , {K1 x̂1 (S),
b . . . , KT x̂T (S)}
b satisfies
u1:T (S)
b u1:T (S))
b − g⋆
h(S,
≤ exp(−γg )
g(∅) − g ⋆
T
= 1, and that
norm of each C̄i,t is 1, i.e., tr C̄i,t C̄i,t
tr[Σt|t (∅)] ≤ λ2max [Σt|t (∅)], γg ’s lower bound is
γg ≥
(12)
where γg is the supermodularity ratio of g(S) in eq. (11).
2) (Running time) Algorithm 1 runs in O(k|V|T n2.4 ) time,
where n , maxt=1,2,...,T (nt ) is the maximum system size
in eq. (1).
Theorem 2 ensures that Algorithm 1 is the first scalable
algorithm for the sensing-constrained LQG control Problem 1.
In particular, Algorithm 1’s running time O(k|V|T n2.4 ) is linear both in the number of available sensors |V|, and the sensor
set cardinality constraint k, as well as, linear in the Kalman
filter’s running time across the time horizon {1, 2 . . . , T }.
Specifically, the contribution n2.4 T in Algorithm 1’s running
time comes from the computational complexity of using the
Kalman filter to compute the state estimation error covariances
Σt|t for each t = 1, 2, . . . , T [19, Appendix E].
Theorem 2 also guarantees that for non-zero ratio γg
Algorithm 1 achieves a value for Problem 1 that is finitely
close to the optimal. In particular, the bound in ineq. (12)
improves as γg increases, since it is decreasing in γg , and
is characterized by the following extreme behaviors: for
γg = 1, the bound in ineq. (12) is e−1 ≃ .37, which
is the minimum for any γg ∈ [0, 1], and hence, the best
bound on Algorithm 1’s approximation performance among
all γg ∈ [0, 1] (ideally, the bound in ineq. (12) would be 0
for γg = 1, in which case Algorithm 1 would be exact,
b u1:T (S))
b = g ⋆ ; however,
since it would be implied h(S,
even for supermodular functions, the best bound one can
achieve in the worst-case is e−1 [21]); for γg = 0, ineq. (12)
b u1:T (S))
b ≤ g(∅) =
is uninformative since it simplifies to h(S,
h(∅, u1:T (∅)), which is trivially satisfied.1
In the remaining
PT of the section, we first prove that if the
strict inequality t=1 Θt ≻ 0 holds, where each Θt is defined
as in eq. (8), then the ratio γg in ineq. (12) is non-zero, and as
result Algorithm 1 achieves a near-optimal approximation performance
(Theorem 3). Then, we prove that the strict inequalP
ity Tt=1 Θt ≻ 0 is equivalent to a verifiable condition (Proposition 1) that holds for appropriate cost matrices Qt and Rt .
Theorem 3 (Lower bound for supermodularity ratio γg ).
Let Θt for all t = 1, 2, . . . , T be defined as in eq. (8), g(S)
be defined as in eq. (11), and for any sensor i ∈ V, C̄i,t be
−1/2
the normalized measurement matrix Vi,t Ci,t .
PT
If t=1 Θt ≻ 0, the supermodularity ratio γg is non-zero.
In addition, if we consider for simplicity that the Frobenius
1 The inequality h(S,
b u1:T (S))
b ≤ h(∅, u1:T (∅)) simply states that a control policy that is informed by the active sensor set S has better performance
than a policy that does not use any sensor; for a more formal proof we refer
the reader to Appendix B.
P
λmin ( Tt=1 Θt ) mint∈{1,2,...,T } λ2min [Σt|t (V)]
PT
λmax ( t=1 Θt ) maxt∈{1,2,...,T } λ2max [Σt|t (∅)]
1 + mini∈V,t∈{1,2...,T } λmin [C̄i Σt|t (V)C̄iT ]
.
2 + maxi∈V,t∈{1,2...,T } λmax [C̄i Σt|t (∅)C̄iT ]
(13)
The supermodularity ratio bound in ineq. (13) suggests two
cases under which γg can increase, and correspondingly, the
performance bound of Algorithm 1 in eq. (12) can improve:
a) Case 1 where γg ’s bound in ineq. (13) increases:
P
P
When the fraction λmin ( Tt=1 Θt )/λmax ( Tt=1 Θt ) increases
to 1, then the right-hand-side in ineq. (13) increases. Equivalently, the right-hand-side in ineq. (13) increases when on
(i)
(i)
average all the directions xt − x̂t of the estimation errors
(1)
(1)
(2)
(2)
(n )
(n )
xt − x̂t = (xt − x̂t , xt − x̂t , . . . , xt t − x̂t t ) become
equally important in selecting the active sensor set. To see this,
consider for example that λmax (Θt ) = λmin (Θt ) = λ; then,
the cost function in eq. (6) that Algorithm 1 minimizes to
select the active sensor set becomes
T
X
tr[Θt Σt|t (S)] = λ
t=1
=λ
T
X
E tr(kxt − x̂t (S)k22 )
t=1
nt
T X
X
t=1 i=1
i
h
(i)
(i)
E tr(kxt − x̂t (S)|22 ) .
Overall, it is easier for Algorithm 1 to approximate a solution
to Problem 1 as the cost function in eq. (6) becomes the cost
function in the standard sensor selection problems where one
minimizes the total estimation covariance as in eq. (10).
b) Case 2 where γg ’s bound in ineq. (13) increases:
When either the numerators of the last two fractions in the
right-hand-side of ineq. (13) increase or the denominators
of the last two fractions in the right-hand-side of ineq. (13)
decrease, then the right-hand-side in ineq. (13) increases.
In particular, the numerators of the last two fractions in righthand-side of ineq. (13) capture the estimation quality when
all available sensors in V are used, via the terms of the
T
form λmin [Σt|t (V)] and λmin [C̄i,t Σt|t (V)C̄i,t
]. Interestingly,
this suggests that the right-hand-side of ineq. (13) increases
when the available sensors in V are inefficient in achieving
low estimation error, that is, when the terms of the form
T
λmin [Σt|t (V)] and λmin [C̄i,t Σt|t (V)C̄i,t
] increase. Similarly,
the denominators of the last two fractions in right-handside of ineq. (13) capture the estimation quality when no
sensors are used, via the terms of the form λmax [Σt|t (∅)] and
T
λmax [C̄i,t Σt|t (∅)C̄i,t
]. This suggests that the right-hand-side of
ineq. (13) increases when the measurement noise increases.
We next give a P
control-level equivalent condition to TheoT
rem 3’s condition t=1 Θt ≻ 0 for non-zero ratio γg .
Theorem 4 (Control-level condition for near-optimal sensor
selection). Consider the LQG problem where for any time t =
1, 2, . . . , T , the state xt is known to each controller ut and
9
the process noise wt is zero, i.e., the optimization problem
PT
(14)
minu1:T t=1 [kxt+1 k2Qt +kut (xt )k2Rt ] Σ =W =0 .
7
6
5
4
3
t
Let At to
PTbe invertible for all t = 1, 2, . . . , T ; the strict
inequality t=1 Θt ≻ 0 holds if and only if for all non-zero
initial conditions x1 ,
PT
0∈
/ arg minu1:T t=1 [kxt+1 k2Qt +kut (xt )k2Rt ] Σ =W =0 .
t|t
t
4 suggests that Theorem 3’s sufficient condition
PTheorem
T
Θ
≻
0 for non-zero ratio γg holds if and only if for
t
t=1
any non-zero initial condition x1 the all-zeroes control policy
u1:T = (0, 0, . . . , 0) is suboptimal for the noiseless perfect
state-information LQG problem in eq. (14).
The all-zeroes control policy is always suboptimal for the
noiseless perfect state-information LQG problem in eq. (14)
for appropriate LQG cost matrices Q1 , Q2 , . . . , QT and
R1 , R2 , . . . , RT , as implied by Proposition 1 below.
Proposition 1 (System-level condition for near-optimal
sensor selection). Let N1 be defined as in eq. (8). The control
policy u1:T , (0, 0, . . . , 0) is suboptimal for the LQG problem
in eq. (14) for all non-zero initial conditions x1 if and only if
PT
T
T
(15)
t=1 A1 · · · At Qt At · · · A1 ≻ N1 .
The strict ineq. (15) holds for appropriate LQG cost matrices Qt and Rt , as, for example, the control cost matrices
Rt are closer to zero than the state cost matrices Qt , that
is, as the LQG cost function penalizes the state vector values
more than the control vector values. For example, consider
the case where the LQG cost matrix R1 tends to zero, i.e.,
consider ǫ such that R1 ǫI, and ǫ −→ 0; then, for
any Q1 , the strict ineq. (15) holds for ǫ small, since N1
−1
T −1
ǫAT
A1 −→ 0 for ǫ −→ 0. In particular,
1 (ǫS1 + B1 B1 )
R1 ǫI, and [22, Proposition 8.5.5],2 imply R1−1 I/ǫ, and
as a result, S1−1 + B1 R1−1 B1T (ǫS1−1 + B1 B1T )/ǫ, which
in turn implies (S1−1 + B1 R1−1 B1T )−1 ǫ(ǫS1−1 + B1 B1T )−1 ,
−1
T −1
and consequently, N1 ǫAT
A1 .
1 (ǫS1 + B1 B1 )
Overall, Algorithm 1 is the first scalable algorithm for the
sensing-constrained LQG control Problem 1, and for appropriate LQG cost matrices Q1 , Q2 , . . . , QT and R1 , R2 , . . . , RT ,
Algorithm 1 achieves near-optimal approximate performance.
V. N UMERICAL E XPERIMENTS
We consider two application scenarios for the proposed sensing-constrained LQG control framework: sensingconstrained formation control and resource-constrained robot
navigation. We present a Monte Carlo analysis for both scenarios, which demonstrates that (i) the proposed sensor selection
strategy is near-optimal, and in particular, the resulting LQGcost (tracking performance) matches the optimal selection in
all tested instances for which the optimal selection could
be computed via a brute-force approach, (ii) a more naive
selection which attempts to minimize the state estimation
2 [22, Proposition 8.5.5] states that for any positive definite matrices M
1
and M2 such that M1 M2 , M2−1 M1−1 .
2
y [meters]
t|t
8
1
0
-1
-2
-3
-4
-5
-6
-7
-8
-9
-9
-8
-7
-6
-5
-4
-3
-2
-1
0
1
2
3
4
5
6
7
8
9
x [meters]
(a) formation control
(b) unmanned aerial robot
Fig. 1. Examples of applications of the proposed sensing-constrained LQG
control framework: (a) sensing-constrained formation control and (b) resourceconstrained robot navigation.
covariance [15] (rather than the LQG cost) has degraded LQG
tracking performance, often comparable to a random selection,
(iii) in the considered instances, a clever selection of a small
subset of sensors can ensure an LQG cost that is close to the
one obtained by using all available sensors, hence providing an
effective alternative for control under sensing constraints [23].
VI. S ENSING - CONSTRAINED
FORMATION CONTROL
Simulation setup. The first application scenario is illustrated in Fig. 1(a). A team of n agents (blue triangles) moves
in a 2D scenario. At time t = 1, the agents are randomly
deployed in a 10m × 10m square and their objective is to
reach a target formation shape (red stars); in the example of
Fig. 1(a) the desired formation has an hexagonal shape, while
in general for a formation of n, the desired formation is an
equilateral polygon with n vertices. Each robot is modeled as
a double-integrator, with state xi = [pi vi ]T ∈ R4 (pi is the 2D
position of agent i, while vi is its velocity), and can control
its own acceleration ui ∈ R2 ; the process noise is chosen
as a diagonal matrix W = diag [1e−2 , 1e−2 , 1e−4 , 1e−4 ] .
Each robot i is equipped with a GPS receiver, which can
measure the agent position pi with a covariance Vgps,i = 2·I2 .
Moreover, the agents are equipped with lidar sensors allowing
each agent i to measure the relative position of another agent j
with covariance Vlidar,ij = 0.1 · I2 . The agents have very
limited on-board resources, hence they can only activate a
subset of k sensors. Hence, the goal is to select the subset of
k sensors, as well as to compute the control policy that ensure
best tracking performance, as measured by the LQG objective.
For our tests, we consider two problem setups. In the first
setup, named homogeneous formation control, the LQG weigh
matrix Q is a block diagonal matrix with 4 × 4 blocks, with
each block i chosen as Qi = 0.1 · I4 ; since each 4 × 4 block
of Q weights the tracking error of a robot, in the homogeneous
case the tracking error of all agents is equally important.
In the second setup, named heterogeneous formation control,
the matrix Q is chose as above, except for one of the agents,
say robot 1, for which we choose Q1 = 10 · I4 ; this setup
models the case in which each agent has a different role or
importance, hence one weights differently the tracking error
of the agents. In both cases the matrix R is chosen to be the
identity matrix. The simulation is carried on over T time steps,
and T is also chosen as LQG horizon. Results are averaged
over 100 Monte Carlo runs: at each run we randomize the
initial estimation covariance Σ1|0 .
Compared techniques. We compare five techniques. All
techniques use an LQG-based estimator and controller, and
they only differ by the selections of the sensors used.
The first approach is the optimal sensor selection, denoted
as optimal, which attains the minimum of the cost function
in eq. (6), and that we compute by enumerating all possible
subsets; this brute-force approach is only viable when the
number of available sensors is small. The second approach
is a pseudo-random sensor selection, denoted as random∗ ,
which selects all the GPS measurements and a random subset
of the lidar measurements; note that we do not consider a
fully random selection since in practice this often leads to an
unobservable system, hence causing divergence of the LQG
cost. The third approach, denoted as logdet, selects sensors
so to minimize the average log det of the estimation covariance
over the horizon; this approach resembles [15] and is agnostic
to the control task. The fourth approach is the proposed sensor
selection strategy, described in Algorithm 2, and is denoted as
s-LQG. Finally, we also report the LQG performance when all
sensors are selected; this is clearly infeasible in practice, due to
the sensing constraints, and it is only reported for comparison
purposes. This approach is denoted as allSensors.
Results. The results of our numerical analysis are reported
in Fig. 2. When not specified otherwise, we consider a
formation of n = 4 agents, which can only use a total of
k = 6 sensors, and a control horizon T = 20. Fig. 2(a)
shows the LQG cost attained by the compared techniques for
increasing control horizon and for the homogeneous case. We
note that, in all tested instance, the proposed approach s-LQG
matches the optimal selection optimal, and both approaches
are relatively close to allSensors, which selects all the
2
available sensors ( n+n
2 ). On the other hand logdet leads
to worse tracking performance, and it is often close to the
pseudo-random selection random∗ . These considerations are
confirmed by the heterogeneous setup, shown in Fig. 2(b).
In this case the separation between the proposed approach
and logdet becomes even larger; the intuition here is that
the heterogeneous case rewards differently the tracking errors
at different agents, hence while logdet attempts to equally
reduce the estimation error across the formation, the proposed
approach s-LQG selects sensors in a task-oriented fashion,
since the matrices Θt for all t = 1, 2, . . . , T in the cost
function in eq. (6) incorporate the LQG weight matrices.
Fig. 2(c) shows the LQG cost attained by the compared
techniques for increasing number of selected sensors k and for
the homogeneous case. We note that for increasing number of
sensors all techniques converge to allSensors (the entire
ground set is selected). As in the previous case, the proposed
approach s-LQG matches the optimal selection optimal.
Interestingly, while the performance of logdet is in general
inferior with respect to s-LQG, when the number of selected
sensors k decreases (for k < n the problem becomes unob-
servable) the approach logdet performs similarly to s-LQG.
Fig. 2(d) shows the same statistics for the heterogeneous case.
We note that in this case logdet is inferior to s-LQG even
in the case with small k. Moreover, an interesting fact is that
s-LQG matches allSensors already for k = 7, meaning
that the LQG performance of the sensing-constraint setup is
indistinguishable from the one using all sensors; intuitively,
in the heterogeneous case, adding more sensors may have
marginal impact on the LQG cost (e.g., if the cost rewards
a small tracking error for robot 1, it may be of little value
to take a lidar measurement between robot 3 and 4). This
further stresses the importance of the proposed framework as a
parsimonious way to control a system with minimal resources.
Fig. 2(e) and Fig. 2(f) show the LQG cost attained by the
compared techniques for increasing number of agents, in the
homogeneous and heterogeneous case, respectively. To ensure
observability, we consider k = round (1.5n), i.e., we select a
number of sensors 50% larger than the smallest set of sensors
that can make the system observable. We note that optimal
quickly becomes intractable to compute, hence we omit values
beyond n = 4. In both figures, the main observation is that
the separation among the techniques increases with the number
of agents, since the set of available sensors quickly increases
with n. Interestingly, in the heterogeneous case s-LQG remains relatively close to allSensors, implying that for the
purpose of LQG control, using a cleverly selected small subset
of sensors still ensures excellent tracking performance.
VII. R ESOURCE - CONSTRAINED
ROBOT NAVIGATION
Simulation setup. The second application scenario is illustrated in Fig. 1(b). An unmanned aerial robot (UAV) moves
in a 3D scenario, starting from a randomly selected initial
location. The objective of the UAV is to land, and more
specifically, it has to reach the position [0, 0, 0] with zero
velocity. The UAV is modeled as a double-integrator, with
state xi = [pi vi ]T ∈ R6 (pi is the 3D position of agent i,
while vi is its velocity), and can control its own acceleration
ui ∈ R3 ; the process noise is chosen as W = I6 . The UAV
is equipped with multiple sensors. It has an on-board GPS
receiver, measuring the UAV position pi with a covariance
2 · I3 , and an altimeter, measuring only the last component
of pi (altitude) with standard deviation 0.5m. Moreover, the
UAV can use a stereo camera to measure the relative position
of ℓ landmarks on the ground; for the sake of the numerical
example, we assume the location of each landmark to be
known only approximately, and we associate to each landmark
an uncertainty covariance (red ellipsoids in Fig. 1(b)), which is
randomly generated at the beginning of each run. The UAV has
limited on-board resources, hence it can only activate a subset
of k sensors. For instance, the resource-constraints may be due
to the power consumption of the GPS and the altimeter, or
may be due to computational constraints that prevent to run
multiple object-detection algorithms to detect all landmarks
on the ground. Similarly to the previous case, we phrase the
problem as a sensing-constraint LQG problem,
and we use
Q = diag [1e−3 , 1e−3 , 10, 1e−3 , 1e−3 , 10] and R = I3 .
12
random*
optimal
logdet
s-LQG
allSensors
6
2450
150
120
random*
optimal
logdet
s-LQG
allSensors
110
100
100
LQG cost
200
random*
optimal
logdet
s-LQG
allSensors
130
LQG cost / T
8
LQG cost
10
LQG cost
2500
250
random*
optimal
logdet
s-LQG
allSensors
4
2400
2350
2300
90
2
50
10
15
20
25
30
10
15
20
30
300
12
10
random*
optimal
logdet
s-LQG
allSensors
250
LQG cost
random*
optimal
logdet
s-LQG
allSensors
14
200
150
8
100
6
4
50
4
5
6
7
8
9
10
4
6
7
8
9
10
(d) heterogeneous
(c) homogeneous
25
180
random*
optimal
logdet
s-LQG
allSensors
15
random*
optimal
logdet
s-LQG
allSensors
160
140
LQG cost
20
LQG cost
5
maxNrUsedSensors
maxNrUsedSensors
10
120
0
100
60
40
3
5
7
nrRobots
(e) homogeneous
9
11
3
5
7
20
30
40
50
2250
4
horizon
6
8
10
maxNrUsedSensors
(b) heterogeneous
Fig. 3. LQG cost for increasing (a)-(b) control horizon T , (c)-(d) number of
selected sensors k, and (e)-(f) number of agents n. Statistics are reported for
the homogeneous formation control setup (left column), and the heterogeneous
setup (right column). Results are averaged over 100 Monte Carlo runs.
optimal selection optimal, while logdet and random∗
have suboptimal performance.
Fig. 3(b) shows the LQG cost attained by the compared
techniques for increasing number of selected sensors k.
Clearly, all techniques converge to allSensors for increasing k, but in the regime in which few sensors are used s-LQG
still outperforms alternative sensor selection schemes, and
matches in all cases the optimal selection optimal.
VIII. C ONCLUDING R EMARKS
80
5
10
(a) homogeneous
(b) heterogeneous
(a) homogeneous
16
LQG cost
25
horizon
horizon
9
11
nrRobots
(f) heterogeneous
Fig. 2. LQG cost for increasing (a)-(b) control horizon T , (c)-(d) number of
selected sensors k, and (e)-(f) number of agents n. Statistics are reported for
the homogeneous formation control setup (left column), and the heterogeneous
setup (right column). Results are averaged over 100 Monte Carlo runs.
Note that the structure of Q reflects the fact that during
landing we are particularly interested in controlling the vertical
direction and the vertical velocity (entries with larger weight
in Q), while we are less interested in controlling accurately the
horizontal position and velocity (assuming a sufficiently large
landing site). In the following, we present results averaged
over 100 Monte Carlo runs: in each run, we randomize the
covariances describing the landmark position uncertainty.
Compared techniques. We consider the five techniques
discussed in the previous section. As in the formation control
case, the pseudo-random selection random∗ always includes
the GPS measurement (which alone ensures observability) and
a random selection of the other available sensors.
Results. The results of our numerical analysis are reported
in Fig. 3. When not specified otherwise, we consider a total of
k = 3 sensors to be selected, and a control horizon T = 20.
Fig. 3(a) shows the LQG cost attained by the compared
techniques for increasing control horizon. For visualization
purposes we plot the cost normalized by the horizon, which
makes more visible the differences among the techniques. Similarly to the formation control example, s-LQG matches the
In this paper, we introduced the sensing-constrained LQG
control Problem 1, which is central in modern control applications that range from large-scale networked systems to
miniaturized robotics networks. While the computation of
the optimal sensing strategy is intractable, We provided the
first scalable algorithm for Problem 1, Algorithm 1, and
under mild conditions on the system and LQG matrices,
proved that Algorithm 1 computes a near-optimal sensing
strategy with provable sub-optimality guarantees. To this end,
we showed that a separation principle holds, which allows the
design of sensing, estimation, and control policies in isolation.
We motivated the importance of the sensing-constrained LQG
Problem 1, and demonstrated the effectiveness of Algorithm 1,
by considering two application scenarios: sensing-constrained
formation control, and resource-constrained robot navigation.
R EFERENCES
[1] N. Elia and S. Mitter, “Stabilization of linear systems with limited
information,” IEEE Trans. on Automatic Control, vol. 46, no. 9, pp.
1384–1400, 2001.
[2] G. Nair and R. Evans, “Stabilizability of stochastic linear systems with
finite feedback data rates,” SIAM Journal on Control and Optimization,
vol. 43, no. 2, pp. 413–436, 2004.
[3] S. Tatikonda and S. Mitter, “Control under communication constraints,”
IEEE Trans. on Automatic Control, vol. 49, no. 7, pp. 1056–1068, 2004.
[4] V. Borkar and S. Mitter, “LQG control with communication constraints,”
Comm., Comp., Control, and Signal Processing, pp. 365–373, 1997.
[5] J. L. Ny and G. Pappas, “Differentially private filtering,” IEEE Trans.
on Automatic Control, vol. 59, no. 2, pp. 341–354, 2014.
[6] G. Nair, F. Fagnani, S. Zampieri, and R. Evans, “Feedback control under
data rate constraints: An overview,” Proceedings of the IEEE, vol. 95,
no. 1, pp. 108–137, 2007.
[7] J. Hespanha, P. Naghshtabrizi, and Y.Xu, “A survey of results in networked control systems,” Pr. of the IEEE, vol. 95, no. 1, p. 138, 2007.
[8] J. Baillieul and P. Antsaklis, “Control and communication challenges in
networked real-time systems,” Proceedings of the IEEE, vol. 95, no. 1,
pp. 9–28, 2007.
12
[9] V. Gupta, T. H. Chung, B. Hassibi, and R. M. Murray, “On a stochastic
sensor selection algorithm with applications in sensor scheduling and
sensor coverage,” Automatica, vol. 42, no. 2, pp. 251–260, 2006.
[10] H. Avron and C. Boutsidis, “Faster subset selection for matrices and
applications,” SIAM Journal on Matrix Analysis and Applications,
vol. 34, no. 4, pp. 1464–1499, 2013.
[11] C. Li, S. Jegelka, and S. Sra, “Polynomial Time Algorithms for Dual
Volume Sampling,” ArXiv e-prints: 1703.02674, 2017.
[12] S. Joshi and S. Boyd, “Sensor selection via convex optimization,” IEEE
Transactions on Signal Processing, vol. 57, no. 2, pp. 451–462, 2009.
[13] J. L. Ny, E. Feron, and M. A. Dahleh, “Scheduling continuous-time
kalman filters,” IEEE Trans. on Aut. Control, vol. 56, no. 6, pp. 1381–
1394, 2011.
[14] M. Shamaiah, S. Banerjee, and H. Vikalo, “Greedy sensor selection:
Leveraging submodularity,” in Proceedings of the 49th IEEE Conference
on Decision and control, 2010, pp. 2572–2577.
[15] S. T. Jawaid and S. L. Smith, “Submodularity and greedy algorithms in
sensor scheduling for linear dynamical systems,” Automatica, vol. 61,
pp. 282–288, 2015.
[16] V. Tzoumas, A. Jadbabaie, and G. J. Pappas, “Sensor placement for
optimal Kalman filtering,” in Amer. Contr. Conf., 2016, pp. 191–196.
[17] E. Shafieepoorfard and M. Raginsky, “Rational inattention in scalar LQG
control,” in Proceedings of the 52th IEEE Conference in Decision and
Control, 2013, pp. 5733–5739.
[18] T. Tanaka and H. Sandberg, “SDP-based joint sensor and controller
design for information-regularized optimal LQG control,” in 54th IEEE
Conference on Decision and Control, 2015, pp. 4486–4491.
[19] D. P. Bertsekas, Dynamic programming and optimal control, Vol. I.
Athena Scientific, 2005.
[20] G. Nemhauser, L. Wolsey, and M. Fisher, “An analysis of approximations
for maximizing submodular set functions – I,” Mathematical Programming, vol. 14, no. 1, pp. 265–294, 1978.
[21] U. Feige, “A threshold of ln(n) for approximating set cover,” Journal
of the ACM, vol. 45, no. 4, pp. 634–652, 1998.
[22] D. S. Bernstein, Matrix mathematics. Princeton University Press, 2005.
[23] L. Carlone and S. Karaman, “Attention and anticipation in fast visualinertial navigation,” in Proceeding of the IEEE International Conference
on Robotics and Automation, 2017, pp. 3886–3893.
[24] P. R. Kumar and P. Varaiya, Stochastic systems: Estimation, identification and adaptive control. Prentice-Hall, Inc., 1986.
[25] Z. Wang, B. Moran, X. Wang, and Q. Pan, “Approximation for maximizing monotone non-decreasing set functions with a greedy method,”
Journal of Combinatorial Optimization, vol. 31, no. 1, pp. 29–43, 2016.
[26] L. F. Chamon and A. Ribeiro, “Near-optimality of greedy set selection
in the sampling of graph signals,” in Proceedings of the IEEE Global
Conference on Signal and Information Processing, 2016, pp. 1265–1269.
[27] D. Coppersmith and S. Winograd, “Matrix multiplication via arithmetic
progressions,” J. of Symbolic Comp., vol. 9, no. 3, pp. 251–280, 1990.
A PPENDIX A: P RELIMINARY
FACTS
This appendix contains a set of lemmata that will be used
to support the proofs in this paper (Appendices B–F).
Lemma 1 ([22, Proposition 8.5.5]). Consider two positive
definite matrices M1 and M2 . If M1 M2 then M2−1 M1−1.
Lemma 2 (Trace inequality [22, Proposition 8.4.13]). Consider a symmetric matrix A, and a positive semi-definite matrix
B of appropriate dimension. Then,
λmin (A)tr (B) ≤ tr (AB) ≤ λmax (A)tr (B) .
Lemma 3 (Woodbury identity [22, Corollary 2.8.8]). Consider matrices A, C, U and V of appropriate dimensions, such
that A, C, and A + U CV are invertible. Then,
(A + U CV )−1 = A−1 − A−1 U (C −1 + V A−1 U )−1 V A−1.
Lemma 4 ([22, Proposition 8.5.12]). Consider two symmetric
matrices A1 and A2 , and a positive semi-definite matrix B.
If A1 A2 , then tr (A1 B) ≤ tr (A2 B).
Lemma 5 ([19, Appendix E]). For any sensor set S ⊆ V,
and for all t = 1, 2, . . . , T , let x̂t (S) be the Kalman estimator
of the state xt , i.e., x̂t (S), and Σt|t (S) be x̂t (S)’s error
covariance, i.e., Σt|t (S) , E[(x̂t (S) − xt )(x̂t (S) − xt )T ].
Then, Σt|t (S) is the solution of the Kalman filtering recursion
Σt|t (S) = [Σt|t−1 (S)−1 + Ct (S)T Vt (S)−1 Ct (S)]−1,
Σt+1|t (S) = At Σt|t (S)AT
t + Wt ,
(16)
with boundary condition the Σ1|0 (S) = Σ1|0 .
Lemma 6. For any sensor set S ⊆ V, let Σ1|1 (S) be defined as
in eq. (16), and consider two sensor sets S1 , S2 ⊆ V. If S1 ⊆
S2 , then Σ1|1 (S1 ) Σ1|1 (S2 ).
Proof of Lemma 6: Let D = S2 \ S1 , and observe that
for all t = 1, 2, . . . , T , the notation in Definition 1 implies
X
T
Ct (S2 )T Vt (S2 )−1 Ct (S2 ) =
Ci,t
Vi,t Ci,t
i∈S2
=
X
T
Ci,t
Vi,t Ci,t +
i∈S1
=
X
X
T
Ci,t
Vi,t Ci,t
i∈D
T
Ci,t
Vi,t Ci,t
i∈S1
Ct (S1 )T Vt (S1 )−1 Ct (S1 ).
(17)
Therefore, Lemma 1 and ineq. (17) imply
T
−1
Σ1|1 (S2 ) = [Σ−1
Ct (S2 )]−1
1|0 + C1 (S2 ) Vt (S2 )
T
−1
[Σ−1
Ct (S1 )]−1 = Σ1|1 (S1 ).
1|0 + C1 (S1 ) Vt (S1 )
Lemma 7. Let Σt|t be defined as in eq. (16) with boundary
condition the Σ1|0 ; similarly, let Σ̄t|t be defined as in eq. (16)
with boundary condition the Σ̄1|0 . If Σt|t Σ̄t|t , then
Σt+1|t Σ̄t+1|t .
Proof of Lemma 7: We complete the proof in two
steps: first, from eq. (16), it its Σt+1|t = At Σt|t AT
t + Wt
+
W
=
Σ̄
.
Then,
from
Σ
Σ̄
,
At Σ̄t|t AT
t
t+1|t
t|t
t|t it follows
t
T
Σ̄
A
.
At Σt|t AT
A
t t|t t
t
Lemma 8. Let Σt|t−1 be defined as in eq. (16) with boundary
condition the Σ1|0 ; similarly, let Σ̄t|t−1 be defined as in
eq. (16) with boundary condition the Σ̄1|0 . If Σt|t−1 Σ̄t|t−1 ,
then Σt|t Σ̄t|t .
Proof of Lemma 8: From eq. (16), it is Σt|t =
T −1
−1
T −1
−1
(Σ−1
(Σ̄−1
= Σ̄t|t ,
t|t−1 + Ct Vt Ct )
t|t−1 + Ct Vt Ct )
since Lemma 1 and the condition Σt|t−1 Σ̄t|t−1 imply
−1
T −1
T −1
Σ−1
t|t−1 + Ct Vt Ct Σ̄t|t−1 + Ct Vt Ct , which in turn
−1
T −1
−1
T −1
−1
implies (Σt|t−1 + Ct Vt Ct )
(Σ̄−1
t|t−1 + Ct Vt Ct ) .
Corollary 1. Let Σt|t be defined as in eq. (16) with boundary
condition the Σ1|0 ; similarly, let Σ̄t|t be defined as in eq. (16)
with boundary condition the Σ̄1|0 . If Σt|t Σ̄t|t , then
Σt+i|t+i Σ̄t+i|t+i for any positive integer i.
Proof of Corollary 1: If Σt|t Σ̄t|t , from Lemma 7,
we get Σt+1|t Σ̄t+1|t , which, from Lemma 8, implies
Σt+1|t+1 Σ̄t+1|t+1 . By repeating the previous argument
another (i − 1) times, the proof is complete.
Corollary 2. Let Σt|t be defined as in eq. (16) with boundary
condition the Σ1|0 ; similarly, let Σ̄t|t be defined as in eq. (16)
with boundary condition the Σ̄1|0 . If Σt|t Σ̄t|t , then
Σt+i|t+i−1 Σ̄t+i|t+i−1 for any positive integer i.
Proof of Corollary 2: If Σt|t Σ̄t|t , from Corollary 1,
we get Σt+i−1|t+i−1 Σ̄t+i−1|t+i−1 , which, from Lemma 7,
implies Σt+i|t+i−1 Σ̄t+i|t+i−1 .
A PPENDIX B: P ROOF
OF
T HEOREM 1
The proof of Theorem 1 follows from the following lemma.
Lemma 9. For any active sensor set S ⊆ V, and admissible
control policies u1:T (S) , {u1 (S), u2 (S), . . . , uT (S)}, let
h[S, u1:T (S)] be Problem 1’s cost function, i.e.,
PT
h[S, u1:T (S)] , t=1 E(kxt+1 (S)k2Qt +kut (S)k2Rt );
Further define the following set-valued function:
g(S) , minu1:T (S) h[S, u1:T (S)],
Consider any sensor set S ⊆ V, and let u⋆1:T,S be the vector of control policies (K1 x̂1,S , K2 x̂2,S , . . . , KT x̂T,S ). Then
u⋆1:T,S is an optimal control policy:
u⋆1:T,S ∈ arg min h[S, u1:T (S)],
(18)
u1:T (S)
i.e., g(S) = h[S, u⋆1:T (S)], and in particular, u⋆1:T,S attains a
(sensor-dependent) LQG cost equal to:
g(S) = E(kx1 kN1 )+
T
X
tr[Θt Σt|t (S)] + tr (Wt St ) . (19)
t=1
Proof of Lemma 9: Let ht [S, ut:T (S)] be the LQG cost
in Problem 1 from time t up to time T , i.e.,
ht [S, ut:T (S)] ,
T
X
E(kxk+1 (S)k2Qt +kuk (S)k2Rt ).
k=t
and define gt (S) , minut:T (S) ht [S, ut:T (S)]. Clearly, g1 (S)
matches the LQG cost in eq. (19).
We complete the proof inductively. In particular, we first
prove Lemma 9 for t = T , and then for any other t ∈
{1, 2, . . . , T − 1}. To this end, we use the following observation: given any sensor set S, and any time t ∈ {1, 2, . . . , T },
gt (S) = min E(kxt+1 (S)k2Qt +kut (S)k2Rt ) + gt+1 (S) ,
ut (S)
(20)
with boundary condition the gT +1 (S) = 0. In particular,
eq. (20) holds since
gt (S) = min E kxt+1 (S)k2Qt +kut (S)k2Rt )+
ut (S)
min
ut+1:T (S)
ht+1 [S, ut+1:T (S)]} ,
where one can easily recognize the second summand to match
the definition of gt+1 (S).
We prove Lemma 9 for t = T . From eq. (20), for t = T ,
gT (S) = minuT (S) E(kxT +1 (S)k2QT +kuT (S)k2RT )
2
= minuT (S) E(kA
T xT + BT uT (S) + wT kQT +
2
kuT (S)kRT ) ,
(21)
since xT +1 (S) = AT xT + BT uT (S) + wT , as per eq. (1);
we note that for notational simplicity we drop henceforth the
dependency of xT on S since xT is independent of uT (S),
which is the variable under optimization in the optimization
problem in (21). Developing eq. (21) we get:
gT (S)
= minuT (S) E(uT (S)T BTT QT BT uT (S) + wTT QT wT +
T
T T
xT
T AT QT AT xT + 2xT AT QT BT uT (S)+
T T
2xT AT QT wT + 2uT (S)T BTT QT wT + kuT (S)k2RT )
= minuT (S) E(uT (S)T BTT QT BT uT (S) + kwT k2QT +
2
T T
T
xT
T AT QT AT xT + 2xT AT QT BT uT (S) + kuT kRT ) ,
(22)
where the latter equality holds since wT has zero mean
and wT , xT , and uT (S) are independent. From eq. (22),
rearranging the terms, and using the notation in eq. (8),
gT (S)
= minuT (S) E(uT (S)T (BTT QT BT + RT )uT (S)+
T T
T
kwT k2QT +xT
T AT QT AT xT + 2xT AT QT BT uT (S)
T
= minuT (S) E(kuT (S)k2MT +kwT k2QT +xT
T AT QT AT xT +
T T
2xT AT QT BT uT (S)
T
= minuT (S) E(kuT (S)k2MT +kwT k2QT +xT
T AT QT AT xT −
−1
T
T
2xT (−AT QT BT MT )MT uT (S)
2
2
T T
= minuT (S) E(kuT (S)k
MT +kwT kQT +xT AT QT AT xT −
T
T
2xT KT M
T uT (S)
= minuT (S) E(kuT (S) − KT xT k2MT +kwT k2QT +
T
T
xT
T (AT QT AT − KT MT KT )xT
= minuT (S) E(kuT (S) − KT xT k2MT +kwT k2QT +
T
xT
T (AT QT AT − ΘT )xT
= minuT (S) E(kuT (S) − KT xT k2MT +kwT k2QT +kxT k2NT
= minuT (S) E(kuT (S) − KT xT k2MT ) + tr (WT QT ) +
E(kxT k2NT ),
(23)
2
)
=
where
the
latter
equality
holds
since
E(kw
k
T QT
T
T
E tr wT QT wT = tr E(wT wT )QT = tr (WT QT ). Now
we note that
min E(kuT (S) − KT xT k2MT )
uT (S)
= E(kKT x̂T (S) − KT xT k2MT )
= tr ΘT ΣT |T (S) ,
(24)
since x̂T (S) is the Kalman estimator of the state xT , i.e.,
the minimum mean square estimator of xT , which implies
that KT x̂T (S) is the minimum mean square estimator of
KT xT (S) [19, Appendix E]. Substituting (24) back into
eq. (23), we get:
gT (S) = E(kxT k2NT ) + tr ΘT ΣT |T (S) + tr (WT QT ) ,
which proves that Lemma 9 holds for t = T .
We now prove that if Lemma 9 holds for t = l + 1, it
also holds for t = l. To this end, assume eq. (20) holds for
t = l + 1. Using the notation in eq. (8),
gl (S) = minul (S) E(kxl+1 (S)k2Ql +kul (S)k2Rl ) + gl+1 (S)
= minul (S) E(kxl+1 (S)k2Ql +kul (S)k2Rl )+
P
E(kxl+1 (S)k2Nl+1 ) + Tk=l+1 tr Θk Σk|k (S) +
tr (Wk Sk )]}
= minul (S) E(kxl+1 (S)k2Sl +kul (S)k2Rl )+
o
PT
k=l+1 [tr Θk Σk|k (S) + tr (Wk Sk )]
PT
= k=l+1 [tr Θk Σk|k (S) + tr (Wk Sk )]+
minul (S) E(kxl+1 (S)k2Sl +kul (S)k2Rl ).
(25)
In eq. (25), for the last summand in the last right-hand-side,
by following the same steps as for the proof of Lemma 9 for
t = T , we have:
2
minul (S) E(kxl+1 (S)k2Sl +ku
l (S)kRl ) =
E(kxl k2Nl ) + tr Θl Σl|l (S) + tr (Wl Ql ) ,
(26)
and ul (S) = Kl x̂l (S). Therefore, by substituting eq. (26) back
to eq. (25), we get:
PT
gl (S) = E(kxl k2Nl ) + k=l [tr Θk Σk|k (S) + tr (Wk Sk )].
(27)
which proves that if Lemma 9 holds for t = l +1, it also holds
for t = l. By induction, this also proves that Lemma 9 holds
for l = 1, and we already observed that g1 (S) matches the
original LQG cost in eq. (19), hence concluding the proof.
Proof of Theorem 1: The proof easily follows from
Lemma 9. Eq. (6) is a direct consequence
eq. (19), since
Pof
T
and
both E(xT
N
x
)
=
tr
Σ
N
1|1 1
1 1 1
t=1 tr (Wt St ) are
independent of the choice of the sensor set S. Second, (7)
directly follows from eq. (18).
A PPENDIX C: P ROOF
OF
T HEOREM 2
The following result is used in the proof of Theorem 2.
Proposition 2 (Monotonicity of cost function in eq. (6)).
Consider the cost function in eq.P
(6), namely, for any
sensor
T
set S ⊆ V the set function
tr
Θ
Σ
(S)
. Then,
t t|t
t=1
for
any
sensor
sets
such
that
S
⊆
S
⊆
V,
it
holds
1
2
PT
PT
tr
Θ
Σ
(S
)
≥
tr
Θ
Σ
(S
)
.
t
1
t
2
t|t
t|t
t=1
t=1
Proof: Lemma 6 implies Σ1|1 (S1 ) Σ1|1 (S2 ), and
then, Corollary 1 implies Σt|t (S1 ) Σt|t (S2 ). Finally, for
any t = 1, 2,. . . , T , Lemma 4 implies tr Θt Σt|t (S1 ) ≥
tr Θt Σt|t (S2 ) , since each Θt is symmetric.
Proof of part (1) of Theorem 2 (Algorithm 2’s approximation quality): Using Proposition 2, and the supermodularity
ratio Definition 4, the proof of the upper bound exp(−γg )
in ineq. (12) follows the same steps as the proof of [26,
Theorem 1].
Proof of part (2) of Theorem 2 (Algorithm 1’s running
time): We compute Algorithm 1’s running time by adding the
running times of Algorithm 1’s lines 1 and 2:
a) Running time of Algorithm 1’s line 1: Algorithm 1’s
line 1 needs O(k|V|T n2.4 ) time. In particular, Algorithm 1’s
line 2 running time is the running time of Algorithm 2, whose
running time we show next to be O(k|V|T n2.4 ). To this end,
we first compute the running time of Algorithm 2’s line 1,
and then the running time of Algorithm 2’s lines 3–15. Algorithm 2’s line 1 needs O(n2.4 ) time, using the Coppersmith
algorithm for both matrix inversion and multiplication [27].
Then, Algorithm 2’s lines 3–15 are repeated k times, due to the
“while loop” between lines 3 and 15. We now need to find the
running time of Algorithm 2’s lines 4–14; to this end, we first
find the running time of Algorithm 2’s lines 4–12, and then the
running time of Algorithm 2’s lines 13 and 14. In more detail,
the running time of Algorithm 2’s lines 4–12 is O(|V|T n2.4 ),
since Algorithm 2’s lines 5–11 are repeated at most |V|
times and Algorithm 2’s lines 6–10, as well as line 11 need
O(T n2.4 ) time, using the Coppersmith-Winograd algorithm
for both matrix inversion and multiplication [27]. Moreover,
Algorithm 2’s lines 13 and 14 need O[|V|log(|V|)] time, since
line 13 asks for the minimum among at most |V| values of the
cost(·) , which takes O[|V|log(|V|)] time to be found, using,
e.g., the merge sort algorithm. In sum, Algorithm 2’s running
time is O[n2.4 + k|V|T n2.4 + k|V|log(|V|)] = O(k|V|T n2.4 ).
b) Running time of Algorithm 1’s line 2: Algorithm 1’s
line 2 needs O(n2.4 ) time, using the Coppersmith algorithm
for both matrix inversion and multiplication [27].
In sum, Algorithm 1’s running time is O(k|V|T n2.4 +
2.4
n ) = O(k|V|T n2.4 ).
A PPENDIX D: P ROOF
OF
T HEOREM 3
Proof of Theorem 3: We complete the proof by first
deriving a lower bound for the numerator of the supermodularity ratio γg , and then, by deriving an upper bound for the
denominator of the supermodularity ratio γg .
use the following notation: c , E(xT
1 N1 x1 ) +
PWe
T
⊆
V,
and time
t=1 tr (Wt St ), and for any sensor set S
t = 1, 2, . . . , T , ft (S) , tr Θt Σt|t (S) . Then, the cost
P
function g(S) in eq. (11) is written as g(S) = c+ Tt=1 ft (S),
due to eq. (19) in Lemma 9.
a) Lower bound for the numerator of the supermodularity ratio γg : Per the supermodularity ratio Definition 4, the
numerator of the submodularity ratio γg is of the form
T
X
[ft (S) − ft (S ∪ {v})],
(28)
t=1
for some sensor set S ⊆ V, and sensor v ∈ V; to lower bound
the sum in (28), we lower bound each ft (S) − ft (S ∪ {v}).
To this end, from eq. (16) in Lemma 5, observe
X
T
C̄i,t
C̄i,t ]−1.
Σt|t (S ∪ {v}) = [Σ−1
t|t−1 (S ∪ {v}) +
i∈S∪{v}
PT
T
=
Define Ωt = Σ−1
i∈S C̄i,t C̄i,t , and Ω̄t
t|t−1 (S) +
P
T
−1
T
Σt|t−1 (S ∪{v})+ i∈S C̄i,t C̄i,t ; using the Woodbury identity
Lemmata 4 and 2,
in Lemma 3,
ft (S ∪ {v}) =
tr
−
tr Θt Ω̄−1
t
−1 T
Θt Ω̄t C̄v,t (I
g(S) − g(S ∪ {v}) ≥ l
T −1
.
) C̄v,t Ω̄−1
+ C̄v,t Ω̄t−1 C̄v,t
t
≥l
T
X
= ltr Φ
−1 T −1
T
C̄v,t Ω̄t−1 ≥
tr Θt Ω̄−1
t C̄v,t (I + C̄v,t Ω̄t C̄v,t )
−1 T −1
T
C̄v,t Ω̄t−1 ,
tr Θt Ω̄−1
t C̄v,t (I + C̄v,t Ω̄t C̄v,t )
≥ lλmin
≥ tr Θt Ω̄−1
where ineq. (29) holds because tr Θt Ω−1
t .
t
≥ tr Θt Ω̄t−1
In particular, the inequality tr Θt Ω−1
t
is implied as follows: Lemma 6 implies Σ1|1 (S)
Σ1|1 (S ∪ {u}). Then, Corollary 2 implies Σt|t−1 (S)
Σt|t−1 (S ∪ {v}), and as a result, Lemma 1 implies
Σt|t−1 (S)−1 Σt|t−1 (S ∪ {u})−1. Now, Σt|t−1 (S)−1
Σt|t−1 (S ∪ {u})−1 and the definition of Ωt and of Ω̄t
imply Ωt Ω̄t . Next, Lemma 1 implies Ωt−1 Ω̄−1
t .
As a result, since also Θt is a symmetric
matrix,
Lem
ma 4 gives the desired inequality tr Θt Ω−1
≥ tr Θt Ω̄−1
.
t
t
Continuing from the ineq. (29),
ft (S) − ft (S ∪ {v}) ≥
−1 T −1
−1 T
≥
tr C̄v,t Ω̄−1
t Θt Ω̄t C̄v,t (I + C̄v,t Ω̄t C̄v,t )
−1 T
T −1
)tr C̄v,t Ω̄−1
λmin ((I + C̄v,t Ω̄−1
t Θt Ω̄t C̄v,t , (30)
t C̄v,t )
where ineq. (30) holds due to Lemma 2. From ineq. (30),
−1 T
−1
T
≥ λ−1
max (I + C̄v,t Σt|t (∅)C̄v,t )tr C̄v,t Ω̄t Θt Ω̄t C̄v,t
T
X
Θt
t=1
T
X
t=1
(29)
−1
−1 T
T
,
= λ−1
max (I + C̄v,t Σt|t (∅)C̄v,t )tr Θt Ω̄t C̄v,t C̄v,t Ω̄t
(31)
where we used Ω̄−1
Σt|t (∅), which holds because of the
t
−1
following: the definition of Ω̄t implies Ω̄t Σt|t−1
(S ∪ {v}),
−1
and as a result, from Lemma 1 we get Ω̄t Σt|t−1 (S ∪{v}).
In addition, Corollary 2 and the fact that Σ1|1 (S ∪ {v})
Σ1|1 (∅), which holds due to Lemma 6, imply Σt|t−1 (S ∪
{v}) Σt|t−1 (∅). Finally, from eq. (16) in Lemma 5 it is
Σt|t−1 (∅) = Σt|t (∅). Overall, the desired inequality Ω̄−1
t
Σt|t (∅) holds.
Consider a time t′ ∈ {1, 2 . . . , T } such that for any time
−1
−1
−1 T
T
t ∈ {1, 2, . . . , T } it is Ω̄−1
t′ C̄v,t′ C̄v,t′ Ω̄t′ Ω̄t C̄v,t C̄v,t Ω̄t ,
−1 T
−1
and let Φ be the matrix Ω̄t′ C̄v,t′ C̄v,t′ Ω̄t′ ; similarly, let l be
T
the mint∈{1,2...,T },u∈V λ−1
max (I + C̄v,t Σt|t (∅)C̄v,t ). Summing
ineq. (31) across all times t ∈ {1, 2 . . . , T }, and using
tr (Θt Φ)
t=1
ft (S) − ft (S ∪ {v}) =
+
− tr Θt Ω̄−1
tr Θt Ω−1
t
t
−1 T
−1
−1 T
= λ−1
max (I + C̄v,t Ω̄t C̄v,t )tr C̄v,t Ω̄t Θt Ω̄t C̄v,t
−1
T
tr Θt Ω̄−1
t C̄v,t C̄v,t Ω̄t
t=1
Therefore, for any time t ∈ {1, 2 . . . , T },
ft (S) − ft (S ∪ {v}) ≥
T
X
!
Θt
!
tr (Φ)
> 0,
PT
which is non-zero because t=1 Θt ≻ 0 and Φ is a non-zero
positive semi-definite matrix.
Finally, we lower bound tr (Φ), using Lemma 2:
−1
T
tr (Φ) = tr Ω̄−1
t′ C̄v,t′ C̄v,t′ Ω̄t′
T
= tr Ω̄−2
t′ C̄v,t′ C̄v,t′
T
≥ λmin (Ω̄−2
t′ )tr C̄v,t′ C̄v,t′
T
= λ2min (Ω̄−1
t′ )tr C̄v,t′ C̄v,t′
T
(32)
≥ λ2min (Σt′ |t′ (V))tr C̄v,t
′ C̄v,t′ ,
where ineq. (32) holds because Ω̄−1
t′ Σt′ |t′ (V). In particular,
′
′
the inequality Ω̄−1
Σ
(S
∪
{v}) is derived by applying
t |t
t′
T
T
′
= Σ−1
C̄v,t
Lemma 1 to the inequality Ω̄t Ω̄t′ + C̄v,t
t′ |t′ (S ∪
{v}), where the equality holds by the definition of Ω̄t′ . In addition, due to Lemma 6 it is Σ1|1 (S ∪ {v}) Σ1|1 (V), and as
a result, from Corollary 1 it also is Σt′ |t′ (S ∪{v}) Σt′ |t′ (V).
Overall, the desired inequality Ω̄−1
t′ Σt′ |t′ (V) holds.
b) Upper bound for the denominator of the supermodularity ratio γg : The denominator of the submodularity ratio γg
is of the form
T
X
[ft (S ′ ) − ft (S ′ ∪ {v})],
t=1
for some sensor set S ′ ⊆ V, and sensor v ∈ V; to upper bound
it, from eq. (16) in Lemma 5 of Appendix A, observe
X
T
′
C̄i,t
C̄i,t ]−1,
Σt|t (S ′ ∪ {v}) = [Σ−1
t|t−1 (S ∪ {v}) +
i∈S ′ ∪{v}
PT
′
T
and let Ht = Σ−1
i∈S ′ C̄i,t C̄i,t , and H̄t =
t|t−1 (S ) +
P
T
T
′
Σ−1
i∈S ′ C̄i,t C̄i,t ; using the Woodbury ident|t−1 (S ∪ {v}) +
tity in Lemma 3,
ft (S ′ ∪ {v}) = tr Θt H̄t−1 −
T −1
T
) C̄v,t H̄t−1 .
(I + C̄v,t H̄t−1 C̄v,t
tr Θt H̄t−1 C̄v,t
Therefore,
T
X
t=1
T
X
t=1
[ft (S ′ ) − ft (S ′ ∪ {v})] =
[tr Θt Ht−1 − tr Θt H̄t−1 +
T
T −1
tr Θt H̄t−1 C̄v,t
(I + C̄v,t H̄t−1 C̄v,t
) C̄v,t H̄t−1 ] ≤
T
X
[tr Θt Ht−1 +
t=1
T −1
T
) C̄v,t H̄t−1 ],
(I + C̄v,t H̄t−1 C̄v,t
tr Θt H̄t−1 C̄v,t
(33)
where ineq. (33) holds since tr Θt H̄t−1 is non-negative. In
eq. (33), the second term in the sum is upper bounded as
follows, using Lemma 2:
T −1
T
) C̄v,t H̄t−1 =
(I + C̄v,t H̄t−1 C̄v,t
tr Θt H̄t−1 C̄v,t
T −1
T
≤
)
(I + C̄v,t H̄t−1 C̄v,t
tr C̄v,t H̄t−1 Θt H̄t−1 C̄v,t
−1 T −1
−1 T
−1
tr C̄v,t H̄t Θt H̄t C̄v,t λmax [(I + C̄v,t H̄t C̄v,t ) ] =
−1
T
T
)≤
λmin (I + C̄v,t H̄t−1 C̄v,t
tr C̄v,t H̄t−1 Θt H̄t−1 C̄v,t
−1
−1 T
−1
T
), (34)
tr C̄v,t H̄t Θt H̄t C̄v,t λmin (I + C̄v,t Σt|t (V)C̄v,t
T
T
since λmin (I + C̄v,t H̄t−1 C̄v,t
) ≥ λmin (I + C̄v,t Σt|t (V)C̄v,t
),
−1
−1
because H̄t Σt|t (V). In particular, the inequality H̄t
T
Σt|t (V) is derived as follows: first, it is H̄t H̄t + C̄v,t
C̄v,t =
′
−1
Σt|t (S ∪ {v}) , where the equality holds by the definition of
H̄t , and now Lemma 1 implies H̄t−1 Σt|t (S ′ ∪ {v}). In addition, Σt|t (S ′ ∪ {v}) Σt|t (V) is implied from Corollary 1,
since Lemma 6 implies Σ1|1 (S ′ ∪ {v}) Σ1|1 (V). Overall,
the desired inequality H̄t−1 Σt|t (V) holds.
T
Let l′ = maxt∈{1,2...,T },v∈V λ−1
min (I + C̄v,t Σt|t (V)C̄v,t ).
From ineqs. (33) and (34),
PT
[ft (S ′ ) − ft (S ′ ∪ {v})] ≤
Pt=1
(35)
T
−1
T
C̄v,t H̄t−1 ].
+ l′ tr Θt H̄t−1 C̄v,t
t=1 [tr Θt Ht
Consider times t′ ∈ {1, 2 . . . , T } and t′′ ∈ {1, 2 . . . , T } such
that for any time t ∈ {1, 2, . . . , T }, it is Ht−1
Ht−1 and
′
−1 T
−1
−1
−1 T
H̄t′′ C̄v,t′′ C̄v,t′′ H̄t′′ H̄t C̄v,t C̄v,t H̄t , and let Ξ = Ht−1
′
−1
T
and Φ′ = H̄t−1
′ C̄v,t′ C̄v,t′ H̄t′ . From ineq. (35), and Lemma 4,
T
X
[ft (S ′ ) − ft (S ′ ∪ {v})] ≤
t=1
T
X
[tr (Θt Ξ) + l′ tr (Θt Φ′ )] ≤
t=1
tr Ξ
T
X
Θt
t=1
!
′
+ l tr Φ
′
T
X
t=1
T
X
(tr (Ξ) + l′ tr (Φ′ ))λmax (
Θt
!
≤
Θt ).
(36)
t=1
Finally, we upper bound tr (Ξ)+l′ tr (Φ′ ) in ineq. (36), using
Lemma 2:
tr (Ξ) + l′ tr (Φ′ ) ≤
tr Ht−1
+
′
l′ λ2max (H̄t−1
′′ )tr
tr Σt′ |t′ (∅) +
T
≤
C̄v,t
′′ C̄v,t′′
′ 2
l λmax (Σt′′ |t′′ (∅))tr
(37)
T
C̄v,t
′′ C̄v,t′′ ,
(38)
where ineq. (38) holds because Ht−1
Σt′ |t′ (∅), and simi′
larly, H̄t−1
Σt′′ |t′′ (∅). In particular, the inequality Ht−1
′′
′
Σt′ |t′ (∅) is implied as follows: first, by the definition of Ht′ ,
it is Ht−1
= Σt′ |t′ (S ′ ); and finally, Corollary 1 and the fact
′
that Σ1|1 (S ′ ) Σ1|1 (∅), which holds due to Lemma 6, imply
Σt′ |t′ (S ′ ) Σt′ |t′ (∅). In addition, the inequality H̄t−1
′′
Σt′′ |t′′ (∅) is implied as follows: first, by the definition of H̄t′′ ,
′
it is H̄t′′ Σ−1
t′′ |t′′ −1 (S ∪ {v}), and as a result, Lemma 1
−1
implies H̄t′′ Σt′′ |t′′ −1 (S ′ ∪ {v}). Moreover, Corollary 2
and the fact that Σ1|1 (S ∪ {v}) Σ1|1 (∅), which holds due to
Lemma 6, imply Σt′′ |t′′ −1 (S ′ ∪ {v}) Σt′′ |t′′ −1 (∅). Finally,
from eq. (16) in Lemma 5 it is Σt′′ |t′′ −1 (∅) = Σt′′ |t′′ (∅).
Overall, the desired inequality H̄t−1
′′ Σt′′ |t′′ (∅) holds.
A PPENDIX E: P ROOF
OF
T HEOREM 4
For the proof of Theorem 4, we use Proposition 1, which
we prove in Appendix E, and Lemmata 10, 11, and 12 below.
Lemma 10. For any t = 1, 2, . . . , T ,
Θt = AT
t St At + Qt−1 − St−1 .
Proof of Lemma 10: Using the Woobury identity in
Lemma 3, and the notation in eq. (8),
−1
+ Bt Rt−1 BtT )−1 At
Nt = AT
t (St
−1 T
= AT
t (St − St Bt Mt Bt St )At
= AT
t St At − Θt .
The latter, gives Θt = AT
t St At −Nt . In addition, from eq. (8),
−Nt = Qt−1 − St−1 , since St = Qt + Nt+1 .
PT
T
T
Lemma 11. t=1 AT
1 A2 · · · At Qt At At−1 · · · A1 ≻ N1 if
and only if
T
X
T
T
AT
1 A2 · · · At−1 Θt At−1 At−2 · · · A1 ≻ 0.
t=1
Proof of Lemma 11: For i = t − 1, t − 2, . . . , 1, we preand post-multiply the identity in Lemma 10 with AT
i and Ai ,
respectively:
Θt = AT
t St At + Qt−1 − St−1 ⇒
T
T
T
AT
t−1 Θt At−1 = At−1 At St At At−1 + At−1 Qt−1 At−1 −
T
At−1 St−1 At−1 ⇒
T
T
T
AT
t−1 Θt At−1 = At−1 At St At At−1 + At−1 Qt−1 At−1 −
Θt−1 + Qt−2 − St−2 ⇒
T
T
Θt−1 + AT
t−1 Θt At−1 = At−1 At St At At−1 +
T
At−1 Qt−1 At−1 + Qt−2 − St−2 ⇒
... ⇒
T
T
Θ2 + AT
2 Θ3 A2 + . . . + A2 · · · At−1 Θt At−1 · · · A2 =
T
T
T
A2 · · · At St At · · · A2 + A2 · · · AT
t−1 Qt−1 At−1 · · · A2 +
. . . + AT
2 Q2 A2 + Q1 − S1 ⇒
T
T
Θ1 + AT
1 Θ2 A1 + . . . + A1 · · · At−1 Θt At−1 · · · A1 =
T
T
T
A1 · · · At St At · · · A1 + A1 · · · AT
t−1 Qt−1 At−1 · · · A1 +
. . . + AT
Q1 A1 − N1 ⇒
1
PT
AT · · · AT
t−1 Θt At−1 · · · A1 =
t=1
PT1
T
T
A
·
t=1 1 · · At Qt At · · · A1 − N1 .
(39)
The last equality in eq. (39) implies Lemma 11.
A PPENDIX F: P ROOF
Lemma 12.
for any t = 1, 2, . . . , T that At is
PConsider
T
T T
T
invertible.
A
A
t=1 1 2 · · · At−1 Θt At−1 At−2 · · · A1 ≻ 0 if
and only if
T
X
Θt ≻ 0.
t=1
Proof of Lemma 12: Let Ut = At−1 At−2 · · · A1 .
We Pfirst prove that for any non-zero vector z, if
T
T T
T
it is
t=1 A1 A2 · · · At−1 Θt At−1 At−2 · · · A1 ≻ 0, then
PT
T
t=1 z Θt z > 0. In particular, since Ut is invertible, —
because for any t ∈ {1, 2, . . . , T }, At is,—
PT
PT
T −⊤ T
T
Ut Θt Ut Ut−1 z
t=1 z Θt z = Pt=1 z Ut
(40)
T
T
= t=1 tr φt φT
t Ut Θ t Ut ,
where we let φt = Ut−1 z. Consider a time t′ such that for any
T
time t ∈ {1, 2 . . . , T }, φt′ φT
t′ φt φt . From eq. (40), using
Lemmata 4 and 2,
T
X
z T Θt z ≥
t=1
T
X
T
tr φt′ φT
t′ Ut Θt Ut
t=1
≥ tr φt′ φT
t′
T
X
UtT Θt Ut
t=1
finally provePthat for any non-zero vector z, if
PWe
T
T
T
T
t=1 Θt ≻ 0, then
t=1 zA1 · · · At−1 Θt At−1 · · · A1 z ≻ 0.
In particular,
tr ξtT Θt ξt ,
(41)
where we let ξt = Ut z. Consider time t′ such that for any
time t ∈ {1, 2 . . . , T }, ξt′ ξtT′ ξt ξtT . From eq. (40), using
Lemmata 4 and 2,
!
T
T
X
X
T
T
tr ξt Θt ξt ≥ tr ξt′ ξt′
Θt
t=1
t=1
T
X
Θt )
≥ tr ξt′ ξt′ λmin (
T
t=1
T
X
= kξt′ k22 λmin (
t=1
Σt|t =Wt =0
= xT
1 N1 x1 ,
(42)
since E(kx1 k2N1 ) = xT
N
x
,
because
x
is
known
(Σ
=
0),
1
1
1
1|1
1
and Σt|t and Wt are zero. In addition, for u1:T = (0, 0, . . . , 0),
the objective function in the noiseless perfect state information
LQG problem in eq. (14) is
PT
2
2
t=1 [kxt+1 kQt +kut (xt )kRt ] Σt|t =Wt =0
PT
(43)
= t=1 xT
Qx
PT t+1 T t Tt+1 T
T
= x1 t=1 A1 A2 · · · At Qt At At−1 · · · A1 x1 ,
since xt+1 = At xt = At At−1 xt−1 = . . . = At At−1 · · · A1 x1
when all u1 , u2 , . . . , uT are zero.
From eqs. (42) and (43), the inequality
T
X
T
X
t=1
t=1
t=1
[kxt+1 k2Qt +kut (xt )k2Rt ]
N1 ≺
> 0.
t=1
u1:T
T
X
T
T
AT
1 A2 · · · At Qt At At−1 · · · A1 x1
holds for any non-zero x1 if and only if
t=1
T
X
P ROPOSITION 1
t=1
!
T
X
UtT Θt Ut )
= kφt′ k22 λmin (
z T UtT Θt Ut z =
min
T
xT
1 N1 x1 < x1
T
X
UtT Θt Ut )
≥ tr φt′ φT
t′ λmin (
T
X
OF
Proof of Proposition 1: For any initial condition x1 ,
eq. (19) in Lemma 9 implies for the noiseless perfect state
information LQG problem in eq. (14):
Θt )
t=1
> 0.
Proof of Theorem 4: Theorem 4 follows from the sequential application of Proposition 1 and Lemmata 11, and 12.
T
AT
1 · · · At Qt At At−1 · · · A1 .
| 3cs.SY
|
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
arXiv:1405.0043v3 [math.RT] 6 Mar 2017
ROBERT GURALNICK, FLORIAN HERZIG, AND PHAM HUU TIEP
Abstract. The notion of adequate subgroups was introduced by Jack Thorne [59]. It
is a weakening of the notion of big subgroups used by Wiles and Taylor in proving automorphy lifting theorems for certain Galois representations. Using this idea, Thorne was
able to strengthen many automorphy lifting theorems. It was shown in [22] and [23] that
if the dimension is smaller than the characteristic then almost all absolutely irreducible
representations are adequate. We extend the results by considering all absolutely irreducible modules in characteristic p of dimension p. This relies on a modified definition
of adequacy, provided by Thorne in [60], which allows p to divide the dimension of the
module. We prove adequacy for almost all irreducible representations of SL2 (pa ) in the
natural characteristic and for finite groups of Lie type as long as the field of definition
is sufficiently large. We also essentially classify indecomposable modules in characteristic p of dimension less than 2p − 2 and answer a question of Serre concerning complete
reducibility of subgroups in classical groups of low dimension.
Contents
1. Introduction
2. Linear groups of low degree
3. Extensions and self-extensions
4. Finite groups with indecomposable modules of dimension ≤ 2p − 2
5. Bounding Ext1G (V, V ) and Ext1G (V, V ∗ )
6. Modules of dimension p
6.1. Imprimitive case
6.2. Chevalley groups in characteristic p
6.3. Remaining cases
7. Certain PIMs for simple groups
7.1. The case H = PSLn (q) with p = (q n − 1)/(q − 1) and n ≥ 2
7.2. The case H = PSUn (q) with p = (q n + 1)/(q + 1), n ≥ 3, (n, q) 6= (3, 2)
2
7
11
15
20
28
29
30
31
38
38
39
Date: March 7, 2017.
2010 Mathematics Subject Classification. Primary 20C20; Secondary 11F80.
Key words and phrases. Artin-Wedderburn theorem, irreducible representations, automorphic representations, Galois representations, adequate representations, complete reducibility, indecomposable module.
The first author was partially supported by NSF grants DMS-1001962, DMS-1302886 and the Simons
Foundation Fellowship 224965. He also thanks the Institute for Advanced Study for its support. The second
author was partially supported by a Sloan Fellowship and an NSERC grant. The third author was partially
supported by the NSF grant DMS-1201374 and the Simons Foundation Fellowship 305247.
We thank Barry Mazur and Jack Thorne for thoughtful discussion of various questions considered in the
paper, and Frank Lübeck and Klaus Lux for help with several computations. We also thank the referee for
careful reading of the paper.
1
2
R. GURALNICK, F. HERZIG, AND P. TIEP
7.3. The case H = SL2 (q) and p = q − 1 ≥ 3
7.4. The case H = Ap with p ≥ 7
8. Indecomposable modules of dimension less than 2p − 2
9. Adequacy for SL2 (q)
10. Adequacy for SLn (q)
11. Asymptotic adequacy
References
40
40
40
48
51
52
55
1. Introduction
Throughout the paper, let k be a field of characteristic p and let V be a finite dimensional
vector space over k. Let ρ : G → GL(V ) be an absolutely irreducible representation. Thorne
[59] called (G, V ) is adequate if the following conditions hold (we rephrase the conditions
slightly by combining two of the properties into one):
(i) p does not divide dim V ;
(ii) Ext1G (V, V ) = 0; and
(iii) End(V ) is spanned by the elements ρ(g) with ρ(g) semisimple.
If G is a finite group of order prime to p, then it is well known that (G, V ) is adequate.
In this case, condition (iii) is often referred to as Burnside’s Lemma, and it is a trivial
consequence of the Artin-Wedderburn Theorem. Furthermore, if G is a connected algebraic
group over k and V is a faithful, absolutely irreducible rational G-module of dimension
coprime to p, then (G, V ) is adequate (cf. [21, Theorem 1.2] and Theorem 11.5).
These conditions are a weakening of the conditions used by Wiles and Taylor in studying the automorphic lifts of certain Galois representations. See [9] for some applications.
Thorne [59] generalized various results assuming the weaker hypotheses for p odd. We
refer the reader to [59] for more references and details. See also [12] for further applications. Recently Thorne [60, Corollary 7.3] has shown that one can relax the condition that
p ∤ (dim V ), still with p odd. So more generally, we say that an absolutely irreducible
representation ρ : G → GL(V ) is adequate if:
(i) H 1 (G, k) = 0;
(ii) H 1 (G, (V ∗ ⊗ V )/k) = 0;
(iii) End(V ) is spanned by the elements ρ(g) with ρ(g) semisimple.
Note that we allow the case p = 2 in the definition. Thorne has used this extended notion
of adequacy to prove an automorphy lifting theorem for 2-adic Galois representations of
unitary type over imaginary CM fields, see [60, Theorem 5.1].
Observe that if p ∤ (dim V ), k is a direct summand of V ∗ ⊗ V . Thus, Ext1G (V, V ) = 0
implies that H 1 (G, k) = 0 in this case. Also note that, by the long exact sequence in
cohomology, if H 2 (G, k) = 0, then H 1 (G, (V ∗ ⊗ V )/k) = 0 follows from Ext1G (V, V ) = 0.
Thus, under the assumption that either p ∤ (dim V ) or H i (G, k) = 0 for i = 1, 2, adequacy
is equivalent to the two conditions:
(i) Ext1G (V, V ) = 0;
(ii) End(V ) is spanned by the elements ρ(g) with ρ(g) semisimple.
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
3
Following [20], we say that the representation ρ : G → GL(V ), respectively the pair
(G, V ), is weakly adequate if End(V ) is spanned by the elements ρ(g) with ρ(g) semisimple.
It was shown in [22, Theorem 9] that:
Theorem 1.1. Let k be a field of characteristic p and G a finite group. Let V be an
absolutely irreducible faithful kG-module. Let G+ denote the subgroup generated by the pelements of G. If dim W ≤ (p − 3)/2 for an absolutely irreducible kG+ -submodule W of V ,
then (G, V ) is adequate.
The example G = SL2 (p) with V irreducible of dimension (p − 1)/2 shows that the
previous theorem is best possible. However, the counterexamples are rare. In fact, as shown
in [23, Corollary 1.5], if dim V < p − 3, then the (p ± 1)/2-dimensional representations of
SL2 (p) are the only two counterexamples. More precisely, in [23] we extend Theorem 1.1 to
the more general situation that dim W < p and show that almost always (G, V ) is adequate:
Theorem 1.2. Let k be a field of characteristic p and G a finite group. Let V be an
absolutely irreducible faithful kG-module, and let G+ denote the subgroup generated by the
p-elements of G. Suppose that the dimension d of any irreducible kG+ -submodule in V is
less than p. Then the following statements hold.
(i) (G, V ) is weakly adequate.
(ii) Let W be an irreducible kG+ -submodule of V ⊗k k. Then (G, V ) is adequate, unless
the group H < GL(W ) induced by the action of G+ on W is as described in one
of the exceptional cases (a), (b)(i)–(vi) listed in [23, Theorem 1.3]. In particular,
if d < p − 3 and (G, V ) is not adequate, then d = (p ± 1)/2 and H ∼
= SL2 (p) or
PSL2 (p).
Above the threshold p−1 for dim W , there are lots of linear groups that are not adequate.
Still, if dim V = p, the situation is very much under control. In this paper, we extend
adequacy results to the case of linear groups of degree p and generalize the asymptotic
result [21, Theorem 1.2] to disconnected algebraic groups G (with p ∤ [G : G 0 ]) allowing
at the same time that p divides the dimension of the G-module. Next, we show that, in
all cases considered in Theorem 1.2, under some additional mild condition (say, G is not
p-solvable if p is a Fermat prime, and p > 5), one in fact has dim Ext1G (V, V ) ≤ 1 – a
result of interest in the deformation theory. An outgrowth of our results leads us to prove
an analogue of the first author’s result [19] and answer a question of Serre on complete
reducibility of finite subgroups of orthogonal and symplectic groups of small degree. In
fact, we essentially classify indecomposable modules in characteristic p of dimension less
than 2p − 2.
Note that if the kernel of ρ has order prime to p, then there is no harm in passing to
the quotient. So we will generally assume that either ρ is faithful or more generally has
kernel of order prime to p. Also, note that the dimension of cohomology groups and the
dimension of the span of the semisimple elements in G in End(V ) does not change under
extension of scalars. Hence, most of the time we will work over an algebraically closed field
k.
Our main results are the following. First we show that the condition that H 1 (G, k) = 0
in the definition of adequacy is not particularly constraining if dim V is small. In particular,
4
R. GURALNICK, F. HERZIG, AND P. TIEP
the next result follows fairly easily from [19] (see [20, Theorem 4.1]). See Theorem 4.10 for
a slightly more general result.
Theorem 1.3. Let G be a finite irreducible subgroup of GLd (k) with k algebraically closed
of characteristic p. Assume that H 1 (G, k) 6= 0 and d < 2p−2. Then G is solvable, d = p−1,
p or p + 1 and one of the following holds:
(i) d = p − 1, p = 2a + 1 is a Fermat prime, [G : Z(G)O2 (G)] = p and O2 (G) is a
group of symplectic type with O2 (G)/Z(O2 (G)) (elementary) abelian of order 22a ;
(ii) d = p and G has a normal abelian p′ -subgroup of index p;
(iii) d = p + 1, p = 2a − 1 is a Mersenne prime, and G contains a normal abelian
p′ -subgroup N such that G/N is a Frobenius group of order dp with kernel of order
d.
The following curious corollary is immediate from Theorem 4.10. We suspect that there
is a proof of this that does not require the classification of finite simple groups.
Corollary 1.4. Let G be a finite irreducible subgroup of GLd (k) with k algebraically closed
of characteristic p and d < 2p − 2. Suppose that G has a composition factor of order p.
Then G is solvable. Moreover, either d = p, or d = 2a with p = d ± 1 (and so p is either a
Mersenne prime or Fermat prime).
In the situation of Theorem 1.2, Ext1G (V, V ) may be nonzero and so G may fail to be
adequate. Nevertheless, we can prove the following two results, which were motivated
by discussions with Mazur and which are of interest in deformation theory. (Recall [47,
Section 1.2], for instance, that the inequality dim Ext1G (V, V ) ≤ n implies that the universal
deformation ring over the ring O of integers of a sufficiently large finite extension of Qp is
a quotient of O[[x1 , . . . , xn ]]. See also [6, Theorem 2.4].)
Theorem 1.5. Let k be a field of characteristic p and G a finite group. Let V be an
absolutely irreducible faithful kG-module, and let G+ denote the subgroup generated by the
p-elements of G. Suppose that the dimension d of any irreducible kG+ -submodule W in V
is less than p, and let H be the image of G+ in GL(W ).
(i) Suppose the following conditions hold:
(a) If p is a Fermat prime, then G is not p-solvable (equivalently, H is not solvable);
(b) If p = 3, then H ∼
6 SL2 (3a ) for all a ≥ 2;
=
(c) If p = 5 and dimk W = 4, then H 6∼
= Ω+
4 (5).
Then dimk Ext1G (V, V ) ≤ 1 and dimk Ext1G (V, V ∗ ) ≤ 1. In particular, H 1 (G, Sym2 (V )) and
H 1 (G, ∧2 (V )) are both at most 1-dimensional.
a
(ii) In the exceptional cases (p, dimk W, H) = (5, 4, Ω+
4 (5)) or (p, H) = (3, SL2 (3 )) with
1
1
∗
a ≥ 2, ExtG (V, V ) and ExtG (V, V ) are at most 2-dimensional.
Note that one cannot remove the conditions (a)–(c) in Theorem 1.5(i). In fact, in the
case G is p-solvable of Theorem 1.5, Ext1G (V, V ) and Ext1G (V, V ∗ ) can be of arbitrarily large
dimension. See Example 5.9. On the other hand, if dimk W < (p − 1)/2 in Theorem 1.5,
then H 1 (G, Sym2 (V )) = H 1 (G, ∧2 (V )) = 0, see Corollary 5.11.
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
5
In fact, we can show that both Ext1G (V, V ) and Ext1G (V, V ∗ ) are at most 1-dimensional in
another situation, without any dimension condition, but instead with a condition on Sylow
p-subgroups.
Theorem 1.6. Let k be a field of characteristic p and G a finite group. Let V be an
absolutely irreducible faithful kG-module, and let G+ denote the subgroup generated by the pelements of G. Suppose that the image of G+ in GL(W ) for some irreducible G+ -submodule
W of V has Sylow p-subgroups of order p, and that G has no composition factor of order
p. Then dimk Ext1G (V, V ) ≤ 1 and dimk Ext1G (V, V ∗ ) ≤ 1. In particular, H 1 (G, Sym2 (V ))
and H 1 (G, ∧2 (V )) are both at most 1-dimensional.
Next we determine adequacy of linear groups of degree p:
Theorem 1.7. Let k be a field of characteristic p and G a finite group. Let V be an absolutely irreducible faithful kG-module with dim V = p. Then precisely one of the following
holds:
(i) (G, V ) is adequate;
(ii) G contains a normal abelian subgroup of index p;
(iii) p = 3 and the image of G in PGL(V ) is PSL2 (9).
Extending the results of [23, §3], we prove in Corollary 9.4 that, aside from some exceptions with p = 2, 3 and with (q, dim(V )) = (p, (p ± 1)/2), all nontrivial irreducible
representations of SL2 (q) over Fq are adequate. This and other results on weak adequacy
and on Ext1 , and the dearth of examples where weak adequacy fails suggest that quite a
lot of irreducible representations are indeed weakly adequate. (Currently, all but one counterexample to weak adequacy are induced modules, and the only primitive counterexample
is given in [20].)
Finally, we classify all low dimensional self-dual indecomposable and non-irreducible kGmodules V with k algebraically closed of characteristic p and G a finite subgroup of GL(V )
with Op (G) = 1.
First we recall one of the main results of [19] which settled a conjecture of Serre.
Theorem 1.8. Let k be a field of positive characteristic p. Let G be a subgroup of GLn (k) =
GL(V ) with no nontrivial normal unipotent subgroup and p ≥ n + 2. Then V is completely
reducible.
Serre asked for an analogous result for the other classical groups. The example Ap <
SOp (k) shows that one cannot do too much better. We also see that there are reducible
indecomposable self-dual SL2 (p)-modules of dimensions p and p ± 1 (contained in Sp for
the dimension p − 1 and SO in the other cases). Building on the methods used in computing Ext1 , we can essentially classify the self-dual reducible indecomposable modules of
dimension less than 2p − 2.
Theorem 1.9. Let k be an algebraically closed field of characteristic p. Let V be a vector
space over k with dim V ≤ 2p − 3. Suppose that G is a finite subgroup of GL(V ) such that
Op (G) = 1, and the kG-module V is indecomposable and self-dual but not irreducible. Then
p > 3, G+ is quasisimple, VG+ is uniserial, and one of the following statements holds for
some U ∼
= U ∗ ∈ IBrp (G+ ).
6
R. GURALNICK, F. HERZIG, AND P. TIEP
(i) VG+ = (k|U |k), and (G+ , p, dim U ) is (SL2 (q), q−1, p+1), (Ap , p, p−2), (SLn (q), (q n −
1)/(q − 1), p − 2), (M11 , 11, 9), (M23 , 23, 21), or (PSL2 (p), p, p − 2).
(ii) VG+ = (U |U ). Furthermore, (G+ , p, dim U ) = (SL2 (q), q + 1, p − 2), (2A7 , 7, 4),
(PSL2 (p), p ≡ ǫ(mod 4), (p + ǫ)/2) or (SL2 (p), p ≡ ǫ(mod 4), (p − ǫ)/2) with ǫ =
±1.
Moreover, V supports a non-degenerate G-invariant bilinear form that is either symmetric
or alternating. Furthermore, all such forms have the same type, which is symmetric in all
cases, except when (G+ , p, dim U ) = ((P) SL2 (p), p, (p−1)/2) in which case it is alternating.
Conversely, all the listed cases give rise to reducible self-dual indecomposable modules of
dimension < 2p − 2.
In particular, this gives a classification of all finite non-G-cr subgroups for G = Sp(V ) or
SO(V ) with dim V < 2p − 2, see Proposition 8.8 (recall that the notion of G-cr subgroups
was introduced by Serre in [56]). It also yields the following variant of the main result of
[19]:
Corollary 1.10. Let k be an algebraically closed field of characteristic p and V a vector
space over k with d := dim V ≤ p − 1. Suppose G is a finite subgroup of GL(V ) such that
Op (G) = 1 and the kG-module V is self-dual. Then either the kG-module V is completely
reducible, or d = p − 1, G+ = (P) SL2 (p), and any G-invariant non-degenerate bilinear
form on V must be alternating.
In Theorem 1.9 and Corollary 1.10, the notation (P) SL2 (p) means SL2 (p) if p ≡ 1(mod
4) and PSL2 (p) if p ≡ 3(mod 4).
This paper is organized as follows. In §2, we describe the structure of quasisimple linear
groups of degree at most 2p. We collect various facts concerning extensions and selfextensions of simple modules in §3 and prove Theorem 1.3 in §4. Theorems 1.5 and 1.6
are proved in §5. Adequacy of linear groups of degree p is discussed in §6; in particular,
we prove Theorem 1.7. In the next §7, we describe the PIMs for various simple modules
of simple groups. These data are used in §8 to classify reducible self-dual indecomposable
modules of dimension at most 2p − 3, cf. Theorem 1.9, and to classify the finite non-G-cr
subgroups of symplectic and orthogonal groups in dimensions at most 2p−3, cf. Proposition
8.8 and Corollary 1.10. §9 is devoted to proving weak adequacy of SL2 (q)-representations,
cf. Proposition 9.1. In §10, we show that almost always the natural module for SLn (q) is
adequate. In §11, we prove Theorem 11.5 concerning adequacy of (possibly disconnected)
reductive algebraic groups and asymptotic adequacy.
Notation. If V is a kG-module and X ≤ G is a subgroup, then VX denote the restriction
of V to X. The containments X ⊂ Y (for sets) and X < Y (for groups) are strict.
Fix a prime p and an algebraically closed field k of characteristic p. Then for any finite
group G, IBrp (G) is the set of isomorphism classes of irreducible kG-representations (or
their Brauer characters, depending on the context), dp (G) denotes the smallest degree of
nontrivial ϕ ∈ IBrp (G), P(ϕ) is the principal indecomposable module (PIM) corresponding
to ϕ, and B0 (G) denotes the principal p-block of G. Sometimes we use 1 to denote the
principal representation. Op (G) is the largest normal p-subgroup of G, Op (G) is the smallest
normal subgroup N of G subject to G/N being a p-group, and similarly for Op′ (G) and
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
7
′
Op (G) = G+ . Furthermore, the Fitting subgroup F (G) is the largest nilpotent normal
subgroup of G, and E(G) is the product of all subnormal quasisimple subgroups of G,
so that F ∗ (G) = F (G)E(G) is the generalized Fitting subgroup of G. Given a finitedimensional kG-representation Φ : G → GL(V ), we denote by M the k-span
hΦ(g) | Φ(g) semisimpleik .
If M is a finite length module over a ring R, then define soci (M ) by soc0 (M ) = 0 and
socj (M )/ socj−1 (M ) = soc(M/ socj−1 (M )). If M = socj (M ) with j minimal, we say that
j is the socle length of M . If V is a vector space endowed with a non-degenerate quadratic
form, then O(V ) denotes the full isometry group of the form. For a linear algebraic group
G, G 0 denotes the connected component containing the identity.
2. Linear groups of low degree
First we recall the description of absolutely irreducible non-solvable linear groups of
degree less than p = char(k), relying on the main result of Blau and Zhang [5]:
Theorem 2.1. [23, Theorem 2.1]. Let W be a faithful, absolutely irreducible kH-module
′
for a finite group H with Op (H) = H. Suppose that 1 < dim W < p. Then one of the
following cases holds, where P ∈ Sylp (H).
(a) p is a Fermat prime, |P | = p, H = Op′ (H)P is solvable, dim W = p − 1, and Op′ (H)
is absolutely irreducible on W .
(b) |P | = p, dim W = p − 1, and one of the following conditions holds:
(b1) (H, p) = (SUn (q), (q n + 1)/(q + 1)), (Sp2n (q), (q n + 1)/2), (2A7 , 5), (3J3 , 19), or
(2Ru, 29).
(b2) p = 7 and H = 61 · PSL3 (4), 61 · PSU4 (3), 2J2 , 3A7 , or 6A7 .
(b3) p = 11 and H = M11 , 2M12 , or 2M22 .
(b4) p = 13 and H = 6 · Suz or 2G2 (4).
(c) |P | = p, dim W = p − 2, and (H, p) = (PSLn (q), (q n − 1)/(q − 1)), (Ap , p), (3A6 , 5),
(3A7 , 5), (M11 , 11), or (M23 , 23).
(d) (H, p, dim W ) = (2A7 , 7, 4), (J1 , 11, 7).
(e) Extraspecial case: |P | = p = 2n + 1 ≥ 5, dim W = 2n , Op′ (H) = RZ(H),
R = [P, R]Z(R) ∈ Syl2 (Op′ (H)), [P, R] is an extraspecial 2-group of order 21+2n , V[P,R]
is absolutely irreducible. Furthermore, S := H/Op′ (H) is simple non-abelian, and either
b ′
S = Sp2a (2b )′ or Ω−
2a (2 ) with ab = n, or S = PSL2 (17) and p = 17.
(f) Lie(p)-case: H/Z(H) is a direct product of simple groups of Lie type in characteristic
p.
Furthermore, in the cases (b)–(d), H is quasisimple with Z(H) a p′ -group.
Now we prove the following result which extends Theorem 2.1 for quasisimple groups
and is of independent interest. Note that the complex analogue of this result is given by
[63, Theorem 8.1].
Theorem 2.2. Let p be a prime, H a finite quasisimple group of order divisible by p.
Suppose W is a faithful, absolutely irreducible kH-module of dimension d, where p ≤ d ≤ 2p.
Then one of the following statements holds.
8
R. GURALNICK, F. HERZIG, AND P. TIEP
(i) H is a quasisimple group of Lie type in characteristic p.
(ii) (H, dim(W ), p) is as listed in Tables I, IIa, IIb, III, where the fourth column lists
the number of isomorphism classes of W for each choice of (H, dim(W ), p).
Proof. Let L be the universal covering group of S := H/Z(H) and let dp (L) denote the
smallest degree of nontrivial absolutely irreducible kL-representations. Then
(2.1)
2p ≥ dim(W ) ≥ dp (L).
This inequality will allow us to rule out the majority of the cases. We will assume that S
is not isomorphic to any finite simple group of Lie type in characteristic p.
First, let S be a sporadic group. Then dp (L) is listed in [33]. Furthermore, p ≤ 71 and
so dim W ≤ 142. Now the result follows from inspecting [31] and [34] (and also [49] for the
three Conway groups), and is listed in Table III.
Assume now that S = An . The cases 5 ≤ n ≤ 13 can be checked by inspecting [34],
and the result is listed in Table I. If 14 ≤ n ≤ 16, then p ≤ 13, dim W ≤ 26, and so the
statement follows by inspecting [31]. So we may assume n ≥ 17. In this case,
dim W ≤ 2p ≤ 2n < (n2 − 5n + 2)/2.
Hence, by [27, Lemma 6.1], W is the heart of the natural permutation module of G = An ,
yielding the first row of Table I.
Next suppose that S is an exceptional group of Lie type defined over Fq . The cases
S ∈ {2 B2 (8), G2 (3), G2 (4), 3 D4 (2), 2 F4 (2)′ , F4 (2)} can be checked using [34] and lead to the
last six rows of Table IIb. For all other groups, dp (L) is bounded below by the LandazuriSeitz-Zalesskii bounds (see [61, Table II] for latest improvements) and one can check that
(2.1) cannot hold. For instance, if S = F4 (q) with q ≥ 3, then p ≤ q 4 + 1 whereas
dp (L) ≥ q 8 + q 4 − 2.
From now on we may assume that S is a finite classical group defined over Fq and p ∤ q.
Suppose first that S = PSL2 (q). Using [34] we may assume that q ≥ 11. If q is even, then
L = SL2 (q) and each ϕ ∈ IBrp (L) has degree q or q ± 1, whereas p | (q ± 1). If in addition
p 6= q ± 1 then 2p ≤ 2(q + 1)/3 < q − 1 ≤ dim(W ), violating (2.1). So p = q ± 1, and
inspecting [7] we arrive at the first multi-row of Table IIa. Assume that q is odd. Then
again L = SL2 (q), and we also get additional possibilities ϕ(1) = (q ± 1)/2 for ϕ ∈ IBrp (L).
Note that p 6= (q ± 1), (q ± 1)/3 (as q ≥ 11 is odd) and (2.1) implies p > (q + 1)/5 as
dp (L) = (q − 1)/2. It follows that p = (q ± 1)/4 or p = (q ± 1)/2. A detailed analysis of
IBrp (L) leads to the 2nd and 3rd multi-rows of Table IIa.
Suppose now that S = PSLn (q) with n ≥ 3. Note that SL4 (2) ∼
= A8 . If (n, q) = (6, 3),
then p ≤ 13 whereas dp (L) ≥ 362 by [26, Table III]. If (n, q) = (6, 2), then p ≤ 31,
so [26, Table III] implies (dim W, p) = (62, 31), as recorded in Table IIa (the 9th row).
The cases (n, q) = (3, q ≤ 7), (4, 3) can be checked using [34]. So we may assume that
(n, q) 6= (3, q ≤ 7), (4, 2), (4, 3), (6, 2), (6, 3). In this case,
2
(q − 1)(q − 1)/ gcd(3, q − 1), n = 3,
2(q n − 1) (q 3 − 1)(q −
gcd(2, q −1), n = 4,
1)/
<
dim W ≤ 2p ≤
n−2
q
−q
q−1
− 1 , n ≥ 5.
(q n−1 − 1)
q−1
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
9
Applying [26, Theorem 1.1], we see that W is one of the Weil modules of L = SLn (q), of
dimension (q n − 1)/(q − 1) − a with a = 0, 1, 2. Note that
qn − 1
q n−2 − 1 q n−1 − 1 2 q n − 1
− 2 > max 2(q + 1), 2 ·
,
, ·
.
q−1
q−1
q−1 3 q−1
Q
If q ≥ 3, then (q n − 1)/(q − 1) − 2 > 2(q n−1 − 1)/(q − 1). Recall that p | ni=1 (q i − 1) and
p ≤ dim W ≤ 2p. So we arrive at one of the following possibilities:
• q = 2, p = 2n−1 − 1, whence W is the unique Weil module of dimension 2p by [26,
Theorem 1.1], leading to the 9th row of Table IIa;
• p = (q n − 1)/(q − 1), whence n is an odd prime, S = L, and W is one of q − 2 Weil
modules of dimension p by [26, Theorem 1.1], leading to the 8th row of Table IIa;
• 2p = (q n − 1)/(q − 1). Here, q is odd and n must be even, but then
qn − 1 q + 1
qn − 1
= 2
·
2(q − 1)
q −1
2
cannot be a prime.
Next suppose that S = PSUn (q) with n ≥ 3. The cases (n, q) = (3, q ≤ 5), (4, q ≤ 3),
(5, 2), (6, 2) can be checked using [34]. So we may assume that none of these cases occurs.
Now observe that if n = p = 3 | (q + 1) then 2p < (q − 1)(q 2 + 3q + 2)/6, and furthermore
(q − 1)(q 2 − q + 1)/3,
n = 3, p 6= 3 | (q + 1),
(2q 3 − q 2 + 2q − 3)/3,
n = 3, 3 ∤ (q + 1)
2(q n − (−1)n ) (q 2 + 1)(q 2 − q + 1)
2p ≤
<
− 1,
n = 4,
q+1
gcd(2, q − 1)
n
(q − (−1)n )(q n−1 − q 2 )
− 1, n ≥ 5.
(q + 1)(q 2 − 1)
Applying [30, Theorem 16] and [24, Theorem 2.7], we conclude that W is one of the Weil
modules of L = SUn (q), of dimension (q n − (−1)n )/(q + 1) − b with b = 0, ±1. Note that
2(q n−2 − (−1)n ) q n−1 + (−1)n 2(q n − (−1)n )
q n − (−1)n
− 1 > max 2(q + 1),
,
,
.
q+1
q+1
q+1
3(q + 1)
If q ≥ 3, then
q n − (−1)n
q n−1 − (−1)n−1
−1>2·
.
q+1
q+1
Qn
Recall that p | i=2 (q i − (−1)i ) and p ≤ dim W ≤ 2p. So we arrive at one of the following
possibilities:
• q = 2, p = (2n−1 − (−1)n−1 )/3; in particular, n − 1 ≥ 5 is a prime. Hence W is either
the unique Weil module of dimension 2p, or one of the two Weil modules of dimension
2p − 1, leading to the 11th and 12th rows of Table IIa;
• p = (q n − (−1)n )/(q + 1), whence n is an odd prime, S = L, and W is one of q Weil
modules of dimension p, yielding the 10th row of Table IIa;
• 2p = (q n − (−1)n )/(q + 1). Here, q is odd and n must be even, but then
q n/2 − (−1)n/2 q n/2 + (−1)n/2
q n − (−1)n
=
·
2(q + 1)
q+1
2
10
R. GURALNICK, F. HERZIG, AND P. TIEP
Table I. Quasisimple linear groups: Alternating groups
H
An
A5
2A5
A6
2A6
3A6
6A6
A7
A7
A7
2A7
2A7
3A7
3A7
6A7
A8
2A8
2A9
2A10
2A11
dim W
2, p | n
n−
1, p ∤ n
3
6
5, resp. 8, 10
10
6
6
4
8, resp. 10
10, resp. 14
4, resp. 6
14
6
9
6
14
8
8
8
16
p
n−1
≤ p≤ n−1
2
3
3
5
5
5
5
2
5
7
3
7
5
7
5
7
5, 7
5, 7
5
11
Class number
1
2
1
2, resp. 1, 1
2
2
4
2
1, resp. 2
1, resp. 2
2
2
2
2
4
1
1
2
2
1
cannot be a prime.
Now let S = PSp2n (q) with n ≥ 2. Note that Sp4 (2)′ ∼
= A6 and PSp4 (3) ∼
= SU4 (2).
Also, the cases (n, q) = (2, 4), (3, 2) can be checked using [34]. So we will assume that
(n, q) 6= (2, q ≤ 4), (3, 2). In this case,
dim W ≤ 2p ≤
(q n − 1)(q n − q)
2(q n + 1)
<
.
gcd(2, q − 1)
2(q + 1)
Using the Landazuri-Seitz-Zalesskii bound for Sp2n (q) with 2 | q and applying [24, Theorem
2.1] to Sp2n (q) with q odd, we now see that q must be odd and W is one of the Weil modules
of L = Sp2n (q), of dimension (q n ± 1)/2. So we arrive at the 4th–7th rows of Table IIa.
Note in addition that if 2 < p = (q n − 1)/4 then q = 5 and n is an odd prime, and if
p = (q n + 1)/4 then q = 3 and n is again an odd prime. Similarly, if p = (q n − 1)/2 then
q = 3 and n is an odd prime, and if p = (q n + 1)/2 then n = 2m .
Now we may assume that S = Ω2n+1 (q) with n ≥ 3, or P Ω±
2n (q) with n ≥ 4. Again, the
±
cases of Ω7 (3) and Ω8 (2) can be checked directly using [34]. Aside from these cases, one
can verify that (2.1) cannot hold.
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
11
Table IIa. Quasisimple linear groups: Groups of Lie type. I
H
dim W
p
p
p+1
p
2p − 2
2p − 1
2p
2p
p+1
2p
2p
2p
SL2 (q)
2|q
PSL2 (q)
2∤q
SL2 (q)
2∤q
PSp2n (q), 2 ∤ q, n ≥ 2
p
Sp2n (3), n odd prime
PSp2n (3), n odd prime
p+1
2p − 1
Sp2n (q), n odd prime
2p
SLn (q)
n odd prime
SLn (2)
n − 1 ≥ 3 prime
SUn (q)
n odd prime
PSUn (2)
n − 1 ≥ 5 prime
SUn (2)
n − 1 ≥ 5 prime
p
2p
p
q+1
q−1
q−1
(q ± 1)/2
(q + 1)/2
(q + 1)/4
(q − 1)/2
(q + 1)/2
(q − 1)/2
(q ± 1)/4
(q − 1)/2
(q + 1)/2
p = (q n − 1)/2, q = 3
p = (q n + 1)/2, n = 2m
(3n − 1)/2
(3n + 1)/4
p = (q n − 1)/4, q = 3
p = (q n + 1)/4, q = 5
qn − 1
q−1
Class number
q/2 − 1
q/2
1
2
1
2
(q − 3)/4
(q − 5)/4
2
2
(q + 1)/4
(q − 1)/4
2n−1 − 1
1
qn
2
2
2
2
q−2
+1
q+1
q
2p
(2n−1 + 1)/3
1
2p − 1
(2n−1 + 1)/3
2
p
3. Extensions and self-extensions
First we recall a convenient criterion concerning self-extensions in blocks of cyclic defect:
Lemma 3.1. [23, Lemma 7.1]. Suppose that G is a finite group and that V is an irreducible
Fp G-representation that belongs to a block of cyclic defect. Then Ext1G (V, V ) 6= 0 if and
only if V admits at least two non-isomorphic lifts to characteristic zero. In this case,
dim Ext1G (V, V ) = 1.
The next observation is useful in various situations:
Lemma 3.2. Let H ≤ G be a subgroup of index coprime to p = char(k). Suppose that V
is a kG-module and VH = V1 ⊕ V2 is a direct sum of two nonzero H-submodules, at least
one of which is also stabilized by G. Then the G-module V is decomposable.
12
R. GURALNICK, F. HERZIG, AND P. TIEP
Table IIb. Quasisimple linear groups: Groups of Lie type. II
H
SL3 (3)
2 · PSL3 (4)
2 · PSL3 (4)
41 · PSL3 (4)
42 · PSL3 (4)
6 · PSL3 (4)
PSL4 (3)
SU3 (3)
SU4 (2)
61 · PSU4 (3)
SU5 (2)
Sp4 (4)
Sp6 (2)
2 · Sp6 (2)
2 · Ω+
8 (2)
Ω−
8 (2)
2 B (8)
2
2 · 2 B2 (8)
G2 (3)
2 · G2 (4)
2 F (2)′
4
3 D (2)
4
dim W
p
Class number
16, resp. 26
13
1, resp. 3
6
3
1
10
5, resp. 7
2, resp. 1
8
5, resp. 7
2, resp. 4
4
3
2
6
5
2
26
13
2
14
7
1
10
2
5
6
1
6
5
2
10
5
1
18, resp. 34
17
1, resp. 2
7
5, 7
1
8
5, 7
1
8
5, 7
1
34
17
1
14
7, 13
2
8
5
1
16, 24
13
14
7, 13
1
12
7
1
26
13
2
26
13
1
Proof. Suppose for instance that V1 is stabilized by G, and consider the natural projection
π : V → V1 along V2 . Write G = ⊔m
i=1 gi H where m := [G : H] is coprime to p, and let
m
π̃ =
1 X
gi πgi−1 .
m
i=1
It is straightforward to check that π̃ is G-equivariant, π̃ 2 = π̃, and Im(π̃) = V1 . Hence the
G-module V decomposes as V1 ⊕ Ker(π̃).
From now on we again assume that k is algebraically closed of characteristic p. First we
record the following consequence of the Künneth formula, cf. [4, 3.5.6].
Lemma 3.3. Let H be a finite group. Assume that H is a central product of subgroups
Hi , 1 ≤ i ≤ t, and that Z(H) is a p′ -group. Let X and Y be irreducible kH-modules. Write
X = X1 ⊗ . . . ⊗ Xt and Y = Y1 ⊗ . . . ⊗ Yt where Xi and Yi are irreducible kHi -modules.
(i) If Xi and Yi are not isomorphic for two distinct i, then Ext1H (X, Y ) = 0.
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
13
Table III. Quasisimple linear groups: Sporadic groups
H
M11
2M12
2J2
M11
2M22
M11
M12
2M12
6Suz
J1
J1
J2
2J2
3J3
M22
3M22
M23
HS
M cL
M24
M23
Co3
Co2
2Co1
dim W
p
Class number
5
3
2
2p
3, 5
2
6
3, resp. 5
2, resp. 1
10
5
3
10
5, resp. 7
2, resp. 1
11, 16
11
1
11, resp. 16
11
2, resp. 1
12
11
1
12
7, 11, 13
2
14
11
1
22, 34
19
1
14
7
2
14
7
1
18
17
4
20
11
1
21
11
2
22
11
1
22
11
1
22
11
1
23, resp. 45
23
1, resp. 2
45
23
2
23
23
1
23
23
1
24
13, 23
1
(ii) If X1 and Y1 are not isomorphic but Xi ∼
= Yi for i > 1, then Ext1H (X, Y ) ∼
=
Ext1H1 (X1 , Y1 ).
(iii) If Xi ∼
= Yi for all i, then Ext1H (X, Y ) ∼
= ⊕i Ext1Hi (Xi , Yi ).
Lemma 3.4. Let G be a finite group with a normal subgroup N = Op (N ) and V be a
kG-module. Suppose that N acts trivially on V . Then H 1 (G, V ) ∼
= H 1 (G/N, V ).
Proof. Since N acts trivially on V , we have that H 1 (N, V ) = Hom(N, V ). Furthermore,
Hom(N, V ) = 0 as Op (N ) = N . Now the inflation-restriction sequence in cohomology
implies that the sequence
0 → H 1 (G/N, V ) → H 1 (G, V ) → 0
is exact, whence the claim follows. (Note that if N is a p′ -group, then the Hochschild-Serre
spectral sequence degenerates, and so H i (G, V ) ∼
= H i (G/N, V N ) for all i. Similarly, if G/N
′
i
i
G/N
is a p -group, then H (G, V ) = H (N, V )
for all i.)
Lemma 3.5. [23, Lemma 7.8]. Let V be a kG-module of finite length.
14
R. GURALNICK, F. HERZIG, AND P. TIEP
(i) Suppose that X is a composition factor of V such that V has no indecomposable
subquotient of length 2 with X as a composition factor. Then V ∼
= X ⊕ M for some
submodule M ⊂ X.
(ii) Suppose that Ext1G (X, Y ) = 0 for any two composition factors X, Y of V . Then V
is semisimple.
Lemma 3.6. [23, Lemma 7.9]. Let V be a kG-module. Suppose that U is a composition
factor of V of multiplicity 1, and that U occurs both in soc(V ) and head(V ). Then V ∼
=
U ⊕ M for some submodule M ⊂ V .
Lemma 3.7. Let G be group with a normal subgroup N of index coprime to p. Let k be
an algebraically closed field of characteristic p, and let V be a kG-module of finite length.
(i) V is semisimple if and only if VN is semisimple. In particular, if V is reducible
indecomposable, then VN cannot be semisimple.
(ii) Suppose V is reducible indecomposable. Then the N -module V has no simple direct
summand.
Proof. (i) The “only if” part is obvious. For the “if” part, suppose U is a G-submodule of
V . Since VN is semisimple, VN = U ⊕ W for some N -submodule W . As U is G-stable, by
Lemma 3.2 there is a G-submodule W ′ such that V = U ⊕ W ′ .
(ii) Consider a decomposition VN = ⊕ti=1 Ui into indecomposable direct summands, and
write V = V1 ⊕ V2 , where V2 is the sum of those Ui ’s which are simple and V1 is the sum
of the non-simple Ui ’s. Assume that V2 6= 0.
Note that if U is any reducible indecomposable N -module, then soc(U ) ⊆ rad(U ). Indeed, suppose a maximal submodule M ⊂ U does not contain soc(U ). Then soc(U ) =
(M ∩ soc(U )) ⊕ W for some N -submodule W 6= 0, and U = M ⊕ W is decomposable, a contradiction. Applying this remark to the summands Ui in V1 , we see that soc(V1 ) ⊆ rad(V1 ).
But V2 is semisimple, so
soc(V1 ) = soc(V1 ) ∩ rad(V1 ) = soc(V ) ∩ rad(V )
is G-stable. By Lemma 3.2, there is a G-submodule V2′ 6= 0 such that soc(V ) = soc(V1 )⊕V2′ .
In this case, VN = V1 ⊕ V2′ . Since V2′ is G-stable, again by Lemma 3.2 we have that
V = V1′ ⊕ V2′ for some G-submodule V1′ . As V is indecomposable and V2′ 6= 0, we must have
that V1′ = 0, whence V1 = 0, VN = V2 is semisimple, contradicting (i).
Lemma 3.8. [23, Lemma 7.11]. Let V be an indecomposable kG-module.
(i) If the G+ -module VG+ admits a composition factor L of dimension 1, then all composition factors of VG+ belong to B0 (G+ ).
(ii) Suppose a normal p′ -subgroup N of G acts by scalars on a composition factor L of
the G-module V . Then N acts by scalars on V . If in addition V is faithful then N ≤ Z(G).
Corollary 3.9. Let V be an indecomposable kG-module of dimension ≤ 2p − 3. Suppose
that dp (G+ ) ≥ p − 3. Then one of the following holds:
(i) The G+ -module V is irreducible.
(ii) All composition factors of the G+ -module V have dimension ≤ p.
(iii) All composition factors of the G+ -module V belong to B0 (G+ ).
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
15
Proof. Suppose that dim U > p for a composition factor U of the G+ -module V but V |G+
is reducible. Since dim(V ) − dim(U ) ≤ p − 4 < dp (G+ ), the G+ -module V must have a
composition factor L of dimension 1. Hence we are done by Lemma 3.8.
Finally, self-dual indecomposable modules of SL2 (q) (where q = pn ) of low dimension are
described in the following statement:
Proposition 3.10. [23, Proposition 8.2]. Suppose that V is a reducible, self-dual, indecomposable representation of SL2 (Fq ) over Fp , where q = pn . If dim V < 2p − 2, then q = p
and either of the following holds:
(i) dim V = p and V ∼
= P(1).
(ii) dim V = p + 1 and V is the unique nonsplit self-extension of L p−1
.
2
(iii) dim V = p − 1 and V is the unique nonsplit self-extension of L p−3
.
2
Conversely, all the listed cases give rise to examples.
4. Finite groups with indecomposable modules of dimension ≤ 2p − 2
Throughout this section, we assume that k is an algebraically closed field of characteristic
p > 3. First we recall several intermediate results proved in [23]:
Lemma 4.1. [23, Lemma 9.1]. Let G be a finite group, p > 3, and V be a faithful kGmodule of dimension < 2p. Suppose that Op (G) = 1 and Op′ (G) ≤ Z(G). Then F (G) =
Op′ (G) = Z(G), F ∗ (G) = E(G)Z(G), and G+ = E(G) is either trivial or a central product
of quasisimple groups of order divisible by p. In particular, G has no composition factor
isomorphic to Cp and so H 1 (G, k) = 0.
Lemma 4.2. [23, Lemma 9.3]. Let V be a faithful indecomposable kG-module with two
composition factors V1 , V2 . Assume that Op (G) = 1 and dim V ≤ 2p−2. If J := Op′ (G+ ) 6≤
Z(G+ ), then the following hold:
(i) p = 2a + 1 is a Fermat prime,
(ii) dim V1 = dim V2 = p − 1,
(iii) J/Z(J) is elementary abelian of order 22a ,
(iv) H 1 (G+ , k) 6= 0.
Lemma 4.3. [23, Lemma 9.5]. Let H be a quasisimple finite group of Lie type in char
p > 3. Assume that V1 , V2 ∈ IBrp (H) satisfy dim V1 + dim V2 < 2p.
(i) If H ∼
6 SL2 (q), PSL2 (q), then Ext1H (V1 , V2 ) = 0. In particular, there is no reducible
=
indecomposable kG-module V with G+ ∼
= H and dim V < 2p.
(ii) Suppose H ∼
= SL2 (q) or PSL2 (q), Ext1H (V1 , V2 ) 6= 0, and dim V1 = dim V2 . Then
q = p and V1 = L((p − 3)/2) or L((p − 1)/2).
Proposition 4.4. [23, Proposition 9.7]. Let p > 3 and let G be a finite group with a faithful,
reducible, indecomposable kG-module V of dimension ≤ 2p − 3. Suppose in addition that
Op (G) = 1. Then G+ = E(G+ ), G has no composition factor isomorphic to Cp , and one
of the following holds.
(i) G+ is quasisimple.
16
R. GURALNICK, F. HERZIG, AND P. TIEP
(ii) G+ is a central product of two quasisimple groups and dim V = 2p − 3. Furthermore,
V has one composition factor of dimension 1, and either one of dimension 2p − 4 or two
of dimension p − 2. In either case, V ∼
6 V ∗.
=
Corollary 4.5. Let k be a field of characteristic p and let V be a faithful reducible indecomposable kG-module of a finite group G with Op (G) = 1. If dim V ≤ 2p − 3, then VG+
is indecomposable.
Proof. Assume the contrary. Then we can pick an indecomposable direct summand U of
dimension ≤ p − 2 of VG+ and let H ≤ GL(U ) be the image of G+ acting on U . By
Proposition 4.4, G has no composition factors isomorphic to Cp . Hence Op (H) = 1. Since
the kH-module U is faithful and indecomposable, U is simple by [19, Theorem A]. But this
contradicts Lemma 3.7(ii).
Recall that a component of a finite group is any subnormal quasisimple subgroup. We
first note that:
Lemma 4.6. Let G be an irreducible subgroup of GL(V ) ∼
= GLd (k) with k algebraically
′
p
closed of characteristic p. Assume that G = O (G) and that G has a component of order
coprime to p. Then d ≥ 2p.
Proof. Assume the contrary: d < 2p. Write E(G) = E1 ∗ E2 , where E1 is a central product
of all components of G of order coprime to p and E2 is the product of the remaining
components; in particular, 1 6= E1 ✁ G. Since G is generated by p-elements, there is a
p-element x not centralizing E1 . Let W be an irreducible constituent of VE1 . Since d < 2p
and dim W ≥ 2, xW ∼
= W.
Now write E1 = Q1 ∗. . .∗Qn as a central product of n components and W ∼
= W1 ⊗. . .⊗Wn ,
where Wi is an irreducible kQi -module. Note that x acts on the set {Q1 , . . . , Qn }. If
this action is nontrivial, then dim W ≥ 2p. (Indeed, we may assume that x permutes
Q1 , . . . , Qm cyclically for some p ≤ m ≤ n, and, replacing W by another E1 -summand of V
if necessary, that Q1 acts nontrivially on W , i.e. dim W1 ≥ 2. Since xW ∼
= W , this implies
that dim Wi ≥ 2 for 1 ≤ i ≤ m, whence dim W ≥ 2m ≥ 2p ≥ 2p.) Thus we may assume
that x normalizes each Qi , but does not centralize Q1 . It follows that Q1 is a quasisimple
p′ -group with a nontrivial outer automorphism of p-power order; in particular, p > 2. If
p = 3, then Q1 ∼
= 2 B2 (22a+1 ) and dim W1 ≥ 14. So p > 3 and, using the description of outer
automorphisms of finite simple groups [18, Theorem 2.5.12], we see that Q1 is a quasisimple
group of Lie type over a field of size q p for some prime power q. Now applying [42], we see
that dim W1 ≥ 2p.
The proof of Lemma 4.6 certainly depends on the classification of finite simple groups.
We note the following result which does not require the classification:
Lemma 4.7. Let k = k̄ be of characteristic p and let G be a finite irreducible p-solvable
subgroup of GL(V ) ∼
= GLn (k) of order divisible by p. Then n ≥ p − 1.
Proof. We may assume that p > 2. By a result of Isaacs [50, Theorem 10.6], V has a
p-rational lift to characteristic 0. But then by [11], the Jordan blocks of any element g ∈ G
of order p acting on V have sizes 1, p − 1, or p. So if n < p − 1, then g acts trivially on V ,
a contradiction.
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
17
In what follows, we will slightly abuse the language by also considering Cp as a Frobenius
group with kernel of order p.
Lemma 4.8. Let p > 2 be a prime and let G be a transitive subgroup of Sn with n < 2p.
Assume that G has a composition factor of order p. Then one of the following holds.
(i) n = p and G is a Frobenius group of order pe for some e | (p − 1) with kernel of
order p.
(ii) p = 2a − 1 is a Mersenne prime and n = 2a = p + 1. Moreover, soc(G) is a regular
elementary abelian subgroup of order n, G = soc(G) ⋊ G1 , and G1 is a Frobenius
group of order pe for some e | a, with kernel of order p. If H 1 (G, k) 6= 0, then
|G| = np.
Proof. Note that G is primitive and contains a p-cycle. Hence we can apply [65] and see
that either (i) holds, or n = 2a = p + 1 and G = soc(G) ⋊ G1 , with soc(G) ∼
= C2a being
regular, and G1 ≤ GLa (2) has Cp as a composition factor. Applying [38] to G1 , we arrive
at (ii).
Lemma 4.9. Let S := Sp2a (2) with p = 2a ± 1 and let V = F2a
2 denote the natural module
for S. Let X ≤ S be a group with Cp as a composition factor and p > 3. Then there is a
normal elementary abelian 2-subgroup E < X such that X/E is a Frobenius group of order
pe with kernel of order p, where e | 2a. Furthermore, if E 6= 1 then p = 2a − 1. Moreover,
X acts reducibly on V precisely when p = 2a − 1 and either E 6= 1 or |X| is odd, in which
case X stabilizes a maximal totally isotropic subspace of V .
′
Proof. (a) It is easy to see that Y := Op (X) can be reducible on V only when p = 2a − 1.
Let P ∈ Sylp (X) and consider the action of X on the natural module V = F2a
2 for S. If
P ✁ X, then X is contained in NS (P ), a Frobenius group of order 2ap, in which case we
set E = 1. It follows that, if p = 2a − 1 in addition, then X is reducible on V precisely
′
when |X| is odd. So we will assume that P 5 X. It follows that P 5 Y := Op (X).
Suppose that the Y -module V is reducible, and so p = 2a − 1 and a ≥ 3 is odd. Then
Y stabilizes a proper subspace U 6= 0 of V of dimension ≤ a. Choosing U minimal, we
see that U is irreducible over Y . If U ∩ U ⊥ = 0, then Y is contained in the p′ -subgroup
Sp(U ) × Sp(U ⊥ ), a contradiction. Hence U ⊆ U ⊥ . Now if dim U < a then Y is contained
in the p′ -subgroup StabS (U ), again a contradiction. Thus dim U = a and U is a maximal
totally isotropic subspace. Setting Q := O2 (R) for R := StabS (U ) and F := Q ∩ Y , we see
that F is an elementary abelian 2-subgroup and Y /F is a subgroup of R/Q ∼
= GLa (2) with
Cp as a composition factor. By the main result of [38], Y /F is a Frobenius group of order
pb for some b | a. In particular, |Y /F | is odd and so F = O2 (Y ) ✁ X. Note that F 6= 1 as
otherwise P ✁ Y , a contradiction. Also, Y /F acts irreducibly on V /U . Since Y /F acts on
CV (F ) ⊇ U and F 6= 1, it follows that CV (F ) = U and so U is fixed by X. Thus X ≤ R
and we are done by setting E := Q ∩ X ≥ F .
(b) From now on we may assume that VY is irreducible. Hence V is absolutely irreducible
over k0 := EndF2 Y (V ). We will consider V as a b-dimensional vector space V ′ over k0 (for
some b | 2a). Thus W := V ′ ⊗k0 k is an irreducible kY -module for k := k̄0 . Observe that
b ≤ 2a ≤ p − 1. Also, Z(Y ) is cyclic by Schur’s lemma.
18
R. GURALNICK, F. HERZIG, AND P. TIEP
Note that if N ✁ Y then the N -module W is homogeneous. Indeed, VN is the direct sum
of t ≤ b < p homogeneous N -components Vi . Hence any p-element 1 6= g ∈ Y stabilizes
′
each Vi , whence Vi is fixed by Y = Op (Y ) and t = 1.
(c) Now we show E(Y ) = 1. Suppose E(Y ) 6= 1 and write E(Y ) = L1 ∗ . . . ∗ Ln , a central
product of n quasisimple groups. Since |S|p = p and Cp is a composition factor of Y , E(Y )
is a p′ -group. By (b), WE(Y ) ∼
= e(W1 ⊗ . . . ⊗ Wn ), where Wi is an irreducible kLi -module
of dimension ≥ 2. Hence b ≥ 2n and so n < p. It follows that every p-element 1 6= g ∈ Y
normalizes each Li , and so does Y . On the other hand, if Y centralizes Li , then Li ≤ Z(Y ),
a contradiction. So some p-element 1 6= g ∈ Y normalizes but does not centralize L1 . As
in the proof of Lemma 4.6, we see that L1 is a quasisimple group of Lie type defined over
Fq with q = r cp for some prime r and some integer c, and conclude that dim W1 > p when
r 6= 2. If r = 2, then by [66], |L1 | is divisible by some prime divisor ℓ of 2p − 1 that does
Q
i
not divide p−1
i=1 (2 − 1), whence ℓ ∤ |S|, again a contradiction.
Next we observe that every normal abelian subgroup A of Y must be central and so
cyclic. Indeed, A acts by scalars on W by (b), and so A ≤ Z(Y ).
(d) We have shown that F ∗ (Y ) = F (Y ). Now if p divides |F (Y )|, then since |S|p = p, P =
Op (F (Y )) ✁ Y , a contradiction. Also, if F (Y ) ≤ Z(Y ), then Y ≤ CY (F (Y )) ≤ F (Y ), and
so Y = F (Y ) is nilpotent, again a contradiction. So F (Y ) is a p′ -group and moreover N :=
Or (F (Y )) is non-central in Y for some prime r 6= p. By (c), every characteristic abelian
subgroup of N is cyclic. Hence by Hall’s theorem, N = F ∗ D, where F is an extraspecial
r-group, and either D is cyclic, or r = 2 and C is dihedral, generalized quaternion, or
semi-dihedral. Arguing as in part (3) of the proof of [25, Theorem 6.7], we can find a
characteristic subgroup L of N such that L = Z(L)E, where E is an extraspecial r-group
of order r 2c+1 for some c ≥ 1 and Z(L) is cyclic. Note that Z(L) ≤ Z(Y ) by (c). It also
follows by (b) that r c | b and r 6= 2, whence a ≥ 3 must be odd (recall that p = 2a ± 1
is prime). As L 6≤ Z(Y ), some p-element 1 6= g ∈ Y normalizes but does not centralize
L, and centralizes Z(L). It follows that p divides the order of the group Outc (L) of outer
automorphisms of L that act trivially on Z(L). Since Outc (L) ֒→ Sp2c (r), we get p | (r d ±1)
for some d ≤ c. Since p ≥ 2a − 1 and r c | a, we arrive at a contradiction.
Now we can prove Theorem 1.3 in a slightly stronger version.
Theorem 4.10. Let G be a finite irreducible subgroup of GL(V ) ∼
= GLd (k) with k algebraically closed of characteristic p > 3. Assume that G has a composition factor of order
p and d < 2p − 2. Then d = p − 1, p or p + 1, a Sylow p-subgroup of G has order p, G is
solvable and one of the following holds:
(i) d = p − 1, p = 2a + 1 is a Fermat prime, F ∗ (G) = Z(G)O2 (G), G/F ∗ (G) is a
Frobenius group of order pe for some e | 2a with kernel of order p, and O2 (G) is a
group of symplectic type with O2 (G)/Z(O2 (G)) ∼
= C22a ;
(ii) d = p and G has a normal abelian p′ -subgroup N = F ∗ (G) such that G/N is a
Frobenius group of order pe for some e | (p − 1) with kernel of order p;
(iii) d = p + 1, p = 2a − 1 is a Mersenne prime and one of
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
19
(a) G has an abelian normal p′ -subgroup N , where the action of G/N on the d
distinct eigenspaces of N induces a subgroup of Sd as described in Lemma 4.8;
or
(b) F ∗ (G) = Z(G)O2 (G), G/F ∗ (G) is a Frobenius group of order 2bp for some
b | a with kernel of order p, and O2 (G) is a group of symplectic type with
O2 (G)/Z(O2 (G)) ∼
= C22a .
Moreover, H 1 (G, k) 6= 0 if and only if one of the conclusions of Theorem 1.3 holds.
Proof. (a) First we show that G+ is irreducible on V . To this end, let W be an irreducible
summand of VG+ . Also let K1 and K2 denote the kernel of the action of G+ on W and
on a G+ -invariant complement V ′ to W in V . Then K1 ∩ K2 = 1 and so K1 embeds in
G+ /K2 as a normal subgroup. Since Cp is a composition factor of G+ , it follows that it
is a composition factor of G+ /K1 or of G+ /K2 . Replacing W by another irreducible G+ summand in V ′ if necessary, we may assume that G+ /K1 has a composition factor of order
p. Then dim W ≥ p−1 by Theorem 2.1. This is true for all other irreducible G+ -summands
in V and dim V < 2p − 2, whence the claim follows. In particular, Z(G+ ) = Z(G) ∩ G+ .
(b) By Lemma 4.1, we have that Q 6≤ Z(G+ ) and so Q 6≤ Z(G) for Q := Op′ (G+ ) ✁ G.
Suppose that Q contains a non-central (in G) abelian subgroup K ✁ G. Decompose V =
⊕ni=1 Vi into K-eigenspaces. By Clifford’s theorem, G acts transitively on {V1 , . . . , Vn }, with
kernel N , and G+ does as well. Note that n > 1 as K 6≤ Z(G). Since G+ is generated by
p-elements, n ≥ p. But d < 2p, so dim Vi = 1 and N is an abelian p′ -group. Now we can
apply Lemma 4.8 to G/N ֒→ Sd . If d = p, we are in case (ii) and F ∗ (G) = N . If d > p,
then d = p + 1 = 2a and G/N has the prescribed structure, i.e. case (iii)(a) holds.
(c) Assume now that Q contains no abelian non-central (in G) subgroup K ✁ G. Let
N be minimal among subgroups of Q that are normal but non-central in G, so N is
non-abelian. By Lemma 4.6, Q contains no components of G. Hence E(N ) = 1 and
F ∗ (N ) = F (N ). If F (N ) < N , the minimality of N implies that F (N ) ≤ Z(G), but
then F (N ) = F ∗ (N ) ≥ CN (F ∗ (N )) = N , a contradiction. So N = F (N ) is nilpotent,
and the minimality of N again implies that N is an r-group for some prime r 6= p. Let
A be any characteristic abelian subgroup of N . Then by the assumption, A ≤ Z(G) and
so it is cyclic. Thus every characteristic abelian subgroup of N is cyclic (and central in
G), and so Hall’s theorem applies to N . Arguing as in part (d) of the proof of Lemma
4.9 and using the minimality of N , we see that N = Z(N )E, where E is an extraspecial
r-group of order r 2a+1 for some a and Z(N ) ≤ Z(G) is cyclic; in particular, r a | d. Since
N 6≤ Z(G+ ), there is a p-element x that induces a nontrivial outer automorphism of N
acting trivially on Z(N ). As Outc (N ) ≤ Sp2a (r), we see that p divides r 2b − 1 for some
1 ≤ b ≤ a. On the other hand, r a ≤ d ≤ 2p − 3. This implies that r = 2, p = 2a ± 1
is either a Mersenne or Fermat prime, d = 2a , and N acts irreducibly on V . The latter
then implies that CG (N ) = Z(G) and X := G/Z(G)N is a subgroup of Outc (N ) ≤ Sp2a (2)
with Cp as a composition factor. Now we can apply Lemma 4.9 to X. Note that if X
stabilizes a maximal totally isotropic subspace of N/Z(N ), then its inverse image in N is
an abelian normal non-central subgroup of G, contrary to our assumptions. Hence either
p = 2a + 1 and we are in case (i), or p = 2a − 1 and we are in case (iii)(b). Also note that
F ∗ (G) = Z(G)N in either case.
20
R. GURALNICK, F. HERZIG, AND P. TIEP
(d) If G satisfies any of the conclusions of Theorem 1.3 then H 1 (G, k) 6= 0. Conversely,
suppose that H 1 (G, k) 6= 0. Then G possesses a normal subgroup L of index p. Thus we
can apply the above results to G. In particular, |G|p = p and so L = Op′ (G). Now the
description of G in (i)–(iii) shows that G must satisfy one of the conclusions of Theorem
1.3.
One can also consider an analogue of Theorem 4.10 for p = 3. In this case, d ≤ 3 and
the analogous result is straightforward by examining subgroups of GL2 and GL3 .
5. Bounding Ext1G (V, V ) and Ext1G (V, V ∗ )
The following result is well known:
Lemma 5.1. Let X be a finite group and let k be an algebraically closed field of characteristic p. Let U and V be irreducible kX-modules belonging to a kX-block B with cyclic
defect subgroups. Then dimk Ext1X (U, V ) ≤ 1.
Proof. By [23, Lemma 7.1], we may assume U ∼
6 V . It is known [53] that P(V ) has simple
=
head and simple socle, both isomorphic to V , and rad(P(V ))/V is a direct sum of at
most two uniserial submodules. Also, note that Ext1 (U, V ) ∼
= HomG (U, P(V )/V ). So if
1
dimk ExtX (U, V ) ≥ 2, then at least two edges of the Brauer tree of B correspond to U ,
which is impossible.
Lemma 5.2. Let H be a finite group with Sylow p-subgroups of order p. Suppose that
′
H = Op (H) and H has no composition factor of order p. Then H/Op′ (H) is a nonabelian simple group.
Proof. Replacing H by H/Op′ (H), we have that Op′ (H) = 1. Together with the condition
that H has no composition factor of order p, this implies that F (H) = 1, and so
F ∗ (H) = E(H) = S1 × . . . × Sn
is a product of non-abelian simple groups. It also follows that Op′ (F ∗ (H)) = 1, whence
′
n = 1 and F ∗ (H) = S1 has order divisible by p. Now H/S1 is a p′ -group and H = Op (H).
It follows that H = S1 .
Proposition 5.3. Let G be a finite group with a faithful kG-module V . Suppose that
V = W1 ⊕ . . . ⊕ Wt is a direct sum of kG+ -submodules, and that for each i the subgroup
Hi ≤ GL(Wi ) induced by the action of G+ on Wi has Sylow p-subgroups of order p. Suppose
in addition that G has no composition factor of order p. Then
G+ /Op′ (G+ ) ∼
= S1 × . . . × Sn
is a direct product of non-abelian simple groups Si , each of order divisible by p.
′
Proof. By assumption, Hi has no composition factor of order p, |Hi |p = p, and Op (Hi ) =
Hi . By Lemma 5.2, Hi /Op′ (Hi ) is simple non-abelian. Hence the claim follows by [23,
Lemma 2.3].
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
21
Proposition 5.4. Let k = k be of characteristic p and let H be a finite group such that
′
H = Op (H) and
H/J = S1 × . . . × Sn
is a direct product of non-abelian simple groups of order divisible by p, where J := Op′ (H).
Suppose that W1 and W2 are irreducible kH-modules such that the image of H in GL(Wi )
has Sylow p-subgroups of order p for i = 1, 2, and that Ext1H (W1 , W2 ) 6= 0. Then the actions
of H on W1 and W2 have the same kernel.
Proof. Let Ki denote the kernel of the action of H on Wi , so that |H/Ki |p = p. Note that H,
and so K1 ∩K2 as well, has no composition factor of order p, whence K1 ∩K2 = Op (K1 ∩K2 ).
Hence by Lemma 3.4 there is no loss to assume that
(5.1)
K1 ∩ K2 = 1.
′
We aim to show in this case that K1 = K2 = 1. Note that the condition H = Op (H)
implies that n ≥ 1.
(i) Suppose for instance that J1 := J ∩ K1 6= 1. This implies by (5.1) that J1 6≤ K2 ,
i.e. J1 does not act trivially on W2 . Since J1 ✁ H, we see that (W2 )J1 is a direct sum of
nontrivial kJ1 -modules. On the other hand, the p′ -group J1 acts trivially on W1 and on
W1∗ . Setting M := W1∗ ⊗k W2 , we then have that M J1 = 0 and so
Ext1H (W1 , W2 ) ∼
= H 1 (H, M ) ∼
= H 1 (H/J1 , M J1 ) = 0,
a contradiction.
(ii) We have shown that J ∩ K1 = J ∩ K2 = 1. Hence
∼ K1 J/J ✁ H/J = S1 × . . . × Sn ,
K1 =
Q
and so K1 is isomorphic to the direct product i∈I Si for some subset I ⊆ {1, 2, . . . , n}.
As |H/K1 |p = p and p | |Si | for all i, we may assume that
(5.2)
K1 ∼
= JK1 /J = S2 × . . . × Sn .
In particular, if n = 1 then K1 = 1 and similarly K2 = 1, whence we are done.
(iii) Now we assume that n ≥ 2. Consider X := JK2 ∩ K1 . Then X ∩ K2 ≤ K1 ∩ K2 = 1
and so
X∼
= XK2 /K2 ✁ JK2 /K2 ∼
=J
Q
′
is a p -group. On the other hand, X ✁ K1 and so by (5.2) we again have that X ∼
= i∈I ′ Si
for some subset I ′ ⊆ {2, . . . , n}. As p | |Si | for all i, we conclude that X = 1. Similarly
JK1 ∩ K2 = 1. Together with (5.2), this implies that
K2 ֒→ H/JK1 ∼
= (H/J)/(JK1 /J) ∼
= S1 .
Q
Furthermore, as shown in (ii), JK2 /J ∼
= K2 ∼
= i∈I ′′ Si for some subset I ′′ ⊆ {1, 2, . . . , n}
of cardinality n − 1. It follows that n = 2, K2 ∼
= S1 , K1 ∼
= S2 , and
H = JK2 K1 ∼
= JK2 × K1 = (J × K2 ) × K1 ∼
= J × K1 × K2 .
Now we can write
W1 ∼
= A1 ⊗k k ⊗k B2 , W2 ∼
= A2 ⊗k B1 ⊗k k,
22
R. GURALNICK, F. HERZIG, AND P. TIEP
where A1 , A2 ∈ IBrp (J) and Bi ∈ IBrp (Ki ) for i = 1, 2. In this case, if B1 ∼
6 k and B2 ∼
6 k,
=
=
1
then ExtH (W1 , W2 ) = 0 by Lemma 3.3(i), a contradiction. So we may assume that B1 ∼
= k,
i.e. K1 acts trivially on W2 . It then follows that K1 ≤ K1 ∩ K2 = 1, contradicting (5.2)
and the equality n = 2.
Proof of Theorem 1.6. We take the convention that V ǫ is V for ǫ = + and V ∗ if ǫ = −,
and the same holds for other modules. Assume that Ext1G (V, V ǫ ) 6= 0 for some ǫ = ±.
Decompose
t
t
M
M
Wiǫ
Wi , VGǫ + = e
VG + = e
i=1
i=1
where W1 , . . . , Wt are pairwise non-isomorphic and G-conjugate irreducible kG+ -modules.
By assumption, the image of G+ in each GL(Wi ) has Sylow p-subgroups of order p, and
G+ has no composition factor of order p. By Proposition 5.3,
(5.3)
G+ /Op′ (G+ ) = S1 × . . . × Sn
is a direct product of non-abelian simple groups of order divisible by p.
′
(i) First we consider the case k = k. Recall that G+ = Op (G+ ). So by Proposition 5.4
1
we have that ExtG+ (Wi , Wjǫ ) = 0, unless G+ has the same kernel on Wi and Wjǫ .
Let Ki denote the kernel of G+ on Wi , and on Wiǫ as well. Relabeling W1 , . . . , Wt , we
may assume that K1 , . . . , Ks are pairwise distinct, with s := |{K1 , . . . , Kt }|. Defining
M
Vi := e
Wj
j:Kj =Ki
for 1 ≤ i ≤ s, we then have that
V = V1 ⊕ . . . ⊕ Vs , V ǫ = V1ǫ ⊕ . . . ⊕ Vsǫ .
Certainly, G acts transitively on {W1 , . . . , Wt }, {V1 , . . . , Vs }, and on {K1 , . . . , Ks } via conjugation. Also, H := NG (K1 ) ✄ G+ stabilizes V1 and has index s in G. It follows that
H = StabG (V1 ), V ∼
= IndG
H (V1 ), and so V1 is an irreducible kH-module.
By the definition of Vi , we have that Ext1G+ (V1 , Viǫ ) = 0 for all i > 1. As H/G+ is a
L
p′ -group, it follows that Ext1H (V1 , i>1 Viǫ ) = 0. Now by Frobenius’ reciprocity,
ǫ ∼
1
ǫ
1
ǫ
Ext1G (V, V ǫ ) = Ext1G (IndG
H (V1 ), V ) = ExtH (V1 , (V )H ) = ExtH (V1 , V1 ).
Recall that K1 acts trivially on V1 and V1ǫ , and |H/K1 |p = |G+ /K1 |p = p. Hence
dimk Ext1H/K1 (V1 , V1ǫ ) ≤ 1 by Lemma 5.1. Finally, K1 has no composition factor of order p by (5.3). So Ext1H (V1 , V1ǫ ) ∼
= Ext1H/K1 (V1 , V1ǫ ) by Lemma 3.4, and so we are done.
(ii) Now we consider the general case. By [32, Theorem 9.21], W1 ⊗k k is a direct sum
of irreducible kG+ -modules W11 , . . . , W1m , which form a Galois conjugacy class over k. By
assumption, |G+ /K1 |p = p, where K1 is the kernel of G+ on W1 . Certainly, K1 is contained
in the kernel K11 of the action of G+ on W11 , whence |G+ /K11 |p ≤ p. If |G+ /K11 |p < p,
′
then the equality G+ = Op (G+ ) implies that K11 = G+ , whence G+ acts trivially on W11
and so on W1 and on V as well, contradicting the faithfulness of V and the assumption
Ext1G (V, V ǫ ) 6= 0. Hence we must have that |G+ /K11 |p = p. Since the dimension of
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
23
Ext1G (V, V ǫ ) does not change under field extensions, we are done by replacing V by V ⊗k k
and applying the result of (i).
A key ingredient of the proof of Theorem 1.5 is the following statement:
′
Proposition 5.5. Let X be a finite group with a normal subgroup Y ≥ Op (X). Let A, B,
W , W ′ be kX-modules, where A and B are absolutely irreducible and Y acts via scalars on
both A and B. Suppose in addition that AY ∼
= BY . Then
dimk Ext1X (A ⊗k W, B ⊗k W ′ ) ≤ dimk Ext1Y (WY , WY′ ).
Proof. Since the dimensions of Ext1 -spaces do not change under field extensions, we may
assume that k is algebraically closed. By assumption, X/Y is a p′ -group and Y acts trivially
on A∗ ⊗k B. Without loss we may assume that dimk B ≤ dimk A. Denoting
H := H 1 (Y, (WY )∗ ⊗k W ′ ) ∼
= Ext1 (WY , W ′ )
Y
Y
Y
we then have dimk B ⊗k H ≤ (dimk A)(dimk H) and so
dimk HomkX (A, B ⊗k H) ≤ dimk H = dimk Ext1Y (WY , WY′ ).
Applying the inflation-restriction spectral sequence, we obtain
dimk Ext1X (A ⊗k W, B ⊗k W ′ ) = dimk (H 1 (Y, A∗ ⊗k B ⊗k W ∗ ⊗k W ′ ))X/Y
= dimk (A∗ ⊗k B ⊗k H)X/Y
= dimk HomkX (A, B ⊗k H)
≤ dimk Ext1Y (WY , WY′ ).
Proposition 5.6. Let k be an algebraically closed field of characteristic p. Assume the
hypothesis of Theorem 1.2 and write VG+ = ⊕ti=1 Vi , where Vi ∼
= eWi and W1 , . . . , Wt are
+
pairwise non-isomorphic irreducible kG -modules. Suppose that there is a unique j ≥ 1
such that Ext1G+ (W1 , Wj ) 6= 0. Then
dimk Ext1G (V, V ) ≤ dimk Ext1G+ (W1 , Wj ).
Proof. Let G1 := StabG (V1 ) be the inertia group of W1 in G. Since G+ ✁G1 , the uniqueness
of j implies that G1 = StabG (Vj ) as well. Next, V ∼
= IndG
G1 ((V1 )G1 ), and so
1
∼
Ext1G (V, V ) = Ext1G (IndG
G1 ((V1 )G1 ), V ) = ExtG1 ((V1 )G1 , VG1 )
L
1
1
∼
= ExtG1 ((V1 )G1 , (Vj )G1 ) ⊕ ExtG1 ((V1 )G1 , i6=j (Vi )G1 ).
L
Since G+ contains a Sylow p-subgroup of G1 , Ext1G1 ((V1 )G1 , i6=j (Vi )G1 ) injects in
M
M
Ext1G+ ((V1 )G+ ,
(Vi )G+ ) ∼
Ext1G+ (W1 , Wi ) = 0
= e2
i6=j
i6=j
and so it is zero.
It remains therefore to show that
dimk Ext1G1 ((V1 )G1 , (Vj )G1 ) ≤ dimk Ext1G+ (W1 , Wj ).
24
R. GURALNICK, F. HERZIG, AND P. TIEP
Let X denote a universal p′ -cover of G1 (so that G1 ∼
= X/Z for some p′ -subgroup Z ≤
′
p
Z(X) ∩ [X, X]), and let Y := O (X). Now we view V1 as an irreducible kX-module by
inflation and note that
dimk Ext1G1 ((V1 )G1 , (Vj )G1 ) = dimk Ext1X ((V1 )X , (Vj )X )
as Z is a p′ -group. Since Z acts trivially on V1 , we also have that (V1 )Y ∼
= e(W1 )Y and also
Y Z/Z ∼
= G+ . Hence (W1 )Y is irreducible, and similarly for Wj . Also,
dimk Ext1Y ((W1 )Y , (Wj )Y ) = dimk Ext1G+ (W1 , Wj ).
Fix a basis of W1 and the corresponding representation Φ of Y on W1 in this basis. By
the Clifford theory, we can decompose the irreducible representation Θ of X on V1 as a
tensor product of an irreducible projective representation Λ of X/Y (of degree e) and an
irreducible projective representation Ψ of X, with
Ψ(y) = Φ(y)
for all y ∈ Y . Since X is
p′ -centrally
closed, there is a function f : X → k× such that
Ψ′ : x 7→ f (x)Ψ(x)
is a linear representation. Note that fY ∈ Hom(Y, k × ) since ΨY = Φ is a linear represen′
tation, and so fY = 1Y as Y = Op (Y ). In particular, Ψ′ (y) = Φ(y) for all y ∈ Y . Now we
inflate Λ to a projective representation of X and define
Λ′ : x 7→ f (x)−1 Λ(x)
so that Θ(x) = Λ′ (x) ⊗ Ψ′ (x) for all x ∈ X. Then Λ′ is also a linear representation of X
and furthermore Λ′Y is trivial (since fY = 1Y ). Thus we can decompose
(V1 )X = A ⊗k W,
where the kX-modules A and W are irreducible, Y acts trivially on A, and WY
Similarly,
(Vj )X = B ⊗k W ′ ,
where the kX-modules B and W ′ are irreducible, Y acts trivially on B, and WY′
Now our statement follows by applying Proposition 5.5.
∼
= (W1 )Y .
∼
= (Wj )Y .
The same proof as above yields:
Proposition 5.7. Let k be an algebraically closed field of characteristic p. Assume the
hypothesis of Theorem 1.2 and write VG+ = ⊕ti=1 Vi , where Vi ∼
= eWi and W1 , . . . , Wt are
pairwise non-isomorphic irreducible kG+ -modules. Suppose that there is a unique j ≥ 1
such that Ext1G+ (W1 , Wj∗ ) 6= 0. Then
dimk Ext1G (V, V ∗ ) ≤ dimk Ext1G+ (W1 , Wj∗ ).
✷
Lemma 5.8. Given the assumption of Theorem 1.5, suppose that H is as in the extraspecial
case (e) of Theorem 2.1. Then
Ext1G (V, V ∗ ) = 0.
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
25
L
Proof. Write VG+ = e ti=1 Wi as usual. It suffices to show that Ext1G+ (Wi , Wj∗ ) = 0 for
all i, j. Recall that J := Op′ (G) acts irreducibly on Wi and Wj∗ by [23, Theorem 2.4(ii)].
Since J is a p′ -group, we have M = CM (J) ⊕ [M, J] for M := Wi∗ ⊗ Wj∗ . As J has no fixed
point on [M, J], H 1 (G+ , [M, J]) = 0. Also,
CM (J) ∼
= HomJ (Wi , W ∗ )
j
is either 0 or k. Hence
Ext1G+ (Wi , Wj∗ ) ∼
= H 1 (G+ , M ) ∼
= H 1 (G+ , CM (J)) ֒→ H 1 (G+ , k)
As G+ is perfect by [23, Theorem 2.4], H 1 (G+ , k) = 0, and so we are done.
Proof of Theorem 1.5. (i) Assume that (G, V ) satisfies all the hypotheses of Theorem 1.5.
We take the convention that V ǫ is V for ǫ = + and V ∗ if ǫ = −, and the same holds for other
modules. Since the dimension of Ext1G (V, V ǫ ) does not change under field extensions, we
will assume that k = k. Assume that Ext1G (V, V ǫ ) 6= 0 for some ǫ = ±. It suffices to show
that G then fulfills the conditions of Propositions 5.6 and 5.7 (with Ext1G+ (W1 , Wjǫ ) ∼
=k
for the index j indicated in these propositions). By [23, Lemma 7.2], there is some j such
that
(5.4)
Ext1G+ (W1 , Wjǫ ) 6= 0.
Note that Ext1G (V, V ǫ ) = 0 in the extra special case (e) of Theorem 2.1, by [23, Proposition
10.4] and Lemma 5.8. So we may assume that the image H of G+ in GL(W ) is a central
product of quasisimple groups, whence, by [23, Theorem 2.4],
G+ = L1 ∗ . . . ∗ Ln
is also a central product of quasisimple groups Li . Moreover, if some Li is not a quasisimple group of Lie type in characteristic p, then by [23, Theorem 2.4], the image of G+ in
each GL(Wi ) has Sylow p-subgroups of order p, and so Theorem 1.6 applies. So in what
follows we may assume that all Li are quasisimple groups of Lie type in characteristic p.
Correspondingly, we can decompose
W1 = A1 ⊗ . . . ⊗ An , Wjǫ = B1 ⊗ . . . ⊗ Bn ,
where Ai and Bi are irreducible kLi -modules and Li′ acts trivially on Ai and Bi whenever
i′ 6= i. By Lemma 3.3 and (5.4), we may assume that
Ai ∼
= Bi
for i > 1, and furthermore Ext1L1 (A1 , B1 ) 6= 0. Since dimk W1 = dimk Wj , it follows that
dimk Ai = dimk Bi for all i.
Note that if dimk Ai = 1, then Ai ∼
= k as Li is perfect, and similarly Bi ∼
= k, whence
1
1
ExtLi (Ai , Bi ) = 0. In fact, ExtLi (Ai , Bi ) = 0 if dimk Ai ≤ (p − 3)/2 by the main result of
[19]. It follows that dimk A1 ≥ (p − 1)/2. Since dimk W1 ≤ p − 1, we arrive at two possible
cases:
(a) dimk Ai = 1 (and so Ai ∼
= Bi ∼
= k) for all i > 1; or
(b) p ≥ 5, dimk Ai = 1 (and so Ai ∼
= Bi ∼
= k) for all i > 2, and {dimk A1 , dimk A2 } =
{(p − 1)/2, 2}.
26
R. GURALNICK, F. HERZIG, AND P. TIEP
(ii) Suppose we are in case (b). Then the quasisimple group Lm (for some m ∈ {1, 2}) is
acting irreducibly on Am ∼
= k2 . As Lm is a Lie-type group in characteristic p, we have that
a
∼
Lm = SL2 (p ) for some a ≥ 1. By Lemma 4.3, L1 ∼
= SL2 (p) (modulo a central subgroup),
∼
∼
dim A1 = (p − 1)/2, Ext1L1 (A1 , B1 ) ∼
k,
and
A
=
1 = B1 . We have shown that Ai = Bi for
1
ǫ
∼
all i; in particular Wj = W1 . Now we have that m = 2, and dimk ExtL2 (A2 , B2 ) equals 0 if
pa > 5 and 1 if pa = 5, see [23, Lemma 8.1]. Again by Lemma 3.3,
dimk Ext1G+ (W1 , Wjǫ ) = 1 + dimk Ext1L2 (A2 , B2 ).
In the case pa = 5, we have (dim W, H) = (4, Ω+
4 (5)) and conclude by Propositions 5.6 and
5.7 that Ext1G (V, V ) and Ext1G (V, V ∗ ) are at most 2-dimensional. Moreover, Example 5.9(i)
shows that the upper bound 2 can indeed be attained. If pa > 5, then Ext1G+ (W1 , Wjǫ ) ∼
=k
and Wj ∼
= W1ǫ for any j satisfying (5.4).
(iii) Now we consider the case (a). Then
∼ Ext1 + (W1 , W ǫ ) 6= 0
(5.5)
Ext1 (A1 , B1 ) =
L1
j
G
by Lemma 3.3.
Suppose first that p = 3. Then L1 ∼
= SL2 (3a ) for some a ≥ 2, and
W1 = A1 ⊗k k ⊗k . . . ⊗k k, Wj = B1ǫ ⊗k k ⊗k . . . ⊗k k.
We may also assume that A1 is the natural kL1 -module. If a = 2, then by (5.5) and
(3)
[2, Corollary 4.5] we have that B1 is isomorphic to the Frobenius twist A1 of A1 , and
Ext1L1 (A1 , B1 ) ∼
= k2 . Thus Wj is uniquely determined, and so dimk Ext1G (V, V ǫ ) ≤ 2 by
Propositions 5.6 and 5.7. Suppose now that a > 2. Since G1 := StabG (V1 ) stabilizes
the isomorphism class of W1 , we see that G1 normalizes each of L1 and L2 ∗ . . . ∗ Ln ,
and induces an inner-diagonal automorphism of L1 . Next, by (5.5) and [2, Corollary 4.5]
(3)
(3a−1 )
we have that B1 is isomorphic to one of the Frobenius twists A1 , A1
of A1 , and
Ext1L1 (A1 , B1 ) ∼
= k. Thus there are at most two possibilities for Wj , each stabilized by G1 .
If only one of them occurs among the submodules Wi , then we have dimk Ext1G (V, V ǫ ) ≤ 1
by Propositions 5.6 and 5.7. Suppose that both of them occur, say for j1 and j2 . It follows
that G1 = StabG (Vj1 ) = StabG (Vj2 ) and furthermore both Vj1 and Vj2 are irreducible over
G1 . Then, arguing as in the proof of Proposition 5.6 we have
1
ǫ ∼
ǫ
Ext1G (V, V ǫ ) = Ext1G (IndG
G1 ((V1 )G1 ), V ) = ExtG1 ((V1 )G1 , VG1 )
∼
= Ext1 ((V1 )G , (V ǫ )G ) ⊕ Ext1 ((V1 )G , (V ǫ )G )
G1
1
j1
1
G1
1
j2
1
has dimension at most 2. In fact, Example 5.9 shows that the upper bound 2 can indeed
be attained.
Suppose now that p > 3. Then by Lemma 4.3, L1 = SL2 (p) (modulo a central subgroup),
A1 ∼
= B1 , Ext1L1 (A1 , B1 ) ∼
= k, Wj ∼
= W1ǫ , and Ext1G+ (W1 , Wjǫ ) ∼
= k.
(iv) We have shown that in the case of Theorem 1.5(i), there is a unique j such that
Ext1G+ (W1 , Wjǫ ) 6= 0, in which case it has dimension 1. Hence we are done by Propositions
5.6 and 5.7.
Example 5.9. (i) Let p = 5 and let S = L1 × L2 , with Li ∼
= SL2 (5), be acting on V =
W1 ⊗ W2 , where Wi ∼
= k2 is an irreducible kLi -module and Li acts trivially on W3−i . Note
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
27
that the kernel of this action is the diagonal cyclic subgroup Z ∼
= C2 of Z(L1 )× Z(L2 ). Now
G = G+ := S/Z ∼
(5)
acts
faithfully
and
irreducibly
on
V
, and dimk Ext1G (V, V ) = 2
= Ω+
4
∗
∼
by Lemma 3.3. Also, V = V .
(ii) Let p = 3, S = SL2 (3a ) for some a ≥ 2 coprime to 3, W1 = k2 be the natural
i
kS-module, and let Wi+1 denote the Frobenius (W1 )(3 ) twist of W1 for 1 ≤ i ≤ a − 1. Then
G = S ⋊ hσi (with σ being the field automorphism of S, of order a) acts irreducibly and
faithfully on V = W1 ⊕ . . . ⊕ Wa , G+ = S, and
a
M
1
∼
Ext1S (W1 , Wi ) ∼
(5.6)
ExtG (V, V ) =
= k2
i=1
by [2, Corollary 4.5]. (Indeed, if a = 2 then Ext1S (W1 , W2 ) ∼
= k2 . If a ≥ 3, then
Ext1S (W1 , W2 ) ∼
= Ext1S (W1 , Wa ) ∼
= k. All other summands in the middle term of (5.6)
are zero.) Also, V ∼
= V ∗.
(iii) Let p = 2f + 1 be a Fermat prime and let H = Op′ (H)P (with P ∼
= Cp and
1+2f
p−1
∼
as in case (i) of
Op′ (H) = 2− ) acting faithfully and absolutely irreducibly on W1 = k
Theorem 2.1. Note that the kH-module W1 is self-dual. Let n be coprime to p and let
G = H1 ≀ Cn = (H1 × . . . × Hn ) ⋊ Cn
∼
with Hi = H1 = H, so that G+ = H1 × . . . × Hn . Inflate W1 to a kG+ -module and consider
+
V := IndG
G+ (W1 ). Note that J := Op′ (G ) acts absolutely irreducibly on W1 , and
W1∗ ⊗ W1 = CW1∗ ⊗W1 (J) ⊕ [W1∗ ⊗ W1 , J]
with CW1∗⊗W1 (J) ∼
= k. Since G+ /J ∼
= Cpn , it now follows that
Ext1G+ (W1 , W1 ) ∼
= H 1 (G+ , W1∗ ⊗ W1 ) ∼
= H 1 (G+ /J, k) ∼
= kn .
On the other hand, the actions of J on W1 and Wj have different kernels for any j > 1,
and so Ext1G+ (W1 , Wj ) = 0. Hence V ∼
= V ∗ and
Ext1 (V, V ) ∼
= Ext1 + (W1 , W1 ) ∼
= Ext1 + (W1 , VG+ ) ∼
= kn .
G
G
G
Next we strengthen Theorem 1.5 in the case dim W is small.
Theorem 5.10. Let k be a field of characteristic p and let V and V ′ be absolutely irreducible faithful kG-modules. Suppose that dimk W + dimk W ′ ≤ p − 2, where W and W ′
are irreducible kG+ -submodules of V and V ′ , respectively. Then H 1 (G, M ) = 0 for any
subquotient M of the G-module V ⊗ V ′ .
Proof. It suffices to prove H 1 (G+ , M ) = 0. Note that VG+ = ⊕ti=1 Wi and VG′ + = ⊕sj=1 Wj′
with Wi , Wj′ ∈ IBrp (G+ ), and Op (G+ ) ≤ Op (G) = 1. Since
(5.7)
dimk Wi + dimk Wj′ ≤ p − 2,
by the main result of [19] we have Ext1G+ (Wi∗ , Wj′ ) = 0. It follows that Ext1G+ (VG∗+ , VG′ + ) =
0, i.e. H 1 (G+ , (V ⊗ V ′ )G+ ) = 0. By Corollary 1 to [55, Theorem 1], (5.7) also implies
that the G+ -module V ⊗ V ′ is semisimple. Thus M is isomorphic to a direct summand of
(V ⊗ V ′ )G+ , whence H 1 (G+ , M ) = 0, as desired.
28
R. GURALNICK, F. HERZIG, AND P. TIEP
Corollary 5.11. Let k be a field of characteristic p and let V be an absolutely irreducible
faithful kG-modules. Suppose that dimk W < (p − 1)/2 for any irreducible kG+ -submodule
of V . Then H 1 (G, Sym2 (V )) = H 1 (G, ∧2 (V )) = 0.
✷
6. Modules of dimension p
Let p be a prime and let k be algebraically closed of characteristic p. The aim of
this section is to show that if G is an irreducible subgroup of GLp (k) = GL(V ), then
almost always (G, V ) is adequate (using Thorne’s new definition). We begin with some
observations.
Remark 6.1. Suppose that G ≤ GL(V ) is a finite irreducible subgroup. Note that, to show
(G, V ) is adequate it suffices to show that G+ is adequate on V . Indeed, any subgroup
being weakly adequate implies that the spanning condition holds for G. Next, adequacy
for any subgroup containing a Sylow p-subgroup of G implies that necessary vanishing of
H 1 for G.
Lemma 6.2. Let G be a finite group with a Sylow p-subgroup P of order p and let V ∈
IBrp (G) be such that p | (dim V ). Then V is projective.
Proof. Assume that V is non-projective and set N := NG (P ). By the Green correspondence
[29, Lemma 4.1.1], in this case we have VN = W ⊕ M , where W is a non-projective
indecomposable N -module and M is a projective N -module (or zero). Now W belongs to
an N -block b of defect 1. By [29, Lemma 4.2.14], W is a uniserial (non-projective) quotient
of P(U ) where U := head(W ) ∈ IBrp (N ). By [29, Lemma 4.2.13], P(U ) has length p,
so W has length l < p. According to [29, Remark 4.2.11], all simple kN -modules in b
are of the same dimension d and have P in their kernel. It follows that d divides |N/P |
and so d is coprime to p. Hence p ∤ dl = dim W and so p ∤ (dim V ) (as p | (dim M )), a
contradiction.
Lemma 6.3. Let G be a finite group with a cyclic Sylow p-subgroup P and p = char(k).
Suppose that G = Op (G). Then H 1 (G, k) = H 2 (G, k) = 0.
Proof. The vanishing of H 1 (G, k) is obvious. Suppose that H 2 (G, k) 6= 0. Since the dimension of H 2 does not change under extension of scalars, we may assume that H 2 (G, Cp ) 6= 0.
As moreover H 1 (G, Cp ) = 0, it follows that p divides the order of the Schur multiplier of
G. It is well known that the latter then implies that Sylow p-subgroups of G are non-cyclic
(see e.g. [32, Corollary (11.21)]).
Next we give an example showing that for modules of dimension 2p, we can satisfy all
conditions aside from the spanning condition.
Example 6.4. Assume that p > 2. Let C be a nontrivial cyclic group of order coprime to p,
with a faithful character λ : C → k× , and let G = C ≀D where D is a dihedral group of order
2p. Let V be the irreducible kG-module of dimension 2p induced from the 1-dimensional
representation with character
λ ⊗ 1C ⊗ . . . ⊗ 1C
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
29
of the abelian normal subgroup A = C × C × . . . × C ∼
= C 2p of G. Let E be the unique
subgroup of D of order p. Note that V = V1 ⊕ V2 , where the Vi are irreducible AEsubmodules of V (of dimension p). Then the following statements hold:
(i) H 1 (G, k) = H 2 (G, k) = 0 by Lemma 6.3;
(ii) Ext1G (V, V ) = 0 (indeed, V is projective by Lemma 6.2);
(iii) The span M of the p′ -elements of G in End(V ) is precisely A ⊕ Hom(V1 , V2 ) ⊕
Hom(V2 , V1 ), where A is the image of kA in End(V1 ) ⊕ End(V2 ).
Now we describe all irreducible linear groups of degree p:
Proposition 6.5. Let k be an algebraically closed field of characteristic p and let G <
GLp (k) be a finite irreducible subgroup. Then one of the following holds:
(i) G is imprimitive on W := kp , G < GL1 (k) ≀ Sp , and furthermore A := G ∩ GL1 (k)p
is non-central in G;
(ii) G is almost quasisimple. Furthermore, H := G(∞) is quasisimple of order divisible
by p acting irreducibly on W , and so (H, W ) is as described in Theorem 2.2.
Proof. By the hypothesis, G acts irreducibly on W = kp . Suppose the action is imprimitive.
Then G permutes transitively the p summands of a decomposition W = W1 ⊕ . . . ⊕ Wp ,
with kernel say A. If A 6≤ Z(G), we arrive at (i). Assume that A ≤ Z(G). Note that
S := G/A is a transitive subgroup of Sp , and so we can apply the main result of [65] to
S. In particular, if S is solvable, then S = P : C with P ∼
= Cp and C ≤ Cp−1 . Then AP
is a normal abelian subgroup of G, whence by Ito’s theorem the degree of any χ ∈ Irr(G)
divides [G : AP ] | (p − 1). On the other hand, G is solvable, and so by the Fong-Swan
theorem, W lifts to an irreducible complex module of dimension p, a contradiction. Thus
S is non-solvable, which implies by [65] that S is almost simple, G is almost quasisimple,
and H := G(∞) is a normal subgroup of index coprime to p. Since dim W = p, the last
condition also implies that H is irreducible on W and so we arrive at (ii).
We may now assume that the G-module W is primitive. Since dim(W ) = p is prime, this
module cannot be tensor decomposable nor tensor induced. Now we can apply Aschbacher’s
theorem in the version given in [28, Proposition 2.8] to (G, W ) to conclude that G is almost
quasisimple: S ✁ G/Z(G) ≤ Aut(S) for some non-abelian simple group S. In particular,
H = G(∞) ✁ G is quasisimple, and moreover irreducible on W by [28, Lemma 2.5]. Hence
we can apply Theorem 2.2 to (H, W ).
6.1. Imprimitive case.
Proposition 6.6. Suppose we are in case (i) of Proposition 6.5. Then (G, W ) is adequate
if and only if |G/A| =
6 p.
Proof. Let P ∈ Sylp (G), so that |P | = p (if P = 1 then G cannot act irreducibly on W ). If
|G/A| = p, then G = AP , A = Op′ (G) contains all the p′ -elements of G but does not act
irreducibly on W (as A is a p′ -group), whence G is not weakly adequate.
Now assume that |G/A| 6= p; in particular, p > 2. Suppose that G has a normal pcomplement K. Then K = Op′ (G) > A (as otherwise G = AP and so |G/A| = p), and
H := G/A ≤ Sp has a normal p-complement K/A 6= 1. Thus H is a transitive subgroup of
Sp with Cp as a composition factor. But then Op′ (H) = 1 by Lemma 4.8, a contradiction.
30
R. GURALNICK, F. HERZIG, AND P. TIEP
Thus G cannot have a normal p-complement, and so H i (G, k) = 0 for i = 1, 2 by Lemma
6.3. Also, W is projective as a G-module by Lemma 6.2, whence Ext1G (W, W ) = 0.
Since A 6≤ Z(G), A has p distinct eigenspaces W1 , . . . , Wp on W permuted transitively
by P . Thus, it remains only to prove that
End(W ) ∼
= ⊕1≤i,j≤p Hom(Wi , Wj )
is spanned by the images of the p′ -elements of G. Given 1 ≤ i 6= j ≤ p, we claim that there
exists a p′ -element x ∈ G with xWi = Wj . Since P is transitive on {W1 , . . . , Wp }, we can
choose y ∈ P with yWi = Wj . Note that N acts on this set as the Frobenius subgroup
Cp ⋊ Cs of Sp , with kernel A ∩ N , and all the elements of (Cp ⋊ Cs ) \ Cp are p′ -elements. So,
since s > 1, we can find z ∈ N \ AP such that zWj = Wj and set x := zy. Then xWi = Wj
and x ∈ N \ AP , whence x is a p′ -element.
Now B := hA, xi is a p′ -group. Note that WA = ⊕pa=1 Wa is a direct sum of p nonisomorphic simple A-submodules. Hence WB = ⊕tb=1 Ub is a direct sum of t ≥ 1 nonisomorphic simple B-modules, with U1 ⊇ Wi ⊕ Wj . By the Artin-Wedderburn theorem,
the image of kB in End(W ) is just ⊕tb=1 End(Ub ), and so it contains
End(U1 ) ⊇ End(Wi ) ⊕ End(Wj ) ⊕ Hom(Wi , Wj ) ⊕ Hom(Wj , Wi ),
and the result follows.
6.2. Chevalley groups in characteristic p. We first point out the following:
Proposition 6.7. Let H be a quasisimple finite group of Lie type in characteristic p. Let k
be an algebraically closed field of characteristic p. Let W be a faithful irreducible kH-module
of prime dimension r ≤ p. Then one of the following statements holds:
(i) H = SL2 (pa ) for r = 2 and H = PSL2 (pa ) for r > 2;
(ii) H = SLr (pa ) or SUr (pa ), and r > 2;
(iii) H = Ωr (pa ) and r ≥ 5; or
(iv) r = 7 and H = G2 (pa ).
Proof. Since char(k) = p and V is faithful, Op (H) = 1. Hence there is a simple simply
connected algebraic group G in characteristic p and a Frobenius endomorphism F : G → G
such that H ∼
= G/Z for G := G F and Z ≤ Z(G). Inflate W to a kG-module. Since
r = dim W is prime, W is tensor indecomposable and in particular is a twist of a restricted
representation. So we may assume that W is restricted and extend W to a kG-module.
By [36], it follows that W = L(λ) where λ is a dominant weight, and dim W equals the
dimension of the Weyl module V (λ) labeled by λ. Thus, we can apply the same result for
characteristic 0 which was proved by Gabber [39, 1.6].
Proposition 6.8. Suppose we are in case (ii) of Proposition 6.5. Assume in addition that
H = G(∞) is a quasisimple group of Lie type in characteristic p > 3. Then (G, W ) is
adequate.
Proof. By the hypothesis, we have that H is quasisimple and irreducible on W . So we
can apply Proposition 6.7 to H; in particular we have that WH = L(λ) is a restricted
module (up to a Frobenius twist; in what follows we will ignore this twist). In the case
H = PSL2 (pa ), we have that λ = (p − 1)̟1 , where ̟1 is the fundamental weight. Since
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
31
WH is G-invariant, we see that G cannot induce nontrivial field automorphisms on H; in
particular, G+ = H. In other cases, applying Propositions 5.4.11 and 5.4.12 of [40], we
see that WH ∼
= N or N ∗ where N is the natural kH-module of dimension p (with highest
weight ̟1 ), and again G+ = H.
By Remark 6.1, without loss we may now assume G = H. Note that all the classical
groups given in the previous proposition when r = p contain an irreducible subgroup
L∼
= PSL2 (p). Indeed, the irreducible kL-representation of degree p embeds L in M ∼
= Ωp (p).
In turn, M embeds in SLp (q) and SUp (q) for any q = pa . The same is true for G2 (p) with
p = 7: G2 (7) > G2 (2) > PSL2 (7). (It is well known, see e.g. [41] that H = G2 (7) contains
a maximal subgroup X ∼
= G2 (2) which acts irreducibly on the minimal 7-dimensional Hmodule W . Next, X contains a maximal subgroup Y ∼
= PSL2 (7), cf. [10]. Using [34] one
can check that Y is irreducible on W .) Thus weak adequacy follows by [23, Proposition
3.1].
It is well known that H 1 (G, k) = H 2 (G, k) = 0 (since p > 3). Thus, it suffices to show
that Ext1G (W, W ) = 0. If G = PSL2 (pa ), the result follows by [2]. If G = Ω5 (5), one
computes directly that Ext1G (W, W ) = 0 (this was done by Klaus Lux). In all other cases,
Ext1G (W, W ) = 0 by the main result of [48].
6.3. Remaining cases.
Lemma 6.9. Let k = k̄, H = Ap+1 with p ≥ 5, and let W be an irreducible kH-module of
dimension p. Then (H, W ) is weakly adequate.
Proof. Note that W is irreducible over a subgroup L ∼
= PSL2 (p) of H. Hence the claim
follows by [23, Proposition 3.1].
We record the following useful observation:
Lemma 6.10. Let X be a finite p′ -subgroup of G < GL(W ) where W is a finite dimensional
vector space over k. Suppose that WX is multiplicity-free. Then (End(W )/M)X = 0.
Proof. Note that the X-module End(W ) is semisimple. Furthermore, the multiplicity-free
assumption implies M ⊇ End(W )X by the Artin-Wedderburn theorem. Hence the claim
follows.
Proposition 6.11. Let k = k̄, H = PSp2n (q) with 2 < p = (q n ± 1)/2, and let W be an
irreducible kH-module of dimension p. Then (H, W ) is weakly adequate.
Proof. (a) Note that W is a Weil module and restricts irreducibly to a subgroup PSL2 (q n )
of H. So without loss we may assume n = 1. We will inflate W to a kL-module for
L := SL2 (q). Note that W is obtained by reducing modulo p one of the four complex
Weil modules of L, with characters ηi of degree (q − 1)/2 and ξi of degree (q + 1)/2, where
i = 1, 2 and ξi + ηi is a reducible Weil character of L, see e.g. [24] and [64]. Let τ denote the
permutation character of L acting on the set of all vectors of the natural module N := F2q .
Using the character table of L as given on [13, p. 155], we see that
(6.1)
(ξi + ηi )(ξ̄i + η̄i ) = τ.
32
R. GURALNICK, F. HERZIG, AND P. TIEP
Let P := StabL (hviFq ) for some 0 6= v ∈ N with normal subgroup Q := StabL (v) of order
q, and let ϕ denote the Brauer character of W . Assume the contrary: M =
6 End(W ), and
let ϑ denote the Brauer character of Q := End(W )/M.
(b) Consider the case p = (q + 1)/2, whence ϕ = ξi◦ and P is a p′ -group. Inspecting
the values of ϕP , we see that WP = W1 ⊕ W2 with W1 , W2 ∈ Irr(P ) of dimension 1 and
(q − 1)/2 > 1. Moreover, (W1 )Q is trivial, and W2Q = 0. By the Artin-Wedderburn theorem
applied to P , M ⊇ End(W1 ) ⊕ End(W2 ); in particular, dim MP ≥ 2 = dim End(W )P .
Hence we conclude that for any composition factor Y of End(W )/M, Y Q = 0 and dim Y ≤
q − 1.
Let ρ denote the permutation character of L acting on the 1-spaces of N . Then ρ =
1L + St, where St is the Steinberg character of L. Moreover, all irreducible constituents of
τ − ρ − 1L have degree q + 1 or (q + 1)/2 and thus have p-defect 0. Note that St◦ = 1L + ψ
with ψ ∈ IBrp (L) (see [7]), and ρ is the character of the PIM P(1L ) of 1L . Since ϕ = ξi◦ has
degree p and (the projective module) End(W ) contains a trivial simple submodule, we see
that End(W ) is the direct sum of P(1L ) and some p-defect 0 modules of dimension q + 1
or (q + 1)/2. In particular, since ψ is the Brauer character of the heart of P(1L ), it cannot
be afforded by a quotient of End(W ), and so ϑ 6= ψ. Since ϑ(1) ≤ q − 1, it follows that
all irreducible constituents of ϑ are of degree 1 (with multiplicity ≤ 2) and (q + 1)/2 (with
multiplicity ≤ 1). But the principal character and the Weil characters of degree (q + 1)/2
of L all contain 1Q when restricted to Q, a contradiction.
(c) Now assume that p = (q − 1)/2; in particular, ϕ = ηi◦ and q ≡ 3(mod 4). Consider a
cyclic subgroup C ∼
= C(q+1)/2 of H. It is straightforward to check that for any χ ∈ Irr(H),
either [χQ , 1Q ]Q 6= 0 or [χC , 1C ]C 6= 0. Since irreducible p-Brauer characters of H lift to
complex characters (cf. [7]), it follows that for any U ∈ IBrp (H), either U Q 6= 0, or U C 6= 0.
Now we may assume ϕ = η1◦ and observe that both ϕQ and ϕC are multiplicity-free.
Hence, each of Q, C has no nonzero fixed points on End(W )/M by Lemma 6.10. Consequently, M = End(W ).
Proposition 6.12. Let k = k̄ and let H = SLn (q), where either 3 < p = (q n − 1)/(q − 1)
or (n, p) = (2, q − 1). Let W be an irreducible kH-module of dimension p. Then (H, W ) is
weakly adequate.
Proof. Let N = he1 , . . . , en iFq denote the natural Fq H-module, and let P := StabH (he1 iFq ).
Since SL2 (4) ∼
= PSL2 (5) and SL3 (2) ∼
= PSL2 (7), we may assume (n, q) 6= (2, 4), (3, 2). Also,
let ϕ denote the Brauer character of W .
(a) First we consider the case p = (q n − 1)/(q − 1). In this case, W is induced from a
one-dimensional kP -module with character say λ. So we can write W = ⊕ω∈PN Wω as a
direct sum of one-dimensional subspaces Wω permuted transitively by H, where PN is the
set of 1-spaces in N .
(a1) Assume in addition that n ≥ 3. It suffices to show that, for any two distinct
ω1 = heiFq , ω2 = hf iFq ∈ PN ,
(6.2)
M ⊇ End(Wω1 ) ⊕ Hom(Wω1 , Wω2 ) ⊕ Hom(Wω2 , Wω1 ).
Since H acts transitively on those pairs (ω1 , ω2 ), we may assume that e = e1 and e = e2 .
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
33
Consider an opposite parabolic subgroup R := StabH (N1 ) for N1 := he1 , . . . , en−1 iFq ,
which is a p′ -subgroup. Then R stabilizes the subspace W1 := ⊕ω∈PN1 Wω of dimension
(q n−1 − 1)/(q − 1). Note that the unipotent radical Q of R acts trivially on W1 . Indeed,
q ≥ 3 since we are assuming n ≥ 3 and (n, q) 6= (3, 2). Now the Levi subgroup L :=
StabH (N1 , hen iFq ) of R acts transitively on q n−1 −1 > dim W1 nontrivial linear characters of
Q, whence the claim follows. Now we identify L with GLn−1 (q) via diag(X, det(X)−1 ) 7→ X.
Then the L-character of W1 is just induced from the character λP ∩L 6= 1P ∩L of the maximal
parabolic subgroup P ∩ L (of index (q n−1 − 1)/(q − 1)) of L. Hence W1 is an irreducible Weil
module of dimension (q n−1 − 1)/(q − 1) for L, and it is irreducible over R. Note that λ is a
linear character of order dividing q − 1 = |P/P ′ | and so it takes value 1 on any unipotent
element of P . Using this, one can check that ϕ(t) = (q n−1 − 1)/(q − 1) for any 1 6= t ∈ Q.
In particular,
ϕQ =
q n−1 − 1
· 1Q +
q−1
X
ν = (dim W1 + 1) · 1Q +
X
ν.
1Q 6=ν∈Irr(Q)
ν∈Irr(Q)
It now follows (by Clifford’s theorem) that WR = W1 ⊕ W2 ⊕ W3 , where W2 has dimension
1, W Q = W1 ⊕ W2 , and W3 is irreducible of dimension q n−1 − 1. Applying the ArtinWedderburn theorem, we see that M ⊇ End(W1 ). Since e1 , e2 ∈ N1 , (6.2) follows.
(a2) Assume now that n = 2, and so p = q + 1 ≥ 17. In this case, ϕ is real, and so W is
self-dual and supports a non-degenerate H-invariant symmetric bilinear form (·, ·). Write
P = QT where Q is elementary abelian of order q and T ∼
= Cq−1 . We also consider another
parabolic subgroup P ♯ = Q♯ T := StabH (he2 iFq ), with T = P ∩ P ♯ . Letting ρ denote the
regular character of T and ν := λT , we see that
X
(6.3)
ϕT = ρ + ν + ν −1 , ϕQ = 1Q +
α.
α∈Irr(Q)
Next, using (6.3) one can see that WP = CW (Q) ⊥ [W, Q], a direct orthogonal sum of
two P -submodules. Here, C := CW (Q) is of dimension 2 and affords
the T -character
P
ν + ν −1 , [W, Q] is of dimension q − 1 and affords the Q-character 1Q 6=α∈Irr(Q) α and the
T -character ρ (as T permutes cyclically and transitively the q − 1 non-principal irreducible
characters of Q). It also follows that these two subspaces are non-degenerate and self-dual
P -submodules. Next, we can further decompose:
[W, Q] = A ⊥ B
as an orthogonal sum of two self-dual T -modules, where A affords the T -character ρ − ν −
ν −1 , and B affords the T -character ν + ν −1 . Summarizing, we have that
WP = A ⊥ B ⊥ C,
where A ⊥ B is an irreducible P -module of dimension q − 1, and C is a sum of two
irreducible P -modules of dimension 1. Applying the Artin-Wedderburn theorem to (P, W ),
we obtain
(6.4)
M ⊃ End(A ⊕ B) := {f ∈ End(W ) | f (A ⊕ B) ⊆ A ⊕ B, f (C) = 0}.
34
R. GURALNICK, F. HERZIG, AND P. TIEP
Repeating the above argument for P ♯ instead of P , we have that
WP ♯ = A♯ ⊥ B ♯ ⊥ C ♯ ,
where A♯ affords the T -character ρ − ν − ν −1 , B ♯ affords the T -character ν + ν −1 , and
C ♯ = CW (Q♯ ) affords the T -character ν + ν −1 , and
(6.5)
M ⊃ End(A♯ ⊕ B ♯ ) := {f ∈ End(W ) | f (A♯ ⊕ B ♯ ) ⊆ A♯ ⊕ B ♯ , f (C ♯ ) = 0}.
Comparing with (6.3), we see that A♯ = A. Furthermore, C ∩ C ♯ is centralized by hQ, Q♯ i =
H, so C ∩C ♯ = 0. But both C and C ♯ are of dimension 2 and orthogonal to A = A♯ , whence
C ⊕ C ♯ = A⊥ . Next, B ∩ B ♯ is a subspace of the non-degenerate subspace A⊥ which is
orthogonal to both C and C ♯ , so B ∩ B ♯ = 0. Since dim B + dim B ♯ = 4 = dim A⊥ , we have
shown that
A♯ = A, A⊥ = B ⊕ B ♯ = C ⊕ C ♯ , W = A ⊥ (B ⊕ B ♯ ).
Suppose now that f ∈ End(W ) belongs to both End(A ⊕ B) and End(A ⊕ B ♯ ) as identified
in (6.4) and (6.5). Then f = 0 on C ⊕ C ♯ = A⊥ , i.e. f (A⊥ ) = 0. Next,
f (A) ⊆ (A ⊕ B) ∩ (A ⊕ B ♯ ) = A.
It follows that
End(A ⊕ B) ∩ End(A ⊕ B ♯ ) ⊆ End(A) := {f ∈ End(W ) | f (A) ⊆ A, f (A⊥ ) = 0},
and so by (6.4), (6.5) we have
dim M ≥ 2(q − 1)2 − (q − 3)2 = q 2 + 2q − 7,
i.e. codimEnd(W ) M ≤ 8.
On the other hand, all non-principal ψ ∈ IBrp (H) have degree ≥ q − 1 ≥ 15. So,
assuming M 6= End(W ), we see that all composition factors of the H-module Q :=
End(W )/M are trivial. Since H is perfect, it follows that H acts trivially on Q. But
dim HomkH (End(W ), 1H ) = 1, so M is contained in the unique submodule E0 := {f ∈
End(W ) | tr(f ) = 0} of codimension 1 in End(W ). But this is a contradiction, since
by (6.4), M contains the map g which acts as identity on A ⊕ B and as 0 on C, with
tr(g) = q − 1 = p − 2 6= 0.
(b) Now we handle the case p = q − 1 (so 2 | q ≥ 8). Then for the unipotent radical Q
of P we have
M
WQ =
Wα
1Q 6=α∈Irr(Q)
with Wα affording the Q-character α. Next, let S ∼
= Cq+1 be a non-split torus in H. Then
there is some 1S 6= γ ∈ Irr(S) such that
M
WS =
Wβ
β∈Irr(S), β6=γ,γ −1
with Wβ affording the S-character β. Thus both Q and S are multiplicity-free on W . By
Lemma 6.10, we see that for any composition factor U of End(W )/M, U Q = U S = 0. On
the other hand, inspecting the (Brauer) character table of H (see [7]), one sees that U Q 6= 0
if dim U = 1, q, q + 1 and U S 6= 0 if dim U = q − 1, for any U ∈ IBrp (H). Hence we conclude
that M = End(W ).
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
35
Proposition 6.13. Let k = k̄ and let H = SUn (q) with 3 < p = (q n + 1)/(q + 1) and n ≥ 3.
Let W be an irreducible kH-module of dimension p. Then (H, W ) is weakly adequate.
Proof. Let ϕ denote the Brauer character of W and let N := Fnq2 denote the natural Fq2 Hmodule. Recall that H possesses the so-called reducible Weil character
dimF
ζn,q : g 7→ (−1)n (−q)
(6.6)
q2
Ker(g−1)
for all g ∈ H, which decomposes as the sum of q + 1 distinct irreducible Weil characters
ζn,q =
q
X
ζni
i=0
(q n − q)/(q + 1)
(q n + 1)/(q + 1)
(of degree
for i = 0 and
for i > 0), see [64]. Then ϕ can be
j
′
obtained by restricting some ζn with j > 0 to p -elements of H. We also let P := StabH (U )
for a maximal totally singular subspace U of N , with unipotent radical Q (so P is a p′ group), and let ρ denote the permutation character of H acting on the set Ω of singular
1-spaces of N .
(a) First we show that
• the only irreducible constituents of ζnj ζ̄nj that are not of p-defect 0 are 1H and (possibly)
another one, σ, of degree
(6.7)
σ(1) =
(q n − q)(q n + q 2 )
;
(q 2 − 1)(q + 1)
• all p-defect 0 constituents of ζnj ζ̄nj have degree > 2p if n > 3, and
• σ is the Steinberg character St of H if n = 3 and it is a constituent of ρ if n > 3.
Indeed, (6.6) implies that (ζn,q )2 is just the permutation character of H acting on the point
set of N , and at the same time it equals the restriction to H of the reducible Weil character
ζ2n,q of SU2n (q), if we embed H diagonally into SU2n (q): X 7→ diag(X, X). Assume n > 3.
Then all irreducible constituents of the latter restriction are described by [44, Proposition
6.3], and their degrees are listed in [44, Table III]. It follows that (ζn,q )2 has exactly two
non-p-defect 0 irreducible constituents, namely 1H (with multiplicity q + 1) and another
one σ of indicated degree (with multiplicity q). Certainly, the permutation representation
of H on (the point set of) N contains the permutation representation of G on Ω as a
subrepresentation (no matter if n > 3 or not). On the other hand, ρ contains an irreducible
constituent of degree as listed in (6.7) (see [57, Table 2]), so ρ = 1H + σ + ψ (and ψ ∈ Irr(H)
has p-defect 0). One also easily checks all defect 0 constituents of (ζn,q )2 have degree > 2p.
Suppose that n = 3; in particular 3 ∤ (q + 1) and q > 2. Inspecting the character table
of H = SU3 (q) as given in [17], we see that the only non-p-defect 0 irreducible characters
of H are 1H , the Weil character ζ30 of degree q 2 − q, the Steinberg character St of degree
(u)
q 3 matching (6.7), and (q 2 − q)/3 characters χ(q+1)2 (q−1) . Direct calculations show that
(u)
[ζ3j ζ̄3j , χ(q+1)2 (q−1) ]H = 0. Next, observe that
H
(ζ3,q )2 = 1H + 1H
Q + (q − 1)1L ,
36
R. GURALNICK, F. HERZIG, AND P. TIEP
where Q = StabH (u) is the unipotent radical of P as above (if U = huiFq2 ) and L =
StabH (v) ∼
= SU2 (q) for some non-singular v ∈ N . Furthermore,
[(ζ30 )Q , 1Q ]Q = 0, (ζ30 )L =
q
X
ζ2i .
i=1
[ζ30 , 1H
Q ]H
The former relation implies
= 0. On the other hand, by [64, Lemma 4.7(ii)],
i
each ζ2 in the latter relation is obtained by restricting an irreducible character of degree
q − 1 ≥ 2 of GU2 (q) ✄ L to L. It follows by Clifford’s theorem that [ζ30 , 1H
L ]H = 0. Thus we
have shown that ζ30 is not a constituent of (ζ3,q )2 , as claimed.
(b) Now we show that, if β is an irreducible constituent of ϕϕ̄ and β 6= 1H , then either
β(1) ≥ 2p, or n = 3, β(1) = p, and [βQ , 1Q ]Q > 0. Indeed, suppose that β(1) < 2p. Suppose
for the moment that n > 3. Then by the results of (a), β is a constituent of the Brauer
character σ ◦ . But according to [43], σ ◦ − 1H ∈ IBrp (H), so β(1) = σ(1) − 1 > 2p by (6.7), a
contradiction. Thus n = 3. If moreover β is in a block of p-defect 0, then using [17, Table
3.1] we see that β(1) = p, β = (ζ3i )◦ for some i > 0 and so βQ contains 1Q . Otherwise, by
the results of (a), β is a constituent of St◦ . In this case, according to [17, Theorem 4.2],
St◦ − 1H ∈ IBrp (H) and so β(1) = St(1) − 1 = q 3 − 1 > 2p, again a contradiction.
(c) When n ≥ 5, according to Lemmas 12.5 and 12.6 of [24], ϕZ(Q) contains a nonprincipal linear character λ, whose P -orbit O has length (q n−1 − 1)/(q + 1); moreover, any
irreducible character of Q above λ has degree q. Since ϕ(1) = (q n + 1)/(q + 1), it follows
that WP = A ⊕ B, where B := CW (Z(Q)) has dimension 1, A := [W,
P Z(Q)] ∈ Irr(P ) has
dimension (q n − q)/(q + 1) = p − 1 and affords the Z(Q)-character q α∈O α. The same is
also true for n = 3, see Tables 2.1 and 3.1 of [17]. Applying the Artin-Wedderburn theorem
to (P, W ), we see that
M ⊇ End(A) ⊕ End(B).
In particular, if M 6= End(W ), then any composition factor X of the H-module End(W )/M
has dimension ≤ 2p − 2 and moreover X Q ⊆ X Z(Q) = 0. But this is impossible by the
results of (b).
Lemma 6.14. Let k = k̄ and let W be an irreducible kH-module of dimension p ≥ 3,
where H is quasisimple and (H, W ) is one of the non-serial examples listed in Tables I,IIa,
IIb, or III. Then (H, W ) is weakly adequate.
Proof. Let ϕ denote the Brauer character of W . Note that the cases (H, p) = (A5 , 3), (A6 , 5)
are covered by Proposition 6.11 since A5 ∼
= PSp2 (5) and A6 ∼
= PSL2 (9).
Suppose that (H, p) = (Sp6 (2), 7). Then H > L ∼
= SL2 (8), and ϕL is irreducible (see
[34]), so we are done by Proposition 6.12.
Assume that (H, p) = (M11 , 11). Then H contains a p′ -subgroup L = M10 ∼
= A6 · 23 ,
and using [16] we can check that ϕL = λ + ψ, where λ, ψ ∈ Irr(L) are rational of degree 1
and 10 (and λ 6= 1L ). It follows that WL = A ⊕ B, where A affords the character λ and
B affords the character ψ. Applying the Artin-Wedderburn theorem to (L, W ) we see that
M ⊇ End(A) ⊕ End(B). In particular, if M 6= End(M ), then any composition factor U of
the H-module End(W )/M has dimension ≤ 20 and moreover all composition factors of UL
afford the character λψ = ψ. The latter condition also implies that dim U = 10 or 20. On
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
37
the other hand, using [34] and [16] we see that any such U must be of dimension 10 and
its character restricted to L yields an irreducible non-rational character, different from ψ.
Hence M = End(W ).
Assume that (H, p) = (M12 , 11). Then H contains a maximal subgroup L ∼
= PSL2 (11),
and using [16] we can check that ϕL is irreducible. So we are done by [23, Proposition 3.1].
Assume that (H, p) = (M24 , 23). Then H contains a maximal subgroup L ∼
= PSL2 (23),
and using [16] we can check that ϕL is irreducible. So we are done by [23, Proposition 3.1].
Assume that (H, p) = (Co2 , 23) or (Co3 , 23). Then H contains a p′ -subgroup L ∼
= M cL,
and using [10] we can check that ϕL = 1L + ψ, with ψ ∈ Irr(L). It follows that WL = A⊕ B,
where A := CW (L) has dimension 1 and B affords the character ψ. Applying the ArtinWedderburn theorem to (L, W ) we see that M ⊇ End(A) ⊕ End(B). In particular, if
M=
6 End(M ), then any composition factor U of the H-module End(W )/M has dimension
≤ 44 and moreover U L = 0. On the other hand, using [49] we see that the only irreducible
kH-modules of dimension ≤ 44 are k and W , and both have nonzero L-fixed points. Hence
we conclude that M = End(W ).
Now we can prove Theorem 1.7 which we restate:
Theorem 6.15. Let k = k̄ be of characteristic p and let G be a finite group with a faithful
irreducible kG-module V of dimension p. Then precisely one of the following holds:
(i) G is adequate on V ;
(ii) G contains an abelian normal subgroup A of index p (and G permutes p onedimensional summands of V with kernel A); or
(iii) p = 3 and the image of G in PGL(V ) is PSL2 (9).
Proof. First assume that p > 3. Apply Proposition 6.5 to G. In the case (i) of the proposition, we are done by Proposition 6.6. So we may assume that G is almost quasisimple,
H := G(∞) is quasisimple, with simple quotient S, and H is irreducible on V . If S is of
Lie type in characteristic p, we can apply Proposition 6.8. Assume we are in the remaining
cases. In all these cases, observe that the outer automorphism group Out(S) is a p′ -group
and the Schur multiplier Mult(S) is a p′ -group as well (as p > 3). The first condition
implies that G+ = H, whence by Remark 6.1 without loss we may assume G = H and so
H 1 (G, k) = 0. The second condition implies that H 2 (G, k) = 0. Furthermore, in all cases
V lifts to a complex module of p-defect 0, whence V is projective and so Ext1G (V, V ) = 0.
Finally, H is weakly adequate on V by Propositions 6.11, 6.12, 6.13, and Lemmas 6.9, 6.14.
Now consider the cases when p = 2 or 3. If V is imprimitive the result follows as above.
So assume that V is primitive. Set H = G(∞) .
Suppose that G = H = SL2 (pa ) or PSL2 (pa ). Then a ≥ 2 and the result follows by
Corollary 9.4. If G > H then as V g ∼
= V as H-modules for all g ∈ G, G/H is a p′ -group
and G is adequate on V whenever H is. Thus the last case to consider is H = PSL2 (9) ∼
=
A6 . The normalizer of H in PGL(V ) is PGL2 (9) (the normalizer is just the subgroup of
the automorphism group which fixes the isomorphism class of V ). If the image of G in
PGL(V ) is PGL2 (9), H 1 (G, k) = H 2 (G, k) = 0 (see the proof of Corollary 9.5) and since
Ext1G (V, V ) = 0, V is adequate in this case.
By Proposition 6.5 and Theorem 2.2, the remaining cases to consider are G almost
quasisimple, p = 3 and H ∈ {A5 , PSL2 (7), SL3 (3a ), SU3 (3a )}. In the first two cases, the
38
R. GURALNICK, F. HERZIG, AND P. TIEP
order of G is not divisible by 9, whence V is projective and so Ext1G (V, V ) = 0. Note also
that H 1 (G, k) = H 2 (G, k) = 0. In the first case, V ⊗ V ∗ is a direct sum of the projective
cover of k and a 3-dimensional module. Elements of order 5 have nonzero trace and 3dimensional fixed space. Since elements of order 5 have only a 2-dimensional fixed space
on the projective cover of k, it follows that the span of the elements of order 5 generate
V ⊗ V ∗ . In the second case, V ⊗ V ∗ is the projective cover of k and since the trace of an
irreducible character cannot be identically 0, it follows that H is weakly adequate on V in
this case as well. Thus, (G, V ) is adequate.
In the last two cases, weak adequacy follows from the fact that V ⊗ V ∗ is a uniserial
module with trivial socle and head. It follows by the main result of [48] that Ext1G (V, V ) = 0
for a > 2. One computes directly that Ext1G (V, V ) = 0 in all other cases (Klaus Lux did
the computation; also see [37] for the case of SL3 (3a )). Since H 1 (G, k) = H 2 (G, k) = 0, the
result follows.
7. Certain PIMs for simple groups
For a finite group X and a fixed prime p, let B0 (X) denote the principal p-block of
X. We will sometimes use the same notation for an irreducible kX-module and its Brauer
character.
First we describe the submodule structure of the PIMs for some non-projective modular
representations of simple groups H described in Theorems 2.1 and 2.2.
Assume that H has a Sylow p-subgroup P of order p and furthermore that P = CH (P ).
In this case, P has a unique block b with defect group P and canonical character 1P , see
[46, Theorem 4.6.12]. According to Brauer’s theorem [46, Theorem 4.12.1], H has a unique
p-block B of defect d > 0 (hence d = 1), and B = bG . In particular, B = B0 (H). Note
that the number of exceptional characters in B equals (p − 1)/|NH (P )/P | in this situation,
and all of them are p-conjugate (and so non-rational if |NH (P )/P | < p − 1), cf. Theorem
4.12.1 and Corollary 4.12.2 of [46]. We will use [46, Theorem 4.12.1] to find PIMs P(ϕ) for
some ϕ ∈ IBrp (B).
7.1. The case H = PSLn (q) with p = (q n − 1)/(q − 1) and n ≥ 2. First suppose that
n ≥ 3. Then B contains unipotent characters χ0 = 1H , χ1 , and χ2 labeled by the partitions
(n), (n − 1, 1), and (n − 2, 12 ), and Brauer characters ϕ0 = 1H , ϕ1 of degree p − 2 (afforded
by D), and ϕ2 of degree
(q n − 2q 2 + 1)(q n − q)
+1
(q 2 − 1)(q − 1)
(afforded by a kH-module say U ) among others (see e.g. [26, Proposition 3.1]). More
precisely,
(7.1)
χ◦0 = ϕ0 , χ◦1 = ϕ0 + ϕ1 , χ◦2 = ϕ1 + ϕ2 .
Note that
dim(U ) = ϕ2 (1) > 2p
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
39
unless (n, q) = (3, ≤ 3). Since χi are rational for i = 0, 1, 2, they are all non-exceptional.
Next, the character of any PIM in B is of the form ψ1 + ψ2 , where ψ1 ∈ Irr(B) is nonexceptional, and either ψ2 ∈ Irr(B), or ψ2 is the sum of all (p − 1)/n exceptional characters
in B. Hence the relations (7.1) show that
• P(ϕ0 ) affords the character (χ0 + χ1 )◦ = 2ϕ0 + ϕ1 . In fact, one can see that P(k) is
the uniserial module (k|D|k). In particular, this shows that Ext1H (k, U ) = Ext1H (U , k) = 0;
• P(ϕ1 ) affords the character (χ1 + χ2 )◦ = ϕ0 + 2ϕ1 + ϕ2 . By [46, Corollary 4.12.5], the
module P(ϕ1 ) = P(D) has socle series (D|k ⊕ U |D). Since ϕ1 is real, P(D) is self-dual.
Furthermore, the only nonzero proper submodules of P(D) are
D = soc(P(D)), (D|k), (D|U ), (D|k ⊕ U ) = rad(P(D)),
(cf. [46, Figure 4.3]), and so none is of the form (D|k|D).
Next, let H = SL2 (q) and p = q + 1. The decomposition numbers for B0 (H) are given
in [7]. In particular, Irr(B) consists of χ0 = 1H , χ1 = St of degree q, and q/2 exceptional
characters θi , 1 ≤ i ≤ q/2, of degree q − 1, and IBrp (B) = {1H , ϕ1 }, with ϕ1 afforded by
D. So as before, we have that P(k) = (k|D|k) is uniserial. Next, P(D) = P(ϕ1 ) affords the
character
q/2
X
q
(St +
θi )◦ = 1H + ( + 1)ϕ1 .
2
i=1
If we let Dj = (D|D| . . . |D) denote a uniserial module with Brauer character jϕ1 , then
P(D) = (D|k ⊕ D(q−2)/2 |D) (cf. [46, Corollary 4.12.5]). Furthermore, the only nonzero
proper submodules of P(D) are Dj or (D|k ⊕ Dj−1 ) with 1 ≤ j ≤ q/2.
7.2. The case H = PSUn (q) with p = (q n + 1)/(q + 1), n ≥ 3, (n, q) 6= (3, 2). Note that
in this case H̃ ∼
= Z(H̃) × H for H̃ := GUn (q); moreover, the unipotent characters of H̃ as
well as the characters in B0 (H̃) are all trivial at Z(H̃). Hence without loss we may assume
H = GUn (q). We will consider the unipotent characters χ0 = 1H , χ1,2,3 , labeled by the
partitions (n), (n − 1, 1), (n − 2, 12 ), and (n − 3, 13 ) (the latter being considered only when
n > 3). Using the description of Brauer trees for H given in [15, §6], we see that, when
n ≥ 5, there exist Brauer characters ϕ0 = 1H , ϕ1,2,3 , such that
(7.2)
χ◦0 = ϕ0 , χ◦1 = ϕ1 , χ◦2 = ϕ0 + ϕ2 , χ◦3 = ϕ1 + ϕ3 .
In particular, ϕ0 = 1H , ϕ1 is a Weil character of degree p − 1,
ϕ2 (1) = p
qn + q2 − q − 1
(q n − q)(q n − q 3 + q 2 − 1)
−
2
>
8p,
ϕ
(1)
=
p
− 2p + 2 > 28p.
3
q2 − 1
(q 2 − 1)(q 3 − 1)
Since χi are all rational, they are all non-exceptional, so as in §7.1, the relations (7.2) and
[46, Corollary 4.12.5] show that both P(ϕ0,1 ) are uniserial:
P(ϕ0 ) = (ϕ0 |ϕ2 |ϕ0 ), P(ϕ1 ) = (ϕ1 |ϕ3 |ϕ1 ),
where we have used the same notation for the module and its Brauer character.
Suppose now that n = 3 and q ≥ 3. Then we still have P(ϕ0 ) = (ϕ0 |ϕ2 |ϕ0 ) is uniserial
for ϕ0 = 1H and ϕ2 (1) > 2p. For the Weil character ϕ1 of degree p − 1, now P(ϕ1 ) affords
40
R. GURALNICK, F. HERZIG, AND P. TIEP
the character
(p−1)/3
(χ1 +
X
θi )◦ =
i=1
p+2
p−1
ϕ1 +
ϕ2 ,
3
3
where θi are exceptional characters in B, of degree (q 2 −1)(q +1) > 2p, for 1 ≤ i ≤ (p−1)/3.
We claim that
P(ϕ1 ) = (ϕ1 |ϕ2 |ϕ1 | . . . |ϕ2 |ϕ1 )
is self-dual, uniserial of length (2p+1)/3, with the composition factors ϕ1 and ϕ2 alternating.
The first statement follows since ϕ1 is real. The second one holds because in this case, both
the socle soc(P(ϕ1 )) and head P(ϕ1 )/rad(P(ϕ1 )) are simple and rad(P(ϕ1 ))/ soc(P(ϕ1 ))
is uniserial. By [63, Theorem 1.1], H has a unique complex character of degree equal to
ϕ0 (1) or ϕ1 (1). Hence the last statement holds by Lemma 3.1 since ϕ0,1 each has a unique
lift to characteristic 0.
7.3. The case H = SL2 (q) and p = q − 1 ≥ 3. The decomposition numbers for B0 (H) are
given in [7]. In particular, Irr(B) consists of χ0 = 1H , χ1 = St of degree q, and (q − 2)/2
exceptional characters θi , 1 ≤ i ≤ q/2, of degree q + 1, and IBrp (B) = {ϕ0 = 1H , ϕ1 }, with
ϕ1 = St◦ afforded by D. Clearly, both ϕ0,1 have a unique complex lift, so by Lemma 3.1
they have no self-extensions. Arguing as in §7.2 and using [46, Corollary 4.12.5], we see
that both
P(ϕ0 ) = (ϕ0 |ϕ1 |ϕ0 | . . . |ϕ1 |ϕ0 ), P(ϕ1 ) = (ϕ1 |ϕ0 |ϕ1 | . . . |ϕ0 |ϕ1 )
are self-dual and uniserial of length p, and with the composition factors ϕ0 and ϕ1 alternating.
7.4. The case H = Ap with p ≥ 7. Consider the irreducible complex characters χ0,1,2 of
Sp labeled by (p), (p − 1, 1), and (p − 2, 12 ). By Peel’s theorem [54],
(7.3)
χ◦0 = ϕ0 , χ◦1 = ϕ0 + ϕ1 , χ◦2 = ϕ1 + ϕ2 ,
where ϕ0,1,2 ∈ IBrp (Sp ), ϕ0 (1) = 1, ϕ1 (1) = p − 2, ϕ2 (1) = (p − 2)(p − 3)/2. It is well known
that χ0,1,2 and ϕ0,1 are all irreducible over H. Restricting to Sp−1 , it is easy to see that
ϕ2 is irreducible over H as well. We will use the same notation for the restrictions of these
characters to H. Since χ0,1,2 are rational, they are non-exceptional in B0 (H). Hence, (7.3)
and [46, Theorem 4.12.1] imply that P(ϕ0 ) = (ϕ0 |ϕ1 |ϕ0 ) is uniserial and that P(ϕ1 ) has
socle series (ϕ1 |ϕ0 ⊕ ϕ2 |ϕ1 ). In particular, there is no kH-module of the form (ϕ1 |ϕ0 |ϕ1 ).
8. Indecomposable modules of dimension less than 2p − 2
First we record a simple observation
Lemma 8.1. Let b ∈ N and let V be a kG-module of dimension ≤ b with a G+ -composition
factor U . Suppose that any quotient of length 2 of P(U ) or P(U ∗ ) has dimension > b. Then
U is a direct summand of the G+ -module V . If moreover U has multiplicity 1, then the
G-module V is either irreducible of dimension dim U , or decomposable.
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
41
Proof. Suppose that W is an indecomposable subquotient of length 2 of the G+ -module
V with U as a composition factor. Replacing W by W ∗ if necessary, we may assume that
head(W ) ∼
= U and so W is a quotient of P(U ). But then by the hypothesis, dim W > b ≥
dim V , a contradiction. So U is a direct summand of the G+ -module V by Lemma 3.5(i):
VG+ = U1 ⊕ M with U1 ∼
= U . The last claim now follows from Lemma 3.7(ii).
Given a nontrivial U ∈ IBrp (X), we call any kX-module V U -special, if V has U or U ∗
as composition factors of total multiplicity ≤ 1, and moreover all other composition factors
of V are trivial.
Lemma 8.2. Let N = Op (N ), b ∈ N, and let U ∈ IBrp (N ) be a nontrivial module. Suppose
that the only U -special quotients of P(k), P(U ), and P(U ∗ ) of dimension ≤ b are uniserial
modules in the list
X := {k, Y, (k|Y ), (Y |k), (k|Y |k) | Y = U or U ∗ }.
Let V be any U -special kN -module of dimension ≤ b. Then V ∼
= ⊕m
i=1 Xi for some Xi ∈ X .
Proof. We induct on the length of V . Suppose V has length ≥ 2. If all composition factors
of V are trivial, then N acts trivially on V since N = Op (N ), and we are done. Replacing
V by V ∗ if necessary, we may assume that V has U as a composition factor of multiplicity
1, and all other composition factors of V are k.
Suppose that U embeds in head(V ) := V /R, where R := rad(V ). Then all composition
factors of R are trivial, and so N acts trivially on R. Assume in addition that V /R is not
simple. Then V /R = M/R ⊕ Y /R for some submodules M, Y with Y /R ∼
= k. Again, N
∼
acts trivially on Y , so we can write Y = R ⊕ Z for some submodule Z = k. It follows
that V = M ⊕ Z, and we are done by induction. Assume now that V /R ∼
= U . Then the
surjection P(U ) → V /R lifts to a surjection P(U ) → V . Since dim V ≤ b, we must then
have that V ∈ X .
The case U ֒→ soc(V ) now follows from the previous case by duality.
Now we may assume that U embeds neither in head(V ) nor in soc(V ). Letting W :=
[N, V ] and T := rad(W ), we see that W has no trivial quotient. But W/T is semisimple,
so W/T ∼
= U . Applying the induction hypothesis to V /T and noting that U is in the socle
but not in the head of V /T , we see that V /T ∼
= L/T ⊕ Y /T , where L/T ∼
= (U |k) and N
acts trivially on Y /T . In this case, N acts trivially on Y as well. If moreover Y 6= T , then
we can decompose Y = T ⊕ Z for some submodule Z 6= 0, whence V = L ⊕ Z and we are
done by induction. Thus we may assume V /T ∼
= (U |k). Consider any maximal submodule
M of V . Since U 6֒→ head(V ), V /M ∼
= k and so M ⊇ W . It follows that R ⊇ T and
R/T = rad(V /T ) ∼
= k. In this final case, the surjection P(k) → V /R lifts
= U . Hence V /R ∼
to a surjection P(k) → V . Since dim V ≤ b, we must again have that V ∈ X .
Corollary 8.3. Let G+ be perfect, b ∈ N, and let U ∈ IBrp (G+ ) be a nontrivial module.
Suppose that the only U -special quotients of dimension ≤ b of P(k), P(U ), and P(U ∗ ) are
uniserial modules in the list
X := {k, Y, (k|Y ), (Y |k), (k|Y |k) | Y = U or U ∗ }.
Let V be any indecomposable kG-module of dimension ≤ b such that VG+ is U -special. Then
VG+ is also indecomposable and belongs to X .
42
R. GURALNICK, F. HERZIG, AND P. TIEP
Proof. We may assume that exactly one indecomposable direct summand A of VG+ has U
as a composition factor, and so VG+ = A ⊕ B with G+ acting trivially on B. Hence B = 0
by Lemma 3.7(ii).
Theorem 8.4. Let G be a finite group, k an algebraically closed field of characteristic
p, Op (G) = 1, and let V be a faithful, indecomposable kG-module of dimension less than
2p − 2. Assume in addition that G+ is quasisimple but not of Lie-type in characteristic p.
Then one of the following statements holds, where U, W ∈ IBrp (G+ ).
(i) V is irreducible.
(ii) (G+ , p, dim U ) = (SL2 (q), q −1, p+1), (Ap , p, p−2), (SLn (q), (q n −1)/(q −1), p−2),
(M11 , 11, 9), (M23 , 23, 21). Furthermore, VG+ is uniserial of the form (k|U ), (U |k),
or (k|U |k), and U ∼
= U ∗.
(iii) (G+ , p, dim U ) = (SL2 (q), q + 1, p − 2), U ∼
= U ∗ , and VG+ is indecomposable of the
form (U |U ), (U |k ⊕ U ), or (k ⊕ U |U ).
(iv) (G+ , p, dim U ) = (2A7 , 7, 4), VG+ = (U |U ) is uniserial, and U ∼
= U ∗.
+
(v) (G , p) = (M11 , 11) and VG+ = (U |W ) is uniserial, {dim U, dim W } = {9, 10}.
(vi) (G+ , p, dim U ) = (3A6 , 5, 3), VG+ = (U |U ) is uniserial, and U ∼
6 U ∗.
=
+
2
(vii) (G , p, dim U ) = ( B2 (8), 13, 14), VG+ is uniserial of the form (k|U ) or (U ∗ |k) for
a fixed U ∼
6 U ∗.
=
Proof. (a) Note that the statement is vacuous for p = 2. Throughout the proof, we assume
that p > 2, V is reducible, and let U be a composition factor of the G+ -module V of
largest dimension. Also set b := 2p − 3 whenever we apply Lemma 8.2 and Corollary
8.3. Note that VG+ is (reducible) indecomposable by Corollary 4.5. Next, G+ must act
irreducibly and non-trivially on some subquotient X of VG+ . Applying Theorems 2.1 and
2.2 to the action of G+ on X, we see that Mult(G+ /Z(G+ )), and so Z(G+ ), has p′ -order.
The indecomposability of VG+ then implies that Z(G+ ) acts via scalars on V and that G+
acts faithfully on U (and so we may identify G+ with its image in GL(U )). In particular, if
k is a composition factor of VG+ , then G+ is simple. This must be the case if dp (G+ ) ≥ p−1.
Also, if Z(G+ ) 6= 1, then G+ acts faithfully on every composition factor of VG+ .
(b) Assume first that (G+ , p) = (J1 , 11). According to [34], the only ϕ ∈ IBrp (G+ ) of
degree < 2p are ϕ1,7,14 . Here we use the notation ϕj to denote the unique ϕ ∈ IBrp (G+ ) of
degree j. Moreover, using [49] we see that
(8.1) P(ϕ1 ) = (ϕ1 |ϕ119 |ϕ1 ), P(ϕ7 ) = (ϕ7 |ϕ49 ⊕ ϕ69 |ϕ7 ), P(ϕ14 ) = (ϕ14 |ϕ106 ⊕ ϕ119 |ϕ14 ).
Since dim V < 2p, each composition factor X of the G+ -module V must afford the Brauer
character ϕi for some i ∈ {1, 7, 14}. Now (8.1) shows that Ext1G+ (X, Y ) = 0 for any two
such composition factors X and Y . Hence the G+ -module V is semisimple by Lemma 3.5,
a contradiction.
From now one we may assume that G+ 6∼
= J1 and so dp (G+ ) ≥ p − 3 by Theorem 2.1. In
particular, dim U ≥ p − 3 and Corollary 3.9 applies.
(c) Here we consider the case where dim U > p. Since dim V ≤ 2p − 3 and VG+ is
reducible, it follows that k is a composition factor of VG+ , and so G+ is simple as noted in
(a). Also, all composition factors of VG+ other than U are trivial. Now we apply Theorem
2.2 to (G+ , U ).
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
43
Suppose that G+ = An with n ≥ p as in the first row of Table I. Since p + 1 ≤ dim U ≤
2p − 3, we have that 5 ≤ p ∤ n and p + 2 ≤ n ≤ 2p − 2. By [51, Lemma 6.10], H 1 (An , U ) = 0,
whence Ext1G+ (k, U ) = Ext1G+ (U, k) = 0. Also, Ext1G+ (k, k) = Ext1G+ (U, U ) = 0 by Lemma
3.1. It follows by Lemma 3.5 that VG+ is semisimple, a contradiction.
Next suppose that (G+ , p, dim U ) = (SL2 (q), q − 1, p + 1) as in Table IIa. Then, as shown
in §7.3, (G+ , U ) satisfies the hypothesis of Corollary 8.3, and so we arrive at (ii).
In the cases where (G+ , p, dim U ) = (A7 , 7, 10), (SL3 (3), 13, 16), (SU4 (2), 5, 6), (Sp4 (4), 17, 18),
(G2 (3), 13, 14), (J1 , 11, 14), (J1 , 19, 22 or 34), (M12 , 11, 16), or (M11 , 11, 16), using the information on decomposition numbers given in [49], one can check that U satisfies the
hypothesis of Lemma 8.1, and so V is decomposable, a contradiction.
Assume that (G+ , p, dim U ) = (2 B2 (8), 13, 14). Then V is U -special, and using [49] one
can check that the only quotients of dimension ≤ 23 of P(k), P(U ), and P(U ∗ ) are k, Y ,
(k|Y ), or (Y |k), with Y = U or U ∗ . Applying Corollary 8.3, we arrive at (vii).
(d) Next we consider the case dim U = p, and apply Theorem 2.2 to (G+ , U ). By Lemma
6.2, U is projective, and so it is a direct summand of VG+ , a contradiction.
(e) Assume now that dim U = p − 1, and apply Theorem 2.1 to (G+ , U ). Note that U
has multiplicity 1 as dim V < 2p − 2. First we consider the case (G+ , p) = (SUn (q), (q n +
1)/(q + 1)). In this case, G+ is simple, dp (G+ ) = p − 1 and so all other composition factors
of the G+ -module V are trivial. As shown in §7.2,
Ext1G+ (k, U ) = Ext1G+ (U, k) = Ext1G+ (k, k) = Ext1G+ (U, U ) = 0.
It follows by Lemma 3.5 that VG+ is semisimple, a contradiction.
Suppose now that (G+ , p) = (Sp2n (q), (q n + 1)/2), (2Ru, 29, 28), (3J3 , 19, 18), (2A7 , 5, 4),
(3A7 , 7, 6), (6A7 , 7, 6), (2J2 , 7, 6), (61 ·PSU4 (3), 7, 6), (6·PSL3 (4), 7, 6), (2M12 , 11, 10), (2M22 , 11, 10),
(6Suz, 13, 12), or (2G2 (4), 13, 12). Since Z(G+ ) 6= 1, G+ acts faithfully on every composition factor X of VG+ as noted in (a), whence dim X ≥ p − 1 > (dim V )/2 by Theorem 2.1.
It follows that VG+ is irreducible, a contradiction.
Assume that (G+ , p, dim U ) = (M11 , 11, 10). The Brauer tree of B0 (G+ ) is given in
[46, Example 4.12.11]. Using this information, we see that the only quotient of length 2 of
dimension ≤ 19 of P(U ) is of form (W |U ), where W ∈ IBrp (M11 ) has dimension 9. Arguing
as in the proof of Lemma 8.1, we arrive at (v).
(f) Next we consider the case dim U = p − 2 and apply Theorem 2.1 to (G+ , U ). We can
exclude the subcase (G+ , p) = (SL3 (2) ∼
= PSL2 (7), 7).
(f1) First we assume that (G+ , p) = (SLn (q), (q n − 1)/(q − 1)) or (Ap , p), and moreover
U is a composition factor of V of multiplicity 2. Since dim V ≤ 2p − 3, we have two cases.
• dim V = 2p − 3 and so k is also a composition factor of the G+ -module V . Suppose
that head(VG+ ) is not simple. Then VG+ contains two maximal submodules A, B of length
2 and A ∩ B ⊆ soc(VG+ ). On the other hand, the indecomposability of VG+ implies that
soc(VG+ ) ⊆ rad(VG+ ) ⊆ A ∩ B, whence soc(VG+ ) = A ∩ B is simple. So up to duality, we
may assume that head(VG+ ) is simple. It follows that VG+ is a quotient of P(U ) or P(k).
The structure of PIMs described in §§7.1, 7.4 shows that (iii) holds.
44
R. GURALNICK, F. HERZIG, AND P. TIEP
• dim V = 2p − 4 and so VG+ has exactly two composition factors, both isomorphic to
U . As VG+ is indecomposable, it is a quotient of P(U ). Using the results of §§7.1, 7.4, we
again arrive at (iii).
(f2) Now we assume that (G+ , p) = (SLn (q), (q n − 1)/(q − 1)) or (Ap , p), and moreover
U is a composition factor of V of multiplicity 1. By Theorem 2.1, U is the only nontrivial
irreducible kG+ -module of dimension ≤ p − 1. Since dim V ≤ 2p − 3, it follows that
all other composition factors of VG+ are trivial, i.e. VG+ is U -special. The structure of
PIMs described in §§7.1, 7.4 shows that the only U -special quotients of dimension at most
b = 2p − 3 of P(U ) and P(k) all belong to {k, U, (U |k), (k|U ), (k|U |k)}. Also, U ∼
= U ∗.
Hence we arrive at (ii) by Corollary 8.3.
(f3) Assume that (G+ , p, dim U ) = (M23 , 23, 21). Then U ∼
= U ∗ , P(k) = (k|U |k), and
the only quotient of length 2 of dimension ≤ 43 of P(U ) is (k|U ). Arguing as in the case
of Ap in (e1) and (e2), we arrive at (ii).
Consider the case (G+ , p, dim U ) = (M11 , 11, 9). Then U ∼
= U ∗ and P(k) = (k|U |k).
Using [46, Example 4.12.11] as above, we see that P(U ) has only two non-simple quotients
of dimension ≤ 19, namely (k|U ) and (W |U ) with W ∈ IBrp (G+ ) of dimension 10. Arguing
as above we arrive at (ii).
Suppose now that (G+ , p, dim U ) = (3A7 , 5, 3). Recall that U is a composition factor of
largest dimension of VG+ and Z(G+ ) acts via scalars on V . Using [34] one can then check
that all composition factors of VG+ are isomorphic to U . But Ext1G+ (U, U ) = 0 by Lemma
3.1. Hence VG+ is semisimple by Lemma 3.5(ii), a contradiction.
Suppose that (G+ , p, dim U ) = (3A6 , 5, 3). As in the case of (3A7 , 5, 3), we see that all
composition factors of VG+ are isomorphic to U . But dim V ≤ 7 and VG+ is indecomposable,
so head(VG+ ) ∼
= U . Inspecting the structure of P(U ) using [49], we conclude that VG+ ∼
=
(U |U ) is uniserial, i.e (vi) holds.
(g) Finally, let dim U = p − 3. By Theorem 2.1, we have that (G+ , p) = (2A7 , 7). As
in the case of (3A7 , 5, 3), we see that all composition factors of VG+ are isomorphic to
U . But dim V ≤ 11 and VG+ is indecomposable, so head(VG+ ) ∼
= U . Note that P(U ) =
(U |U ⊕ W |U ), where W ∈ IBrp (G+ ) has dimension 16 (as one can see using [49]). It follows
that VG+ ∼
= (U |U ), the unique quotient of dimension 8 of P(U ), and we arrive at (iv).
Lemma 8.5. Suppose that p = 3 and V is a reducible, faithful, indecomposable kG-module
of dimension ≤ 2p − 3. Then Op (G) 6= 1.
∼k
Proof. Suppose first that every G+ -composition factor of V is of dimension 1 and so =
′
+
+
+
+
p
(as G = O (G )). By faithfulness, G is a p-group; moreover G 6= 1 as otherwise G is
a p′ -group. Thus 1 6= G+ = Op (G).
Since VG+ is reducible, it remains to consider the case where VG+ has exactly two composition factors, U of dimension 2 and W of dimension 1, and moreover Op (G) = 1. Let
′
K denote the kernel of the action of G+ on U . Again, G+ = Op (G+ ) acts trivially on
W . It follows by faithfulness of G on V that K ≤ Op (G+ ) ≤ Op (G) = 1. Next, since
′
G+ = Op (G+ ), the image of G+ in GL(U ) is contained in SL(U ). Now if |G+ | is odd, then
G+ is solvable and so by the Fong-Swan theorem cannot act irreducibly on U of dimension
2. So G+ contains an element of order 2 which must then act as −1U and belong to Z(G+ ).
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
45
Thus U and W have different central characters and so VG+ is semisimple, contradicting
Lemma 3.7.
Next we will prove some criteria to decide the type of a self-dual indecomposable module.
Lemma 8.6. Let k = k be of characteristic p > 2 and let V be a self-dual indecomposable kG-module with dim EndkG (V ) ≤ 2. Then V supports a non-degenerate G-invariant
bilinear form that is either symmetric or alternating. Furthermore, all such forms have the
same, symmetric or alternating, type.
Proof. Let Φ denote the matrix representation of G on V relative to a fixed basis (e1 , . . . , en )
of V . Since V ∼
= V ∗ as G-modules, we can find b ∈ GLn (k) such that t Φ(g)−1 = bΦ(g)b−1 ,
and so b yields a non-degenerate G-invariant bilinear form on V . Note that the map
π : X 7→ bX yields a k-space isomorphism between EndkG (V ) and the space B of Ginvariant bilinear forms on V . In particular, dim B ≤ 2, and, since p > 2, it is a direct
sum S ⊕ A of symmetric and alternating G-invariant forms. Hence the claims follow if
dim EndkG (V ) = 1. Assume dim EndkG (V ) = 2. Since V is indecomposable, EndkG (V )
is a local algebra, cf. [46, Corollary 1.6.5], and its unique maximal ideal J, which then
has dimension 1, consists of (nilpotent) non-units. Thus π(J) is contained in the subset
D of degenerate G-invariant bilinear forms on V . But π −1 (D) is obviously contained in
J. It follows that D = π(J) is a subspace and dim D = 1. Hence we are also done if S
or A is zero. Assume S, A 6= 0, whence both of them are 1-dimensional. Now if Y ∈ D,
then t Y ∈ B and it is degenerate. As p > 2 and dim D = 1, it follows that t Y = ±Y .
Thus D is either S or A, and so the nonzero forms in the other subspace are precisely the
non-degenerate G-invariant forms on V that are either symmetric or alternating.
Lemma 8.7. Suppose that G is a finite group with a Sylow p-subgroup P of order p > 2
such that NG (P )/P is abelian. Let V be a reducible self-dual indecomposable G-module
over k = k of characteristic p, of even dimension d < 2p. Then V is not orthogonal if
d < p and V is not symplectic if d > p.
Proof. For 1 ≤ i ≤ p, let Xi denote the unique indecomposable kP -module of dimension
i (so Xp is projective). By the Green correspondence (see e.g. [46, Theorem 4.9.2]),
VN = X⊕Y for N := NG (P ), where X is non-projective indecomposable and Y is projective
(if nonzero). Let M denote any indecomposable kN -module. According to [1, p. 42], M
is uniserial. Also, Lemma 8 of [1, §5] says that the P -radical filtration agrees with the
N -radical filtration on M ; in particular, rad(M ) = rad(MP ). As N/P is abelian, any
irreducible kN -module remains irreducible as over P . It follows that MP is indecomposable.
Applying this to X and Y , we see that VP = Xd if d < p, respectively VP = Xd−p ⊕ Xp if
d > p. Now suppose that V is equipped with a non-degenerate G-invariant bilinear form of
a fixed parity. The claim then follows by using the description of Jordan forms of unipotent
elements in classical groups, see e.g. [45, Theorem 3.1].
Proof of Theorem 1.9. There is nothing to prove for p = 2; furthermore p 6= 3 by
Lemma 8.5. So we may assume p > 3. By Proposition 4.4, the self-duality of V implies
that G+ is quasisimple. If furthermore G+ is not a Lie-type group in characteristic p, then
by Theorem 8.4 we arrive at (i) and (ii). Assume that G+ is of Lie-type in characteristic
46
R. GURALNICK, F. HERZIG, AND P. TIEP
p. By Lemma 4.3(i), G+ ∼
= SL2 (q) or PSL2 (q) for some q = pa . By Corollary 4.5, VG+ is
indecomposable of length ≥ 2. Applying Proposition 3.10, we arrive at (i) and (ii).
Note that in each of the listed cases, there is a unique (up to isomorphism) reducible
indecomposable G+ -module V of the indicated shape (indeed, if W := head(VG+ ) then there
is a unique quotient of P(W ) of this shape). Since W ∗ ∼
=W ∼
= soc(VG+ ), it follows that
VG+ is self-dual. Thus all the listed cases give rise to examples of reducible indecomposable
self-dual modules (at least for G+ ).
It remains to determine the type of each indecomposable module. Note that in all cases
dim EndkG+ (V ) = 2, whence dim EndkG (V ) ≤ 2 and Lemma 8.6 applies to both G and G+ .
Thus V supports a non-degenerate G-invariant form that is either symmetric or alternating.
If dim V is odd, then all such forms must be symmetric. Consider the case dim V is even.
Note that in all cases |P | = p and NG+ (P )/P is abelian for P ∈ Sylp (G+ ). So by Lemma
8.7, all such forms are symmetric when dim V > p, respectively alternating when dim V < p.
✷
Recall from [56] that for G a connected reductive group over an algebraically closed field k
and for G ≤ G a subgroup we say that G is G-cr if whenever G ≤ P for a parabolic subgroup
P of G, then G is contained in a Levi subgroup of P. If G = Sp(V ) or SO(V ) for some
finite-dimensional vector space equipped with a non-degenerate alternating or symmetric
bilinear form, then this is equivalent to saying that for any G-stable isotropic subspace
W ⊂ V there exists a G-stable isotropic subspace W ′ ⊂ V with W + W ′ non-degenerate.
For these G and provided p > 2, a subgroup G ≤ G is G-cr if and only if the kG-module V
is completely reducible [56, §3.2.2].
We can extend Serre’s notion to the disconnected group G = O(V ) by saying that a
subgroup G is O(V )-cr if for any G-stable isotropic subspace W ⊂ V there exists a Gstable isotropic subspace W ′ ⊂ V with W + W ′ non-degenerate. We then see using the
same argument as in [56, §3.2.2] as well as Lemma 3.7(i), that for G ≤ O(V ) and p > 2 the
following are equivalent:
(i) G is O(V )-cr,
(ii) G ∩ SO(V ) is SO(V )-cr,
(iii) the kG-module V is completely reducible.
The next result shows that for G = Sp(V ) or O(V ), the finite non-G-cr subgroups of G
are made up from the groups with a nontrivial unipotent normal subgroup and the groups
described in Theorem 1.9.
Proposition 8.8. Let k = k be of characteristic p > 0 and let G be either Sp(V ) or O(V )
with dimk V ≤ 2p − 3. Suppose that G < G is a finite subgroup such that the G-module V is
not completely reducible. Then there is a G-invariant decomposition V = V1 ⊕ V2 ⊕ V3 of V
into an orthogonal direct sum of three subspaces, where Vi is either zero or non-degenerate,
at least one of the Vi ’s is zero and at least one of V1 and V2 is nonzero, and the following
conditions hold for the images Gi of G in GL(Vi ).
(i) If V1 6= 0, then Op (G1 ) = 1, the kG-module V1 is reducible indecomposable, and
(G1 , V1 ) is as described in Theorem 1.9.
(ii) If V2 6= 0, then Op (G2 ) 6= 1.
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
47
(iii) If V3 6= 0, then V3 is an orthogonal direct sum of non-degenerate subspaces, each
being an irreducible G-module.
Proof. (a) First note that p > 2. Setting V2 = V when Op (G) 6= 1, we may assume
Op (G) = 1. Setting V1 = V when VG is indecomposable, we may assume that VG is
decomposable.
First we consider the case where no composition factor of G has order p. Choose a
decomposition VG = A ⊕ B with A, B 6= 0 being G-invariant and A of smallest possible
dimension. Then dim A ≤ p − 2 and the image X of G in GL(A) has Op (X) = 1. By
[19], the X-module A is completely reducible, whence it is irreducible by its choice. If A is
non-degenerate, then VG = A ⊕ A⊥ . Consider the case A ∩ A⊥ 6= 0. By the irreducibility
of A, A ⊆ A⊥ and so A⊥ = A ⊕ C for C := B ∩ A⊥ . It is easy to see that C ∩ C ⊥ = 0
and so VG = C ⊕ C ⊥ . Note that C ⊥ 6= 0. Also, C 6= 0 as otherwise A⊥ = A, B ∼
= V /A =
V /A⊥ ∼
= A∗ is an irreducible G-module and so VG is semisimple, a contradiction. Thus in
either case V is an orthogonal direct sum of nonzero non-degenerate G-invariant subspaces.
Repeating this process for the summands, we obtain an orthogonal direct sum V = ⊕ni=1 Ui ,
where each Ui is non-degenerate and indecomposable as a kG-module, and n ≥ 2. Since
VG is not semisimple, we may assume that U1 is reducible. Again the image Yi of G in
GL(Ui ) has Op (Yi ) = 1. By Theorem 1.9, dim U1 ≥ p − 1, whence all Ui with i ≥ 2 must
be irreducible over G. Setting V1 = U1 and V3 = ⊕ni=2 Ui , we are done. Note that in this
case dim V > dim U1 ≥ p − 1.
(b) Let W1 , . . . , Wm denote all the composition factors of VG+ (with counting multiplicities) and let J := Op′ (G+ ).
Consider the case p = 3. If dim Wi = 1 for all i, then the first paragraph of the proof of
Lemma 8.5 shows that Op (G) 6= 1, contrary to our hypotheses. As VG is decomposable of
dimension ≤ 3, it follows that VG+ = W1 ⊕ W2 with {dim W1 , dim W2 } = {1, 2} and this
decomposition is G-invariant. Thus VG is completely reducible, again a contradiction. So
we must have that p > 3.
Suppose that J acts by scalars on each of the Wi ’s. Then, in a suitable basis of V ,
[J, G+ ] is represented by unitriangular matrices, and so it is a p-subgroup. But Op (G) = 1,
so J ≤ Z(G+ ). Applying Lemma 4.1 to G+ , we see that G+ , and so G as well, has no
composition factors of order p. Thus we are done by (a).
So we may now assume that J does not act by scalars on W1 . It follows that the image
of G+ in GL(W1 ) contains a non-scalar normal p′ -subgroup. Applying Theorem 2.1, we see
that dim W1 ≥ p − 1. Since dim V ≤ 2p − 3, it follows that J acts by scalars on each Wi
with i > 1, and m ≥ 2 as VG is reducible. Since J is a p′ -group, VJ ∼
= W1 ⊕ (⊕m
i=2 Wi ), and
this decomposition is G-invariant. It follows that G fixes a decomposition V = W1 ⊕ U
where UG+ has composition factors Wi , 2 ≤ i ≤ m. Since J acts by scalars on each Wi
with i > 1 but not on W1 , it also follows that U = W1⊥ , whence W1 is non-degenerate. If
Op (Y ) = 1 for the image Y of G in GL(U ), then UG is semisimple by [19] (as dim U ≤
p − 2), a contradiction. So we can now set V2 = U and V3 = W1 . Note that in this case
dim V > dim W1 ≥ p − 1.
48
R. GURALNICK, F. HERZIG, AND P. TIEP
Proof of Corollary 1.10. Suppose that the kG-module V is not completely reducible.
If VG is indecomposable, then we are done by Theorem 1.9. Otherwise, the proof of Proposition 8.8 shows that dim V ≥ p.
✷
9. Adequacy for SL2 (q)
The aim of this section is to prove the following statement which extends the results of
[23, §3]:
Proposition 9.1. Any nontrivial irreducible representation V of G := SL2 (pr ) over Fp is
weakly adequate, except when q := pr ≤ 3.
By the Steinberg tensor product theorem we can write
r−1
L(ai )(i)
V = L(a) := ⊗i=0
Pr−1
for some a = i=0
ai pi , 0 ≤ ai ≤ p − 1, where L(1) is the natural 2-dimensional Fp Grepresentation, L(b) = Symb (L(1)), and (i) denote the ith Frobenius twist. Also, G ∼
= SL2
denotes the underlying algebraic group for G.
Lemma 9.2. We have that
headG (End(V )) ∼
=
M
r−1
⊗i=0
L(2bi )(i) .
b0 ,...,br−1 : 0≤bi ≤min( p−1
,ai )
2
Moreover, if a < q − 1 then
headG (End(V )) = headG (End(V )),
whereas if a = q − 1, then
headG (End(V )) = headG (End(V )) ⊕ L(q − 1).
Proof. As End(V ) is self-dual, we may replace “head” by ”socle”. By [14, Lemmas 1.1 and
1.3], we have for 0 ≤ b ≤ (p − 1)/2
L(b) ⊗ L(b) ∼
= ⊕b T (2i),
i=0
and for (p − 1)/2 ≤ b ≤ p − 1,
⌊(p−1)/2⌋
p−2−b
L(b) ⊗ L(b) ∼
= ⊕i=0 T (2i) ⊕ ⊕i=p−1−b T (2p − 2 − 2i),
where T (λ) denotes the tilting module of G with highest weight λ ≥ 0. Recall that T (λ) =
L(λ) if λ ≤ p − 1, and that, when 0 ≤ λ ≤ p − 2, T (2p − 2 − λ) is uniserial of shape
(L(λ)|L(2p − 2 − λ)|L(λ)) and T (2p − 2 − λ) ∼
= Q1 (λ) in the notation of [2, §3].
r−1
The statement will follow if we can show for any 0 ≤ bi ≤ 2p−2 that (i) socG (⊗i=0
T (bi )(i) )
r−1
is simple, and (ii) socG (⊗i=0
T (bi )(i) ) is simple if bi < 2p−2 for at least one i and isomorphic
Pr−1 i
to L(0) ⊕ L(q − 1) otherwise. Let ci := min(bi , 2p − 2 − bi ) ≤ p − 1 and c := i=0
ci p .
Then
r−1
r−1
⊗i=0
T (bi )(i) ֒→ ⊗i=0
Q1 (ci )(i) = Qr (c)
in the notation of [2, §3]. By [2, Theorem 3.7], socG (Qr (c)) = L(c). Furthermore, by [2,
Lemma 4.1], socG (Qr (c)) = L(c) if c 6= 0 and socG (Qr (c)) = L(0) ⊕ L(q − 1) if c = 0 (note
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
49
that “⊗” should be “⊕” in [2, Lemma 4.1(b)]). Finally, if c = 0 then it is easy to check
r−1
T (bi )(i) , unless bi = 2p − 2 for all i.
that L(q − 1) does not occur in ⊗i=0
Proof of Proposition 9.1. By [23, Proposition 3.1] we may assume that r > 1. We will
follow the same strategy of proof. It suffices to show M = headG (End(V )), where M
denotes the span of the images of all p′ -elements of G in headG (End(V )).
Pr−1
Suppose that k = i=0
ki pi with 0 ≤ ki ≤ min((p − 1)/2, ai ). By [23, Lemma 3.5], the
r−1 (i)
G-subrepresentation L(2k) in End(V ) is generated by the weight 0 element ∆k := ⊗i=0
∆ki ,
where ∆ki ∈ End(L(ai )) is defined in [23, Lemma 3.5]. Let δk := tr(− ◦ ∆k ) ∈ (End(V ))∗ .
Pr−1 i
r−1 (i)
For ℓ = i=0
ℓi p with 0 ≤ ℓi ≤ ai , let πℓ := ⊗i=0
πℓi ∈ End(V ), where πℓi ∈ End(L(ai ))
is the projection X j Y ai −j 7→ δjℓi X j Y ai −j . For any other ℓ let πℓ := 0. Also let pk (ℓ) :=
Qr−1
pki (ℓi ), where pki (ℓi ) agrees with a polynomial of degree
δk (πℓ ) ∈ Fp . Then pk (ℓ) = i=0
ki for 0 ≤ ℓi ≤ ai . In particular, as ki ≤ ai , there exist 0 ≤ ℓi ≤ ai such that pki (ℓi ) 6= 0 for
all i. Thus pk (ℓ) 6= 0 for some ℓ.
Pr−1
(a) Suppose a < q − 1 and p > 2. Also, suppose that there exists a k = i=0
ki pi with
0 ≤ ki ≤ min((p − 1)/2, ai ) such that M does not contain L(2k). Then δk (M ) = 0, so the
action of the split Cartan subgroup gives
X
pk (ℓ) = 0, ∀ℓ′ .
ℓ≡ℓ′ (mod (q−1)/2)
As in [23, §3], the action of a non-split Cartan subgroup similarly gives
X
pk (ℓ) = 0, ∀ℓ′ .
ℓ≡ℓ′ (mod (q+1)/2)
Therefore, if 0 ≤ ℓ < (q − 1)/2, pk (ℓ) = −pk (ℓ + (q − 1)/2) = pk (ℓ − 1), and so by induction
pk (ℓ) = pk (ℓ − 1) = . . . = pk (−1) = 0. Similarly, pk (ℓ) = 0 for ℓ > a − (q − 1)/2, and so
pk (ℓ) = 0 for all ℓ, a contradiction.
(b) Now we consider the case a = q − 1 and p > 2.
(b1) Suppose that M does not contain L(2k) for some k < (q − 1)/2. The same argument
as in (a) shows that
pk (0) = pk (1) = . . . = pk ((q − 3)/2) = −pk ((q + 1)/2) = . . . = −pk (q − 1)
and pk ((q − 1)/2) = 0. Hence pki ((p − 1)/2) = 0 for some i. As r > 1 we deduce that
pk (ℓ) = 0 for some 0 ≤ ℓ < (q − 1)/2 (e.g. ℓ = pi (p − 1)/2), so pk (ℓ) = 0 for all ℓ, again a
contradiction.
(b2) Suppose that M does not contain L(q − 1)⊕2 . By [23, §3], the G-representation
i(i)
h
r−1
∂ (p−1)/2
is the unique G-subrepresentation L(q − 1) in
X ∂Y
generated by v := ⊗i=0
End(V ). Note that the upper-triangular Borel subgroup B := ( ∗ ∗∗ ) ⊂ G fixes v 2 =
h
i(i)
r−1
∂ p−1
X ∂Y
⊗i=0
and that v and v 2 are linearly independent. As v 2 ∈
/ (End(V ))G = Fp ,
the G-representation generated by v 2 is isomorphic to L(q − 1) or to L(q − 1) ⊕ L(0) ∼
=
2
IndG
B (1). In particular, for some c ∈ Fp , v + c generates the second copy of L(q − 1)
50
R. GURALNICK, F. HERZIG, AND P. TIEP
in End(V ). A calculation as in [23, §3] shows that c = (−1)r . Now we can deduce that
pk (ℓ) = 0 for all ℓ exactly as in part (b2) of the proof of [23, Proposition 3.1].
(c) Suppose now that p = 2. Note that headG (End(V )) is multiplicity-free. If M does not
contain L(0), then the argument in (a) (but using only a non-split Cartan subgroup) shows
that p0 (ℓ) = 0 for all ℓ (as q + 1 > a), a contradiction. (In fact, we could alternatively
use only a split Cartan subgroup, even when a = q = 1.) Suppose that a = q − 1 and
r−1
∂ (i)
+ 1 generates the unique Gthat M does not contain L(q − 1). By (b2), ⊗i=0
X ∂Y
subrepresentation L(q − 1) of End(V ). However, as in (b2) of the proof of [23, Proposition
3.1],
∂ (i)
α
r−1
)
6= 0
tr
◦ ⊗i=0 (X
α−1
∂Y
6 ∅, and this gives a final contradiction.
for any α ∈ F×
q \ {1} =
Remark 9.3. The results of [2] play a key role in our analysis of SL2 (q)-representations.
We should also point out some minor inaccuracies in [2, §4]. The first line of the displayed
formula right before [2, Corollary 4.5] should have the extra condition λ, µ 6= p − 1. Furthermore, in the case n = 2 of [2, Corollary 4.5(b)], there are four (not just two as stated)
cases when dim Ext1 = 2, namely when λ0 , λ1 ∈ {(p − 3)/2, (p − 1)/2} and µi = p − 2 − λi
for all i = 0, 1. (Also, the k and i in [2, Corollary 4.5(a)] satisfy 0 ≤ i, k ≤ n − 1.)
Corollary 9.4. Let V be nontrivial absolutely irreducible representation of G = SL2 (pr ) in
characteristic p. Then either V is adequate, or one of the following holds:
(i) r = 1, 1 < dim(V ) = (p ± 1)/2, and dim Ext1G (V, V ) = 1.
(ii) pr = 2, 3, 4 and dim V = pr .
(iii) pr = 9 and dim V = 3, 6, 9.
Proof. The case r = 1 is already treated by [23, Corollary 1.4], so we will assume r > 1.
In this case, Ext1G (V, V ) = 0 by [2, Corollary 4.5(a)]. Suppose that pr 6= 4, 9. Then
H 2 (G, k) = 0, and furthermore H 1 (G, k) = 0 as G is perfect. It follows that V is adequate.
The same conclusion holds if p ∤ dim(V ).
Now we consider the case pr = 4, 9 and keep the notation of the proof of Proposition 9.1.
The proof of Lemma 9.2 shows that the one-dimensional subspace End(V )G is contained
in the direct summand W := T (b0 ) ⊗ T (b1 )(1) , where bi = 0 if ai < p − 1 and bi = 2p − 2 if
ai = p − 1. As H 1 (G, End(V )) = 0, we deduce that H 1 (G, End(V )/k) = H 1 (G, W/k).
(a) If a0 = a1 = p − 1, then W = Q2 (0) ∼
= P(1) ⊕ L(p2 − 1) by [2, Lemma 4.1]. Hence
H 1 (G, End(V )/k) ∼
= H 1 (G, P(1)/k) ∼
= H 2 (G, k), which is one-dimensional.
(b) Suppose that precisely one of a0 , a1 is p − 1. Without loss we may assume that
a0 = p − 1 > a1 . Then W ∼
= T (2p − 2). Note that T (2p − 2) has composition factors
L(0) = 1 (twice) and L(p − 2) ⊕ L(1)(1) . As T (2p − 2) is self-dual and injects into Q2 (0),
we deduce that it is uniserial with trivial socle and head. Thus the sequence
0 → k → H 1 (G, L(p − 2) ⊗ L(1)(1) ) → H 1 (G, End(V )/k) → H 1 (G, k) = 0
is exact, whence
1
dim H (G, End(V )/k) =
dim Ext1G (k, L(p
(1)
− 2) ⊗ L(1)
)−1=
1, if p = 3,
0, if p = 2,
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
by [2, Corollary 4.5].
51
If one replaces SL2 (q) by GL2 (q), then in fact there are no exceptions to adequacy for
q > 3 odd (if q = 3 and dim V = 3, then weak adequacy fails).
Corollary 9.5. Let G be a finite group and V be a faithful absolutely irreducible representation of G in odd characteristic p. If the image of G in PGL(V ) is PGL2 (pa ) with pa > 3,
then (G, V ) is adequate.
Proof. Without loss we may assume that V is an irreducible kG-module with k = k. Let H
be the inverse image of PSL2 (pa ) under the projection from G onto PGL2 (pa ) < PGL(V ).
Then H/Z(H) ∼
= PSL2 (pa ) and Z(H) = Z(G) is a p′ -group. Since the universal p′ -cover
a
of PSL2 (p ) is SL2 (pa ), it follows that H = Z(H)L, where L := [H, H] is the quotient of
SL2 (pa ) by a central subgroup (of order 1 or 2). Moreover, G normalizes L and centralizes
Z(H), and in fact G induces the full subgroup of inner-diagonal automorphisms of L. It
is well known that any irreducible kL-representation is invariant under any inner-diagonal
automorphism. It follows that, if W is an irreducible kH-summand of V |H , then G preserves
the isomorphism class of W . But G/H ∼
= C2 , hence W extends to a kG-module W̃ (see
e.g. [50, Theorem 8.12]), and so by Frobenius’ reciprocity, V ∼
= W̃ ⊗k A for some onedimensional k(G/H)-module A. We have shown that VH is irreducible. By Proposition
9.1, (H, V ) is weakly adequate, so is (G, V ). Also, as Op (H) = H and G/H ∼
= C2 , we
see that Op (G) = G and so H 1 (G, k) = 0. Moreover, since p 6= 2, the inflation-restriction
sequence in cohomology implies that adequacy of (H, V ) yields the same for (G, V ). So by
Corollary 9.4, it suffices to consider the following cases:
(a) a = 1 and dim V = (p ± 1)/2;
(b) pa = 9 and dim V = 3, 6, 9.
In the first case, G has a cyclic Sylow p-subgroup P of order p. It follows that H 2 (G, k) = 0
and so it it suffices to show that Ext1G (V, V ) = 0. Note that V has no lifts to characteristic
0 (since NG (P ) acts transitively on the p − 1 nontrivial elements of P , an element g ∈ P
of order p would have at least p − 1 distinct eigenvalues in any characteristic 0 lift). Thus,
Ext1G (V, V ) = 0 by Lemma 3.1 and (G, V ) is adequate.
In the second case, note that H 2 (PGL2 (9), k) = 0 [10]. Since G/Z(G) ∼
= PGL2 (9)
and Z(G) is a p′ -group, it follows that H 2 (G, k) = 0. So again it suffices to show that
Ext1G (V, V ) = 0. Note that the p′ -group Z(H) acts trivially on V ⊗k V ∗ and H 1 (L, V ⊗k
V ∗ ) = Ext1L (V, V ) = 0 by [23, Lemma 8.1]. Hence, H 1 (H, V ⊗k V ∗ ) = 0. Since G/H ∼
= C2 ,
it follows that H 1 (G, V ⊗k V ∗ ) = 0, and so we are done.
10. Adequacy for SLn (q)
In this section, we give another family of examples of modules that are adequate. The
result follows from [9, Lemma 2.5.6] if p > n. Also, recall that the case n = 2 was considered
in the previous section.
Theorem 10.1. Let p be a prime and q a power of p. Let k be an algebraically closed field
of characteristic p and V = kn with n > 2. Suppose that G < GL(V ) is a finite group that
contains a normal subgroup S ∼
= SLn (q). Then (G, V ) is adequate.
52
R. GURALNICK, F. HERZIG, AND P. TIEP
Proof. By [40, Proposition 5.4.11], any nontrivial irreducible kS-representations of dimension ≤ n is quasi-equivalent to the natural n-dimensional kS-module U . It follows that the
kS-module V is irreducible and quasi-equivalent to U . Next, the only automorphisms of
S that preserve the isomorphism class of U (hence of V ) are the inner-diagonal automorphisms. Therefore, G induces only inner-diagonal automorphisms of S, and so p ∤ [G : S].
This implies that it is enough to prove the statement for S with V being the standard
representation.
Note that V ⊗ V ∗ = W ⊕ k with W irreducible if p ∤ n and that V ⊗ V ∗ is uniserial
(of length three) with trivial head and socle if p | n. Let W denote the unique nontrivial
irreducible composition factor of V ⊗ V ∗ . Since there are semisimple elements in S with
nonzero trace on V and since not all semisimple elements of S are scalars, it follows that
End(V ) is spanned by the images of the semisimple elements of S.
By the table in [37] or by [62, Theorem 9], it follows that Ext1S (V, V ) = 0 whence the
result holds as long as p ∤ n. If p | n, then using the fact that H 1 (S, k) = 0 and H 0 (S, W ) = 0
and the long exact sequence for cohomology we see that H 1 (S, V ⊗ V ∗ /k) = 0 if and only
if dim H 1 (S, W ) = 1. By [37], this is the case and so the result follows (one can give an
alternate proof using [62] as well).
A slight modification of the proof shows that if gcd(n, q) = 1, then (G, V ) is in fact
big. Indeed, we only need observe the obvious fact that there exists a semisimple regular
element g ∈ SLn (q) with nonzero trace.
11. Asymptotic adequacy
In this section, we extend [21, Theorem 1.2] to include disconnected groups as well as
to allow the possibility that p divides dim V . First we prove some statements relating
discrete cohomology (i.e. of abstract groups) and rational cohomology (i.e. in the category
of rational modules), of linear algebraic groups on the one side, and cohomology of finite
groups of Lie type on the other side. We use subscripts disc and rat to make distinction
between these two types of cohomology groups.
First we record the following result (which essentially is a special case of a result of van
der Kallen, cf. [52, p. 239]):
Lemma 11.1. Let k = Fp and let G be a linear algebraic group defined over Fq ⊂ k. Let
V be a finite-dimensional rational kG(k)-module. If H 1 (G(Fqf ), V ) = 0 for large enough f ,
1 (G(k), V ) = 0.
then Hdisc
Proof. First we note that if U is any finite-dimensional kG(k)-module, then G(Fqf ) and G(k)
have the same subspace of fixed points on U when f is divisible by some integer N = N (U ).
Indeed, let Uj denote the fixed point subspace for G(Fqj! ) on U . Then U1 ⊇ U2 ⊇ . . ., and
so Uj stabilizes when j ≥ j0 for some j0 . But each element of G(k) is contained in G(Fqj! )
for some j ≥ j0 . It follows that CU (G(k)) = Uj0 = CU (G(Fqj! )) for all j ≥ j0 . In particular,
we can choose N = j0 !.
Consider any exact sequence 0 → V → W → k → 0 of kG(k)-modules. By assumption,
it is split over G(Fqf ) for f large enough. Hence CW (G(Fqf )) has dimension equal to
dimk CV (G(Fqf ))+1, which by our claim is equal to dimk CV (G(k))+1 when N (U )|f . Again
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
53
by our claim, CW (G(Fqf )) = CW (G(k)) for N (W )|f . It follows that dimk CW (G(k)) =
1 (G(k), V ) = 0.
dimk CV (G(k)) + 1, whence W is split over G(k). Hence Hdisc
Next we observe that the results of [35, Prop. II.2.14] and [8, Theorem 6.6] hold in more
generality than they were stated.
Proposition 11.2. Let k be an algebraically closed field of characteristic p and let G be
a (not necessarily connected) reductive algebraic group defined over k. Let V be a finitedimensional rational kG-module.
(i) Suppose that p ∤ [G : G 0 ] and V is irreducible. Then Ext1G (V, V )rat = 0.
(ii) Suppose G is connected and defined over Fq ⊂ k. Then for e and f large enough
(depending on V and n),
n
Hrat
(G, V (e)) ∼
= H n (G(Fqf ), V ),
= H n (G(Fqf ), V (e)) ∼
where V (e) is the e-th Frobenius twist of V .
Proof. (i) Since p ∤ [G : G 0 ], it suffices to show that Ext1G 0 (X, Y )rat = 0 for any two
irreducible G 0 -submodules X, Y of V . Assume the contrary: Ext1G 0 (X, Y )rat 6= 0 for some
such X and Y . By Clifford’s theorem, Y ∼
= g(X) for some g ∈ G. Given a pair (T , B) of a
maximal torus T and a Borel subgroup B containing T of G 0 , we have that g−1 (T , B)g =
h−1 (T , B)h for a suitable h ∈ G 0 . Replacing g by gh−1 , we get that g normalizes both T
and B, and Y ∼
= g(X). Suppose that X = L(λ) for some dominant weight λ with respect to
T . Then τ (λ) is dominant and Y = L(τ (λ)), where τ is the outer automorphism (possibly
trivial) of G 0 induced by g. By [35, Proposition II.2.14], Ext1G 0 (X, Y )rat 6= 0 implies that
λ 6= τ (λ) but λ and τ (λ) are comparable, say λ > τ (λ). Since τ fixes (T , B), it fixes the set
of positive roots with respect to T , whence τ i (λ) > τ i+1 (λ) for all i ≥ 0. Also, note that
the action of τ on the weight lattice X(T ) has finite order N . Thus we arrive at the chain
λ > τ (λ) > τ 2 (λ) > . . . > τ N (λ) = λ,
a contradiction.
(ii) Consider the action of the central torus Z := Z(G)0 on V and decompose V =
n (G, [Z, V ]) = 0, and so H n (G, V ) ∼ H n (G, V ′ ).
V ⊕ [Z, V ] with V ′ := CV (Z). Then Hrat
= rat
rat
The same holds for Frobenius twists V (e) of V ; moreover, V ′ (e) ∼
= CV (e) (Z). Applying [8,
Theorem 6.6] to the semisimple group H := G/Z (and recalling that Z is a torus), we get
for e and f large enough that
′
n
n
Hrat
(G, V ′ (e)) ∼
(H, V ′ (e)) ∼
= H n (H(Fqf ), V ′ ).
= Hrat
= H n (H(Fqf ), V ′ (e)) ∼
On the other hand, H(Fqf ) is isomorphic to G(Fqf )/Z(Fqf ) (by the Lang-Steinberg theorem). Moreover, for f large enough, Z and Z(Fqf ) have the same eigenspaces on V . It
follows that [Z, V ] = [Z(Fqf ), V ] and V ′ = CV (Z(Fqf )), whence
H n (G(Fqf ), V ) = H n (G(Fqf ), V ′ ) = H n (H(Fqf ), V ′ )
(as Z(Fqf ) is a p′ -group). The same holds for V (e), and so the statement follows.
54
R. GURALNICK, F. HERZIG, AND P. TIEP
Lemma 11.3. Let p be a prime and let k be an algebraically closed field of characteristic p. Let G be a connected reductive algebraic group over k and V be a rational G1 (G, V (e)) = 0 for all Frobenius twists V (e) of V with e large enough, then
module. If Hrat
1
Hdisc (G(k), V ) = 0.
Proof. If the result fails, then there exists a (possibly non-rational) kG(k)-module W and
a non-split extension 0 → V → W → k → 0.
Let K be the algebraic closure of Fp in k. Note that G can be defined over Fq ⊂ k for
q sufficiently large (as it can be defined over K by the isomorphism theorem for reductive
groups). Also, let T (k) be a maximal torus of G(k) containing a maximal torus T (K) of
G(K). For e and f large enough, we have by assumption and by Proposition 11.2(ii) that
1 (G, V (e)) = 0. It follows by Lemma 11.1 that W is split over G(K),
H 1 (G(Fqf ), V ) = Hrat
whence
(11.1) dimk CW (G(K)) = dimk CV (G(K)) + 1, dimk CW (T (K)) = dimk CV (T (K)) + 1.
We claim that CW (T (K)) = CW (T (k)). Clearly, the fixed point subspace U := CW (T (K))
is T (k)-invariant. Also, since V is a rational kG(k)-module, T (k) acts trivially on U ∩ V =
CV (T (K)), which has codimension 1 in U by (11.1). Thus, T (k) maps into a unipotent
subgroup of GL(U ). Note that T (k) is p-divisible, and so is any homomorphic image of it.
It follows that T (k) acts trivially on U , as stated.
Thus, the fixed point subspace of hT (k), G(K)i on W is the fixed point subspace of
hT (K), G(K)i = G(K) on W . Observe that G(k) = hT (k), G(K)i. (Indeed, if Uα (k) ⊇
Uα (K) are root subgroups corresponding to a root α with respect to T , then T (k) acts
transitively on Uα (k) \ {1}. Since G(K) ⊃ Uα (K) and G(k) is generated by T (k) and all
the root subgroups Uα (k), the claim follows.) Hence, CW (G(k)) = CW (G(K)) and so
dimk CW (G(k)) ≥ dimk CV (G(k)) + 1
by (11.1). Hence W is split as a G(k)-module, a contradiction.
Corollary 11.4. Let k be an algebraically closed field of characteristic p and let G be a
reductive algebraic group over k. Let V be an irreducible rational kG(k)-module. Assume
that p ∤ [G : G 0 ]. Then
H 1 (H(k), k) = Ext1H(k) (V, V ) = H 1 (H(k), (V ∗ ⊗ V )/k) = 0,
both as rational and discrete cohomology groups, and for both H = G, G 0 .
Proof. Since p ∤ [G : G 0 ], it suffices to prove the statement for G 0 . As shown in Proposition
i (G 0 , k) = 0 for i > 0 by [35, Corollary II.4.11]. We
11.2(i), Ext1G 0 (V, V )rat = 0. Also, Hrat
1 (G 0 , k) = H 1 (G 0 , (V ∗ ⊗ V )/k) = 0. The same applies to
have therefore shown that Hrat
rat
1 (G 0 (k), k) = H 1 (G 0 (k), (V ∗ ⊗
Frobenius twists. Hence Ext1G 0 (k) (V, V )disc = 0 and Hdisc
disc
V )/k) = 0 by Lemma 11.3.
We finally show that adequacy holds over a sufficiently large field and also for (not
necessarily connected) reductive algebraic groups (whether one uses rational cohomology
or discrete cohomology in the definition). Note that if p does divide [G : G 0 ], then adequacy
may fail (the spanning may fail as well as the cohomological conditions even assuming that
p ∤ (dim V ) – one can construct examples precisely as in [20]).
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
55
Theorem 11.5. Let k be an algebraically closed field k of characteristic p. Let G be a
reductive algebraic group defined over Fq ⊂ k such that p ∤ [G : G 0 ], and let V be a finitedimensional faithful irreducible rational kG-module. Then the following statements hold.
(i) (G, V ) is adequate.
(ii) Assume that every coset of G 0 in G is defined over Fq . Then (G(Fqf ), V ) is adequate
for f sufficiently large (with f possibly depending upon V ).
Proof. (a) Arguing precisely as in [22], we see that the set of semisimple elements in any
coset of G 0 is Zariski dense in G. It follows that the linear span of semisimple elements of
G is Zariski dense in the linear span of G in End(V ). Thus, the span of the semisimple
elements in G is all of End(V ).
Let T ⊂ B be a maximal torus and a Borel subgroup of G 0 that are defined over Fq .
Choose f large enough so that T (Fqf ) has exactly the same weight spaces on End(V ) and
V as does T . Let N := NG (T , B) denote the simultaneous normalizer of (T , B) in G. Then
N ∩ G 0 = T , hence every element of N is semisimple (as p ∤ [G : G 0 ]). Conversely, if
g ∈ G is semisimple, then by [58, 7.5] g normalizes some pair (T ′ , B ′ ) of a maximal torus
T ′ contained in a Borel subgroup B ′ . We deduce that
(11.2)
Gss = ∪x∈G xN x−1 ,
where Gss denotes the set of semisimple elements in G.
Let W be the linear span of G(Fqf )ss := G(Fqf ) ∩ Gss in End(V ). Then W is G(Fqf )stable and hence in particular T -stable (as T and T (Fqf ) have the same eigenspaces on
End(V )). Arguing as in [21], we see that hT, G(Fqf )i is Zariski dense in G. It follows that
W is G-stable. Since N /T ֒→ G/G 0 and every coset of G 0 is defined over Fq , we deduce
by Lang’s theorem that N = N (Fqf ) · T . Moreover, as T (Fqf ) and T have the same
eigenspaces on V , T (Fqf ) and T span the same subspace of End(V ). Now W contains the
span of N (Fqf ) ⊂ G(Fqf )ss , hence contains the span of N . Since W is G-stable, we deduce
from (11.2) that W contains the span of Gss . Thus for f sufficiently large we have that
W = End(V ); in particular, G(Fqf ) acts absolutely irreducibly on V .
1 (G(k), k) = H 1 (G(k), (V ∗ ⊗ V )/k) = 0.
(b) From Corollary 11.4 we get that Hdisc
disc
Together with (a), this implies (i).
1 (G 0 , k) = H 1 (G 0 , (V ∗ ⊗ V )/k) = 0, and the same holds for
We also have that Hrat
rat
all Frobenius twists. Applying Proposition 11.2(ii) we obtain (for f large enough) that
H 1 (G 0 (Fqf ), k) = H 1 (G 0 (Fqf ), (V ∗ ⊗ V )/k) = 0, whence H 1 (G(Fqf ), k) = H 1 (G(Fqf ), (V ∗ ⊗
V )/k) = 0 as well since p ∤ [G : G 0 ]. Hence (ii) holds.
References
[1] J. L. Alperin, ‘Local Representation Theory. Modular Representations as an Introduction to the
Local Representation Theory of Finite Groups, Cambridge Studies in Advanced Mathematics, 11.
Cambridge University Press, Cambridge, 1986.
[2] H. Andersen, J. Jorgensen, and P. Landrock, The projective indecomposable modules of SL(2, pn ).
Proc. London Math. Soc. (3) 46 (1983), 38–52.
[3] C. Bendel, D. Nakano, and C. Pillen, On the vanishing ranges for the cohomology of finite groups
of Lie type II. Recent developments in Lie algebras, groups and representation theory, 25–73, Proc.
Sympos. Pure Math., 86, Amer. Math. Soc., Providence, RI, 2012.
56
R. GURALNICK, F. HERZIG, AND P. TIEP
[4] D. J. Benson, ‘Representations and Cohomology. I’ 2nd ed., Cambridge Stud. Adv. Math. 30,
Cambridge University Press, Cambridge 1998.
[5] H. I. Blau and J. Zhang, Linear groups of small degree over fields of finite characteristic, J. Algebra
159 (1993), 358–386.
[6] G. Böckle, A local-to-global principle for deformations of Galois representations, J. reine angew
Math. 509 (1999), 199–236.
[7] R. Burkhardt, Die Zerlegungsmatrizen der Gruppen PSL(2, pf ), J. Algebra 40 (1976), 75–96.
[8] E. Cline, B. Parshall, L. Scott, and Wilberd van der Kallen, Rational and generic cohomology.
Invent. Math. 39 (1977), 143–163.
[9] L. Clozel, M. Harris, and R. Taylor, Automorphy for some l-adic lifts of automorphic mod l Galois
representations. With Appendix A, summarizing unpublished work of Russ Mann, and Appendix
B by Marie-France Vignéras, Publ. Math. Inst. Hautes Études Sci. No. 108 (2008), 1–181.
[10] J. H. Conway, R. T. Curtis, S. P. Norton, R. A. Parker, and R. A. Wilson, ‘An ATLAS of Finite
Groups’, Clarendon Press, Oxford, 1985.
[11] F. E. Diederichsen, Über die Ausreduction ganzzahliger Gruppendarstellungen bei arithmetischer
Äquivalenz, Abh. Math. Sem. Univ. Hamburg 13 (1938), 357–412.
[12] L. Dieulefait, Automorphy of Symm5 (GL(2)) and base change, J. Math. Pures Appl. 104 (2015),
619–656.
[13] F. Digne and J. Michel, ‘Representations of Finite Groups of Lie Type’, London Mathematical
Society Student Texts 21, Cambridge University Press, 1991.
[14] S. Doty and A. Henke, Decomposition of tensor products of modular irreducibles for SL2 , Quart.
J. Math. 56 (2005), 189–207.
[15] P. Fong and B. Srinivasan, Brauer trees in classical groups, J. Algebra 131 (1990), 179–225.
[16] The GAP group, ‘GAP - groups, algorithms, and programming’, Version 4.4, 2004, http://www.gapsystem.org.
[17] M. Geck, Irreducible Brauer characters of the 3-dimensional special unitary groups in non-describing
characteristic, Comm. Algebra 18 (1990), 563–584.
[18] D. Gorenstein, R. Lyons, and R. Solomon, ‘The Classification of the Finite Simple Groups’, Number
3, Mathematical Surveys and Monographs Volume 40, American Math. Soc. 1998.
[19] R. M. Guralnick, Small dimensional representations are semisimple, J. Algebra 220 (1999), 531–541.
[20] R. M. Guralnick, Adequate subgroups II, Bull. Math. Sci. 2 (2012), 193–203.
[21] R. M. Guralnick, Adequacy of representations of finite groups of Lie type, Appendix A to: L.
Dieulefait, Automorphy of Symm5 (GL(2)) and base change, J. Math. Pures Appl. 104 (2015),
619–656.
[22] R. M. Guralnick, F. Herzig, R. Taylor, and J. Thorne, Adequate subgroups, J. Inst. Math. Jussieu
11 (2012), 907–920.
[23] R. M. Guralnick, F. Herzig, and Pham Huu Tiep, Adequate groups of low degree, Algebra & Number
Theory 9 (2015), 77–147.
[24] R M. Guralnick, K. Magaard, J. Saxl, and Pham Huu Tiep, Cross characteristic representations of
odd characteristic symplectic groups and unitary groups, J. Algebra 257 (2002), 291–347.
[25] R. M. Guralnick, G. Navarro, and Pham Huu Tiep, Real class sizes and real character degrees,
Math. Proc. Camb. Philos. Soc. 150 (2011), 47–71.
[26] R. M. Guralnick and Pham Huu Tiep, Low-dimensional representations of special linear groups in
cross characteristic, Proc. London Math. Soc. 78 (1999), 116–138.
[27] R. M. Guralnick and Pham Huu Tiep, The non-coprime k(GV ) problem, J. Algebra 293 (2005),
185–242.
[28] R. M. Guralnick and Pham Huu Tiep, Symmetric powers and a problem of Kollár and Larsen,
Invent. Math. 174 (2008), 505–554.
[29] G. Hiss and K. Lux, ‘Brauer Trees of Sporadic Groups’, Clarendon Press, 1989, 525 pp.
[30] G. Hiss and G. Malle, Low dimensional representations of special unitary groups, J. Algebra 236
(2001), 745–767.
ADEQUATE SUBGROUPS AND INDECOMPOSABLE MODULES
57
[31] G. Hiss and G. Malle, Corrigenda: Low-dimensional representations of quasi-simple groups, LMS
J. Comput. Math. 5 (2002), 95–126.
[32] I. M. Isaacs, ‘Character Theory of Finite Groups’, AMS-Chelsea, Providence, 2006.
[33] C. Jansen, The minimal degrees of faithful representations of the sporadic simple groups and their
covering groups, LMS J. Comput. Math. 8 (2005), 122–144.
[34] C. Jansen, K. Lux, R. A. Parker, and R. A. Wilson, ‘An ATLAS of Brauer Characters’, Oxford
University Press, Oxford, 1995.
[35] J. C. Jantzen, Representations of algebraic groups. Second edition. Mathematical Surveys and
Monographs, 107. American Mathematical Society, Providence, RI, 2003.
[36] J. C. Jantzen, Low-dimensional representations of reductive groups are semisimple, in: ‘Algebraic
Groups and Lie Groups’, 255–266, Austral. Math. Soc. Lect. Ser., 9, Cambridge Univ. Press, Cambridge, 1997.
[37] W. Jones and B. Parshall, On the 1-cohomology of finite groups of Lie type. Proceedings of the
Conference on Finite Groups (Univ. Utah, Park City, Utah, 1975), pp. 313–328. Academic Press,
New York, 1976.
[38] W. M. Kantor, Linear groups containing a Singer cycle, J. Algebra 62 (1980), 232–234.
[39] N. M. Katz, Exponential sums and differential equations, Annals of Math. Studies 124, Princeton
Univ. Press, Princeton, NJ 1999.
[40] P. B. Kleidman and M. W. Liebeck, ‘The Subgroup Structure of the Finite Classical Groups’, London
Math. Soc. Lecture Note Ser. no. 129, Cambridge University Press, 1990.
[41] P. B. Kleidman, The maximal subgroups of the Chevalley groups G2 (q) with q odd, the Ree groups
2
G2 (q), and their automorphism groups, J. Algebra 117 (1988), 30–71.
[42] V. Landazuri and G. Seitz, On the minimal degrees of projective representations of the finite Chevalley groups, J. Algebra 32 (1974), 418–443.
[43] M. W. Liebeck, Permutation modules for rank 3 unitary groups, J. Algebra 88 (1984), 317- 329.
[44] M. W. Liebeck, E. O’Brien, A. Shalev, and Pham Huu Tiep, The Ore conjecture, J. Eur. Math.
Soc. 12 (2010), 939–1008.
[45] M. W. Liebeck and G. M. Seitz, ‘Unipotent and Nilpotent Classes in Simple Algebraic Groups and
Lie Algebras’, Math. Surveys and Monographs, vol. 180, AMS, Providence, RI, 2012.
[46] K. Lux and H. Pahlings, ‘Representations of Groups: A Computational Approach’, Cambridge Univ.
Press, 2010.
[47] B. Mazur, Deforming Galois representations, in: ‘Galois Groups over Q (Berkeley, 1987)’, SpringerVerlag, Berlin-Heidelberg-New York, 1989, pp. 385–437.
[48] G. McNinch, Dimensional criteria for semisimplicity of representations. Proc. London Math. Soc.
76 (1998), 95–149.
[49] Decomposition matrices, http://www.math.rwth-aachen.de/homes/MOC/decomposition/
[50] G. Navarro, ‘Characters and Blocks of Finite Groups’, Cambridge University Press, 1998.
[51] G. Navarro and Pham Huu Tiep, Characters of p′ -degree over normal subgroups, Annals of Math.
178 (2013), 1135–1171.
[52] B. J. Parshall, Cohomology of algebraic groups, in: The Arcata Conference on Representations of
Finite Groups (Arcata, Calif., 1986), 233–248, Proc. Sympos. Pure Math. 47, Part 1, Amer. Math.
Soc., Providence, RI, 1987.
[53] R. M. Peacock, Blocks with a cyclic defect group, J. Algebra 34 (1975), 232–259.
[54] M. H. Peel, Hook representations of the symmetric groups, Glasgow Math. J. 12 (1971), 136–149.
[55] J.-P. Serre, Sur la semi-simplicité des produits tensoriels de représentations de groupes, Invent.
Math. 116 (1994), 513–530.
[56] J.-P. Serre, Complète réductibilité, Séminaire Bourbaki Vol. 2003/2004, Astérisque No. 299 (2005),
Exp. No. 932, viii, 195–217.
[57] P. Sin and Pham Huu Tiep, Rank 3 permutation module of the finite classical groups, J. Algebra
291 (2005), 551–606.
[58] R. Steinberg, Endomorphisms of linear algebraic groups. Memoirs of the American Mathematical
Society, No. 80, American Mathematical Society, Providence, R.I. 1968
58
R. GURALNICK, F. HERZIG, AND P. TIEP
[59] J. Thorne, On the automorphy of l-adic Galois representations with small residual image. With an
appendix by Robert Guralnick, Florian Herzig, Richard Taylor and Thorne, J. Inst. Math. Jussieu
11 (2012), 855–920.
[60] J. Thorne, A 2-adic automorphy lifting theorem for unitary groups over CM fields, Math. Z. (2016),
DOI:10.1007/s00209-016-1681-2.
[61] Pham Huu Tiep, Finite groups admitting grassmannian 4-designs, J. Algebra 306 (2006), 227–243.
[62] O. Taussky and H. Zassenhaus, On the 1-cohomology of the general and special linear groups,
Aequationes Math. 5 (1970), 129–201.
[63] Pham Huu Tiep and A. E. Zalesskii, Minimal characters of the finite classical groups, Commun.
Algebra 24 (1996), 2093–2167.
[64] Pham Huu Tiep and A. E. Zalesskii, Some characterizations of the Weil representations of symplectic
and unitary groups, J. Algebra 192 (1997), 130–165.
[65] T. E. Zieschang, Primitive permutation groups containing a p-cycle, Arch. Math. 64 (1995), 471–474.
[66] K. Zsigmondy, Zur Theorie der Potenzreste, Monatsh. Math. Phys. 3 (1892), 265–284.
Department of Mathematics, University of Southern California, Los Angeles, CA 900892532, USA
E-mail address: [email protected]
Department of Mathematics, University of Toronto, 40 St. George Street, Room 6290,
Toronto, ON M5S 2E4, Canada
E-mail address: [email protected]
Department of Mathematics, University of Arizona, Tucson, AZ 85721-0089, USA
E-mail address: [email protected]
| 4math.GR
|
1
Hydra: an Ensemble of Convolutional Neural
Networks for Geospatial Land Classification
arXiv:1802.03518v1 [cs.CV] 10 Feb 2018
Rodrigo Minetto, Maurı́cio Pamplona Segundo, Sudeep Sarkar
Abstract—We describe in this paper Hydra, an ensemble of convolutional neural networks (CNN) for geospatial land classification.
The idea behind Hydra is to create an initial CNN that is coarsely optimized but provides a good starting pointing for further
optimization, which will serve as the Hydra’s body. Then, the obtained weights are fine tuned multiple times to form an ensemble of
CNNs that represent the Hydra’s heads. By doing so, we were able to reduce the training time while maintaining the classification
performance of the ensemble. We created ensembles using two state-of-the-art CNN architectures, ResNet and DenseNet, to
participate in the Functional Map of the World challenge. With this approach, we finished the competition in third place. We also applied
the proposed framework to the NWPU-RESISC45 database and achieved the best reported performance so far. Code and CNN
models are available at https://github.com/maups/hydra-fmow.
Index Terms—Geospatial land classification, remote sensing image classification, functional map of world, ensemble learning, on-line
data augmentation, convolutional neural network.
F
1
I NTRODUCTION
Land use is a critical piece of information for a wide
range of applications, from humanitarian to military purposes. For this reason, automatic land use classification
from satellite images has been drawing increasing attention
from academia, industry and government agencies. One of
the most recent initiatives in this area is the Functional
Map of the World (FMOW) challenge [1] sponsored by the
Intelligence Advanced Research Projects Activity (IARPA),
an organization within the Office of the USA Director of
National Intelligence.
The FMOW challenge consists of creating automatic
solutions to classify a specific given location as one of the
62 target classes (e.g. airport, flooded road, nuclear power
plant and so on) or as none of them (false detections).
FMOW images vary in quality and are distributed over
more than 100,000 globe locations, which leads to high
intraclass variations and considerable interclass confusion.
This, added to traditional satellite imaging problems like
viewpoint, weather, shadow and scale variations, makes this
classification problem a lot harder than previous land use
datasets, such as UC Merced Land Use Dataset [2], WHURS19 [3] and NWPU-RESISC45 [4]. Finally, the FMOW
challenge also limits time and computational resources for
training and testing to minimize the disparity among participants’ solutions.
To cope with the difficulties and restrictions of the
FMOW challenge, as our main contribution, in this work
•
•
•
•
R. Minetto is with Universidade Tecnológica Federal do Paraná (UTFPR),
Brazil. E-mail: [email protected]
M. P. Segundo is with Universidade Federal da Bahia (UFBA), Brazil.
E-mail: [email protected]
S. Sarkar is with Department of Computer Science and Engineering, University of South Florida (USF), Tampa, FL, USA. E-mail: [email protected]
The research in this paper was conducted while the authors were at the
Computer Vision and Pattern Recognition Group, USF.
we present a framework that creates ensembles of Convolutional Neural Networks (CNN) for land use classification in satellite images, which we called Hydra. The idea
behind Hydra is to create an initial CNN that is coarsely
optimized but provides a good starting pointing for further
optimization, which will serve as the Hydra’s body. Then,
the obtained weights are fine tuned multiple times to form
an ensemble of CNNs that represent the Hydra’s heads. The
Hydra framework tackles one of the most common problem
in multiclass classification, which is the existence of several
local minima that prioritize some classes over others and the
eventual absence of a global minimum within the classifier
search space. The ensemble ends up expanding this space
by combining multiple classifiers that converged to local
minima and reaches a better global approximation. Figure 1
illustrates this process, where a black line represents the
body optimization part and the red lines represent the heads
reaching different local minima. To stimulate convergence to
different end points, we exploit different strategies, such as
using online data augmentation, variations in the size of the
region of interest, and different image formats released by
FMOW.
2
T HE FMOW C HALLENGE
The main goal of the FMOW challenge was to encourage
researchers and machine learning experts to design automatic classification solutions for land use interpretation
in satellite images. The competition was hosted by TopCoder1 , where participants had one submission every three
hours to keep track of their ongoing performance. Details
regarding dataset, scoring, and restrictions are given in
Sections 2.1, 2.2, and 2.3, respectively.
1. https://www.topcoder.com/
2
(a)
(b)
(c)
Fig. 1. Illustration of the optimization process in the Hydra framework.
The black line represents the coarse parameter optimization that forms
the Hydra’s body. The red lines represent the Hydra’s heads, which are
obtained after fine tuning the body parameters multiple times seeking to
reach different local minima.
2.1
Dataset
The FMOW dataset, as detailed by Christie et al. [1], contains
almost half of a million images split into training, evaluation
and testing subsets, as presented in Table 1. It was publicly
released2 in two image formats: JPEG (194 GB) and TIFF (2.8
TB). For both formats, high-resolution pan-sharpened [5]
and multi-spectral images were available, although the latter one varies with the format. While TIFF images provided
the original 4/8-band multi-spectral information, the color
space was compressed in JPG images to fit their 3-channel
limitation.
Fig. 2. Different challenges in FMOW satellite images: (a) shadow and
viewpoint variations, (b) arbitrary weather conditions, (c) inaccurate
annotations and (d) scale variations.
TABLE 1
Table of contents of the FMOW dataset.
Training
Evaluation
Testing
Total
# of images
# of boxes
# of distinct boxes
363,572
53,041
53,473
470,086
363,572
63,422
76,988
503,982
83,412
12,006
16,948
112,366
Pan-sharpened image dimensions range from 293 × 230
to 16184 × 16288 pixels, and multi-spectral from 74 × 58
to 4051 × 4077. Each image has one or more annotated
boxes that correspond to regions of interest. Some of these
regions may appear in several images taken in different
time periods, adding a temporal element to the problem,
as shown in Figures 2(a) and 2(b). Traditional problems
arising from that include, but are not limited to, variations
in shadows, satellite viewpoint and weather conditions. Figures 2(c) and 2(d) illustrate other recurrent problems in the
FMOW dataset: inaccurate annotations and large variations
in region size.
Each region is an instance of one of the 62 classes
or a false detection. All classes considered in the FMOW
challenge are illustrated in Figure 3. Those classes have an
unbalanced distribution, as shown in Figure 4. Images in
the training subset contain a single region that is never a
2. https://www.iarpa.gov/challenges/fmow.html
(d)
false detection. The evaluation subset has images with one
or two regions, with at least one not being a false detection.
Testing images may have several regions but their labels are
unknown (see Figure 2(d)). From 53,473 images in the test
subset, 82.60% have a single region, 4.96% have two regions,
5.66% have three regions and 6.78% have four regions or
more, with a maximum number of fourteen regions per
image.
The FMOW dataset also provides different metadata
along with each region, such as bounding box and image
size, ground sample distance, region visibility, cloud cover,
timestamps, wavelengths, scan direction, sun elevation,
Universal Transverse Mercator (UTM) zone, and so on.
2.2
Scoring
Submissions consisted of a list with one label prediction
— among the m possible classes (m = 63, i.e. 62 land use
classes plus false detection) — for each distinct region in the
test set. When the same region appeared in several images,
like the ones in Figures 2(a) and 2(b), the competitors
were asked to merge all these information into a single
label outcome. The quantitative criteria used by the FMOW
challenge to rank its competitors was a weighted average of
the F -measure [6] for each class.
3
Airport
(w: 0.6)
Airport hangar
(w: 1.0)
Airport terminal
(w: 1.0)
Amusement
park (w: 1.0)
Aquaculture
(w: 1.0)
Archaeological
site (w: 1.0)
Barn
(w: 1.0)
Border checkpoint
(w: 1.4)
Burial site
(w: 1.0)
Car dealership
(w: 1.0)
Construction
site (w: 1.4)
Crop field
(w: 0.6)
Dam
(w: 1.0)
Debris or rubble
(w: 0.6)
Educational institution (w: 1.4)
Electric substation
(w: 1.0)
Factory or power
plant (w: 1.4)
Fire station
(w: 1.4)
Flooded road
(w: 0.6)
Fountain
(w: 1.0)
Gas station
(w: 1.4)
Golf course
(w: 1.0)
Ground transp.
station (w: 1.0)
Helipad
(w: 1.0)
Hospital
(w: 1.0)
Impoverished settlement (w: 1.0)
Interchange
(w: 1.0)
Lake or pond
(w: 1.0)
Lighthouse
(w: 1.0)
Military facility
(w: 0.6)
Multi-unit residen- Nuclear powerplant
tial (w: 1.0)
(w: 0.6)
Office building
(w: 1.0)
Oil or gas facility
(w: 1.0)
Park
(w: 1.0)
Parking lot or
garage (w: 1.0)
Place of worship
(w: 1.0)
Police station
(w: 1.4)
Port
(w: 1.0)
Prison
(w: 1.0)
Race track
(w: 1.0)
Railway bridge
(w: 1.0)
Recreational
facility (w: 1.0)
Road bridge
(w: 1.4)
Runway
(w: 1.0)
Shipyard
(w: 1.0)
Shopping mall
(w: 1.0)
Single-unit residential (w: 0.6)
Smokestack
(w: 1.4)
Solar farm
(w: 0.6)
Space facility
(w: 1.0)
Stadium
(w: 1.0)
Storage tank
(w: 1.0)
Surface mine
(w: 1.0)
Swimming pool
(w: 1.0)
Toll booth
(w: 1.0)
Tower
(w: 1.4)
Tunnel opening
(w: 0.6)
Waste disposal
(w: 1.0)
Water treatment
facility (w: 1.0)
Wind farm
(w: 0.6)
Zoo
(w: 1.0)
False detection
(w: 0.0)
Fig. 3. List of classes of the FMOW challenge. An image example, a label and an associated weight are presented for each class. The least
challenging classes have a lower weight (0.6) while the most difficult ones have a higher weight (1.4). These weights impact the final classification
score.
4
Fig. 4. The class histogram distribution for the training (gray bars) and evaluation sets (black bars) by using the 3-band pansharpened
JPEG/RGB set. Note the highly unbalanced class distribution.
class 2
class m
class 1
C=
Ground-truth label
Let P = he1 , e2 , . . . , en i be the predicted labels provided
by a competitor and G = hg1 , g2 , . . . , gn i the ground-truth
labels for n distinct test regions. From P and G, one can
derive a confusion matrix C :
Predicted label
···
class 1
C1,1
C1,2
...
C1,m
class 2
..
.
C2,1
..
.
C2,2
..
.
...
..
.
C2,m
..
.
class m-1
Cm−1,1
Cm−1,2
...
Cm−1,m
class m
Cm,1
Cm,2
...
Cm,m
where Ci,i is the number of correct classifications for the ith class and Ci,j the number of images from the i-th class
that were mistakenly classified as the j -th class. With these
values we compute the precision Pi and recall Ri scores for
the i-th class as follows:
Pi =
tpi
tpi + f pi
Ri =
tpi
tpi + f ni
Pm
Fi × w i
Pm
F̄ = i=1
i=1 wi
where wi is the weight of i-th class. The weights for each
class are provided in Figure 3. It is worth noting that the
weight for false detections is 0, but it does not mean it will
not affect F̄ . If a region is misclassified as a false detection,
its class will have an additional false negative. If a false
detection is misclassified as another class, one more false
positive will be taken into account.
2.3 Hardware and time restrictions
The competition ran from September 21st to December 31st,
2017. The provisional standings were used to select the top
10 competitors as finalists. For the final round, competitors
were asked to provide a dockerized version3 of their solutions, and the final standings were determined based on
the results for a separate sequestered dataset (i.e. none of
the competitors had access to these images). The dockerized
solution had to cope with the following limitations:
•
•
where:
tpi = Ci,i
f pi =
m
X
Cj,i
j=1,j6=i
f ni =
m
X
•
Ci,j
All models required by the solution should be generated during training using raw data only (i.e. preprocessed data was not allowed) within 7 days.
The prediction for all test images should be performed in no more than 24 hours. Prebuilt models
should be provided to enable testing before training.
Hardware was restricted to two types of Linux AWS
instances: m4.10xlarge and g3.16xlarge.
j=1,j6=i
are respectively the number of true positives, false positives
and false negatives for the i-th class. The F -measure for each
class is then computed as the harmonic mean of precision
and recall:
Fi =
(1)
2 × Pi × R i
Pi + R i
The final score F̄ is the weighted sum of the F -measure
for all classes:
Instances of the type m4.10xlarge have 40 Intel Xeon E52686 v4 (Broadwell) processors and 160GB of memory and
were recommended for CPU-based systems. Instances of the
type g3.16xlarge have 4 NVIDIA Tesla M60 GPUs, 64 similar
processors and 488GB of memory and were recommended
for GPU-based systems. However, as can be noticed, a
g3.16xlarge instance was best suited for both CPU-based and
GPU-based systems.
3. https://www.docker.com/
5
2.4
Baseline
4
The organizers released a baseline classification code in
October 14th, 2017, and updated that solution in November
17th of the same year. The most recent and accurate version
of the baseline consists in extracting discriminant features
from images and metadata using deep neural networks.
To this end, they concatenated image features obtained
by a CNN (DenseNet-161 [7]) to a preprocessed vector of
metadata (mean subtraction) and fed this to a multilayer
perceptron (MLP) with 2 hidden layers. A softmax loss is
then used by the Adam optimizer [8] during training. They
used both training and validation subsets for training, and
no augmentation was performed. More details can be found
in the work of Christie et al. [1]. The classifiers employed in
our Hydra framework are variations of this baseline code.
3
on I MAGE N ET [15] as a starting point. Then, the obtained
weights serve as starting point for several copies of these
CNNs, which are further optimized to form the heads of the
Hydra. During testing, an image is fed to all Hydra’s heads
and their outputs are fused to generate a single decision, as
illustrated in Figure 6.
FMoW dataset
...
T HE H YDRA F RAMEWORK
Ensembles of learning algorithms have been effectively used
to improve the classification performance in many computer vision problems [9], [10], [11]. The reasons for that,
as pointed out by Dietterich [12] are: the training phase
could not provide sufficient data to build a single best
classifier; the optimization algorithm fails to converge to the
global minimum, but an ensemble using distinct starting
points could better approximate the optimal result; or the
space being searched may not contain the optimal global
minimum, but an ensemble may expand this space for a
better approximation.
Formally,
a
supervised
machine
learning
algorithm receives a list of training samples
{(x1 , y1 ), (x2 , y2 ), . . . , (xm , ym )} drawn from some
unknown function y = f (x) — where x is the feature
vector and y the class — and searches for a function h
through a space of functions H, also known as hypotheses,
that best matches y . As described by Dietterich [12],
an ensemble of learning algorithms build a set of
hypotheses {h1 , h2 , . . . , hk } ∈ H with respective weights
{w1 , w2 , . . . , wk }, and then provides a classification decision
ȳ through hypotheses fusion. Different fusion approaches
Pk
may be exploited, such as weighted sum ȳ = i=1 wi hi (x),
majority voting, and so on.
Hydra is a framework to create ensembles of CNNs. As
pointed out by Sharkey [13], a neural network ensemble
can be build by: varying the initial weights; varying the
network architecture; and varying the training set. Varying
the initial weights, however, requires a full optimization for
each starting point and thus consumes too much computational resources. For the same reason, varying the network
architecture demands caution when it is not possible to
share learned information between different architectures.
Unlike previous possibilities, varying the training set can be
done in any point of the optimization process for one or
multiple network architectures, and, depending on how its
done, it may not impact the final training cost.
Our framework uses two state-of-the-art CNN architectures, ResNet [14] and DenseNet [7], as shown in Figures 5 and 6. The training process depicted in Figure 5 consists in first creating the body of the Hydra by coarsely optimizing the chosen architectures using weights pretrained
4. https://github.com/fmow/baseline
ResNet1
ResNet
FMoW
Weights
ResNet2
ResNet3
...
ImageNet
Weights
ResNetm
DenseNet1
DenseNet
FMoW
Weights
DenseNet2
DenseNet3
...
Hydra’s body
DenseNetn
Hydra’s heads
Fig. 5. Hydra’s training flowchart: two CNNs, ResNet and DenseNet,
are optimized starting from the ImageNet weights to create the Hydra’s
body. The obtained weights are then used as starting point for several
copies of these CNNs, which are further optimized to form the heads of
the Hydra.
ResNet1
O UTPUT
Shipyard
ResNet2
I NPUT I MAGE
ResNet3
...
ResNetm
Fusion
DenseNet1
...
DenseNet2
DenseNet3
...
63 classes
DenseNetn
Fig. 6. Hydra’s inference flowchart: an input image is fed to all heads of
the Hydra, and each head outputs the probability of each class being the
correct one. A majority voting is then used to determine the final label,
considering that each head votes for its most probable label.
6
A key point on ensemble learning, as observed by
Krogh et al. [16] and Xandra et al. [17], is that the hypotheses should be as accurate and as diverse as possible. In Hydra, such properties were prompted by applying different geometric transformations over the training
samples, like zoom, shift, reflection and rotations. More
details regarding the proposed framework are presented in
Sections 3.1, 3.2 and 3.3.
3.1
Input
ResNet
or
DenseNet
I MAGE
CNN
x` = F` (x`−1 ) + x`−1
(2)
where x`−1 and x` are the input and output of layer `, and
x0 is the input image. ResNets are built by stacking several
of these units, creating bypasses for gradients that improve
the back-propagation optimization in very deep networks.
As a consequence, ResNets may end up generating redundant layers. As a different solution to improve the flow of
gradients in deep networks, Huang et al. [7] proposed to
connect each layer to all preceding layers, which was named
a DenseNet unit:
x` = F` ([x0 , x1 , . . . , x`−1 ])
(3)
where [x0 , x1 , . . . , x`−1 ] is the concatenation of the ` previous layer outputs. DenseNets also stack several units,
but they use transition layers to limit the number of connections. In the end, DenseNets have less parameters than
ResNets and compensate this facts through feature reuse.
This difference in behavior may be one of the causes for
their complementarity.
Both networks are initialized with weights from ImageNet [15] and trained for six epochs using the Adam optimization algorithm [8] with a learning rate of 10−4 . Fully
connected layers with ImageNet weights are discarded, and
three new layers with 4096 neurons are trained from scratch
using a 50% dropout rate, as illustrated in Figure 7. Any
given image metadata that is useful for classification (e.g.
ground sample distance, sun angle) is concatenated to the
input of the first fully connected layer.
The obtained CNN weights from the previous step will
then be used as starting point to multiple copies of the
chosen networks, which will be trained for five more epochs
using progressive drops in the learning rate as follows: a
learning rate of 10−4 is used for the first epoch, 10−5 for the
following three epochs, and 10−6 for the last epoch.
During training, we added different class weights to
the loss function for different heads seeking to enhance
the complementarity between them. Four different weighting schemes were used: unweighted classes, FMOW class
weights as presented in Figure 3 (false detection weight was
set to 1), a frequency-based weighting using the balanced
heuristic [18], and a manual adjustment of the latter.
Metadata
FC
FC
Softmax
...
...
...
...
...
CNN architectures
Both CNN architectures used in this work, ResNet (ResNet50) [14] and DenseNet (DenseNet-161) [7], were chosen for
achieving state-of-the-art results in different classification
problems while being complementary to each other. Let
F` (·) a non-linear transformation — a composite function
of convolution, relu, pooling, batch normalization and so on
— for layer `. As described by He et al. [14], a ResNet unit
can be defined as:
FC
...
Fig. 7. The convolutional neural network architecture used by the Hydra
framework.
3.2
On-line data augmentation
The process of artificially augmenting the image samples in
machine learning by using class-preserving transformations
is meant to reduce the overfitting in imbalanced classification problems or to improve the generalization ability in the
absence of enough data [19], [20], [21], [22], [23]. The breakthroughs achieved by many state-of-the-art algorithms that
benefit from data augmentation has lead many authors [24],
[25], [26], [27] to study the effectiveness of this technique.
Different techniques can be considered, such as geometric (e.g. reflection, scaling, warping, rotation and translation)
and photometric (e.g. brightness and saturation enhancements, noise perturbation, edge enhancement and color
palette changing) transformations. However, as pointed out
by Ratner et al. [28], selecting transformations and tuning
their parameters can be a tricky and time-consuming task.
A careless augmentation may produce unrealistic or meaningless images that will negatively affect the performance.
The process of data augmentation can be conducted offline, as illustrated in Figure 8. In this case, training images
undergo geometric and/or photometric transformations before the training. Although this approach avoids repeating
the same transformation several times, it also considerably
increases the size of the training set. For instance, if we
only consider basic rotations (90, 180 and 270 degrees) and
reflections, the FMOW training set will have more than 2.8
million samples. This would not only increase the amount of
required storage space, but the time consumed by a training
epoch as well.
In our framework we apply an on-line augmentation,
as illustrated in Figure 9. To this end, different random
transformations were applied to the training samples every
epoch. This way, the number of transformations seen during
training increase with the number of epochs. In addition,
different heads of the Hydra experience different versions
of the training set and hardly converge to the same local
minima.
On-line data augmentations were carried out by the
ImageDataGenerator function from Keras5 . We used random flips in vertical and horizontal directions, random
zooming and random shifts over different image crops. We
did not use any photometric distortion to avoid creating
5. https://keras.io/
7
Training samples (n regions)
Training samples (n regions)
...
...
CNN-Training
Off-line data augmentation
Randomly transformed samples for the 1st epoch (n regions)
...
CNN-Training
Image samples for all epochs (8n regions)
CNN-Training
...
Fig. 8. Illustration of an off-line data augmentation process. In this
example considering basic rotations (90, 180 and 270 degrees) and
reflections only, each image is augmented to a set of 8 images before
the training starts. After that, the augmented training set is static for all
CNN epochs.
meaningless training samples. Three image crop styles were
considered, as illustrated in Figure 10: the original bounding
box that was provided as a metadata (Figure 10(a)), and
an expanded bounding box extracted from both JPG pansharpened images (Figure 10(b)) and JPG multi-spectral
images (Figure 10(c)). As multi-spectral images have low
resolution, we discarded samples with less than 96 pixels
of width or height. Table 2 presents the configuration of
each head in terms of network architecture, augmentation
technique and class weighting.
Randomly transformed samples (n regions)
for the 2nd, 3rd and following epochs
Fig. 9. Illustration of an on-line data augmentation process. In this
example considering basic rotations (90, 180 and 270 degrees) and
reflections only, each image is randomly transformed into a new image
before each training epoch starts.
(a)
(c)
Fig. 10. Different image crops used for training: (a) original bounding
box and (b) expanded bounding box extracted from JPG pan-sharpened
images, and (c) expanded bounding box extracted from JPG multispectral images.
4
3.3
(b)
E XPERIMENTAL RESULTS
Fusion and decision
To determine the most appropriate label for an input region
according to the models generated by Hydra, we have to
combine the results from each head. Each head produces
a vector of score values that indicate how well the input
region fits each class. As some regions may appear in
more than one image, their vectors are summed so that
each region ends up with only one vector of scores. The
softmax function is then applied to convert these scores to
probabilities, and the most probable class is selected for each
head. A majority voting is the used to select the final label.
However, if the number of votes is below or equal to half of
the head count, the region is considered a false detection.
4.1
FMOW dataset
In this section we report and compare the Hydra performance in the FMOW dataset. First, to visualize the benefits
of using the Hydra framework, we show the accuracy of
the trained CNNs (body and heads) over the epochs in
Figure 11. We also show the ensemble accuracy when the
heads are available. In this experiment, we use the entire
training subset and false detections of the evaluation subset
for training, and the remaining images of the evaluation
subset for accuracy computation. As can be observed, the
body of the Hydra provides a good starting point for its
8
TABLE 2
Hydra’s heads description. For each head are given the network
architecture, image crop style, augmentation technique and class
weighting method. Three different image crop styles were considered:
ORIG-PAN (Figure 10(a)), EXT-PAN (Figure 10(b)) and EXT-MULTI
(Figure 10(c)). Augmentation techniques included random flips in
vertical and horizontal directions (Flip), random zooming (Zoom) and
random shifts (Shift). Class weighting methods were: no weighting,
FMOW weights (Figure 3) and frequency-based weighting using the
balanced heuristic [18] before (Frequency #2) and after (Frequency #1)
a manual adjustment.
Head
#1
#2
#3
#4
#5
#6
#7
#8
#9
#10
#11
#12
CNN
DenseNet
DenseNet
DenseNet
DenseNet
DenseNet
DenseNet
DenseNet
DenseNet
ResNet
ResNet
ResNet
ResNet
Crop
EXT-PAN
ORIG-PAN
EXT-MULTI
EXT-PAN
EXT-PAN
EXT-MULTI
EXT-PAN
ORIG-PAN
EXT-PAN
EXT-MULTI
ORIG-PAN
EXT-MULTI
Augment
Flip
Flip
Flip
Zoom
Shift
Shift
Flip
Flip
Flip
Flip
Flip
Flip
Class weighting
Unweighted
Frequency #1
Frequency #1
Frequency #2
Unweighted
FMOW weights
Frequency #2
Frequency #2
Unweighted
Frequency #1
Frequency #1
Frequency #2
heads, which have slightly different accuracy in the end
of the training process. When the heads are combined, the
obtained accuracy is considerably higher than the accuracy
of the best performing head.
We then reran the training process using all images from
the training and evaluation subsets. The TopCoder submission system was used to obtain the F̄ -measure (Equation 1)
for the testing subset, as it has no ground truth available
so far. The obtained F̄ -measure is shown in Table 3 along
with results reported by Christie et al. [1] and results for
our best individual classifiers. As previously reported by
Christie et al. [1], using metadata for classification purposes
is better than just using the image information. In Table 3
we show that on-line data augmentation has also a big
impact, increasing the accuracy of their respective versions
without augmentation in more than 2%. Still, our ensemble
can improve the performance even more, with a gap of at
least 3% to all individual classifiers.
TABLE 3
F̄ -measure for the FMOW testing subset. Evaluated classifiers may
use metadata information (M ) and/or on-line data augmentation (O).
F̄ -measure values for classifiers marked with † were reported by
Christie et al. [1].
F̄ -measure
Hydra (DenseNet + ResNet)M,O
0.781
DenseNetM,O
0.749
ResNetM,O
0.737
†
LSTMM
0.734
†
DenseNetM
0.722
ResNetM
0.712
LSTM
†
0.688
DenseNet†
Fig. 11. Evaluation of the ensemble accuracy. The first six 6 epochs
show the accuracy of the Hydra’s body using DenseNet and Resnet.
The following epochs show the accuracy for each head separately as
well as for the fusion of all heads (ensemble).
The confusion matrix for the best ensemble result in
Figure 11 (epoch 11) is presented on Figure 12. The hardest
classes for Hydra were: shipyard (mainly confused with
port), hospital (confused with educational institution),
multi unit residential (confused with single unit residential), police station (confused with educational institution),
and office building (confused with fire station, police station, etc). We also had high confusion between nuclear powerplant and port, between prison and educational institution, between space facility and border checkpoint, and
between stadium and recreational facility. These, however,
are natural mistakes that could easily be repeated by human
experts. Most of these classes look alike in aerial images (e.g.
shipyard and port) or have similar features that can trick
an automatic classifier (e.g. both stadium and recreational
facility have a sport court). In Figure 13 we show examples
of mislabeled regions that cannot be easily identified, even
by a human.
0.679
With this performance, we managed to finish the competition among the top ten competitors in the provisional
scoreboard (5th place). We then submitted a dockerized
version of Hydra, which was evaluated by the FMOW
organizers using a sequestered dataset. As we show in
Table 4, the Hydra ensemble was very stable, with nearly
the same performance in the testing subset and sequestered
dataset. Thanks to this, our final ranking was third place,
with accuracy 1% lower than the best result achieved by
another competitor.
TABLE 4
Final ranking: F̄ -measure for the top five participants of the FMOW
challenge using the available testing subset (Open) and the
sequestered dataset (Sequestered).
Rank
Handle
1
pfr
0.7934
0.7917
2
jianminsun
0.7940
0.7886
usf bulls (Hydra) 0.7811
0.7817
3
Open Sequestered
4
zaq1xsw2tktk
0.7814
0.7739
5
prittm
0.7681
0.7670
9
Fig. 12. Confusion matrix for the Hydra ensemble after 11 training epochs (Figure 11). Each row represents the ground truth label, while the
column shows the label obtained by the Hydra. Large values outside of the main diagonal indicate that the corresponding classes are hard to be
discriminated.
4.2
NWPU-RESISC45 dataset
We also reported the Hydra performance in the NWPURESISC45 dataset, which was recently compiled by
Cheng et al. [4]. This dataset has 45 classes, with 700 images
per class, each one with resolution of 256 × 256 pixels,
totaling 31,500 images. The intersection of classes between
FMOW and NWPU-RESISC45 is reasonable, with some
land uses being overcategorized in NWPU-RESISC45 (e.g.
basketball court, baseball diamond and tennis court are
all labeled as recreational facility in the FMOW dataset) and
some classes not representing land use, such as objects (e.g.
airplane and ship) and vegetation (e.g. desert, forest and
wetland). This dataset does not provide satellite metadata.
We followed the experimental setup proposed by
Cheng et al. [4] and randomly chose 10% or 20% of the
samples in each class for training. The remaining samples
were used for testing. We repeated this experiment five
times for each configuration, and reported the average accuracy and standard deviation. As can be shown in Table 5,
Cheng et al. [4] compared six classifiers based on handcrafted features and three different CNN architectures. In
their experiments, the best performance was achieved by
a VGG16 network [10] that was trained using ImageNet
weights as a starting point.
The configuration of the Hydra was slightly different
in this experiment. We used 4 DenseNets and 4 ResNets,
which were trained for 16 epochs (8 for the body and 8 for
the heads). Each head for each architecture has a different
augmentation: 1) no augmentation; 2) random vertical and
horizontal flips; 3) random zoomimg; and 4) random shifts.
The accuracy results for the NWPU-RESISC45 dataset are
shown in Table 5. As NWPU-RESISC45 is less challenging
than FMOW, the results are higher and the improvements
are lower. The on-line data augmentation increased ResNet
and DenseNet accuracies in about 1.0% and 1.6%, respectively. Individual classifiers for both architectures outperformed all previous results reported by Cheng et al. [4], and
the Hydra ensemble was able to achieve the highest accuracy. The improvement on accuracy, however, was slightly
greater than 1% when compared to individual CNNs. It is
worth noting that we outperformed the best accuracy results
reported by Cheng et al. [4] by more than 4%.
10
5
(a) Shipyard mislabeled as port
(b) Hospital mislabeled as educational institution.
(c) Multi-unit residential misla- (d) Office building mislabeled as
beled as single-unit residential
gas station
Fig. 13. Regions from the evaluation subset that were mislabeled by our
Hydra ensemble.
TABLE 5
Accuracy for the NWPU-RESISC45 dataset using training/testing splits
of 10%/90% and 20%/80%. Average accuracy and standard deviation
were reported by using five different random splits. Evaluated
classifiers may use on-line data augmentation (O) and/or ImageNet
weights (I ). Accuracy values for classifiers marked with † were
reported by Cheng et al. [4].
Hydra (DenseNet + ResNet)O,I
DenseNetO,I
ResNetO,I
VGG16I †
AlexNetI †
GoogleNetI †
AlexNet†
VGG16†
GoogleNet†
BoVW†
LLC†
BoVW + SPM†
Color histograms†
LBP†
GIST†
Accuracy
10%/90% 20%/80%
92.44 ± 0.34 94.51 ± 0.21
91.06 ± 0.61 93.33 ± 0.55
89.24 ± 0.75 91.96 ± 0.71
87.15 ± 0.45 90.36 ± 0.18
81.22 ± 0.19 85.16 ± 0.18
82.57 ± 0.12 86.02 ± 0.18
76.69 ± 0.21 79.85 ± 0.13
76.47 ± 0.18 79.79 ± 0.15
76.19 ± 0.38 78.48 ± 0.26
41.72 ± 0.21 44.97 ± 0.28
38.81 ± 0.23 40.03 ± 0.34
27.83 ± 0.61 32.96 ± 0.47
24.84 ± 0.22 27.52 ± 0.14
19.20 ± 0.41 21.74 ± 0.18
15.90 ± 0.23 17.88 ± 0.22
D ISCUSSION
There is an extensive literature on algorithms for geospatial
land image processing — often called remote sensing image
processing. This broad and active field of research has many
branches, such as semantic segmentation [29], [30], [31],
[32], [33], target location [34] and region classification [35],
[36], [37], [38], [39]. An exhaustive literature review was
carried out by Cheng and Han [40] and covered most of the
literature works based on handcrafted features. However, as
can be seen in Table 5, handcrafted approaches are no longer
competitive due to the recent advances on deep learning.
The FMOW challenge is focused on the classification
branch and currently constitutes the most difficult task of
its kind (63 classes, 470086 images). Other smaller datasets,
such as UC Merced Land Use Dataset [2] (UCMerced, 21
classes, 2100 images), WHU-RS19 [3] (19 classes, 950 image),
Brazilian Coffe Scenes [41] (BCS, 2 class, 2876 images)
and Aerial Image Dataset et al. [42] (AID, 30 classes, 10000
images), were used in the literature but not necessarily
represent a challenging task (i.e. reported results are reaching a nearly perfect score). Even NWPU-RESISC45 [4] (45
classes, 31500 images), the most recent dataset other than
FMOW, is not as challenging. We were able to achieve nearly
95% accuracy using only 20% of its images for training. In
comparison, the training/testing split for FMOW is about
85%/15%, but the top performance achieved by a competitor was still below 80%.
Anyway, there is no doubt that deep CNNs are becoming more and more popular for land use classification.
Nogueira et al. [38] compared different state-of-the-art CNN
architectures, such as AlexNet [19], GoogleNet [22] and
VGG [10], and evaluated different training setups, such as
using ImageNet weights as starting point and Support Vector Machines (SVM) for classification. They reported results
for UCMerced, WHU-RS19 and BCS datasets, which were
very successful when compared to many other approaches
based on handcrafted features. Chaib et al. [39] improved
regular VGG networks [10] by fusing the output of the last
two fully connected layers. The work was validated using
UCMerced, WHU-RS19 and AID. We also have CNN-based
results reported by Christie et al. [1] and Cheng et al. [4]
for NWPU-RESISC45 and FMOW datasets, respectively,
as shown in Tables 3 and 5. Thus, CNN classifiers were
definitely the best choice for our Hydra framework. After
experimenting different architectures, we ended up using
DenseNets and ResNets. The main reasons are that they
perform better than previously mentioned architectures,
have a reasonable level of complementarity, and achieve
similar classification results.
In terms of computational resources consumption, Hydra has some disadvantages when compared to individual
classifiers. If N heads are used, its training and inference
costs tend to be N/2 and N times more expensive, respectively. On the other hand, when compared to other ensemble
techniques with the same number of classifiers, Hydra’s
training is approximately two times faster because half of
the epochs are carried out for the Hydra’s body only.
Overall, the Hydra framework was a good fit to the
FMOW restrictions described in Section 2.3. The time limit
for testing was large enough to allow using many classi-
11
fiers, making the training time to generate them the main
problem. Hydra succeeded in increasing the number of classifiers for inference without losing complementarity, a key
requirement for ensembles. As a consequence, we finished
the competition in the third place, with accuracy 1% lower
than the best result achieved by another competitor.
6
C ONCLUSIONS
Automatic and robust classification of aerial scenes is
critical for decision-making by governments and intelligence/military agencies in order to support fast disaster
response. This research problem, although extensively addressed by many authors, is far from being solved, as
evidenced by the final results of the FMOW challenge.
The main contribution of this work was the development
of a framework that automatically creates ensembles of
CNNs to perform land use classification in satellite images. We called this framework Hydra due to its training
procedure: the Hydra’s body is a CNN that is coarsely
optimized and then fine tuned multiple times to form the
Hydra’s heads. The resulting ensemble achieved the fifth
best performance for the open testing subset of the FMOW
challenge, qualifying our solution to the second phase of
the competition. Then, the same ensemble was applied to
a sequestered testing set that was used to officially rank
the competitors, and it obtained the third best performance.
Given that there were more than 50 active participants, including teams and individuals from academia and industry,
and most of them probably were experimenting different
CNN-based setups, we believe that Hydra is well aligned
with the most recent state-of-the-art developments.
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
ACKNOWLEDGMENTS
Part of the equipments used in this project are supported by
a grant (CNS-1513126) from the National Science Foundation (NSF) of USA. The Titan Xp used for this research was
donated by the NVIDIA Corporation.
[23]
[24]
R EFERENCES
[25]
[1]
[26]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
G. Christie, N. Fendley, J. Wilson, and R. Mukherjee, “Functional
map of the world,” arXiv preprint arXiv:1711.07846, 2017.
Y. Yang and S. Newsam, “Bag-of-visual-words and spatial extensions for land-use classification,” in Proceedings of the 18th SIGSPATIAL International Conference on Advances in Geographic Information
Systems, ser. GIS ’10. ACM, 2010, pp. 270–279.
D. Dai and W. Yang, “Satellite image classification via two-layer
sparse coding with biased image representation,” IEEE Geoscience
and Remote Sensing Letters, vol. 8, no. 1, pp. 173–176, 2011.
G. Cheng, J. Han, and X. Lu, “Remote sensing image scene
classification: Benchmark and state of the art,” Proceedings of the
IEEE, vol. 105, no. 10, pp. 1865–1883, Oct 2017.
H. Li, L. Jing, Y. Tang, Q. Liu, H. Ding, Z. Sun, and Y. Chen,
“Assessment of pan-sharpening methods applied to worldview-2
image fusion,” in IEEE International Geoscience and Remote Sensing
Symposium (IGARSS), 2015, pp. 3302–3305.
D. M. W. Powers, “Evaluation: From precision, recall and fmeasure to roc., informedness, markedness & correlation,” Journal
of Machine Learning Technologies, vol. 2, no. 1, pp. 37–63, 2011.
G. Huang, Z. Liu, L. van der Maaten, and K. Q. Weinberger,
“Densely connected convolutional networks,” in IEEE Conference
on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 4700–
4708.
D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” CoRR, vol. abs/1412.6980, 2014.
[27]
[28]
[29]
[30]
[31]
[32]
H. Li, Z. Lin, X. Shen, J. Brandt, and G. Hua, “A convolutional
neural network cascade for face detection,” in IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), 2015, pp. 5325–
5334.
K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” pp. 1–14, 2015.
C. Ding and D. Tao, “Trunk-branch ensemble convolutional neural
networks for video-based face recognition,” IEEE Transactions on
Pattern Analysis and Machine Intelligence (TPAMI), vol. PP, no. 99,
pp. 1–11, 2017.
T. G. Dietterich, Ensemble Methods in Machine Learning. Berlin,
Heidelberg: Springer Berlin Heidelberg, 2000, pp. 1–15.
A. J. Sharkey, Ed., Combining Artificial Neural Nets: Ensemble and
Modular Multi-Net Systems, 1st ed. Secaucus, NJ, USA: SpringerVerlag New York, Inc., 1999.
K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning
for image recognition,” in IEEE Conference on Computer Vision and
Pattern Recognition (CVPR), 2016.
J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei,
“ImageNet: A Large-Scale Hierarchical Image Database,” in IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), 2009.
A. Krogh and J. Vedelsby, “Neural network ensembles, cross validation and active learning,” in International Conference on Neural
Information Processing Systems (NIPS), 1994, pp. 231–238.
A. Chandra and X. Yao, “Evolving hybrid ensembles of learning
machines for better generalisation,” Neurocomputing, vol. 69, no. 7,
pp. 686–700, 2006.
G. King and L. Zeng, “Logistic regression in rare events data,”
Political Analysis, vol. 9, pp. 137–163, 2001.
A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” Commun. ACM,
vol. 60, no. 6, pp. 84–90, May 2017.
A. G. Howard, “Some improvements on deep convolutional neural network based image classification,” CoRR, vol. abs/1312.5402,
2013. [Online]. Available: http://arxiv.org/abs/1312.5402
K. He, X. Zhang, S. Ren, and J. Sun, “Spatial pyramid pooling
in deep convolutional networks for visual recognition,” IEEE
Transactions on Pattern Analysis and Machine Intelligence (TPAMI),
vol. 37, no. 9, pp. 1904–1916, 2015.
C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov,
D. Erhan, V. Vanhoucke, and A. Rabinovich, “Going deeper with
convolutions,” in IEEE Computer Vision and Pattern Recognition
(CVPR), 2015, pp. 1–9.
E. Ahmed, M. Jones, and T. K. Marks, “An improved deep learning
architecture for person re-identification,” in IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), 2015, pp. 3908–
3916.
K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman, “Return
of the devil in the details: Delving deep into convolutional nets.”
in British Machine Vision Conference (BMVC). BMVA Press, 2014.
T. DeVries and G. W. Taylor, “Dataset Augmentation in Feature
Space,” ArXiv e-prints, 2017.
L. Perez and J. Wang, “The Effectiveness of Data Augmentation
in Image Classification using Deep Learning,” ArXiv e-prints, Dec.
2017.
S. C. Wong, A. Gatt, V. Stamatescu, and M. D. McDonnell, “Understanding data augmentation for classification: When to warp?” in
International Conference on Digital Image Computing: Techniques and
Applications, 2016, pp. 1–6.
A. J. Ratner, H. R. Ehrenberg, J. D. Zeshan Hussain, and C. Re,
“Learning to compose domain-specific transformations for data
augmentation,” in International Conference on Neural Information
Processing Systems (NIPS), 2017, pp. 1–11.
Q. Shi, B. Du, and L. Zhang, “Spatial coherence-based batchmode active learning for remote sensing image classification,”
IEEE Transactions on Image Processing (TIP), vol. 24, no. 7, pp. 2037–
2050, 2015.
H. Lee and H. Kwon, “Going deeper with contextual cnn for
hyperspectral image classification,” IEEE Transactions on Image
Processing (TIP), vol. 26, no. 10, pp. 4843–4855, 2017.
H. Wu and S. Prasad, “Semi-supervised deep learning using
pseudo labels for hyperspectral image classification,” IEEE Transactions on Image Processing (TIP), vol. 27, no. 3, pp. 1259–1270, 2018.
J. Zhao, Y. Zhong, H. Shu, and L. Zhang, “High-resolution image
classification integrating spectral-spatial-location cues by conditional random fields,” IEEE Transactions on Image Processing (TIP),
vol. 25, no. 9, pp. 4033–4045, 2016.
12
[33] X. Xu, W. Li, Q. Ran, Q. Du, L. Gao, and B. Zhang, “Multisource
remote sensing data classification based on convolutional neural
network,” IEEE Transactions on Geoscience and Remote Sensing,
vol. PP, no. 99, pp. 1–13, 2017.
[34] Z. Zou and Z. Shi, “Random access memories: A new paradigm for
target detection in high resolution aerial remote sensing images,”
IEEE Transactions on Image Processing (TIP), vol. 27, no. 3, pp. 1100–
1111, 2018.
[35] Y. Xia, L. Zhang, Z. Liu, L. Nie, and X. Li, “Weakly supervised
multimodal kernel for categorizing aerial photographs,” IEEE
Transactions on Image Processing (TIP), vol. 26, no. 8, pp. 3748–3758,
2017.
[36] E. Li, J. Xia, P. Du, C. Lin, and A. Samat, “Integrating multilayer
features of convolutional neural networks for remote sensing
scene classification,” IEEE Transactions on Geoscience and Remote
Sensing, vol. 55, no. 10, pp. 5653–5665, 2017.
[37] X. Lu, X. Zheng, and Y. Yuan, “Remote sensing scene classification
by unsupervised representation learning,” IEEE Transactions on
Geoscience and Remote Sensing, vol. 55, no. 9, pp. 5148–5157, 2017.
[38] K. Nogueira, O. A. Penatti, and J. A. dos Santos, “Towards better
exploiting convolutional neural networks for remote sensing scene
classification,” Pattern Recognition - Elsevier, vol. 61, pp. 539–556,
2017.
[39] S. Chaib, H. Liu, Y. Gu, and H. Yao, “Deep feature fusion for
vhr remote sensing scene classification,” IEEE Transactions on
Geoscience and Remote Sensing, vol. 55, no. 8, pp. 4775–4784, 2017.
[40] G. Cheng and J. Han, “A survey on object detection in optical
remote sensing images,” ISPRS Journal of Photogrammetry and
Remote Sensing - Elsevier, vol. 117, pp. 11–28, 2016.
[41] O. A. B. Penatti, K. Nogueira, and J. A. dos Santos, “Do deep
features generalize from everyday objects to remote sensing and
aerial scenes domains?” in IEEE Conference on Computer Vision and
Pattern Recognition Workshops (CVPRW), 2015, pp. 44–51.
[42] G. S. Xia, J. Hu, F. Hu, B. Shi, X. Bai, Y. Zhong, L. Zhang, and
X. Lu, “Aid: A benchmark data set for performance evaluation
of aerial scene classification,” IEEE Transactions on Geoscience and
Remote Sensing, vol. 55, no. 7, pp. 3965–3981, 2017.
| 1cs.CV
|
arXiv:1405.2322v2 [math.ST] 4 Oct 2016
Parameter estimation of a two-colored urn model
class
Line Le Goff
Philippe Soulier
January 29, 2018
Abstract
Though widely used in applications, reinforced random walk on graphs
have never been the subject of a valid statistical inference. We develop in
this paper a statistical framework for a general two-colored urn model. The
probability to draw a ball at each step depends on the number of balls of
each color and on a multidimensional parameter θ through a function f ,
called a choice function. We introduce two estimators of θ: the maximum
likelihood estimator and a weighted least squares estimator which is less
efficient, but is closer to the calibration techniques used in the applied literature. In general, the model is an inhomogeneous Markov chain and
this property makes the estimation of the parameter impossible on a single path, even if it were infinite. Therefore we assume that we observe
i.i.d. experiments, each of a predetermined finite length. This is coherent
with the usual experimental set-ups. We apply the statistical framework to
a real life experiment: the selection of a path among pre-existing channels
by an ant colony. We performed experiments, which consisted of letting
ants pass through the branches of a fork. We consider the particular urn
model proposed by J.-L. Deneubourg in 1990 to describe this phenomenon.
We simulate this model for several parameter values in order to assess the
accuracy of the MLE and the WLSE. Then we estimate the parameter
from the experimental data and evaluate confident regions with Bootstrap
algorithms. The findings of this paper do not contradict the biological literature, but give statistical significance to the values of the parameter found
therein.
1
1
Introduction
Urn models have been studied for nearly one century. In 1931, G. Pólya provided
the first probabilistic result on the game consisting in drawing a ball from an urn
initially containing one red ball and one black ball (see Pólya, 1931). At each time
step, a ball is drawn and put back in the urn with an additional ball of the same
color. The probability to draw a red ball is the proportion of red balls in the urn.
G. Pólya proved that, as the number of draws tends to infinity, the proportion of
red balls tends to a random variable following the uniform distribution on [0, 1].
The Pólya urn is easily generalizable to a large class of two-colored urn models
characterized by a choice function f : Θ × N × N → [0, 1] which itself depends on
a parameter θ ∈ Θ ⊂ Rd , d ≤ 1. Let Rn and Bn be the numbers of red and black
balls in the urn after n draws. Note that, by construction, Rn + Bn = n. The
probability that the (n + 1)-th ball is red is given by
probR
n+1 = f (θ, Rn , Bn ) = f (θ, Rn , n − Rn ) .
(1)
Consequently, the probability to draw a black ball at time n + 1 is 1 − probR
n+1 .
The goal of this paper is to propose a valid statistical methodology to estimate
the parameter θ. We suppose that we observe N independent paths, each consisting of a sequence of n colors drawn by the model defined in (1). The statistical
theory is developed as n is fixed and N tends to infinity. We choose this framework since in some models, it is not possible to obtain consistent estimators of the
parameter with only one path, even if its length n increases to infinity. Moreover
data from real experiments comprises a set of finite paths. We define estimators
for θ that we prove to be consistent and asymptotically normal under some usual
regularity assumptions on the model (1). We study more precisely two particular
cases: the maximum likelihood estimator (MLE) and the weighted least squares
estimators (WLSE).
We have applied these statistical tools to the problem of path formation by
an ant colony. One of the fundamental factors affecting an organism’s survival is
its ability to optimally and dynamically exploit its environment. For example, in
order to take advantage of the best sites of resources, housing or reproduction,
these areas must be discovered and exploited at the earliest opportunity. Many
species of ants rise to this challenge by developing a network of paths, which
connects different strategic sites such as nests and food sources. These paths
consist of pheromones, attractive chemical substances. We focus on a specific
aspect of this phenomenon: the selection of a path among pre-existing channels.
When exploring their environment, ants often face bifurcations and must bypass
obstacles. As shown by experimental studies, the laying of pheromones by ants
passing successively through a bifurcation results in two possible outcomes: either
one branch is eventually selected and the other abandoned, or both branches end
2
up being uniformly chosen (see Deneubourg, Aron, Goss, and Pasteels, 1990).
The analysis of the spontaneous path formation by a colony of ants is made
difficult by the absence of any means to measure precisely the quantity of, or even
detect, the pheromones laid by the ants.
It is commonly assumed that when approaching a bifurcation, ants choose a
branch and lay a certain constant amount of pheromone without ever turning back.
Consequently, the quantity of pheromone laid on a branch is proportional to the
number of ants which passed through it. Thus this phenomenon can be described
by a urn model as proposed by J.-L. Deneubourg et al. in 1990 (see Deneubourg
et al., 1990). More precisely they define a choice function with a two-dimension
parameter (α, c) ∈ (0, 1)2 such that the probability for an ant to choose the right
branch after Rn passages through the right branch and n passages in total is given
by
(c + Rn )α
.
(2)
P robR
n+1 =
(c + Rn )α + (c + n − Rn )α
The probability to choose the left branch is consequently 1 − P robR
n+1 . The parameter α makes this model non linear with respect to the proportion of passages
through one branch. It models the sensitivity of the ant to the concentration of
pheromone. The parameter c is the intrinsic attractiveness of each branch and
can also be interpreted as the inverse of the attractiveness (or the strength) of the
pheromone deposit laid by each ant.
Several probabilistic studies provide the asymptotic behavior of Rn /n in terms
of α and c (see Pólya, 1931; Tarrès, 2011; Davis, 1990). The influence of the two
parameters is on different time scales, but they can contribute to the same effect
(selection of one branch or unifomization of the traffic on the two branches) or have
antagonistic effects. The model is thus characterized by four phases, according to
the values of α and c: slow or fast uniformization; slow or fast selection. The phase
most commonly considered in the literature is slow selection. This corresponds
in our model to α and c larger than 1: selection of one branch will eventually
happen, though slowly because of weak pheromone deposits (see Deneubourg
et al., 1990; Vittori, Talbot, Gautrais, Fourcassié, Araujo, and Theraulaz, 2006;
Garnier, Guérécheau, Combe, Fourcassié, and Theraulaz, 2009). However, the
model may account for other possibilities, such as fast uniformization, which occurs when α < 1 and c > 1. One purpose of this paper is to investigate more
thoroughly these possibilities which have been more or less overlooked in the previous literature.
Ethological studies have already provided values for the parameter (α, c), but
the methods used mainly consisted of calibration without control of the statistical
validity of these methods and results (see Deneubourg et al., 1990; Vittori et al.,
2006; Garnier et al., 2009; Thienen, Metzler, Choe, and Witte, 2014) and in particular these methods do not produce confidence regions. However this type of
3
information supplies interesting and important elements to the behavioral discussion. However this type of information supply interesting and important elements
to the behavioral discussion. In this paper, we define the MLE and the WLSE for
the ant behavior model (2). We assess the quality of these estimators in a simulation experiment. We then use them on experimental data provided by a real
life experiment performed to this purpose with ants. We also compute confidence
region by a Bootstrap algorithm.
The model (2) can be applied, mutatis mutandis, to many other fields. For
instance we can consider a fork with two branches as a neuron having two axons.
During each period of time, the length of one of the two axons increases. The
longer an axon is, the higher the probability that it will further grow. Thus
the dynamic of this biological system may also be modeled by (2) (see Khanin
and Khanin, 2001). Furthermore, using the notion of a choice function (see
Section 2.3), the model (1) can be adapted and applied to many situations where
a binary choice occurs, or to even more complex situations such as networks with
several nodes (see Jeanson, Ratnieks, and Deneubourg, 2003; Pemantle, 2007), for
instance network exploration by an ant colony (see Aron, Deneubourg, Goss, and
Pasteels, 1990; Beckers, Deneubourg, and Goss, 1993; Nicolis and Deneubourg,
1999; Dussutour, Deneubourg, and Fourcassié, 2005; Nicolis and Dussutour, 2008;
Thienen et al., 2014; Arganda, Nicolis, Perochain, Péchabadens, Latil, and Dussutour, 2014).
The paper is organized as follows. The model (1) and the statistical framework is rigorously defined in Section 2. Section 2.1 is focused on the MLE and
Section 2.2 on the WLSE. Under some usual regularity conditions, we prove that
both estimators are consistent and asymptotically normal. Moreover the MLE
is asymptotically efficient. The numerical implementation of the MLE may be
difficult and unstable (and lengthy), therefore, a WLSE is considered and theoretically studied. This estimator does not match the theoretical performances of
the MLE, but is easier to compute and is popular among practitioners. Section 2.3
proposes an extension of the general urn model to a class of vertex reinforced random walk on graphs. Section 3 is an adaptation of the statistical framework to
the ethological problem. In Section 3.1, we first introduce the assumptions on
the ant behavior and then we define the model proposed by J.-L. Deneubourg.
The four phases of the model are described more precisely and are interpreted
from the point of view of ethology. Section 3.2 rewrites the MLE and the WLSE
for this particular case. In Section 3.3, we show that it is not possible to consistently estimate the parameter of the model on a single experiment of length n,
even if n tends to infinity. In order to assess the performance of the estimators,
a short simulation experiment is reported in Section 3.4. Section 4 reports the
study on the experimental data. The experimental protocol and the data produced are described in Sections 4.1 and 4.2. Section 4.3 supplies the estimation
4
results (computation of the estimators and their confidence regions). We provide
some concluding remarks in Section 5 and prove the theoretical statistical results
of this paper in Section 6.
2
Parameter estimation
Let us first write precisely the model studied and statistical framework used. We
assume that {Xk , k ≥ 1} is a sequence of Bernoulli random variables (representing
the colors of the balls drawn: 1 for red and 0 for black) and that there exists a
function f0 : N2 → [0, 1] such that, for all integers k,
P(Xk+1 = 1 | Fk ) = f0 (Zk , k − Zk ) .
where Z0 = 0 and for k ≥ 1, Zk = X1 + · · · + Xk and Fk is the sigma-field
generated by Z0 , X1 , . . . , Xk . The random walk {Zk , k ≥ 0} is an inhomogeneous
Markov chain. For n ≥ 1 and a sequence (e1 , . . . , en ) ∈ {0, 1}n , applying the
Markov property, we obtain
P(X1 = e1 , . . . , Xn = en ) =
n−1
Y
f0 (zk , k − zk )ek+1 {1 − f0 (zk , k − zk )}1−ek+1 , (3)
k=0
where zk = e1 + · · · + ek with z0 = 0, by convention.
We assume that we observe N experiments, each consisting in a path of
length n of the model (3). In other words, we have a set of N sequences of n consecutive draws. For j = 1, . . . , N and k = 1, . . . , n, let Xkj ∈ {0, 1} denote the
color of the k-th ball drawn in the j-th experiment. Let Z0j = 0 and Zkj =
Pk
j
i=1 Xi , k ≥ 1 be the total number of red balls drawn at time k during the j-th
j
experiment, so that Xkj = Zkj − Zk−1
. In all the paper, n will be fixed and our
asymptotic results will be obtained with N (the number of experiments) tending
to ∞.
Let Θ ⊂ Rd and f : Θ × N2 → (0, 1) be a function, called the choice function
of the urn. We assume that there exists θ0 ∈ Θ such that f0 (·, ·) = f (θ0 , ·, ·),
i.e. for k = 0, . . . , n − 1 and i = 0, . . . , k,
j
P(Xk+1
= 1 | Zkj = i) = f (θ0 , i, k − i) .
To proof the consistency and the asymptotic normality of the estimators introduced below, we need the following assumptions on the choice function f . For
any function g defined on Θ, we denote ġ and g̈ the gradient and Hessian matrix
with respect to θ, ∂s g the partial derivative with respect to the s-th component θs
of θ, 1 ≤ s ≤ d and A0 the transpose of the vector or matrix A.
5
Assumption 1.
(i) (Regularity) The set Θ is a compact with non empty interior. For 0 ≤
i ≤ k ≤ n − 1, f0 (i, k − i) > 0 and the function θ → f (θ, i, k − i) is twice
continuously differentiable on Θ.
(ii) (Identifiability) If f (θ1 , i, k − i) = f (θ2 , i, k − i) for all 0 ≤ i ≤ k ≤ n − 1,
then θ1 = θ2 ,
(iii) The n(n − 1)-dimensional vectors {∂s f (θ0 , i, k − i), 0 ≤ i ≤ k ≤ n − 1},
1 ≤ s ≤ d, are linearly independent in Rn(n−1) .
Assumption 1 ensures that θ0 is the unique maximizer of L and that the Fisher
information matrix
In (θ0 ) = −L̈(θ0 ) =
n−1 X
k
X
k=0 i=0
P(Zk = i)
f˙0 (i, k − i)f˙0 (i, k − i)0
¯
f0 (i, k − i)f0 (i, k − i)
is invertible, where we denote f˙0 (i, k − i) = f˙(θ0 , i, k − i) and f¯0 (i, k − i) =
1 − f0 (i, k − i). The explicit expression of the probabilities P(Zk = i), i, k ∈ N, is
given in Section 6.1, Equation (13).
2.1
Maximum likelihood estimation (MLE)
The structure of the model (3) allows to have an explicit expression of likelihood VN (θ). The independence of the N experiments yields the following multiplicative form
VN (θ) =
N n−1
Y
Y
j
j
f (θ, Zkj , k − Zkj )Xk+1 {1 − f (θ, Zkj , k − Zkj )}1−Xk+1 .
j=1 k=0
The log-likelihood function LN based on N paths, is thus given by
LN (θ) =
N X
n−1
X
j
Xk+1
log f (θ, Zkj , k − Zkj )
j=1 k=0
j
+ (1 − Xk+1
) log{1 − f (θ, Zkj , k − Zkj )} . (4)
Let θ̂N be the maximum likelihood estimator of θ0 , that is
θ̂N = arg max LN (θ) .
θ∈Θ
6
(5)
Define L(θ) = N −1 E[LN (θ)] (where the dependence in n is omitted). Then,
L(θ) =
n−1
X
E f0 (Zk , k − Zk ) log f (θ, Zk , k − Zk )
k=0
+ {1 − f0 (Zk , k − Zk )} log{1 − f (θ, Zk , k − Zk )} . (6)
Let N (m, Σ) denote the Gaussian distribution with mean m and covariance Σ.
Theorem 2. If Assumptions 1-(i) and 1-(ii) hold then the maximum likelihood
estimator θ̂N is a strongly consistent estimator of θ0 . If moreover√Assumption 1-(iii)
holds and θ0 is an interior point of Θ, then, as N tends to ∞, N (θ̂N − θ0 ) converges weakly towards N (0, In−1 (θ0 )).
The proof is in Section 6.4. It is the consequence of a more general result
stated and proved therein.
2.2
Weighted least squares estimation (WLSE)
Least squares estimators are very popular among practitioners. Moreover for some
urn models, the MLE may be numerically unstable hence difficult (and lengthy) to
compute so the WLSE constitutes a convenient alternative. In order to describe
this estimator, we introduce some notation. For 0 ≤ i ≤ k ≤ n − 1, define
PN
j
1
N
1 X
j=1 1{Zkj =i} Xk+1
N
, (7)
aN (i, k − i) =
1 j
and pN (i, k − i) =
N j=1 {Zk =i}
aN (i, k − i)
with the convention 00 = 0. The quantity aN (i, k − i) is the empirical probability
that i red balls have been drawn at time k and pN (i, k − i) is the empirical
conditional probability that a red ball is again chosen at time k + 1 given i red
balls were drawn at time k.
We further define qN (i, k−i) = 1−pN (i, k−i) and f¯(θ, i, k−i) = 1−f (θ, i, k−i).
Let {wN (i, k − i), 0 ≤ i ≤ k} be a sequence of weights and define the contrast
function
WN (θ) =
n−1 X
k
X
wN (i, k − i){pN (i, k − i) − f (θ, i, k − i)}2 .
(8)
k=0 i=0
The weighted least squares estimator minimizes WN , that is
W
θ̂N
= arg min WN (θ) .
θ∈Θ
7
(9)
Theorem 3. If Assumptions 1-(i) and 1-(ii) hold and if the weights wN converge
almost surely to a sequence of positive weights w0 , then the weighted least squares
W
estimator θ̂N
is a strongly consistent estimator of θ0 .
If moreover
1-(iii) holds and θ0 is an interior point of Θ, then as N
√ Assumption
W
tends to ∞, N (θ̂N − θ0 ) converges weakly towards N (0, Σn (θ0 )), where Σn (θ0 )
is a definite positive covariance matrix.
If moreover wN (i, k − i) = pN −1 qN −1 aN (i, k − i), for all 0 ≤ i ≤ k ≤ n − 1, the
W
estimator θ̂N
is asymptotically efficient, i.e. Σn (θ0 ) = In (θ0 ).
The proof is in Section 6.4, where a explicit expression of Σn (θ0 )−1 is supplied.
2.3
Generalization
It is possible to extend the statistical framework introduced in the previous
sections to a large class of reinforced random walks on graphs. For instance,
let G = (X
, E) be a locally finite non oriented graph, with X the set of its vertices
and E ⊂ {i, j} : i, j ∈ X , i 6= j the set of its non oriented edges. We denote
by x ∼ y, if {x, y} ∈ E. Let NX be the set of integer vectors indexed on X . We
define X = (Xn )n≥0 a random walk on G, i.e. a sequence of vertices, such that,
for all n ≥ 0, Xn+1 ∼ Xn . The vector Zn ∈ NX is such that, for all vertex x ∈ X ,
Zn (x) is the number of times the walk X has visited x up to time n. Let Θ ⊂ Rd ,
d ≥ 1, we suppose that the walk X is vertex-reinforced and that there exists a
choice function f : Θ × X × NX → [0, 1] and a parameter θ ∈ Θ such that, for
all n ≥ 0 and x ∈ X ,
P(Xn+1 = x|X0 , · · · , Xn ) = f (θ, x, Zn )1{x∼Xn } .
For instance, the choice function can be similar to the one proposed by J.-L.
Deneubourg (see Deneubourg et al., 1990), for n ≥ 0 and x ∈ X ,
α
c + Zn (x)
P(Xn+1 = x|X0 , · · · , Xn ) = X
α 1{x∼Xn } ,
c + Zn (y)
y∼Xn
where Θ = (0, +∞)2 and θ = (α, c).
The statistical framework introduced for the general urn model is easily adaptable to this vertex reinforced random work. Under some adequate regular conditions, it would be not difficult to prove the consistency and the asymptotic
normality of the MLE and the WLSE by establishing a theorem similar of the
general result proved in Section 6.3.
8
3
Application to an ethological problem
In 1990, J.-L. Deneubourg et al. used a particular urn model to reproduce the
sequences of consecutive choices made by ants at a fork (see Deneubourg et al.,
1990). Let replace the urn filled with balls of two colors by a fork with two
branches. Drawing a ball and adding a ball of the same color in the urn is
equivalent to an ant choosing a branch and reinforcing it with pheromone by going
throw it. Then the probability to draw a red ball depending on the previous draws
is equal to the probability to choose the right branch depending on the previous
passages.
We first introduce the model proposed by J.-L. Deneubourg et al. in the statistical framework described in the previous section. Futher, we provide a description of the model behavior depending on the parameter value. We also supply an
ethological interpretation of the parameter. We then prove that it is impossible
to estimate the parameter on a single path. Finally we report the study of the
estimator accuracy that we perform on simulated data.
3.1
Formalization and results
Behavioral assumptions We first introduce the hypotheses assumed on the
ant behavior.
1. Each ant regularly deposits a constant amount of pheromone as it walks.
2. The ants are strictly identical which means that every ant has the same
reaction to the same amount of pheromone.
3. Pheromone trails do not evaporate during the experiment.
4. Each ant reaches the fork alone, chooses a branch and leaves the bifurcation
by crossing only once into the chosen branch without passing through or
reinforcing the non-chosen branch.
Under these assumptions, the quantity of pheromone laid on each branch is proportional to the number of passages through it. In path formation modeling, these
assumptions are commonly made. However, because of the inter-individual variability in ant behavior, the first two assumptions are unrealistic. For instance, the
pheromone perception noise implies that each ant could detect a different signal
from the same quantity of pheromone. These assumptions are an approximation
of the real ant behavior. The implicit hypothesis here is that the inter-individual
variability is small enough to consider that all ants are identical. For the third
assumption, we suppose that the persistence of the pheromone trails allows to
ignore the evaporation of the pheromone. Experimental protocols are designed to
9
make the four assumptions more acceptable by choosing the ant species adequately
and by placing them in an appropriate situation (see Section 4).
The model The random variable Xk , introduced previously, is the choice of
the k-th ant going through the fork (1 for right and 0 for left). Consequently, for
k ≥ 1, Zk is the number of passages through the right branch after k passages.
For θ = (α, c) ∈ (0, ∞)2 and all integers 0 ≤ i ≤ k, we define choice function
f (θ, i, k − i) =
(c + i)α
.
(c + i)α + (c + k − i)α
(10)
We assume that the probability that an ant chooses the right branch at time k + 1
given the first k choices is given by:
P(Xk+1
(c + Zk )α
= 1|Fk ) = f (θ, Zk , k − Zk ) =
,
(c + Zk )α + (c + k − Zk )α
(11)
where c > 0 is the intrinsic attractiveness of each branch and α > 0 is the
possible non-linearity of the choice. This process is an urn model which has been
exhaustively investigated in the probabilistic literature. We recall here its main
features.
Theorem 4.
(i) If α < 1, then
lim
n→∞
1
Zn
= a.s.
n
2
(ii) If α = 1, then Zn /n converges almost surely to a random limit with a
Beta(c, c) distribution with density x → Γ−2 (c)Γ(2c)xc−1 (1 − x)c−1 with respect to Lebesgue’s measure on [0, 1], and Γ is the Gamma function.
(iii) If α > 1, then eventually only one branch will be chosen, i.e.
∃x ∈ {0, 1} , ∃n0 ∈ N , ∀n ≥ n0 , Xn = x .
The case α < 1 is due to Tarrès (2011); the case α > 1 to Davis (1990) and
the case α = 1 to Pólya (1931) (see Freedman (1965) for an online access).
Ethological interpretation of the parameters and properties of the model
The parameter α characterizes the ant’s differential sensitivity to the pheromone.
When α > 1, the ants can detect better and better increasing amounts of pheromones laid on each branch and thus are more likely to choose the branch with the
10
most pheromones. Moreover, after a random but almost surely finite number of
passages, one branch will eventually be selected, i.e. all ants will afterwards choose
this branch (see Theorem 4-(iii)). In the opposite case, when α < 1, the ants are
less able to perceive the differences between the amounts of pheromones laid on
each branch as these amounts increase. This minimization effect is so strong that
the proportion of passages on each branch converges to 1/2 (see Theorem 4-(i)). It
is important to note that, when α 6= 1, the asymptotic behavior of the proportion
of passages through each branch depends only on α and not on c. Furthermore
the larger α is or the closer α is to zero, the faster these effects will happen.
The role of c is clearer when (11) is rewritten as follows:
P(Xk+1 = 1|Fk ) =
(1 + Zk /c)α
.
(1 + Zk c)α + (1 + (k − Zk )/c)α
The parameter c is the inverse of the reinforcement incrementation and thus can be
interpreted as the inverse of the attractiveness (or the strength) of the pheromones
laid at each passage. Consequently when α is neither very close to zero nor very
large, c has a strong short term influence. When c is small compared to 1, the
first passage strongly reinforces the first chosen branch. Thus during the first
few passages, a branch will be highly favored even if α < 1 (in which case the
branches will eventually be uniformly crossed). When c is large, the first passages
weakly reinforce the chosen branches. Then if α > 1 (in which case a branch will
eventually be selected), a large number of passages must be observed before the
clear emergence of a preference. Naturally, the larger c is or the closer c is to zero,
the longer these effects will be seen.
When α = 1, the asymptotic behavior of the passage proportion is determined
by c (see Theorem 4-(ii)). As c grows from zero to infinity, the limiting distribution
of Zn /n (as n → ∞) evolves continuously from two Dirac point masses at 0 and 1
to a single Dirac mass at 1/2. To illustrate this point, we show in Figure 1 the
density of the Beta distribution for c = 0.1 and c = 10. We make some further
comments.
• If c < 1, a strong asymmetry in the choices of the branches appears. One
branch is eventually chosen much more frequently than the other. Furthermore as c tends to 0, the Beta distribution tends to the distribution with
two point masses at 0 and 1. This limit case corresponds to the situation in
which a branch is selected, i.e. α > 1.
• If c = 1, the limiting distribution is uniform on [0, 1].
• If c > 1, Zn /n appears to be much more concentrated around 1/2. This is
similar to what is observed in the case α < 1.
11
Figure 1: Graph of the Beta distribution for parameters (c, c) with c = 0.1, c = 1
and c = 10
To summarize, the model possesses four phases: fast and slow selections and
fast and slow uniformizations. These phases are delimited by two phase transitions: a discontinuous one between α > 1 (branch selection) and α < 1 (branch
uniformization) with a critical state α = 1 and a smooth one between c < 1 (strong
pheromone deposits) and c > 1 (weak pheromone deposits). These properties are
summarized in the phase diagram in Figure 2. When c is small, it is very likely
that one branch will be favored during the first passages, thus the empirical distribution of the choices resembles the Beta distribution with a small c (the grey
solid line in Figure 1). As the number of experiments increases, the shape of the
empirical distribution will be closer and closer to its limit: a Dirac mass at 1/2
if α < 1 and two Dirac masses at 0 and 1 if α > 1. When c is large, the earlier passages do not show any preference between the branch. Again, when the number
of experiments increases, the asymptotic behavior is progressively revealed.
To date, the most commonly used behavioral state in the model is the slow selection of a branch (see Deneubourg et al., 1990; Beckers, Deneubourg, and Goss,
1992). But at least two other states are interesting. The fast uniformization
can describe the case where none of the branches are preferred. The slow uniformization could reproduce the saturation phenomenon. There exists a threshold
concentration of pheromone upon which ants can no longer detect the pheromone
concentration variations (see Pasteels, Deneubourg, and Goss, 1987). In experiments involving many ants, one can first observe the favorization of a branch.
But when this branch is saturated (its attractiveness stops increasing), ants go
more and more through the other branch, whose attractiveness still increases.
Eventually, the two branches are uniformly chosen.
12
Remark 5. In the context of an estimation procedure, the similar effects of α
and c (favorization/selection of a branch or not) induce an identifiability issue.
Indeed, we observe a finite number of choices and consequently we only see the
short term behavior. We have seen that the favorization of a branch in the first
passages could be due to a pair of parameter values (α, c) with c small compared
to 1 and α close to 1 or to a pair (α, c) with c close to 1 and α large compared to 1.
On the other hand, if the first passages are nearly uniform on the two branches,
it could be the result of c close to 1 and α small, or of c large and α close to 1.
Thus, we can expect that the estimation of the parameters will be difficult
when both parameters contribute to the same effect, e.g. α large and c small (fast
selection of one branch) or α small and c large (no selection); and also when
the parameters have competing effects: very small c and α < 1, or very large c
and α > 1. The statistical procedure that we introduce in this paper partially
circumvents this difficulty, since it focuses on the transition probabilities rather
than on the general shape of a curve, which is what calibration methods do. This
will be illustrated in Section 3.4.
Figure 2: Phase diagram of the model. The graphs in the shaded boxes show
the shape of the empirical distribution of Zn /n for small n (left) and its limiting
distribution (as n → ∞, right).
13
3.2
Parameter estimation
We start with the maximum likelihood estimator, that is θ̂N = (α̂N , ĉN ) defined
by (5). The Fisher information matrix has the following expression.
In (θ) =
n−1 X
k
X
P(Zk = i)f (α, c, i, k − i)f¯(α, c, i, k − i)J (α, c, i, k − i) ,
(12)
k=0 i=0
with θ = (α, c) and for 0 ≤ i ≤ k ≤ n − 1,
c+i
log
log2 c+k−i
α(k−2i)
J (α, c, i, k − i) =
c+i
log c+k−i (c+i)(c+k−i)
α(k−2i)
c+i
c+k−i (c+i)(c+k−i)
α2 (k−2i)2
(c+i)2 (c+k−i)2
!
.
It is important to note that the Fisher information matrix is not diagonal. Thus
the estimation of each parameter has an effect on the estimation of the other.
Corollary 6. Let Θ be a compact subset of (0, ∞)2 which contains (α0 , c0 ). Then
the maximum√
likelihood estimator θ̂N is consistent and asymptotically normal and
efficient, i.e. N (θ̂N − θ0 ) converges weakly to N (0, In−1 (θ0 )).
As mentioned above, we also use weighted least squared estimators defined
by (9) for several different weight sequences wN , such that wN (i, k − i) converges
almost surely to w0 (i, k − i) > 0, for all 0 ≤ k ≤ n − 1 and 0 ≤ i ≤ k.
Corollary 7. Let Θ be a compact subset of (0, ∞)2 which contains (α0 , c0 ) and
assume that wN converges almost surely to positive weights w0 . Then the
√weighted
W
W
least squared estimator θ̂N
is consistent and asymptotically normal, i.e. N (θ̂N
− θ0 )
converges weakly to N (0, Σn (θ0 )), where Σn (θ0 ) is a positive definite covariance
−1
matrix. It is efficient, if wN (i, k − i) = p−1
N qN aN (i, k − i), for all 0 ≤ k ≤ n − 1
and 0 ≤ i ≤ k.
The proofs of these corollaries are in Section 6.5. They are an application of
Theorems 2 and 3. Since Assumption 1-(i) obviously holds, it remains only to
check (ii) and (iii) of Assumption 1.
3.3
Estimation on a single path
The main feature of the binary choice model for α0 > 1 is that only one branch
will be crossed eventually. It seems clear then that a statistical procedure based
on only one path (one sequence of choices) cannot be consistent, since no new
information will be obtained after one branch is eventually abandoned. This
intuition is true and more surprisingly, it is also true in the case α0 = 1. This
is translated in statistical terms in the following theorem. Let `n denote the loglikelihood based on a single path of length n and `˙n its gradient. The model is
regular, so the Fisher information is varθ (`˙n (θ)).
14
Theorem 8.
(i) If α0 = 1 and c0 > 0, then limn→∞ In (c0 ) < ∞.
(ii) If α0 < 1, n−1 `n (θ0 ) → − log 2.
(iii) If α0 > 1, then `n (θ0 ) converges almost surely to a random variable as n → ∞.
The proof is in Section 6.6. Statement (i) means that, when α0 = 1, the
Fisher information is bounded. This implies that the parameter c0 cannot be
estimated on a single path. This also implies that the length n of each path
should be taken as large as possible (theoretically infinite) in order to minimize
the asymptotic variance of the estimators. Statements (ii) and (iii) imply that the
maximum likelihood estimator is inconsistent, since the likelihood does not tend
to a constant.
3.4
Simulation experiment
In order to assess the quality of the estimators proposed, we have made a short
simulation study. For several pairs (α, c), we have simulated 1000 experiments
of N = 50 paths of length n = 100 (recall that n is the number of ants going
through the bifurcation). These are reasonable values in view of the practical
experiments with actual ants. We compare the performance of the maximum
likelihood estimator θ̂N (MLE) defined in (5) and of the weighted least squares
W
estimator θ̂N
(WLSE) defined in (9) with the weights wN (i, k − i) = aN (i, k − i)
defined in (7). The asymptotically efficient WLSE with the weights wN (i, k − i) =
aN (i, k − i)pN (i, k − i)−1 qN (i, k − i)−1 provides a severely biased estimation of α
and always estimates a very small value of c with a very small dispersion. This
is caused by the fact that the empirical pN and qN vanish frequently, so that the
weights are infinite. We will not report the study for this estimator.
The theoretical standard deviation We first evaluate numerically some values of the theoretical standard deviations of both estimators for several values of α
and c. We have chosen arbitrary values of α and c in the range 0.5, 2. We have
also chosen values of α and c which correspond to those found in the literature
cited and to those that we have estimated in the real life experiment described in
Section 4. These results are reported in Table 1 and in Figure 3 and their features
are summarized in the following points.
• As theoretically expected, the asymptotic variance of the MLE, which is the
Fisher information bound, is smaller than the variance of the WLSE, but
the ratio between the variances of the two estimators is never less than one
fourth. Moreover, their overall behavior is similar.
15
• The variance of the estimators of α is smaller when both parameters do not
contribute to the same effect. The worst variance is for α large and c small,
that is when the values of both parameters imply fast selection of a branch.
The variance tend to infinity when α tends to infinity.
• The variance of the estimators of c increases with c and tends to infinity
when α tends to 0 and to ∞.
• These effects are explained by the fact that the coefficients of the Fisher
information matrix tend to zero when α tends to zero, except the coefficient
corresponding to α. See Formula (12).
(a) TSD of α
(b) TSD of c
Figure 3: Theoretical standard deviation for N = 50 paths of length n = 100 of
the MLE of α (a) and of c (b), for α in (0, 2] and fixed values of c. Theoretical
standard deviation of the WLSE has the same shape, but with higher convergence
speed as α tend to 0 or ∞.
Performance of the estimators Recall that we have simulated 1000 experiments, each of N = 50 paths of length n = 100. Because of the length of the
computations, each MLE was computed only 500 times. Table 1 reports root
mean squared error (MSE) of both estimators based on the simulated data for the
same values of the parameters and their features are summarized in the following
points.
• For most values of the parameters, the MSE are close to the theoretical
standard deviation.
16
• The MSE increase significantly when c is large or when both parameters
contribute to the same effect.
• This increase is more noticeable for the WLSE than for the MLE.
• This increase is in part due to the skewness of these estimators. For some
values of the parameters, both estimators tend to overestimate the parameters.
• For the MLE, the MSE is much larger in the case of non selection than in
the case of selection where the empirical performance of the MLE nearly
matches the theoretical value.
• These effects are always stronger for the estimation of c than for the estimation of α.
This degraded performance for some specific or extreme values of the parameters is in part due to numerical issues.
• In the case where selection of a branch is fast, many of the empirical weights
used to compute the WLSE vanish, and the least squares method uses very
few points to fit the curve. The MLE is not affected by this problem.
• In the case where both parameters concur to non selection, the probability
of choosing one branch converges very fast to 1/2, and thus the experiment
brings very little information. This affects both the MLE and the WLSE,
and in addition, many of the empirical weights vanish so the WLSE is even
less efficient.
The degraded performance may also be caused by the identifiability problem
explained in Remark 5, i.e. the similar effects of the two parameters makes the
estimation more difficult.
Bootstrap confidence intervals Since the asymptotic variance depends on
the unknown parameters, we have computed the pivotal Bootstrap 95% confidence
intervals for the parameters based on one simulation of N = 50 paths of length
n = 100 and a Bootstrap sample size of 500 (see Wasserman (2004), Section 8.3,
for details on this method). We have compared these Bootstrap intervals with
the corresponding Monte-Carlo intervals, based on 500 simulations (see Table 2).
The match is nearly perfect for the MLE for α, but as before, the performance is
poorer for the estimation of c. The intervals for c are noticeably skewed to the
right but always contain the true value. For further comparison, we only show
here the results corresponding to the values of the parameters estimated in the
real life experiment reported below and those corresponding to values found in
the earlier literature.
17
Table 1:
Theoretical standard deviations (TSD) for N = 50 paths of
length n = 100 and square root of the mean squared errors (MSE) for 500 (for
the MLE) or 1000 (for the WLSE) simulated experiments of N = 50 paths of
length n = 100. All figures of this table must be mutiply by 0.01.
(α, c)
(0.5, 0.5)
(0.5, 1.0)
(0.5, 2.0)
(1.0, 0.5)
(1.0, 1.0)
(1.0, 2.0)
(1.5, 0.5)
(1.5, 1.0)
(1.5, 2.0)
(2.0, 0.5)
(2.0, 1.0)
(2.0, 2.0)
(2.0, 20.0)
(2.6, 60.0)
(1.1, 3.0)
(1.1, 7.0)
α
c
MLE
WLSE
MLE
WLSE
√
√
√
√
TSD
M SE TSD
M SE TSD
M SE TSD
M SE
5.02
5.60
6.25
7.60
25.4
54.1
31.7
58.8
6.45
7.39
8.56
10.2
59.8
137
79.4
169
8.80
16.9
12.9
356
146
651
214
29500
3.81
3.90
4.97
6.25
11.8
12.4
15.9
21.0
4.34
4.98
5.91
7.34
25.2
31.3
34.9
47.3
5.83
5.80
8.72
9.38
59.3
64.9
89.1
111
7.83
7.85
12.8
19.4
12.3
13.0
18.7
24.6
6.69
6.59
10.8
15.1
21.0
21.4
33.1
40.3
6.88
7.24
11.5
13.9
41.7
45.9
68.8
78.4
19.4
26.1
41.8
2690
16.4
19.8
28.6
1410
13.5
14.2
28.7
985
23.9
26.0
44.3
988
11.2
11.7
23.1
35.4
40.7
43.4
77.2
101
35.5
58.8
129
124
794
1420
2880
3240
166
864
1230
3590
5780
32500
43200 142000
7.20
7.80
11.9
11.3
89.9
111
148
201
13.8
17.4
29.5
28.3
297
433
633
758
All figures must be multiply by 0.01
Table 2: Monte-Carlo 95% confidence intervals for 500 simulated experiment of
N = 50 paths of length n = 100 and Bootstrap 95% confidence intervals for one
simulated experiment of 50 paths of length 100
IDC 95% for α
IDC 95% for c
(α, c)
Monte-Carlo Bootstrap Monte-Carlo Bootstrap
MLE (2.0, 20) (1.50, 3.38) (1.90, 5.10) (9.90, 55.5) (19.1, 94.4)
(2.6, 60) (1.23, 29.3) (1.12, 38.7) (15.9, 1053) (19.9, 1637)
(1.1, 3.0) (0.98, 1.29) (1.01, 1.23) (1.81, 6.22) (1.34, 3.99)
WLSE (2.0, 20) (1.26, 4.85) (1.06, 3.99) (5.92, 88.1) (2.76, 69.3)
(2.6, 60) (0.74, 88.3) (0.24, 87.7) (3.86, 3754) (0.17, 4132)
(1.1, 7.0) (0.75, 1.83) (0.55, 2.56) (1.68, 30.3) (1.48, 66.0)
18
4
Real life experiment with ants
In this section, we apply the previous estimators on data from a path selection
experiment by a colony of ants.
4.1
Experiment description
This experiment was done in the Research Center on Animal Cognition (UMR 5169)
of Paul Sabatier University Toulouse under the supervision of Guy Theraulaz,
Hugues Chaté and the first author. A small laboratory colony (approximately
200 workers) of Argentine ants Linepithema humile was starved for two days before the experiment. During the experiment, the colony had access to a fork
carved in a white PVC slab, partially covered by a Plexiglas plate (see Figure 4).
The angle between the branches was 60◦ . The fork galleries had a 0.5 cm square
section. The entrance of the maze was controlled by a door. Food was never
present during the experiment. The maze was initially free of any pheromone
trail.
Figure 4: The experimental set-up: a fork carved in a white PVC slab, partially
covered by a Plexiglass plate.
Each trial (N = 50) consisted in introducing separately each ant to the entrance of the fork (see Figure 4) one at a time. Once inside, an ant must choose
between the left or the right branch of the fork. As soon as the ant had made a
choice and stepped into one branch, it was removed from the set-up and another
ant was introduced. All the choices were recorded and a trial ended when 100 ants
had passed through the fork.
This experimental protocol was designed to strengthen the behavioral assumptions described in Section 3.1. Any return to the fork is forbidden so that we can
19
consider that each ant passed only one time. There was never more than one
ant in the set-up. This implies that each ant in the maze received no other
cue about the previous passages than the pheromone that was been laid. The
species Linepithema humile was in part chosen to justify the assumption of identical pheromone deposits. Indeed, these ant may deposit regularly the same type
of pheromone on their trajectory (see Van Vorhis Key and Baker, 1982; Aron,
Pasteels, and Deneubourg, 1989). All ants were prepared the same way before
the experiments to increase the credibility of the assumption stating that each
ant behaved by the same way. The length of the experiments was limited to stay
close to the half-life duration of the pheromone trails (see Jeanson et al., 2003).
4.2
Data representation
Figures 5(a) shows the 50 paths of length n = 100, that is, 50 choice sequences of
100 ants that went through the bifurcation. The paths are represented as random
walks with increment +1 when the right branch is chosen, and −1 when the left
one is chosen. In less than ten experiments, a branch seemed to be selected,
whereas in the others, selection of a branch was not obvious. Figure 5(b) shows
the histogram of the distribution of Z100 /100, that is the final proportion of the
choices of the right branch. There is no clear visual evidence that α > 1 as it is
claim in the literature (see Deneubourg et al., 1990; Vittori et al., 2006; Garnier
et al., 2009).
(a) The 50 paths of n = 100 ants choosing
either left (+1) or right (-1)
(b) Histogram of the final proportion of right
passages (Z100 /100)
Figure 5: Data representation
20
4.3
Parameter estimation
Several values of these parameters have been proposed in the applied literature.
Deneubourg et al. (1990) proposed α = 2, c = 20 and more recently Garnier
et al. (2009) suggested α = 2.6 and c = 60. It must be noted however that
these values are not obtained by a statistical method but by the calibration of
a curve to a plot. Therefore, these methods do not lead to confidence intervals.
Moreover, a calibration method has an inherent risk of over fitting, because of the
identifiability problem explained in Remark 5. As illustrated in Figure 2, if for
instance α and c are both small, then both branches will be asymptotically equally
chosen, but paths of finite length n might be misleading and the calibration will
suggest values of α and c corresponding to the selection of a branch. The statistical
procedure is based on the dynamics of the process and is thus less prone to this
type of error. Nevertheless, we will see that our results do not contradict those of
Deneubourg et al. (1990) and Garnier et al. (2009), but complement them.
Table 3: The MLE and the WLSE for the 50 paths of real ants and their
Boostrap 95% confidence intervals
α̂
Bootstrap 95% CI
ĉ
Bootstrap 95% CI
MLE 1.07
(0.80, 1.99)
3.26
(1.14, 23.0)
WLSE 1.10
(0.62, 3.81)
6.91
(0.94, 85.4)
Table 3 shows the results of the maximum likelihood estimation and the
weighted least squares estimation. Both estimates of α are close to 1.1 and the
estimates of c are between 3 and 7. The 95% confidence intervals are slightly
larger than the simulated ones (see Table 2). This increased variability may be
due to the extreme paths which seem to show a very fast selection of one branch
(see Figures 5(a) and 5(b)). This may suggest that the ants did not have the
same behavior and that the distribution of Z100 /100 could be a mixture of two
distributions.
For both methods, the 95% Bootstrap confidence intervals of α contain the
value 1. More precisely, as shown in Figure 6, approximately 1/3 of the bootstrap parameters gives weak pheromone deposits (c > 1) and a weak differential
sensitivity (α < 1), which means that branches are eventually uniformly crossed.
In almost all the others cases, we conclude for weak pheromone deposits (c > 1)
and a strong differential sensitivity (α > 1), which means that a branch will be
eventually, though slowly, selected. In only a few cases do the estimators give
strong pheromone deposits (c < 1), but a weak differential sensitivity (α < 1),
which means that a branch is chosen more than the other at the beginning of the
experiment, but branches are eventually uniformly crossed. Finally, there are no
values which imply both strong pheromone deposits (c < 1) and a strong diffe21
rential sensitivity (α > 1). Therefore, we can conclude that pheromone deposits
are weak with a good confidence but we cannot confidently decide for α.
The values obtained by Deneubourg et al. (1990) (α = 2, c = 20) and more
recently by Garnier et al. (2009) (α = 2.6, c = 60) are both in the confidence
intervals for the WLSE found in Table 3. But the values of α suggested by
these authors are out of the 95% confidence interval for the MLE. Thus these
parameters, which decide for a slow branch selection, are no more likely than a
parameter set which would yield non selection of a path.
Figure 6 illustrates the fact that the two estimators are strongly positively
correlated. There seems to be two cutoff values for c: if ĉ∗ > 8, then α̂∗ > 1, and
if ĉ∗ < 1.5, then α̂∗ < 1. The above mentioned values reported by Deneubourg
et al. (1990) and Garnier et al. (2009) exhibit these features: they both have c > 8
and α > 1 and α increase with c.
If we fix the value of c and estimate only α, then the 95% Bootstrap confidence
intervals for α are smaller. Figure 7 shows the estimated values of α and the
confidence intervals as functions of the fixed value of c. We see that if c is greater
than 6 for the MLE (or than 12 for the WLSE), then the confidence intervals of α
lie entirely above 1. Furthermore if c is less than 0.8 for the MLE (or than 2 for the
WLSE), then the confidence intervals of α lie entirely under 1. This shows that if
the deposits are weak enough, i.e. c > 12, we can conclude that a slow selection
of a branch will occur with probability 1. On the other hand, if the deposits
are strong enough, i.e. c < 0.8, we can conclude that branches will eventually be
uniformly crossed with probability 1.
(a) MLE
(b) WLSE
Figure 6: Log-log scatterplots of the estimates (α̂∗ , ĉ∗ ) for the 500 Bootstrap
samples for the MLE (a) and the WLSE (b) for the 50 paths of real ants.
22
(a) MLE
(b) WLSE
Figure 7: Graph of the estimates α̂ and their Bootstrap 95% confidence interval
for the MLE (a) and the WLSE (b) for the 50 paths of real ants as a function of
fixed value of c.
5
Concluding remarks
In the literature no parameter estimation methods for reinforced random walks
can be found. To partially fill this void, this article proposes a statistical framework to estimate the parameter of a general two-colored urn model. We define
the maximum likelihood estimator (MLE) and the weighted least squares estimators (WLSE) for the parameter of this model and prove their consistency and
their asymptotically normality under some usual regularity conditions. The proof
lies on a general result for a large class of estimators called minimum contrast
estimators. The MLE is asymptotically efficient, but can be difficult (lengthy)
to compute, which can be an issue specially while using Bootstrap algorithms.
The WLSE is a suitable alternative. Moreover this estimator is popular among
practitioners.
We apply this statistical tools to the problem of path selection by an ant
colony. To this purpose, we performed experiments with actual ants to collect
data. The experiment consisted of introducing one hundred ants into a Y shaped
device, one at a time, and observing their successive choices. We also consider
the particular urn model introduced by Deneubourg et al. (1990) to describe this
phenomenon. This urn has two parameters, α and c, which have distinct biological
interpretations, but contribute to the same effect: either selection of a branch or
uniformization of the choices. The parameter c influences the short term behavior,
whereas α determines the asymptotic behavior. Consequently the model exhibits
four phases which are illustrated by Figure 2. The case most commonly considered
in the literature is the case of slow selection, which corresponds to α > 1 and c > 1:
the ants will eventually always choose the same branch, but this selection will take
23
a long time. For instance, Deneubourg et al. (1990) provides the values α = 2
and c = 20. However other phases can be relevant to describe the ant behavior.
For instance the fast uniformization, corresponding to α < 1 and c > 1, can model
the less likely but not negligible case in which ants do not select a branch.
After assessing the accuracy of the MLE and the WLSE on simulated data,
we estimate the value of α and c with the two estimators. We also evaluate
confidence regions by Bootstrap proceeding. The estimated values of α and c
ranged between 1.1 and 3 and between 3 and 7, respectively. This tends to imply
that slow selection of a branch will occur. However, the Bootstrap sample gives a
confidence level of 65% for the hypothesis of slow selection, while the hypothesis
of fast uniformization has a confidence of 35%.
This low level of confidence for the commonly assumed slow selection phase
might be explained by technical reasons. The number of experiments (50) is relatively small; increasing the number of replicas will reduce the confidence regions.
Moreover the competition between the parameters for the same effect induces an
identifiability issue. For instance the apparent preference of a branch may be due
to α > 1 or c small with respect to 1. Therefore, the model, which is biologically
relevant, is statistically difficult to estimate. Indeed, for an ethological study, discriminating the ant pheromone sensitivity from the pheromone deposit strength
is meaningful. But for a statistical procedure, the similarity of effect of the two
parameters scales down the estimation performance.
However, the uncertainty may not come from an inefficiency of the statistical
procedure, but from shortcomings of the ethological hypotheses. Indeed, the estimated confidence intervals computed from the experimental data are larger than
the ones computed from the simulated data (for similar parameter values). Moreover the assumption that the inter-individual variability is negligible is strong.
For instance, it may be necessary to consider that the pheromone deposit varies
at each passage, i.e. that c is random.
These ethological considerations will be further discussed in a forthcoming
paper which will analyze more elaborated experimental designs. The ants will be
observed while freely evolving in a network with several nodes. In addition of a
data analysis, we will model the experiment with a reinforced random walk on
a finite graph for which we have provided probabilistic results (see Le Goff and
Raimond, 2015). The statistical methodology introduced in this paper will be
extended to a larger class of reinforced random walks.
24
6
6.1
Proofs
Distribution of Zk , for k ∈ N
In order to compute the distribution of Zk , we introduce some notation. Let Sk
be the set of sequences of length k + 1 of integers i0 , . . . , ik such that i0 = 0 and
ij − ij−1 ∈ {0, 1} for j = 1, . . . , k. For i ≤ k let Sk (i) = {(i0 , . . . , ik ) ∈ Sk | ik = i}.
Then we have
P(Zk = i) =
X
k−1
Y
f0 (iq , q − iq )iq+1 −iq (1 − f0 (iq , q − iq ))1−iq+1 +iq . (13)
(i0 ,...,ik )∈Sk (i) q=0
6.2
A central limit theorem for the empirical conditional
probabilities
For 0 ≤ i ≤ k ≤ n − 1, recall the definition of aN (i, k − i) and pN (i, k − i) in (7)
and that f¯0 (i, k − i) = 1 − f0 (i, k − i).
√
Lemma 9. { N (pN (i, k − i) − f0 (i, k − i)), 0 ≤ i ≤ k ≤ n − 1} converges weakly
to a Gaussian vector with diagonal covariance matrix Γ0 with diagonal elements
γ0 (i, k − i) =
f0 (i, k − i)f¯0 (i, k − i)
.
P(Zk = i)
(14)
P
j
Proof. Define bN (i, k − i) = N −1 N
j=1 1{Zkj =i} Xk+1 , the empirical estimate of
b(i, k − i) = P(Zk = i, Xk+1 = 1) and a(i, k − i) = E[aN (i, k − i)] = P(Zk = i).
Write then
pN (i, k − i) − f0 (i, k − i) =
bN (i, k − i) − b(i, k − i)
aN (i, k − i)
b(i, k − i)
−
(aN (i, k − i) − a(i, k − i)) .
aN (i, k − i)a(i, k − i)
Since the paths (Z1j , . . . , Znj ), 1 ≤ j ≤ N are i.i.d., the multivariate central limit
holds for the sequence of 2n(n − 1) dimensional vectors {(bN (i, k − i) − b(i, k −
i), aN (i, k − i) − a(i, k − i)), 0 ≤ i ≤ k ≤ n − 1}. The proof is concluded by tedious
computations using the Markov property, which we omit.
Remark 10. We can prove that the covariance matrix Γ0 is diagonal by a statistical argument. If we consider the tautological model {f (i, k − i), 0 ≤ i ≤ k ≤
25
n − 1}, i.e. θ = f and f0 is the true value. Then the likelihood is
LN (f ) =
n−1 X
k
X
aN (i, k − i){pN (i, k − i) log f (i, k − i)
k=0 i=0
+ qN (i, k − i) log f¯(i, k − i)} ,
where f¯(i, k−i) = 1−f (i, k−i). Thus we see that {pN (i, k − i), 0 ≤ i ≤ k ≤ n − 1}
is the maximum
likelihood estimator of f0 . This model is a regular statistical
√
model, thus N (pN − f0 ) converges weakly to the Gaussian distribution with covariance matrix In−1 (f0 ), where In (f ) is the Fisher information matrix of the
model. It is easily seen that In (f0 ) is the n(n − 1) dimensional diagonal matrix
with diagonal elements given by (14).
6.3
A general result for minimum contrast estimators
Theorems 2 and 3 are a consequence of the general result we prove in this section.
More precisely we demonstrate the consistency and the asymptotic normality of
a general estimator of which the MLE and the WLSE are particular cases.
For 0 ≤ i ≤ k ≤ n − 1, recall the definition of aN (i, k − i), pN (i, k − i) in (7)
and that f¯(θ, i, k − i) = 1 − f (θ, i, k − i). Let wN (i, k − i), 0 ≤ i ≤ k ≤ n − 1
be a sequence of random weights and let G be function defined on [0, 1] × (0, 1).
Define the empirical contrast function by
WN (θ) =
n−1 X
k
X
wN (i, k − i)G(pN (i, k − i), f (θ, i, k − i)) .
k=0 i=0
For instance, choosing G(p, q) = −p log q − (1 − p) log(1 − q) and wN (i, k − i) =
aN (i, k − i) yields
WN (θ) = −
n−1 X
k
X
aN (i, k − i) pN (i, k − i) log f (θ, i, k − i)
k=0 i=0
+ qN (i, k − i) log f¯(θ, i, k − i)
= − N −1 LN (θ) ,
so that minimizing WN is equivalent to maximizing the likelihood LN , defined
in (4). Choosing G(p, q) = (p − q)2 yields the weighted least squares contrast
function WN , defined in (8). We now define the minimum contrast estimator of θ0
by
W
θ̂N
= arg min WN (θ) .
θ∈Θ
26
W
In order to prove the consistency and asymptotic normality of θ̂N
, we make
the following assumptions on G and on the weights wN (i, k − i). Let ∂2 G and ∂22 G
denote the first and second derivatives of G with respect to its second argument.
Assumption 11. The function G is non negative, twice continuously differentiable
on [0, 1]×(0, 1) with G(p, q)−G(p, p) > 0 if p 6= q, ∂2 G(p, p) = 0 and ∂22 G(p, p) > 0.
Assumption 12. For all 0 ≤ i ≤ k ≤ n − 1, wN (i, k − i) converge almost surely
to w0 (i, k − i) and w0 (i, k − i) > 0.
W
Theorem 13. If Assumptions 1-(i), 1-(ii), 11 and 12 hold, then θ̂N
is consistent.
If moreover θ0 is an interior point of Θ and Assumption 1-(iii) holds, then
√
W
N (θ̂N
− θ0 ) converges weakly to a Gaussian distribution with zero mean.
The exact expression of the variance is given in the proof.
Proof. Under Assumption 11, the strong law of large numbers shows that WN (θ)
converges almost surely to
W(θ) =
k
n−1 X
X
w0 (i, k − i)G(f0 (i, k − i), f (θ, i, k − i)) .
k=0 i=0
Assumptions 1-(ii) and 11 ensure that θ0 is the unique minimum of W. Indeed,
G(p, q) > 0 if p 6= q and G(p, p) = 0. Thus, W is minimized by any value of θ
such that f (θ, i, k − i) = f (θ0 , i, k − i). By Assumption 1-(ii), this implies θ = θ0 .
Moreover the convergence of WN to W is uniform, since Θ is compact and the
function f is twice continuously differential with respect to θ, its first variable.
W
. For the sake of completeness, we give a brief
This yields the consistency of θ̂N
W
minimizes WN , we have
proof. Since θ0 minimizes W and θ̂N
W
0 ≤ W(θ̂N
) − W(θ0 )
W
W
W
) − WN (θ0 ) + WN (θ0 ) − W(θ0 )
= W(θ̂N
) − WN (θ̂N
) + WN (θ̂N
W
W
) − WN (θ̂N
) + WN (θ0 ) − W(θ0 ) ≤ 2 sup |WN (θ) − W(θ)| .
≤ W(θ̂N
θ∈Θ
Since θ0 is the unique minimizer of W, for > 0, we can find δ such that if θ ∈ Θ
and kθ − θ0 k > , then W(θ) − W(θ0 ) ≥ δ. Thus
P(kθ̂N − θ0 k > ) ≤ P(W(θ̂N ) − W(θ0 ) ≥ δ)
≤ P 2 sup |WN (θ) − W(θ)| ≥ δ → 0 .
θ∈Θ
27
The central limit theorem is a consequence of the consistency and Lemma 9. A
first order Taylor extension of ẆN (θ) at θ0 yields
W
W
0 = ẆN (θ̂N
) = ẆN (θ0 ) + ẄN (θ̃N )(θ̂N
− θ0 ) ,
W
where θ̃N ∈ [θ0 , θ̂N
]. Setting f˙0 (i, k − i) = f˙(θ0 , i, k − i), we have
ẆN (θ0 ) =
n−1 X
k
X
wN (i, k − i)∂2 G(pN (i, k − i), f0 (i, k − i))f˙0 (i, k − i) .
k=0 i=0
2
G be the mixed second derivative of G. Note that
Let ∂12
∂2 G(f0 (i, k − i), f0 (i, k − i)) = 0 .
Thus, by the delta-method (see Dacunha-Castelle and Duflo,√
1986, Theorem 3.3.11)
and since wN converges almost surely to w0 , we obtain that N ẆN (θ0 ) converges
weakly towards
k
n−1 X
X
2
w0 (i, k − i)∂12
G(f0 (i, k − i), f0 (i, k − i))Λ0 (i, k − i)f˙0 (i, k − i) ,
k=0 i=0
where Λ0 (i, k) are independent Gaussian random
√ variables with zero mean and
variance γ0 (i, k) defined in 14. Equivalently, N ẆN (θ0 ) converges weakly to a
Gaussian vector with zero mean and covariance matrix H(θ0 ) defined by
H(θ0 ) =
n−1 X
k
X
2
w02 (i, k − i){∂12
G(f0 (i, k − i), f0 (i, k − i))}2
k=0 i=0
× γ0 (i, k)f˙0 (i, k − i)(f˙0 (i, k − i))0 .
By the law of large numbers, ẄN (θ) converges almost surely to Ẅ(θ) and this convergence is also locally uniform. Thus, ẄN (θ̃N ) converges almost surely to Ẅ(θ0 ).
Using again the fact that ∂2 G(p, p) = 0, we obtain
Ẅ(θ0 ) =
n−1 X
k
X
w0 (i, k − i)∂22 G(f0 (i, k − i), f0 (i, k − i))
k=0 i=0
× f˙0 (i, k − i)(f˙0 (i, k − i))0 .
Denote for brevity g(i, k − i) = w0 (i, k − i)∂22 G(f0 (i, k − i), f0 (i, k − i)). Then, for
any u ∈ Rd , we have
!2
n−1 X
k
d
X
X
uẄ(θ0 )u0 =
g(i, k − i)
us ∂s f (θ0 , i, k − i)
.
(15)
s=1
k=0 i=0
28
By assumption 12, g(i, k − i) > 0 for all 0 ≤ i ≤ k ≤P
n − 1, thus (15) is zero only
if for all k = 0, . . . , n − 1 and i = 0, . . . , k, we have ds=1 us ∂s f (θ0 , i, k) = 0. By
Assumption 1-(iii), this is possible only if us = 0 for all s = 1, . . . , d. Thus Ẅ(θ0 )
is positive definite.
We can now conclude that for large enough N , ẄN (θ̃N ) is invertible and we
can write
√
√
W
N (θ̂N
− θ0 ) = −Ẅ−1
(
θ̃
)
N ẆN (θ0 ) .
N
N
The right hand side converges weakly to the Gaussian distribution with zero mean
and covariance matrix Ẅ−1 (θ0 )H(θ0 )Ẅ−1 (θ0 ).
6.4
Proofs of theorems 2 and 3
Theorems 2 and 3 are a consequence of Theorem 13.
Lemma 14. Assumption 12 holds for the weights wN (i, k − i) = aN (i, k − i) and
−1
wN (i, k − i) = aN (i, k − i)p−1
N (i, k − i)qN (i, k − i), 0 ≤ i ≤ k ≤ n − 1.
Proof. For all 0 ≤ i ≤ k ≤ n − 1, aN (i, k − i) converges almost surely to P(Zk = i)
−1
−1
−1 ¯
and aN (i, k − i)p−1
N (i, k − i)qN (i, k − i) to P(Zk = i)f0 (i, k − i) f0 (i, k − i) .
Moreover Assumption 1-(i) implies that f0 (i, k − i) > 0 and f¯0 (i, k − i) > 0 for all
0 ≤ i ≤ k ≤ n − 1. Using Formula (13), this in turn implies that P(Zk = i) > 0
for all 0 ≤ i ≤ k ≤ n − 1.
Proof of Theorem 2. As mentioned above, the maximum likelihood estimator minimizes the contrast function W written with the function G(p, q) = −p log q −
(1 − p)log(1 − q) and the weights aN (i, k − i). Thus the proof of Theorem 2 consists in checking Assumption 11 and 12 to apply Theorem 13. Lemma 14 implies
that Assumption 12 holds.
The function G considered here satisfies Assumption 11. Indeed, for p, q ∈ (0, 1),
define
K(p, q) = G(p, q) − G(p, p) = p log(p/q) + (1 − p) log((1 − p)/(1 − q) .
Remark that K(p, q) is the Kullback-Leibler distance between the Bernoulli measures with respective success probabilities p and q. Then it is well known that
K(p, q) > 0 except if p = q. Indeed, by Jensen’s inequality,
K(p, q) ≥ − log(pq/p + (1 − p)(1 − q)/(1 − p)) = log 1 = 0 ,
and by strict concavity of the log function, equality holds only if p = q. Moreover,
∂2 G(p, q) = −p/q+(1−p)/(1−q) so ∂2 G(p, p) = 0 and ∂22 G(p, p) = p−1 (1 − p)−1 > 0.
29
Proof of Theorem 3. Again, the proof consists in checking Assumption 11 and 12
to apply Theorem 13. The latter holds by virtue of Lemma 14 and Assumption 11 trivially holds for the function G(p, q) = (p − q)2 . If wN (i, k − i) =
−1
p−1
N (i, k − i)qN (i, k − i)aN (i, k − i), then
H(θ0 ) = 2Ẅ(θ0 ) = 4In (θ0 )
=4
n−1 X
k
X
k=0 i=0
P(Zk = i)
f˙0 (i, k − i)(f˙0 (i, k − i))0 . (16)
f0 (i, k − i)f¯0 (i, k − i)
So that Σn (θ0 ) = Ẅ−1 (θ0 )H(θ0 )Ẅ−1 (θ0 ) = In−1 (θ0 ).
If the weights are chosen as wN (i, k) = aN (i, k), then w0 (i, k − i) = P(Zk = i)
and
H(θ0 ) = 4
Ẅ (θ0 ) = 2
k
n−1 X
X
k=0 i=0
n−1
k
XX
P(Zk = i)f0 (i, k − i)f¯0 (i, k − i)f˙0 (i, k − i)(f˙0 (i, k − i))0 , (17)
P(Zk = i)f˙0 (i, k − i)(f˙0 (i, k − i))0 .
(18)
k=0 i=0
6.5
Proofs of Corollaries 6 and 7
Corollaries 6 and 7 are a consequence of Theorem 13. The assumptions on the
weights wN and on the functions G have been already verified in the previous
section. We have to prove the Assumption 1 on the choice function f defined
in (10). Hypothesis 1-(i) is obvious.
By elementary computations, we have, for 0 ≤ i ≤ k ≤ n − 1,
α
α0
c+i
c0 + i
f (α, c, i, k − i) = f (α0 , c0 , i, k − i) ⇔
=
c+k−i
c0 + k − i
log(c0 + i) − log(c0 + k − i)
α
=
. (19)
⇔
α0
log(c + i) − log(c + k − i)
Plugging the pairs (i, k) = (0, 1) and (i, k) = (0, 2) into (19) yields
log(c0 ) − log(c0 + 1)
log(c0 ) − log(c0 + 2)
=
,
log(c) − log(c + 1)
log(c) − log(c + 2)
or equivalently
log(1 + 1/c0 )
log(1 + 1/c)
=
.
log(1 + 2/c0 )
log(1 + 2/c)
30
(20)
It is easily checked that the function x → log(1 + x)/ log(1/2x) is strictly increasing on (0, ∞). Thus (20) implies that c = c0 . Plugging this equality into (19)
yields α = α0 . This proves Assumption 1-(ii).
We now prove that if n ≥ 2, the vectors {∂α f (θ0 , i, k − i), 0 ≤ i ≤ k ≤ n − 1}
and {∂c f (θ0 , i, k − i), 0 ≤ i ≤ k ≤ n − 1} are linearly independent in Rn(n−1) .
For 0 ≤ i ≤ k ≤ n − 1, we have,
c+i
,
∂α f (α, c, i, k − i) = f (α, c, i, k − i)f (α, c, k − i, i) log
c+k−i
α(k − 2i)
∂c f (α, c, i, k − i) = f (α, c, i, k − i)f (α, c, k − i, i)
.
(c + i)(c + k − i)
Let (u, v) ∈ R2 and assume that for all i, j ≤ n − 1 such that i + j ≤ n − 1, it
holds that
c0 + i
α0 (j − i)
=0.
u log
+v
c0 + j
(c0 + i)(c0 + j)
Replacing (i, j) for instance successively by (0, 1) and (0, 2) yields
c
α0
0
+v
=0,
u log
c0 (c0 + 1)
c0 + 1
c0
2α0
=0.
+v
u log
c0 + 2
c0 (c0 + 2)
If (u, v) 6= (0, 0), this implies
c0 + 2
c0 + 2
c0 + 1
c0 + 1
log
+2
log
=0.
c0
c0
c0
c0
By strict convexity of the function x → x log x on (0, ∞), this is impossible.
Thus u = v = 0 and Assumption 1-(iii) holds.
6.6
Proof of Theorem 8
Proof of Theorem 8, case α0 = 1. In this case the model is Pólya’s urn, and we
have
n−1
X
1
1
4
1
E
+E
−
.
(21)
In (c) =
2c + k
c + Zk
c + k − Zk
2c + k
k=0
The distribution of Zk is given by
k c(c + 1) · · · (c + i − 1) × c(c + 1) · · · (c + k − i − 1)
P(Zk = i) =
.
2c(2c + 1) · · · (2c + k − 1)
i
31
Thus,
X
k
k c(c + 1) · · · (c + i − 1) × c(c + 1) · · · (c + k − i − 1) 1
1
=
.
E
c + Zk
i
2c(2c + 1) · · · (2c + k − 1)
c+i
i=0
For any c > 0, there exists constants C1 < C2 such that, for all integers h ≥ 1,
h
Y
C1 h ≤
(1 + c/i) ≤ C2 hc .
c
i=1
Therefore, there exists a constant C > 0 such that for all k ≥ 1,
−1
k−1
c−2
c−1
O(k ) if c > 1 ,
X i
i
1
= O(k −1 log k) if c = 1 ,
≤ Ck −2
1−
E
c + Zk
k
k
i=1
O(k −c ) if c < 1 .
In all three cases, we obtain that the first series in (21) is summable. By symmetry,
the sum of the second expectations is also finite.
Proof of Theorem 8, case α0 < 1. In this case, we know by Theorem 4 that Zn /n
converges almost surely to 1/2. This implies that f (θ, Zn , n−Zn ) converges almost
surely to 1/2 for all θ. By Cesaro’s Lemma, this implies that n−1 `n (θ) → − log 2
a.s.
Proof of Theorem 8, case α0 > 1. Let Ω1 be the event that color 1 is eventually selected, which happens with probability 1/2 by Theorem 4. Then, on Ω1 , Zn /n → 1
and if k > T∞ , then Xk+1 = 1 and Zk = k − Q∞ . Thus for large enough n, the
log-likelihood on one path becomes
`n (θ) =
T∞
X
Xk+1 log f (θ, Zk , k − Zk ) + (1 − Xk+1 ) log{1 − f (θ, Zk , k − Zk )}
k=0
+
n
X
log f (θ, k − Q∞ , Q∞ ) .
k=T∞ +1
As k → ∞, for any α > 0,
(c + Q∞ )α
log f (θ, k − Q∞ , Q∞ ) = − log 1 +
(c + k − Q∞ )α
32
∼−
(c + Q∞ )α
.
(c + k − Q∞ )α
If α ≤ 1 the series is divergent and thus limn→∞ `n (θ) = −∞. If α > 1 then the
series is convergent and thus, on Ω1 ,
lim `n (θ) =
n→∞
=
+
∞
X
k=0
T∞
X
Xk+1 log f (θ, Zk , k − Zk ) + (1 − Xk+1 ) log{1 − f (θ, Zk , k − Zk )}
Xk+1 log f (θ, Zk , k − Zk ) + (1 − Xk+1 ) log{1 − f (θ, Zk , k − Zk )}
k=0
∞
X
log f (θ, k − Q∞ , Q∞ ) .
T∞ +1
This implies that arg maxθ∈Θ `n (θ) = arg maxθ∈Θ,α>1 `n (θ) and that this argmax
is a random variable which is a function of the whole path, and does not depend
on the true value θ0 .
Acknowledgment We thanks Guy Theraulaz and Hugues Chaté for providing
their material framework and their field expertise to allow the first author to
collect the data of the Argentine ants experiments. These experiments are part
of the project TRACES supported by the CNRS. They were done during two
visits in April and July 2012 of the first author to the Centre de Recherches sur
la Cognition Animale (CRCA, Centre de Recherches sur la Cognition Animale,
UMR 5169, Paul Sabatier University, Toulouse), whose hospitality is gratefully
acknowledged.
References
Arganda, S., S. Nicolis, A. Perochain, C. Péchabadens, G. Latil, and A. Dussutour (2014): “Collective choice in ants: The role of protein and carbohydrates
ratios,” J. Insect Physiol., 69, 19–26.
Aron, S., J.-L. Deneubourg, S. Goss, and J. Pasteels (1990): Functional Selforganisation illustrated by Inter-nest Traffic in Ants : the Case of the Argentine
Ant, Springer-Verlag, chapter Biological Motion, 533–547.
Aron, S., J. Pasteels, and J.-L. Deneubourg (1989): “Trail-laying behaviour during
exploratory recruitment in the argentine ant, iridomyrmex humilis (mayr),”
Biol. of Behav., 14, 207–217.
Beckers, R., J.-L. Deneubourg, and S. Goss (1992): “Trails and u-turns in the
selection of a path by the ant lasius niger,” J. Theor. Biol., 159, 397–415.
33
Beckers, R., J.-L. Deneubourg, and S. Goss (1993): “Modulation of trail laying
in the ant lasius niger (hymenoptera: formicidae) and its role in the collective
selection of a food source,” J. of Insect Behav., 6, 751–759.
Dacunha-Castelle, D. and M. Duflo (1986): Probability and statistics. Vol. II, New
York: Springer-Verlag.
Davis, B. (1990): “Reinforced random walk,” Probab. Theory Related Fields, 84,
203–229.
Deneubourg, J.-L., S. Aron, S. Goss, and J. Pasteels (1990): “The self-organizing
exploratory pattern of the argentine ant,” J. of Insect Behav., 3, 159–168.
Dussutour, A., J.-L. Deneubourg, and V. Fourcassié (2005): “Amplification of
individual preferences in a social context: the case of wall-following in ants,”
Proc. R. Soc. London, Ser. B, 272, 705–714.
Freedman, D. (1965): “Bernard friedman’s urn,” Ann. Math. Stat., 36, 956–970.
Garnier, S., A. Guérécheau, M. Combe, V. Fourcassié, and G. Theraulaz (2009):
“Path selection and foraging efficiency in argentine ant transport networks,”
Behav. Ecol. Sociobiol., 63, 1167–1179.
Jeanson, R., F. Ratnieks, and J.-L. Deneubourg (2003): “Pheromone trail decay
rates on different substrates in the pharaohs ant, monomorium pharaonis,”
Physiol. Entomol., 28, 192–198.
Khanin, K. and R. Khanin (2001): “A probabilistic model for the establishment
of neuron polarity,” J. Math. Biol., 42, 26–40.
Le Goff, L. and O. Raimond (2015): “Vertex reinforced non-backtracking random
walks: an example of path formation,” URL http://arxiv.org/abs/1506.
01239, arXiv:1506.01239.
Nicolis, S. and J.-L. Deneubourg (1999): “Emerging patterns and food recruitment in ants: an analytical study,” J. Theor. Biol., 198, 575–592.
Nicolis, S. and A. Dussutour (2008): “Self-organization, collective decision making
and resource exploitation strategies in social insects,” Eur. Phys. J. B, 65, 379–
385.
Pasteels, J., J.-L. Deneubourg, and S. Goss (1987): “Transmission and amplification of information in a changing environment: The case of insect societies,”
Eds I. Prigogine & M. Sanglier. Gordes, Bruxelles.
34
Pemantle, R. (2007): “A survey of random processes with reinforcement,” Probability Surveys, 4, 1–79.
Pólya, G. (1931): “Sur quelques points de la théorie des probabilités,” Ann.
I.H.P., 1, 117–161.
Tarrès, P. (2011): “Localization of reinforced random walks,” URL http://
arxiv.org/abs/1103.5536, arXiv:1103.5536.
Thienen, W., D. Metzler, D.-H. Choe, and V. Witte (2014): “Pheromone communication in ants: a detailed analysis of concentration-dependent decisions in
three species,” Behav. Ecol. Sociobiol., 68, 1611–1627.
Van Vorhis Key, S. and T. Baker (1982): “Trail-following responses of the argentine ant, iridomyrmex humilis (mayr), to a synthetic trail pheromone component
and analogs,” J. Chem. Ecol., 8, 3–14.
Vittori, K., G. Talbot, J. Gautrais, V. Fourcassié, A. Araujo, and G. Theraulaz
(2006): “Path efficiency of ant foraging trails in an artificial network,” J. Theor.
Biol., 239, 507–515.
Wasserman, L. (2004): All of Statistics: A Concise Course in Statistical Inference,
Springer.
35
| 10math.ST
|
Sound and Automated Synthesis of Digital Stabilizing
Controllers for Continuous Plants
∗
Alessandro Abate1 , Iury Bessa2 , Dario Cattaruzza1 , Lucas Cordeiro1,2 ,
Cristina David1 , Pascal Kesseli1 and Daniel Kroening1
arXiv:1610.04761v3 [cs.SY] 16 Feb 2017
1
University of Oxford, Oxford, United Kingdom
ABSTRACT
Modern control is implemented with digital microcontrollers,
embedded within a dynamical plant that represents physical
components. We present a new algorithm based on counterexample guided inductive synthesis that automates the design
of digital controllers that are correct by construction. The
synthesis result is sound with respect to the complete range
of approximations, including time discretization, quantization effects, and finite-precision arithmetic and its rounding
errors. We have implemented our new algorithm in a tool
called DSSynth, and are able to automatically generate stable
controllers for a set of intricate plant models taken from the
literature within minutes.
Keywords
Digital control synthesis, CEGIS, finite-word-length representation, time sampling, quantization
1.
INTRODUCTION
Modern implementations of embedded control systems
have proliferated with the availability of low-cost devices that
can perform highly non-trivial control tasks, with significant
impact in numerous application areas such as environmental
control and robotics [4, 17]. Correct control is non-trivial,
however. The problem is exacerbated by artifacts specific
to digital control, such as the effects of finite-precision arithmetic, time discretization, and quantization noise introduced
by A/D and D/A conversion. Thus, programming expertise
is a key barrier to broad adoption of correct digital controllers, and requires considerable knowledge outside of the
expertise of many control engineers.
Beyond classical a-posteriori validation in digital control,
there has been plenty of previous work aiming at verifying
a given designed controller, which however broadly lack automation. Recent work has studied the stability of digital
∗
Supported by EPSRC grant EP/J012564/1, ERC project
280053 (CPROVER) and the H2020 FET OPEN SC2 .
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from [email protected].
c 2017 ACM. ISBN .
DOI:
2
Federal University of Amazonas, Manaus, Brazil
controllers considering implementation aspects, i.e., fixedpoint arithmetic and the word length [8]. They exploit
advances in bit-accurate verification of C programs to obtain
a verifier for software-implemented digital control.
By contrast, we leverage a very recent step-change in the
automation and scalability of program synthesis. Program
synthesis engines use a specification as the starting point, and
subsequently generate a sequence of candidate programs from
a given template. The candidate programs are iteratively
refined to eventually satisfy the specification. Program synthesizers implementing Counter-Example Guided Inductive
Synthesis (CEGIS) [34] are now able to generate programs
for highly non-trivial specifications with a very high degree of
automation. Modern synthesis engines combine automated
testing, genetic algorithms, and SMT-based automated reasoning [1, 11].
By combining and applying state-of-the-art synthesis engines we present a tool that automatically generates digital
controllers for a given continuous plant model that are correct by construction. This approach delivers a high degree of
automation, promises to reduce the cost and time of development of digital control dramatically, and requires considerably less expertise than a-posteriori verification. Specifically,
we synthesize stable, software-implemented embedded controllers along with a model of a physical plant. Due to the
complexity of such closed-loop systems, in this work we focus
on linear models with known configurations, and perform
parametric synthesis of stabilizing digital controllers (further closed-loop performance requirements are left to future
work).
Our work addresses challenging aspects of the control synthesis problem. We perform digital control synthesis over a
hybrid model, where the plant exhibits continuous behavior
whereas the controller operates in discrete time and over a
quantized domain. Inspired by a classical approach [4], we
translate the problem into a single digital domain, i.e., we
model a digital equivalent of the continuous plant by evaluating the effects of the quantizers (A/D and D/A converters)
and of time discretization. We further account for uncertainties in the plant model. The resulting closed-loop system is
a program with a loop that operates on bit-vectors encoded
using fixed-point arithmetic with finite word length (FWL).
The three effects of 1. uncertainties, 2. FWL representation
and 3. quantization errors are incorporated into the model,
and are taken into account during the CEGIS-based synthesis
of the control software for the plant.
In summary, this paper makes the following original contributions.
• We automatically generate correct-by-construction digital controllers using an inductive synthesis approach.
Our application of program synthesis is non-trivial and
addresses challenges specific to control systems, such
as the effects of quantizers and FWL. In particular, we
have found that a two-stage verification engine that
continuously refines the precision of the fixed-point representation of the plant yields a speed-up of two orders
of magnitude over a conventional one-stage verification
engine.
• Experimental results show that DSSynth is able to efficiently synthesize stable controllers for a set of intricate
benchmarks taken from the literature: the median runtime for our benchmark set considering the faster engine
is 48 s, i.e., half of the controllers can be synthesized in
less than one minute.
2.
2.1
PRELIMINARIES
Discretization of the Plant
The digital controllers synthesized using the algorithm
we present in this paper are typically used in closed loops
with continuous (physical) plants. Thus, we consider continuous dynamics (the plant) and discrete parts (the digital
controller). In order to obtain an overall model for the synthesis, we discretize the continuous plant and particularly
look at the plant dynamics from the perspective of the digital
controller.
As we only consider transfer function models, and require
a z-domain transfer function G(z) that captures all aspects
of the continuous plant, which is naturally described via
a Laplace-domain transfer function G(s). The continuous
model of the plant must be discretized to obtain the corresponding coefficients of G(z).
Among the discretization methods in the literature [17],
we consider the sample-and-hold processes in complex systems [20]. On the other hand, the ZOH discretization models
the exact effect of sampling and DAC interpolation over the
plant.
Assumption 1. The sample-and-hold effects of the ADC
and the presence of the ZOH of the DAC are synchronized,
namely there is no delay between sampling the plant output
at the ADC and updating the DAC accordingly. The DAC
interpolator is an ideal ZOH.
Lemma 1. [4] Given a synchronized ZOH input and sampleand-hold output on the plant, with a sample time T satisfying
the Nyquist criterion, the discrete pulse transfer function
G(z, T ) is an exact z-domain representation of G(s), and can
be computed using the following formula:
G(s)
G(z, T ) = (1 − z −1 )Z L−1
.
(1)
s
t=kT
In this study, for the sake of brevity, we will use the notation G(z) to represent the pulse transfer function G(z, T ).
Lemma
1 ensures that the poles and zeros match under the
Z L−1 {·}t=kT operations, and it includes the ZOH dynamics in the (1 − z −1 ) term. This is sufficient for stability
studies over G(s) [15], i.e., if there is any unstable pole (in
the complex domain <{s} > 0), the pulse transfer function
in (1) will also present the same number of unstable poles
(|z| > 1) [17].
2.2
Model Imprecision, Finite Word Length
Representation and Quantization Effects
Let C(z) be a digital controller and G(z) be a discrete-time
representation of the plant, given as
C(z) =
β0 + β1 z −1 + ... + βMC z −MC
Cn (z)
=
,
Cd (z)
α0 + α1 z −1 + ... + αNC z −NC
(2)
G(z) =
b0 + b1 z −1 + ... + bMG z −MG
Gn (z)
=
.
Gd (z)
a0 + a1 z −1 + ... + aNG z −NG
(3)
~ and α
where β
~ are vectors containing the controller’s coefficients; similarly, ~b and ~a denote the plant’s coefficients; and
finally N(·) and M(·) indicate the order of the polynomials,
and we require in particular that NG ≥ MG .
Uncertainties in G(z) may appear owing to: 1. uncertainties in G(s) (we denote the uncertain continuous plant by
G (s)+∆ G (s)
Ĝ(s) = Gnd (s)+∆pp Gnd (s) to explicitly encompass the effects of
the uncertainty terms ∆p G(·) (s)) arising from tolerances/imprecision in the original model; 2. errors in the numerical
calculations due to FWL effects (e.g., coefficient truncation
and round-off, which will be denoted as ∆b Gn (s), ∆b Gd (s));
and 3. errors caused by quantization (which we model later
as as external disturbances ν1 and ν2 ). These uncertainties
are parametrically expressed by additive terms, eventually
resulting in an uncertain model Ĝ(z), such that:
Ĝ(z) =
Gn (z) + ∆Gn (z)
,
Gd (z) + ∆Gd (z)
(4)
which will be represented by the following transfer function:
Ĝ(z) =
b̂0 + b̂1 z −1 + ... + b̂MG z −MG
.
â0 + â1 z −1 + ... + âNG z −NG
(5)
Notice that, due to the nature of the methods we use for
the stability check, we require that the parametric errors
in the plant have the same polynomial order as the plant
itself (indeed, all other errors described in this paper fulfill this property). We also remark that, due to its native digital implementation, there are no parametric errors
(∆p Cn (z), ∆p Cd (z)) in the controller. Thus Ĉ(z) ≡ C(z).
We introduce next a notation based on the coefficients
of the polynomial to simplify the presentation. Let P N be
the space of polynomials of order N . Let P ∈ P M,N be a
rational polynomial PPnd , where Pn ∈ P M and Pd ∈ P N . For
a vector of coefficients
~ ∈ RN +M +2 = [n0 n1 . . . nM d0 d1 . . . dN ]T
P
(6)
and an uncertainty vector
~ ∈ RN +M +2 = [∆n0 . . . ∆nM ∆d0 . . . ∆dN ]T
∆P
(7)
we write
~
~ + ∆G,
~ where
Ĝ = G
(8)
~ ∈ RNG +MG +2 = [b0 . . . bM a0 . . . aN ]T ,
G
G
G
~ ∈ RNG +MG +2 = [∆b0 . . . ∆bM ∆a0 . . . ∆aN ]T .
∆G
G
G
In the following we will either manipulate the transfer functions G(z), C(z) directly, or work over their respective coef~ C
~ in vector form.
ficients G,
A typical digital control system with a continuous plant
and a discrete controller is illustrated in Figure 1. The DAC
and ADC converters introduce quantization errors (notice
ν2 (z)
ν1 (z)
DAC
ADC
Controller
R(z)
+
e(z)
Ĉ(z)
-
Plant
U (z)
+
+
U (s)
ZOH
Ŷ (s)
Ĝ(s)
Q2
+
+
Ŷ (z)
Q1
Figure 1: Closed-loop digital control system (cf. Section 2.2 for notation)
that each of them may have a different FWL representation
than the controller), which are modeled as disturbances ν1 (z)
and ν2 (z); G(s) is the continuous-time plant model with
parametric additive uncertainty ∆p Gn (s) and ∆p Gd (s) (as
mentioned above); R(z) is a given reference signal; U (z) is
the control signal; and Ŷ (z) is the output signal affected by
the disturbances and uncertainties in the closed-loop system
The ADC and DAC may be abstracted by transforming the
closed-loop system in Figure 1 into the digital system in
Figure 2, where the effect of ν1 and ν2 in the output Y (z) is
additive noise.
Controller
R(z)
+
e(z)
-
Ĉ(z)
U (z)
ν2 (z)
+
+
Plant
Ĝ(z)
where P n is the space of polynomials of n-th order, P n hI, F i
is the space of polynomials with coefficients in RhI, F i, and
(as a special case) F0 hI, F i(x) returns the element x̃ ∈ RhI, F i
that is closest to the real parameter x.
Similarly, Fn,m hI, F i(·) : P n,m → P n,m hI, F i applies the
same effect to a ratio of polynomials, where P n,m , P n,m hI, F i
are rational polynomial domains.
Thus, the perturbed controller model C̃(z) may be obn (z)
tained from the original model Ĉ(z) = C(z) = C
as
Cd (z)
follows:
ν1 (z)
+
+
Ŷ (z)
Figure 2: Fully digital equivalent to system in Figure 1
In Figure 2, two sources of uncertainty are illustrated:
parametric uncertainties model the errors (which are rep~ and uncertainties for the quantizations
resented by ∆p G),
in the ADC and DAC conversions (ν1 and ν2 ), which are
assumed to be non-deterministic. Recall that we discussed
how the quantization noise is an additive term, which means
it does not enter parametrically in the transfer function. Instead, we later show that the system is stable given these
non-deterministic disturbance.
The uncertain model may be rewritten as a vector of
~
~+
coefficients in the z-domain using equation (8) as Ĝ = G
~ The parametric uncertainties in the plant are assumed
∆p G.
to have the same order as the plant model, since errors of
higher order can move the closed-loop poles by large amounts,
thus preventing any given controller from stabilizing such a
setup. This is a reasonable assumption since most tolerances
do not change the architecture of the plant.
Direct use of controllers in fixed-point representation.
Since the controller is implemented using finite representation, C(z) also suffers disturbances from the FWL effects,
with roundoffs in coefficients that may change closed-loop
poles and zeros position, and consequently affect its stability,
as argued in [8].
Let Ĉ(z) be the digital controller transfer function represented using this FWL with integer size I and fractional
size F . The term I affects the range of the representation
and is set to avoid overflows, while F affects the precision
and the truncation after arithmetic operations. We shall
denote the FWL domain of the coefficients by RhI, F i and
define a function
Fn hI, F i(P ∈ P) : P n → P n hI, F i
~ = F hI, F i(c ),
~ ∧ c̃i ∈ P̃
, P̃ ∈ P n hI, F i : ci ∈ P
0
i
(9)
C̃(z) = FMC ,NC hI, F i(C(z)) =
FMC hI, F i(Ĉn (z))
FNC hI, F i(Ĉd (z))
. (10)
In the case of a digitally synthesized controller (as it is the
case in this work), C̃(z) ≡ Ĉ(z) ≡ C(z) because the synthesis
is performed directly using FWL representation. In other
words, we synthesize a controller that is already in the domain
RhI, F i and has therefore no uncertainties entering because
of FWL representations, that is, ∆b Cn (z) = ∆b Cd (z) = 0.
Fixed-point computation in program synthesis.
The program synthesis engine uses fixed-point arithmetic.
Specifically, we use the domain RhI, F i for the controller’s coefficients and the domain RhIp , Fp i for the plant’s coefficients,
where I and F , as well as Ip and Fp , denote the number of bits
for the integer and fractional parts, respectively, and where
it is practically motivated to consider RhIp , Fp i ⊇ RhI, F i.
Given the use of fixed-point arithmetic, we examine the
discretization effect during these operations. Let C̃(z) and
G̃(z) be transfer functions represented using fixed-point bitvectors.
C̃(z) =
β̃0 + β̃1 z −1 + ... + β̃MC z −MC
,
α̃0 + α̃1 z −1 + ... + α̃NC z −NC
(11)
G̃(z) =
b̃0 + b̃1 z −1 + ... + b̃MG z −MG
.
ã0 + ã1 z −1 + ... + ãNG z −NG
(12)
Recall that since the controller is synthesized in the RhI, F i
domain, C̃(z) ≡ Ĉ(z) ≡ C(z). However, given a real plant
Ĝ(z), we need to introduce G̃(z) = FMG ,NG hIp , Fp i(Ĝ(z)),
where
(b̂0 + ∆b b̂0 ) + ... + (b̂MG + ∆b b̂MG )z −MG
(â0 + ∆b â0 ) + ... + (âNG + ∆b âNG )z −NG
~
~
~ =G
~ + ∆p G
~ + ∆b G,
~
G̃ = Ĝ + ∆b G
(13)
G̃(z) =
where ∆b ci = c̃i − ĉi , and ∆b G represents the plant uncertainty caused by the rounding off effect. We capture the
~ = ∆p G
~ + ∆b G.
~
global uncertainty as ∆G
2.3
Closed-Loop Stability Verification under
Parametric Uncertainties, FWL Representation and Quantization Noise
Sound synthesis of the digital controller requires the consideration of the effect of FWL on the controller, and of
quantization disturbances in the closed-loop system. Let the
quantizer Q1 (ADC) be the source of a white noise ν1 , and
Q2 (DAC) be the source of a white noise ν2 . The following equation models the system in Figure 1, including the
~ and the FWL effects on the
parametric uncertainties ∆G
controller C̃(z):
Ŷ (z) = ν1 (z) + Ĝ(z)C̃(z)R(z) + Ĝ(z)ν2 (z) − Ĝ(z)C̃(z)Ŷ (z).
(14)
The above can be rewritten as follows:
Ŷ (z) = H1 (z)ν1 (z) + H2 (z)ν2 (z) + H3 (z)R(z),
(15)
where
H1 (z) =
1
,
1 + Ĝ(z)C̃(z)
H2 (z) =
Ĝ(z)
,
1 + Ĝ(z)C̃(z)
H3 (z) =
Ĝ(z)C̃(z)
.
1 + Ĝ(z)C̃(z)
Assumption 2. The quantization noises ν1 (from Q1)
and ν2 (from Q2) are uncorrelated white noises and their
amplitudes are always bounded by the half of quantization
step [4], i.e., |ν1 | ≤ q21 and |ν2 | ≤ q22 , where q1 and q2 are
the quantization steps of ADC and DAC, respectively.
A discrete-time dynamical system is said to be BoundedInput and Bounded-Output (BIBO) stable if bounded inputs
necessarily result in bounded outputs. This condition holds
true over an LTI model if and only if every pole of its transfer
function lies inside the unit circle [5]. Analyzing Eq. (15),
the following proposition provides conditions for the BIBO
stability of the system in Figure 1, with regards to the
exogenous signals R(z), ν1 , and ν2 , which are all bounded
(in particular, the bound on the quantization noise is given
by Assumption 2).
Proposition 1. [8, 15] Consider a feedback closed-loop
control system as given in Figure 1 with a FWL implementation of the digital controller C̃(z) = FMC ,NC hI, F i(C(z))
and uncertain discrete model of the plant from (6), (7)
Ĝ(z) =
Ĝn (z)
,
Ĝd (z)
~
~ + ∆p G.
~
Ĝ = G
Then Ĝ(z) is BIBO-stable if and only if:
• the roots of characteristic polynomial S(z) are inside
the open unit circle, where S(z) is:
S(z) = C̃n (z)Ĝn (z) + C̃d (z)Ĝd (z);
(16)
• the direct loop product C̃(z)Ĝ(z) has no pole-zero cancellation on or outside the unit circle.
Proposition 1 provides necessary (and sufficient) conditions for the controller to stabilize the closed-loop system,
~ quanconsidering plant parametric uncertainties (i.e., ∆p G),
tization noises (ν1 and ν2 ) and FWL effects in the control
software. In particular, note that the model for quantization
noise enters as a signal to be stabilized: in practice, if the
quantization noise is bounded, the noise may be disregarded
if the conditions on Proposition 1 are satisfied.
If the verification is performed using FWL arithmetic,
the above equations must use G̃(z) instead of Ĝ(z). The
former will provide sufficient conditions for the latter to be
stabilized.
3.
3.1
AUTOMATED PROGRAM SYNTHESIS
FOR DIGITAL CONTROL
Overview of the Synthesis Process
In order to synthesize closed-loop digital control systems,
we use a program synthesis engine. Our program synthesizer
implements Counter-Example Guided Inductive Synthesis
(CEGIS) [34]. We start by presenting its general architecture
followed by describing the parts specific to closed-loop control
systems. A high-level view of the synthesis process is given
in Figure 3. Steps 1 to 3 are performed by the user and
Steps A to D are automatically performed by our tool for
Digital Systems Synthesis, named DSSynth.
CEGIS-based control synthesis requires a formal verifier
to check whether a candidate controller meets the requirements when combined with the plant. We use the DigitalSystem Verifier (DSVerifier) [19] in the verification module
for DSSynth. It checks the stability of closed-loop control
systems and considers finite-word length (FWL) effects in the
digital controller, and uncertainty parameters in the plant
model (plant intervals) [8].
Given a plant model in ANSI-C syntax as input (Steps 1–3),
DSSynth constructs a non-deterministic model to represent
the plant family, i.e., it addresses plant variations as interval
sets (Step A), and formulates a function (Step B) using implementation details provided in Steps 2 and 3 to calculate
the controller parameters to be synthesized (Step C). Note
that DSSynth synthesizes the controller for the desired numerical representation and realization form. Finally, DSSynth
builds an intermediate ANSI-C code for the digital system
implementation, which is used as input for the CEGIS engine
(Step D).
This intermediate ANSI-C code model contains a specification φ for the property of interest (i.e., robust stability)
and is passed to the Counterexample-Guided Inductive Synthesis (CEGIS) module of CBMC [9], where the controller
is marked as the input variable to synthesize. CEGIS employs an iterative, counterexample-guided refinement process,
which is explained in detail in Section 3.2. CEGIS reports
a successful synthesis result if it generates a controller that
is safe with respect to φ. In particular, the ANSI-C code
model guarantees that a synthesized solution is complete
and sound with respect to the stability property φ, since
it does not depend on system inputs and outputs. In the
case of stability, the specification φ consists of a number of
assumptions on the polynomial coefficients, following Jury’s
Criteria, as well as the restrictions on the representation of
these coefficients as discussed in detail in Section 3.3.
3.2
Architecture of the Program Synthesizer
The input specification provided to the program synthesizer
~ .∀~
~ ) where P
~ ranges over functions,
is of the form ∃P
x.σ(~
x, P
User
Step 1
Determine plant
model and intervals
DSSynth
Step 2
Define controller
numerical representation
Step B
Formulate a FWL
effect function
Step A
Construct a nondeterministic plant model
Step 3
Define controller
realization form
Step C
Compute FWL
controller model
ANSI-C input
file
Intermediate
ANSI-C code
Step D
Synthesise digital
controller
Figure 3: Overview of the synthesis process
~
x ranges over ground terms and σ is a quantifier-free formula.
We interpret the ground terms over some finite domain D.
The design of our synthesizer is given in Figure 4 and
consists of two phases, Synthesize and Verify, which interact via a finite set of test vectors inputs that is updated
incrementally. Given the aforementioned specification σ,
~
the synth procedure tries to find an existential witness P
~
satisfying the specification σ(~
x, P ) for all ~
x in inputs (as
opposed to all ~x ∈ D). If synthesize succeeds in finding
~ , this witness is a candidate solution to the full
a witness P
synthesis formula. We pass this candidate solution to verify,
~ satisfies the
which checks whether it is a full solution (i.e., P
~
specification σ(~x, P ) for all ~
x ∈ D). If this is the case, then
the algorithm terminates. Otherwise, additional information
is provided to the synthesize phase in the form of a new
counterexample that is added to the inputs set and the loop
iterates again (the second feedback signal “Increase Precision” provided by the Verify phase in Figure 4 is specific to
control synthesis and will be described in the next section).
Each iteration of the loop adds a new input to the finite
set inputs that is used for synthesis. Given that the full set
of inputs D is finite, this means that the refinement loop can
only iterate a potentially very large, but finite number of
times.
3.3
Synthesis for Control
Formal specification of the stability property.
Next, we describe the specific property that we pass to
the program synthesizer as the specification σ. There are
a number of algorithms in our verification engine that can
be used for stability analysis [7, 8]. Here we choose Jury’s
criterion [4] in view of its efficiency and ease of integration
within DSSynth: we employ this method to check the stability
in the z-domain for the characteristic polynomial S(z) defined
in (16). We consider the following form for S(z):
S(z) = a0 z N + a1 z N −1 + · · · + aN −1 z + aN = 0, a0 6= 0.
Next, the following matrix M = [mij ](2N −2)×N is built
from S(z) coefficients:
M =
V (0)
V (1)
..
.
,
V (N −2)
(k)
where V (k) = [vij ]2×N such that:
(0)
vij =
aj−1 ,
if i = 1
0
v(1)(N
−j+1) , if i = 2
(k)
vij
0,
if j > n − k
(k−1)
(k−1)
(k−1) v11
−
v
.
v
,
if
j ≤ n − k and i = 1
=
(k−1)
2j
1j
v21
vk
if j ≤ n − k and i = 2
(1)(N −j+1) ,
and where k ∈ Z is such that 0 < k < N −2. We have that [4]
S(z) is the characteristic polynomial of a stable system if
and only if the following four conditions hold: R1 : S(1) > 0;
R2 : (1)N S(1) > 0; R3 : |a0 | < aN ; R4 : m11 > 0 ∧ m31 > 0 ∧
m51 > 0 ∧ . . . ∧ m(2N −3)(1) > 0. The stability property
is then encoded by a constraint of the form: φstability ≡
(R1 ∧ R2 ∧ R3 ∧ R4 ).
The synthesis problem.
The synthesis problem we are trying to solve is the following: find a digital controller C̃(z) that makes the closed-loop
system stable for all possible uncertainties G̃(z) (13). When
mapping back to the notation used for describing the general
architecture of the program synthesizer, the controller C̃(z)
denotes P and G̃(z) represents x.
As mentioned above, we compute the coefficients for C̃(z)
in the domain RhI, F i, and those for G̃(z) in the domain
RhIp , Fp i. While the controller’s precision hI, F i is given, we
can vary hIp , Fp i such that RhIp , Fp i ⊇ RhI, F i. As the cost of
SAT solving increases with in the size of the problem instance,
our algorithm tries to solve the problem first for small Ip ,
Fp , iteratively increasing the precision if it is insufficient.
3.4
The Synthesize and Verify phases
The synthesize phase uses BMC to compute a solution
C̃(z). There are two alternatives for the verify phase. The
first approach uses interval arithmetic [28] to represent the
coefficients [ci − ∆p ci − ∆b ci , ci + ∆p ci − ∆b ci + (2−Fp )]
and rounds outwards. This0 allows us to simultaneously
evaluate the full collection of plants Ĝ(s) (i.e., all concrete
plants G(s) in the range G(s) ± ∆p G(s)) plus the effects of
numeric calculations. Synthesized controllers are stable for
all plants in the family. Preliminary experiments show that
a synthesis approach using this verification engine has poor
performance and we therefore designed a second approach.
Our experimental results in Section 4 show that the speedup
yielded by the second approach is in most cases of at least
two orders of magnitude.
The second approach is illustrated in Figure 4 and uses
a two-stage verification approach: the first stage performs
potentially unsound fixed-point operations assuming a plant
precision hIp , Fp i, and the second stage restores soundness
by validating these operations using interval arithmetic on
the synthesized controller. In more detail, in the first stage,
denoted by Uncertainty in Figure 4, assuming a precision
hIp , Fp i we check whether the system is unstable for the current candidate solution, i.e., if ¬φstability is satisfiable for S(z).
If this is the case, then we obtain a counterexample G̃(z),
Increase Precision
Verify
Candidate solution
Synthesize
Uncertainty
Counterexample
Candidate
UNSAT/
Inputs
P
model
UNSAT/
candidate
Closed-loop System Search
BMC-based Verifier
Precision
Done
Candidate
P
True/
False
Fixed-point Arithmetic Verifier
Figure 4: Counterexample-Guided Inductive Synthesis of Closed-loop Systems (Step D)
which makes the closed-loop system unstable. This uncertainty is added to the set inputs such that, in the subsequent
synthesize phase, we obtain a candidate solution consisting
of a controller C(z), which makes the closed-loop system
stable for all the uncertainties accumulated in inputs.
If the Uncertainty verification stage concludes that the
system is stable for the current candidate solution, then we
pass this solution to the second verification stage, Precision,
which checks the propagation of the error in the fixed-point
calculations using a Fixed-point Arithmetic Verifier based
on interval arithmetic.
If the precision verification returns false, then we increase
the precision of hIp , Fp i and re-start the synthesize phase
with an empty inputs set. Otherwise, we found a full sound
solution for our synthesis problem and we are done.
In the rest of the paper, we will refer to the two approaches
for the verify phase as one-stage and two-stage, respectively.
3.5
Soundness
The synthesise phase generates potentially unsound candidate solutions. The soundness of the model is ensured by
the verify phase. If a candidate solution passes verification,
it is necessarily sound.
The verify phase has two stages. The first stage ensures
that no counterexample plant with an unstable closed loop
exists over finite-precision arithmetic. Since the actual plant
uses reals, we need to ensure we do not miss a counterexample because of rounding errors. For this reason, the second
verification stage uses an overapproximation with interval
arithmetic with outward rounding. Thus, the first verification
stage underapproximates and is used to generate counterexamples, and the second stage overapproximates and provides
proof that no counterexample exists.
3.6
Illustrative Example
We illustrate our approach with a classical cruise control
example from the literature [5]. It highlights the challenges
that arise when using finite-precision arithmetic in digital
control. We are given a discrete plant model (with a time
step of 0.2 s), represented by the following z-expression:
G(z) =
0.0264
.
z − 0.9998
(17)
Using an optimization tool, the authors of [36] have designed a high-performance controller for this plant, which is
characterized by the following z-domain transfer function:
C(z) =
2.72z 2 − 4.153z + 1.896
.
z 2 − 1.844z + 0.8496
(18)
The authors of [36] claim that the controller C(z) in (18)
stabilizes the closed-loop system for the discrete plant model
G(z) in (17). However, if the effects of finite-precision arithmetic are considered, then this closed-loop system becomes
unstable. For instance, an implementation of C(z) using
Rh4, 16i fixed-point numbers (i.e., 4 bits for the integer part
and 16 bits for the fractional part) can be modeled as:
2
−4.1529998779296875z+1.89599609375
C̃(z):= 2.7199859619140625z
.
z 2 −1.843994140625z+0.8495941162109375
(19)
The resulting system, where C̃(z) and G(z) are in the forward
path, is unstable. Notice that this is disregarding further
approximation effects on the plant caused by quantization in
the verifier (i.e., G̃(z)). Figure 5a gives the Bode diagram for
the digital controller represented in (18): as the phase margin
is negative, the controller is unstable when considering the
FWL effects.
3.7
Program Synthesis for the Example
We now demonstrate how our approach solves the synthesis problem for the example given in the previous section.
Assuming a precision of Ip = 16, Fp = 24, we start with
an a-priori candidate solution with all coefficients zero (the
controller performs FWL arithmetic, hence we use C̃(z)):
C̃(z) =
0z 2 +0z+0
.
0z 2 +0z+0
In the first verify stage, the uncertainty check finds the
following counterexample:
G̃(z) =
0.026506
.
1.000610z + 1.002838
We add this counterexample to inputs and initiate the synthesize phase, where we obtain the following candidate
solution:
C̃(z) =
12.402664z 2 −11.439667z+0.596756
.
4.003906z 2 −0.287949z+0.015625
This time, the uncertainty check does not find any counterexample and we pass the current candidate solution to
the precision verification stage. We obtain the result false,
meaning that the current precision is insufficient. Consequently, we increase our precision to Ip = 20, Fp = 28.
Since the previous counterexamples were obtained at lower
precision, we remove them from the set of counterexamples. Back in the synthesize phase, we re-start the process
with a candidate solution with all coefficients 0, as above.
Next, the uncertainty verification stage provides the first
counterexample at higher precision:
G̃(z) =
0.026314
.
0.999024z−1.004785
Bode Diagram
Gm = 16.7 dB (at 2.12 rad/s) , Pm = −57.2 deg (at 1.29 rad/s)
Magnitude (dB)
10
0
−10
−20
Phase (deg)
−30
−90
−180
−270
−360
−450
−1
10
0
10
1
10
2
10
Frequency (rad/s)
(a) Original controller [36]
Bode Diagram
Gm = 17.8 dB (at 15.7 rad/s) , Pm = Inf
Magnitude (dB)
0
−5
−10
−15
−20
Phase (deg)
−25
0
−45
−90
−135
−180
−225
−1
10
0
10
1
10
2
10
Frequency (rad/s)
(b) Controller synthesized by DSSynth
Figure 5: Bode diagram for original controller in [36] and
for newly synthesized closed-loop system
In the synthesize phase, we get a new candidate solution
that eliminates the new, higher precision counterexample:
C̃(z) =
11.035202z 2 +5.846100z+4.901855
.
1.097901z 2 +0.063110z+0.128357
This candidate solution is validated as the final solution
by both stages uncertainty and precision in the verify
phase. Figure 5 compares the Bode diagram using the digital
controller represented by Eq. (18) from [36] (Figure 5a) and
the final candidate solution from our synthesizer (Figure 5b).
The DSSynth final solution is stable since it presents an
infinite phase margin and a gain margin of 17.8 dB.
Figure 6 illustrates the step responses of the closed-loop
system with the original controller represented by Eq. (18)
(Figure 6a), the first (Figure 6b) and final (Figure 6c) candidate solutions provided by DSSynth. The step response
in Figure 6a confirms the stability loss if we consider FWL
effects. Figure 6b shows that the first candidate controller is
able to stabilize the closed-loop system without uncertainties,
but it is rejected during the precision phase by DSSynth
since this solution is not sound. Finally, Figure 6c shows a
stable behavior for the final (sound) solution, which presents
a lower settling time (hence the digitization effects).
4.
EXPERIMENTAL EVALUATION
4.1
Description of the Benchmarks
The first set of benchmarks uses the discrete model G1 of
a cruise control system for a car, and accounts for rolling
friction, aerodynamic drag, and the gravitational disturbance
force [5]. The second set of benchmarks considers the discrete
model G2 of a simple spring-mass damper plant [36]. A third
set of benchmarks uses the discrete model G3 for satellite
attitude dynamics [17], which require attitude control for
orientation of antennas and sensors w.r.t. Earth. The fourth
set of benchmarks presents an alternative discrete model G4
of a cruise control system [36]. The fifth and sixth set of
benchmarks describe the discrete model of a DC servo motor
velocity dynamics [27, 35]. The seventh set of benchmarks
contains a well-studied discrete non-minimal phase model
G7 . Non-minimal phase models cause additional difficulties
for the design of stable controllers [12]. The eighth set of
benchmarks describes the discrete model G8 for the Helicopter Longitudinal Motion, which provides the longitudinal
motion dynamics of a helicopter [17]. The ninth set of benchmarks contains the discrete model G9 for the known Inverted
Pendulum, which describes a pendulum dynamics with its
center of mass above its pivot point [17]. The tenth set of
benchmarks contains the Magnetic Suspension discrete model
G10 , which describes the dynamics of a mass that levitates
with support only of a magnetic field [17]. The eleventh set
of benchmarks contains the Computer Tape Driver discrete
model G11 , which describes a system to read and write data
on a storage device [17]. The last set of benchmarks considers
a discrete model G12 that is typically used for evaluating
stability margins and controller fragility [23, 24].
Additional benchmarks were created for the Cruise Control
System, Spring-mass damper, and Satellite considering parametric additive in the nominal plant model (represented by
~ in Eq. (13)). The uncertainties are deviations bounded
∆p G
to a maximum magnitude of 0.5 in each coefficient. These
uncertain models are respectively represented by G1b , G2b ,
G3b and G3d .
All experiments have been conducted on a 12-core 2.40 GHz
Intel Xeon E5-2440 with 96 GB of RAM and Linux OS. All
times given are wall clock times in seconds, as measured
by the UNIX date command. For the two-stage verification
engine in Figure 4 we have applied a timeout of 8 hours per
benchmark, whereas 24 hours have been set for the approach
using a one-stage engine.
4.2
Objectives
Using the closed-loop control system benchmarks given in
Section 4.1, our experimental evaluation aims to answer two
research questions:
RQ1 (performance) does the CEGIS approach generate a
FWL digital controller in a reasonable amount of time?
RQ2 (sanity check) are the synthesized controllers sound
and can their stability be confirmed outside of our
model?
4.3
Results
We give the run-times required to synthesize a stable
controller for each benchmark in Table 1. Here, Plant is
the discrete or continuous plant model, Benchmark is the
name of the employed benchmark, I and F represent the
Step Response
Step Response
28
1.5
Step Response
x 10
1.4
1
1.2
0.9
1
0.8
1
0.5
0
Amplitude
Amplitude
Amplitude
0.7
0.6
0.5
0.8
0.6
0.4
0.4
0.3
−0.5
0.2
0.2
0.1
−1
0
50
100
150
200
Time (seconds)
250
(a) Original controller
300
350
0
0
0
50
100
150
0
0.5
Time (seconds)
(b) First solution by DSSynth
1
1.5
2
2.5
Time (seconds)
3
3.5
4
(c) Final solution by DSSynth
Figure 6: Step responses for original [36] closed-loop system with FWL effects and for each synthesize iteration of DSSynth
number of integer and fractional bits of the stable controller,
respectively, while the two right columns display the total
time (in seconds) required to synthesize a stable controller
for the given plant.
For the majority of the benchmarks, the conjecture explained in Section 3.3 holds and the two-stage verification
engine is able to find a stable solution in less than one minute
for half of the benchmarks. This is possible if the inductive
solutions need to be refined with few counterexamples and
increments of the fixed-point precision. However, the benchmark SatelliteB2 with uncertainty (G3b ) has required too
many counterexamples to refine its solution. For this particular case, the one-stage engine is able to complement the
two-stage approach and synthesizes a solution. It is important to reiterate that the one-stage verification engine does
not take advantage of the inductive conjecture inherent to
CEGIS, but instead fully explores the counterexample space
in a single SAT instance. As expected, this approach is significantly slower on average and is only useful for benchmarks
where the CEGIS approach requires too many refinement
iterations such that exploring all counterexamples in a single
SAT instance performs better. Our results suggest an average
performance difference of at least two orders of magnitude,
leading to the one-stage engine timing out on the majority
of our benchmarks. Table 1 lists the results for both engines,
where in 16 out of 23 benchmarks, the two-stage engine is
faster.
The presence of uncertainty in some particular benchmarks
(2, 4, 6, and 8) leads to harder verification conditions to be
checked by the verify phase, which impacts the overall
synthesis time. However, considering the faster engine for
each benchmark (marked in bold in Table 1), the median
run-time is 48 s, implying that DSSynth can synthesize half of
the controllers in less than one minute. Overall, the average
fastest synthesis time considering both engines is approximately 42 minutes. We consider these times short enough
to be of practical use to control engineers, and thus affirm
RQ1. We further observe that the two-stage verification
engine is able to synthesize stable controllers for 19 out of
the 23 benchmarks, and can be complemented using the onestage engine, which is faster for two benchmarks where the
inductive conjectures fail. Both verification engines together
enable controller synthesis for 20 out of 23 benchmarks. For
the remaining benchmarks our approach failed to synthesize a stable controller within the time limits. This can be
addressed by either increasing either the time limit or the
fixed-point word widths considered, or by using floating-point
arithmetic instead. The synthesized controllers have been
#
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
Plant
G1a
G1b
G2a
G2b
G3a
G3b
G3c
G3d
G4
G5
G6
G7
G8
G9
G10
G11
G12a
G12a
G12a
G12b
G12b
G12b
G12c
Benchmark
CruiseControl02
CruiseControl02†
SpgMsDamper
SpgMsDamper†
SatelliteB2
SatelliteB2†
SatelliteC2
SatelliteC2†
Cruise
DCMotor
DCServomotor
Doyleetal
Helicopter
Pendulum
Suspension
Tapedriver
a ST1 IMPL1
a ST1 IMPL2
a ST1 IMPL3
a ST2 IMPL1
a ST2 IMPL2
a ST2 IMPL3
a ST3 IMPL1
I
4
4
15
15
3
3
3
3
3
3
4
4
3
3
3
3
16
16
16
16
16
16
16
F
16
16
16
16
7
7
5
5
7
7
11
11
7
7
7
7
4
8
12
4
8
12
4
2-stage
12 s
14600 s
52 s
7
36 s
7
3s
50 s
1s
1s
46 s
8769 s
44 s
1s
1s
1s
11748 s
351 s
8772 s
1128 s
7
15183 s
7
1-stage
67 s
52 s
318 s
7
7
4111 s
205 s
1315 s
1s
10 s
7
7
7
14826 s
5s
1s
7
7
7
7
7
7
7
Table 1: DSSynth results (7 = time-out, † = uncertainty)
confirmed to be stable outside of our model representation
using MATLAB, positively answering RQ2. A link to the
full experimental environment, including scripts to reproduce
the results, all benchmarks and the DSSynth tool, is provided
in the footnote.1
4.4
Threats to Validity
We have reported a favorable assessment of DSSynth over
a diverse set of real-world benchmarks. Nevertheless, this
set of benchmarks is limited within the scope of this paper
and DSSynth’s performance needs to be assessed on a larger
benchmark set in future work.
Furthermore, our approach to select suitable FWL word
widths to model plant behavior employs a heuristic based
on user-provided controller word-width specifications. Given
the encouraging results of our benchmarks, this heuristic
appears to be strong enough for the current benchmark set,
but this may not generalize. Further experiments towards
determining suitable plant FWL configurations may thus be
necessary in future work.
Finally, the experimental results obtained using DSSynth
for stability properties may not generalize to other properties.
1
http://www.cprover.org/DSSynth/experiment.tar.gz
CBMC (SHA-1 hash) version:
7a6cec1dd0eb8843559591105235f1f2c4678801
The inductive nature of the two-stage back-end of DSSynth
increases performance significantly compared to the onestage back-end, but this performance benefit introduced
by CEGIS inductive generalizations may not be observed
for other controller properties. Additional experiments are
necessary to confirm that the performance of our inductive
synthesis approach can be leveraged in those scenarios.
5.
RELATED WORK
Robust Synthesis of Linear Systems.
The problem of parametric control synthesis based on stability measures for continuous Linear Time Invariant (LTI)
Single Input-Single Output (SISO) systems has been researched for several decades. On a theoretical level it is
a solved problem [37], for which researchers continuously
seek better results for a number of aspects in addition to
stability. A vast range of pole placement techniques such as
Moore’s algorithm for eigenstructure assignment [25] or the
more recent Linear Quadratic Regulator (LQR) [6] have been
used with increasing degrees of success. The latter approach
highlights the importance of conserving energy during the
control process, which results in lower running costs. Since
real systems are subject to tolerance and noise as well as
the need for economy, more recent studies focus on the problem of achieving robust stability with minimum gain [33,
26]. However, when applied with the aim of synthesizing a
digital controller, many of these techniques lack the ability
to produce sound or stable results because they disregard
the effects of quantization and rounding. Recent papers on
implementations/synthesis of LTI digital controllers [10, 18]
focus on time discretization, failing to account for these errorinducing effects and can result in digital systems that are
unstable even though they have been proven to be robustly
stable in a continuous space.
Formal Verification of Linear Digital Controllers.
Various effects of discretizing dynamics, including delayed
response [13] and Finite Word Length (FWL) semantics [3]
have been studied, with the goal to either verify [7] or to
optimize [29] given implementations.
There are two different problems that arise from FWL
semantics. The first is the error in the dynamics caused
by the inability to represent the exact state of the physical
system while the second relates to rounding errors during
computation. In [16], a stability measure based on the error
of the digital dynamics ensures that the deviation introduced
by FWL does not make the digital system unstable. A recent approach [38] uses the µ-calculus to directly model the
digital controller so that the selected parameters are stable
by design. Most work in verification focuses on finding a
correct variant of a known controller, looking for optimal
parameter representations using FWL, but ignore the effects
of rounding errors due to issues of mathematical tractability.
The analyses in [32, 36] rely on an invariant computation
on the discrete system dynamics using Semi-Definite Programming (SDP). While the former uses BIBO properties to
determine stability, the latter uses Lyapunov-based quadratic
invariants. In both cases, the SDP solver uses floating-point
arithmetic and soundness is checked by bounding the error.
An alternative approach is taken by [30], where the verification of existing code is performed against a known model
by extracting an LTI model of the code through symbolic
execution. In order to account for rounding errors, an upper
bound is introduced in the verification phase. If the error of
the implementation is lower than this tolerance level, then
the verification is successful.
Robust Synthesis of FWL Digital Controllers.
There is no technique in the existing literature for automatic synthesis of fixed-point digital controllers that considers FWL effects.
Other tools such as [14] are aimed at robust stability
problems, but they fail to take the FWL effects into account.
In order to provide a correct-by-design digital controller, [2]
requires a user-defined finite-state abstraction to synthesize
a digital controller based on high-level specifications. While
this approach overcomes the challenges presented by the
FWL problem, it still requires error-prone user intervention.
A different solution that uses FWL as the starting point is
an approach that synthesizes word lengths for known control
problems [22]; however, this provides neither an optimal
result nor a comprehensive solution for the problem.
The CEGIS Architecture.
Program synthesis is the problem of computing correctby-design programs from high-level specifications, and algorithms for this problem have made substantial progress in
recent years. One such approach [21] inductively synthesizes
invariants to generate the desired programs.
Program synthesizers are an ideal fit for synthesis of parametric controllers since the semantics of programs capture
effects such as FWL precisely. In [31], the authors use
CEGIS for the synthesis of switching controllers for stabilizing continuous-time plants with polynomial dynamics. The
work extends to its application on affine systems, finding its
major challenge in the hardness of solving linear arithmetic
with the state-of-the-art SMT solvers. Since this approach
uses switching states instead of linear dynamics in the digital
controller, it entirely circumvents the FWL problem. It is
also not suitable for the kind of control we seek to synthesize.
We require a combination of a synthesis engine with a control
verification tool that addresses the challenges presented here
in the form of FWL effects and stability measures for LTI
SISO controllers. We take the former from [11] and the latter
from [7] while enhancing the procedure by evaluating the
quantization effects of the Hardware interfaces (ADC/DAC)
to obtain an accurate discrete-time FWL representation of
the continuous dynamics.
6.
CONCLUSIONS
We have presented a method for synthesizing stable controllers and an implementation in a tool called DSSynth. The
novelty in our approach is that it is fully automated and algorithmically and numerically sound. In particular, DSSynth
marks the first use of the CEGIS that handles plants with
uncertain models and FWL effects over the digital controller.
Implementing this architecture requires transforming the
traditional CEGIS refinement loop into a two-stage engine:
here, the first stage performs fast, but potentially unsound
fixed-point operations, whereas the second stage restores
soundness by validating the operations performed by the first
stage using interval arithmetic. Our experimental results
show that DSSynth is able to synthesize stable controllers
for most benchmarks within a reasonable amount of time
fully automatically. Future work will be the extension of this
CEGIS-based approach to further classes of systems, including those with state space. We will also consider performance
requirements while synthesizing the digital controller.
7.
[19]
REFERENCES
[1] R. Alur, D. Fisman, R. Singh, and A. Solar-Lezama.
SyGuS-Comp 2016: Results and analysis. In Workshop
on Synthesis, volume 229 of EPTCS, 2016.
[2] R. Alur, S. Moarref, and U. Topcu. Compositional
synthesis with parametric reactive controllers. In
HSCC. ACM, 2016.
[3] A. Anta, R. Majumdar, I. Saha, and P. Tabuada.
Automatic verification of control system
implementations. In Embedded Software (EMSOFT),
pages 9–18, 2010.
[4] K. Åström and B. Wittenmark. Computer-controlled
systems: theory and design. 1997.
[5] K. J. Astrom and R. M. Murray. Feedback Systems: An
Introduction for Scientists and Engineers. 2008.
[6] A. Bemporad, M. Morari, V. Dua, and E. N.
Pistikopoulos. The explicit linear quadratic regulator
for constrained systems. Automatica, 38(1), 2002.
[7] I. Bessa, H. Ismail, L. Cordeiro, and J. Filho.
Verification of fixed-point digital controllers using
direct and delta forms realizations. Design Autom. for
Emb. Sys., 20(2), 2016.
[8] I. Bessa, H. Ismail, R. Palhares, L. Cordeiro, and
J. E. C. Filho. Formal non-fragile stability verification
of digital control systems with uncertainty. IEEE
Transactions on Computers, 66(3):545–552, 2017.
[9] E. M. Clarke, D. Kroening, and F. Lerda. A tool for
checking ANSI-C programs. In TACAS, volume 2988,
2004.
[10] S. Das, I. Pan, K. Halder, S. Das, and A. Gupta. LQR
based improved discrete PID controller design via
optimum selection of weighting matrices using
fractional order integral performance index. Applied
Mathematical Modelling, 37(6), 2013.
[11] C. David, D. Kroening, and M. Lewis. Using program
synthesis for program analysis. In LPAR, LNCS, 2015.
[12] J. C. Doyle, B. A. Francis, and A. R. Tannenbaum.
Feedback Control Theory. 1991.
[13] P. S. Duggirala and M. Viswanathan. Analyzing real
time linear control systems using software verification.
In IEEE Real-Time Systems Symposium, Dec 2015.
[14] C. Economakos, G. Economakos, M. Skarpetis, and
M. Tzamtzi. Automated synthesis of an FPGA-based
controller for vehicle lateral control. In MATEC Web of
Conferences, volume 41, 2016.
[15] S. Fadali and A. Visioli. Digital Control Engineering:
Analysis and Design, volume 303 of Electronics &
Electrical. 2009.
[16] I. J. Fialho and T. T. Georgiou. On stability and
performance of sampled-data systems subject to
wordlength constraint. IEEE Trans. on Automatic
Control, 39(12), 1994.
[17] G. Franklin, D. Powell, and A. Emami-Naeini. Feedback
Control of Dynamic Systems. 7th edition, 2015.
[18] S. Ghosh, R. K. Barai, S. Bhattarcharya,
P. Bhattacharyya, S. Rudra, A. Dutta, and R. Pyne.
An FPGA based implementation of a flexible digital
PID controller for a motion control system. In
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
Computer Communication and Informatics (ICCCI).
IEEE, 2013.
H. Ismail, I. Bessa, L. C. Cordeiro, E. B. de Lima Filho,
and J. E. C. Filho. DSVerifier: A bounded model
checking tool for digital systems. In SPIN, volume 9232,
2015.
R. Istepanian and J. F. Whidborne. Digital controller
implementation and fragility: A modern perspective.
2012.
S. Itzhaky, S. Gulwani, N. Immerman, and M. Sagiv.
A simple inductive synthesis methodology and its
applications. In ACM Sigplan Notices, volume 45.
ACM, 2010.
S. Jha and S. A. Seshia. SWATI: Synthesizing
wordlengths automatically using testing and induction.
arXiv preprint arXiv:1302.1920, 2013.
L. Keel and S. Bhattacharyya. Robust, fragile, or
optimal? IEEE Trans. on Automatic Control, 42(8),
1997.
L. Keel and S. Bhattacharyya. Stability margins and
digital implementation of controllers. In Proc.
American Control Conference, volume 5, 1998.
G. Klein and B. Moore. Eigenvalue-generalized
eigenvector assignment with state feedback. IEEE
Trans. on Automatic Control, 22(1), 1977.
U. Konigorski. Pole placement by parametric output
feedback. Systems & Control Letters, 61(2), 2012.
Y. Li, K. Ang, G. Chong, W. Feng, K. Tan, and
H. Kashiwagi. CAutoCSD–evolutionary search and
optimisation enabled computer automated control
system design. Int J Automat Comput, 1(1), 2004.
R. E. Moore. Interval analysis, volume 4. 1966.
A. K. Oudjida, N. Chaillet, A. Liacha, M. L.
Berrandjia, and M. Hamerlain. Design of high-speed
and low-power finite-word-length PID controllers.
Control Theory and Technology, 12(1), 2014.
J. Park, M. Pajic, I. Lee, and O. Sokolsky. Scalable
verification of linear controller software. In TACAS.
Springer, 2016.
H. Ravanbakhsh and S. Sankaranarayanan.
Counter-example guided synthesis of control Lyapunov
functions for switched systems. In Conference on
Decision and Control, CDC, 2015.
P. Roux, R. Jobredeaux, and P. Garoche. Closed loop
analysis of control command software. In HSCC, 2015.
R. Schmid, L. Ntogramatzidis, T. Nguyen, and
A. Pandey. A unified method for optimal arbitrary pole
placement. Automatica, 50(8), 2014.
A. Solar-Lezama. Program sketching. STTT, 15(5-6),
2013.
K. Tan and Y. Li. Performance-based control system
design automation via evolutionary computing.
Engineering Applications of Artificial Intelligence,
14(4), 2001.
T. E. Wang, P. Garoche, P. Roux, R. Jobredeaux, and
E. Feron. Formal analysis of robustness at model and
code level. In HSCC, 2016.
W. Wonham. On pole assignment in multi-input
controllable linear systems. IEEE Trans. on Automatic
Control, 12(6), 1967.
J. Wu, G. Li, S. Chen, and J. Chu. Robust finite word
length controller design. Automatica, 45(12), 2009.
| 3cs.SY
|
On Generalization and Regularization
in Deep Learning
An Introduction for Data Scientists
arXiv:1704.01312v2 [stat.ML] 6 Apr 2017
Pirmin Lemberger
[email protected]
Weave Business Technology
37 rue du Rocher, 75008 Paris
weave.eu
April 10, 2017
Abstract
Why do large neural network generalize so well on complex tasks such
as image classification or speech recognition? What exactly is the role regularization for them? These are arguably among the most important open
questions in machine learning today. In a recent and thought provoking
paper [1] several authors performed a number of numerical experiments
that hint at the need for novel theoretical concepts to account for this
phenomenon. The paper stirred quit a lot of excitement among the machine learning community but at the same time it created some confusion
as discussions in [2] testifies. The aim of this pedagogical paper is to make
this debate accessible to a wider audience of data scientists without advanced theoretical knowledge in statistical learning. The focus here is on
explicit mathematical definitions and on a discussion of relevant concepts,
not on proofs for which we provide references.
1
Motivation
The heart and essence of machine learning is the idea of generalization. Put
in intuitive terms this is the ability for an algorithm to be trained on samples
S = {(x1 , y1 ), . . . , (xn , yn )} of n examples of an unknown relation between some
features x = (x1 . . . , xp ) and a target response y and to provide accurate predictions on new, unseen data. To achieve this goal an algorithm tries to pick up an
estimator fˆ among a class of function F, called the hypothesis class hereafter,
which makes small prediction errors |yi − fˆ(xi )| when evaluated on the train
set S. Of course such an endeavor makes no sense unless we can ascertain that
the train set error is close in some sense to the true error we would make on
the whole population. Practically this discrepancy is estimated using validation
and test sets, a procedure familiar to any data scientist.
1
Again, on intuitive grounds we expect that in order to make good predictions we need to select a hypothesis class F that is appropriate for the problem
at hand. More precisely we should use some prior knowledge about the nature
of the link between between the features x and the target y to choose which
functions the class F should possess. For instance if, for any reason, we know
that with high probability the relation between x and y is approximately linear we better choose F to contain only such functions fw (x) = w · x. In the
most general setting this relationship is encoded in a complicated and unknown
probability distribution P on labeled observations (x, y). In many cases all we
know is that the relation between x and y has some smoothness properties.
The set of techniques that data scientists use to adapt the hypothesis class
F to a specific problem is know as regularization. Some of these are explicit in
the sense that they constrain estimators f in some way as we shall describe in
section 2. Some are implicit meaning that it is the dynamics of the algorithm
which walks its way through the set F in search for a good f (typically using
stochastic gradient descent) that provides the regularization. Some of these
regularization techniques actually pertain more to art than to mathematics as
they rely more on experience and intuition than on theorems.
Figure 1: The architecture of AlexNet which is one of the networks used by the authors
in [1]
Deep Learning is a a very popular class of machine learning models, roughly
inspired by biology, that are particularly well suited for tackling complex, AIlike tasks such as image classification, NLP or automatic translation. Roughly
speaking these models are defined by stacking layers that, each, combine linear
combinations of the input with non-linear activation functions (and perhaps
some regularization). We won’t enter into defining them in detail here as many
excellent textbooks [3, 4] will do the job. Figure 1 shows the architecture of
AlexNet a deep network used in the experiment [1]. For our purpose, which is a
discussion of the issue of generalization and regularization, suffice it to say here
that these Deep Learning problems share the following facts:
• The number n of samples available for training these networks is typically
much smaller than the number k of parameters w = (w1 , . . . , wk ) that
define the functions fw ∈ F 1 .
• The probability distribution P (x, y) is impossible to describe in any sensible way in practice. For concreteness, think of x as the pixels of and
1 The number of parameters k of a Deep Learning network such as AlexNet can be over a
hundred of millions while being trained on “only” a few millions of images in image-net.
2
image an of y as the name of an animal represented in the picture. Popular wisdom tells us nevertheless that conditional distributions P (x|y) are
located in the vicinity of some sort of “manifold” My ⊂ Rp .
• The hypothesis class F is defined implicitly by some neural network architecture together with one or more regularization procedures. In section
4 we shall briefly describe the more common such procedures.
As a matter of fact, from a practitioner’s point of view, Deep Learning with
its many tips and tricks [5] generalizes almost unreasonably well. The question
thus is: why? There are two kind of attempts at explaining this mystery.
1. On the one hand there are heuristic explanations, which rely mostly on
variants of physicist’s multi-scale analysis [6]. Proponents of these explanations argue that the efficiency of deep learning models F has its roots
in the physical generating process P of features x (such as image pixels)
that describe objects found in nature. They claim this process is generally
a hierarchy of simpler processes and that the multi-layered is well adapted
to learn such signals. We won’t touch upon this further here.
2. On the other there are classical results from Empirical Risk Minimization
theory [7] which substantiates the idea that the class F of hypothesis
functions should not be too flexible in order for a model to generalize
properly. This is our subject.
Numerical experiments in [1] are thought provoking because they challenge the
second class of explanations which all interpret regularization as a way of limiting some specific measures of class complexity. In the sequel we shall use
hypothesis class and model interchangeably.
The main results from Empirical Risk Minimization (ERM) theory which form
the necessary background for interpreting the experimental results in [1] are
presented in section 3. However, to put things in perspective we first review the
concepts of generalization and regularization in the context of the bias-variance
trade-off which is more familiar for data scientists. Section 2 and 3 are actually
logically independent but the underlying theme of the rigidity of a class F of
functions unifies them. In the context of bias-variance trade-off the rigidity of
a model is optimized to minimize the so called expected loss, whereas in the
context of ERM the rigidity of a model allows to bound population errors with
sample errors. Finally, section 4 explains in what sense the experiments in [1]
are challenging ERM.
2
The Bias-Variance Trade-off
Readers in a hurry to learn the basics of ERM can skip this section except for
definitions (1) and (3).
2.1
The Expected Loss
In the most general setting the relationship between the features x and the
target y is given by an unknown probability distribution P from which labeled
3
observations (x, y) are drawn. Assume we use a function f ∈ F from our
hypothesis class to make predictions and that we pick a loss function `(y, f (x)) ≡
`f (x, y) to measure the error between the true value y and our prediction f (x).
The population error errP [f ], also called the expected loss (EL), can then be
defined as
Z
errP [f ] ≡ `(y, f (x))P (x, y) dx dy
(1)
≡ EP [`f ].
This is the quantity that we would ideally like to minimize as a functional
over F. The special case of a square loss function `(y, f (x)) = [y − f (x)]2 is
interesting, not because it is particularly useful in practice, but because it leads
to explicit expressions of various quantities. In this special case, the optimal h
which minimizes (1) is given by the conditional mean
Z
h(x) = yP (y |x)dy ≡ EP [y|x].
(2)
As P will forever remain unknown to us we must contend ourselves with estimates ĥS of the optimal h based on finite samples S = {(x1 , y1 ), . . . , (xn , yn )}.
An estimate ĥS can be defined as a minimizer over f ∈ F of the sample version
of errP [f ]
n
err
c S [f ] ≡
1X
`(yi , f (xi ))
n i=1
(3)
bS [`f ].
≡E
In a frequentist perspective, the accuracy of these empirical minimizers ĥS ∈ F
can then be defined by taking an average of the EL over train sets S. We call it
the Average Expected Loss (AEL) even if this naming doesn’t particularly shine
with elegance. We denote the averaging operation over training samples S by
Ave to distinguish it from the expectation EP over labeled observations
AEL ≡ Ave[errP [ĥS ]].
2.2
(4)
The Bias-Variance Decomposition
It is instructive to examine the AEL for the special case of a square loss function `(y, f (x)) = [y − f (x)]2 because it allows an explicit decomposition of the
AEL into three components that are easy to interpret. Indeed, let us denote
the Average Expected Square Loss by AESL, the best possible estimator by
h(x) = EP [y|x] and the minimizer over F of the empirical error err
c S [.] by ĥS .
A straightforward application of the above definitions leads to the following
decomposition, see for instance [3]
AESL = (bias)2 + variance + noise,
4
(5)
where
Z
h
i
2
Ave ĥS (x) − h(x) P (x)dx,
Z
2
variance = Ave ĥS (x) − Ave[ĥS (x)]
P (x)dx,
Z
noise = [h(x) − y]2 P (x, y)dx dy.
2
(bias) =
(6)
The (bias)2 measures how much our predictions ĥS (x) averaged over all train
sets S deviate from the optimal prediction h(x). The variance measures how
much our predictions ĥS (x) fluctuate around their average value when the train
set S varies. Finally the noise measures how much the true values y fluctuate
around the best possible prediction h(x), this is only term which is independent
of our predictions ĥS (x).
2.3
Regularization as Constraints
Remember that all prediction functions ĥS belong to some hypothesis class F
that defines our model. We can control the variance by putting some constraints
on functions which belong to F, thus preventing too large fluctuations of ĥS
when training the model on different samples S. Stronger constraint means
a more rigid model and thus typically a smaller variance but a larger (bias)2 .
Regularization of the model amounts to choosing the amount of rigidity that
will minimize AEL using an optimal trade-off between the bias and the variance.
For the AESL it is simply their sum which should be minimized as (5) shows.
If we consider a parametric model, each function fw ∈ F is characterized
by a vector w ∈ Rp of parameters and in particular ĥ = fw
b . One way then to
make the regularization procedure explicit in this context is to put a constraint
on w such as for instance kwk2 < c (Ridge regularization) or kwk1 < c (Lasso
regularization) where the constant c can be tuned for optimal rigidity. To minimize err
c S [fw ] under such a constraint one proceeds the usual way by defining
a Lagrange function and look for its minimum
reg
err
c S [fw ] ≡ err
c S [fw ] + λkwkp
(7)
One can show that for each constraint c there is indeed a corresponding λ for
b p < c. Summarizing the intuition gained so far and denoting Fλ the
which kwk
class of regularized functions
larger λ ⇔ smaller c ⇔ stronger constraints ⇔
“smaller” Fλ ⇔ (smaller variance and larger bias).
If we want to explicit the dependence of our minimizer on the regularization
parameter λ we use ĥS,λ .
2.4
Regularization in Practice
Practically the proper amount of regularization λ in (7) is chosen by using a
procedure such as cross-validation (CV) well known to data scientist. This
5
[ for the true AEL defined
procedure can be viewed as giving an estimate AEL
in (4) which involve two distinct sampling procedures:
1. The first procedure estimates Ave in (4) by an arithmetic mean over r
train sets S1 , ..., Sr that correspond to a splitting of the original data S
into r folds as depicted in figure 2.
2. The second procedure estimates the population errors errP [·] in (4) by
their sample versions err
c Tj [·] on the test sets Tj = S\Sj .
Figure 2: A data set S split into r = 4 folds. The red rectangles are the test sets
Tj = S\Sj associated to the train sets Sj .
We thus define the estimate of AEL as a function of λ
r
1X
[
AEL(λ)
≡
err
c S\Sj [ĥSj ,λ ].
r j=1
(8)
d as a function of the regularization λ allows to find an estimator
Plotting ESL
b
λ of the optimal rigidity of the model. When r = 1 CV amounts to splitting
the data between a train set S and a test set T . The train error err
c S [ĥS,λ ] is
usually smaller than the test error err
c T [ĥS,λ ] which typically has an U-shaped
graph as shown in figure 3.
Remark 1 : The bias-variance decomposition in (5) is strictly valid only for a
Figure 3: Empirical estimation of the optimal regularization λ
6
squared loss function `(y, ŷ) = |y − ŷ|2 . However on experimental grounds the
heuristics behind the trade-off, namely the set of equivalences above and the
intuition behind figure 3 are expected to remain valid for other loss functions `
as well.
Remark 2 : Universality theorems for neural networks tell us that there exists
functions f , defined by even the simplest neural network (NN) architectures,
that will make the sample error err
c S [f ] as small as we want when we minimize
without regularization constraints (λ = 0). In other words, neural networks can
possibly over-fit any sample S provided the network is large enough. These are
mere existence theorems however and they tell us nothing neither about how to
optimize regularization λ to minimize EL nor on our practical ability to find a
minimizer in Fλ with these constraints.
3
Empirical Risk Minimization
The whole analysis in last section rests on our ability to approximate the expected loss errP [f ] = EP [`f ] defined in (1) with its sample estimate err
c S [f ] =
bS [`f ] defined in (3). But how good exactly is this estimate? This is the quesE
tion we examine in the current section. It is also the subject matter of Empirical
Risk Minimization in general.
We expect the discrepancy between the population expectation EP [h] of a
bS [h] to depend on the class of function
function h and its empirical estimate E
H from which we pick h. Intuition suggests that if the set F is too flexible
then there are high chances that our empirical estimates could be way off the
population expectation. There are several ways to make such hand waving
statements rigorous but they all rely on some measure of complexity of a set
H of functions. We restrict in this paper to the Rademacher complexity for its
intuitive character and because its one of the more recent concepts.
3.1
Rademacher Complexity
To introduce this concept consider a class H of real valued function h defined on
an arbitrary space Z and let S = {z1 , · · · , zn } be a set of examples zi drawn independently from a distribution P over Z. Our aim is to measure how well functions in H are able to match any prescribed binary sequence σ = {σ1 , . . . , σn }
when they are evaluated on samples S drawn from P . How well a single function h fits thePprescribed sequence σ on the sample S can be defined as the
n
correlation n1 i=1 σi h(zi ) between σ and the values h(S) ≡ (h(z1 ), ..., h(zn )).
It equals 1 for a perfect fit between h(S) and σ. How well the class of functions
H as a whole can fit a specific sequence σ on the sample S can naturally be
defined as the highest correlation we can achieve using functions h ∈ H
!
n
1X
sup
σi h(zi ) .
(9)
n i=1
h∈H
A measure of how well functions in H can fit any sequence σ can be defined
as the expectation Eσ of (9) over sequences σ (sampled uniformly). This is by
7
definition the empirical Rademacher complexity of the class of functions H on
the sample S
"
!#
n
X
1
d n (H) ≡ Eσ sup
Rad
σi h(zi )
.
(10)
n i=1
h∈H
Finally, the Rademacher complexity of the class H is defined as the expectation
of (10) over samples S of size n drawn from P
d n (H)].
Radn (H) ≡ EP [Rad
(11)
If we assume that functions h ∈ H are binary classifiers, which means h(z) ∈
d n (H) ≤ 1. When Rad
d n (H) is close
{−1, +1}, we see that (9) implies 0 < Rad
to its upper bound 1 the binary classification model defined by H can literally
“store” any binary assignment σ to the examples in S.
3.2
Bounding Population Expectations
The Rademacher complexity is the key ingredient for bounding a population
bS [h]. Assume we draw samples
expectation EP [h] with an empirical average E
S = {z1 , . . . , zn } of size n from a distribution P and that the functions h we
are interested in belong to a class H. Chose a small positive number 0 < δ < 1.
Then the following bound holds with a probability larger than 1 − δ for any
h ∈ H, see [7]
r
bS [h] + 2 Rad
d n (F) + 3 ln(2/δ) .
(12)
EP [h] ≤ E
n
The second term on the right hand side substantiates our intuition that the
discrepancy between the population expectation EP [h] and the empirical exbS [h] can grow when the set of functions H becomes more and more
pectation E
flexible. The last term is a price we pay for requiring small chances to be wrong,
fortunately it grows only logarithmicamlly as δ → 0.
3.3
Application to Binary Classification Errors
Let us apply (12) to a binary classification problem. In this case the set Z is
simply the set Rp × {−1, +1} of observations z = (x, σ). As errP [f ] = EP [`f ]
and err
c S [f ] = ES [`f ] the functions whose expectations we want to bound are
the errors function `f associated with binary classifiers f : Rp → {−1, +1}.
Therefore here H = {`f |f ∈ F } is the class of loss functions. We select the
miss-classification rate `f (x, σ) = If (x)6=σ as our loss function. Using If (x)6=σ =
[1 − σf (x)]/2 and the (almost) obvious property, see [7]
d n (aF + b) = |a| Rad
d n (F)
Rad
d n (H) = 1 Rad
d n (F). Using (12) we can thus express our basic inwe get Rad
2
equality directly in terms of the complexity of the hypothesis class F
r
ln(2/δ)
d
errP [f ] ≤ err
c S [f ] + Radn (F) + 3
(13)
n
with probability > 1 − η.
8
Inequality (13) is the basic prerequisite for interpreting the
challenging results in paper [1]
Remark 1: It is important to realize that inequality (13) holds, of course, uniformy for any f ∈ F. In particular f need not be a minimizer of the expected
loss which was our main concern in section 2.
Remark 2: Inequality (13) assumes that f is selected from a fixed class F
b
whereas in the previous section we empirically estimated an optimal rigidity λ
using a train set S and test set T . Although it is very tempting (and morally
right!) to substitute Fλb for F in (13) strictly speaking this is not correct.
d n (F) as a funcFor parametric models the next step would be to bound Rad
tion of the sample size n and the number k of parameters (w1 , . . . , wk ) defining
fw ∈ F. To our knowledge there are no explicit bounds of this kind. There
exists however an interesting bound for binary classification models which exd n (F) as n → ∞ for a fixed class F. It
emplifies the asymptotic behavior of Rad
involves an alternative notion of complexity namely the VC dimension d of F
defined as follows: it is the size of the largest set S = {x1 , . . . , xd } such that
for any binary sequence σ = (σ1 , . . . , σd ) we can find an f ∈ F that takes these
values on S. The bound reads
r
2d ln n
d
,
(14)
Radn (F) ≤
n
which vanishes when the sample size n grows larger which should come as no
surprise.
4
Regularization for Deep Neural Networks
Now asymptotic behavior is of interest for mathematicians but data scientist
are truly interested in finite samples! Here we can only speculate about the
d n (F) as a function of the size n of the sample and the number k of
value of Rad
parameters. Or we can make experiments. In the context of Deep Learning the
case of particular interest is n k. Recall the following:
Classical interpretation of regularization: Regularization is an explicit
or an implicit constraints on predictors f that shrinks the Rademacher complexity of a neural network model F towards zero so that inequality (13) nearly
saturates.
The most common such mechanisms in Deep Learning are, see [3, 4]
• L2 penalty on weights was shortly discussed in 2.3
• Dropout consists in randomly dropping some links of the neural network
during training thus preventing the model to adapt to closely to the train
data.
• Early stopping consists in monitoring the prediction error on a validation
set while performing gradient descent and stopping it when the validation
error start to increase. Conceptually it is close to the Lp penalty [3].
9
• Data augmentation is used in image recognition problems. The original
data set is augmented with artificial images obtained by distorting (the
size, the orientation, the hue of) the original images. It thus acts to
d n (F) as (14) illustrates.
increase n without changing F, thus reducing Rad
• Implicit regularization effectively shrinks the class of F as a consequence of the dynamics of how the algorithm explores F when optimizing
the network parameters w. Why this happens is more mysterious than
for the previous techniques.
The numerical experiments described in [1] involve large neural networks with
a number of parameters p > 1 000 000 such as AlexNet and deep Multilayer Perceptrons. The data sets S = {(x1 , y1 ), . . . , (xn , yn )} that where used are image
data sets such as CIFAR-10 which contains n = 60 000 images categorized in 10
classes. The authors also trained the models on pure noise images.
They then created artificial data sets S σ = {(x1 , σ1 , . . . , (xn , σn )} in which
the labels σi where generated randomly. They noticed the surprising fact that
for these large networks, even without any explicit regularization mechanism
switched on, the train error err
c S σ [hS σ ] was close to zero (< 0.18%) for any binary assignment σ and even for images made of random pixels. Assuming this
multi-class problem is cast into series of equivalent binary classification problems this strongly suggests:
d n (F) of these large models is very close to 1. In
Fact 1: The complexity Rad
other words these large networks are able to “learn” or “store” these artificial
data sets almost perfectly. But remember:
Fact 2: Deep Learning models are known to generalize very well in practice.
This is the reason why people use them! In other words minimizers of the sample error err
c S [.] are also verified to usefully approximate minimizers of the true
population error errP [.].
Conclusion: Considering the basic bound (13), one way to reconcile fact 1
and fact 2 is simply to give up the classical interpretation of regularization that
d n (F) towards zero. Anpretends its role is to shrink the model complexity Rad
other possibility is that we need more refined bounds than (13). Remember that
this inequality assumed strictly nothing about the distribution P from which
samples S are drawn. It should be no surprise then that these kind of bounds
are way too rough. They really don’t take into account at all what makes the
specificity of deep neural networks! The whole point of using techniques such as
deep learning is that the models F they define are somehow well adapted to the
actual samples S that nature produces. What is a stake is better understanding
of the subtle interplay between the dynamics of the learning process of a multilayered network, which defines an effective F, and the nature of the physical
processes P that generate physical samples S on which they are trained. But
for the moment deep neural networks keep their mystery.
10
Acknowledgments
I would like thank my colleagues Olivier Reisse, the founder of Weave Business
Technology entity, and Christophe Vallet who is leading the Weave Data entity
for their consistent support for all activities at the Weave data lab where I work.
I would also like to thank Marc-Antoine Giuliani who is an expert in statistical
learning theory and data scientist at the Weave data lab for pointing our the
many intricacies of his field to a somewhat carefree physicist like me.
References
[1] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, Oriol Vinyals,
Understanding Deep Learning Requires Rethinking Generalization, ICLR
2017, arxiv.org/abs/1611.03530v2.
[2] Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, Oriol
Vinyals, Understanding Deep Learning Requires Rethinking Generalization, OpenReview.net, https://openreview.net/forum?id=Sy8gdB9xx&
noteId=Sy8gdB9xx.
[3] Christopher M. Bishop, Pattern Recognition And Machine Learning,
Springer-Verlag New York Inc, second edition, 2011.
[4] Ian Goodfellow, Yoshua Bengio and Aaron Courville, Deep Learning, MIT
Press, 2016, http://www.deeplearningbook.org.
[5] Nikolas Markou, The Black Magic of Deep Learning - Tips and Tricks for
the practitioner, EnVision blog, http://nmarkou.blogspot.fr/2017/02/
the-black-magic-of-deep-learning-tips.html.
[6] Henry W. Lin (Harvard), Max Tegmark (MIT), Why does deep and cheap
learning work so well? https://arxiv.org/abs/1608.08225v2.
[7] Maria-Florina Balcan, Rademacher Complexity, Lecture Notes in Machine Learning Theory CS 8803 - Georgia Tech, http://www.cs.cmu.edu/
~ninamf/ML11/lect1117.pdf
11
| 10math.ST
|
Scheduling Wireless Ad Hoc Networks in
Polynomial Time Using Claw-free Conflict
Graphs
arXiv:1711.01620v1 [cs.IT] 5 Nov 2017
Alper Köse∗† , Muriel Médard∗
∗ Research Laboratory of Electronics, Massachusetts Institute of Technology,
77 Massachusetts Avenue, Cambridge, MA 02139, USA
† Department of Electrical Engineering, École Polytechnique Fédérale de Lausanne
Route Cantonale, 1015 Lausanne, Switzerland
{akose, medard}@mit.edu
Abstract—In this paper, we address the scheduling
problem in wireless ad hoc networks by exploiting the
computational advantage that comes when such scheduling
problems can be represented by claw-free conflict graphs.
It is possible to formulate a scheduling problem of network
coded flows as finding maximum weighted independent
set (MWIS) in the conflict graph of the network. We
consider activation of hyperedges in a hypergraph to model
a wireless broadcast medium. We show that the conflict
graph of certain wireless ad hoc networks are claw-free.
It is known that finding MWIS of a general graph is NPhard, but in a claw-free conflict graph, it is possible to
apply Minty’s or Faenza et al.’s algorithms in polynomial
time. We discuss our approach on some sample networks.
Keywords—Wireless ad hoc networks; Claw-free graph;
Conflict graph; Independent Set
I.
I NTRODUCTION
The problem of scheduling in wireless ad hoc networks is generally computationally expensive. In [1],
Arikan proves that scheduling is NP-complete in general
packet radio networks when both primary and secondary interference are considered. In spread spectrum
networks, complexity reduces as shown in [2], where
link-based scheduling can be done in polynomial time
when each node can converse with at most one other
node at a time. We schedule subsets of hyperedges in
the hypergraph model of the network instead of doing
classical link-based scheduling due to inability of linkbased scheduling to capture the character of network
coded flows. In [3], authors prove the NP-completeness
of scheduling broadcasts without tolerance of secondary
conflicts. Another perspective for evaluation of interference is K-hop interference models in which no two links
within K hops can successfully transmit simultaneously.
It is proved in [4] and [5] that link scheduling can be
completed in polynomial time when K = 1, otherwise
it is NP-hard. In [6], scheduling of network coded flows
with conflict graphs is considered, but without exploiting
any graph characteristic. The model we illustrate is
∗ This
work has been presented at the 2017 IEEE PIMRC.
closely related to [6]. Novelty of this paper comes
from the fact that polynomial time scheduling property
of wireless networks having claw-free conflict graphs
is not investigated, and we contribute to literature by
introducing some families of networks which can be
scheduled in polynomial time.
In a conflict graph of a network, each node symbolizes a valid transmission and an edge between two
nodes represents that these two transmissions cannot
be scheduled simultaneously, owing to interference constraints which can differ based on the assumptions on
the network. The rate region of the network with random
linear network coding is determined by three different
constraints on capacity, flow and scheduling as in [6].
Among these constraints, the determinant for optimization complexity is the scheduling constraint. Therefore,
in this work, we concentrate on the scheduling constraint
rather than the others. The scheduling problem can be
formulated as finding maximum weighted independent
set (MWIS) in the conflict graph where independent
set is defined as follows. Given an undirected graph
G = (V, E), a subset of vertices S ⊆ V is an independent set if {i, j} ∈
/ E is satisfied for all i and j
in S. There are several examples for independent set
formulation including the scheduling for routed traffic
[7] and network code construction in a wireline setup
[8]. Finding MWIS dominates the complexity of the
rate maximization problem in the rate region. MWIS
problem is NP-hard for general graphs, therefore this
makes scheduling NP-hard in general networks. On other
hand, it is known that MWIS problem is solvable in
polynomial time in claw-free graphs [9], [10], [11]. A
graph G = (V, E) is claw-free if none of its vertices V
has three pairwise nonadjacent neighbours [12].
The remainder of the paper is organized as follows.
We detail on the conflict graph construction and present
different scenarios, which are on line and tree networks
for which the conflict graphs are claw-free, in Section II.
As a next step, we explain the rate region constraints and
discuss some results in a butterfly network in Section III.
Finally, we conclude the paper in Section IV.
II.
N ETWORK M ODEL AND C ONFLICT G RAPH
E
Throughout the scenarios, we use the Protocol model
[13] and K-hop interference model [4] with small variations to represent networks instead of the Physical model
[13], which takes SIN R levels into account.
r
A. Scenario I
r
B
r
A
C
r
D
r
r
F
G
We consider a wireless ad hoc network with n nodes.
We have a set of assumptions for the network:
A1.1) A node cannot receive from multiple nodes at Fig. 1: A wireless network example with the weights r
the same time and a node cannot receive if it is in the which show the Euclidean distance between nodes.
interference range of another transmitting node.
A1.2) Time division duplex transceivers are used in
network.
(E, B)
A1.3) The Protocol model [13] is used to decide interference relationships between transmissions. Suppose node
A is transmitting to node B where node C is simultaneously making a transmission to another node. Then,
the transmission between A and B will be successful if
(A, B)
and only if inequalities (1) and (2) are satisfied:
|A − B| ≤ rT
|B − C| ≥ (1 + ∆)|A − B|
(1)
(F, C)
(G, D)
Fig. 2: An induced subgraph of the conflict graph of the
(2) network in Fig. 1.
where rT is the maximum transmission range of
nodes and ∆ > 0 is the guard zone which can be
chosen arbitrarily small for simplicity without loss of
generality. Inequality (1) means that every node employs
a common range of transmission rT , and (2) says that
in a communication pair, among all transmitting nodes,
the receiver of this pair must be closest to its transmitter
with a guard zone ∆.
Without making additional assumptions, the conflict
graph of a network is not guaranteed to be claw-free.
Construction 1 (Construction of Conflict Graph):
Reference [6] shows the construction of conflict graph
G = (V, E) in a wireless network. Here, we use a similar
strategy. It consists of the steps below:
C1.1) We can denote transmission v as (i, J) where i
is the source and J as the set of receivers and N (i)
as the set of nodes that are in the transmission range
of node i. The condition for a transmission being valid
is given by J ⊂ N (i), which means that the set of
receiver nodes must be in the transmission range of the
transmitter node. A valid transmission v is modeled as
a vertex, v ∈ V.
C1.2) Say v1 , v2 ∈ V, then {v1 , v2 } ∈ E if v1 and v2
cannot be scheduled simultaneously.
Let v1 , v2 ∈ V be in the conflict graph.{v1, v2 } ∈ E if
any of the below conditions hold:
C1.2.1) i1 = i2 ;
C1.2.2) (i1 ∈ J2 )||(i2 ∈ J1 );
C1.2.3) J1 ∩ J2 6= ∅;
C1.2.4) |i2 − j| ≤ (1 + ∆)|i1 − j| for ∃j ∈ J1 ;
C1.2.5) |i1 − j| ≤ (1 + ∆)|i2 − j| for ∃j ∈ J2 .
Example 1: Consider the topology given in Fig.
1. Assume a transmission range rT and an Euclidean
distance r where rT ≥ r . In this case, when A is a transmitting, for instance to B, then there will be interference
to (E, B), (F, C), (G, D) due to the Protocol model [13]
and hence transmission node (A,B) shares edges with
these three possible transmissions in the conflict graph.
On the other hand, suppose ∠BAC, ∠CAD, ∠DAB =
120◦ , then there will not be any interference between any
pairs of (E, B), (F, C), (G, D) because these transmissions satisfy constraint (2) as well as (1). Consequently,
this will lead to a claw in the conflict graph which can
be seen in Fig. 2.
Example 2: A possible arrangement of wireless
nodes to have a claw-free conflict graph is shown in
Fig. 3. Source and sink can be thought as the nodes A
and E, respectively. Let the maximum possible transmission distance be rT and ∆ be very small. Then,
we can model the conflict graph of this network as
seen in Fig. 4. The independent set polytope is the
convex hull of the incidence vectors of the five independent sets {(A, B), (D, E)}, (A, C), {(B, C), (D, E)},
(C, D) and (A, {B, C}).
We can generalize Example 2 by realizing that
the conflict graphs of the line networks are claw-free
provided that there is enough distance between nodes
to make 3 node away transmission impossible. We
should physically satisfy rT < |µi − µi+3 | for all
i ∈ {1, 2, ..., n − 3} where µi is the ith node in the
A
2rT /3
B
rT /3
C
rT
D
rT
E
Fig. 3: A possible physical arrangement of the wireless
network which leads to a claw-free conflict graph for
Scenario I.
(A, B)
B
C
D
E
F
Fig. 6: Illustration for proof of Theorem 1.
v1
v2
...
vn−2
vn−1
Fig. 7: Conflict graph of a line network where only
possible transmission for a node i is vi = (i, i + 1).
(A, C)
(C, D)
A
(D, E)
(A, {B, C})
(B, C)
Fig. 4: Conflict graph of the network seen in Fig. 3.
network. This inequality can be easily satisfied by many
different positioning scenarios, so, for simplicity, we
can use a hypergraph H = (N , A) to represent our
model, where N denotes the nodes and A denotes the
hyperedges to symbolize valid transmissions between
wireless nodes. Hypergraph of a line network, which
has claw-free conflict graph, changes depending on the
number of nodes to which a node i can transmit, which
can be 1 or 2 for every i ∈ {1, 2, ..., n − 2} and 1
for i ∈ {n − 1}. We have 2n−2 different possible
hypergraphs of a line network, with n nodes, all of
which lead to claw-free conflict graphs. For instance,
hypergraph of the network in Fig. 3 can be seen in Fig.
5. Here, first node is able to transmit up to next two
neighbours whereas the other nodes are only able to
transmit to next node. Transmissions (A, B) and (C, D)
are not simultaneously possible, because we use the
Protocol model [13] to decide interference relations.
Receiver B is closer to C, a transmitter of another
transmission, than to A and this violates the constraint
(2). Directed antennas, which we do not assume in our
scenario, could be used to avoid the interference.
To have a claw in the conflict graph, a transmission
v1 should have conflicts with three other transmissions
v2 , v3 , v4 whereas those three should not have any
conflicts between them. Let us use Fig. 6 for ease of
understanding. Since transmissions are from µ1 to µn ,
from source to sink, assume that a node only transmits
to nodes that are located closer to sink in terms of hop
distance. To this end, assume v1 as the central node
of a possible claw, v1 = (C, D). Then, we can have
one interfering transmission from the source side of C,
say v2 = (A, B), and one from the sink side of D,
v3 = (E, F ) not to have interference between v2 and
v3 . B and E are chosen to be as far as possible. In such
situation, transmitting nodes C and E cause interference
on receiving nodes B and D, respectively. Now, we
have to place the transmitter and the receiver of the
last transmission. This one has to interfere with (C, D)
without interfering with (A, B) and (E, F ) to induce a
claw in the conflict graph of the network. If we place
the transmitter on the source side of C, this leads to
an interference with (A, B) which will break the claw,
so this option is not possible. Also, we cannot place
the receiver in the sink side of D since this leads to
an interference with (E, F ) which will again break the
claw. Therefore, since the receiver must be on the sink
side relative to the transmitter, the only remaining option
is to place both the transmitter and the receiver between
C and D. However, this option makes C able to transmit
to 3 different nodes where we assume each node is able
to transmit to at most 2 nodes.
Let us consider a line network to see the effect of
increasing number of nodes on end-to-end throughput,
first when all transmissions are orthogonal and second
when MWIS scheduling policy is used. To be very
simple, suppose we have a line network with n nodes
and each node can only transmit to next node where node
Proof: Let us prove this theorem by contradiction. µ1 is the source and node µn is the sink. Suppose n is
odd. Also, assume that the transmission (i, i + 1) does
not lead to an interference with (i + 2, i + 3). We name
the transmission (i, i+1) as vi . Under these assumptions,
B
our conflict graph is shown in Fig. 7. So we have two
different maximum independent sets, where weights are
A
C
D
E
assumed equal, which are S1 = {v1 , v3 , ..., vn−2 } and
S2 = {v2 , v4 , ..., vn−1 }. Consider the situation where
transmissions do not have any errors or loss. Since we
Fig. 5: Hypergraph of the network seen in Fig. 3.
only need 2 time slots to combine S1 and S2 to complete
Theorem 1: Under the assumptions A1.1-3 and
Construction 1, the conflict graph of a line network
where a node is able to transmit to at most 2 nodes
and all nodes convey information in the direction from
µ1 to µn is guaranteed to be claw-free.
end-to-end transmissions, we conclude that end-to-end
F
throughput is 1/2 bits/sec if we assume one time slot
D
has a duration of 1 second and the link capacity between
B
G
nodes is 1 bit/sec. The throughput does not depend
A
E
on n. However, if we assume all transmissions to be
orthogonal, we can say that throughput will decrease
C
linearly with the number of nodes in the network as it
equals 1/(n − 1) bits/sec per frequency use. As seen,
we have 0.5(n − 1) gain on throughput by using MWIS Fig. 8: An example of a tree network which has clawapproach. Using network coding does not provide any free conflict graph for Scenario II.
advantage in the end-to-end throughput of this system.
B. Scenario II
We can have other network topologies that induce
a claw-free conflict graph. One of them is a tree representation of the network, but since it is harder to get
claw-freeness with the same assumptions for the line
topology model, we propose new set of assumptions:
A2.1) A node cannot receive from multiple nodes at the
same time.
A2.2) Time division duplex transceivers are used in
network.
A2.3) Nodes are arranged as a tree topology.
A2.4) We directly work on hypergraph model without
any consideration on physical locations of nodes and
assume an interference model based on hops instead of
the Protocol model [13].
A2.5) Transmissions are in the direction from root node
to leaves of tree.
A2.6) Directed antennas are used, so interference can
only occur in the forward direction along the tree.
A2.7) A transmitting node does not lead to any interference to the receivers which are 3-hops or more away
from it.
A2.8) Only one node in every level can have children.
(A, B)
(B, D)
(D, F )
(A, C)
(B, E)
(D, G)
(A, {B, C})
(B, {D, E})
(D, {F, G})
Fig. 9: Conflict graph of the network seen in Fig. 8.
ik ’s parent. Since only one of these transmissions can
be scheduled in one time slot, they induce a complete
subgraph in the conflict graph. Therefore, we can only
choose one transmission, say v2 , from level k − 1 to k
which has interference with v1 because the cardinality
of maximum independent set of the complete graph is
1. Assuming that we have the node ik+1 , which belongs
to level k + 1 and has children, in the receiver set of the
transmission v1 (otherwise, it is easier to say that we will
not have a claw.), we have another complete subgraph
which contains the transmissions from the node ik+1 to
its children in level k + 2. One of these transmissions
can be selected for a possible claw, say v3 . Since, it is
not possible to find another independent transmission v4 ,
we have a claw-free conflict graph for the network.
Construction 2 (Construction of Conflict Graph):
In the construction of conflict graph, we implement the
steps below:
C2.1) Model valid transmissions as the vertices V of
conflict graph G = (V, E) as in C1.1.
C2.2) Say v1 , v2 ∈ V, then {v1 , v2 } ∈ E if v1 and v2
cannot be scheduled simultaneously.
Let v1 , v2 ∈ V be in the conflict graph.{v1, v2 } ∈ E if
any of the below conditions hold:
C2.2.1) i1 = i2 ;
An example of a tree network which has a claw-free
C2.2.2) (i1 ∈ J2 )||(i2 ∈ J1 );
conflict
graph and its conflict graph can be seen in Fig.
C2.2.3) J1 ∩ J2 6= ∅;
8
and
Fig.
9, respectively.
C2.2.4) (i1 is a child of i2 ) || (i2 is a child of i1 ).
Theorem 2: Under the assumptions A2.1-8 and C. Scenario III
Construction 2, conflict graph of a wireless network is
In this scenario, transmission and interference
guaranteed to be claw-free.
schemes are less restricted, since we only allow for
Proof: Let us assume that the root node belongs to full-duplex transceivers and do not allow for 2-hop
level 1 and a node, which has a distance k to root in interference, compared to Scenario II, but the network
terms of hyperedge number, belongs to level k +1. Now, topology is relaxed by letting every node have children.
consider the transmission v1 , from a node ik in level k Assumptions are below:
to its children that reside in level k +1. We have 2Nk −1 A3.1) A node cannot receive from multiple nodes at the
different interfering transmissions to v1 which are from same time.
level k − 1 to k where Nk is the number of children of A3.2) Full duplex transceivers are used in network.
A3.3) Nodes are arranged as a tree topology.
F
A3.4) We directly work on hypergraph model without
D
any consideration on physical locations of nodes and
G
assume an interference model based on hops instead of
B
the Protocol model [13].
A3.5) Transmissions are in the direction from root node
A
H
to leaves of tree.
C
E
A3.6) Directed antennas are used, so interference can
I
only occur in the forward direction along the tree.
A3.7) A transmitting node does not lead to any interference to the receivers which are 2-hops or more away Fig. 10: An example of a tree network which has a clawfrom it.
free conflict graph for Scenario III but not for Scenario
Construction 3 (Construction of Conflict Graph): II.
In the construction of conflict graph, we implement the
steps below:
C3.1) Model valid transmissions as the vertices V of
conflict graph G = (V, E) as in C1.1.
C3.2) Say v1 , v2 ∈ V, then {v1 , v2 } ∈ E if v1 and v2
cannot be scheduled simultaneously.
Let v1 , v2 ∈ V be in the conflict graph.{v1, v2 } ∈ E if
any of the below conditions hold:
C3.2.1) i1 = i2 ;
C3.2.2) J1 ∩ J2 6= ∅;
Theorem 3: Under the assumptions A3.1-7 and
Construction 3, conflict graph of a wireless network is
guaranteed to be claw-free.
Proof: Let us again assume that the root node
belongs to level 1 and a node, which has a distance k
to root in terms of hyperedge number, belongs to level
k + 1. In these settings, we eliminate the possibility
of having interference between any transmissions from
different levels. Let us denote a transmission from a node
in level k to a node in level k + 1 as k → k + 1. Any
transmission k → k+1 interferes neither with k−1 → k
nor with k + 1 → k + 2 because we use full-duplex
transceivers which makes a node eligible to transmit
and receive at the same time. Therefore, we only have
interference between the transmissions which initiate
from the same node. A node with j children has 2j − 1
possible transmissions where each pair has interference
between them which then leads to a complete graph with
2j − 1 nodes in the conflict graph. Assume we have m
non-leaf tree nodes, then we have m disjoint complete
subgraphs in the overall conflict graph which implies
that the conflict graph is claw-free.
An example of a tree network which has a claw-free
conflict graph and its conflict graph can be seen in Fig.
10 and Fig. 11, respectively.
(A, B)
(B, D)
(D, F )
(E, H)
(A, C)
(B, E)
(D, G)
(E, I)
(A, {B, C})
(B, {D, E})
(D, {F, G})
(E, {H, I})
Fig. 11: Conflict graph of the network seen in Fig. 10.
T ∈ N where all sinks T ∈ N demand the same
information and network coding is used. We can extend
it to multiple connections using intra-session coding.
When we use i to denote a wireless node, N (i) denotes
the set of nodes to which node i can transmit, namely
there is a hyperedge from i to set N (i). Recall that
MWIS problem dominates the complexity of the rate
maximization problem in the rate region. In this section,
we review the three constraints of the rate region to
highlight the importance of the scheduling problem that
we address by revisiting [6] and [14] and give the
formulation of MWIS problem for the conflict graph.
Firstly, we have capacity constraints given in (3).
X
(t)
xij ≤
j∈K
X
zij
(3)
j⊂N (i)
∀i ∈ N , K ⊂ N (i), t ∈ T .
(t)
In constraint (3), xij shows the flow rate of data
packets from node i to node j in the way of sink t. ziJ
III. R ATE R EGION
is the average packet injection rate to the output set J of
hyperedge
(i, J) and ziJ is guaranteed to exist and be
A. Rate Region Constraints
finite if the injection process is assumed to be stationary
To find the rate region, we assume that we have a and ergodic.
hypergraph H = (N , A) model of the network as in
Second, we have flow constraints given in (4) and
Fig. 8 where N symbolizes the nodes and A symbolizes
the hyperedges. We consider the multicast connection (5).
with rate R, we have a source s ∈ N and sinks
network has claw-free conflict graph. There are 12 valid
transmissions, 3 for each node being source except E
if i = s
R,
X
X (t)
and F . Therefore, possible transmissions from a node
(t)
xji = −R, if i = t
xij −
(4) lead to a complete subgraph of 3 nodes. To have a claw,
{j|i∈N (j)}
j∈N (i)
0,
otherwise
a transmission (i1 , J1 ) should have three conflicts, each
with a different source ia 6= ib for a, b ∈ {1, 2, 3, 4}
where these three should not have any conflicts between
∀i ∈ N , t ∈ T .
them. However, this is not possible in the given butterfly
network which lead to a claw-free conflict graph. So,
(t)
(5) the scheduling can be completed in polynomial time.
xij ≥ 0
Complement of the conflict graph can be seen in Fig.
∀i ∈ N , j ∈ N (i), t ∈ T .
13 for a more clear illustration.
Constraint (4) should be satisfied to guarantee a flow
Assume all transmissions are done without any errors
of rate R from the source to sinks and it is trivial that or loss from source to sinks and every link has a unit
flow rates cannot be negative as seen in (5).
capacity, say 1 bit/sec. We have 2 bits to send in the
multicast
connection. We consider four different conLastly, we have the scheduling constraint given in
ditions where the constraints (3)-(6) from the previous
(6).
subsection only apply to MWIS approach with network
coding:
z ∈ PIN D (G).
(6) 1) All transmissions are orthogonal without network
coding.
In (6), vector z = (ziJ ) is called network coding 2) All transmissions are orthogonal with network coding.
subgraph where (i, J) ∈ A. PIN D (G) denotes the 3) MWIS approach without network coding.
independent set polytope of G = (V, E), which is the 4) MWIS approach with network coding.
conflict graph of the considered network. In overall, the
In the fully orthogonal model [14], all transmission
network coding subgraph should lie in the independent
signals are orthogonal, making the network interferenceset polytope of the conflict graph.
free but this leads to a suboptimal bandwith usage.
The maximum weighted independent set problem We maximize bandwith efficiency by frequency reuse
of a weighted undirected graph G = (V, w, E) with n whenever possible. We could also consider the two-hop
vertices where w : V(G) → R can be formulated as constraints model [6] in which if a node i transmits,
none of the nodes in two-hop neighbourhood of i can
below:
P
transmit, but it adds up to the same thing with orthogonal
Maximize i∈V wi vi
model in this butterfly network. In 1-3, our conflict graph
construction, Construction 1, does not represent the
Subject to vi + vj ≤ 1 {i, j} ∈ E
conditions correctly owing to some additional constraints
where vi ∈ {0, 1} for 1 ≤ i ≤ n.
that show up. For instance, when we adopt orthogonal
We maximize R subject to (3)-(6). Scheduling con- model, conflict graph becomes a complete graph where
straint (6) can be formulated as MWIS problem, and we can choose only one transmission for a time slot
it is known to be NP-hard for general graphs to solve due to our bandwith limited regime. Our conflict graph
which makes our optimization problem NP-hard. How- formulation can only be used for MWIS approach with
ever, there are algorithms proposed for solving MWIS network coding. Throughput analysis can be seen below:
problem in claw-free graphs in polynomial time. For 1) We can achieve an end-to-end throughput of 1/3
instance, Minty’s [9], [10] algorithm can be used for bits/sec per frequency use.
solvability in O(n6 ) or Faenza et al.’s [11] algorithm 2) We can achieve an end-to-end throughput of 2/5
can be used for further improvement to O(n3 ). Since we bits/sec per frequency use which is better than the first
are able to find MWIS of the claw-free conflict graph in approach.
polynomial time, it is possible to solve the joint subgraph 3) We are now scheduling {(A, B), (C, {D, F })},
optimization and scheduling problem, i.e. maximizing R {(A, C), (B, {D, E})}, (D, E) and (D, F ) in four
time slots, thus have an end-to-end throughput of 1/2
subject to (3)-(6), in polynomial time [15].
bits/sec which is better than second approach.
4) We first set the conflict graph as seen in Fig. 13,
B. Illustration
then find the maximum weighted independent sets, and
Networks having claw-free conflict graphs are not finally use convex combination of them for scheduling.
limited with the discussed ones in Section II. Let us We have {(A, B), (C, {D, F })}, {(A, C), (B, {D, E})}
consider a butterfly network as seen in Fig. 12 with and (D, {E, F }) which we can combine in three time
the assumptions A1.1-3 which lead to Construction 1 slots. End-to-end throughput is 2/3 bits/sec which is
where r ≤ rT , |AD| > rT and ∆ is negligible. A the highest among four conditions.
is the source node where E and F are the sinks of
a multicast connection. First, we show that butterfly
r
C
r
A
r
r
r
D
r
B
[2]
F
r
[3]
r
[4]
E
Fig. 12: A butterfly network with the weights r which
show the Euclidean distance between nodes.
(A, C)
(C, D)
(C, F )
(C, {D, F })
(D, F )
(A, {B, C})
[5]
[6]
(D, {E, F })
[7]
(A, B)
(B, D)
(B, E)
(B, {D, E})
(D, E)
[8]
Fig. 13: Complement of the conflict graph of the network
seen in Fig. 12.
[9]
IV.
C ONCLUSION
We have investigated wireless ad hoc networks which
have claw-free conflict graphs. To this end, we used
conflict graph to model possible valid transmissions
and their interference relations in the network, then
we exploited the claw-freeness of the conflict graph of
certain networks and MWIS algorithm for graphs to
do the scheduling in polynomial time. We believe that
this paper can be used as a guide to set up wireless
networks which allow jointly solving the network coding
subgraph and the scheduling problem in polynomial
time. However, most of the real life networks do not have
claw-free conflict graphs and our approach can only be
applied to limited number of networks.
For future work, we work on a strategy which
breaks all of the claws in a conflict graph by iteratively
introducing edges. We believe this new strategy will
allow us to do polynomial time scheduling for more
general networks with a very small throughput loss.
Besides this new strategy, the Physical model [13] can
be considered to form the conflict graph of networks
which may complicate scenarios but give a more detailed
view at the same time. Lastly, a superset of clawfree graphs in which MWIS problem can be solved in
polynomial time can be searched to extend throughput
optimal polynomial time scheduling to more general
networks.
R EFERENCES
[1]
E. Arikan, “Some complexity results about packet radio networks (corresp.),” IEEE Transactions on Information Theory,
vol. 30, no. 4, pp. 681–685, 1984.
[10]
[11]
[12]
[13]
[14]
[15]
B. Hajek and G. Sasaki, “Link scheduling in polynomial time,”
IEEE transactions on Information Theory, vol. 34, no. 5, pp.
910–917, 1988.
A. Ephremides and T. V. Truong, “Scheduling broadcasts in
multihop radio networks,” IEEE Transactions on communications, vol. 38, no. 4, pp. 456–460, 1990.
G. Sharma, R. R. Mazumdar, and N. B. Shroff, “On the complexity of scheduling in wireless networks,” in Proceedings of
the 12th annual international conference on Mobile computing
and networking. ACM, 2006, pp. 227–238.
G. Sharma, N. B. Shroff, and R. R. Mazumdar, “Maximum
weighted matching with interference constraints,” in Pervasive
Computing and Communications Workshops, 2006. PerCom
Workshops 2006. Fourth Annual IEEE International Conference
on. IEEE, 2006, pp. 5–pp.
D. Traskov, M. Heindlmaier, M. Médard, and R. Koetter,
“Scheduling for network-coded multicast,” IEEE/ACM Transactions on Networking (TON), vol. 20, no. 5, pp. 1479–1488,
2012.
L. Tassiulas and A. Ephremides, “Stability properties of constrained queueing systems and scheduling policies for maximum
throughput in multihop radio networks,” IEEE transactions on
automatic control, vol. 37, no. 12, pp. 1936–1948, 1992.
J. K. Sundararajan, M. Médard, R. Koetter, and E. Erez, “A
systematic approach to network coding problems using conflict
graphs,” in Proceedings of the UCSD Workshop on Information
Theory and its Applications, 2006.
G. J. Minty, “On maximal independent sets of vertices in clawfree graphs,” Journal of Combinatorial Theory, Series B, vol. 28,
no. 3, pp. 284–304, 1980.
D. Nakamura and A. Tamura, “A revision of minty’s algorithm
for finding a maximum weight stable set of a claw-free graph,”
Journal of the Operations Research Society of Japan, vol. 44,
no. 2, pp. 194–204, 2001.
Y. Faenza, G. Oriolo, and G. Stauffer, “Solving the weighted
stable set problem in claw-free graphs via decomposition,”
Journal of the ACM (JACM), vol. 61, no. 4, p. 20, 2014.
M. Chudnovsky and P. D. Seymour, “The structure of claw-free
graphs.” Surveys in combinatorics, vol. 327, pp. 153–171, 2005.
P. Gupta and P. R. Kumar, “The capacity of wireless networks,”
IEEE Transactions on information theory, vol. 46, no. 2, pp.
388–404, 2000.
D. S. Lun, N. Ratnakar, M. Médard, R. Koetter, D. R. Karger,
T. Ho, E. Ahmed, and F. Zhao, “Minimum-cost multicast over
coded packet networks,” IEEE Transactions on information
theory, vol. 52, no. 6, pp. 2608–2623, 2006.
D. Traskov, M. Heindlmaier, M. Médard, R. Koetter, and D. S.
Lun, “Scheduling for network coded multicast: A conflict graph
formulation,” in GLOBECOM Workshops, 2008 IEEE. IEEE,
2008, pp. 1–5.
| 7cs.IT
|
Predicting Signed Edges with O(n1+o(1) log n) Queries
Michael Mitzenmacher∗
Charalampos E. Tsourakakis†
arXiv:1609.00750v2 [cs.DS] 4 Oct 2016
Abstract
Social networks and interactions in social media involve both positive and negative relationships. Signed graphs capture both types of relationships: positive edges correspond to pairs of
“friends”, and negative edges to pairs of “foes”. The edge sign prediction problem, which aims
to predict whether an interaction between a pair of nodes will be positive or negative, is an
important graph mining task for which many heuristics have recently been proposed [LHK10a;
LHK10b].
Motivated by social balance theory, we model the edge sign prediction problem as a noisy
correlation clustering problem with two clusters. We are allowed to query each pair of nodes
whether they belong to the same cluster or not, but the answer to the query is corrupted
with some probability 0 < q < 12 . Let c = 12 − q be the gap. We provide an algorithm that
recovers the clustering with high probability in the presence of noise for any constant gap c
1+
1
with O(n log log n log n) queries. Our algorithm uses simple breadth first search as its main
algorithmic primitive. Finally, we provide a novel generalization to k ≥ 3 clusters and prove that
our techniques can recover the clustering if the gap is constant in this generalized setting.
1
Introduction
With the rise of social media, where both positive and negative interactions take place, signed
graphs, whose study was initiated by Heider, Cartwright, and Harary [CH56; Hei46; Har53], have
become prevalent in graph mining. A key graph mining problem is the edge sign prediction problem,
which aims to predict whether an interaction between a pair of nodes will be positive or negative
[LHK10a; LHK10b]. Recent works have developed numerous heuristics for this task that perform
relatively well in practice [LHK10a; LHK10b].
In this work we propose a theoretical model for the edge sign prediction problem that highlights
its intimate connections with the famous planted partition problem [ABH16; CK01; HWX16; McS01].
Specifically, we model the edge sign prediction problem as a noisy correlation clustering problem,
where we are able to query a pair of nodes (u, v) to test whether they belong to the same cluster
(edge sign f (u, v) = +1) or not (edge sign f (u, v) = −1). The query fails to return the correct answer
with some probability 0 < q < 12 . Correlation clustering is a basic data mining primitive with a large
number of applications ranging from social network analysis [Har53; LHK10a] to computational
biology [HEP+16]. Our theoretical model is inspired by the famous balance theory: “the friend of
my enemy is my friend” [CH56; EK10; Hei46]. The details of our model follow.
Model I. Let V = [n] be the set of n items that belong to two clusters, call them red and blue.
Set f : V → {red, blue}, R = {v ∈ V (G) : f (v) = red} and B = {v ∈ V (G) : f (v) = blue}, where
∗
†
Harvard University, [email protected]
Harvard University, [email protected]
1
0 ≤ |R| ≤ n. The function f is unknown and we wish to recover the two clusters R, B by querying
pairs of items. (We need not recover the labels, just the clusters.) For each query we receive the
correct answer with probability 1 − q, where q > 0 is the corruption probability. That is, for a pair
of items u, v such that f (u) = f (v), with probability q it is reported that f˜(u) 6= f˜(v), and similarly
if f (u) 6= f (v) with probability q it is reported that f˜(u) = f˜(v). Our goal is to perform as few
queries as possible while recovering the underlying cluster structure.
Main result. Our main theoretical result is that we can recover the clusters (R, B) with high
probability1 in polynomial time. Our algorithm uses breadth first search (BFS) as its main
algorithmic primitive. Our result is stated as Theorem 1.
Theorem 1. There exists a polynomial time algorithm that performs Θ(n log n(2c)
queries and recovers the clustering (R, B) whp for any gap 0 < c = 21 − q < 12 .
1+
log n
− log
log n
) edge
1
When c is constant, then the number of queries is O(n log log n log n). A natural follow-up question
that we address here is whether our results generalize to the case of more than two clusters. We
provide a general model and show that our techniques recover the cluster structure whp as long as
the gap c is constant.
Model II. Our model is now that there are k groups, that we number {0, 1, ..., k − 1} and that we
think of as being arranged modulo k. Let g(u) refer to the group number associated with a vertex
u. We start by noting that if when querying an edge we returned only whether the the groups of
the two edges were equal, it would be difficult to reconstruct the clusters; indeed, even with no
errors, a chain of such responses along a path would not generally allow us to determine whether
the endpoints of a path were in the same group or not. A model that provides more information
and naturally generalizes the two cluster case is the following: when we query an edge e = (x, y),
we obtain
f˜(e) =
g(x) − g(y) mod k,
with probability 1 − q;
g(x) − g(y) + 1 mod k,
with probability q/2;
g(x) − g(y) − 1 mod k,
with probability q/2.
(1)
That is, we obtain the difference between the groups when no error occurs, and with probability
q we obtain an error that adds or subtracts one to this gap with equal probability. When q = 0, so
there are no errors from f˜(e), the edge queries would allow us to determine the difference between
the group numbers of vertices at the start and end of any path, and in particular would allow us to
determine if the groups were the same. We also note that we choose this description for ease of
exposition. More generally we could handle queries governed by more general error models, of the
form:
f˜(e) = g(x) − g(y) + i with probability qi , 0 ≤ i < k.
That is, the error does not depend on the group values x and y, but is simply independent and
identically distributed over the values 0 to k − 1.
1+
1
Theorem 2. There exists a polynomial time algorithm that performs O(n log log n log n) edge
queries and recovers the k clusters under the model of equation (1) whp for any constant gap
0 < c < 12 .
1
An event An holds with high probability (whp) if
lim Pr [An ] = 1.
n→+∞
2
Our proof techniques extend naturally to this model.
Roadmap. Section 2 presents some theoretical preliminaries. Section 3 presents our algorithmic
contributions. Section 4 briefly reviews related work. Finally, Section 5 concludes the paper.
2
Theoretical Preliminaries
We use the following powerful probabilistic results for the proofs in Section 3.
Theorem 3 (Chernoff bound, Theorem 2.1 [JLR11]). Let X ∼ Bin (n, p), µ = np, a ≥ 0 and
ϕ(x) = (1 + x) ln(1 + x) − x (for x ≥ −1, or ∞ otherwise). Then the following inequalities hold:
−µϕ
−a
µ
−µϕ
−a
µ
Pr [X ≤ µ − a] ≤ e
Pr [X ≥ µ + a] ≤ e
2
−a
2µ
≤e
,
(2)
a2
− 2(µ+a/3)
≤e
.
(3)
We define the notion of read-k families, a useful concept when proving concentration results for
weakly dependent variables.
Definition 1 (Read-k families). Let X1 , . . . , Xm be independent random variables. For j ∈ [r], let
Pj ⊆ [m] and let fj be a Boolean function of {Xi }i∈Pj . Assume that |{j|i ∈ Pj }| ≤ k for every
i ∈ [m]. Then, the random variables Yj = fj ({Xi }i∈Pj ) are called a read-k family.
Theorem 4 (Concentration of Read-k families [GLS+15]). Let Y1 , . . . , Yr be a family of read-k
indicator variables with Pr [Yi = 1] = q. Then for any > 0,
Pr
" r
X
#
Yi ≥ (q + )r ≤ e−DKL (q+||q)·r/k
(4)
i=1
and
Pr
" r
X
#
Yi ≤ (q − )r ≤ e−DKL (q−||q)·r/k .
(5)
i=1
Here, DKL is Kullback-Leibler divergence defined as
q
1−q
+ (1 − q) log
.
p
1−p
We will use the following corollary of Theorem 5, which provides Chernoff-type bounds for
read-k families. This is derived in a similar way that Chernoff multiplicative bounds are derived
from Equations (3) and (2), see [McD98]. Notice that the main difference compared to the standard
Chernoff bounds is the extra k factor in denominator of the exponent.
DKL (q||p) = q log
Theorem 5 (Concentration of Read-k families [GLS+15]). Let Y1 , . . . , Yr be a family of read-k
P
indicator variables with Pr [Yi = 1] = q. Also, let Y = ri=1 Yi . Then for any > 0,
Pr [Y ≥ (1 + )E [Y ]] ≤ e
Pr [Y ≤ (1 − )E [Y ]] ≤ e−
3
2 E[Y ]
− 2k(1+/3)
2 E[Y ]
2k
.
(6)
(7)
3
Proposed Method
We prove our main result through a sequence of claims and lemmas. For completeness we include all
proofs even if some claims are classic, e.g., Claim 1. At a high level, our proof strategy is as follows:
1. We compute the probability that a simple path between u and v provides us with the correct
information on whether f (u) = f (v) or not.
4L
2. Let L = logloglogn n . We show that there exist N = (2c)−L e 5 almost edge-disjoint paths of length
(1 + o(1))L between any pair of vertices with probability at least 1 − n13 . The reader can think
of the paths as being edge-disjoint, if that is helpful; we shall clarify both what we mean by
almost edge-disjoint paths and how it affects the proof later in the paper.
3. For each path from the collection of N almost edge-disjoint paths, we compute the product
of the sign of the edges along the path. Since the paths are not entirely edge disjoint, the
corresponding random variables are weakly dependent. We use concentration of multivariate
polynomials [GLS+15], see also [AS04; KV00], in combination with Claim 1 to show that
using the majority of the N resulting signs to decide whether f (u) = f (v) or not for a pair of
nodes u, v ∈ V (G) gives the
correct answer with probability lower bounded by 1 − n13 . Taking
n
the union bound over 2 pairs concludes the proof.
The pseudo-code is shown as Algorithm 1. The algorithm runs over each pair of nodes, and
it invokes Algorithm 2 to construct almost edge-disjoint paths for each pair of nodes u, v using
Breadth First Search. Note that since we perform 20n log n(2c)−L queries uniformly at random,
−L
the resulting graph is is asymptotically equivalent to G ∼ G(n, 40 log n(2c)
), see [FK15, Chapter 1].
n
Here, G(n, p) is the classic Erdös-Rényi model
(a.k.a
random
binomial
graph
model) where each
[n]
possible edge between each pair (u, v) ∈ 2 is included in the graph with probability p independent
from every other edge.
log n
It turns out that our algorithm needs an average degree O (2c)
only for the first level of the
L
trees Tu , Tv that we grow from u and v when we invoke Algorithm 2. For all other levels of the grown
trees, we need the degree to be only O(log n). This difference in the branching factors exists in order
1
to ensure that the number of leaves of trees Tu , Tv in Algorithm 2 is amplified by a factor of (2c)
L,
which then allows us to apply Theorem 5. Using appropriate data structures, Algorithm 1 runs
in O(n2 (n + m)) = O(n3 log n(2c)−L ). One can improve the run time in expectation by sampling
O(log n) neighbors for each node in O(log n) time instead of O(log n(2c)−L ) time using a standard
sublinear sampling technique that generates geometric random variables between successive successes,
see [Knu07; TKM11]. This results in total expected run time O(n3 log n). Since we use a branching
factor of O(log n) for all except the first two levels of Tu , Tv , we work with the G(n, p) model with
n
p = 40 log
to construct the set of almost edge disjoint paths. (Alternatively, one can think that we
n
start with the larger random graph with more edges, and then in the construction of the almost
edge disjoint paths we subsample a smaller collection of edges to use in this stage.) The diameter of
−L
this graph whp grows asymptotically as L [Bol98] for this value of p. We use the G(n, 40 log n(2c)
)
n
model only in Lemma 1 to prove that every node has degree at least 5 log n(2c)−L .
The following result is well known but we present a proof for completeness.
4
Algorithm 1 2-Correlation Clustering with Noise
L ← logloglogn n
Perform 20n log n(2c)−L queries uniformly at random.
Let G(V, E, f˜) be the resulting graph, f˜ : E → {+1, −1}
for each item pair u, v do
Pu,v = {P1 , . . . , PN } ← Almost-Edge-Disjoint-Paths(u, v)
Q
Yi ← e∈Pi f˜(e) for i = 1, . . . , N
P
Yuv ← P ∈Pu,v YP
if Yuv ≥ 0 then
f˜(u) = f˜(v)
else
f˜(u) 6= f˜(v)
end if
end for
Algorithm 2 Almost-Edge-Disjoint-Paths(u, v)
Require: G(V, E, f˜), u, v ∈ V (G)
← √log1log n
Using Breadth First Search (BFS) grow a tree Tu starting from u as follows.
−
log n
For the first level of the tree, we choose 4 log n(2c) log log n neighbors of u.
For the rest of the tree we use a branching factor equal to 4 log n until it reaches depth equal to
L. Similarly, grow a tree Tv rooted at v, node disjoint from Tu of equal depth.
From each leaf ui (vi ) of Tu (Tv ) for i = 1, . . . , N grow node disjoint trees until they reach depth
( 12 + )L with branching factor 4 log n. Finally, find an edge between Tui , Tvi
Claim 1. Consider a path Puv between nodes u, v of length L. Let Ruv =
Pr [Ruv = 1|f (u) = f (v)] = Pr [Ruv = −1|f (u) 6= f (v)] =
Q
e∈Puv
f˜(e). Then,
1 + (1 − 2q)L
2
Proof. Here Ruv = 1 iff f˜ agrees with the unknown clustering function f on u, v. This happens when
the number of corrupted edges along that path Puv is even. Let Zuv ∼ Bin(L, q) be the number of
corrupted edges/sign flips along the Puv path. Clearly, Pr [Zuv is even] + Pr [Zuv is odd] = 1. Also,
L
(1 − 2q) =
L
X
k=0
!
bLc
bLc
!
!
2
2
X
X
L
L 2k
L
(−q)k (1 − q)L−k =
q (1 − q)L−2k −
q 2k+1 (1 − q)L−(2k+1) =
k
2k
2k
+
1
k=0
k=0
Pr [Zuv is even] − Pr [Zuv is odd] .
Therefore Pr [Zuv is even] =
1+(1−2q)L
,
2
and the result follows.
The next lemma is a direct corollary of the lower tail multiplicative Chernoff bound.
log n
Lemma 1. Let G ∼ G(n, 40
) be a random binomial graph. Then whp all vertices have degree
(2c)L n
greater than 5 log n(2c)−L .
5
log n
Proof. The degree deg(u) of a node u ∈ V (G) follows the binomial distribution Bin(n − 1, 40
).
(2c)L n
Set δ = 34 . Then
δ2
i
h
−L
Pr deg(u) < 5 log n(2c)−L < e− 2 40 log n(2c)
n−1 .
Taking a union bound over n vertices gives the result.
Now we proceed to our construction of sufficiently enough almost edge-disjoint paths. Our construction is based on standard techniques in random graph theory [BFS+98; DFT15; FT12; Tso13], we
include the full proofs for completeness.
n
Lemma 2. Let G ∼ G(n, p) where p = 40 log
. Fix t ∈ Z+ and 0 < α < 1. Then, whp there does
n
not exist a subset S ⊆ [n], such that |S| ≤ αtL and e[S] ≥ |S| + t.
Proof. Set s = |S|.Then,
Pr [∃S : s ≤ αtL and e[S] ≥ s + t] ≤
X
s≤αtL
≤
X
(e
s≤αtL
2+o(1)
log n)
s
20es log n
n
t
n
s
s
2
!
s+t
2+o(1)
≤ αtL (e
!
s+t
p
≤
X ne s
s≤αtL
αL
log n)
s
20eαt log2 n
n log log n
es2 p
2(s + t)
!!t
<
!s+t
1
.
n(1−α−o(1))t
Lemma 3. Let T be a rooted tree of depth at most 4L
7 and let v be a vertex not in T . Then with
probability 1 − o(n−3 ), v has at most 10 neighbors in T , i.e., |N (v) ∩ T | ≤ 10.
Proof. Let T be a rooted tree of depth at most 4L
7 and let S consist of v, the neighbors of v in
T plus the ancestors of these neighbors. Set b = |N (v) ∩ T |. Then |S| ≤ 4bL/7 + 1 ≤ 3bL/5 and
e(S) = |S| + b − 2. It follows from Lemma 2 with α = 3/5 and t = 8, that we must have b ≤ 10
with probability 1 − o(n−3 ).
We show that by growing trees iteratively we can construct sufficiently many edge-disjoint paths for
n sufficiently large.
Lemma 4. Let = √log1log n , and k = L. For all pairs of vertices x, y ∈ [n] there exists a
subgraph Gx,y (Vx,y , Ex,y ) of G as shown in figure 1, whp. The subgraph consists of two isomorphic
vertex disjoint trees Tx , Ty rooted at x, y each of depth k. Tx and Ty both have a branching factor
of 4 log n(2c)−L for the first level, and 4 log n for the remaining levels. If the leaves of Tx are
x1 , x2 , . . . , xτ , τ ≥ (2c)−L n4/5 then yi = f (xi ) where f is a natural isomorphism. Between each pair
of leaves (xi , yi ), i = 1, 2, . . . , m there is a path Pi of length (1+2)L. The paths Pi , i = 1, 2, . . . , τ, . . .
are edge disjoint.
Proof. Because we have to do this for all pairs x, y, we note without further comment that likely
(resp. unlikely) events will be shown to occur with probability 1 − o(n−2 ) (resp. o(n−2 )).
To find the subgraph shown in Figure 1(b) we grow tree structures as shown in Figure 1(a).
Specifically, we first grow a tree from x using BFS until it reaches depth k. Then, we grow a tree
starting from y again using BFS until it reaches depth k. For the first level of both trees, we choose
5 log n
neighbors of x, y respectively. For all other levels we use a branching factor equal to 4 log n.
(2c)L
6
Before we show how to continue our construction, we need to prune down the degree of G([n]\{x, y})
so that the remaining subgraph behave as G(n, p) with p = Θ( logn n ). This can be achieved for
n
example either by considering a random subgraph of G according to G([n]\{x, y}, 40 log
), applying
n
Chernoff bounds as in Lemma 1 to show that each node has degree at least 5 log n, or by letting
each node choose 5 log n neighbors uniformly at random.
Finally, once trees Tx , Ty have been constructed, we grow trees from the leaves of Tx and Ty
using BFS for depth γ = ( 12 + )L. Now we analyze these processes. Since the argument is the same
we explain it in detail for Tx and we outline the differences for the other trees. We use the notation
(ρ)
Di for the number of vertices at depth i of the BFS tree rooted at ρ.
First we grow Tx . As we grow the tree via BFS from a vertex v at depth i to vertices at depth
i + 1 certain bad edges from v may point to vertices already in Tx . Lemma 3 shows with probability
1 − o(n−3 ) there can be at most 10 bad edges emanating from v.
Hence, we obtain the recursion
(x)
(x)
Di+1 ≥ (5 log n − 10) (Di
(x)
− 1) ≥ 4 log nDi .
(8)
Therefore the number of leaves satisfies
(x)
Dk ≥
L
1
1
≥
4
log
n
n4/5 .
(2c)L
(2c)L
(9)
We can make the branching factors exactly 4 log n(2c)−L for the first level and 4 log n for all remaining
levels by pruning. We do this so that the trees Tx are isomorphic to each other. With a similar
argument
4
1
(y)
n 5 .
Dk ≥
(10)
L
(2c)
The only difference is that now we also say an edge is bad if the other endpoint is in Tx . This
immediately gives
(y)
(y)
(y)
Di+1 ≥ (5 log n − 20) (Di − 1) ≥ 4 log nDi
and the required conclusion (10).
Similarly, from each leaf xi ∈ Tx and yi ∈ Ty we grow trees Tbxi , Tbyi of depth γ = 21 + L using
the same procedure and arguments as above. Lemma 3 implies that there are at most 20 edges from
the vertex v being explored to vertices in any of the trees already constructed (at most 10 to Tx
plus any trees rooted at an xi and another 10 for y). The number of leaves of each Tbxi now satisfies
1
4
b (xi ) ≥ (4 log n)γ+1 ≥ n 2 + 5 .
D
γ
(y )
bγ i .
The result is similar for D
Observe next that BFS does not condition on the edges between the leaves Xi , Yi of the trees Tbxi
and Tbyi . That is, we do not need to look at these edges in order to carry out our construction. On
the other hand we have conditioned on the occurrence of certain events to imply a certain growth
rate. We handle this technicality as follows. We go through the above construction and halt if ever
we find that we cannot expand by the required amount. Let A be the event that we do not halt the
construction i.e. we fail the conditions of Lemmas 2 or 3. We have Pr [A] = 1 − o(1) and so,
Pr [∃i : e(Xi , Yi ) = 0 | A] ≤
1+ 8
4
Pr [∃i : e(Xi , Yi ) = 0]
≤ 2n 5 (1 − p)n 5 ≤ n−n .
Pr(A)
7
(a)
(b)
Figure 1: Illustration of construction in Lemma 4. (a) By repeatedly growing trees appropriately,
(b) we create for each pair of nodes x, y two node disjoint trees Tx , Ty of depth k = L whose leaves
can be matched via a natural isomorphism and linked with edge disjoint paths of length (1 + o(1))L
We conclude that whp there is always an edge between each Xi , Yi and thus a path of length at
most (1 + 2)L between each xi , yi .
The proof of Theorem 1 follows. Set q =
1
2
− c.
Proof of Theorem 1. Fix a pair of nodes x, y ∈ V (G). Let Y1 , . . . , YN be the signs of the N edge
P
disjoint paths connecting them, i.e., Yi ∈ {−1, +1} for all i. Also let Y = N
i=1 Yi . Notice that
N
{Y1 , . . . , YN } is a read-k family where k = 4 log n(2c)
.
−L
By the linearity of expectation
4
E [Y ] = N (2c)L ≥ n 5 (2c)L .
By applying Theorem 5 we obtain
n4/5 (2c)L
= o(n−2 ).
Pr [Y < 0] = Pr [Y − E [Y ] < −E [Y ]] ≤ exp −
4/5
2n
4(2c)−L log n
Planted bisection model. Before we prove Theorem 2 we discuss the connection between our
formulation and the well studied planted bisection model. Suppose that n is even, and the graph
has two clusters of equal size. The probabilities of connecting are p within each cluster, and q < p
across the clusters. Now recall our setting as described in Model I, and consider just the edges
that correspond to queries that return +1. These form a graph drawn from the planted bisection
40 log n
40 log n
model where p = 1+c
, q = 1−c
. Therefore, one can apply existing methods for
2 ×
n
2 ×
n
exact recovery, e.g., [ABH16; McS01] instead of our method when the sizes of the two clusters
are (roughly) equal. It is worth noting that despite the wide variety of techniques that appear in
the context of the planted partition model, including the EM algorithm [SN97], spectral methods
[McS01; Vu14], semidefinite programming [ABH16; HWX16; MS15], hill-climbing [CI01], Metropolis
algorithm [JS98], modularity based methods [BC09], our edge-disjoint path technique is novel in
this context.
Hajek, Wu, and Xu proved that when each cluster has Θ(n) nodes, the average degree has
log n
√ 2 for exact recovery [HWX16]. Also, they showed that using semidefinite
to scale as (√1−q−
q)
8
programming (SDP) exact recovery is achievable at this threshold [HWX16]. Note that as q → 12 ,
log n
log n
this lower bound scales as Θ( (1−2q)
2 ) = Θ( (2c)2 ). It is an interesting theoretical problem to explore
if the techniques we develop in this work, or similar techniques can get closer to this lower bound.
Proof of Theorem 2. Since the proof of Theorem 2 overlaps with the proof of Theorem 1, we outline
the main differences. Let us return to the basic version of Model II, and let X(e) ∈ {−1, 0, 1} for
e = (x, y) be
f˜(e) − (g(x) − g(y)) mod k.
Then given a path between two vertices u and v,
g(v) = g(u) +
X
f˜(e) −
e∈Puv
X
X(e) mod k.
e∈Puv
P
Our question is now what is Zuv = e∈Puv X(e) mod k. We would like that Zuv be (even slightly)
more highly concentrated on 0 than on other values, so that when g(u) = g(v), we find that the
P
sum of the return values from our algorithm, e∈Puv f˜(e) mod k, is most likely to be 0. We could
then conclude by looking over many almost edge-disjoint paths that if this sum is 0 over a plurality
of the paths, then u and v are in the same group whp.
P
For our simple error model, the sum e∈Puv X(e) mod k behaves like a simple lazy random walk
on the cycle of values modulo k, where the probability of remaining in the same state at each step
is q. Let us consider this Markov chain on the values modulo k; we refer to the values as states.
Let ptij be the probability of going from state i to state j after t steps in such a walk. It is well
known than one can derive explicit formulae for ptij ; see e.g. [Fel68, Chapter XVI.2]. It also follows
by simply finding the eigenvalues and eigenvectors of the matrix corresponding to the Markov chain
and using that representation. One can check the resulting forms to determine that pt0j is maximized
when j = 0, and to determine the corresponding gap maxj∈[1,k−1] |pt00 − p0j |t . Based on this gap, we
can apply Chernoff-type bounds as in Theorem 5 to show that the plurality of almost edge-disjoint
paths will have error 0, allowing us to determine whether the endpoints of the path x and y are in
the same group with high probability.
The simplest example is with k = 3 groups, where we find
pt00 =
1 2
+ (1 − 3q/2)t ,
3 3
and
1 1
− (1 − 3q/2)t .
3 3
In our case t = L, and we see that for any q < 2/3, pt00 is large enough that we can detect paths
using the same argument as in Model I.
For general k, we use that the eigenvalues of the matrix
pt01 = pt02 =
1−q
q
0 ...
1 − q q ...
q
.
..
..
..
.
.
.
.
.
q
0
0 ...
q
0
..
.
1−q
are 1−q+q cos(2πj/k), j = 0, . . . , k−1, with the j-th corresponding eigenvector being [1, ω j , ω 2j , . . . , ω j(k−1) ]
where ω = e2πi/k is a primitive kth root of unity. Here, i is not an index but the square root of -1,
9
i.e., i =
√
−1. In this case we have
pt00 =
X
t
1 1 k−1
+
1 − q + q cos(2πj/k) .
k k j=1
Note that pt00 > 1/k. Some algebra reveals that the next largest value of pt0j belongs to pt01 , and
equals
X
t
1 1 k−1
pt01 = +
ω −j 1 − q + q cos(2πj/k) .
k k j=1
We therefore see that the error between ends of a path again have the plurality value 0, with a gap
of at least
pt00 − pt01 ≥ 2(1 − cos(2π/k))(1 − q + q cos(2π/k))t .
This gap is constant for any constant k ≥ 3 and q ≤ 1/2.
As we have already mentioned, the same approach could be used for the more general setting where
f˜(e) = g(x) − g(y) + j with probability qj , 0 ≤ j < k,
but now one works with the Markov chain matrix
q0
qk−1
.
.
.
q1
4
q1 q2 . . .
q0 q1 . . .
..
.. . .
.
.
.
q2 q3 . . .
qk−1
qk−2
..
.
.
q0
Related Work
Fritz Heider introduced the notion of a signed graph, with +1 or −1 labels on the edges, in the
context of balance theory [Hei46]. The key subgraph in balance theory is the triangle: any set
of three fully interconnected nodes whose product of edge signs is negative is not balanced. The
complete graph is balanced if every one of its triangles is balanced. Early work on signed graphs
focused on graph theoretic properties of balanced graphs [CH56]. Harary proved the famous balance
theorem which characterizes balanced graphs [Har53].
Bansal et al. [BBC04] studied Correlation Clustering: given an undirected signed graph partition the nodes into clusters so that the total number of disagreements is minimized. This problem
is NP-hard [BBC04; SST04]. Here, a disagreement can be either a positive edge between vertices in two clusters or a negative edge between two vertices in the same cluster. Note that in
Correlation Clustering the number of clusters is not specified as part of the input. The case when
the number of clusters is constrained to be at most two is known as 2-Correlation-Clustering.
Minimizing disagreements is equivalent to maximizing the number of agreements. However, from an approximation perspective these two versions are different: minimizing is harder.
For minimizing disagreements, Bansal et al. [BBC04] provide a 3-approximation algorithm for
2-Correlation-Clustering, and Giotis and Guruswami provide a polynomial time approximation
scheme (PTAS) [GG06]. Ailon et al. designed a 2.5-approximation algorithm [ACN08] that was
further improved by Coleman et al. to a 2-approximation [CSW08]. We remark that the notion of
imbalance studied by Harary is the 2-Correlation-Clustering cost of the signed graph. Mathieu and
10
Schudy initiated the study of noisy correlation clustering [MS10]. They develop various algorithms
when the graph is complete, both for the cases of a random and a semi-random model. Later,
Makarychev, Makarychev, and Vijayaraghavan proposed an algorithm for graps with O(npoly log n)
edges under a semi-random model [MMV15].
When the graph is not complete Correlation Clustering and Minimum Multicut reduce to
one another leading to a O(log n) approximations [CGW03; DI03]. The case of constrained size
clusters has recently been studied by Puleo and Mileknovic [PM15]. Finally, by using the GoemansWilliamson SDP relaxation for Max Cut [GW95], one can obtain a 0.878 approximation guarantee
for 2-Correlation-Clustering problem as noted by [CSW08].
Chen et al. [CJS+14; CSX12] consider also model I as described in Section 1 and provide a
method that can reconstruct the clustering for random binomial graphs with O(npoly log n edges.
Their method exploits low rank properties of the cluster matrix, and requires certain conditions,
including conditions on the imbalance between clusters, see [CSX12, Theorem 1, Table 1] to be true.
Their methods is based on a convex relaxation of a low rank problem. Also, closely related to our
work lies the work of Cesa-Bianchi et al. [CBGV+12] that take a learning-theoretic perspective
on the problem of predicting signs. They consider three types of models: batch, online, and active
learning, and provide theoretical bounds for prediction mistakes for each setting. They use the
correlation clustering objective as their learning bias, and they show that the risk of the empirical
risk minimizer is controlled by the correlation clustering objective. Chian et al. point out that the
work of Candès and Tao [CRT06] can be used to predict signs of edges, and also provide various other
methods, including singular value decomposition based methods, for the sign prediction problem
[CHN+14]. The incoherence is the key parameter that determines the number of queries, and is
n
equal to the group imbalance τ = max |C|
. The number of queries needed for exact recovery
cluster C
under Model I is O(τ 4 n log2 n).
5
Conclusion
In this work we have studied the edge sign prediction problem, showing that under our proposed
correlation clustering formulation and a fully random noise model querying O(n1+o(1) log n) pairs
of nodes uniformly at random suffices to recover the clusters efficiently, whp. We also provided a
generalization of our model and proof approach to more than two clusters. While our work here is
theoretical, in future work we plan to apply our method and additional heuristics to real data, and
compare against related alternatives. From a theoretical perspective, it is an interesting problem
log n
to close the gap between our upper bound and the known (2c)
2 lower bound for exact recovery
[HWX16] using techniques based on BFS.
Acknowledgment
We would like to thank Bruce Hajek and Zeyu Zhou for detecting an error in an earlier version
of our work. We also want to thank Yury Makarychev, Konstantin Makarychev, and Aravindan
Vijayaraghavan for their feedback.
This work was supported in part by NSF grants CNS-1228598, CCF-1320231, CCF-1563710,
and CCF-1535795.
11
References
[ABH16]
Emmanuel Abbe, Afonso S Bandeira, and Georgina Hall. “Exact recovery in the
stochastic block model”. In: IEEE Transactions on Information Theory 62.1 (2016),
pp. 471–487 (cit. on pp. 1, 8).
[ACN08]
Nir Ailon, Moses Charikar, and Alantha Newman. “Aggregating inconsistent information: ranking and clustering”. In: Journal of the ACM (JACM) 55.5 (2008), p. 23
(cit. on p. 10).
[AS04]
Noga Alon and Joel H Spencer. The probabilistic method. John Wiley & Sons, 2004
(cit. on p. 4).
[BBC04]
Nikhil Bansal, Avrim Blum, and Shuchi Chawla. “Correlation clustering”. In: Machine
Learning 56.1-3 (2004), pp. 89–113 (cit. on p. 10).
[BC09]
Peter J Bickel and Aiyou Chen. “A nonparametric view of network models and
Newman–Girvan and other modularities”. In: Proceedings of the National Academy
of Sciences 106.50 (2009), pp. 21068–21073 (cit. on p. 8).
[BFS+98]
Andrei Z Broder, Alan M Frieze, Stephen Suen, and Eli Upfal. “Optimal construction
of edge-disjoint paths in random graphs”. In: SIAM Journal on Computing 28.2
(1998), pp. 541–573 (cit. on p. 6).
[Bol98]
Béla Bollobás. “Random graphs”. In: Modern Graph Theory. Springer, 1998 (cit. on
p. 4).
[CBGV+12] Nicolo Cesa-Bianchi, Claudio Gentile, Fabio Vitale, Giovanni Zappella, et al. “A
Correlation Clustering Approach to Link Classification in Signed Networks.” In:
COLT. 2012, pp. 34–1 (cit. on p. 11).
[CGW03]
Moses Charikar, Venkatesan Guruswami, and Anthony Wirth. “Clustering with qualitative information”. In: Proceedings. 44th Annual IEEE Symposium on Foundations
of Computer Science (FOCS). IEEE. 2003, pp. 524–533 (cit. on p. 11).
[CH56]
Dorwin Cartwright and Frank Harary. “Structural balance: a generalization of Heider’s
theory.” In: Psychological review 63.5 (1956), p. 277 (cit. on pp. 1, 10).
[CHN+14]
Kai-Yang Chiang, Cho-Jui Hsieh, Nagarajan Natarajan, Inderjit S Dhillon, and Ambuj
Tewari. “Prediction and clustering in signed networks: a local to global perspective.”
In: Journal of Machine Learning Research 15.1 (2014), pp. 1177–1213 (cit. on p. 11).
[CI01]
Ted Carson and Russell Impagliazzo. “Hill-climbing finds random planted bisections”.
In: Proceedings of the twelfth annual ACM-SIAM Symposium on Discrete Algorithms
(SODA). Society for Industrial and Applied Mathematics. 2001, pp. 903–909 (cit. on
p. 8).
[CJS+14]
Yudong Chen, Ali Jalali, Sujay Sanghavi, and Huan Xu. “Clustering partially observed
graphs via convex optimization.” In: Journal of Machine Learning Research 15.1
(2014), pp. 2213–2238 (cit. on p. 11).
[CK01]
Anne Condon and Richard M Karp. “Algorithms for graph partitioning on the planted
partition model”. In: Random Structures and Algorithms 18.2 (2001), pp. 116–140
(cit. on p. 1).
12
[CRT06]
Emmanuel J Candès, Justin Romberg, and Terence Tao. “Robust uncertainty principles: Exact signal reconstruction from highly incomplete frequency information”. In:
IEEE Transactions on information theory 52.2 (2006), pp. 489–509 (cit. on p. 11).
[CSW08]
Tom Coleman, James Saunderson, and Anthony Wirth. “A local-search 2-approximation
for 2-correlation-clustering”. In: European Symposium on Algorithms (ESA). Springer,
2008, pp. 308–319 (cit. on pp. 10, 11).
[CSX12]
Yudong Chen, Sujay Sanghavi, and Huan Xu. “Clustering sparse graphs”. In: Advances
in neural information processing systems. 2012, pp. 2204–2212 (cit. on p. 11).
[DFT15]
Andrzej Dudek, Alan M Frieze, and Charalampos E Tsourakakis. “Rainbow connection
of random regular graphs”. In: SIAM Journal on Discrete Mathematics 29.4 (2015),
pp. 2255–2266 (cit. on p. 6).
[DI03]
Erik D Demaine and Nicole Immorlica. “Correlation clustering with partial information”. In: Approximation, Randomization, and Combinatorial Optimization
(APPROX-RANDOM). Springer, 2003, pp. 1–13 (cit. on p. 11).
[EK10]
David Easley and Jon Kleinberg. Networks, crowds, and markets: Reasoning about a
highly connected world. Cambridge University Press, 2010 (cit. on p. 1).
[FK15]
Alan Frieze and Michał Karoński. Introduction to random graphs. Cambridge University Press, 2015 (cit. on p. 4).
[FT12]
Alan Frieze and Charalampos E Tsourakakis. “Rainbow connectivity of sparse random graphs”. In: Approximation, Randomization, and Combinatorial Optimization
(APPROX-RANDOM). Springer, 2012, pp. 541–552 (cit. on p. 6).
[Fel68]
William Feller. An introduction to probability theory and its applications: volume I.
Vol. 3. John Wiley & Sons London-New York-Sydney-Toronto, 1968 (cit. on p. 9).
[GG06]
Ioannis Giotis and Venkatesan Guruswami. “Correlation clustering with a fixed
number of clusters”. In: Proceedings of the seventeenth annual ACM-SIAM Symposium
on Discrete Algorithms (SODA). Society for Industrial and Applied Mathematics.
2006, pp. 1167–1176 (cit. on p. 10).
[GLS+15]
Dmitry Gavinsky, Shachar Lovett, Michael Saks, and Srikanth Srinivasan. “A tail
bound for read-k families of functions”. In: Random Structures & Algorithms 47.1
(2015), pp. 99–108 (cit. on pp. 3, 4).
[GW95]
Michel X Goemans and David P Williamson. “Improved approximation algorithms
for maximum cut and satisfiability problems using semidefinite programming”. In:
Journal of the ACM (JACM) 42.6 (1995), pp. 1115–1145 (cit. on p. 11).
[HEP+16]
Jack P Hou, Amin Emad, Gregory J Puleo, Jian Ma, and Olgica Milenkovic. “A
new correlation clustering method for cancer mutation analysis”. In: arXiv preprint
arXiv:1601.06476 (2016) (cit. on p. 1).
[HWX16]
Bruce Hajek, Yihong Wu, and Jiaming Xu. “Achieving exact cluster recovery threshold
via semidefinite programming”. In: IEEE Transactions on Information Theory 62.5
(2016), pp. 2788–2797 (cit. on pp. 1, 8, 9, 11).
[Har53]
Frank Harary. “On the notion of balance of a signed graph.” In: The Michigan
Mathematical Journal 2.2 (1953), pp. 143–146 (cit. on pp. 1, 10).
13
[Hei46]
Fritz Heider. “Attitudes and cognitive organization”. In: The Journal of psychology
21.1 (1946), pp. 107–112 (cit. on pp. 1, 10).
[JLR11]
Svante Janson, Tomasz Luczak, and Andrzej Rucinski. Random Graphs. Vol. 45. John
Wiley & Sons, 2011 (cit. on p. 3).
[JS98]
Mark Jerrum and Gregory B Sorkin. “The Metropolis algorithm for graph bisection”.
In: Discrete Applied Mathematics 82.1 (1998), pp. 155–175 (cit. on p. 8).
[KV00]
Jeong Han Kim and Van H Vu. “Concentration of multivariate polynomials and its
applications”. In: Combinatorica 20.3 (2000), pp. 417–434 (cit. on p. 4).
[Knu07]
Donald E Knuth. “Seminumerical algorithms”. In: (2007) (cit. on p. 4).
[LHK10a]
Jure Leskovec, Daniel Huttenlocher, and Jon Kleinberg. “Predicting positive and
negative links in online social networks”. In: Proceedings of the 19th international
conference on World Wide Web (WWW). ACM. 2010, pp. 641–650 (cit. on p. 1).
[LHK10b]
Jure Leskovec, Daniel Huttenlocher, and Jon Kleinberg. “Signed networks in social
media”. In: Proceedings of the SIGCHI conference on Human Factors in Computing
Systems. ACM. 2010, pp. 1361–1370 (cit. on p. 1).
[MMV15]
Konstantin Makarychev, Yury Makarychev, and Aravindan Vijayaraghavan. “Correlation clustering with noisy partial information”. In: Proceedings of the Conference
on Learning Theory (COLT). Vol. 6. 2015, p. 12 (cit. on p. 11).
[MS10]
Claire Mathieu and Warren Schudy. “Correlation clustering with noisy input”. In:
Proceedings of the twenty-first annual ACM-SIAM symposium on Discrete Algorithms.
Society for Industrial and Applied Mathematics. 2010, pp. 712–728 (cit. on p. 11).
[MS15]
Andrea Montanari and Subhabrata Sen. “Semidefinite programs on sparse random graphs and their application to community detection”. In: arXiv preprint
arXiv:1504.05910 (2015) (cit. on p. 8).
[McD98]
Colin McDiarmid. “Concentration”. In: Probabilistic methods for algorithmic discrete
mathematics. Springer, 1998, pp. 195–248 (cit. on p. 3).
[McS01]
Frank McSherry. “Spectral partitioning of random graphs”. In: Proceedings. 42nd
IEEE Symposium on Foundations of Computer Science (FOCS). IEEE. 2001, pp. 529–
537 (cit. on pp. 1, 8).
[PM15]
Gregory J Puleo and Olgica Milenkovic. “Correlation clustering with constrained
cluster sizes and extended weights bounds”. In: SIAM Journal on Optimization 25.3
(2015), pp. 1857–1872 (cit. on p. 11).
[SN97]
Tom AB Snijders and Krzysztof Nowicki. “Estimation and prediction for stochastic
blockmodels for graphs with latent block structure”. In: Journal of classification 14.1
(1997), pp. 75–100 (cit. on p. 8).
[SST04]
Ron Shamir, Roded Sharan, and Dekel Tsur. “Cluster graph modification problems”.
In: Discrete Applied Mathematics 144.1 (2004), pp. 173–182 (cit. on p. 10).
[TKM11]
Charalampos E Tsourakakis, Mihail N Kolountzakis, and Gary L Miller. “Triangle
Sparsifiers.” In: J. Graph Algorithms Appl. 15.6 (2011), pp. 703–726 (cit. on p. 4).
[Tso13]
Charalampos E. Tsourakakis. “Mathematical and Algorithmic Analysis of Network
and Biological Data”. PhD thesis. Carnegie Mellon University, 2013 (cit. on p. 6).
14
[Vu14]
Van Vu. “A simple SVD algorithm for finding hidden partitions”. In: arXiv preprint
arXiv:1404.3918 (2014) (cit. on p. 8).
15
| 8cs.DS
|
Under review as a conference paper at ICLR 2018
E VIDENCE AGGREGATION FOR A NSWER R E -R ANKING
IN O PEN -D OMAIN Q UESTION A NSWERING
arXiv:1711.05116v1 [cs.CL] 14 Nov 2017
Shuohang Wang1 , Mo Yu2 , Jing Jiang1 ,Wei Zhang2 , Xiaoxiao Guo2 , Shiyu Chang2 , Zhiguo Wang2
Tim Klinger2 , Gerald Tesauro2 and Murray Campbell2
1
School of Information System, Singapore Management University
2
AI foundations, IBM Research. Yorktown Heights NY, USA
[email protected], [email protected], [email protected]
A BSTRACT
A popular recent approach to answering open-domain questions is to first search
for question-related passages and then apply reading comprehension models to
extract answers. Existing methods usually extract answers from single passages
independently. But some questions require a combination of evidence from across
different sources to answer correctly. In this paper, we propose two models which
make use of multiple passages to generate their answers. Both use an answerreranking approach which reorders the answer candidates generated by an existing state-of-the-art QA model. We propose two methods, namely, strengthbased re-ranking and coverage-based re-ranking, to make use of the aggregated
evidence from different passages to better determine the answer. Our models
have achieved state-of-the-art results on three public open-domain QA datasets:
Quasar-T, SearchQA and the open-domain version of TriviaQA, with about 8 percentage points of improvement over the former two datasets.
1
I NTRODUCTION
Open-domain question answering (QA) aims to answer questions from a broad range of domains
by effectively marshalling evidence from large open-domain knowledge sources. Such resources
can be Wikipedia (Chen et al., 2017), the whole web (Ferrucci et al., 2010), structured knowledge
bases (Berant et al., 2013; Yu et al., 2017) or combinations of the above (Baudiš & Šedivỳ, 2015).
Recent work on open-domain QA has focused on using unstructured text retrieved from the web
to build machine comprehension models (Chen et al., 2017; Dhingra et al., 2017b; Wang et al.,
2017). These studies adopt a two-step process: an information retrieval (IR) model to coarsely
select relevant passages to a question, followed by a reading comprehension (RC) model (Wang
& Jiang, 2017; Seo et al., 2017; Chen et al., 2017) to infer an answer from the passages. These
studies have made progress in bringing together evidence from large data sources, but they predict
an answer to the question with only a single retrieved passage at a time. However, answer accuracy
can often be improved by using multiple passages. In some cases, the answer can only be determined
by combining multiple passages.
In this paper, we propose a method to improve open-domain QA by explicitly aggregating evidence
from across multiple passages. Our method is inspired by two notable observations from previous
open-domain QA results analysis:
• First, compared with incorrect answers, the correct answer is often suggested by more passages
repeatedly. For example, in Figure 1(a), the correct answer “danny boy” has more passages
providing evidence relevant to the question compared to the incorrect one. This observation can
be seen as multiple passages collaboratively enhancing the evidence for the correct answer.
• Second, sometimes the question covers multiple answer aspects, which spreads over multiple
passages. In order to infer the correct answer, one has to find ways to aggregate those multiple
passages in an effective yet sensible way to try to cover all aspects. In Figure 1(b), for example,
the correct answer “Galileo Galilei” at the bottom has passages P1, “Galileo was a physicist ...”
1
Under review as a conference paper at ICLR 2018
Question1: What is the more popular name for the
londonderry air？
A1: tune from county
Question2: Which physicist , mathematician and astronomer
discovered the first 4 moons of Jupiter
A1: Isaac Newton
P1: the best known title for this melody is londonderry air lrb- sometimes also called the tune from county derry -rrb- .
P1: Sir Isaac Newton was an English physicist , mathematician ,
astronomer , natural philosopher , alchemist and theologian …
P2: Sir Isaac Newton was an English mathematician, astronomer, and
physicist who is widely recognized as one of the most influential
scientists …
A2: danny boy
P1: londonderry air words : this melody is more commonly
known with the words `` danny boy ''
P2: londonderry air danny boy music ftse london i love you .
P3: danny boy limavady is most famous for the tune
londonderry air collected by jane ross in the mid-19th century
from a local fiddle player .
P4: it was here that jane ross noted down the famous
londonderry air -lrb- ` danny boy ' -rrb- from a passing fiddler .
Question2: Which physicist , mathematician and astronomer
discovered the first 4 moons of Jupiter
A2: Galileo Galilei
P1: Galileo Galilei was an Italian physicist , mathematician ,
astronomer , and philosopher who played a major role in the Scientific
Revolution .
P2: Galileo Galilei is credited with discovering the first four moons
of Jupiter .
(a)
(b)
Figure 1: Two examples of questions and candidate answers. (a) A question benefiting from the repetition of
evidence. Correct answer A2 has multiple passages that could support A2 as answer. The wrong answer A1
has only a single supporting passage. (b) A question benefiting from the union of multiple pieces of evidence
to support the answer. The correct answer A2 has evidence passages that can match both the first half and the
second half of the question. The wrong answer A1 has evidence passages covering only the first half.
and P2, “Galileo discovered the first 4 moons of Jupiter”, mentioning two pieces of evidence to
match the question. In this case, the aggregation of these two pieces of evidence can help entail
the ground-truth answer “Galileo Galilei”. In comparison, the incorrect answer “Isaac Newton”
has passages providing partial evidence on only “physicist, mathematician and astronomer”.
This observation illustrates the way in which multiple passages may provide complementary
evidence to better infer the correct answer to a question.
To provide more accurate answers for open-domain QA, we hope to make better use of multiple
passages for the same question by aggregating both the strengthened and the complementary evidence from all the passages. We formulate the above evidence aggregation as an answer re-ranking
problem. Re-ranking has been commonly used in NLP problems, such as in parsing and translation, in order to make use of high-order or global features that are too expensive for decoding
algorithms (Collins & Koo, 2005; Shen et al., 2004; Huang, 2008; Dyer et al., 2016). Here we apply
the idea of re-ranking; for each answer candidate, we efficiently incorporate global information from
multiple pieces of textual evidence without significantly increasing the complexity of the prediction
of the RC model. Specifically, we first collect the top-K candidate answers based on their probabilities computed by a standard RC/QA system, and then we use two proposed re-rankers to re-score
the answer candidates by aggregating each candidate’s evidence in different ways. The re-rankers
are:
• A strength-based re-ranker, which ranks the answer candidates according to how often their
evidence occurs in different passages. The re-ranker is based on the first observation if an answer
candidate has multiple pieces of evidence, and each passage containing some evidence tends to
predict the answer with a relatively high score (although it may not be the top score), then the
candidate is more likely to be correct. The passage count of each candidate, and the aggregated
probabilities for the candidate, reflect how strong its evidence is, and thus in turn suggest how
likely the candidate is the corrected answer.
• A coverage-based re-ranker, which aims to rank an answer candidate higher if the union of all its
contexts in different passages could cover more aspects included in the question. To achieve this,
for each answer we concatenate all the passages that contain the answer together. The result is a
new context that aggregates all the evidence necessary to entail the answer for the question. We
then treat the new context as one sequence to represent the answer, and build an attention-based
match-LSTM model (Wang & Jiang, 2017) between the sequence and the question to measure
how well the new aggregated context could entail the question.
Overall, our contributions are as follows: 1) We propose a re-ranking-based framework to make use
of the evidence from multiple passages in open-domain QA, and two re-rankers, namely, a strength2
Under review as a conference paper at ICLR 2018
based re-ranker and a coverage-based re-ranker, to perform evidence aggregation in existing opendomain QA datasets. We find the second re-ranker performs better than the first one on two of the
three public datasets. 2) Our proposed approach leads to the state-of-the-art results on three different
datasets (Quasar-T (Dhingra et al., 2017b), SearchQA (Dunn et al., 2017) and TriviaQA (Joshi
et al., 2017)) and outperforms previous state of the art by large margins. In particular, we achieved
up to 8% improvement on F1 on both Quasar-T and SearchQA compared to the previous best results.
2
M ETHOD
Given an open-domain question q, the problem we are solving is to find the correct answer ag to q
based on all the information from the web. To retrieve all the contexts most related to the question,
an IR model (with the help of a search engine such as Google or Bing) is used to provide the top-N
candidate passages, p1 , p2 , . . . , pN , from the web. Then a reading comprehension (RC) model is
used to extract the answer from these passages.
This setting is different from closed-domain QA (Rajpurkar et al., 2016), where a single fixed passage is given, from which the answer is to be extracted. When developing a closed-domain QA
system, we can use the specific positions of the answer sequence in the given passage for training.
In comparison, in the open-domain setting, the RC models are usually trained under distant supervision (Chen et al., 2017; Dhingra et al., 2017b; Joshi et al., 2017). Specifically, as the training data
does not label the positions of the answer spans in passages, during the training stage, the RC model
will match all passages that contain the ground-truth answer with the question one by one. Throughout the paper, we will first directly apply an existing RC model called R3 (Wang et al., 2017) to
extract the candidate answers.
After the candidate answers are extracted, we start to aggregate evidence from multiple passages
for better answer prediction by adopting a re-ranking approach over answer candidates. Given a
question q, suppose we have a baseline open-domain QA system that could generate the top-K
answer candidates a1 , . . . , aK , each being a text span in some passage pi . The goal of the re-ranker
is to rank this list of candidates such that the top-ranked candidates have more chance to be the
correct answer ag .
This re-ranking step considers more powerful features and more context information that cannot
be easily handled by the baseline system. In this paper, we investigate two types of such more
powerful information based on the multiple pieces of evidence for each answer: evidence strength
and evidence coverage, as aggregated from multiple passages. An overview of our method is shown
in Figure 2.
2.1
E VIDENCE AGGREGATION FOR S TRENGTH - BASED R E - RANKER
For open-domain QA, unlike the closed-domain setting, we have more passages retrieved by the IR
model and the ground-truth answer may appear in different passages, which means different answer
spans may be referring to the same answer. Considering this property, we provide two features to
further re-rank the top-K answers generated by the RC model.
Measuring Strength by Count For this re-ranking method, we hypothesize that the more passages we have that could entail a particular answer for the question, the stronger the evidence is for
that answer. In the top-K answer spans generated by the RC model, each time an answer appears
in the top-K answer list, we increase the counter for that answer by one. In this way, in the end the
answer with the highest count will be predicted to be the final answer.
Measuring Strength by Probability Since we can get the probability of each answer span in a
passages based on the RC model, we can also sum up the probabilities of the answer spans that are
referring to the same answer. Then the answer with the highest probability would be the final prediction 1 . In the re-ranking scenario, it is not necessary to exhaustively consider all the probabilities
of all the spans in the passages, as there may be a large number of different answer spans and most
of them are irrelevant to the ground-truth answer.
1
This is an extension of the Attention Sum method in (Kadlec et al., 2016) from single-token answers to
phrase answers.
3
Under review as a conference paper at ICLR 2018
Figure 2: An overview of the full re-ranker. It consists of strength-based and coverage-based reranking.
Remark: Note that neither method above requires new model training. Both methods just take
the candidate predictions from the baseline QA system and then perform counting or probability
calculation. During testing the time complexity of strength-based re-ranking is negligible.
2.2
E VIDENCE AGGREGATION FOR C OVERAGE - BASED R E - RANKER
This section focuses on another type of evidence aggregation which needs to take the union of
complementary information from different evidence passages. As shown in Figure 1, the two answer
candidates both have evidence matching the first half of the question. However, only the correct
answer has evidence that could also match the second half. In this case, the strength-based re-ranker
will treat both answer candidates the same due to the equal amount of supporting evidence, while
actually the second answer has complementary evidence satisfying all aspects of the question. To
handle such case, we propose a coverage-based re-ranker that could rank the answer candidates
according to whether the unions of their evidence from different passages could well cover the
question.
In order to take the union of evidence into consideration, we first concatenate the passages containing
the answer into a single pseudo passage. Then we measure how well the new passage could entail
the answer for the question. As examples shown in Figure 1(b), we hope the textual entailment
model could reflect (i) how each aspect of the question is matched by the union of multiple passages,
where we simply define the aspect to be one hidden state of bi-directional LSTM; (ii) whether all the
aspects of the question can be matched by the union of multiple passages. The match-LSTM (Wang
& Jiang, 2016) model is one way to achieve the above effect in entailment. Therefore we build our
coverage-based re-ranker on top of the concatenated pseudo passages through match-LSTM. The
detailed method is described below.
Passage Aggregation We consider the top-K answers, a1 , . . . , aK , provided by the baseline QA
system. For each answer ak , k ∈ [1, K], we concatenate all the passages that contain ak , {pn |ak ∈
pn , n ∈ [1, N ]}, to form the union passage p̂k . Our further model is to identify which union
passage, e.g., p̂k , could better entail its answer, e.g., ak , for the question.
Measuring Aspect(Word)-Level Matching As discussed earlier, the first mission of the
coverage-based re-ranker is to measure how each aspect of the question is matched by the union
4
Under review as a conference paper at ICLR 2018
of multiple passages. We achieve this with word-by-word attention followed by a comparison module.
First, we write the answer candidate a, question q and the union passage p̂ of a as matrices A, Q, P̂,
with each column being the embedding of a word in the sequence. We then feed them to the bidirectional LSTM as follows:
Ha = BiLSTM(A), Hq = BiLSTM(Q),
Hp = BiLSTM(P̂),
(1)
where Ha ∈ Rl×A , Hq ∈ Rl×Q and Hp ∈ Rl×P are the hidden states for the answer candidate,
question and passage respectively; l is the dimension of the hidden states, and A,Q and P are the
length of the three sequences, respectively.
Next, we enhance the question representation Hq with Ha :
Haq
[Ha ; Hq ],
=
(2)
where [·; ·] is the concatenation of two matrices in row and Haq ∈ Rl×(A+Q) . As most of the answer
candidates do not appear in the question, this is for better matching with the passage and finding
more answer-related information from the passage. Now we can view each aspect of the question
as a column vector (i.e. a hidden state at each word position in the answer-question concatenation)
in the enhanced question representation Haq . Then the task becomes to measure how well each
column vector can be matched by the union passage; and we achieve this by computing the attention
vector Parikh et al. (2016) for each hidden state of sequences a and q as follows:
aq
(3)
α = SoftMax (Hp )T Haq , H = Hp α,
where α ∈ RP ×(A+Q) is the attention weight matrix which is normalized in column through softaq
max. H ∈ Rl×(A+Q) are the attention vectors for each word of the answer and the question by
weighted summing all the hidden states of the passage p̂. Now in order to see whether the aspects
in the question can be matched by the union passage, we use the following matching function:
aq J aq
H
H
m Haq − Haq
+ bm ⊗ e(A+Q) ,
(4)
M = ReLU
aq
W
H
aq
H
J
where · ⊗ e(A+Q) is to repeat the vector (or scalar) on the left A + Q times; (· ·) and (· − ·) are the
element-wise operations for checking whether the word in the answer and question can be matched
by the evidence in the passage. We also concatenate these matching representations with the hidden
aq
state representations Haq and H , so that some stop words or punctuation representations may not
affect the final aspect-level matching representations M ∈ R2l×(A+Q) , which is computed through
the non-linear transformation on four different representations with parameters Wm ∈ R2l×4l and
b m ∈ Rl .
Measuring the Entire Question Matching Next, in order to measure how the entire question is
matched by the union passage p̂ by taking into consideration of the matching result at each aspect,
we add another bi-directional LSTM on top of it to aggregate the aspect-level matching information 2 :
Hm = BiLSTM(M), hs = MaxPooling(Hm ),
(5)
where Hm ∈ Rl×(A+Q) is to denote all the hidden states and hs ∈ Rl , the max of pooling of each
dimension of Hm , is the entire matching representation which reflects how well the evidences in
questions could be matched by the union passage.
2
Note that we use LSTM here to capture the conjunction information (the dependency) among aspects,
i.e. how all the aspects are jointly matched. In comparison simple pooling methods will treat the aspects
independently. Low-rank tensor inspired neural architectures like in (Lei et al., 2017) could be another choice
and we will investigate them in future work.
5
Under review as a conference paper at ICLR 2018
Re-ranking Objective Function Our re-ranking is based on the entire matching representation.
For each candidate answer ak , k ∈ [1, K], we can get a matching representation hsk between the
answer ak , question q and the union passage p̂k through Eqn. (1-5). Then we transform all representations into scalar values followed by a normalization process for ranking as follows:
R = Tanh (Wr [hs1 ; hs2 ; ...; hsK ] + br ⊗ eK ) , o = Softmax(wo R + bo ⊗ eK ),
(6)
where we concatenate the match representations for each answer in row through [·; ·], and do a
non-linear transformation by parameters Wr ∈ Rl×l and br ∈ Rl to get hidden representation
R ∈ Rl×K . Finally, we map the transformed matching representations into scalar values through
parameters wo ∈ Rl and wo ∈ R. o ∈ RK is the normalized probability for the candidate answers
to be ground-truth. Due to the aliases of the ground-truth answer, there may be multiple answers in
the candidates are ground-truth, we use KL distance as our objective function as follows:
K
X
yk (log(yk ) − log(ok )) ,
(7)
k=1
where yk indicates whether ak the ground-truth answer or not and is normalized by
ok is the ranking output of our model for ak .
2.3
PK
k=1
yk and
C OMBINATION OF D IFFERENT T YPES OF AGGREGATIONS
Although the coverage-based re-ranker tries to deal with more difficult cases compared to the
strength-based re-ranker, the strength-based re-ranker works on more common cases according to
the distributions of most open-domain QA datasets. As a result we hope to make the best of both
methods.
We achieve our full re-ranker by weighted combination of the outputs of the above different rerankers without further training. Specifically, we first use softmax to re-normalize the top-5 answer
scores provided by the two strength-based rankers and the one coverage-based re-ranker; we then
weighted sum up the scores for the same answer and select the answer with the largest score as the
final prediction.
3
E XPERIMENTAL S ETTINGS
We conduct experiments on three publicly available open-domain QA datasets, namely, QuasarT (Dhingra et al., 2017b), SearchQA (Dunn et al., 2017) and TriviaQA (Joshi et al., 2017). These
datasets contain passages retrieved for all questions using a search engine such as Google or Bing.
We do not retrieve more passages but use the provided passages only.
3.1
DATASETS
The statistics of the three datasets are shown in Table 1.
Quasar-T 3 (Dhingra et al., 2017b) is based on a trivia question set and uses Lucene to collect
100 sentence-level passages for each question from the ClueWeb09 data source. The human performance is evaluated in an open-book setting, i.e., the human subjects had access to the same passages
retrieved by the IR model and tried to find the answers from the passages.
SearchQA 4 (Dunn et al., 2017) is based on Jeopardy! questions and uses Google to collect about 50
web page snippets as passages for each question. The human performance is evaluated in a similar
way to the Quasar-T dataset.
TriviaQA (Open-Domain Setting) 5 (Joshi et al., 2017) collected trivia questions coming from
14 trivia and quiz-league websites, and makes use of the Bing Web search API to collect the top 50
3
https://github.com/bdhingra/quasar
https://github.com/nyu-dl/SearchQA
5
http://nlp.cs.washington.edu/triviaqa/data/triviaqa-unfiltered.tar.gz
4
6
Under review as a conference paper at ICLR 2018
#q(train) #q(dev) #q(test)
Quasar-T
SearchQA
TriviaQA
28,496
99,811
66,828
3,000
13,893
11,313
#p
3,000 100
27,247 50
10,832 100
#p(truth) #p(aggregated)
14.8
16.5
16.0
5.2
5.4
5.6
Table 1: Statistics of the datasets. #q represents the number of questions for training (not counting
the questions that don’t have ground-truth answer in the corresponding passages for training set),
development, and testing datasets. #p is the number of passages for each question. For TriviaQA,
we split the raw documents into sentence level passages and select the top 100 passages based on the
its overlaps with the corresponding question. #p(golden) means the number of passages that contain
the ground-truth answer in average. #p(aggregated) is the number of passages we aggregated in
average for top 10 candidate answers provided by RC model.
webpages most related to the questions. We focus on the open domain setting (the unfiltered passage
set) of the dataset 6 and our model uses all the information retrieved by the IR model.
3.2
BASELINES
Our baseline models 7 include the following: GA (Dhingra et al., 2017a;b), a reading comprehension
model with gated-attention; BiDAF (Seo et al., 2017), a RC model with bidirectional attention flow;
AQA (Buck et al., 2017), a reinforced system learning to aggregate the answers generated by the
re-written questions; R3 (Wang et al., 2017), a reinforced model making use of a ranker for selecting
passages to train the RC model. As R3 is the first step of our system for generating candidate
answers, the improvement of our re-ranking methods can be directly compared to this baseline.
TriviaQA does not provide the leaderboard under the open-domain setting. As a result, there is no
public baselines in this setting and we only compare with the R3 baseline.8
3.3
I MPLEMENTATION D ETAILS
We first use a pre-trained R3 model (Wang et al., 2017), which gets the state-of-the-art performance
on the three public datasets we consider, to generate the top 50 candidate spans for the training,
development and test datasets, and we use them for further ranking. During training, if the groundtruth answer does not appear in the answer candidates, we will manually add it into the answer
candidate list.
For the coverage-based re-ranker, we use Adamax (Kingma & Ba, 2015) to optimize the model.
Word embeddings are initialized by GloVe (Pennington et al., 2014) and are not updated during
training. We set all the words beyond Glove as zero vectors. We set l to 300, batch size to 30,
learning rate to 0.002. We tune the dropout probability from 0 to 0.5 and the number of candidate
answers for re-ranking (K) in [3, 5, 10] 9 .
4
R ESULTS AND A NALYSIS
In this section, we will show the results of our different re-ranking methods on the three different
public datasets. We will also show some further analysis on our models.
6
Despite the open-domain QA data provided, the leaderboard of TriviaQA focuses on evaluation of RC
models over filtered passages that is guaranteed to contain the correct answers (i.e. more like closed-domain
setting). The evaluation is also passage-wise, different from the open-domain QA setting.
7
Most of the results of different models come from the public paper. While we re-run model R3 (Wang
et al., 2017) based on the authors’ source code and extend the model to the datasets of SearchQA and TriviaQA
datasets.
8
To demonstrate that R3 serves as a strong baseline on the TriviaQA data, we generate the R3 results
following the leaderboard setting. The results showed that R3 achieved F1 56.0, EM 50.9 on Wiki domain and
F1 68.5, EM 63.0 on Web domain, which is competitive to the state-of-the-arts. This confirms that R3 is a
competitive baseline when extending the TriviaQA questions to open-domain setting.
9
Our code will be released.
7
Under review as a conference paper at ICLR 2018
Quasar-T
EM
F1
SearchQA
EM
F1
TriviaQA (open)
EM
F1
GA (Dhingra et al., 2017a)
BiDAF (Seo et al., 2017)
AQA (Buck et al., 2017)
R3 (Wang et al., 2017)
26.4
25.9
35.3
26.4
28.5
41.7
28.6
40.5
49.0
34.6
47.4
55.3
47.3
53.7
Full Re-Ranker
Strength-Based Re-Ranker (Probability)
Strength-Based Re-Ranker (Counting)
Coverage-Based Re-Ranker
42.3
36.1
37.1
40.6
49.6
42.4
46.7
49.1
57.0
50.4
54.2
53.6
63.2
56.5
61.6
60.6
50.6
49.2
46.1
50.0
57.3
55.1
55.8
57.0
Human Performance
51.5
60.6
43.9
-
-
-
Table 2: Experiment results on three open-domain QA test datasets: Quasar-T, SearchQA and TriviaQA (open-domain setting). EM: Exact Match. Full Re-ranker is the combination of three different
re-rankers.
Figure 3: Performance decomposition according to the length of answers and the question types.
4.1
OVERALL R ESULTS
The performance of our models are shown in Table 2. We use F1 score and Exact Match (EM) as our
evaluation metrics 10 . From the results, we can clearly see that our full re-ranker, the combination of
different re-rankers, can significantly outperform the previous best performance with a large margin,
especially on the datasets of Quasar-T and SearchQA. Moreover, our model is much better than the
human performance on the SearchQA dataset. Besides, we can see that our coverage-based re-ranker
generally achieves good performance on the three datasets, even though its performance is 1% lower
than the strength-based re-ranker on the SearchQA dataset.
4.2
A NALYSIS
In this subsection, we analyze the benefits of our re-ranking models.
Re-ranking performance versus answer lengths and question types Figure 3 decomposes the
performance according to the length of the ground truth answers and the types of questions on
TriviaQA and Quasar-T. We do not include the analysis on SearchQA because the Jeopardy! style
10
Our evaluation is based on the tool from SQuAD (Rajpurkar et al., 2016).
8
Under review as a conference paper at ICLR 2018
Top-K
Quasar-T
EM
F1
SearchQA
EM
F1
TriviaQA (open)
EM
F1
1
3
5
10
35.1
46.2
51.0
56.1
51.2
63.9
69.1
75.5
47.6
54.1
58.0
62.1
41.6
53.5
58.9
64.8
57.3
68.9
73.9
79.6
53.5
60.4
64.5
69.0
Table 3: The upper bound (recall) of the Top-K answer candidates generated by the baseline R3
system (on dev set), which indicates the potential of the coverage-based re-ranker.
Candidate Set
Re-Ranker Results
EM
F1
Upper Bound
EM
F1
top-3
top-5
top-10
40.5
41.8
41.3
46.2
51.0
56.1
47.8
50.1
50.8
53.5
58.9
64.8
Table 4: Results of running coverage-based re-ranker on different number of the top-K answer
candidates on Quasar-T (dev set).
questions are more difficult to distinguish the questions types, and the range of answer lengths is
smaller.
According to the results, the coverage-based re-ranker could outperform the baseline in different
lengths of answers and different types of questions. The strength-based re-ranker (counting) also
gives improvement but is less stable across different datasets, while the strength-based re-ranker
(probability) tends to have results and trends that are close to the baseline curves, which is probably
because the method is dominated by the probabilities predicted by the baseline.
The coverage-based re-ranker and the strength-based re-ranker (counting) have similar trends on
most of the question types. The only exception is that the strength-based re-ranker performs significantly worse compared to the coverage-based re-ranker on the “why” questions. This is possibly
because those questions usually have non-factoid answers, which are less likely to have exactly the
same text spans predicted on different passages by the baseline.
Potential improvement of re-rankers Table 3 shows the percentage of times the correct answer is
included in the top-K answer predictions of the baseline R3 method. More concretely, the scores are
computed by selecting the answer from the top-K predictions with the best EM/F1 score. Therefore
the final top-K EM and F1 can be viewed as recall or upper bound of the top-K predictions. From
the results, we can see that although the top-1 prediction of R3 is not very accurate, there is high
chance that a moderate top-K list could cover the correct answer. This explains why our re-ranking
approach could achieve large improvement. Also by comparing the upper bound performance of
top-5 and our re-ranking performance in Table 2, we can see there is still a clear gap of about 10%
on both datasets and on both F1 and EM, showing the great potential improvement for the re-ranking
model in future work.
Effect of the selection of K for the coverage-based re-ranker As shown in Table 3, with the
increase of K from 1 to 10, the recall of top-K predictions from the baseline R3 system increase
significantly. Ideally, if we use a larger K, then the candidate lists will be more likely to contain
good answers. At the same time, the lists to be ranked are longer thus the re-ranking problem is
harder. Therefore, there is a trade-off between the coverage of rank lists and the difficulty of reranking; and selecting an appropriate K becomes important. Table 4 shows the effects of K on
the performance of coverage-based re-ranker. We train and test the coverage-based re-ranker on the
top-K predictions from the baseline, where K ∈ {3, 5, 10}. The upper bound results are the same
ones from Table 3. The results show that when K is small, like K=3, the performance is not very
good due to the low coverage (thus low upper bound) of the candidate list. With the increase of K,
the performance becomes better, but the top-5 and top-10 results are on par with each other. This
is because the higher upper bound of top-10 results counteracts the harder problem of re-ranking
9
Under review as a conference paper at ICLR 2018
Candidate Set
Re-Ranker Results
EM
F1
Upper Bound
EM
F1
top-10
top-50
top-100
top-200
37.9
37.8
36.4
33.7
56.1
64.1
66.5
68.7
46.1
47.8
47.3
45.8
64.8
74.1
77.1
79.5
Table 5: Results of running strength-based re-ranker (counting) on different number of top-K answer
candidates on Quasar-T (dev set).
Q:
A1:
P1
P2
Which children ’s TV programme , which first appeared in November 1969 , has won a record
122 Emmy Awards in all categories ?
Great Dane
A2: Sesame Street
The world ’s most famous Great Dane first P1: In its long history , Sesame Street has received more Emmy Awards than any other
appeared on television screens on Sept. 13 ,
program , ...
1969 .
premiered on broadcast television ( CBS ) P2: Sesame Street ... is recognized as a pioneer
Saturday morning , Sept. 13 , 1969 , ... yet
of the contemporary standard which combeloved great Dane .
bines education and entertainment in children ’s television shows .
Table 6: An example from Quasar-T dataset. The ground-truth answer is ”Sesame Street”. Q:
question, A: answer, P: passages containing corresponding answer.
longer lists. Since there is no significant advantage of the usage of K=10 while the computation
cost is higher, we report all testing results with K=5.
Effect of the selection of K for the strength-based re-ranker Similar to Table 4, we conduct
experiments to show the effects of K on the performance of the strength-based re-ranker. We run
the strength-based re-ranker (counting) on the top-K predictions from the baseline, where K ∈
{10, 50, 100, 200}. We also evaluate the upper bound results for these Ks. Note that the strengthbased re-ranker runs very fast and the different values of K do not affect the computation speed very
much compared to the other QA components. The results are shown in Table 5, where we achieve
the best results when K=50. The performance drops significantly when K increases to 200. This is
because the ratio of incorrect answers increases a lot, making incorrect answers also likely to have
high counts. When K is smaller, such incorrect answers appear less because statistically they have
lower prediction scores. We report all testing results with K=50.
Examples Table 6 shows an example from Quasar-T where the re-ranker successfully corrected
the wrong answer predicted by the baseline. This is a case where the coverage-based re-ranker
helped: the correct answer “Sesame Street” has evidence from different passages that covers the
aspects “Emmy Award” and “children ’s television shows”. Although it still does not fully match
all the facts in the question, it still helps to rank the correct answer higher than the top-1 prediction
“Great Dane” from the R3 baseline, which only has evidence covering “TV” and “1969” in the
question.
5
R ELATED W ORK
Open Domain Question Answering The task of open domain question answering dates back
to as early as (Green Jr et al., 1961) and was popularized by TREC-8 (Voorhees, 1999). The task
is to produce the answer to a question by exploiting resources such as documents (Voorhees, 1999),
webpages (Kwok et al., 2001) or structured knowledge bases (Berant et al., 2013; Bordes et al.,
2015; Yu et al., 2017).
Recent efforts (Chen et al., 2017; Dunn et al., 2017; Dhingra et al., 2017b; Wang et al., 2017) benefit
from the advances of machine reading comprehension (RC) and follow the search-and-read QA
direction. These deep learning based methods usually rely on a document retrieval module to retrieve
a list of passages for RC models to extract answers. As there is no passage-level annotation about
10
Under review as a conference paper at ICLR 2018
which passages entail the answer, the model has to find proper ways to handle the noise introduced
in the IR step. Chen et al. (2017) uses bi-gram passage index to improve the retrieval step; Dunn
et al. (2017); Dhingra et al. (2017b) propose to reduce the length of the retrieved passages. Wang
et al. (2017) focus more on noise reduction in the passage ranking step, in which a ranker module is
jointly trained with the RC model with reinforcement learning.
To the best of our knowledge, our work is the first to improve neural open-domain QA systems by
using multiple passages for evidence aggregation.
Multi-Step Approaches for Reading Comprehension We are the first to introduce re-ranking
methods to neural open-domain QA and multi-passage RC. Meanwhile, our two-step approach
shares some similarities to the previous multi-step approaches proposed for standard single-passage
RC, in terms of the purposes of either using additional information or re-fining answer predictions
that are not easily handled by the standard answer extraction models for RC.
On cloze-test tasks (Hermann et al., 2015), Epireader Trischler et al. (2016) relates to our work in
the sense that it is a two-step extractor-reasoner model, which first extracts K most probable singletoken answer candidates and then constructs a hypothesis by combining each answer candidate to the
question and compares the hypothesis with all the sentences in the passage. Their model differs from
ours in several aspects: (i) Epireader matches a hypothesis to every single sentence, including all the
“noisy” ones that does not contain the answer, that makes the model inappropriate for open-domain
QA setting; (ii) The sentence matching is based on the sentence embedding vectors computed by
a convolutional neural network, which makes it hard to distinguish redundant and complementary
evidence in aggregation; (iii) Epireader passes the probabilities predicted by the extractor to the
reasoner directly to sustain differentiability, which cannot be easily adapted to our problem to handle
phrases as answers or to use part of passages.
Similarly, (Cui et al., 2017) also combined answer candidates to the question to form hypotheses, and
then explicitly use language models trained on documents to re-rank the hypotheses. This method
benefits from the consistency between the documents and gold hypotheses (which are titles of the
documents) in cloze-test datasets, but does not handle multiple evidence aggregation like our work.
S-Net (Tan et al., 2017) proposes a two-step approach for generative QA. The model first extracts an
text span as the answer clue and then generates the answer according to the question, passage and
the text span. Besides the different goal on answer generation instead of re-ranking like this work,
their approach also differs from ours on that it extracts only one text span from a single selected
passage.
6
C ONCLUSIONS
We have observed that open-domain QA can be improved by explicitly combining evidence from
multiple retrieved passages. We experimented with two types of re-rankers, one for the case where
evidence is consistent and another when evidence is complementary. Both re-rankers helped to
significantly improve our results individually, and even more together. Our results considerably
advance the state-of-the-art on three open-domain QA datasets.
R EFERENCES
Petr Baudiš and Jan Šedivỳ. Modeling of the question answering task in the yodaqa system. In
International Conference of the Cross-Language Evaluation Forum for European Languages, pp.
222–228. Springer, 2015.
Jonathan Berant, Andrew Chou, Roy Frostig, and Percy Liang. Semantic parsing on freebase from
question-answer pairs. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, 2013.
Antoine Bordes, Nicolas Usunier, Sumit Chopra, and Jason Weston. Large-scale simple question
answering with memory networks. Proceedings of the International Conference on Learning
Representations, 2015.
11
Under review as a conference paper at ICLR 2018
Christian Buck, Jannis Bulian, Massimiliano Ciaramita, Andrea Gesmundo, Neil Houlsby, Wojciech Gajewski, and Wei Wang. Ask the right questions: Active question reformulation with
reinforcement learning. arXiv preprint arXiv:1705.07830, 2017.
Danqi Chen, Adam Fisch, Jason Weston, and Antoine Bordes. Reading Wikipedia to answer opendomain questions. In Proceedings of the Conference on Association for Computational Linguistics, 2017.
Michael Collins and Terry Koo. Discriminative reranking for natural language parsing. Computational Linguistics, 31(1):25–70, 2005.
Yiming Cui, Zhipeng Chen, Si Wei, Shijin Wang, Ting Liu, and Guoping Hu. Attention-overattention neural networks for reading comprehension. Proceedings of the Conference on Association for Computational Linguistics, 2017.
Bhuwan Dhingra, Hanxiao Liu, Zhilin Yang, William W Cohen, and Ruslan Salakhutdinov. Gatedattention readers for text comprehension. Proceedings of the Conference on Association for Computational Linguistics, 2017a.
Bhuwan Dhingra, Kathryn Mazaitis, and William W Cohen. QUASAR: Datasets for question answering by search and reading. arXiv preprint arXiv:1707.03904, 2017b.
Matthew Dunn, Levent Sagun, Mike Higgins, Ugur Guney, Volkan Cirik, and Kyunghyun Cho.
SearchQA: A new q&a dataset augmented with context from a search engine. arXiv preprint
arXiv:1704.05179, 2017.
Chris Dyer, Adhiguna Kuncoro, Miguel Ballesteros, and Noah A Smith. Recurrent neural network
grammars. In Proceedings of the Conference on the North American Chapter of the Association
for Computational Linguistics, 2016.
David Ferrucci, Eric Brown, Jennifer Chu-Carroll, James Fan, David Gondek, Aditya A Kalyanpur,
Adam Lally, J William Murdock, Eric Nyberg, John Prager, et al. Building watson: An overview
of the deepqa project. AI magazine, 31(3):59–79, 2010.
Bert F Green Jr, Alice K Wolf, Carol Chomsky, and Kenneth Laughery. Baseball: an automatic
question-answerer. In Papers presented at the May 9-11, 1961, western joint IRE-AIEE-ACM
computer conference, pp. 219–224. ACM, 1961.
Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa
Suleyman, and Phil Blunsom. Teaching machines to read and comprehend. In Advances in
Neural Information Processing Systems, pp. 1693–1701, 2015.
Liang Huang. Forest reranking: Discriminative parsing with non-local features. In Proceedings of
the Conference on Association for Computational Linguistics, 2008.
Mandar Joshi, Eunsol Choi, Daniel S. Weld, and Luke Zettlemoyer. Triviaqa: A large scale distantly
supervised challenge dataset for reading comprehension. In Proceedings of the Annual Meeting
of the Association for Computational Linguistics, 2017.
Rudolf Kadlec, Martin Schmid, Ondrej Bajgar, and Jan Kleindienst. Text understanding with the
attention sum reader network. In Proceedings of the Annual Meeting of the Association for Computational Linguistics, 2016.
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of
the International Conference on Learning Representations, 2015.
Cody Kwok, Oren Etzioni, and Daniel S Weld. Scaling question answering to the web. ACM
Transactions on Information Systems (TOIS), 19(3):242–262, 2001.
Tao Lei, Wengong Jin, Regina Barzilay, and Tommi Jaakkola. Deriving neural architectures from
sequence and graph kernels. International Conference on Machine Learning, 2017.
Ankur P Parikh, Oscar Täckström, Dipanjan Das, and Jakob Uszkoreit. A decomposable attention
model for natural language inference. In Proceedings of the Conference on Empirical Methods in
Natural Language Processing, 2016.
12
Under review as a conference paper at ICLR 2018
Jeffrey Pennington, Richard Socher, and Christopher D Manning. GloVe: Global vectors for word
representation. In Proceedings of the Conference on Empirical Methods in Natural Language
Processing, 2014.
Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. SQuAD: 100,000+ questions
for machine comprehension of text. In Proceedings of the Conference on Empirical Methods in
Natural Language Processing, 2016.
Minjoon Seo, Aniruddha Kembhavi, Ali Farhadi, and Hannaneh Hajishirzi. Bidirectional attention
flow for machine comprehension. In Proceedings of the International Conference on Learning
Representations, 2017.
Libin Shen, Anoop Sarkar, and Franz Josef Och. Discriminative reranking for machine translation. In Proceedings of the Conference on the North American Chapter of the Association for
Computational Linguistics, 2004.
Chuanqi Tan, Furu Wei, Nan Yang, Weifeng Lv, and Ming Zhou. S-net: From answer extraction to
answer generation for machine reading comprehension. arXiv preprint arXiv:1706.04815, 2017.
Adam Trischler, Zheng Ye, Xingdi Yuan, and Kaheer Suleman. Natural language comprehension
with the epireader. Proceedings of the Conference on Empirical Methods in Natural Language
Processing, 2016.
Ellen M. Voorhees. The trec-8 question answering track report. In Trec, volume 99, pp. 77–82,
1999.
Shuohang Wang and Jing Jiang. Learning natural language inference with LSTM. In Proceedings of
the Conference on the North American Chapter of the Association for Computational Linguistics,
2016.
Shuohang Wang and Jing Jiang. Machine comprehension using match-LSTM and answer pointer.
In Proceedings of the International Conference on Learning Representations, 2017.
Shuohang Wang, Mo Yu, Xiaoxiao Guo, Zhiguo Wang, Tim Klinger, Wei Zhang, Shiyu Chang,
Gerald Tesauro, Bowen Zhou, and Jing Jiang. R3: Reinforced reader-ranker for open-domain
question answering. arXiv preprint arXiv:1709.00023, 2017.
Mo Yu, Wenpeng Yin, Kazi Saidul Hasan, Cicero dos Santos, Bing Xiang, and Bowen Zhou. Improved neural relation detection for knowledge base question answering. Proceedings of the
Conference on Association for Computational Linguistics, 2017.
13
| 2cs.AI
|
Unconstrained Scene Text and Video Text
Recognition for Arabic Script
Mohit Jain, Minesh Mathew and C. V. Jawahar
arXiv:1711.02396v1 [cs.CV] 7 Nov 2017
Center for Visual Information Technology, IIIT Hyderabad, India.
[email protected], [email protected], [email protected]
Abstract—Building robust recognizers for Arabic has always
been challenging. We demonstrate the effectiveness of an end-toend trainable CNN - RNN hybrid architecture in recognizing Arabic
text in videos and natural scenes. We outperform previous stateof-the-art on two publicly available video text datasets - ALIF
and ACTIV. For the scene text recognition task, we introduce
a new Arabic scene text dataset and establish baseline results.
For scripts like Arabic, a major challenge in developing robust
recognizers is the lack of large quantity of annotated data. We
overcome this by synthesizing millions of Arabic text images
from a large vocabulary of Arabic words and phrases. Our
implementation is built on top of the model introduced here [37]
which is proven quite effective for English scene text recognition.
The model follows a segmentation-free, sequence to sequence
transcription approach. The network transcribes a sequence of
convolutional features from the input image to a sequence of
target labels. This does away with the need for segmenting input
image into constituent characters/glyphs, which is often difficult
for Arabic script. Further, the ability of RNNs to model contextual
dependencies yields superior recognition results.
Keywords—Arabic, Arabic Scene Text, Arabic Video Text, Synthetic Data, Deep Learning, Text Recognition.
I.
I NTRODUCTION
For many years, the focus of research on text recognition
in Arabic has been on printed and handwritten documents [7],
[11], [12], [13]. Majority of the works in this space were on
individual character recognition [12], [13]. Since segmenting
an Arabic word or line image into its sub units is challenging, such models did not scale well. In recent years there
has been a shift towards segmentation-free text recognition
models, mostly based on Hidden Markov Models (HMM) or
Recurrent Neural Networks (RNN). Such models generally
follow a sequence-to-sequence approach wherein the input
line/word image is directly transcribed to a sequence of labels.
Methods such as [9] and [10] use RNN based models for
recognizing printed/handwritten Arabic script. Our attempt
to recognize Arabic text on videos (video text recognition)
and natural scenes (scene text recognition) follows the same
paradigm. Video text recognition in Arabic has gained interest
recently [24], [25], [38]. There are now two benchmarking
datasets available for the task ALIF [24] and ACTIV [25]. In
the scene text domain, [8] uses sparse coding of SIFT features
for a bag-of-features model using spatial pyramid matching
for the recognition task at the character level in Arabic. To the
best of our knowledge, there has been no work so far on scene
text recognition at word level in Arabic.
The computer vision community experienced a strong
revival of neural networks based solutions in the form of
Deep Learning in recent years. This process was stimulated
Fig. 1: Examples of Arabic Scene Text (left) and Video Text
(right) recognized by our model. Our work deals only with
the recognition of cropped words/lines. The bounding boxes
were provided manually.
by the success of models like Deep Convolutional Neural
Networks (DCNNs) in feature extraction, object detection and
classification tasks as seen in [1], [3]. These tasks however
pertain to a set of problems where the subjects appear in
isolation, rather than appearing as a sequence. Recognising
a sequence of objects often requires the model to predict a
series of description labels instead of a single label. DCNNs are
not well suited for such tasks as they generally work well in
scenarios where the inputs and outputs are bounded by fixed
dimensions. Also, the lengths of these sequence-like objects
might vary drastically and that escalates the problem difficulty. Recurrent neural networks (RNNs) tackle the problems
faced by DCNNs for sequence-based learning by performing
a forward pass for each segment of the sequence-like-input.
Such models often involve a pre-processing/feature-extraction
step, where the input is first converted into a sequence of
feature vectors [4], [5]. These feature extraction stage would
be independent of RNN-pipeline and hence they are not endto-end trainable.
The upsurge in video sharing on social networking websites
and the increasing number of TV channels in today’s world
reveals videos to be a fundamental source of information.
Effectively managing, storing and retrieving such video data is
not a trivial task. Video text recognition can greatly aid video
content analysis and understanding, with the recognized text
giving a direct and concise description of the stories being
depicted in the videos. In news videos, superimposed tickers
running on the edges of the video frame are generally highly
correlated to the people involved or the story being portrayed
and hence provide a brief summary of the news event.
Reading text in natural scenes is relatively harder task
compared to printed text recognition. The problem has been
drawing increasing research interest in recent years. This can
partially be attributed to the rapid development of wearable and
mobile devices such as smart phones, google glass and selfdriving cars, where scene text is a key module to a wide range
of practical and useful applications. While the recognition
of text in printed documents has been studied extensively in
the form of Optical Character Recognition (OCR) Systems,
these methods generally do not generalize very well to a
natural scene setting where factors like inconsistent lighting
conditions, variable fonts, orientations, background noise and
image distortions add to the problem complexity. Even though
Arabic is the fourth most spoken language in the world after
Chinese, English and Spanish, to the best of our knowledge,
there has not been any work in the field of word-level scene
text recognition, which makes this area of research ripe for
exploration.
A. Intricacies of Arabic Script
Automatic recognition of Arabic is a pretty challenging
task due to various intricate properties of the script. There
are only 28 letters in the script and it is written in a semicursive fashion from right to left. The letters of the script
are generally connected by a line at its bottom to the letters
preceding and following it, except for 6 letters which can
be linked only to their preceding letters. Such an instance
in a word creates a paw (part-of Arabic word) in the word.
Arabic script has another exceptional characteristic where the
shape of a letter changes depending on whether the letter
appears in the word in isolation, at the beginning, middle
or at the end. In general, each letter can have one to four
possible shapes which might have little or no similarity in
shape whatsoever. Another frequent occurrence in Arabic is
that of dots. The addition of a single dot to a letter can
change the character completely. Arabic also follows ligature,
where two letters when combined form a third different letter
in a way that they cannot be separated by a simple baseline (i.e. a complex shape represents this combined letter).
These distinctive characteristics make automated Arabic script
recognition more challenging than most other scripts.
B. Scene Text and Video Text Recognition
Though there has been a lot of work done in the field of
text transcription in natural scenes and videos for the English
script [4], [5], [14], [30], it is still in a nascent state as far as
Arabic script is considered. Previous attempts made to address
similar problems in English [30], [31] first detect individual
characters and then character-specific DCNN models are used
to recognize these detected characters. The short-comings of
such methods are that they require training a strong character
detector for accurately detecting and cropping each character
out from the original word. In case of Arabic, this task becomes
even more difficult due to the intricacies of the script, as
discussed earlier. Another approach by Jaderberg et al. [14],
was to treat scene text recognition as an image classification
problem instead. To each image, they assign a class label from
a lexicon of words spanning the English language (90K most
frequent words were chosen to put a bound on the span-set).
However, this approach is limited to the size of lexicon used
for its possible unique transcriptions and the large number of
output classes add to training complexity. Hence the model
is not scalable to inflectional languages like Arabic where
number of unique words is much higher compared to English.
Another category of solutions typically embed image and its
text label in a common subspace and retrieval/recognition is
performed on the learnt common representations. For example,
Almazan et al. [32] embed word images and text strings in a
common vector-subspace, and thus convert the task of word
recognition into a retrieval problem. Yao et al. [35] and Gordo
et al. [36] used mid-level features for scene text recognition.
The segmentation-free, transcription approach was proved
quite effective for Indian Scripts OCR [21], [22] where segmentation is often problematic. Similar approach was used in [5]
for scene text recognition. Handcrafted features derived from
image gradients were used with a RNN to map the sequence of
features to a sequence of labels. Unlike the problem of OCR,
scene text recognition required more robust features to yield
results comparable to the transcription based solutions for OCR.
A novel approach combining the robust convolutonal features
and transcription abilities of RNN was introduced by [37]. We
employ this particular architecture where the hybrid CNN - RNN
network with a Connectionist Temporal Classification (CTC)
loss is trained end-to-end.
Works on text extraction from videos has generally been
in four broad categories, edge detection methods [39], [40],
[41], extraction using connected components [42], [43], texture
classification methods [44] and correlation based methods [45],
[46]. The previous attempts at solving video text for Arabic
used two separate routines; one for extracting relevant image
features and another for classifying features to script labels for
obtaining the target transcription. In [38] text-segmentation and
statistical feature extraction are followed by fuzzy k-nearest
neighbour techniques to obtain transcriptions. Similarly, [23]
experiment with two feature extraction routines; CNN based
feature extraction and Deep Belief Network (DBN) based
feature extraction, followed by a Bi-directional Long Short
Term Memory (LSTM) layer [16], [17].
The rest of the paper is organized as follows; Section II
describes the hybrid CNN - RNN architecture we use. Section
III focuses on the Arabic text recognition from a transcription
perspective. Section IV presents the results on benchmarking
datasets, followed by a discussion based on the qualitative
results. Section V concludes with the findings of our work.
II.
H YBRID CNN - RNN A RCHITECTURE
The network consists of three components. The initial
convolutional layers, the middle recurrent layers and a final
transcription layer. This architecture we employ was introduced in [37].
The role of the convolutional layers is to extract robust
feature representations from the input image and feed it into the
recurrent layers, which in turn will transcribe them to an output
sequence of labels or characters. The sequence to sequence
transcription is achieved by a CTC layer at the output.
All the images are scaled to a fixed height before being
fed to the convolutional layers. The convolutional components
then create a sequence of feature vectors from the feature maps
Conv 1
W: 3*3
S: 1
P:1,1
Input
Image
W*32
W*32*64
Conv 2
Conv 3
W: 3*3
W: 3*3
S: 1
S: 1
P:1,1
P:1,1
Max
Max
Pool
Pool
W: 2*2
W: 2*2
S: 2
S: 2
P : 0,0
P : 0,0
W/2*16*128
W/4*8*256
W/4*8*256
Conv 4
W: 3*3
S: 1
P:1,1
Conv 5 Conv 6
W: 3*3
W: 3*3
S: 1
S: 1
P:1,1
P:1,1
Max
Pool
W: 1*2
S: 2
P : 1,0
Batch
Norm.
W/4*4*512
(W/4+1)*4*
512
Conv 7
W: 3*3
S: 1
P: 0,0
Max
Pool
W: 1*2
S: 2
P : 1,0
Batch
Norm.
W/4*2*512
(W/4+2)*1*
512
Bidirectional
LSTM
Bidirectional
LSTM
softmax
Feature
Sequence
(W/4+1) *
512
CTC
1*256
1*256
L
A
B
E
L
S
E
Q
U
E
N
C
E
1* #labels
Fig. 2: The deep neural network architecture employed in our work, which was devised by [37].
The symbols ‘k’, ‘s’ and ‘p’ stand for kernel size, stride and padding size respectively.
by splitting them column-wise, which then act as inputs to
the recurrent layers (Fig. 2). This feature descriptor is highly
robust and most importantly can be trained to be adopted to a
wide variety of problems [1], [3], [14].
On top of the convolutional layers, the recurrent layers
take each frame from the feature sequence generated by the
convolutional layers and make predictions. The recurrent layers
consist of deep bidirectional LSTM nets. Number of parameters
in a RNN is independent of the length of the sequence, hence
we can simply unroll the network as many times as the number
of time-steps in the input sequence. This in our case helps
perform unconstrained recognition as the predicted output can
be any sequence of labels derived from the entire label set.
Traditional RNN units (vanilla RNNs) face the problem of
vanishing gradients [15] and hence (LSTM) units are used
which were specifically designed to tackle the vanishinggradients problem [16], [17]. In a text recognition problem,
contexts from both directions (left-to-right and right-to-left) are
useful and complementary to each other in outputting the right
transcription. Therefore, a combined forward and backward
oriented LSTM is used to create a bi-directional LSTM unit.
Multiple such bi-directional LSTM layers can be stacked to
make the network deeper and gain higher levels of abstractions
over the image-sequences as shown in [18].
The transcription layer at the top of the network is used
to translate the predictions generated by the recurrent layers
into label sequences for the target language. The CTC layer’s
conditional probability is used in the objective function as
shown in [19]. The complete network configuration used for
the experiments can be seen in (Fig. 2).
III.
A RABIC T EXT T RANSCRIPTION
The efficacy of a seq2seq approach lies in the fact that
the input sequence can be transcribed directly to the output
sequence without the need for a target defined at each timestep. This was proven quite effective in recognizing complex
scripts where sub-word segmentation is often difficult [22].
In addition, contextual modelling of the sequence in both
the directions, forward and backward, is achieved by the
bidirectional LSTM block. Contextual information is critical
in making accurate predictions for a Script like Arabic where
there are many similar looking characters/glyphs. The hybrid
architecture replaces the need for any handcrafted features to
be extracted from the image. Instead, more robust convolutional features are fed to the RNN layers.
Fig. 3: Sample images from the rendered synthetic Arabic
scene text dataset. The images closely resemble real world
scene images.
A. Datasets
Models for both video text and scene text recognition
problems are trained using synthetic images rendered from a
large vocabulary using freely available Arabic Unicode fonts
(Fig. 3). For scene text, the images were rendered from a
vocabulary of a quarter million most commonly used words
in Arabic Wikipedia. A random word from the vocabulary is
first rendered into the foreground layer of the image by varying
the font, stroke color, stroke thickness, kerning, skew and
rotation. Later a random perspective projective transformation
is applied to the image, which is then blended with a random
crop from a natural scene image. Finally the foreground layer
is alpha composed with a background layer, which is again a
random crop from a random natural image. The synthetic line
images for video text recognition are rendered from random
text lines crawled from Arabic news websites. The rendering
process is much simpler since real video text usually have
uniform color for the text and background and lacks any
perspective distortion. Details on the rendering process are
described here [20]. Around 2 million video text line images
and 4 million scene text word images were used for training the
respective models. The model for video text recognition task
was initially trained on the Synthetic video text dataset and
then fine-tuned on the train partitions of real-world datasets,
ALIF and ACTIV .
ALIF dataset consists of 6,532 cropped text line images
from 5 popular Arabic News channels. ACTIV dataset is larger
than ALIF and contains 21,520 line images from popular
Arabic News channels. The dataset contains video frames
wherein bounding boxes of text lines are annotated.
A new Arabic scene text dataset was curated by downloading freely available images containing Arabic script from
Google Images. The dataset consists of 2000 word images
of Arabic script occurring in various scenarios like local
CRR =
(nCharacters −
Fig. 4: Sample images from the Arabic scenetext dataset.
B. Implementation Details
Convolutional block in the network follows the VGG architecture [6]. In the 3rd and 4th max-pooling layers, the
pooling windows used are rectangular instead of the usual
square windows used in VGG. The advantage of doing this is
that the feature maps obtained after the convolutional layers are
wider and hence we obtain longer feature sequences as inputs
for the recurrent layers that follow. To enable faster batch
learning, all input images are re-sized to a fixed width and
height (32x100 for scene text and 32x504 for video text). We
have observed that resizing all images to a fixed width is not
affecting the performance much. The images are horizontally
flipped before feeding them to the convolutional layers since
Arabic is read from right to left.
To tackle the problems of training such deep convolutional
and recurrent layers, we used the batch normalization [27]
technique. The network is trained with stochastic gradient descent (SGD). Gradients are calculated by the back-propagation
algorithm. Precisely, the transcription layers’ error differentials
are back-propagated with the forward-backward algorithm,
as shown in [19]. While in the recurrent layers, the BackPropagation Through Time (BPTT) [28] algorithm is applied
to calculate the error differentials. The hassle of manually
setting the learning-rate parameter is taken care of by using
ADADELTA optimization [29].
IV.
R ESULTS AND D ISCUSSION
In this section we demonstrate the efficacy of above architecture in recognizing Arabic script appearing in video frames
and natural scene images. It is assumed that input images
are cropped word images from natural scenes, not full scene
images. Since there were no previous works in Arabic scene
text recognition at word level, the baseline results are reported
on the new Arabic scene text we introduce. For video text
recognition problem the results are reported on two existing
video text datasets - ALIF [23] and ACTIV [25].
Results on video text recognition are presented in Table
I. Since there has been no work done in word-level Arabic
scene text recognition, we compare the results obtained on the
Arabic scene text dataset using a popular free English OCR Tesseract [47]. The performance has been evaluated using the
following metrics; CRR - Character Recognition Rate, WRR Word Recognition Rate, LRR - Line Recognition Rate. In the
below equations, RT and GT stand for recognized text and
ground truth respectively.
EditDistance(RT, GT ))
nCharacters
W RR =
markets & shops, billboards, navigation signs, graffiti, etc, and
spans a large variety of naturally occurring image-noises and
distortions (Fig. 4). The images were manually annotated by
human experts of the script and the transcriptions as well as the
image data is being made publicly available for future research
groups to compare and improve performance upon.
P
LRR =
nW ordsCorrectlyRecognized
nW ords
nT extImagesCorrectlyRecognized
nImages
It should be noted that even though the methods compared
on for video text recognition use a separate convolutional
architecture for feature extraction, unlike the end-to-end trainable architecture, we obtain better character and line-level
accuracies for the Arabic video text recognition task and set
the new state-of-the-art for the same.
TABLE I: Accuracy for Video Text Recognition.
ALIF Test1
CRR(%)
LRR(%)
94.36
55.03
90.73
39.39
83.26
26.91
98.17
79.67
VideoText
ConvNet-BLSTM [23]
DBN-BLSTM [23]
ABBYY [26]
CNN - RNN hybrid network
The hybrid
ALIF Test2
CRR(%)
LRR(%)
90.71
44.90
87.64
31.54
81.51
27.03
97.84
77.98
AcTiV
CRR(%)
LRR(%)
–
–
–
–
–
–
97.44
67.08
CNN - RNN
architecture we employ outperforms previous
video text recognition benchmarks.
TABLE II: Accuracy for Scene Text Recognition.
SceneText
Tesseract [47]
Baseline using the hybrid CNN - RNN
Arabic SceneText Dataset
CRR(%)
WRR(%)
17.07
5.92
75.05
39.43
Lower accuracies on scene text recognition problem testify
the inherent difficulty associated with the problem compared
to printed or video text recognition.
On video text recognition, we report the best results so
far on both the datasets. Scene text recognition is a much
harder problem compared to OCR or video text recognition.
The variability in terms of lighting, distortions and typography
make the learning pretty hard and we provide baseline results
for the same. A qualitative analysis of the recognition tasks
can be seen in Fig. 5.
V.
C ONCLUSION
We demonstrate that the state-of-the-art deep learning
techniques can be successfully adapted to some rather challenging tasks like Arabic text recognition. The newer script
and language agnostic approaches are well suited for low
resource languages like Arabic where the traditional methods
often involved language specific modules. The success of
RNN s in sequence learning problems has been instrumental
in the recent advances in speech recognition and image to text
transcription problems. This came as a boon for languages
like Arabic where the segmentation of words into sub word
units was often troublesome. The sequence learning approach
could directly transcribe the images and also model the context
in both forward and backward directions. With better feature
representations and learning algorithms available, we believe
the focus should now shift to harder problems like scene text
recognition. We hope the introduction of a new Arabic scene
text dataset and the initial results would instill an interest
among the Arabic computer vision community to pursue this
field of research further.
Fig. 5: Qualitative results of Scene Text Recognition. For each image, the annotations on bottom-left and bottom-right are the
label and model prediction, respectively.
ACKNOWLEDGMENT
The authors would like to thank Maaz Anwar, Anjali,
Saumya, Vignesh and Rohan for helping annotate the Arabic
scenetext dataset and Dr. Girish Varma for his timely help and
discussions.
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
A. Krizhevsky, I. Sutskever and G.E. Hinton, Imagenet classification with
deep convolutional neural networks. NIPS 2012.
Y. LeCun, L. Bottou, Y. Bengio and P.Haffiner, Gradient-based learning
applied to document recognition. Proceedings of the IEEE, 1998.
R.B. Girshick, J. Donahue, T. Darrel and J. Malik, Rich feature hierarchies for accurate object detection and semantic segmentation. CVPR
2014.
A. Graves, M. Liwicki, S. Fernandez, R. Bertolami, H. Bunke and J.
Schidhuber, A novel connectionist system for unconstrained handwriting
recognition. TPAMI, 2009.
B. Su and S. Lu, Accuracte scene text recognition based on recurrent
neural network. ACCV 2014.
K. Simonyan and A. Zisserman. Very deep convolutional networks for
large-scale image recognition. CoRR 2014.
Prasad, Rohit, et al. Improvements in hidden Markov model based Arabic
OCR, ICPR 2008.
M. Tounsi, I. Moalla, A.M. Alimi, F. Lebouregois, Arabic characters
recognition in natural scenes using sparse coding for feature representations, ICDAR 2015.
Yousefi, Mohammad Reza, et al. A comparison of 1D and 2D LSTM
architectures for the recognition of handwritten Arabic, SPIE 2015.
Ul-Hasan, Adnan, et al, Offline printed Urdu Nastaleeq script recognition with bidirectional LSTM networks, ICDAR 2013.
El-Hajj, Ramy, Laurence Likforman-Sulem, and Chafic Mokbel, Arabic
handwriting recognition using baseline dependant features and hidden
Markov modeling. ICDAR 2005.
Amin, Adnan, Humoud Al-Sadoun, and Stephen Fischer, Hand-printed
Arabic character recognition system using an artificial network. Pattern
recognition 29.4 (1996): 663-675.
Amin, Adnan. Off-line Arabic character recognition: the state of the
art. Pattern recognition 31.5 (1998): 517-530.
M. Jaderberg, K. Simoyan, A. Vedaldi and A. Zisserman, Reading text
in the wild with convolutional neural networks. IJCV 2015.
Y. Bengio, P.Y. Simard and P. Frasconi, Learning long-term dependencies with gradient descent is difficult. NN 1994.
F.A. Gers, N.N. Schraudolph and J. Schmidhuber, Learning precise
timing with LSTM recurrent networks. JMLR 2002.
S. Hochreiter and J. Schmidhuber, Long short-term memory. Neural
Computation 1997.
A. Graves, A. Mohamed and G.E. Hinton, Speech recognition with deep
recurrent neural networks. ICASSP 2013.
A. Graves, S. Fernandez, F.J. Gomez and J. Schmidhuber, Connectionist temporal classification: labelling unsegmented sequence data with
recurrent neural networks. ICML 2006.
Praveen Krishnan and C.V. Jawahar, Generating Synthetic Data for Text
Recognition. arXiv 2016.
Naveen Sankaran and C.V. Jawahar, Recognition of printed Devanagari
text using BLSTM Neural Network. ICPR 2012.
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
Minesh Mathew, Ajeet Kumar Singh and C.V. Jawahar, Multilingual
OCR for Indic Scripts. DAS 2016.
S. Yousfi, B. Sid-Ahmed and G. Christophe, ALIF: A dataset for Arabic
embedded text recognition in TV broadcast. ICDAR 2015.
Yousfi, Sonia, Sid-Ahmed Berrani, and Christophe Garcia, Arabic
text detection in videos using neural and boosting-based approaches:
Application to video indexing. ICIP 2014.
O. Zayene, J. Hennebert, S.M. Touj, R. Ingold, N.E. Amara, A dataset
for Arabic text detection, tracking and recognition in news videos-AcTiV.
ICDAR 2015.
K. Wang, B. Babenko and S. Belongie, End-to-end scene text recognition. In ICCV, 2011.
S. Ioffe and C. Szegedy, Batch normalization: Accelerating deep
network training by reducing internal covariate shift. ICML 2015.
P.J. Werbos, Backpropagation through time: what it does and how to
do it. Proceedings of the IEEE 1990.
M.D. Zeiler, ADADELTA: an adaptive learning rate method. CoRR
2012.
T. Wang, D.J. Wu, A. Coates and A.Y. Ng, End-to-end text recognition
with convolutional neural networks. CVPR 2014.
A. Bissacco, M. Cummins, Y. Netzer and H. Neven, Photoocr: Reading
text in uncontrolled conditions. ICCV 2013.
Anand Mishra, Karteek Alahari and C.V. Jawahar, Word spotting and
recognition with embedded attributes. TPAMI 2014.
J. Almazan, A. Gordo, A. Fornes and E. Valveny, Scene Text Recognition using Higher Order Langauge Priors. BMVC 2012.
J.A. Rodriguez-Serrano, A. Gordo and F. Perronnin, Label embedding:
A frugal baseline for text recognition. IJCV 2015.
C. Yao, X. Bai, B. Shi and W. Liu, Strokelets : A learned multi-scale
representation for scene text recognition. ICPR 2012.
A. Gordo, Supervised mid-level features for word image representation.
CVPR 2015.
B. Shi, X. Bai and C. Yao, An end-to-end trainable neural network
for image-based sequence recognition and its application to scene text
recognition. arXiv 2015.
Halima, M. Ben, Hichem Karray, and Adel M. Alimi, Arabic text
recognition in video sequences. arXiv 2013.
L. Agnihotri and N. Dimitrova, Text detection for video analysis. CVPR
workshop, 1999.
L. Agnihotri., N.Dimitrova, M.Soletic, Multi-layered Videotext Extraction Method. ICME 2002.
S. Hua,X., X.-R. Chen., Automatic Location of Text in Video Frames.
MIR 2001.
A. K. Jain and B. Yu, Automatic text location in images and video
frames. PR 1998.
R. Lienhart and F. Stuber, Automatic text recognition indigital videos.
Proceedings of SPIE Image and Video 1996.
H. Karray, M. Ellouze and M.A. Alimi, Indexing Video Summaries for
Quick Video Browsing. Computer Communications and Networks 2009.
H. Karray , A.M. Alimi, Detection and Extraction of the Text in a video
sequence. ICES 2005.
M. Kherallah, H. Karray, M. Ellouze, A.M. Alimi, Toward an Interactive Device for Quick News Story Browsing. ICPR 2008,
”Tesseract OCR”, https://github.com/tesseract-ocr/tesseract.
| 1cs.CV
|
arXiv:1801.06634v1 [math.ST] 20 Jan 2018
Joint CLT for eigenvalue statistics from
several dependent large dimensional sample
covariance matrices with application
Weiming Li1,* Zeng Li2,** and Jianfeng Yao3,†
1
School of Statistics and
Management
Shanghai University of Finance
and Economics
e-mail:
*
[email protected]
2
Department of Statistics
Pennsylvania State University
e-mail: ** [email protected]
3
Department of Statistics and
Actuarial Science
The University of Hong Kong
e-mail: † [email protected]
Abstract: Let Xn = (xij ) be a k × n data matrix with complex-valued, independent
and standardized entries satisfying a Lindeberg-type moment condition. We consider
simultaneously R sample covariance matrices Bnr = n1 Qr Xn X∗n Q>
r , 1 ≤ r ≤ R,
where the Qr ’s are nonrandom real matrices with common dimensions p × k (k ≥
p). Assuming that both the dimension p and the sample size n grow to infinity,
the limiting distributions of the eigenvalues of the matrices {Bnr } are identified,
and as the main result of the paper, we establish a joint central limit theorem for
linear spectral statistics of the R matrices {Bnr }. Next, this new CLT is applied
to the problem of testing a high dimensional white noise in time series modelling.
In experiments the derived test has a controlled size and is significantly faster than
the classical permutation test, though it does have lower power. This application
highlights the necessity of such joint CLT in the presence of several dependent sample
covariance matrices. In contrast, all the existing works on CLT for linear spectral
statistics of large sample covariance matrices deal with a single sample covariance
matrix (R = 1).
Keywords and phrases: Large sample covariance matrices, central limit theorem,
linear spectral statistics, white noise test, high-dimensional times series.
MSC 2010 Mathematics Subject Classifications: Primary 62H10; secondary 60B12, 60B20.
1. Introduction
Modern information technology tremendously accelerates computing speed and greatly enlarges the amount of data storage, which enables us to collect, store and analyze data
1
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
2
of large dimensions. Classical limit theorems in multivariate analysis, which normally assume fixed dimensions, become no longer applicable for dealing with high dimensional
problems. Random matrix theory investigates the spectral properties of random matrices when their dimensions tend to infinity and hence provides a powerful framework for
solving high dimensional problems. This theory has made systematic corrections to many
classical multivariate statistical procedures in the past decades, see the monographs of Bai
and Silverstein (2010), Yao et al. (2015) and the review papers Johnstone (2007) and Paul
and Aue (2014). It has found diverse applications in various research areas, including signal processing, network security, image processing, statistical genetics and other financial
econometrics problems.
The sample covariance matrix is of central importance in multivariate analysis. Many
fundamental statistics in multivariate analysis can be written as functionals of eigenvalues
of a sample covariance matrix Sn such as linear spectral statistics (LSSs) of the form
f (λ1 ) + · · · + f (λp ) where the λj ’s are eigenvalues of Sn and f (·) is a smooth function.
The wide range of creditable applications in high dimensional statistics triggered an uptick
in the demand for CLTs of such LSSs. Actually one of the most widely used results in
this area is Bai and Silverstein (2004), which considers a sample covariance matrix of the
form Bn = n1 T1/2 Xn X∗n T1/2 , where Xn = (xij ) is p × n matrix consisting of i.i.d. complex
standardized entries and T is a p × p nonnegative Hermitian matrix. A CLT for LSSs
of Bn is established under the so-called Marčenko-Pastur regime, i.e. n, p → ∞, p/n →
c > 0. Further refinement and extensions can be found in Zheng et al. (2015), Chen and
Pan (2015), Zheng et al. (2017a), and Zheng et al. (2017b). Among them, Zheng et al.
(2015) studied the unbiased sample covariance matrix when the population mean vector
is unknown. Chen and Pan (2015) looked into the ultra-high dimensional case when the
dimension p is much larger than the sample size n, that is p/n → ∞ as n → ∞. Zheng et
al. (2017a) derived the CLT for LSSs of large dimensional general Fisher matrices. Zheng
et al. (2017b) attempted to relax the fourth order moment condition in Bai and Silverstein
(2004) and incorporated it into the limiting parameters.
However, this rich literature all deals with a single sample covariance matrix Bn . This
paper, on the contrary, aims at the joint limiting behaviour of functionals of several groups
of eigenvalues coming from different yet closely related sample covariance matrices. Specifically, we consider data samples {yjr }1≤j≤n,1≤r≤R of the form yjr = Qr xj where
(M1) {xj , 1 ≤ j ≤ n} is a sequence of k-dimensional independent and complex-valued
random vectors with independent standardized components (xij ), i.e. Exij = 0 and
E|xij |2 = 1, and the dimension k ≥ p;
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
3
(M2) {Qr , 1 ≤ r ≤ R} are R nonrandom real matrices with common dimensions p × k.
The R population covariance matrices {Tnr = Qr Q>
r , r = 1, . . . R} are assumed
product-commutative.
We thus consider R sample covariance matrices given by
n
Bnr
1
1X
∗
yjr yjr
= Qr Xn X∗n Q>
=
r , 1 ≤ r ≤ R,
n j=1
n
(1.1)
where Xn = (x1 , . . . , xn ) is of size k × n, ∗ denotes the conjugate transpose of matrices or
vectors, and > stands for the transpose of real ones. Let (λjr )1≤j≤p be the eigenvalues of
Bnr (1 ≤ r ≤ R), and consider L × R real-valued functions (flr )1≤l≤L, 1≤r≤R . This leads to
the family of L × R LSSs
ϕlr = flr (λ1r ) + · · · + flr (λpr ), 1 ≤ l ≤ L, 1 ≤ r ≤ R .
This paper establishes a joint CLT for these L × R statistics {ϕlr } under appropriate
conditions.
The importance of such joint CLT for LSSs is best explained and illustrated with the
following problem of testing a high dimensional white noise. Indeed, our motivation for the
joint CLT originates from this application to high-dimensional time series analysis. Testing
for white noise is a classical yet important problem in statistics, especially for diagnostic
checks in time series modelling. For high dimensional time series, current literatures focus
on estimation and dimension-reduction aspects of the modelling, including high dimensional VAR models and various factor models. Yet model diagnostics have largely been
untouched. Classical omnibus tests such as the multivariate Hosking and Li-McLeod tests
are no longer suitable for handling high dimensional time series. They become extremely
conservative, losing size and power dramatically. In a very recent work, Li et al. (2016)
looked into this high dimensional portmanteau test problem and proposed several new test
statistics based on single-lagged and multi-lagged sample auto-covariance matrices. More
precisely, let’s consider a p-dimensional time series modelled as a linear process
xt =
X
Al zt−l ,
(1.2)
l≥0
where {zt } is a sequence of independent p-dimensional random vectors with independent
components zt = (zit ) satisfying Ezit = 0, E|zit |2 = 1, E|zit |4 < ∞. Hence {xt } has
Ext = 0, and its lag-τ auto-covariance matrix Στ = Cov(xt+τ , xt ) depends on τ only. In
particular, Σ0 = Var(xt ) denotes the population covariance matrix of the series. The goal
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
4
is to test whether xt is a white noise, i.e.
H0 : Cov(xt+τ , xt ) = 0, τ = 1, · · · , q,
(1.3)
where q ≥ 1 is a prescribed constant integer. Let x1 , . . . , xn be a sample generated from
the model (1.2). The lag-τ sample auto-covariance matrix is
n
X
bτ = 1
Σ
xt x∗t−τ ,
n t=1
(1.4)
b τ . For any
where xt = xn+t when t ≤ 0. Li et al. (2016) proposed a test statistic based on Σ
eτ was designed to test the specific
given constant integer 1 ≤ τ ≤ q, their test statistic L
lag-τ autocorrelation of the sequence, i.e.
eτ =
L
p
X
f∗ M
f τ ),
λ2j,τ = Tr(M
τ
j=1
where {λj,τ , j = 1, · · · , p} are the eigenvalues of
n
1 X
1 b
∗
b
f
Στ + Στ =
xt x∗t−τ + xt−τ x∗t ,
Mτ =
2
2n t=1
which is the symmetrized lag-τ sample auto-covariance matrix.
f τ = 1 Xn (Dτ + D> )X∗ , where
Notice that in matrix form M
τ
n
2n
Dτ =
0 In−τ
Iτ
0
!
where Im denotes the mth order unit matrix. They have proved that, under the null
hypothesis, in the simplest setting when xt = zt , the limiting distribution of the test
eτ is
statistic L
ne
p d
1
3c(ν4 − 1)
Lτ − →
− N
, 1+
.
p
2
2
2
Here, p, n → ∞ and p/n → c > 0 and ν4 = E|zit |4 . The null hypothesis should be rejected
eτ . Simulation results also show that this test statistic is consistently
for large values of L
more powerful than the Hosking and Li-McLeod tests even when the latter two have been
size adjusted.
f τ , which can be studied with the
eτ is an LSS of M
It can be seen that the test statistic L
CLT in Bai and Silverstein (2004). Indeed, the non-null eigenvalues of the sample covariance
1/2
1/2
matrix Sn = n1 Tp Xn X∗n Tp considered there are the same as the matrix Sn = n1 Xn Tp X∗n
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
5
f τ . However, the test statistic L
eτ can only detect serial
which resembles to the matrix M
dependence in a single lag each time. In order to capture a multi-lag dependence structure,
a naturally more effective way would be accumulating the lags and consider the statistic
Lq =
q
X
τ =1
eτ =
L
q
X
fτ M
f ∗ ).
Tr(M
τ
(1.5)
τ =1
Note that the CLT in Bai and Silverstein (2004) (or in its recent extensions) can only
f τ , while to derive
be used to study the correlations between different LSSs of a given M
the null distribution of Lq , we need to study the correlations between LSSs of several
f τ , 1 ≤ τ ≤ q. Consequently, we need to resort to the joint CLT
covariance matrices M
eτ , 1 ≤ τ ≤ q}. It is worth
studied in this paper to characterize the correlations among {L
noticing that Li et al. (2016) proposed another multi-lagged test statistic Uq by stacking
a number of consecutive observation vectors. It will be shown in this paper that this test
statistic Uq is essentially much less powerful than Lq considered here due to the data loss
caused by observation stacking. This superiority of Lq over Uq demonstrates the necessity
and significance of studying a joint CLT for LSSs of several dependent sample covariance
matrices as proposed in this paper.
The rest of the paper is organized as follows. The main results of the joint CLT of LSSs
of different sample covariance matrices are presented in Section 2. The application on high
dimensional white noise test is provided in Section 3 to demonstrate the utility of this joint
CLT. Numerical studies have also lent full support to the theoretical derivations. Technical
lemmas and proofs are left to Section 4. Finally, Matlab codes for reproducing simulations
in the paper are available at: http://web.hku.hk/~jeffyao/papersInfo.html.
2. Joint CLT for linear spectral statistics of {Bnr }1≤r≤R
2.1. Preliminary knowledge on LSDs of {Bnr }1≤r≤R
Recall that the empirical spectral distribution (ESD) of a p × p square matrix A is the
P
probability measure F A = p−1 pi=1 δλi , where the λi ’s are eigenvalues of A and δa denotes
the Dirac mass at point a. For any probability measure F on the real line, its Stieltjes
transform is defined by
Z
1
m(z) =
dF (x), z ∈ C+ ,
x−z
+
where C denotes the upper complex plane.
The assumptions needed for the existence of limiting spectral distributions (LSDs) of
{Bnr }1≤r≤R are as follows. The setup as well as the following Lemma 2.1 are established
in Zheng et al. (2017b).
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
6
Assumption (a) Both dimensions p and n tend to infinity such that cn = p/n → c > 0
as n → ∞.
Assumption (b) Samples are {yjr = Qr xj , j = 1, . . . , n, r = 1, . . . , R}, where Qr is p×k,
xj = (x1j , . . . , xkj )> is k × 1, and the dimension k (k ≥ p) is arbitrary. Moreover,
{xij , i = 1, . . . , k, j = 1, . . . , n} is a k × n array of independent random variables, not
necessarily identically distributed, with common moments
Exij = 0, E|x2ij | = 1,
and satisfying the following Lindeberg-type condition: for each η > 0,
n
R
k
√
1 XXX
2
2
kqir k E|xij |I |xij | > η n/kqir k → 0,
pnη 2 i=1 j=1 r=1
where kqir k is the Euclidean norm of the i-th column vector qir of Qr .
Assumption (c) The ESD Hnr of the population covariance matrix Tnr = Qr Q>
r converges weakly to a probability distribution Hr , r = 1, . . . , R. Also the sequence of
the spectral norm of (Tnr ) is bounded in n and r.
Lemma 2.1. [Theorem 2.1 of Zheng et al. (2017b)] Under Assumptions (a)-(c), almost
surely, the ESD Fnr of Bnr weakly converges to a nonrandom LSD F c,Hr . Moreover, the
Stieltjes transform mr (z) of F c,Hr is the unique solution to the following Marčenko-Pastur
equation
Z
1
mr (z) =
dHr (t) ,
(2.1)
t[1 − c − czmr (z)] − z
on the set {mr (z) ∈ C : −(1 − c)/z + cmr (z) ∈ C+ }.
Define the companion LSD of Bnr as
F c,Hr = (1 − c)δ0 + cF c,Hr .
It is readily checked that F c,Hr is the LSD of the companion sample covariance matrix Bnr =
n−1 X∗n Q>
r Qr Xn (which is n × n), and its Stieltjes transform mr (z) = −(1 − c)/z + cmr (z)
satisfies the so-called Silverstein equation
Z
t
1
+c
dHr (t).
(2.2)
z=−
mr (z)
1 + tmr (z)
2.2. Main Results
Let A and B be two real symmetric p × p matrices satisfying AB = BA. The two matrices
can then be diagonalized simultaneously. We define the joint spectral distribution of (A, B)
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
7
as the two-dimensional spectral distribution of the complex matrix A + iB, i.e.,
1
G(x, y) = #{i ≤ p, <(si ) ≤ x, =(si ) ≤ y},
p
where (si ) are the p eigenvalues of A + iB and #E denotes the cardinality of a set E.
Recall the random vector of L × R LSSs of Bnr ’s
Z
f`r (x)dFnr (x)
,
(2.3)
1≤`≤L,1≤r≤R
where (Fnr ) are the corresponding empirical spectral distributions of (Bnr ) and (f`r ) are
L × R measurable functions on the real line. Our aim in this section is to establish the joint
distribution of (2.3) under suitable conditions. The main results are presented as follows.
Assumption (d) The variables {xij , i = 1, . . . , k, j = 1, . . . , n} are independent, with
common moments
Exij = 0, E|x2ij | = 1, βx = E|x4ij | − |Ex2ij |2 − 2, and αx = |Ex2ij |2 ,
and satisfying the following Lindeberg-type condition: for each η > 0
k
n
R
p
1 XXX
2
4
n/kq
k
→ 0.
kq
k
E|x
|I
|x
|
>
η
ir
ir
ij
ij
pnη 6 i=1 j=1 r=1
(2.4)
Assumption (e) Either βx = 0, or the mixing matrices {Qr } are such that the matrices
{Q>
r Qr } are diagonal (therefore with arbitrary βx ).
Assumption (f ) The joint spectral distribution Hnrs of Tnr and Tns converges weakly
to a probability distribution Hrs , 1 ≤ r, s ≤ R.
The framework with Assumptions (d)-(e)-(f) is inspired by the one advocated in Zheng
et al. (2017b). However, an extension is necessary here since we are dealing with several
random matrices simultaneously while only one matrix is considered in the reference.
Theorem 2.1. Under Assumptions (a)-(f ), let f11 , . . . , fLR be L × R functions analytic
on a complex domain containing
√
√ 2
[I(0<c<1) (1 − c)2 lim inf λT
c) lim sup λT
(2.5)
min , (1 +
max ]
n
n
T
with T = {Tnr }, and λT
min and λmax denoting the smallest and the largest eigenvalue of all
the matrices in T, respectively. Then, the random vector
Z
Z
cn ,Hnr
p
f`r (x)dFnr (x) − f`r (x)dF
(x)
.
(2.6)
1≤`≤L,1≤r≤R
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
8
converges to an (L × R)-dimensional Gaussian random vector (Xf11 , . . . , XfLR ). The mean
function is
I
1
βx
αx
+
dz,
EXf`r = −
f`r (z)g1 (z)
2πi C1
(1 − g2 (z))(1 − αx g2 (z)) 1 − g2 (z)
where
Z
g1 (z) =
cm3r (z)t2
dHr (t) and
(1 + tmr (z))3
cm2r (z)t2
dHr (t).
(1 + tmr (z))2
Z
g2 (z) =
The covariance function is
1
Cov(Xf`0 r , Xf`s ) =
4π 2
I I
f`0 (z1 )f` (z2 )
∂ 2 g(z1 , z2 )
dz1 dz2 ,
∂z1 ∂z2
(2.7)
C1 C2
where g(z1 , z2 ) = log(1 − a(z1 , z2 )) + log(1 − αx a(z1 , z2 )) − βx a(z1 , z2 ) with
ZZ
cmr (z1 )ms (z2 )t1 t2
dHrs (t1 , t2 ).
a(z1 , z2 ) =
(1 + t1 mr (z1 ))(1 + t2 ms (z2 ))
The contours C1 and C2 are non-overlapping, closed, positively orientated in the complex
plane, and enclosing both the supports of F c,Hr and of F c,Hs .
Remark 1. The centralization term in (2.6) is the expectation of f with respect to the
distribution F cn ,Hnr . This distribution is a finite dimensional version of the LSD F c,Hr ,
which is defined by (2.1) with the parameters (c, Hr ) replaced with (cn , Hnr ). The use of
F cn ,Hnr instead of F c,Hr aims to eliminate the effect of the convergence rate of (cn , Hnr ) to
(c, Hr ).
As an illustrative example of Theorem 2.1, we consider a simplified case where only
two sample covariance matrices are involved, i.e. Xn X0n /n and QXn X0n Q> /n, where Xn
is a p × n matrix of i.i.d. real standard Gaussian variables. The corresponding population
covariance matrices are Ip and Tn := QQ> , respectively. It’s clear that the ESD and its
limit of the identity matrix Ip are both the Dirac measure δ1 . Those of Tn are general
and denoted by Hn and H, respectively. Moreover, the joint spectral distribution function
Hn12 (t1 , t2 ) of Ip and Tn is equal to Hn (t2 ) for t1 = 1 and zero otherwise. Denote the ESDs
of the two sample covariance matrices by Fn1 and Fn2 , respectively, and let
Gn1 (x) = Fn1 (x) − F cn ,δ1 (x) and Gn2 (x) = Fn2 (x) − F cn ,Hn2 (x).
Then for any analytic function f , we have
Z
>
Z
d
p
f (x)dGn1 (x), f (x)dGn2 (x)
→
− N
!
!!
v1
ψ11 ψ12
,
.
v2
ψ12 ψ22
(2.8)
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
9
The parameters (v1 , v2 , ψ11 , ψ22 ) of the marginal distributions in (2.8) have been derived by
many authors, see Bai and Silverstein (2004) and Zheng et al. (2017b) for example. While
the covariance parameter ψ12 is new and, from Theorem 2.1, it can be formulated as
I I
c
f (z1 )f (z2 )(m1 (z1 ) + z1 m01 (z1 ))(m2 (z2 ) + z2 m02 (z2 ))
dz1 dz2 ,
v12 = − 2
2π C1 C2
[c − (1 + z1 m1 (z1 )(1 + z2 m2 (z2 )]2
where m1 (z) and m2 (z) are the companion Stieltjes transforms of the LSDs F c,δ1 (x) and
F c,H (x), respectively, and m0 (z) denotes the derivative of m(z) with respect to z. For the
R
simplest function f (z) = z, one may figure out v12 = 2c tdH(t) by the residual theorem.
3. Application to high dimensional white noise test
As discussed in the introduction, a notable application of the joint CLT presented in this
paper is to the high dimensional white noise test. In particular, it is expected that testing
power could be gained by accumulating information across different lags, that is, by using
P
fτ M
f ∗ ) defined in (1.5).
the test statistic Lq = qτ =1 Tr(M
τ
Define the scaled statistic
n
qp
φq = Lq − .
(3.1)
p
2
The null hypothesis will be rejected for large values of φq . We consider high-dimensional
situations where the dimension p is large compared to the sample size n. By applying the
CLT in Theorem 2.1, the asymptotic null distribution of φq is derived as follows.
Theorem 3.1. Let q ≥ 1 be a fixed integer, and assume that
1. {zit , i = 1, · · · , p, t = 1, · · · , n} is a set of i.i.d. real-valued variables satisfying
Ezit = 0, Ezit2 = 1, Ezit4 = ν4 < ∞;
2. Relaxed Marčenko-Pastur regime: both the sample size n and the dimension p grow
to infinity such that
p
p
0 < lim inf ≤ lim sup < ∞.
n→∞ n
n→∞ n
Then in the simplest setting where xt = zt , we have
q d
s(cn )−1/2 {φq − } →
− N (0, 1),
(3.2)
2
where s(u) = q + u(ν4 − 1)(q 2 + q/2).
The proof of this theorem is given in Section 4.
Let Zα be the upper-α quantile of the standard normal distribution at level α. Based on
Theorem 3.1, we obtain a procedure for testing the null hypothesis in (1.3) as follows.
p
q
Multi-Lag-q test: Reject H0 if φq − > Zα s(cn ).
(3.3)
2
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
10
3.1. Simulation Experiments
Most of the experiments of this section are designed to compare our test procedure in (3.3)
and the procedure based on the test statistic Uq from Li et al. (2016) using Simes’ method
(Simes, 1986). In Li et al. (2016), several testing procedures are discussed and the test Uq
performs quite satisfactorily in terms of both size and power across different scenarios.
More precisely, let q ≥ 1 be a fixed integer, define p(q + 1)-dimensional vectors yj =
xj(q+1)−q
h i
..
n
. Since Ext = 0 and Στ = Cov(xt+τ , xt ), we have
,
j
=
1,
.
.
.
,
N
,
N
=
.
q+1
xj(q+1)
Σ0 Σ1 · · · Σq
.
.
Σ1 Σ0 . . ..
.
.. . .
..
. Σ1
.
.
Σq · · · Σ1 Σ0 (q+1)p×(q+1)p
Cov(yj ) =
The null hypothesis H0 : Cov(xt+τ , xt ) = 0, τ = 1, · · · , q becomes H0 : Σ1 = · · · = Σq =
0, a test for a block diagonal covariance structure of the stacked sequence {yj }.
When Σ0 = σ 2 Ip , the white noise test of {xt } reduces to a sphericity test of {yj }.
The well known John’s test statistic Uq can be adopted for this purpose. In our case, the
corresponding John’s test statistic Uq is defined as
Uq =
1
p(q+1)
Pp(q+1)
i=1
lq
li,q − lq
2
2
,
where {li,q , i = 1, . . ., p(q + 1)} are the eigenvalues of the sample covariance matrix Sq =
PN
1
∗
j=1 yj yj , and l q is their average.
N
Notice however
h ithat the use of blocks above reduces the sample size n to the number of
n
blocks N = q+1
. This may result in a certain loss of power for the test. To limit such
loss of power, we adopt Simes’ method for multiple hypothesis testing in Simes (1986). To
implement Simes’ method, we denote
(0)
yj = yj (x1 , . . . , xn )
as the previously defined stacked sample. Then we rotate the sample (xj ) and define a
(k)
series of new stacked samples yj for k = 1, . . . , q, that is,
(k)
yj = yj (xk+1 , . . . , xn , x1 , . . . , xk )
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
11
Then John’s test statistic Uq can be calculated based on the q + 1 samples, which results
(k)
in q + 1 different statistics {Uq }. Moreover, let Pk , 0 ≤ k ≤ q, denote the (asymptotic)
P-value for the John’s test with the k-th set of yj ’s, i.e.
Pk = 1 − Φ (N Uq(k) − p(q + 1) − ν4 + 2)/2 ,
where Φ(·) is the cumulative distribution function of the standard normal distribution. Let
P(1) ≤ · · · ≤ P(q+1) be a permutation of P0 , . . . , Pq . Then by the Simes method, we reject
k
α at least for one 1 ≤ k ≤ q + 1 for the nominal level α.
H0 if P(k) ≤ q+1
To compare our test statistic φq with multi-lag-q John’s test statistic Uq , we set the
significance level α = 5% and the critical regions of the two tests are
p
(1) Our test φq : {φq > 2q + Z0.95 s(cn ) };
(2) Multi-lag-q John’s test Uq (using Simes’ method): {at least for one 1 ≤ k ≤ q +
k
1, P(k) ≤ q+1
0.05}.
Data are generated following four different scenarios for comparison:
i.i.d.
(I) Test size under Gaussian white noise: xt = zt , (zt ) ∼ Np (0, Ip );
i.i.d.
(II) Test size under Non-Gaussian white noise: xt = zt − 2, (zit ) ∼ Gamma(4, 0.5),
E(zit ) = 2, Var(zit ) = 1, ν4 (zit ) = 4.5;
(III) Test power under a Gaussian spherical AR(1) process: xt = Axt−1 + zt , A = aIp ,
i.i.d.
a = 0.1, (zt ) ∼ Np (0, Ip );
(IV) Test power under a Non-Gaussian spherical AR(1) process: xt = Axt−1 + (zt − 2),
i.i.d.
A = aIp , a = 0.1, (zit ) ∼ Gamma(4, 0.5), E(zit ) = 2, Var(zit ) = 1, ν4 (zit ) = 4.5.
Various (p, n)-combinations are tested to show the suitability of our test statistic for both
low and high dimensional settings. Empirical statistics are obtained using 2000 independent
replications. Table 1 compares the empirical sizes of the two tests φq and Uq . It can be
seen that both of them have reasonable sizes compared to the 5% nominal level across
all the tested (p, n)-combinations. Still, the two tests become slightly conservative under
Non-Gaussian distributions in Scenario (II) compared to the Gaussian case in Scenario (I).
A sample display of these sizes is given in Figure 1 (left panel).
In Table 2, we compare the power of the two tests. Our test φq displays a generally
much higher power than the multi-lag-q John’s test Uq , especially when the dimensions
(p, n) become larger. On the other hand, both tests have slightly lower power under the
Non-Gaussian distribution than under the Gaussian distribution, which is consistent with
the previous observation that the two tests become more conservative with Non-Gaussian
populations. A sample display of these powers is given in Figure 1 (right panel).
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
0.08
12
1
φq(q=1)
φq(q=1)
φq(q=3)
φq(q=3)
Uq(q=1)
0.07
0.8
Uq(q=3)
Uq(q=1)
Uq(q=3)
Size
Power
0.06
0.5
0.05
0.04
0.03
50
0.2
200
300
500
(p/n=0.5) dimension p
0
50
200
300
500
(p/n=0.5) dimension p
Fig 1. Sample plots of empirical sizes (left panel) from Table 1 (Scenario II with cn = 0.5) and
empirical powers (right panel) from Table 2 (Scenario III with cn = 0.5).
To further explore the powers of the two tests, we varied the AR coefficient a in Scenario (III) and (IV) from -0.1 to 0.1 (a = 0 corresponds to testing size). Smaller values
of the AR(1) coefficient a are used here leading to a more difficult testing problem and a
generally decreased power for both tests. Three dimensional settings are considered with
p/n ∈ {0.1, 0.5, 1.5} while the sample size is fixed as n = 600. The number of independent
replications is still 2000 in each case. Results for Scenario (III) and (IV) are plotted in
Figure 1. This Figure further consolidates that our test φq dominates Uq under all tested
scenarios. A nonnegligible increase in the testing power of both test statistics as the dimension p becomes larger sheds more light on the blessings of high dimensionality. Still both
tests are more conservative with Non-Gaussian population distribution than with Gaussian
distribution.
3.2. Comparison to a permutation test
As many complex analytic tools are employed to derive the asymptotic null distributions of
the test statistic φq , it is natural to wonder about the performance of a “simple-minded” test
procedure, namely the permutation test. Under the null hypothesis of white noise, since the
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
1
1
φq(q=1)
φq(q=1)
0.8
φq(q=3)
Power Values
Power Values
0.8
Uq(q=1)
0.5
Uq(q=3)
0.2
0.05
0
−0.1
Uq(q=1)
0.5
−0.05
0
0.05
(p/n=0.1) AR coefficient a
0.05
0
−0.1
0.1
Power Values
Power Values
Uq(q=3)
0.2
−0.05
0
0.05
(p/n=0.5) AR coefficient a
0.5
0.05
0
−0.1
0.1
−0.05
0
0.05
(p/n=0.5) AR coefficient a
0.1
φq(q=1)
0.8
φq(q=3)
Power Values
Power Values
Uq(q=3)
1
Uq(q=1)
Uq(q=3)
0.2
0.05
0
−0.1
φq(q=3)
Uq(q=1)
φq(q=1)
0.5
0.1
0.2
1
0.8
−0.05
0
0.05
(p/n=0.1) AR coefficient a
φq(q=1)
0.8
φq(q=3)
Uq(q=1)
0.05
0
−0.1
Uq(q=3)
1
φq(q=1)
0.5
φq(q=3)
0.2
1
0.8
13
φq(q=3)
Uq(q=1)
0.5
Uq(q=3)
0.2
−0.05
0
0.05
(p/n=1.5) AR coefficient a
0.1
0.05
0
−0.1
−0.05
0
0.05
(p/n=1.5) AR coefficient a
0.1
Fig 2. Empirical Powers for the two tests with varying AR coefficient a from -0.1 to 0.1. Left panel:
Scenario(III) for Gaussian Distribution. Right panel: Scenario(IV) for Gamma Distribution
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
14
sample vectors x1 , . . . , xn have an i.i.d. structure, one can permute these n sample vectors
say B times to obtain an empirical upper 5% quantiles of the test statistic φq . The null
hypothesis will be rejected if the observed statistic φq from the original (non permuted)
sample vectors x1 , . . . , xn is larger than this empirical quantile.
Data are generated following the spherical AR(1) process in Scenario (III) and (IV) to
compare this straightforward test with our test statistic φq . In order to compare the power
performance of two tests, the AR coefficient a takes different values, a = [0, 0.05, 0.09, 0.1]
(a = 0 corresponds to testing size). The sample size is fixed as n = 300 yet data dimension
p varies. As for the permutation test, the permutation times is set as B = 500. The nominal
level is α = 5%. Testing size and power of two tests are shown in Tables 3 and 4 based on
500 replicates for all (p, n) configurations.
It can be seen that the sizes of both tests are well controlled. As for their power, our test
offers an acceptable level while the permutation test consistently performs better in the
tested cases. However, the permutation test is extremely time consuming compared to our
test. For instance, to run one set of (p, n) = (150, 300) combination for 500 replicates, it
takes only 25 seconds with our test, while almost 3 hours for the permutation test with
permutation times B = 500. Particularly the computation time increases greatly when the
sample size n grows. Therefore, our test statistic φq provides a very competitive choice for
testing high dimensional white noise while the classical permutation test is simpler, more
powerful though much slower.
4. Proofs of the main theorems
4.1. Proof of Theorem 2.1
The general strategy for our main Theorem 2.1 follows the methods advocated in Bai and
Silverstein (2004), with its most recent update in Zheng et al. (2017b). However, as we
are dealing with several random matrices simultaneously, all the technical steps for the
implementation of this strategy have to be carefully rewritten. They are presented in this
section.
4.1.1. Sketch of the proof of Theorem 2.1
Let v0 > 0 be arbitrary, xr be any number greater than the right end point of interval
(2.5), and xl be any negative number if the left end point of (2.5) is zero, otherwise choose
√ 2
c) ). Define a contour C as
xl ∈ (0, lim inf p→∞ λT
min (1 −
C = {x + iv : x ∈ {xr , xl }, v ∈ [−v0 , v0 ]} ∪ Cu with Cu = {x ± iv0 : x ∈ [xl , xr ]},
(4.1)
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
15
and let Cn = Cu ∪ {x ± iv : x ∈ {xl , xr }, v ∈ [n−1 εn , v0 ]} with εn ≥ n−α for some α ∈ (0, 1).
By definition, the contour C encloses a rectangular region in the complex plane, which
contains the union of the support sets of all the LSDs F c,Hr , 1 ≤ r ≤ R. As a regularized
version of C, Cn excludes a small segment near the real line.
Let mnr (z), mnr (z), m0nr (z), m0nr (z) be the Stieltjes transforms of Fnr , F nr , F cn ,Hnr , and
F cn ,Hnr , respectively, where Fnr is the ESD of Bnr , F cn ,Hnr is the LSD defined in Remark
1, F and F are linked by the equation F = (1 − cn )δ0 + cn F . A major task of proving
Theorem 2.1 is to study the convergence of the empirical process
Mnr (z) := p[mnr (z) − m0nr (z)] = n[mnr (z) − m0nr (z)].
To this end, we need to truncate Mnr (z) as
z ∈ Cn ,
Mnr (z)
cnr (z) = M (x + in−1 ε ) x ∈ {x , x } and v ∈ [0, n−1 ε ],
M
nr
n
l
r
n
−1
−1
Mnr (x − in εn ) x ∈ {xl , xr } and v ∈ [−n εn , 0],
which agrees to Mnr (z) on Cn . This truncation is essential when proving the tightness
cnr (z) on C. Write
M
c
c
c
Mn (z) = Mn1 (z), . . . , MnR (z) ,
we will establish its convergence as stated in the following lemma.
c n (·) converges weakly to a Gaussian process
Lemma 4.1. Under Assumptions (a)-(f ), M
M(·) = (M1 , . . . , MR )(·) on C. The mean function is
EMr (z) =
αx g1 (z)
βx g1 (z)
+
,
(1 − g2 (z))(1 − αx g2 (z)) 1 − g2 (z)
where
Z
g1 (z) =
cm3r (z)t2
dHr (t) and
(1 + tmr (z))3
Z
g2 (z) =
cm2r (z)t2
dHr (t).
(1 + tmr (z))2
The covariance function is
Cov(Mr (z1 ), Ms (z2 )) = −
∂2
[log(1 − a(z1 , z2 )) + log(1 − αx a(z1 , z2 )) − βx a(z1 , z2 )]
∂z1 ∂z2
where
ZZ
a(z1 , z2 ) = c
t1 t2 mr (z1 )ms (z2 )
dHrs (t1 , t2 ).
(1 + t1 mr (z1 ))(1 + t2 ms (z2 ))
From this lemma, Theorem 2.1 follows by similar arguments on Pages 562 and 563 in
Bai and Silverstein (2004).
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
16
4.1.2. Proof of Lemma 4.1
Following closely the steps of truncation, centralization and rescaling in Appendix B of
Zheng et al. (2017b), one may find that it is sufficient to prove this lemma under the
assumption that
√
ηn n
|xij | <
,
(4.2)
max1≤r≤R {kqir k}
where the constant ηn → 0 as n → ∞.
Write for r ∈ {1, . . . , R} and z ∈ Cn ,
Mnr (z) = p[mnr (z) − Emnr (z)] + p[Emnr (z) − m0nr (z)]
1
2
:= Mnr
(z) + Mnr
(z).
The Lemma can be proved by verifying three conditions (Bai and Silverstein, 2004):
Condition 1: Finite dimensional convergence of Mn1 (z) in distribution;
Condition 2: Tightness of Mn1 (z) on Cn ;
Condition 3: Convergence of Mn2 (z).
Since the second and third conditions can be obtained directly from Lemma 5.1 in Zheng
et al. (2017b), we only consider the first one by showing that, for any W × R complex
1
(zjr ))1≤j≤W,1≤r≤R converges to a Gaussian
numbers z11 , . . . , zW R , the random vector (Mnr
vector. Without loss of generality, we assume max{kQr k} ≤ 1. We will also denote by K
any constants appearing in inequalities and K may take on different values for different
expressions.
√
With the notation rjr = (1/ n)Qr xj , we define some quantities:
Dr (z) = Bnr − zI, Djr (z) = Dr (z) − rjr r∗jr , Dijr (z) = Dr (z) − rir r∗ir − rjr r∗jr ,
−1
−1
jr (z) = r∗jr D−1
jr (z)rjr − n trTnr Djr (z),
−2
−1
δjr (z) = r∗jr D−2
jr (z)rjr − n trTnr Djr (z),
1
1
1
, βijr (z) =
, β̄jr (z) =
,
βjr (z) =
−1
−1
∗
∗
−1
1 + rjr Djr (z)rjr
1 + rir Dijr (z)rir
1 + n trTnr D−1
jr (z)
1
1
bnr (z) =
,
b
(z)
=
,
12r
1 + n−1 EtrTnr D−1
1 + n−1 EtrTnr D−1
r (z)
12r (z)
which will be frequently used in the sequel. Note that quantities in the last two rows are
all bounded in absolute value by |z|/=(z).
1
By martingale difference decomposition, the process Mnr
(z) can be expressed as
−1
p[mnr (z) − Emnr (z)] = tr[D−1
r (z) − EDr (z)]
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
=
n
X
17
−1
−1
−1
trEj [D−1
r (z) − Djr (z)] − trEj−1 [Dr (z) − Djr (z)]
j=1
n
X
= −
(Ej − Ej−1 )βjr (z)r∗jr D−2
jr (z)rjr
j=1
n
dX
(Ej − Ej−1 ) log βjr (z)
= −
dz j=1
n
dX
=
(Ej − Ej−1 ) log 1 + jr (z)β̄jr (z) ,
dz j=1
−1
−1
−1
∗
where the third equality is from the identity D−1
r (z) = Djr (z) − Djr (z)rjr rjr Djr (z)βjr (z)
and the last one is obtained using the identity βjr (z) = β̄jr (z)[1 + β̄jr (z)jr (z)]−1 . We next
show that
n
n
dX
dX
(Ej − Ej−1 ) log 1 + jr (z)β̄jr (z) −
(Ej − Ej−1 )jr (z)β̄jr (z) = op (1).
dz j=1
dz j=1
Considering the second moment of the above difference, by the Cauchy integral formula,
one may get
2
n
dX
E
(Ej − Ej−1 ) log 1 + jr (z)β̄jr (z) − jr (z)β̄jr (z)
dz j=1
I
n
X
log 1 + jr (ζ)β̄jr (ζ) − jr (ζ)β̄jr (ζ)
1
=E
dζ
2
2πi
(z
−
ζ)
|ζ−z|=v/2
j=1
n I
K X
E|jr (ζ)β̄jr (ζ)|4 |dζ|.
≤ 2 4
π v j=1 |ζ−z|=v/2
2
(4.3)
From Lemma A.1 and the truncation in (4.2), we have
k
2 X
K
−1
−1
4
8
> −1
E|jr (ζ)| ≤ 4 E trTnr Djr (ζ)Tnr Djr (ζ̄) +
E|xij | E|qir Djr (ζ)qir |
n
i=1
4
≤ Kn−2 + Kηn4 n−1 ,
by which the right hand side of (4.3) tends to zero. Therefore, we need only to consider
the limiting distribution of
n
n
dX
dX
(Ej − Ej−1 )jr (z)β̄jr (z) =
Ej jr (z)β̄jr (z)
dz j=1
dz j=1
(4.4)
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
18
in finite dimensional situations. To verify the Lyapunov condition, one can show that
n
2
X
d
d
E (Ej − Ej−1 ) jr (z)β̄jr (z) I (Ej − Ej−1 ) jr (z)β̄jr (z) ≥
dz
dz
j=1
n
1X
d
≤ 2
E Ej jr (z)β̄jr (z)
j=1
dz
4
→ 0,
where the convergence is again from Lemma A.1 and (4.2). Hence, from the martingale
1
(Billingsley, 1995, Theorem 35.12), the random vector (Mnr
(zjr )) will tend to a Gaussian
vector (Mr (zjr )) with covariance function
n
X
∂
∂
jr (z1 )β̄jr (z1 ) · Ej
js (z2 )β̄js (z2 ) . (4.5)
Cov(Mr (z1 ), Ms (z2 )) = lim
Ej−1 Ej
n→∞
∂z1
∂z2
j=1
We note that the referenced martingale CLT applies also to multidimensional martingale
by considering arbitrary linear combination of its components.
Using the same approach of Bai and Silverstein (2004) on Page 571, one may replace
β̄jr (z) by bnr (z). Then, by (1.15) of Bai and Silverstein (2004), we have
Γnrs (z1 , z2 ):=
n
X
bnr (z1 )bns (z2 )Ej−1 [Ej jr (z1 )Ej js (z2 )]
j=1
"
n
1 X
−1
> −1
= 2
bnr (z1 )bns (z2 ) trEj Q>
r Djr (z1 )Qr Ej Qs Djs (z2 )Qs
n j=1
−1
> −1
+ αx trEj Q>
r Djr (z1 )Qr Ej Qs Djs (z2 )Qs
+ βx
k
X
#
−1
−1
>
q>
ir Ej Djr (z1 )qir qis Ej Djs (z2 )qis
i=1
:= Γ1 + αx Γ2 + βx Γ3 ,
(4.6)
where αx = |Ex211 |2 and βx = E|x411 | − |Ex211 |2 − 2.
Now we derive the limit of the first term in (4.6). The means is to replace D−1
jr (z) (and
−1
similarly Djs (z)) by a proper nonrandom matrix. For this, we introduce such a one
Lr (z) = zI −
n−1
b12r (z)Tnr ,
n
whose inverse spectral norm is bounded, that is,
||Lr (z)||−1 ≤
|b−1
1 + p/(nv)
12r (z)|
≤
.
−1
v
=(zb12r (z))
(4.7)
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
19
−1
−1
∗
We will show that the major part of D−1
jr (z) is just −Lr (z). From the identity rir Djr (z) =
βijr (z)r∗ir D−1
ijr (z), we get
−1
−1
−1
D−1
jr (z) + Lr (z) = Lr (z) (Djr (z) + Lr (z)) Djr (z)
X
= L−1
r (z)
i6=j
X
= L−1
r (z)
n−1
b12r (z)Tnr
rir r∗ir −
n
βijr (z)rir r∗ir D−1
ijr (z) −
i6=j
!
D−1
jr (z)
n−1
b12r (z)Tnr D−1
jr (z)
n
= b12r (z)R1r (z) + R2r (z) + R3r (z),
!
(4.8)
where
R1r (z) =
X
−1
∗
−1
L−1
r (z)(rir rir − n Tnr )Dijr (z),
i6=j
X
−1
∗
(βijr (z) − b12r (z))L−1
R2r (z) =
r (z)rir rir Dijr (z),
i6=j
R3r (z) = n−1 b12r (z)L−1
r (z)Tnr
X
−1
D−1
ijr (z) − Djr (z) .
i6=j
−1
From this decomposition, after substituting −L−1
r (z) for Djr (z) in the first term in (4.6),
there are three remaining quantities. Let’s check which one (or ones) of them can be
omitted. From Lemma A.3, (4.7), and (4.3) of Bai and Silverstein (1998), for any p × p
matrix M, we have
2
1/2
E|trR2r (z)M| ≤ nE
2
1/2
(|β12r (z) − b12r (z)| )E
≤ n1/2 K|||M|||
−1
r∗1r D−1
12r MLr (z)r1r
|z|2 (1 + p/(nv))
,
v5
(4.9)
where |||M||| denotes a nonrandom bound on the spectral norm of M. From Lemma A.2,
|trR3r (z)M| ≤ |||M|||
|z|(1 + p/(nv))
.
v3
(4.10)
Again from Lemma A.3 and (4.7), for nonrandom M,
−1
−1
−1
−1
2
E|trR1r (z)M| ≤ nE1/2 |r∗ir D−1
ijr (z)MLr (z)rir − n trTnr Dijr (z)MLr (z)|
≤ n1/2 K||M||
(1 + p/(nv))
.
v2
(4.11)
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
20
Therefore, quantities containing R2r (z) and R3r (z) are both negligible. For the quantity
−1
−1
−1
∗
involving R1r (z), applying the identity D−1
js (z) = Dijs (z) − Dijs (z)rjs rjs Dijs (z)βijs (z), it
can be divided into three parts, that is,
> −1
trQ>
r Ej (R1r (z1 ))Qr Qs Djs (z2 )Qs = R11 (z1 , z2 ) + R12 (z1 , z2 ) + R13 (z1 , z2 ),
(4.12)
where
R11 (z1 , z2 ) = −
X
−1
> −1
∗
> −1
βijs (z2 )r∗ir Ej (D−1
ijr (z1 ))Qr Qs Dijs (z2 )ris ris Dijs (z2 )Qs Qr Lr (z1 )rir ,
i<j
R12 (z1 , z2 ) = −tr
X
−1
−1
−1
−1
>
>
L−1
r (z1 )n Tnr Ej (Dijr (z1 ))Qr Qs (Djs (z2 ) − Dijs (z2 ))Qs Qr ,
i<j
R13 (z1 , z2 ) = tr
X
−1
∗
−1
> −1
>
L−1
r (z1 )(rir rir − n Tnr )Ej (Dijr (z1 ))Qr Qs Dijs (z2 )Qs Qr .
i<j
From Lemma A.2 and (4.7) we get |R12 (z1 , z2 )| ≤ (1 + p/(nv0 ))/v03 , and similar to (4.9),
E|R13 (z1 , z2 )| ≤ n1/2 (1 + p/(nv0 ))/v03 . Thus these two parts are trivial. We then turn to
dealing with R11 (z1 , z2 ). Using Lemma A.1, Lemma A.3, and (4.3) of Bai and Silverstein
(1998) we get, for i < j,
−1
> −1
∗
> −1
E βijs (z2 )r∗ir Ej (D−1
ijr (z1 ))Qr Qs Dijs (z2 )ris ris Dijs (z2 )Qs Qr Lr (z1 )rir
−1
> −1
>
> −1
>
−b12s (z2 )n−2 tr Ej (D−1
ijr (z1 ))Qr Qs Dijs (z2 )Qs Qr tr Dijs (z2 )Qs Qr Lr (z1 )Qr Qs }
≤ Kn−1/2 .
So we may simplify R11 (z1 , z2 ) by replacing βijs (z2 ) with b12s (z2 ) and remove the random
parts of rir and ris . By Lemma A.2, we have
−1
> −1
>
> −1
>
tr Ej (D−1
ijr (z1 ))Qr Qs Dijs (z2 )Qs Qr tr Dijs (z2 )Qs Qr Lr (z1 )Qr Qs }
−1
> −1
>
> −1
>
≤ Kn.
−tr Ej (D−1
jr (z1 ))Qr Qs Djs (z2 )Qs Qr tr Djs (z2 )Qs Qr Lr (z1 )Qr Qs }
−1
−1
It implies that we may further replace D−1
ijr (z1 ) and Dijs (z2 ) in R11 (z1 , z2 ) with Djr (z1 )
and D−1
js (z2 ), respectively, which yields
E R11 (z1 , z2 )
+
j−1
−1
> −1
> −1
> −1
b12s (z2 )tr Q>
r Ej (Djr (z1 ))Qr Qs Djs (z2 )Qs tr Qs Djs (z2 )Qs Qr Lr (z1 )Qr
2
n
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
≤ Kn1/2 .
21
(4.13)
Integrating the results in (4.8)-(4.13), we obtain
−1
> −1
tr Q>
r Ej (Djr (z1 ))Qr Qs Djs (z2 )Qs
j−1
> −1
> −1
b12r (z1 )b12s (z2 )tr Qs Djs (z2 )Qs Qr Lr (z1 )Qr
× 1+
n2
> −1
−1
= −trQ>
r Lr (z1 )Qr Qs Djs (z2 )Qs + R14 (z1 , z2 ),
(4.14)
where E|R14 (z1 , z2 )| ≤ Kn1/2 . Furthermore, from this and (4.8)-(4.11), we may substitute
−1
for the second and third D−1
js (z2 ) in (4.14) with −Ls (z2 ) and then get
−1
> −1
tr Q>
E
(D
(z
))Q
Q
D
(z
)Q
j
1
r
2
s
r
s
jr
js
j−1
> −1
> −1
b12r (z1 )b12s (z2 )tr Qs Ls (z2 )Qs Qr Lr (z1 )Qr
× 1−
n2
−1
> −1
= trQ>
r Lr (z1 )Qr Qs Ls (z2 )Qs + R15 (z1 , z2 ),
(4.15)
where E|R15 (z1 , z2 )| ≤ Kn1/2 .
From Lemma A.2 and (4.3) of Bai and Silverstein (1998), we have
|b12r (z) − bnr (z)| ≤ Kn−1
and |bnr (z) − Eβ1r (z)| ≤ Kn−1/2 ,
respectively. By (2.2) of Silverstein (1995) and discussions in Section 5 of Bai and Silverstein
(1998), we have
Eβ1r (z) = −zEmnr (z) and |Emnr (z) − m0nr (z)| ≤ Kn−1 ,
respectively. Therefore, we get
|b12r (z) + zm0nr (z)| ≤ Kn−1/2 .
Combining this and (4.15), it follows that
j−1 0
tr
1−
mnr (z1 )m0ns (z2 )
n2
>
0
−1
>
0
−1
×tr Qs (I + mns (z2 )Tns ) Qs Qr (I + mnr (z1 )Tnr ) Qr
j−1
−1
>
> −1
= tr Qr Ej (Djr (z1 ))Qr Qs Djs (z2 )Qs 1 −
cn m0nr (z1 )m0ns (z2 )
n
Z Z
t1 t2 dHprs (t1 , t2 )
×
(1 + t1 m0nr (z1 ))(1 + t2 m0ns (z2 ))
−1
> −1
Q>
r Ej (Djr (z1 ))Qr Qs Djs (z2 )Qs
(4.16)
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
ncn
=
z1 z2
Z Z
t1 t2 dHprs (t1 , t2 )
+ R16 (z1 , z2 ),
(1 + t1 m0nr (z1 ))(1 + t2 m0ns (z2 ))
22
(4.17)
where E|R16 (z1 , z2 )| ≤ Kn1/2 . Similar to the arguments in Bai and Silverstein (2004) (page
577), using (4.17) and letting
Z Z
t1 t2 dHprs (t1 , t2 )
0
0
an (z1 , z2 ) = cn mnr (z1 )mns (z2 )
,
(1 + t1 m0nr (z1 ))(1 + t2 m0ns (z2 ))
we get
n
1X
an (z1 , z2 )
i.p.
Γ1 =
+ op (1) −→ − log(1 − a(z1 , z2 )),
n j=1 1 − ((j − 1)/n)an (z1 , z2 )
(4.18)
where
Z Z
a(z1 , z2 ) = cmr (z1 )ms (z2 )
t1 t2 dHrs (t1 , t2 )
.
(1 + tr mr (z1 ))(1 + ts ms (z2 ))
Similar to the derivation of Γ1 , one can easily show that
n
1X
αx an (z1 , z2 )
i.p.
αx Γ2 =
+ op (1) −→ − log(1 − αx a(z1 , z2 )).
n j=1 1 − αx ((j − 1)/n)an (z1 , z2 )
(4.19)
Considering the third term of (4.6), βx Γ3 with βx 6= 0, from Assumption (e), the matrix
> −1
Q>
r Qr is diagonal, so is Qr Lr Qr . Using (4.8)-(4.11) and (4.16), we have
n
k
1 XX > >
−1
> >
βx Γ3 = βx bnr (z1 )bns (z2 ) 2
e Q Ej (D−1
jr (z1 ))Qr ei ei Qs Ej (Djs (z2 ))Qs ei
n j=1 i=1 i r
n
k
1 X X > > −1
> −1
= βx bnr (z1 )bns (z2 ) 2
ei Qr Lr (z1 )Qr ei e>
i Qs Ls (z2 )Qs ei + op (1)
n j=1 i=1
n
1 X
−1
> −1
tr Q>
= βx bnr (z1 )bns (z2 ) 2
r Lr (z1 )Qr Qs Ls (z2 )Qs + op (1)
n j=1
i.p.
−→ βx a(z1 , z2 ).
(4.20)
Collecting the results in (4.5), (4.18)-(4.20), we finally get the covariance function in the
Lemma and the proof of this Lemma is completed.
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
23
4.2. Proof of Theorem 3.1
First we show that it is it is enough to establish the following claim: under the (classical)
Marčenko-Pastur regime, i.e., n → ∞, p = pn → ∞ such that p/n → c > 0, it holds that
φq,n −
q d
→
− N (0, s(c)) ,
2
(4.21)
where recall that s(u) = q + u(ν4 − 1)(q 2 + 2q ). Here we use φq,n for φq to signify the
dependence in n. So assume this claim is true. Under the relaxed Marčenko-Pastur regime,
the sequence {pn /n} is bounded below and above. For any subsequence (pnk , nk )k of (pn , n),
we can extract a further subsequence (pnk` , nk` )` such that the ratios pnk` /nk` converge to
α > 0 when ` → ∞. On this subsequence, by Claim (4.21),
φq,nk` −
q d
→
− N (0, s(α)) ,
2
` → ∞.
By continuity of the function u → s(u), we have
q d
−1/2
→
− N (0, 1) ,
s(pnk` /nk` )
φq,nk` −
2
` → ∞.
As this limit is independent of the subsequence and it holds for all such subsequences, the
same limit holds for the whole sequence, that is,
q d
→
− N (0, 1) , ` → ∞.
s(pn /n)−1/2 φq,n −
2
The required asymptotic normality is thus established.
The remaining of the section is devoted to a proof of Claim (4.21) assuming p/n → c > 0.
Define the banded Toeplitz matrix
0 · · · 21 · · · 0
.. . .
..
1
.
. 0
.
2
1
.
1
.
Cn,τ =
. 0 2
0
2
.
.. 1
..
. ..
0
. 2
0 · · · 12 · · · 0 n×n
and
n
1
1 X
b
Nτ =
xt x∗t−τ + xt−τ x∗t = Xn Cn,τ X∗n .
2p t=1+τ
p
Define the associated Fourier series f (λ) of the banded Toeplitz matrix Cn,τ as
f (λ) = lim
n→∞
τ
X
k=−τ
tk eikλ =
1 iτ λ
e + e−iτ λ = cos(τ λ),
2
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
24
where tk is entry on the k-diagonal of Cn,τ .
According to the fundamental eigenvalue distribution theorem of Szegö for Toeplitz
forms, see Section 5.2 in Grenander and Szegö (1958) and Theorem 4.1 in Gray (2006), we
can infer that
Lemma 4.2. Suppose {lt , t = 1, · · · , n} are eigenvalues of Cn,τ with Fourier series f (λ),
then
(1) For any positive integer s,
n
1
1X s
lt =
lim
n→∞ n
2π
t=1
Z
0
2π
1
f (λ) dλ =
2π
s
Z
2π
(cos(τ λ))s dλ,
0
(2) For any continuous function on support of {lt , t = 1, · · · , n},
n
1
1X
F (lt ) =
lim
n→∞ n
2π
t=1
Z
2π
F (cos(τ λ)) dλ,
0
(3) Sequence {lt , t = 1, · · · , n} and
2πτ t
, t = 1, · · · , n
cos
n
are asymptotically equally distributed.
The limiting spectral distribution of Cn,τ is also derived in Lemma 3.1 of Bai and Wang
(2015), this useful lemma is stated as follows:
Lemma 4.3. As T → ∞, the ESD of Cn,τ tends to H, which is an Arcsine distribution
with density function
1
H 0 (t) = √
, t ∈ (−1, 1).
π 1 − t2
Recall for the permutation matrices D1 and Dτ = Dτ1 defined in Introduction, it holds
that
>
>
>
D1 D>
1 = D1 D1 = Dτ Dτ = Dτ Dτ = In .
Meanwhile, from the properties of Chebyshev polynomials, we can derive the following
lemma.
has eigenvalue
Lemma 4.4. (1) 12 D1 + D>
1
2πt
cos
, t = 1, · · · , n ,
n
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
(2)
1
2
25
Dτ + D>
has eigenvalue
τ
2πt
Tτ cos
, t = 1, · · · , n ,
n
where Tτ (·) stands for the Chebyshev polynomial of order τ .
shares the same asymptotic spectral distribution with Cn,τ as n → ∞.
(3) 12 Dτ + D>
τ
Since
n
X
eτ = 1
xt x∗t−τ + xt−τ x∗t
N
2p t=1
!
q
X
1
Dτ + D>
(x1 , · · · , xn )∗
(x1 , · · · , xn )
=
τ
2p
τ =1
!
q
X
1
= Xn
Dτ + D>
X∗n ,
τ
2p
τ =1
here for t ≤ 0, xt = xn+t , by Lemmas 4.2 and 4.4, it doesn’t take too much effort to see
e τ and N
b τ share the same limiting spectral distribution.
that N
e τ , by
Consider the Stieltjes transform mτ (z) of the limiting spectral distribution of N
implementing the Silverstein equation (2.2), we can infer that mτ (z) satisfies
Z
1
1
t
z=−
+
dHτ (t)
mτ (z) c
1 + tmτ (z)
!
1
1
1
−1 + − p
=
,
mτ (z)
c c 1 − m2τ (z)
where p/n → c > 0 as n → ∞, which coincides with the results in Bai and Wang (2015)
and Li et al. (2016).
Note that our test statistic
Lq =
q
X
τ =1
eτ =
L
q
X
τ =1
fτ M
f∗ ) =
Tr(M
τ
q
p 2 X
n
e τN
e ∗ ),
Tr(N
τ
τ =1
fτ = 1 Σ
bτ + Σ
b ∗ = 1 Pn xt x∗ + xt−τ x∗ , thus the asymptotic properties
where M
τ
t−τ
t
t=1
2
2n
f
e τ since M
fτ = p N
e .
of Mτ can be inferred from those of N
n τ
eτ
Actually in Li et al. (2016), the asymptotic behavior of the single-lag-τ test statistic L
has already been thoroughly explored and characterized. Theorem 2.1 in Li et al. (2016) is
stated as follows:
Lemma 4.5. Let τ ≥ 1 be a fixed integer, and assume that
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
26
1. {zit , i = 1, · · · , p, t = 1, · · · , n} are all independently distributed satisfying Ezit =
0, Ezit2 = 1, Ezit4 = ν4 < ∞;
2. (Marčenko-Pastur regime). The dimension p and the sample size n grow to infinity
in a related way such that cn := p/n → c > 0.
eτ
Then in the simplest setting when xt = zt , the limiting distribution of the test statistic L
is
3(ν4 − 1)
p d
1
ne
, 1+
c .
Lτ − →
− N
p
2
2
2
Pq e
Now consider the multi-lag-q test statistic Lq =
τ =1 Lτ , combining with Lemma
4.5,
all we need is the joint distribution of any two different single-lag test statistic, i.e.
er , L
es , 1 ≤ r 6= s ≤ q.
L
For a given integer q > 0, 1 ≤ r 6= s ≤ q, let fr (x) = fs (x) = x2 ,
∗
∗
1
1
Dr + D>
Ds + D>
r Xn Xn , Bns =
s Xn Xn ,
2p
2p
e r = 1 Xn Dr + D> X∗ , B ns = N
e s = 1 Xn Ds + D> X∗ ,
=N
r
n
s
n
2p
2p
Bnr =
Bnr
then both the LSDs of Bnr and Bns , i.e. F c,Hr , F c,Hs have Stieltjes transform mr (z) and
ms (z) satisfying the equation
!
1
1
1
.
z=
−1 + − p
m(z)
c c 1 − m2 (z)
Meanwhile,
e rN
e ∗) =
Tr(N
r
Z
e sN
e ∗) =
fr (x) dFnr (x), Tr(N
s
Z
fs (x) dFns (x).
where Fnr and Fns are the ESDs of Bnr and Bns . Thus, by directly implementing our joint
CLT for linear spectral statistics ofthe samplecovariance matrices,
i.e. Theorem 2.1, we
e1 , · · · , L
eq or any pair of L
er , L
es , 1 ≤ r 6= s ≤ q.
can derive the joint distribution of L
Precisely, the covariance function in Theorem 2.1 for the present case can be calculated to
be
3
1+ 2 c(β2 x +2) , if r = s,
c
(4.22)
Cov (Xfr , Xfs ) =
βx +2 ,
if
r
=
6
s.
c
The details of this lengthy derivation are postponed to Appendix B. Combining with
Lemma 4.5, for any given integer q ≤ 1, it can be inferred that, under the same assumptions
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
27
e1 , · · · , L
eq is
in Lemma 4.5, the joint limiting distribution of L
e1
c(ν4 − 1)
1 + 3(ν42−1) c · · ·
L
n . p
d
1
..
..
...
− N 1q ,
,
.. − 1q →
.
.
p
2
2
eq
c(ν4 − 1) · · · 1 + 3(ν42−1) c
L
where 1q = (1, · · · , 1)> is a q dimensional vector with q ones.
Recall that
q
q
p 2 X
X
e ∗ ),
e τN
e
Tr(N
Lq =
Lτ =
τ
n
τ =1
τ =1
then by the Delta method, we can derive the limiting distribution of our test statistic Lq ,
i.e.,
q
n
pq d
q
Lq −
→
− N
, q + c(ν4 − 1)(q 2 + ) .
p
2
2
2
Claim (4.21) is thus established.
5. Discussions
In this paper we have introduced, for the first time in the literature on eigenvalues of
large sample covariance matrices, a joint central limit theorem involving several population covariance matrices. This theorem is believed to provide wide applications to current
problems in high-dimensional statistics, especially for testing on structures of population
covariances. As a show-case, we treated the problem of testing for a high-dimensional
white noise in time series modelling. The derived new test shows very promising performance compared to existing competitors For future study, it would be worth investigating
other significant applications of this CLT.
Acknowledgments. We thank a Referee for the suggestion of the comparison to a
permutation test made in Section 3.1. Weiming Li’s research is partially supported by
National Natural Science Foundation of China, No. 11401037, MOE Project of Humanities
and Social Sciences, No. 17YJC790057, and Program of IRTSHUFE. Jianfeng Yao thanks
support from HKSAR GRF Grant 17305814.
References
Anderson, T. W. (2003). An Introduction to Multivariate Statistical Analysis. 3rd Edition.
John Wiley & Sons.
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
28
Bai, Z. D., (1999) Methodologies in spectral analysis of large dimensional random matrices,
a review. Statistica Sinica, 9, 611-677.
Bai, Z. D., Jiang, D. D., Yao, J. F. and Zheng, S. R. (2009). Corrections to LRT on large
dimensional covariance matrix by RMT. Annals of Statistics, 37, 3822-3840.
Bai, Z. D., Jiang, D. D., Yao, J. F., and Zheng, S. R. (2013). Testing linear hypotheses in
high-dimensional regressions. Statistics, 47, 1207-1223.
Bai, Z. D. and Silverstein, J. W. (1998). No eigenvalues outside the support of the limiting
spectral distribution of large-dimensional sample covariance matrices. Ann. Probab., 26,
316-345.
Bai, Z. D. and Silverstein, J. W. (2004). CLT for linear spectral statistics of large dimensional sample covariance matrices. Ann. Probab., 32, 553-605.
Bai, Z. D. and Silverstein, J. W. (2010). Spectral Analysis of Large Dimensional Random
Matrices. Science Press: Beijing.
Bai, Z. D. and Wang, C. (2015). A note on the limiting spectral distribution of a symmetrized auto-cross covariance matrix. Statistics & Probability Letters, 96, 333-340.
Bai, Z. D. and Zhou, W. (2008). Large sample covariance matrices without independence
structures in columns. Statistica Sinica, 28, 425-442.
Chen, B. B. and Pan, G. M. (2015) CLT for linear spectral statistics of normalized sample
covariance matrices with the dimension much larger than the sample size. Bernoulli, 21,
1089-1133.
Billingsley, P. (1995). Probability and Measure. John Wiley & Sons: New York
Burkholder, D. L. (1973). Distribution function inequalities for martingales. Ann. Probab.,
1, 19-42.
Gray, R. M. (2006). Toeplitz and Circulant Matrices: a Review. Now Publishers Inc.
Grenander, U. and Szegö, G.(1958). Toeplitz Forms and Their Applications. In: California
Monographs in Mathematical Sciences. University of California Press, Berkeley.
Johnstone, I. M. (2007). High dimensional statistical inference and random matrices. Int.
Cong. Mathematicians, Vol. I, 307-333. Zürich, Switzerland: European Mathematical
Society.
Li, Z., Yao, J., Lam, C., and Yao, Q. (2016). On testing a high-dimensional white noise.
Manuscript.
Marčenko, V. A. and Pastur, L. A. (1967). Distribution of eigenvalues for some sets of
random matrices. Math. USSR-Sb, 1, 457-483.
Pan, G. M. (2012). Comparison between two types of large sample covariance matrices.
Annales de l’Institut Henri Poincare-Probabiliteset Statistiques, 50, 655-677.
Paul, D. and Aue, A. (2014). Random matrix theory in statistics: A review, Journal of
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
29
Statistical Planning and Inference, 150, 1-29.
Silverstein, J. W. (1995). Strong convergence of the empirical distribution of eigenvalues
of large dimensional random matrices. J. Multivariate Anal., 5, 331-339.
Silverstein, J. W. and Bai, Z. D. (1995). On the empirical distribution of eigenvalues of a
class of large dimensional random matrices. J. Multivariate Anal., 54, 175-192.
Silverstein, J. W. and Choi, S. I. (1995). Analysis of the limiting spectral distribution of
large dimensional random matrices. J. Multivariate Anal., 54, 295-309.
Simes, R. J. (1986). An improved Bonferroni procedure for multiple tests of significance.
Biometrika, 73, 751-754.
Srivastava, M. S. (2005). Some tests concerning the covariance matrix in high dimensional
data. J. Japan Statist. Soc., 35, 251–272.
Yao, J. F., Bai, Z. D., and Zheng, S. R. (2015). Large Sample Covariance Matrices and
High-Dimensional Data Analysis, Cambridge University Press.
Zheng, S. R., Bai, Z. D., and Yao, J. F. (2015). Substitution principle for CLT of linear
spectral statistics of high-dimensional sample covariance matrices with applications to
hypothesis testing. Ann. Statist., 43, 546-591.
Zheng, S. R., Bai, Z. D., and Yao, J. F. (2017a). CLT for large dimensional general Fisher
matrices and its applications in high-dimensional data analysis. Bernoulli, 23, 1130-1178.
Zheng, S. R., Bai, Z. D., Yao, J. F., and Zhu, H. T. (2017b). CLT for linear spectral statistics
of large dimensional sample covariance matrices with dependent data. Manuscript.
Appendix A: Mathematical Tools
Lemma A.1. Let X = (X1 , . . . , Xn )> be a (complex) random vector with independent and
standardized entries having finite fourth moment and C = (cij ) be n × n (complex) matrix.
We have
!
n
X
E|X∗ CX − trC|4 ≤ K [tr(CC∗ )]2 +
E|Xii8 ||cii |4 .
i=1
The proof of the lemma follows easily by simple calculus and thus omitted.
Lemma A.2. (Lemma 2.6 of Silverstein and Bai (1995)). Let z ∈ C+ with v = = z, A
and B being n × n with B Hermitian, and r ∈ Cn . Then
kAk
r∗ (B − zI)−1 A(B − zI)−1 r
≤
.
tr (B − zI)−1 − (B + rr∗ − zI)−1 A =
∗
−1
1 + r (B − zI) r
v
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
30
Lemma A.3. [Formula 2.3 of Bai and Silverstein (2004)] For any nonrandom p × p
e ` , ` = 1, . . . q2 .
matrices Ck , k = 1, . . . , q1 and C
!
q2
q1
Y
Y
e ` r1r − n−1 trTnr C
e `)
E
(r∗ C
r∗ Ck r1r
1r
1r
`=1
q1
q2
Y
Y
−(1∧q2 ) (q2 −2)∨0
e ` ||,
Kn
δn
||Ck ||
||C
k=1
`=1
k=1
≤
q1 , q2 ≥ 0,
(A.1)
where K is a positive constant depending on q1 and q2 .
Appendix B: Derivation of the covariance (4.22)
Applying Theorem 2.1 to the functions fr (x) = fs (x) where 1 ≤ r 6= s ≤ q, the
corresponding covariance function is
I I
∂ 2 g(z1 , z2 )
1
dz1 dz2 ,
(B.1)
f
(z
)f
(z
)
Cov(Xfr , Xfs ) =
r 1 s 2
4π 2
∂z1 ∂z2
C1 C 2
where g(z1 , z2 ) = log(1 − a(z1 , z2 )) + log(1 − αx a(z2 , z2 )) − βx a(z1 , z2 ) with
ZZ
cmr (z1 )ms (z2 )t1 t2
a(z1 , z2 ) =
dHrs (t1 , t2 ).
(1 + t1 mr (z1 ))(1 + t2 ms (z2 ))
Mapping into our case, we have
1
p ↔ n, n ↔ p, c ↔ , αx = 1, βx = ν4 − 3, fr (x) = fs (x) = x2 ,
c
1
a(z1 , z2 ) =
c
1
=
c
ZZ
mr (z1 )ms (z2 )t1 t2
dHrs (t1 , t2 )
(1 + t1 mr (z1 ))(1 + t2 ms (z2 ))
1
m1 m2 Tr (t)Ts (t)
−1
(1 + Tr (t)m1 )(1 + Ts (t)m2 )
Z
dH(t),
where m1 , mr (z1 ), m2 , ms (z2 ), Tr (t), Ts (t) are Chebyshev polynomial of order r and
s respectively. Furthermore, we have
∂a (z1 , z2 )
1
=
∂z1
c
Z
∂a (z1 , z2 )
1
=
∂z2
c
Z
1
−1
1
−1
Ts (t) m2
Tr (t)
∂m1
·
dH(t),
2 ·
1 + Ts (t) m2 (1 + Tr (t) m1 )
∂z1
Tr (t) m1
Ts (t)
∂m2
·
dH(t),
2 ·
1 + Tr (t) m1 (1 + Ts (t) m2 )
∂z2
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
1
∂ 2 a (z1 , z2 )
=
∂z1 ∂z2
c
Z
1
Tr (t)
Ts (t)
∂m1 ∂m2
·
dH(t).
2 ·
2 ·
∂z1 ∂z2
(1 + Tr (t) m1 ) (1 + Ts (t) m2 )
−1
Since g(z1 , z2 ) = 2 log (1 − a(z1 , z2 )) − βx a(z1 , z2 ),
2
∂ a
∂a
· ∂a
∂ 2 log (1 − a(z1 , z2 ))
∂z1 ∂z2
∂z1 ∂z2
,
=−
−
∂z1 ∂z2
1 − a (1 − a)2
Thus
1
Cov (Xfr , Xfs ) = 2
4π
I I
C1
z12 z22
C2
∂ 2 g (z1 , z2 )
dz1 dz2
∂z1 ∂z2
I I
∂2a
∂a
· ∂a
1
1
2 2 ∂z1 ∂z2
=− 2
dz1 dz2 − 2
z1 z2
z12 z22 ∂z1 ∂z22 dz1 dz2
2π C1 C2
1−a
2π C1 C2
(1 − a)
I I
2
βx
∂ a
− 2
dz1 dz2 , M1 + M2 + M3 .
z12 z22
4π C1 C2
∂z1 ∂z2
I I
Consider M1 first, by Cauchy’s residue theorem, we have
I
1
z12
∂ 2a
·
dz1
2πi C1 1 − a ∂z1 ∂z2
I
Z
1
Ts (t)
∂m1 ∂m2
z12
1 1
Tr (t)
=
·
dz1
·
2 ·
2 dH(t) ·
2πi C1 1 − a c −1 (1 + Tr (t) m1 ) (1 + Ts (t) m2 )
∂z1 ∂z2
Z
I
1
∂m2 1 1
Tr (t)Ts (t)
z12
·
dm1 dH(t)
=−
·
∂z2 c −1 (1 + Ts (t) m2 )2 2πi C1 (1 − a) (1 + Tr (t) m1 )2
2
1
1
Z
I
−1 + c − √
c 1−m21
1
∂m2 1 1
Tr (t)Ts (t)
=−
·
·
dm
1 dH(t)
∂z2 c −1 (1 + Ts (t) m2 )2 2πi C1 m21 (1 − a) (1 + Tr (t) m1 )2
∂m2 1
=−
·
∂z2 c
T (t)T (t)
r
s
2
−1 (1 + Ts (t) m2 )
Z
1
2 (1)
1
1
−1 + c − c√1−m2
1
·
(1 − a) (1 + Tr (t) m1 )2
dH(t).
m1 =0
Note that
∂a
1
=
∂m1
c
Z
1
−1
Ts (t)m2
Tr (t)
·
dH(t),
1 + Ts (t) m2 (1 + Tr (t) m1 )2
31
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
32
then
2 (1)
1
1
−1 + c − c√1−m2
1
2
(1 − a) (1 + Tr (t) m1 )
1
= −2Tr (t) +
c
Z
1
−1
Ts (u)Tr (u)m2
dH(u),
1 + Ts (u) m2
m1 =0
therefore,
I I
∂ 2a
1
1
2 2
·
dz1 dz2
z z ·
M1 = − 2
2π C1 C2 1 2 1 − a ∂z1 ∂z2
I
Z 1
Z
2
Tr (t)Ts (t)
1 1 Ts (u)Tr (u)m2
2 1
=
z ·
−2Tr (t) +
dH(u) dH(t)dm2
2πi C2 2 c −1 (1 + Ts (t) m2 )2
c −1 1 + Ts (u) m2
2
I −1 + 1c − √ 1 2
Z
c 1−m2
1 1 −2Tr2 (t)Ts (t)
2
·
dH(t)
=
dm2
2
2
2πi C2
m2
c −1 (1 + Ts (t) m2 )
1
+
c
1
=
c
Z
1
2 I
−1 2πi C2
Z
1
−1
−1 + 1c − √ 1
2
Tr (t)Ts (t) Z 1
1
Ts (u)Tr (u)
·
dH(u)
dm
2 dH(t)
2
c
1
+
T
(u)
m
m2 (1 + Ts (t) m2 )
s
−1
2
c
1−m22
2 (1)
1
1
2 −1 + c − √
c 1−m22
2
−2Tr (t)Ts (t) ·
2
(1 + Ts (t) m2 )
dH(t)
m =0
2
Z 1
Z 1
2
1
Ts (u)Tr (u)dH(u) ·
Tr (t)Ts (t)dH(t)
+
c −1
c −1
Z 1
2
Z
8 1 2
2
2
=
T (t)Ts (t)dH(t) + 2
Ts (u)Tr (u)dH(u) .
c −1 r
c
−1
Similarly, for M2 , considering
I
1
1
∂a ∂a
z12 ·
·
dz1
2·
2πi C1
(1 − a) ∂z1 ∂z2
2
1
1
Z 1
I
−1 + c − √
c 1−m21
1
1
Ts (t) m2
Tr (t)
∂m1
=
·
·
·
dH(t)
2πi C1
c −1 1 + Ts (t) m2 (1 + Tr (t) m1 )2 ∂z1
m21 (1 − a)2
Z 1
1
Tr (t) m1
Ts (t)
∂m2
·
·
·
dH(t) dz1
c −1 1 + Tr (t) m1 (1 + Ts (t) m2 )2 ∂z2
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
∂m2 1
=−
·
∂z2 c
1
Z
1 I
−1 2πi C1
2
−1 + 1c − √ 1
c
33
1−m21
2
m1 (1 − a)
·
Tr (t)Ts (t) m2
(1 + Ts (t) m2 ) (1 + Tr (t) m1 )2
Z 1
1
Tr (u) Ts (u)
·
dH(u) dm1 dH(t)
c −1 (1 + Tr (u) m1 ) (1 + Ts (u) m2 )2
Z
Z
1 1 Tr (u) Ts (u)
∂m2 1 1 Tr (t)Ts (t) m2
·
dH(t) ·
dH(u).
=−
∂z2 c −1 1 + Ts (t) m2
c −1 (1 + Ts (u) m2 )2
Then,
I I
1
1
∂a ∂a
M2 = − 2
z12 z22 ·
·
dz1 dz2
2·
2π C1 C2
(1 − a) ∂z1 ∂z2
I
Z
Z
1 1 Tr (u) Ts (u)
2
∂m2 1 1 Tr (t)Ts (t) m2
2
dH(t) ·
dH(u) dz2
=
z −
·
2πi C2 2
∂z2 c −1 1 + Ts (t) m2
c −1 (1 + Ts (u) m2 )2
2
1
1
Z 1
Z
I Tr (t)Ts (t) −1 + c − √
c 1−m22
1 1
Tr (u) Ts (u)
1
2
=
dH(u)
dm
2 dH(t)
c −1 2πi C2
m2 (1 + Ts (t) m2 )
c −1 (1 + Ts (u) m2 )2
2
= 2
c
Z
2
1
Tr (t)Ts (t) dH(t)
.
−1
As for M3 ,
I
1
∂ 2a
z12
dz1
2πi C1 ∂z1 ∂z2
2
1
1
Z 1
I
−1 + c − √
c 1−m21
Tr (t)Ts (t)
∂m1 ∂m2
1
1
=
·
dH(t)
·
dz1
2
2
2πi C1
m21
c −1 (1 + Tr (t) m1 ) (1 + Ts (t) m2 )
∂z1 ∂z2
2
1
1
Z
−1 + c − √
1 I
c 1−m21
Tr (t)Ts (t)
∂m2 1 1
=−
·
·
dm
1 dH(t)
2
2
2
∂z2 c −1 (1 + Ts (t) m2 ) 2πi C1 m1 (1 + Tr (t) m1 )
∂m2 1
= −
·
∂z2 c
Z
1
−1
Tr (t)Ts (t)
(1 + Ts (t) m2 )2
1
c
√1
c 1−m21
−1 + −
·
(1 + Tr (t) m1 )2
2 (1)
dH(t)
m1 =0
=
∂m2 1
·
∂z2 c
Z
1
−1
2Tr2 (t)Ts (t)
(1 + Ts (t) m2 )2
dH(t),
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
thus
βx
M3 =
2πi
I
z22
C2
∂m2 1
·
∂z2 c
Z
1
1 I
−1 2πi C2
βx
= −
c
1
2Tr2 (t)Ts (t)
2 dH(t) dz2
−1 (1 + Ts (t) m2 )
2
1
1
−1 + c − √
c 1−m22
2Tr2 (t)Ts (t)
·
dm
2 dH(t)
2
m22
(1 + Ts (t) m2 )
Z
βx
= −
c
Z
Z
1
1
c
√1
c 1−m22
2 (1)
−1 + −
(1 + Ts (t) m2 )2
1
2Tr2 (t)Ts (t)
−1
dH(t)
m2 =0
=
4βx
c
Tr2 (t)Ts2 (t)dH(t).
−1
Combining the three terms gives
4(βx + 2)
Cov (Xfr , Xfs ) =
c
Z
1
Tr2 (t)Ts2 (t)dH(t)
−1
1
Note that for dH(t) = √
dt, we have
π 1 − t2
Z 1
Tr (t)Ts (t) dH(t) =
−1
Z 1
Tr2 (t)Ts2 (t) dH(t) =
−1
4
+ 2
c
Z
1
2
Ts (u)Tr (u)dH(u) .
−1
1
1{r=s} ,
2
3
1
1{r=s} + 1{r6=s} .
8
4
Therefore,
Cov (Xfr , Xfs ) =
The required formula is established.
1+ 23 c(βx +2)
,
c2
if r = s,
βx +2
,
c
if r 6= s.
34
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
Table 1
Test Size under Scenario (I) and (II)
p
5
10
25
40
10
20
50
80
50
100
250
400
10
40
60
100
50
200
300
500
90
360
540
900
150
600
900
1500
200
800
1200
2000
500
2000
3000
5000
n
1000
2000
5000
8000
1000
2000
5000
8000
1000
2000
5000
8000
100
400
600
1000
100
400
600
1000
100
400
600
1000
100
400
600
1000
100
400
600
1000
100
400
600
1000
cn
0.005
0.005
0.005
0.005
0.01
0.01
0.01
0.01
0.05
0.05
0.05
0.05
0.1
0.1
0.1
0.1
0.5
0.5
0.5
0.5
0.9
0.9
0.9
0.9
1.5
1.5
1.5
1.5
2
2
2
2
5
5
5
5
φq (I)
q=1 q=3
0.081 0.078
0.059 0.060
0.051 0.054
0.050 0.049
0.072 0.067
0.066 0.059
0.046 0.053
0.056 0.049
0.063 0.056
0.058 0.052
0.056 0.055
0.049 0.047
0.073 0.075
0.053 0.062
0.049 0.043
0.062 0.058
0.066 0.067
0.053 0.052
0.053 0.054
0.052 0.050
0.051 0.055
0.056 0.051
0.058 0.050
0.048 0.053
0.042 0.047
0.048 0.054
0.056 0.055
0.051 0.052
0.051 0.051
0.047 0.056
0.055 0.050
0.054 0.053
0.063 0.053
0.049 0.054
0.052 0.056
0.052 0.052
Uq (I)
q=1 q=3
0.065 0.049
0.052 0.050
0.047 0.043
0.054 0.036
0.057 0.048
0.050 0.040
0.045 0.044
0.051 0.046
0.051 0.048
0.061 0.047
0.050 0.047
0.047 0.040
0.062 0.061
0.050 0.043
0.055 0.047
0.053 0.045
0.066 0.050
0.051 0.045
0.044 0.046
0.045 0.051
0.050 0.057
0.050 0.050
0.061 0.049
0.058 0.046
0.065 0.064
0.049 0.040
0.040 0.046
0.041 0.043
0.057 0.049
0.052 0.049
0.051 0.052
0.050 0.051
0.049 0.044
0.042 0.044
0.052 0.042
0.047 0.049
φq (II)
q=1 q=3
0.081 0.074
0.058 0.055
0.054 0.059
0.062 0.055
0.063 0.060
0.057 0.058
0.048 0.053
0.046 0.046
0.046 0.058
0.048 0.052
0.046 0.048
0.055 0.042
0.072 0.074
0.051 0.059
0.052 0.053
0.051 0.054
0.051 0.059
0.048 0.045
0.040 0.048
0.041 0.045
0.048 0.051
0.047 0.048
0.037 0.046
0.045 0.045
0.039 0.048
0.036 0.045
0.039 0.043
0.042 0.049
0.045 0.045
0.043 0.052
0.043 0.052
0.048 0.048
0.051 0.052
0.040 0.048
0.034 0.044
0.043 0.045
Uq (II)
q=1 q=3
0.090 0.081
0.068 0.067
0.055 0.050
0.057 0.051
0.068 0.052
0.065 0.049
0.055 0.044
0.052 0.050
0.055 0.058
0.049 0.047
0.047 0.037
0.046 0.044
0.079 0.083
0.056 0.055
0.053 0.049
0.054 0.044
0.069 0.070
0.053 0.059
0.043 0.038
0.048 0.038
0.069 0.078
0.047 0.047
0.054 0.057
0.048 0.045
0.059 0.070
0.048 0.049
0.048 0.046
0.049 0.047
0.067 0.066
0.046 0.048
0.053 0.040
0.049 0.040
0.051 0.068
0.055 0.048
0.049 0.051
0.049 0.057
35
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
Table 2
Test Power under Scenario (III) and (IV)
p
5
10
25
40
10
20
50
80
50
100
250
400
10
40
60
100
50
200
300
500
90
360
540
900
150
600
900
1500
200
800
1200
2000
500
2000
3000
5000
n
1000
2000
5000
8000
1000
2000
5000
8000
1000
2000
5000
8000
100
400
600
1000
100
400
600
1000
100
400
600
1000
100
400
600
1000
100
400
600
1000
100
400
600
1000
cn
0.005
0.005
0.005
0.005
0.01
0.01
0.01
0.01
0.05
0.05
0.05
0.05
0.1
0.1
0.1
0.1
0.5
0.5
0.5
0.5
0.9
0.9
0.9
0.9
1.5
1.5
1.5
1.5
2
2
2
2
5
5
5
5
φq (III)
q=1
q=3
0.999
0.990
1
1
1
1
1
1
1
0.998
1
1
1
1
1
1
1
0.9995
1
1
1
1
1
1
0.335
0.213
0.983
0.764
1
0.973
1
1
0.416
0.245
1
0.957
1
1
1
1
0.506
0.311
1
0.991
1
1
1
1
0.607
0.365
1
1
1
1
1
1
0.694
0.428
1
1
1
1
1
1
0.9375 0.704
1
1
1
1
1
1
Uq (III)
q=1 q=3
0.807 0.635
0.9995 0.986
1
1
1
1
0.824 0.647
1
0.992
1
1
1
1
0.859 0.681
1
0.993
1
1
1
1
0.085 0.077
0.294 0.207
0.482 0.338
0.851 0.651
0.093 0.093
0.287 0.190
0.497 0.327
0.878 0.669
0.090 0.079
0.323 0.208
0.527 0.338
0.897 0.655
0.089 0.078
0.337 0.203
0.573 0.346
0.922 0.674
0.092 0.082
0.351 0.207
0.609 0.351
0.923 0.657
0.102 0.082
0.452 0.190
0.734 0.353
0.986 0.659
φq (IV)
q=1
q=3
0.999
0.986
1
1
1
1
1
1
1
0.997
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0.3055 0.1955
0.968 0.6935
1
0.94
1
1
0.3075 0.1805
0.9995 0.8585
1
0.995
1
1
0.3785 0.217
1
0.952
1
0.9995
1
1
0.4545 0.2745
1
0.992
1
1
1
1
0.5165 0.2985
1
0.9985
1
1
1
1
0.809
0.519
1
1
1
1
1
1
Uq (IV)
q=1 q=3
0.761 0.633
0.999 0.984
1
1
1
1
0.79 0.636
1
0.989
1
1
1
1
0.835 0.660
1
0.996
1
1
1
1
0.129 0.128
0.283 0.225
0.483 0.331
0.852 0.652
0.094 0.091
0.310 0.223
0.491 0.322
0.886 0.646
0.096 0.097
0.315 0.214
0.504 0.331
0.894 0.671
0.102 0.098
0.319 0.212
0.576 0.326
0.906 0.656
0.108 0.098
0.352 0.204
0.592 0.342
0.933 0.656
0.108 0.097
0.430 0.202
0.712 0.329
0.984 0.679
36
Weiming Li, Zeng Li and Jianfeng Yao/Joint CLT for linear spectral statistics
37
Table 3
Size and power comparison of permutation test and φq for Gaussian distribution under Scenario (III),
nominal level α = 5%, testing size is shown in the first row of each (p, n) configuration block.
Permutation
φq
p
n
cn
a
q=1
q=3
q=1
q=3
150
300
0.5
0
0.056
0.062
0.052
0.054
150
300
0.5
0.05
0.360
0.256
0.254
0.146
150
300
0.5
0.09
0.992
0.948
0.934
0.694
150
300
0.5
0.1
1
0.988
0.976
0.820
270
300
0.9
0
0.068
0.074
0.060
0.050
270
300
0.9
0.05
0.510
0.408
0.376
0.216
270
300
0.9
0.09
1
1
0.990
0.820
270
300
0.9
0.1
1
1
0.998
0.914
600
300
2
0
0.056
0.066
0.062
0.062
600
300
2
0.05
0.808
0.708
0.536
0.294
600
300
2
0.09
1
1
1
0.974
600
300
2
0.1
1
1
1
0.994
Table 4
Size and power comparison of permutation test and φq for Gamma distribution under Scenario (IV),
nominal level α = 5%, testing size is shown in the first row of each (p, n) configuration block.
Permutation
φq
p
n
cn
a
q=1
q=3
q=1
q=3
150
300
0.5
0
0.052
0.044
0.050
0.058
150
300
0.5
0.05
0.344
0.276
0.228
0.142
150
300
0.5
0.09
0.996
0.962
0.892
0.534
150
300
0.5
0.1
1
0.988
0.956
0.692
270
300
0.9
0
0.058
0.048
0.046
0.052
270
300
0.9
0.05
0.470
0.372
0.244
0.144
270
300
0.9
0.09
1
1
0.960
0.662
270
300
0.9
0.1
1
1
0.988
0.802
600
300
2
0
0.058
0.054
0.042
0.056
600
300
2
0.05
0.766
0.692
0.366
0.202
600
300
2
0.09
1
1
0.998
0.884
600
300
2
0.1
1
1
1
0.944
| 10math.ST
|
A Multi-Layer Market for Vehicle-to-Grid Energy Trading in the Smart Grid
Albert Y.S. Lam
∗1
, Longbo Huang
1
†1
arXiv:1609.01437v1 [cs.GT] 6 Sep 2016
‡1
, and Walid Saad
§2
University of California, Berkeley
2
University of Miami
Abstract
cies [2, 3, 4, 5, 6] and is expected to lie at the heart of the
emerging smart grid system.
Enabling V2G interactions has several advantages such as
serving as backup power sources during outages or supplying ancillary services back to the grid for regulation services.
However, in order to fully reap the benefits of V2G systems,
several key challenges must be addressed at different level
such as control, communications, charging behavior, implementation, and market mechanisms. In [7], the authors propose a scheme that uses PHEV batteries to absorb the randomness in intermittent wind power generation. The authors
in [8] study the use of game theory for providing frequency
regulation through V2G operation. Using the PHEVs as
storage units is studied and analyzed in [9] while communication architectures suitable for V2G systems are discussed
in [10]. Further, the authors in [11] considers the problem
of optimally providing energy and ancillary services using
electric vehicles.
Clearly, most existing work are focused on implementation, communication, and energy transfer in V2G systems.
However, the need for energy transfer and exchange from
PHEVs to the grid has also an economical aspect that must
be addressed. In this respect, the work in [12] sheds a light
on this aspect by investigating the price and quantities exchanged if the PHEVs and the grid elements form an energy
trade market. Beyond [12], little work seems to have been
focused on the economics of V2G exchanges which are essential for a better understanding on the potential of using V2G
in future power grid systems.
The main contribution of this paper is to propose a general framework and algorithm for studying the economics of
the market emerging between PHEVs, aggregators, and the
smart grid elements. To address this problem, we propose
a multi-layered market mechanisms in which the agreggators, the PHEVs, and the grid elements can decide on the
quantity and prices at which they wish to trade energy while
optimizing the tradeoff between the benefits (e.g., revenues)
and costs from this energy exchange. In this proposed market, first, the aggregators and the smart grid elements (e.g.,
substations) submit their reservation prices and bids so as to
agree on a price and energy trading mechanism. These interactions are modeled using a double auction whose result
is then fed back into the second layer, which deals with the
management of PHEV resources at each aggregator. In this
In this work, we propose a multi-layer market for vehicleto-grid energy trading. In the macro layer, we consider a
double auction mechanism, under which the utility company
act as an auctioneer and energy buyers and sellers interact.
This double auction mechanism is strategy-proof and converges asymptotically. In the micro layer, the aggregators,
which are the sellers in the macro layer, are paid with commissions to sell the energy of plug-in hybrid electric vehicles
(PHEVs) and to maximize their utilities. We analyze the
interaction between the macro and micro layers and study
some simplified cases. Depending on the elasticity of supply
and demand, the utility is analyzed under different scenarios.
Simulation results show that our approach can significantly
increase the utility of PHEVs.
1
, Alonso Silva
Introduction
The rising oil prices combined with the ongoing trend for developing environmental-friendly technological solutions, implies that electrically-operated vehicles will lie at the heart of
future transportation systems. In particular, it is envisioned
that plug-in hybrid electric vehicles (PHEVs), which are essentially electric vehicles equipped with storage devices, will
constitute one key component towards realizing the vision
of green, environment-friendly transportation networks. For
instance, it is forecast that up to 2.7 million electric vehicles
will be put on the road in the United States, by 2020 [1].
The presence of energy storage devices implies that PHEVs
can not only serve as a green means of transportation, but
also, if properly configured, they can function as a “moving”
energy reservoir that can store and, possibly, supply power
back to the power grid. While current PHEV deployment
are mostly concerned with grid-to-vehicle interactions, enabling two-way vehicle-to-grid (V2G) interactions between
the grid and PHEVs, has recently started to receive considerable attention both in research and standardization agen-
∗ Email: [email protected] To whom correspondence should
be addressed.
† Email: [email protected]
‡ Email: [email protected] To whom correspondence should
be addressed.
§ Email: [email protected]
1
PHEVs denoted by In = {1, . . . , In }. In this model, we are
particularly interested in smart grid elements that are unable
to meet their demand and, hence, need to buy energy from
(Xk , Bk ) 1
alternative sources such as the PHEV aggregators. Hence,
hereinafter, all grid elements are referred to as buyers while
Macro layer
Auctioneer
the aggregators are referred to as sellers.
The energy exchange process between the aggregators and
(An , Sn ) 4 5 (P, Qn )
the grid elements is modeled using a two-layered market
8
model as shown in Fig. 1 with the utility company’s control
Aggregator n
center acting as a middleman that handles the prospective
energy trading mechanisms. On the one hand, at the first
pn 6 7 a i
layer, referred to as the macro layer, the aggregators and the
Micro layer
grid elements interact so as to trade energy. At this layer,
PHEV i
the buyers wish to optimize their performance and meet their
demand by buying energy from the PHEVs through the aggregators while the aggregators wish to strategically choose
Figure 1: The two-layer market model. The numbers in the
their price and quantity to trade so as to optimize their revfigure correspond to the step numbers in Algorithm 1.
enues. On the other hand, at the second layer, referred to
as the micro layer, the aggregators must interact with the
layer, within each PHEV group, the aggregator negotiates PHEVs so as to optimize the energy resources and provide
with the PHEVs to settle for an agreement on the resource incentives for the PHEVs to actually participate in the trade.
usage. In particular, the aggregator will announce its energy Clearly, the outcomes of these two layers are coupled and,
buying price to the PHEVs, and each PHEV determines how thus, any market solution must take into account this intermuch energy it is willing to supply to the aggregator. The layer dependence. Below, we discuss and analyze, in details,
outcome of this layer is directly linked to the previous market the market operation at each layer.
due to the fact that the aggregator has to carefully balance
its earnings from the energy market and its payments to the 2.1 Macro Layer
PHEVs within its group. Hence, unlike existing work such
as in [12], the proposed scheme depends, not only on pricing An auction is a mechanism for buying or selling goods or
issues, but also on modeling the PHEVs-to-aggregator inter- services by taking bids and then selling them to the highactions (from an economical perspective) as well as on pro- est bidders. The main approach to study this mechanism
viding incentives for the PHEVs to participate in the foreseen is game theory by considering the bidders as the players of
market and, hence, it can leverage V2G, for improving the this game, and their bids as their strategies. For more deoverall smart grid performance. We characterize the equilib- tails on auction theory see [13]. For the completeness of
ria resulting from the proposed multi-layered market and we this work, here we give an overview of the double auction
show their existence. Further, using linear approximation mechanism [14, 15, 12]. In this layer, each potential seller
techniques we provide a large-system analysis on the eco- or aggregator n ∈ N sends the quantity of energy An that
nomics of V2G energy trading. Using simulations, we assess it intends to supply and its reservation price Sn to the aucthe properties and performance of the energy trading mecha- tioneer. The reservation price sent by the potential sellers
nisms resulting from proposed scheme and we show that our corresponds to the minimum price at which the seller is willing to sell its offered amount of energy. Each buyer k ∈ K
approach can significantly increase the utility of PHEVs.
The paper is organized as follows: In Section 2, we present proposes a bid Bk and the quantity it requests, denoted by
the proposed system model. In Section 3, we analyze the Xk , to the auctioneer. Here, we are mainly focused on the
system for different PHEV energy supply costs. Simulation interactions between buyers and sellers in a certain window
results are discussed and analyzed in Section 4 while conclu- of time during which the bids and reservation prices do not
vary. This can correspond to an energy trading market in
sions are drawn in Section 5.
which decisions are based on medium or long-term energy
needs such as in a day-ahead market. In each round, each
2 The Model and the Market Mech- aggregator n decides its own An with the fixed Sn . After
receiving all An ’s, the auctioneer determines the price P (A)
anism
of the energy, where A = (An , 1 ≤ n ≤ N ), and Qn (A),
which corresponds to the total quantity sold by aggregator
Consider a smart grid system consisting of K grid elements n, ∀n ∈ N , by a double auction.
(e.g., substations) with K = {1, . . . , K} denoting the set of
In this work, we adopt a double auction mechanism based
all such elements. In this grid, N electric vehicles aggregators on [14, 12, 16, 15]. In this auction:
are deployed. We let N = {1, . . . , N } denote the set of all
aggregators. Each aggregator n ∈ N manages a group of
• we order the sellers in an increasing order of their reserBuyer k
2
2.2
vation price. W.l.o.g. we consider
S1 < S2 < . . . < SN .
Micro Layer
In this layer, each aggregator announces a price pn to its
PHEVs by
(1)
pn = γn P (A), ∀n ∈ N ,
• we order the buyers in a decreasing order of their reservation bids. W.l.o.g. we consider
•
•
•
•
(5)
where 0 < γn < 1 is the commission rate. The price differB1 > B2 > . . . > BK .
(2) ence, P (A)−pn , is the commission (i.e., cost of management)
earned by aggregator n from its managed PHEVs. Each
aggregator can make a profit in this way and thus it has
if two sellers (respectively, buyers) have equal reservaan incentive to help its PHEVS participate in the market.
tion prices (bids), they are aggregated into one single
This provides an incentive for an aggregator to maximize
“virtual” seller (or buyer).
its PHEVs’ profits. By utility maximization, each PHEV
max
we generate the supply curve (selling reservation price i ∈ In determines its available supply ai ∈ [0, ai ] according
max
to its utility where ai
characterizes the amount of energy
Sn versus the amount of energy put out for sale An )
the owner of PHEV i can afford to sell, i.e., after reserving
we generate the demand curve (offered bids Bk versus enough for its own use. For the same type of vehicle, a PHEV
which needs more reservation for its proper operation has a
quantity needed Xk ).
lower amax
. We have
i
we find an intersection point.
This intersection is at the level of a certain seller L and
buyer M , such that BM ≥ SL and BM +1 < SL+1 . We
find seller L and buyer M and the double auction dictates
that the first L − 1 and M − 1 buyers will participate in the
energy trading. We don’t consider seller L and buyer M in
this trade since it’s a necessary condition for matching the
total supply and demand while maintaining a strategy-proof
mechanism [15].
Thus all sellers with index n < L and all buyers with
index k < M become the participants in the double auction
so as to clear the market. In consequence, the trading prices
for the sellers and the buyers can be chosen within any point
in the range [SL , BM ] [16]. In this work, for any strategy
choice A (or An , ∀n ∈ N ) by the sellers (i.e. aggregators),
given seller L and buyer M at the intersection, we consider
all sellers i < L and buyers k < M trade at a price P (A),
given by
SL (A) + BM (A)
P (A) =
,
(3)
2
where the dependence on A is due to the fact that each A
can give different interaction points with the demand and
supply curves, and thus we can have different L and M .
At the end of the auction, numerous criteria can be used
for determining the amount of traded energy between each
one of the L − 1 sellers and M − 1 buyers. We adopt the
approach of [15] in which the volume is divided in such a
way as to ensure a strategy-proof auction. Using the method
in [15] the total quantity Qn (A) sold by any PHEV group n,
for a given choice A is:
PM −1
Pn
if
An
k=1 Xk ≥
j=1 Aj ,
Qn (A) =
(4)
(An − Ψ)+ if n = L − 1,
0
if n > L − 1,
An =
In
X
i=1
ai , ∀n ∈ N .
(6)
Let an = (ai , 1 ≤ i ≤ In ) be the strategy of the PHEVs
in group i. Aggregator n determines the actual quantity
qi (an , Qn (A)) allocated to PHEV i in group n proportional
to ai by
ai
× Qn (A),
(7)
qi (an , Qn (A)), ∀n ∈ N .
(8)
qi (an , Qn (A)) = P
i∈In
ai
and thus
Qn (A) =
In
X
i=1
Consider that each PHEV i ∈ In imposes a cost of discharging its battery to supply energy to its aggregator. We
denote this cost by ci (a), where a is the amount of energy
sold by PHEV i. We assume that during each transaction,
each PHEV will try to maximize its profit by choosing the
amount of energy ai to be
ai = arg max{pn (an ) − ci (ai )}.
ai
2.3
Market Mechanism
The interaction between the macro and micro layers can be
seen as a market mechanism or an algorithm, shown in Algorithm 1, which is used to find the market equilibrium P ∗ , A∗n .
It terminates when a pre-determined number of iterations
tmax has been reached or the percentage change of the market price is less than a certain threshold ξ.
As we can see in simulation later, Algorithm 1 performs
very efficiently and converges within a few steps.
where (x)+ = max{0, x} and Ψ is the oversupply, i.e., Ψ =
PL−1
PM −1
j=1 Aj −
k=1 Xk .
3
price(p)
price(p)
Algorithm 1 Market Mechanism
1: For all k, buyer k submits its bid Bk and its requested
qd = q0 − βp
demand
amount Xk to the auctioneer
supply
qs = αp
2: t ← 1
3: repeat
4:
For all n, aggregator n submits its reservation price Sn
and its proposed supply An to the auctioneer
5:
The auctioneer implements a double auction and determines the market price P (t) and the allocated quantity
quantity(q)
quantity(q)
Qn for aggregator n for all n
6:
For each aggregator n, the selling price pn is announced
(a) Real situation
(b) Linear approximation
to its PHEVs
7:
For each PHEV i, it determines its proposed selling
Figure 2: Supply and demand curves
amount ai by maximizing its utility according to pn
and returns ai back to its aggregator
We note that, a PHEV can always decide not to participate in
8:
Each aggregator n sums up all ai ’s from its PHEVs
the market in which case it maximum utility will be u∗i = 0.
9:
t←t+1
P (t) −P (t−1)
max
As a result, the maximum utility in (9) is always nonnegative.
|<ξ
10: until t > t
or |
P (t)
In this case, the solution a∗i of (9) is
max
ai
if ηi ≥ pn ,
∗
ai =
(10)
3 Linear Approximation for Large
0
else.
systems
We can then subdivide the set of PHEVs In in two dis(2)
(1)
(2)
(1)
joint sets In and In , such that In ∪ In = In , where
(2)
(1)
, ∀i ∈ In . Thus, we obtain
a∗i = 0, ∀i ∈ In , and a∗i = amax
i
X
X
X
An =
ai =
ai +
ai .
(11)
In a practical smart grid, both the number of PHEVs and
smart grid element are expected to be large. In this respect,
it is of interest to analyze the impact of the presence of large
numbers of buyers and aggregators on the overall proposed
market mechanism. Therefore, here, we consider the situations when N and M are sufficiently large such that the
corresponding supply and demand curves can be approximated by a linear function, depending on the distribution
of aggregators’ reservation price (buyers’ bids). We assume
that the reservation prices Sn from each one can be ordered
nicely into a line with each point being one unit apart. Fig. 2
shows one example of the supply and demand curves, each
of which is approximated by one single linear function. In
the following, since the focus of this paper is on the supply
side, we study different types of supply curves while always
keeping the demand curve of the same type.
i∈In
(1)
i∈In
P
(1)
Since for all i ∈ In , a∗i = 0, then i∈I (1) ai = 0, and in
n
consequence
X
X
An =
ai =
amax
.
(12)
i
(2)
i∈In
3.1.1
(2)
i∈In
Linear Approximation
Consider the linear case as in Fig. 2(b). The PHEVs are
price takers and the supply and demand curves are given by
Supply(P, Q)
3.1
(2)
i∈In
= αP,
(13)
Demand(P, Q) = Q0 − βP.
(14)
Linear Cost Function with Homogeneous PHEVs
Here Q0 is the total demand from all the buyers and P is
the price determined from the double auction in the macro
(2)
layer. Also, α will depend on In for all aggregator n. In
that case, the equilibrium price P ∗ and the quantity sold Q∗
can be determined when the supply meets the demand, i.e.,
In this subsection, we consider that the cost function of
PHEV i is a linear function with respect to the amount of
energy it provides, i.e., ci (a) = ηi a, within a certain interval [0, amax
]. If the amount of energy to be sold exceeds amax
,
i
i
the cost will become infinite. Note that this case also takes
into account the inconvenience cost (e.g., when the PHEV
does not have enough energy to operate normally). Then its
utility function is given by ui (a) = (γn P −ηi )a. The problem
PHEV i needs to address is the following:
u∗i = maximize
ui (ai ).
max
ai ∈[0,ai
]
Supply(P ∗ , Q∗ ) = Demand(P ∗ , Q∗ ) ⇒ αP ∗ = Q0 − βP ∗ ,
or equivalently, when
P∗ =
(9)
Q0
.
α+β
(15)
The utility for each PHEV i in group n will be thus given by
4
1. If ηi ≥
γn Q 0
(α+β)
then a∗i = 0 and u∗i = 0.
2. If ηi <
γn Q 0
(α+β)
and
then a∗i = amax
i
u∗i = (γn P ∗ − ηi ) a∗i =
γn Q0
− ηi amax
.
i
(α + β)
The total utility Un of aggregator n is given by
X γn Q0
Un =
− ηi amax
.
i
(α
+
β)
(2)
Based on the above observation, we have the following simple
lemma characterizing the necessary conditions under which
there exists market equilibriums, i.e., (19) and (21) both
hold.
(16) Lemma 1. The following conditions are necessary for the
above iteration process converges to an market equilibrium
(P ∗ , A∗n ):
(β + dn )2 − 4(dn β − Cn Q0 ) ≥
(17)
2
(β − dn ) + 4Cn Q0
i∈In
≥
0, ∀n ∈ N
0, ∀n ∈ N . 3
(22)
(23)
Proof. In equilibrium, this price must result in a supply that
From this simple example, we can deduce the following is exactly equal to the resulting An , i.e.,
properties:
Cn Q0 γn
− dn .
(24)
An =
An + β
1. If the supply slope of a market 1 is higher than that
of a market 2, i.e., α1 > α2 , then the utility gained in
This gives rise to the following condition on An :
market 1 is smaller than that in market 2.
A2n + (β + dn )An + βdn − Cn γn Q0 = 0.
(25)
2. If the demand slope of a market 1 is higher than that
of a market 2, i.e., β1 > β2 , then the utility gained in Note that the above argument also shows an interesting fact
market 1 is smaller than that in market 2.
that there is an implicit iteration for the market price P as
follows:
3. If the total possible demand of a market 1 is higher than
1
2
Q0
that of a market 2, i.e., Q0 > Q0 , then the utility gained
P∗ =
.
(26)
in market 1 is greater than that in market 2.
C n γn P ∗ − dn + β
3.2
This similarly implies that the fixed point should be:
Quadratic Cost Function
Cn γn P ∗2 + (β − dn )P ∗ − Q0 = 0.
We first recall that under the quadratic cost model, each
PHEV i will supply the following amount of energy:
a∗i =
pn (an ) − ηi
2υi
Since both (25) and (27) are quadratic functions, we see that
the only way there exists a market equilibrium is when (22)
and (23) hold.
amax
i
.
(18)
0
In the next section, we will show how the algorithm evolves
through simulation and demonstrate the intuition behind the
iteration process.
In order to make the analysis easier, below we assume that
a∗i ≤ amax
for all i. Then, we have:
i
a∗i =
where
pn (an ) − ηi
,
2υi
and A∗n = Cn pn (an ) − dn ,
X 1
Cn =
2υi
i
X ηi
and dn =
.
2υi
i
4
(19)
Q0
.
A∗n + β
Simulation Results
For simulations, we consider a smart grid network in which
a number of aggregators sell their energy surplus to smart
grid elements (buyers) through a utility company. The simulation setting is as follows. Each aggregator manages a
certain number of PHEVs, randomly generated in the range
of [500, 1000]. Each PHEV has a maximum battery capacity
of 250 miles with power consumption 22kWh per 100 miles
[17, 18] out of which an arbitrary amount of energy, between
30 and 100 miles, is reserved for the PHEV’s private use. The
reservation prices of the aggregators are uniformly selected
in [10, 50] dollars/MWh while the buyers’ bids are randomly
chosen from [15, 60] dollars/MWh. Each buyer requests energy demand with the amount chosen in [20, 60] MWh. The
commission rate is set to γn = 0.91, ∀n ∈ N . Each PHEV
i has random cost function parameters ηi ∈ [10, 50] and
υi ∈ [1000, 2000] for quadratic cost function and βi = 0 for
the linear cost function. The algorithm always starts by setting ai = amax
, ∀i ∈ In , ∀n ∈ N .
i
(20)
In this case, A∗n is the supply from one aggregator and A∗n is
the slope of the supply curve, i.e., α = A∗n . Thus, we have
by (15) that:
P∗ =
(27)
(21)
Hence, we see that there is an iteration process of the market
price, which approximates the actual price iteration:
• In iteration t, with An (t) computed, we obtain the op0
timal price P ∗ (t) = AnQ
(t)+β .
• In iteration t + 1, we compute the new supply quantity
by (19), i.e., An (t + 1) = Cn γn P ∗ (t) − dn .
5
Greedy
approach
Price
50
5
40
50
30
20
0
4
6
-‐50
8
10
15
20
4
Price
(P)
Average
u.lity
per
aggregator
Proposed
2-‐layer
approach
100
2
3
Demand
1
10
0
Number
of
aggregators
(N)
(a) Total utility in each iteration
(b) Supply and demand curves
(a) Total utility and price for linear cost
Greedy
approach/100
4
48.5
2
48
47.5
0
-‐2
4
6
8
10
15
20
47
46.5
-‐4
Figure 4: Linear cost function with 1000 buyers (K) and 1000
aggregators (N)
Price
when N ≥ 6, an increase in the number of aggregators N will
yield a decrease in the settled price which leads to a decrease
in the utility. In the cases with quadratic cost, the average
utility increases with N on the grounds that more aggregators participate in the market resulting in a larger amount
of total energy sold. Hence, in general, the equilibrium trading price decreases with N as more aggregators lead to an
increased competition, which subsequently imposes a lower
market price.
Fig. 3(c) shows the average number of iterations required
to reach an equilibrium. Clearly, as more aggregators participate in the market, the number of iterations till convergence
increases. Moreover, Fig. 3(c) shows that the convergence
time is faster in the case with quadratic cost. With a linear
cost function, each PHEV i takes either 0 or amax
in each
i
iteration, and thus, the algorithm is more like to oscillate
more around the equilibrium point before convergence. With
a quadratic cost function, due to the concavity of the utility,
the algorithm moves toward the equilibrium in a smoother
manner which is further corroborated in the subsequent simulations.
Price
(P)
Average
u.lity
per
aggregator
Proposed
2-‐layer
approach
46
-‐6
Number
of
aggregators
(N)
45.5
(b) Total utility and price for quadratic cost
Number
of
itera-ons
Linear
cost
Quadra6c
cost
15
10
5
0
4
6
8
10
15
Number
of
aggregators
(N)
20
(c) Iterations required for convergence
Figure 3: Basic simulations for small numbers of buyers and
aggregators.
4.1
4.2
Large Numbers of Buyers and Aggregators
Small Numbers of Buyers and AggregaHere, we simulate cases in which a large number of buyers
tors
(K = 1000) and aggregators (N = 1000) are deployed so
as to verify the analytical results induced from the linear
approximation studied in Section 3.1. As previously mentioned, when N (K) increases, the supply (demand) curve
approaches a linear function as all random numbers are generated uniformly. We first study the results with a fixed
demand curve (i.e. with β and Q0 fixed) and they correspond to the situations when the value evolves in a particular simulation. Fig. 41 shows the results for linear PHEV
cost functions when the algorithm iterates. Fig. 4(a) gives
the total utility computed from double auction for each iteration while we have the corresponding supply and demand
curves in Fig. 4(b) ( the part shows intersection points only).
We can see that the total utility decreases with the supply
We simulate the cases with small numbers of buyers and aggregators. We consider 5 buyers (K = 5) in each case and we
compare our two-layer approach with a greedy approach, in
which each PHEV always proposes to sell amax
. Fig. 3 shows
i
the average results of 1000 independent simulation runs for
each case. Figs. 3(a) and 3(b) present the average utility per
aggregator corresponding to the linear and quadratic cost
functions, respectively. We can see that our approach always gives higher average utility than the greedy one, as
shown in Figs. 3(a) and 3(b). In the cases of linear cost, we
can see that the average utility start by increasing with N
but then, it starts to decrease when the number of aggregators reaches that of buyers (N = 6). This result is due to the
fact that, for small N , an increase in the number of partic1 The numbers in Subfigure (b) correspond to iteration/case numbers
ipating aggregators leads to a larger amount of energy sold in Subfigure (a). For clearer demonstration, In Subfigure (b), only the
which, subsequently, improves the average utility. However, curves corresponding to the first few iterations/cases are given.
6
References
2
(a) Total utility in each iteration
4 3
[1] T. A. Becker and I. Sidhu, “Electric vehicles in the
united states: A new model with forecasts to 2030,”
Tech. Rep. 2009.1.v.2.0, Center for Entrepreneurship
& Technology, University of California, Berkeley, Aug.
2009.
Demand
1
[2] A. N. Brooks and S. H. Thesen, “PG&E and Tesla
motors: Vehicle to grid demonstration and evaluation
program,” in Proc. 23rd International Electric Vehicles
Symposium and Exposition, (Anaheim, CA, USA), Dec.
2007.
(b) Supply and demand curves
Figure 5: Quadratic cost function with 1000 buyers (K) and
1000 aggregators (N)
[3] Electric Power Research Institute (EPRI), “PAP11electric vehicle roaming scenarios,” EPRI Research Program, Use Case.
5
4
Supply
3
2
[4] National Institute for Standards and Technology
(NIST), “SGIP CoS: SAE J2836 use cases for communication between plug-in vehicles and the utility grid,”
NIST Standards.
1
(a) Total Utility in each case
(b) Supply and demand curves
[5] C. Guille and G. Gross, “A conceptual framework for the
vehicle-to-grid (V2G) implementation,” Energy Policy,
vol. 37, pp. 4379–4390, Nov. 2009.
Figure 6: Fixed Q0 with 1000 buyers (K) and 1000 aggregators (N)
[6] B. K. Sovacool and R. F. Hirsh, “Beyond batteries:
An examination of the benefits and barriers to plug-in
hybrid electric vehicles (PHEVs) and a vehicle-to-grid
(V2G) transition,” Energy Policy, vol. 37, pp. 1095–
1103, Mar. 2009.
slope (α). For example, we consider iterations 1 and 2. α1 is
larger than α2 and we have the utility gained in iteration 1 is
smaller than that in iteration 2. We also study the quadratic
PHEV cost functions and the similar results are shown in
Fig. 5.1 However, with the quadratic cost function, the algorithm converges faster and smoother. Next we study the
results for six different demand curves with a fixed supply
curve (i.e. α fixed). Fig. 61 shows the six cases with a fixed
Q0 . The simulation also aligns the results deduced in Section
3.1: the larger β, the smaller the utility.
5
[7] J. Pillai and B. Bak-Jensen, “Vehicle-to-grid for islanded
power system operation in bornholm,” in Power and
Energy Society General Meeting, 2010 IEEE, pp. 1 –8,
july 2010.
[8] C. Wu, H. Mohsenian-rad, and J. Huang, “Vehicle-toaggregator interaction game,” IEEE Trans. Smart Grid,
to appear 2011.
Conclusion
[9] M. Erol-Kantarci and H. Mouftah, “Management of
We have studied the V2G energy trading market, in which
phev batteries in the smart grid: Towards a cyberPHEVs sell their excessive energy to some loads in the disphysical power infrastructure,” in Wireless Communitribution network. We have proposed a multi-layer system
cations and Mobile Computing Conference (IWCMC),
to coordinate the trading where those PHEVs in a close ge2011 7th International, pp. 795 –800, july 2011.
ographical area are represented by an aggregator. In the
macro layer, the energy buyers and the aggregators engage [10] N. Matta, R. Rahim-Amoud, L. Merghem-Boulahia,
in a double auction to determine the trading price and the
and A. Jrad, “A cooperative aggregation-based architecactual energy quantity sold for each aggregator. In the micro
ture for vehicle-to-grid communications,” in Global Inlayer, each aggregator helps its managed PHEVs maximize
formation Infrastructure Symposium (GIIS), 2011, pp. 1
their utilities. A mechanism is proposed to coordinate the
–6, aug. 2011.
interaction between the macro and micro layers. We have
analyzed the system performance when the numbers of buy- [11] E. Sortomme and M. A. El-Sharkawi, “Optimal schedulers and aggregators are large. Simulation results show that
ing of vehicle-to-grid energy and ancillary services,”
our mechanism is always better than the greedy approach
Smart Grid, IEEE Transactions on, vol. PP, no. 99, p. 1,
and verify some of our analytical results.
2011.
7
[12] W. Saad, Z. Han, H. V. Poor, and T. Basar, “A noncooperative game for double auction-based energy trading between phevs and distribution grids,” in Proc.
IEEE International Conference on Smart Grid Communications (SmartGridComm), (Brussels, Belgium), Oct.
2011.
[13] V. Krishna, Auction Theory. 2002.
[14] P. Huang, A. Scheller-Wolf, and K. Sycara, “A strategyproof multiunit double auction mechanism,” in Proceedings of the first international joint conference on Autonomous agents and multiagent systems: part 1, AAMAS ’02, 2002.
[15] P. Huang, A. Scheller-Wolf, and K. Sycara, “Design of
a multi-unit double auction e-market,” Computational
Intelligence, vol. 18, pp. 596–617, Feb. 2002.
[16] D. Friedman and J. Rust, The Double Auction Market: Institutions, Theories, and Evidence. Boulder, CO,
USA: Westview Press, 1993.
[17] C. Silva, M. Ross, and T. Farias, “Evaluation of energy
consumption, emissions and cost of plug-in hybrid vehicles,” Elsevier Energy Conversion and Management,
vol. 50, pp. 1635–1643, Jul. 2009.
[18] T. Motors, “Roadster innovations: Motor,” Feb. 2011.
8
| 3cs.SY
|
1
On a class of optimization-based robust estimators
arXiv:1705.02837v1 [cs.SY] 8 May 2017
Laurent Bako
Abstract—We consider in this paper the problem of estimating
a parameter matrix from observations which are affected by two
types of noise components: (i) a sparse noise sequence which,
whenever nonzero can have arbitrarily large amplitude (ii) and a
dense and bounded noise sequence of "moderate" amount. This is
termed a robust regression problem. To tackle it, a quite general
optimization-based framework is proposed and analyzed. When
only the sparse noise is present, a sufficient bound is derived on
the number of nonzero elements in the sparse noise sequence that
can be accommodated by the estimator while still returning the
true parameter matrix. While almost all the restricted isometrybased bounds from the literature are not verifiable, our bound
can be easily computed through solving a convex optimization
problem. Moreover, empirical evidence tends to suggest that it is
generally tight. If in addition to the sparse noise sequence, the
training data are affected by a bounded dense noise, we derive
an upper bound on the estimation error.
I. I NTRODUCTION
In many engineering fields such as control system design,
signal processing, machine learning or statistics, one is frequently confronted with the problem of empirically uncovering
a mathematical relationship between a number of signals of
interest. The usual method to achieve this goal is to run an
experiment during which one measures (a finite number of)
samples of the relevant signals and proceed with fitting a
certain model structure to the experimental data samples. This
process is known as system identification [11], [19]. A issue of
critical importance during this process is that the experimental
data samples might be contaminated by a measurement noise
of relatively high level due for example to intermittent sensor
failures or various communication disruptions. To cope with
the troublesome effects of the noise, the model estimation must
be designed with care.
In this paper we consider the situation where the data are
corrupted by two types of noise: a sparse noise sequence which
shows up only intermittently in time but can take on arbitrarily
large values whenever it is nonzero; and a more standard dense
noise component of moderate amount.
II. T HE
ROBUST REGRESSION PROBLEM
Consider a system described by an equation of the form
yt = Ao xt + ft + et
(1)
where yt ∈ Rm and xt ∈ Rn are respectively the output and
the regressor vector at time t; Ao ∈ Rm×n is an unknown
parameter matrix; ft and et are some noise terms which are
unobserved.
L. Bako is with Laboratoire Ampère – Ecole Centrale de Lyon – Université
de Lyon – 36 avenue Guy de Collongue, 69134 Ecully, France – E-mail:
[email protected]
N
Problem. Given a finite collection {xt , yt }t=1 of measurements obeying the relation (1), the robust regression problem
of interest here is the one of finding an estimate of the
parameter matrix Ao under the assumptions that {et } and {ft }
are unknown but enjoy the following (informal) properties:
• {et } is a dense noise sequence with bounded elements
accounting for moderate model mismatches or measurement noise.
• {ft } is such that the majority of its elements are equal
to zero while the remaining nonzero elements can be of
arbitrarily large magnitude. The nonzero elements of that
sequence are usually termed gross errors or outliers. They
can account for possible intermittent sensor faults. We
will refer to {ft } as the sequence of sparse noise.
For the time being, these are just informal descriptions of the
characteristics of the sequences {ft } and {et }. They will be
made more precise whenever necessary in the sequel for the
need of stating more formal results.
Let Y ∈ Rm×N and X ∈ Rn×N be data matrices formed respectively with N output measurements and regressor vectors.
Then it follows from (1) that
Y = Ao X + E + F,
m×N
(2)
m×N
where E ∈ R
and F ∈ R
are unknown noise
components. The matrices Y and X can be structured or
not, depending on whether the system (1) is dynamic or not.
For example, when the model (1) is of MIMO FIR type, Y
contains a finite collection of output measurements while X
is a Hankel matrix containing lagged inputs of the system. In
this case Y and X take the form
Y = y1 y2 · · · yN ,
u1
u2
···
uN
u0
u1
· · · uN −1
X= .
.
.. .
..
..
···
.
u1−nf
u2−nf
···
uN −nf
where {ut } and {yt } stand respectively for the input and
output of the system and the maximum lag nf is called the
order of the model. In the sequel, the notations of the type yt
and xt with subindex t ∈ I , {1, . . . , N } refer to the columns
of Y and X respectively.
Relevant prior works. The so formulated regression problem
is called a robust regression problem in connection with the
fact that the error matrix F assume columns of (possibly)
arbitrarily large amplitude. It has applications in e.g., the identification of switched linear systems [1], [15], [14], subspace
clustering [2], etc. Existing approaches for solving the robust
regression problem can be roughly divided into two groups:
methods from the field of robust statistics [17], [12], [9] which
have been developed since the early 60s and a class of more
2
recent methods inspired by the compressed sensing paradigm
[3], [4], [18], [21], [13]. The first group comprises methods
such as the least absolute deviation (LAD) estimator [8], the
least median of squares [16], the least trimmed squares [17],
the family of M-estimators [9]. The latter group can be viewed
essentially as a refreshed look at the so-called least absolute
deviation method. There has been however a fundamental
shift of philosophy in the analysis. While in the framework
of robust statistics, robustness of an estimator is measured
in terms of the breakdown point (the asymptotic minimum
proportion of points which cause the estimation error induced
by an estimator to be unbounded if they were to be arbitrarily
corrupted by gross errors), in the compressed-sensing-inspired
category of robust methods, the analysis aims generally at
characterizing properties of the data that favor exact recovery
of the true parameter matrix Ao . In this latter group, the LAD
estimator is sometimes regarded as a convex relaxation of a
combinatorial sparse optimization problem.
To the best of our knowledge, only the papers [18] provides
an explicit bound on the estimation error induced by the
LAD estimator. However that bound does not fully apply to
the current setting since the estimators although similar are
of different natures. Indeed, the LAD estimator stands only
as a special case of the current framework. Moreover the
bound in [18] is not easily computable while ours is. The
references [4] and [13] provide some bounds for a noise-aware
version of the LAD estimator which are based respectively
on the Restricted Isometry Property (RIP) and a measure of
subspace angles. Unfortunately numerical evaluation of those
bounds is a process of exponential complexity, a price that is
unaffordable in practice.
A related but different problem from the regression problem
considered here is that of sparse signal recovery studied in the
field of compressed sensing [5], [7]. This is about finding the
sparsest solution to an underdetermined set of linear equations.
Various analysis approaches have been devised which rely on
the RIP constant, the mutual coherence, the nullspace property,
to name but a few. Again, these analysis results either cannot
be extended efficiently to the robust regression problem or lead
to bounds that are NP-hard to compute [20], [10], [6].
Contributions. In this paper we propose and analyze a class
of optimization-based robust estimators. It is shown that the
robust properties of the estimators are essentially inherited
from a key property of the to-be-optimized performance function (or loss function) called column-wise summability. The
proposed framework admits the LAD estimator and its usual
variants as special cases. Moreover it applies to both SISO and
MIMO systems. When the dense noise component E in (2)
is identically equal to zero, we derive bounds on the number
of gross errors (nonzero columns of F ) that the estimator is
able to accommodate while still returning the true parameter
matrix Ao . In comparison with the existing literature, the
proposed bounds have the important advantage that they are
numerically computable through convex optimization. When
both E and F are active, exact recovery of the true parameter
matrix is no longer possible. In this scenario, we derive upper
bounds on the parametric estimation error in function of the
amplitude of E and the number of nonzero columns of F .
Again, computable but (possibly) looser versions of those
bounds are obtainable.
The current paper can be viewed as a generalization of our
previous work reported in [3]. While [3] provides an analysis
of mostly a single estimator (namely the LAD estimator)
relying on nonsmooth optimization theory, we focus here
on a much larger class of optimization-based robust estimators by highlighting some key robustness-inducing properties.
Moreover, we provide, for the considered class of estimators,
stability results which permit the estimation of parametric error
bounds.
Outline. The rest of the paper is organized as follows. Section
III defines the optimization-based approach to the robust
regression problem. Section IV discusses the properties of the
proposed estimation framework. Section V provides further
comments. Section VI reports some numerical experiments.
Lastly, Section VII contains some concluding remarks.
Notations. R is the set of real numbers; R≥0 (respectively
R>0 ) is the set of nonnegative (respectively positive) real
numbers ; RN is the space of N -tuples (vectors) of real
numbers. For any vector x = [x1 · · · xN ]⊤ ∈ RN , the
p-norm of x with p ∈ {1, . . . , ∞} is defined by kxkp =
PN
p 1/p
. A special case is the limit case p = ∞
i=1 |xi |
in which kxk∞ = maxi=1,...,N |xi |. For any matrix A =
[a1 · · · aN ] with ai ∈ Rm , the induced p-norm of A is
defined by kAkp = supx∈RN ,kxkp =1 kAxkp .
Cardinality of a finite set. Throughout the paper, whenever S
is a finite set, the notation |S| will refer to the cardinality of
S. However, for a real number x, |x| will denote the absolute
value of x.
Submatrices and subvectors. Let X ∈ Rn×N and I =
{1, . . . , N } be the index set for the columns of X. If I ⊂ I,
the notation XI denotes a matrix in Rn×|I| formed with the
columns of X indexed by I. We will use the convention that
XI = 0 ∈ Rn when the index set I is empty.
III. A
CLASS OF ROBUST ESTIMATORS
Let DN be the set of N data points generated by system
(1) for any possible values of the noise sequences, i.e.,
n
DN = (Y, X) ∈ Rm×N × Rn×N :
o
f
e
∃(E, F ) ∈ GN
× GN
, (2) holds ,
f
e
with GN
⊂ Rm×N and GN
⊂ Rm×N denoting the set of dense
and sparse noise matrices respectively. The estimation problem
aims at determining the unknown parameter matrix Ao given
a point (Y, X) in DN . Of course, this quest would not make
much sense if the noises E and F were completely arbitrary
since in this case, we would have DN = Rm×N × Rn×N
hence losing any informativity concerning the data-generating
system. Therefore some minimum constraints need to be put
on E and F as informally described above.
With respect to the estimation problem just stated, an estimator
is a set-valued map Ψ : DN → P(Rm×n ), (Y, X) 7→
Ψ(Y, X) which is defined from the data space DN to the power
3
set P(Rm×n ) of the parameter space. For (Y, X) generated
by a system of the form (1), one would like to design an
estimator achieving, whenever possible, the ideal property that
Ψ(Y, X) = {Ao }. In default of that ideal situation, a more
pragmatic goal is to search for a Ψ so that Ao ∈ Ψ(Y, X) and
Ψ(Y, X) is of small size in some sense despite the troublesome
effects of the unknown noise components E and F . The design
of an optimal estimator requires specifying a performance
index (usually called a loss function) which is to be minimized.
In this paper, we study the properties of the estimator of
the parameter matrix Ao in (2) defined by
Ψ(Y, X) = arg min ϕ(Y − AX)
(3)
A∈Rm×n
where ϕ : M (R) → R≥0 is a convex function defined on the
set M (R) of (all) real matrices. It is assumed that ϕ has the
following properties:
P1. For all B, C ∈ M (R) of compatible dimensions,
ϕ([B
C]) = ϕ(B) + ϕ(C)
(5)
P3. There exists a constant real number ε ≥ 0 such that for
all B ∈ M (R) with n rows and N columns,
ℓ(B) − |Iεc (B)| ε ≤ ϕ(B) ≤ ℓ(B)
IV. P ROPERTIES
(6)
where
We first study the conditions under which the true parameter
matrix Ao in (1) can be exactly recovered. Theorem 1 and
Theorem 2 stated next show that if the number of nonzero
columns in the matrix V , E + F is less than a certain
threshold, then Ψ(Y, X) = {Ao }.
Theorem 1 (A necessary and sufficient condition). Let ϕ
be a function satisfying (4)-(6) with ε = 0 and Ψ be
defined as in (3). Let d be an integer and assume that
rank(X) = n. For any A ∈ Rm×n and Y ∈ Rm×N , let
Ic (Y − AX) = {t ∈ I : yt − Axt 6= 0}. Then the following
statements are equivalent.
(i)
Iεc (B) = i ∈ {1, . . . , N } : ℓ(bi ) > ε
∀A ∈ Rm×n , ∀Y ∈ Rm×N , |Ic (Y − AX)| ≤ d
and |Iεc (B)| is the cardinality of Iεc (B) and bi ∈ Rn is
the ith column of the (n, N )-matrix B.
The property (4) will be called column-wise summability.
Since ϕ is a function defined over the space of real matrices of any dimensions, it is also defined for n-dimensional
vectors of real numbers. Hence according to property (4), if
B = [b1 · · · bN ] with column vectors bi ∈ Rn , then
ϕ(B) =
N
X
ϕ(bi ).
i=1
The so-defined function ϕ is not necessarily a norm. For any
εo ≥ 0 and any vector norm ℓo , it can be verified that the
function ϕ defined by
ϕ(B) =
N
X
max(0, ℓo (bi ) − εo )
OF THE ROBUST ESTIMATORS
A. Exact recoverability
(4)
with [B C] denoting the matrix formed by concatenating column-wise B and C.
P2. There exists a matrix norm ℓ : M (R) → R≥0 such that
for all B, C ∈ M (R), conformable for addition,
ϕ(B) ≤ ϕ(B − C) + ℓ(C)
enjoys some impressive robustness properties with respect to
the sparse noise matrix F . The term sparse is used here to
mean that a relatively large proportion of the column vectors
of F are equal to zero. And saying that Ψ is robust with
respect to F means that Ψ(Y, X) does not depend on (or
is insensitive to) the magnitudes of the nonzero columns of
F under the sparsity condition. Therefore those few columns
which are nonzero can have arbitrarily large magnitude. As
will be shown in the sequel, the robustness properties of Ψ are
inherited from the properties P1-P3 of the objective function
ϕ. In the special case where ϕ is a norm, the properties P2P3 are automatically satisfied so that P1 becomes the only
key property required. As to the convexity of ϕ, it is intended
just for computational reasons as it eases the solving of the
optimization problem in (3).
(7)
i=1
is positive and convex and satisfies properties (4)-(6) but it
is not a norm for εo > 0 since in this case, ϕ(B) = 0 does
not imply that B = 0. But if εo = 0 in (7), then ϕ = ℓ
by (6) soP
that ϕ corresponds to the matrix norm defined by
N
ϕ(B) = i=1 ℓo (bi ). We note in this latter case that (6) is
trivial while (5) reduces to the triangle inequality.
We will show in the sequel that the estimator Ψ in (3)
⇒
(ii)
max
c
max
I ⊂I: Λ∈Rm×n
|I c |=d
Λ6=0
Ψ(Y, X) = A
(8)
1
ϕ(ΛXI c )
<
ϕ(ΛX)
2
(9)
Here and in the following, the notation I , {1, . . . , N } is used
to denote the index set for the columns of the data matrices.
Proof: We first note that the rank assumption on X is
intended to insure that (9) is well-defined since then, with ϕ
being a norm, ϕ(ΛX) 6= 0 whenever Λ 6= 0.
(i) ⇒ (ii): Assume that (i) holds.
Consider an arbitrary subset I c of I such that |I c | = d.
Let Λ be any matrix in Rm×n satisfying Λ 6= 0. Finally,
consider a matrix Y ∈ Rm×N defined by YI c = 0 and
YI 0 = ΛXI 0 where I 0 = I \ I c . Then Ic (Y − ΛX) ⊂ I c and
so |Ic (Y − ΛX)| ≤ d. Hence by (i) {Λ} = arg minH ϕ(Y −
HX) which means that ϕ(Y − ΛX) < ϕ(Y − HX) for any
H ∈ Rm×n , H 6= Λ. In particular, by taking H = 0 we get
ϕ(Y − ΛX) < ϕ(Y ). It follows from the property (4) that
ϕ(YI c − ΛXI c ) + ϕ(YI 0 − ΛXI 0 ) < ϕ(YI c ) + ϕ(YI 0 ).
Using now the relations YI c = 0 and YI 0 = ΛXI 0
yields ϕ(ΛXI c ) < ϕ(ΛXI 0 ) or, equivalently, ϕ(ΛXI c ) <
4
1/2ϕ(ΛX). Eq. (9) then follows from the fact that I c and
Λ are arbitrary.
(ii) ⇒ (i): To begin with, note that if Eq. (9) holds for some
d, then it holds also for any d0 ≤ d. As a result, the equality
|I c | = d in (9) can be changed to |I c | ≤ d. Assuming
(ii), let A ∈ Rm×n and Y ∈ Rm×N be matrices satisfying
|Ic (Y − AX)| ≤ d. Set I c = Ic (Y − AX) and I 0 = I \ I c .
Then for all Λ ∈ Rm×n such that Λ 6= 0,
2ϕ(ΛXI c ) < ϕ(ΛX) = ϕ(ΛXI c ) + ϕ(ΛXI 0 ),
where the equality is obtained by the property (4) of ϕ. It
follows that
ϕ(ΛXI c ) < ϕ(YI 0 − (A + Λ)XI 0 ).
(10)
On the other hand, we know by (5) that
ϕ(YI c − AXI c )−ϕ(YI c − (A + Λ)XI c ) ≤ ϕ(ΛXI c ).
Combining with the inequality (10) yields
ϕ(Y − AX) < ϕ(Y − (A + Λ)X).
Since Λ is an arbitrary nonzero matrix, this inequality says
that A is the unique minimizer of V (H) = ϕ(Y − HX).
Consider a data pair (Y, X) generated by (1). By letting
πϕc (X) = max d : Eq. (9) holds ,
(11)
and assuming that πϕc (X) > 0 we can see that whenever
|Ic (Y − Ao X)| ≤ πϕc (X), Ao can be exactly recovered by
computing Ψ(Y, X). Of course this is likely to hold only
if the dense noise component E does not exist. So in the
situation where E = 0, the theorem says that Ao can be
uniquely obtained by convex optimization provided that the
number of outliers (nonzero columns of F ) is less than or
equal to πϕc (X). For the condition of exact recoverability to
be checkable we must be able to compute πϕc (X). The bad
news are that evaluating numerically such a number is likely
to be NP-hard in most cases.
In the sequel, we investigate sufficient conditions of exact recovery which are more tractable from a numerical standpoint.
For this purpose let us introduce some definitions.
Definition 1. A matrix X = [x1 · · · xN ] ∈ Rn×N is said
to be self-decomposable if rank(X) = n and for all k ∈ I,
xk ∈ im(X6=k ) where X6=k , XI\{k} is the matrix obtained
from X by removing its k-th column and im(·) refers to range
space.
For a matrix to be self-decomposable it is enough that X6=k
be full row rank for any k ∈ I. Achieving this condition in
practice seems easy provided that the number N of measurements is large enough compared to the dimension n of X.
Definition 2 (self-decomposability amplitude). Let X ∈
Rn×N be a self-decomposable matrix. We call selfdecomposability amplitude of X, the number ξ(X) defined
by
o
n
(12)
kγk k∞ : xk = X6=k γk .
ξ(X) = max min
k∈I γk ∈RN −1
The so-defined ξ(X) constitutes a quantitative measure of
richness (or genericity) of the regressor matrix X. By richness
it is meant here how much, in a global sense, the columns of
X are linearly independent. ξ(X) is expected to be small if
the columns of X are somehow strongly linearly independent.
Remark 1. If for some k the norm of xk was to be considerably large in comparison to the norm of the other columns of
X, then ξ(X) would get large hence reducing recoverability
capacity of the considered class of estimators (see also Eq.
(9)). Such situations can be alleviated by normalizing each
column of X, i.e., for example by replacing (yk , xk ) by
(ỹk , x̃k ) , (yk / kxk k , xk / kxk k) under the assumption that
xk 6= 0 for all k ∈ I.
With the help of the device of self-decomposability amplitude (12), we can state a condition for exact recovery of the
parameter matrix Ao by solving the optimization problem in
(3). A similar result was proven in [3] for the Least Absolute
Deviation (LAD) estimator.
Theorem 2 (A sufficient condition for exact recovery). Let ϕ
be a function satisfying (4)-(6) with ε = 0 and Ψ be defined as
in (3). Assume that X is self-decomposable. Then the following
statement is true:
∀A ∈ Rm×n , ∀Y ∈ Rm×N ,
|Ic (Y − AX)| < T ξ(X)
⇒
(13)
Ψ(Y, X) = A .
where T : R>0 → R>0 is the function defined by T (α) =
1
1
1+
.
2
α
Proof: The proof is completely parallel to that of Theorem
11 in [3]. From the assumptions, each xk , k ∈ I, can be
written as a linear combination of the columns of X6=k . Let
γk ∈ RN −1 be any vector satisfying xk = X6=k γk . It follows
that for any Λ ∈ Rm×n ,
X
γk,t Λxt
ϕ(Λxk ) = ϕ
t∈I\{k}
with γk,t denoting the entry of γk ∈ RN −1 indexed by t.
Under the assumptions of the theorem, ϕ is a norm. So, it
is positive and satisfies the triangle inequality property. As a
result we can write
X
ϕ(Λxk ) ≤
|γk,t | ϕ(Λxt ) ≤ kγk k∞ (ϕ(ΛX) − ϕ(Λxk ))
t6=k
where the rightmost term follows from the property (4) of ϕ.
Since this holds for any γk such that xk = X6=k γk , it holds
also for
n
o
γk⋆ = arg min kγk∞ : xk = X6=k γ .
γ∈RN −1
Hence,
ϕ(Λxk ) ≤ ξ(X) (ϕ(ΛX) − ϕ(Λxk )) ∀k ∈ I, ∀Λ ∈ Rm×n .
(14)
or equivalently,
ϕ(Λxk ) ≤
ξ(X)
ϕ(ΛX) ∀k ∈ I, ∀Λ ∈ Rm×n .
1 + ξ(X)
Let I c be any subset of I and pose |I c | = d. Summing the
5
previous inequality over the set I c yields
max
Λ6=0
ϕ(ΛXI c )
1
|I c |
≤
ϕ(ΛX)
2T ξ(X)
relation
(15)
Note that the term on the right hand side is well-defined since
by the self-decomposability assumption, rank(X) = n which
implies that ϕ(ΛX)
6= 0 whenever Λ 6= 0. Therefore (9) holds
if |I c | < T ξ(X) and the conclusion follows from Theorem
1.
It is worth noting that the threshold T (ξ(X)) on the number
of correctable outliers does not depend on ϕ. Hence this
threshold is valid when the estimator is defined from any
matrix norm obeying (4).
Remark 2. The statement of Theorem 2 still holds true if we
replace ξ(X) with the ϕ-dependent number δϕ (X) defined by
ϕ(Λxk )
δϕ (X) = max sup
k∈I Λ6=0 ϕ(ΛX6=k )
(16)
when it is assumed that ϕ is a norm and rank(X6=k ) = n
for all k. Doing so will give a less conservative condition
for exact recovery. However δϕ (X) seems much harder to
evaluate numerically than ξ(X).
Remark 3 (A few useful properties of ξ(X)).
•
•
For any nonsingular matrix R ∈ Rn×n , ξ(RX) = ξ(X).
It follows that the number ξ(X) depends only on the
subspace spanned by the rows of the regressor matrix X.
For any self-decomposable X ∈ Rn×N , ξ(X) is lowerbounded in the following sense
1
,
N −1
This follows from the more general observation that
ξ(X) ≥
ξ(X) ≥ max P
k∈I
kxk k
t6=k kxt k
for any vector norm k·k. As a result, T (ξ(X) is upperbounded as follows
N
.
2
Theorem 2 provides a sufficient condition for exact recovery
in the situation where the function ϕ is a norm. Next, another
condition is stated which holds in the general case.
T (ξ(X)) ≤
ϕ(YI c − AXI c ) − ϕ(YI c − (A + Λ)XI c ) < ϕ(ΛXI 0 ).
By (5), we can note that ϕ(YI c − AXI c ) − ϕ(YI c − (A +
Λ)XI c ) ≤ ℓ(ΛXI c ). It then follows that
ℓ(ΛXI c ) < ϕ(ΛXI 0 )
is a sufficient condition for Ψ(Y, X) = {A}. Finally, invoking
(6) allows us to observe that ℓ(ΛXI 0 ) − |Iεc (ΛXI 0 )| ε ≤
ϕ(ΛXI 0 ) which implies that ℓ(ΛXI c ) < ℓ(ΛXI 0 ) −
|Iεc (ΛXI 0 )| ε is a sufficient condition for Ψ(Y, X) = {A}.
We have hence proved the proposition.
B. Uncertainty set induced by dense noise
When both E and F are nonzero in the data-generating
system (1), Ψ(Y, X) is likely to be a non-singleton subset of
P(Rm×n ) especially if we consider all possible realizations of
the unknown components E and F . In this case the desirable
properties of the estimator are in default of better (i) that it
contains Ao and (ii) that its size with respect to some metric
is as small as possible. In this section we are interested in
estimating the size of Ψ(Y, X) when both dense noise E and
sparse noise F are active in the data-generating system (1).
A notion of estimator gain. Similarly to the concept of
system gain in control [22], one could define the gain of an
estimator, that is, a quantitative measure of the sensitivity
of the estimator with respect to the perturbations affecting
the measurements. Consider a data pair (Y, X) generated
by a system of the form (1) with Ao being the parameter
matrix sought for. Let us fix the sparse noise matrix F or
view it somehow as part of the data-generating system. This
consideration proceeds from the fact that Ψ can be insensitive
to F (when acting alone) under, for example, the condition
derived in Theorem 2. Let E be bounded in the sense that
ℓ(E) is finite with ℓ being the norm appearing in (6). Then
we can define a gain of the estimator with respect to the dense
noise component E. More specifically, an (ℓ, q)-gain of the
estimator Ψ with respect to the dense noise E may be defined
by
kA⋆ − Ao kq
.
(18)
gℓ,q (Y, X) =
sup
ℓ(E)
A⋆ ∈Ψ(Y,X)
0<ℓ(E)<∞
F sparse
Proposition 1. Consider a triplet (ϕ, ℓ, ε) satisfying (4)-(6).
For A ∈ Rm×n and Y ∈ Rm×N , pose I c = Ic (Y − AX),
I 0 = I \ I c = {t ∈ I : yt − Axt = 0} and Iεc (ΛXI 0 ) =
t ∈ I 0 : ℓ(Λxt ) > ε . Then Ψ(Y, X) = {A} if
Here k·kq denotes matrix q-norm. The so-defined number
gℓ,q (Y, X) provides an upper bound on the distance from the
set Ψ(Y, X) to Ao in function of the amount of dense noise.
The following theorem and its corollaries show that if the
number of nonzero columns in F is no larger than a certain
threshold, then gℓ,q (Y, X) exists and is finite.
∀Λ ∈ Rm×n , Λ 6= 0.
Theorem 3. Let (Y, X) be the data generated by system (1)
subject to the noise components E and F . Consider a triplet
(ϕ, ℓ, ε) satisfying (4)-(6). Let S 0 ⊂ I be a set such that FS 0 =
0 and let S c = I \ S 0 . Assume that the matrix X and the
partition (S 0 , S c ) are such that there exists α > 0 such that
|Iεc (ΛXI 0 )| ε < ℓ(ΛXI 0 ) − ℓ(ΛXI c )
(17)
Proof: Ψ(Y, X) = {A} is equivalent to
ϕ(Y − AX) < ϕ(Y − (A + Λ)X)
for any Λ ∈ Rm×n , Λ 6= 0. Using the definitions of the sets
I 0 and I c and applying property (4) of ϕ yields the equivalent
ℓ(ΛXS 0 ) − ℓ(ΛXS c ) ≥ α kΛkq ∀Λ ∈ Rm×n ,
(19)
6
with k·kq denoting some matrix q-norm.
Then for any A⋆ ∈ Ψ(Y, X), it holds that
1
kA⋆ − Ao kq ≤
2ℓ(ES 0 ) + |Iεc | ε
(20)
c
γℓ,q (X, S )
with1 Iεc = Iεc (YS 0 − A⋆ XS 0 ) = t ∈ S 0 : ℓ(yt − A⋆ xt ) > ε
and
ℓ(ΛXS 0 ) − ℓ(ΛXS c )
(21)
γℓ,q (X, S c ) = inf
m×n
kΛkq
Λ∈R
Λ6=0
if the dense noise matrix E is such that ℓo (et ) ≤ εo for
all t ∈ I, then by taking ε = εo the set Iεc defined in the
statement of Theorem 3 corresponds to the empty set so that
(22) holds as well in this case. In connection with the concept
of estimator gain discussed earlier, one can interpret the factor
2/γℓ,q (X, S c ) as an estimate of the gain (of the estimator Ψ)
with respect to dense noise.
Lastly, it is interesting to see that when ϕ is a norm, if E =
0 then the result of Theorem 3 implies that Ψ(Y, X) = {Ao }
provided (19) is true.
where k·kq refers to matrix q-norm.
V. D ISCUSSIONS
Proof: By definition of Ψ(Y, X) in (3),
For the purpose of illustrating the extent of the results above,
let us discuss further the situation where ϕ reduces to a norm.
ϕ(Y − A⋆ X) ≤ ϕ(Y − AX) ∀A ∈ Rm×n
By letting Λ = A − Ao , Λ⋆ = A⋆ − Ao and applying (2), the
last inequality takes the form
⋆
ϕ(F + E − Λ X) ≤ ϕ(F + E − ΛX) ∀Λ ∈ R
m×n
.
In particular, for Λ = 0, we get ϕ(F + E − Λ⋆ X) ≤ ϕ(F + E)
which, thanks to property (4) of ϕ, takes the form
ϕ(FS c + ES c − Λ⋆ XS c )+ϕ(ES 0 − Λ⋆ XS 0 )
≤ ϕ(FS c + ES c ) + ϕ(ES 0 ).
Now applying property (5) to the first member of the left hand
side and rearranging yields
⋆
A. Scenario when the loss function is a norm
Corollary 1. Let (Y, X) be the data generated by system (1)
subject to the noise components E and F . Let S 0 and S c be
defined as in the statement of Theorem 3. Assume that ϕ is a
norm i.e., it satisfies (4)-(6) with ε = 0.
If X is self-decomposable and |S c | < T ξ(X) , then for any
A⋆ ∈ Ψ(Y, X),
kA⋆ − Ao kq ≤ Bϕ,q (|S 0 |, X)ϕ(ES 0 )
2
,
h
N −r i
σϕ,q (X) 1 −
T (ξ(X))
ϕ(ΛX)
σϕ,q (X) = inf
Λ6=0 kΛkq
Bϕ,q (r, X) =
Using (6) gives
ℓ(ES 0 − Λ⋆ XS 0 ) − |Iεc | ε − ℓ(Λ⋆ XS c ) ≤ ϕ(ES 0 ) ≤ ℓ(ES 0 ).
Here we used the fact that Iεc (ES 0 − Λ⋆ XS 0 ) is equal to the
set Iεc defined in the statement of the theorem.
Applying the triangle inequality property of ℓ, it can be seen
that ℓ(Λ⋆ XS 0 ) − ℓ(ES 0 ) ≤ ℓ(ES 0 − Λ⋆ XS 0 ). Combining with
the previous inequality yields
ℓ(Λ⋆ XS 0 ) − ℓ(Λ⋆ XS c ) ≤ 2ℓ(ES 0 ) + |Iεc | ε.
Finally, it follows from the definition of γℓ,q (X, S c ) in (21)
that
γℓ,q (X, S c ) kΛ⋆ kq ≤ 2ℓ(ES 0 ) + |Iεc | ε .
The condition (19) guarantees that γℓ,q (X, S c ) is well-defined
and is positive. Hence the statement of the theorem is established.
Theorem 3 constitutes an interesting stability result in that
it provides a finite upper bound on the distance from Ao to
the set Ψ(Y, X) as a function of the amplitude of the dense
noise matrix E. It applies to any estimator Ψ defined as in (3)
with ϕ a function obeying (4)-(6). In particular, in the situation
where ϕ is a norm (in which case ε can be taken equal to zero
in (6)), the inequality in (20) simplifies to
2
ℓ(ES 0 ).
γℓ,q (X, S c )
(22)
If ϕ is defined as in (7) (which, recall, is not a norm) and
1 The
notation Iεc is used for simplicity reasons.
(23)
where
⋆
ϕ(ES 0 − Λ XS 0 ) − ℓ(Λ XS c ) ≤ ϕ(ES 0 ).
kA⋆ − Ao kq ≤
ON SOME SPECIAL CASES
(24)
(25)
Proof: The principle of the proof is to show that
γℓ,q (X, S c ) is well-defined and then find a positive underestimate of it. Using the property (4) of ϕ and the fact that
ϕ = ℓ, we can write
2ϕ(ΛX) 1 ϕ(ΛXS c )
ℓ(ΛXS 0 ) − ℓ(ΛXS c )
.
=
−
kΛkq
kΛkq
2
ϕ(ΛX)
On the other hand we know from the proof of Theorem 2 (see
Eq. (15)) that
1
ϕ(ΛXS c )
≤
|S c |
ϕ(ΛX)
2T (ξ(X))
so that
1−
|S c |
ϕ(ΛX)
ℓ(ΛXS 0 ) − ℓ(ΛXS c )
≤
T (ξ(X)) kΛkq
kΛkq
Taking now the infimum on both sides of the inequality symbol
over all nonzero matrices Λ ∈ Rm×n yields
|S c |
≤ γℓ,q (X, S c ).
σϕ,q (X) 1 −
T (ξ(X))
It follows from the rank condition imposed on X (by the selfdecomposability assumption) that σϕ,q (X) > 0. This shows
that γℓ,q (X, S c ) is well defined and is strictly positive. Finally,
since ϕ = ℓ, invoking (22) gives the result.
Two important comments can be made at this stage.
7
First it is interesting to note that the bound Bϕ,q (r, X)
is an increasing function of ξ(X). Therefore it is all the
smaller as ξ(X) is small. That is, the error bound will be
small if the data matrix X is rich enough.
• Second, Bϕ,q (r, X) is a decreasing function of r. This
means that the upper bound on the estimation error
decreases when the number of gross error columns in F
decreases. In the extreme case where S 0 = N (no gross
error), Bϕ,q (|S 0 |, X) in (23) reduces to 2/σϕ,q (X).
Beyond these observations it should be noted that
a key
assumption of Corollary 1 is that |S c | < T ξ(X) with S c
being the index set of the nonzero columns in F . Realizing
this condition requires on the one hand that the number of
nonzero columns in the sparse noise matrix F be small and
on the other hand that ξ(X) be small2 (which means that the
data must be generic). Indeed this condition is not necessarily
as strong as it might appear to be at first sight. For example,
it can be relaxed as follows. Observe that the sum E + F
is not uniquely defined from model (2). Taking advantage of
this, one can always absorb in E all nonzero columns of F
whose magnitude does not exceed a certain level. To see this,
let I = {t ∈ S c : ℓ(et + ft ) ≤ εo } where εo = maxt∈I ℓ(et ).
Then we can define Ẽ and F̃ such that E + F = Ẽ + F̃ and
F̃S 0 ∪I = 0 that is, we set ẽt = ft +et and f˜t = 0 for any t ∈ I
and (ẽt , f˜t ) = (et , ft ) otherwise. As a consequence, E and F
in Corollary 1 can be replaced by Ẽ and F̃ respectively so
that |S| and |S c | are replaced by |S| + |I| and |S c | − |I|. The
condition of the corollary then becomes |S c |−|I| < T ξ(X) ,
which is potentially easier to fulfill.
•
Remark 4 (sum of p-norms). Evaluating numerically the
bound Bϕ (r, X) might prove to be a hard problem due to
the potential difficulty in computing the term σϕ,q (X) in (25).
A particular case of interest is when ϕ consists of a sum
of p-norms
the column vectors, i.e. when it is defined by
Pof
N
bN ]. In this case
ϕ(B) =
i=1 kbi kp for B = [b1 · · ·
if we take q = 2 in (23) and (25), it is easy to see that
1/2
1/2
λmin (XX ⊤) ≤ σϕ,2 (X) with λmin (·) denoting the square
root of the minimum eigenvalue. Replacing σϕ,2 (X) with
1/2
λmin (XX ⊤) in (24) yields an overestimate of Bϕ (r, X) which
is computable.
Remark 5. Corollary 1 still holds true if one replaces
T (ξ(X)) with πϕc (X) defined in (11). As shown in [18], the
number πϕc (X) in (11) is computable although at the price
of a combinatorial complexity. However if the n-dimension of
X is small enough the complexity of the algorithm proposed
there can be affordable. Then by using our formula (24) and
Remark 4 above, it is possible therefore to obtain a smaller
bound on the estimation error.
B. Single output case: ℓ1 norm
In this section, we discuss for an illustrative purpose, the
applicability of Theorem 3 to the case of single-output systems. This is an interesting case to highlight since it represents
2 Recall that T is a decreasing function hence implying that T (ξ(X)) is
large when ξ(X) is small.
the most classical situation. Consider the single-output system
defined by
yt = (θo )⊤ xt + ft + et
(26)
where yt , et , ft are scalars and xt and θo are n-dimensional
vectors. By letting Y = [y1 · · · yN ] ∈ R1×N and defining
E and F similarly, we obtain
Y = (θo )⊤ X + F + E.
(27)
This last equation corresponds indeed to (2) where the matrix
Ao reduces
to the row vector (θo )⊤ . In this case, if we let
PN
n
ϕ(B) =
i=1 kbi k2 then for any θ ∈ R , the columns of
(the row vector) Y − AX are scalars so that
ϕ(Y − θ⊤ X) =
N
X
yt − θ⊤ xt
t=1
2
=
N
X
yt − θ⊤ xt . (28)
t=1
As a result, Ψ coincides in this case with the Least Absolute
Deviation (LAD) estimator. The following corollary specializes the result of Theorem 3 to the LAD estimator.
1×N
Corollary 2. Let (Y, X)
× Rn×N be generated by
∈R
c
c
model (26). Let S = t ∈ I : ft 6= 0 , S 0 = I \ S
. Assume
c
that X is self-decomposable and |S | < T ξ(X) . Then for
any θ⋆ ∈ arg min Y − θ⊤ X 1 ,
θ∈Rn
where
kθ⋆ − θo k2 ≤ B1,2 |S 0 |, X kES 0 k1
2
,
h
N −r i
σ1,2 (X) 1 −
T (ξ(X))
X ⊤η 1
.
σ1,2 (X) = inf
η6=0
kηk2
B1,2 (r, X) =
Again here the bound B1,2 (r, X) can be numerically overestimated by following the idea of Remark 4.
VI. N UMERICAL
ILLUSTRATIONS
The performance of the estimator Ψ has been extensively
tested in some existing papers in the special case of the LAD
(see e.g., [3]) . We therefore concentrate here on evaluating
numerically an estimate of the gain of the estimator based on
Corollary 1 and Remark 4. The estimation is carried out for
the case where ϕ consists in the sum of 2-norms and q = 2.
Four different cases are studied:
(a) Static data: X ∈ R2×200 is sampled from a Gaussian distribution N (0, I2 ) with zero-mean and identitycovariance.
(b) Dynamic data generated by a switched linear system:
X ∈ R2×200 is formed with the regressors (yt−1 , ut−1 )
generated by a switched linear system composed of 3
subsystems of order 1. This is a switched ARX system defined by yt = aσ(t) yt−1 + bσ(t) ut−1 with the
switching signal σ(t) ∈ {1, 2, 3} generated from a
uniform distribution and input ut being a white noise
with Gaussian distribution; (a1 , b1 ) = (−0.40, −0.15),
(a2 , b2 ) = (1.55, −2.10) and (a3 , b3 ) = (1, −0.65).
8
3
3
T (ξ(X))
c (X)
πϕ
T (ξ(X))
c (X)
πϕ
Bound
Bound
2
2
1
1
0
0
2
4
6
8
10
12
14
16
18
20
22
24
26
28
0
2
Percentage of nonzeros [%]
(a) static system: ξ(X) = 0.0083
4
6
8
10
12
14
3
18
20
3
T (ξ(X))
c (X)
πϕ
T (ξ(X))
c (X)
πϕ
2
2
Bound
Bound
16
Percentage of nonzeros [%]
(b) switched system: ξ(X) = 0.0127
1
1
0
2
4
6
8
10
12
Percentage of nonzeros [%]
(c) linear system: ξ(X) = 0.0188
0
0
2
4
6
8
10
12
14
16
18
20
22
Percentage of nonzeros [%]
(d) nonlinear system: ξ(X) = 0.0107
Fig. 1: An overestimate of Bϕ using respectively πϕc (X) and T (ξ(X)) for a data matrix X ∈ R2×200 : (a) static data sampled
from a Gaussian distribution; (b) data generated by a switched system; (c) data generated by a linear dynamic system ; (d) data
generated by a dynamic nonlinear system. In each case, the x-axis is limited to the range of nonzero gross errors proportions
which statisfy the stability condition |S c | /N < T ξ(X) /N (see e.g., Corollary 1).
then the so-defined estimator will inherit robustness properties.
Considering a data set generated by a linear model subject to
both sparse and dense noises, we showed that the estimator is
insensitive to the sparse noise when this latter is acting alone
and provided that the number of its nonzero components is no
larger than a certain (computable) threshold. Conditions are
proposed for the exact recovery of the true parameter matrix
when only the sparse noise is active. When both types of noises
affect the measurements we propose computable bounds on
the parametric estimation error. By assuming stochasticity of
the dense noise sequence, the obtained bounds are probably
improvable by exploiting appropriately the statistics of the
dense noise. This is a matter than can be investigated in future
research.
(c) Dynamic data generated by a linear ARX system defined
by yt = a1 yt−1 + b1 ut−1 with the (a1 , b1 ) defined above
in case (b).
(d) Dynamic data generated by a nonlinear NARX system
2
defined by yt = (yt−1 + 2.5)/(1 + yt−1
) + ut−1 .
Following Remark 1, the columns of all data matrices X have
been normalized to unit 2-norm before being processed.
Figure 1 plots the obtained estimate of the estimator gain
against the proportion of correctable outliers. As remarked in
Section V, the gain estimate increases as the proportion of
outliers gets larger. But the growth rate of the gain estimate
depends on the genericity of the data matrix X. The more
generic the columns of X are, the smaller the growth rate
of the estimation error is when regarded as a function of
the proportion of outliers. The experiment confirms also the
intuition according to which static data tend to be more generic
than data generated by a dynamic system. Among the three
cases of dynamic systems, the linear system appears to be the
one generating the least generic data.
The author is grateful to the Associate Editor and the
anonymous reviewers for constructive feedback.
VII. C ONCLUSIONS
R EFERENCES
In this paper we have discussed a somewhat general framework for designing a robust estimator. Given the training data,
the estimator is defined as the minimizing set of a certain
performance index applying to the data. We have shown that
if the performance function possesses some key properties,
[1] L. Bako. Identification of switched linear systems via sparse optimization. Automatica, 47:668–677, 2011.
[2] L. Bako. Subspace clustering through parametric representation and
sparse optimization. IEEE Signal Processing Letters, 21:356–360, 2014.
[3] L. Bako and H. Ohlsson. Analysis of a nonsmooth optimization approach
to robust estimation. Automatica, 66:132–145, 2016.
ACKNOWLEDGEMENT
9
[4] E. Candès and P. A. Randall. Highly robust error correction by convex
programming. IEEE Transactions on Information Theory, 54:2829–
2840, 2006.
[5] E. J. Candès and M. B. Wakin. An introduction to compressive sampling.
IEEE Signal Processing Society, 25:21–30, 2008.
[6] D. L. Donoho, M. Elad, and V. Temlyakov. Stable recovery of
sparse overcomplete representations in the presence of noise. IEEE
Transactions on Information Theory, 52:6–18, 2006.
[7] S. Foucart and H. Rauhut. A mathematical introduction to compressive
sensing. Birkhäuser, 2013.
[8] P. J. Huber. The place of l1 -norm in robust estimation. Computational
Statistics and Data Analysis, 5:255–262, 1987.
[9] P. J. Huber and E. M. Ronchetti. Robust Statistics. A. John Wiley &
Sons, Inc. Publication (2nd Ed), 2009.
[10] A. Juditsky and A. Nemirovski. On verifiable sufficient conditions for
sparse signal recovery via ℓ1 minimization. Mathematical Programming,
127:57–88, 2011.
[11] L. Ljung. System Identification: Theory for the user (2nd Ed.). PTR
Prentice Hall., Upper Saddle River, USA, 1999.
[12] R. A. Maronna, R. D. Martin, and V. J. Yohai. Robust Statistics: Theory
and Methods. John Wiley & Sons, Inc., 2006.
[13] K. Mitra, A. Veeraraghavan, and R. Chellappa. Analysis of sparse
regularization based robust regression approaches. IEEE Transactions
on Signal Processing, 61:1249–1257, 2013.
[14] N. Ozay and M. Sznaier. Hybrid system identification with faulty measurements and its application to activity analysis. In IEEE Conference
on Decision and Control and European Control Conference, Orlando,
FL, USA, 2011.
[15] N. Ozay, M. Sznaier, C. Lagoa, and O. Camps. A sparsification approach
to set membership identification of a class of affine hybrid systems. IEEE
Transactions on Automatic Control, 57:634–648, 2012.
[16] P. J. Rousseeuw. Least median of squares regression. Journal of the
American Statistical Association, 79:871–880, 1984.
[17] P. J. Rousseeuw and A. M. Leroy. Robust Regression and Outlier
Detection. John Wiley & Sons, Inc., 2005.
[18] Y. Sharon, J. Wright, and Y. Ma. Minimum sum of distances estimator:
Robustness and stability. In American Control Conference, St Louis,
MO, USA, 2009.
[19] T. Soderstrom and P. Stoica. System identification. Prentice Hall, Upper
Saddle River, USA, 1989.
[20] A. M. Tillmann and M. E. Pfetsch. The computational complexity
of the restricted isometry property, the nullspace property, and related
concepts in compressed sensing. IEEE Transactions on Information
Theory, 60:1248–1259, 2014.
[21] W. Xu, E.-W. Bai, and M. Cho. System identification in the presence
of outliers and random noises: A compressed sensing approach. Automatica, 50:2905–2911, 2014.
[22] G. Zames. Feedback and optimal sensitivity: Model reference transformations, multiplicative seminorms, and approximate inverses. IEEE
Transactions on Automatic Control, 26:301–320, 1981.
| 3cs.SY
|
Systematic N-tuple Networks for Position
Evaluation: Exceeding 90% in the Othello League
Wojciech Jaśkowski
arXiv:1406.1509v3 [cs.NE] 25 Jun 2014
Institute of Computing Science, Poznan University of Technology, Poznań, Poland
[email protected]
Abstract—N-tuple networks have been successfully used as
position evaluation functions for board games such as Othello
or Connect Four. The effectiveness of such networks depends
on their architecture, which is determined by the placement
of constituent n-tuples, sequences of board locations, providing
input to the network. The most popular method of placing ntuples consists in randomly generating a small number of long,
snake-shaped board location sequences. In comparison, we show
that learning n-tuple networks is significantly more effective
if they involve a large number of systematically placed, short,
straight n-tuples. Moreover, we demonstrate that in order to
obtain the best performance and the steepest learning curve for
Othello it is enough to use n-tuples of size just 2, yielding a
network consisting of only 288 weights. The best such network
evolved in this study has been evaluated in the online Othello
League, obtaining the performance of nearly 96% — more than
any other player to date.
Keywords: Othello, Reversi, evolution strategy, n-tuple networks, Othello League, tabular value functions, strategy representation
I. I NTRODUCTION
Board games have always attracted attention in AI due
to they clear rules, mathematical elegance and simplicity.
Since the early works of Claude Shannon on Chess [1]
and Arthur Samuel on Checkers [2], a lot of research have
been conducted in the area of board games towards finding
either perfect players (Connect-4, [3]), or stronger than human
players (Othello, [4]). The bottom line is that board games still
constitute valuable test-beds for improving general artificial
and computational intelligence game playing methods such as
reinforcement learning, Monte Carlo tree search, branch and
bound, and (co)evolutionary algorithms.
Most of these techniques employ a position evaluation function to quantify the value of a given game state. In the context
of Othello, one of the most successful position evaluation
functions is tabular value function [5] or n-tuple network [6]. It
consists of a number of n-tuples, each associated with a look
up table, which maps contents of n board fields into a real
value. The effectiveness of n-tuple network highly depends on
the placement of n-tuples [7]. Typically, n-tuples architectures
consist of a small number of long, randomly generated, snakeshaped n-tuples [8], [7], [9].
Despite the importance of network architecture, to the best
of our knowledge no study exist that studies and evaluates
different ways of placing n-tuples on the board.
In this paper, we propose an n-tuple network architecture
consisting of a large number of short, straight n-tuples, gen-
a
b
c
d
e
f
g
h
1
2
3
4
5
6
7
8
Figure 1: An Othello position, where white has 6 legal moves
(dashed gray circles). If white places a piece on e3, the pieces
on d3, d4, and e4 are reversed to white.
erated in a systematic way. In the extensive computational
experiments, we show that for learning position evaluation for
Othello, such an architecture is significantly more effective
than the one involving randomly generated n-tuples. We also
investigate how the length of n-tuples affects the learning
results. Finally, the performance of the best evolved n-tuple
network is evaluated in the online Othello League.
II. M ETHODS
A. Othello
Othello (a.k.a. Reversi) is a two player, deterministic, perfect information strategic game played on an 8 × 8 board.
There are 64 pieces being black on one side and white on
the other. The game starts with two white and two black
pieces forming an askew cross in the center on the board.
The players take turns putting one piece on the board with
their color facing up. A legal move consists in placing a piece
on a field so that it forms a vertical, horizontal, or diagonal
line with another player’s piece, with a continuous, non-empty
sequence of opponent’s pieces in between (see Fig. 1), which
are reversed after the piece is placed. Player passes if and only
if it cannot make a legal move. The game ends when both
players passed consecutively. Then, the player having more
pieces with their color facing up wins.
Table I: The weights of the Standard WPC Heuristic player
(SWH)
1
−0.25
0.1
0.05
0.05
0.1
−0.25
1
−0.25
−0.25
0.01
0.01
0.01
0.01
−0.25
−0.25
0.1
0.01
0.05
0.02
0.02
0.05
0.01
0.1
0.05
0.01
0.02
0.01
0.01
0.02
0.01
0.05
0.05
0.01
0.02
0.01
0.01
0.02
0.01
0.05
0.1
0.01
0.05
0.02
0.02
0.05
0.01
0.1
−0.25
1
−0.25 −0.25
0.01 0.1
0.01 0.05
0.01 0.05
0.01 0.1
−0.25 −0.25
−0.25 1
a
B. Position Evaluation Functions
In this paper, our goal is not to design state-of-the-art
Othello players, but to evaluate position evaluation functions.
That is why our players are simple state evaluators in a 1-ply
setup: given the current state of the board, a player generates
all legal moves and applies the position evaluation function
to the resulting states. The state gauged as the most desirable
determines the move to be played. Ties are resolved at random.
The simplest position evaluation function is positionweighted piece counter (WPC), which is a linear weighted
board function. It assigns a weight wrc to a board location
(r, c) and uses scalar product to calculate the utility f of
a board state b = (brc )r,c=1...8 :
f (b) =
8 X
8
X
wrc brc ,
r=1 c=1
where bij is 0 in the case of an empty location, +1 if a black
piece is present or −1 in the case of a white piece.
A WPC player often used in Othello research as an expert
opponent [18], [6], [19], [13], [16], [7] is Standard WPC
Heuristic Player (SWH). Its weights, hand-crafted by Yoshioka
et al. [20], are presented in Table I.
C. Othello Position Evaluation Function League
WPC is only one of the possible position evaluation functions. Others popular ones include neural networks and n-tuple
networks. To allow direct comparison between various position
evaluation functions and algorithms capable of learning their
parameters, Lucas and Runarsson [18] have appointed the Othello Position Evaluation Function League 1 . Othello League,
for short, is an on-line ranking of Othello 1-ply state evaluator
players. The players submitted to the league are evaluated
against SWH (the Standard WPC Heuristic Player).
Both the game itself and the players are deterministic (with
an exception of the rare situation when at least two positions
have the same evaluation value). Therefore, to provide more
continuous performance measure, Othello League introduces
some randomization to Othello. Both players are forced to
make random moves with the probability of = 0.1. As
1 http://algoval.essex.ac.uk:8080/othello/League.jsp
c
d
3
1
2
3
4
5
Othello has been found to have around 1028 legal positions
[10] and has is not been solved; this is one reason why it has
become such a popular domain for computational intelligence
methods [11], [12], [13], [14], [15], [16], [7], [17].
b
6
3
e
g
f
2
2
1
1
0
0
0123 weight
h
3
3
2
1
0
0
1
2
2
1
0
0
1
2
3
7
3
8
0
0
1
1
2
2
3
3
0000
0001
0010
..
.
0111
..
.
1111
..
.
2211
..
.
2222
3.04
−3.90
−2.14
..
.
5.89
..
.
−1.01
..
.
−9.18
..
.
3.14
Figure 2: An 4-tuple employed eight times to take advantage of
board symmetries (symmetric sampling). The eight symmetric
expansions of the 4-tuple return, in total, 5×−1.01+2×5.89−
9.18 = −2.45 for the given board state.
a consequence the players no longer play (deterministic)
Othello, but stochastic -Othello. However, it was argued that
the ability to play -Othello is highly correlated with the ability
to play Othello [18].
The performance in Othello League is determined by the
number of wins against SWH player in -Othello in 50 double
games, each consisting of two single games played once white
and once black. To aggregate the performance into a scalar
value, we assume that a win counts as 1 point, while a draw
0.5 points. The average score obtained in this way against
SWH constitutes the Othello League performance measure,
which we incorporate in this paper.
D. N-tuple Network
The best performing evaluation function in the Othello
League is n-tuple network [16]. N-tuple networks have been
first applied to optical character recognition problem by Bledsoe and Browning [21]. For games, it have been used first
by Buro under the name of tabular value functions [5], and
later popularized by Lucas [6]. According to Szubert et al.
their main advantages of n-tuple networks “include conceptual
simplicity, speed of operation, and capability of realizing
nonlinear mappings to spaces of higher dimensionality” [7].
N-tuple network consists of m ni -tuples, where ni is tuple’s
size. For a given board position b, it returns the sum of
values returned by the individual n-tuples. The ith ni -tuple,
for i = 1 . . . m, consists of a predetermined sequence of board
locations (locij )j=1...ni , and a look up table LUTi . The latter
contains values for each board pattern that can be observed
on the sequence of board locations. Thus, n-tuple network is
a function
f (b) =
m
X
i=1
fi (b) =
m
X
i=1
LUTi index bloci1 , . . . , blocini .
a
b
c
d
e
f
g
h
a
1
1
2
2
3
3
4
4
5
5
6
6
7
7
8
8
(a) rand-8×4 network consisting of 8 randomly generated
snake-shaped 4-tuples (648 weights).
b
c
d
e
f
g
h
(b) all-3 network consisting of all 24 straight 3-tuples (648
weights).
Figure 3: Comparison of rand-* and all-* n-tuple network architectures. “Main” n-tuples have been shown by red, while their
symmetric expansions by light gray.
Among possible ways to map the sequence to an index in the
look up table, the following one is arguably convenient and
computationally efficient:
index(v) =
|v|
X
vj cj−1 ,
j=1
where c is a constant denoting the number of possible values
on a single board square, and v is the sequence of board values
(the observed pattern) such that 0 ≤ vj < c for j = 1 . . . |v|.
In the case of Othello, c = 3, and white, empty, and black
squares are encoded as 0, 1, and 2, respectively.
The effectiveness of n-tuple networks is improved by using
symmetric sampling, which exploits the inherent symmetries of
the Othello board [11]. In symmetric sampling, a single n-tuple
is employed eight times, returning one value for each possible
board rotation and reflection. See Fig. 2 for an illustration.
E. N-tuple Network Architecture
Due to the spatial nature of game boards, n-tuples are usually consecutive snake-shaped sequences of locations, although
this is not a formal requirement. If each n-tuple in a network is
of the same size, we denote it as m × n-tuple network, having
m × 3n weights. Apart from choosing n and m, an important
design issue of n-tuples network architecture is the location of
individual n-tuples on the board [7].
1) Random Snake-shaped N-tuple Network: Thus it is surprising that so many investigations in game strategy learning have involved randomly generated snake-shaped n-tuple
networks. Lucas [6] generated individual n-tuples by starting
from a random board location, then taking a random walk of 6
steps in any of the eight orthogonal or diagonal directions. The
repeated locations were ignored, thus the resulting n-tuples
were from 2 to 6 squares long. The same method Krawiec
and Szubert used for generating 7 × 4, 9 × 5 and 12 × 6-tuple
networks [22], [7], and Thill et al. [23] for generating 70 × 8
tuple networks playing Connect Four.
An m × n-tuple network generated in this way we will
denote as rand-m × n (see Fig. 3a for an example).
2) Systematic Straight N-tuple Network: Alternatively, we
propose a deterministic method of constructing n-tuple networks. Our systematic straight n-tuple networks consist of all
possible vertical, horizontal or diagonal n-tuples placed on
the board. Its smallest representative is a network of 1-tuples.
Thanks to symmetric sampling, only 10 of them is required
for an 8 × 8 Othello board, and such 10 × 1-tuple network,
which we denote as all-1 contains 3 × 101 = 30 weights. all2 network containing 32 2-tuples is shown in Fig. 3b. Table
II shows the number of weights in selected architectures of
rand-* and all-* networks.
3) Other Approaches: Logistello [4], computer player,
which beat the human Othello world champion in 1997, used
11 n-tuples of n ∈ {3, 10}, hand-crafted by an expert. External
knowledge has also been used by Manning [8], who, generated
a diverse 12 × 6-tuple network using random inputs method
from Breiman’s Random Forests basing on a set of 10 000
labeled random games.
F. Learning to Play Both Sides
When a single player defined by its evaluation function is
meant to play both as black and white, it must interpret the
result of the evaluation function complementary depending
on the color it plays. There are three methods serving this
purpose.
The first one is doubled function (e.g., [23]), which simply
employs two separate functions: one for playing white and
the other for playing black. It allows to fully separate the
strategy for white and black players. However, its disadvantage
consists in that two times more weights must be learned, and
all-2 (32 × 2)
all-3 (24 × 3)
all-4 (21 × 4)
weights
architecture
weights
288
648
1701
rand-10 × 3
rand-8 × 4
rand-7 × 5
270
648
1701
Table II: The number of weights for three pairs of systematic
straight (all-*) and random snake-shaped (rand-*) n-tuple
networks architectures.
the experience learned when playing as black does not used
when playing as white and vice versa.
Output negation and board inversion (e.g., [9]) are alternatives to doubled function. They use only single set of
weights, reducing the search space and allowing to transfer the
experience between the white and black player. When using
output negation, black selects the move leading to a position
with the maximal value of the evaluation function whereas
white selects the move leading to a position with the minimal
value.
If a player uses board inversion it learns only to play black.
As the best black move it selects the one leading to the position
with the maximum value. If it has to play white, it temporarily
flips all the pieces on the board, so it can interpret the board as
if it played black. Then it selects the best ‘black’ move, flips
all the pieces back, and plays the white piece in the selected
location.
The SWH player uses output negation.
III. E XPERIMENTS AND R ESULTS
A. Common Settings
1) Evolutionary Setup: In order to compare different ntuple network architectures, we performed several computational experiments. In each of them the weights of n-tuple
networks have been learned by (10 + 90) evolution strategy
[24] for 5000 generations. The weights of individuals in the
initial population were drawn from the [−0.1, 0.1] interval.
Evolution strategy used Gaussian mutation with σ = 1.0. The
individual’s fitness was calculated using the Othello League
performance measure estimated over 1000 double games (cf.
II-C).
In total, 1010 games were played in each evolutionary run.
This makes our experiments exceptionally large compared to
the previous studies. For example, in a recent study concerning
n-tuple networks [7] 3×106 games were played. Also, despite
using the much simpler WPC representation, Samothrakis et
al. [16] performed 108 games per run.
Such extensive experiment was possible due to efficient
n-tuple network and Othello implementation in Java, which
is capable of running about 1000 games per second on a
single CPU core. Thanks to it, we were able to finish one
evolutionary run in 28 hours on a 6-core Intel(R) Core(TM)
i7-2600 CPU @3.40GHz.
2) Performance Evaluation: We repeated each evolutionary
10 times. Every 10 generations, we measured the (Othello
League) performance of the fittest individual in the population
0.8
performance
architecture
0.7
0.6
v
-in
4
x
8
nd-
ra
ran
g
-ne
4
x
d-8
nv
1-i
l
l
a
g
-ne
1
all-
Figure 4: Comparison of output negation against board inversion for two n-tuples architectures. The performance is measured as the average score obtained against the Standard WPC
Heuristic Player at = 0.1 (Othello League performance). In
each violin shape, the white dot marks the median, the black
boxes range from the lower to the upper quartile, while the thin
black lines represent 1.5 interquartile range. Outliers beyond
this range are denoted by black dots. The outer shape shows
the probability density of the data.
using 50 000 double games. The performance of the fittest
individual from the last generation is identified with method’s
performance. Since, the sample size is only 10 per method, for
statistical analysis of the following experiments, we used nonparametric Wilcoxon rank sum test (a.k.a. the Mann-Whitney
U test) with the significance level α = 0.01 and Holm’s
correction when comparing more than two methods at once.
B. Preliminary: Board Inversion vs. Output Negation
Figure 4 presents the results of learning with board inversion
against output negation for representatives of both types of ntuple networks architectures: rand-8 × 4 having 8 × 43 = 648,
and all-1 with 10 × 31 = 30 weights.
The figure shows that board inversion surpasses output
negation regardless of the player architecture, which confirms
a previous study of the two methods for preference learning
[9]. The differences between the methods are statistically
significant (see also the detailed results in Table IV).
Moreover, visual inspection of the violin plots reveals that
board inversion leads to more robust learning, since the
variance of performances is lower. Therefore, in the following
experiments we employ exclusively board inversion.
C. All Short Straight vs. Random Long Snake-shaped N-tuples
In the main experiment, we compare n-tuple networks
consisting of all possible short straight n-tuples (all-2, all-3,
and all-4) with long random snake-shaped ones (rand-10 × 3,
rand-8 × 4 and rand-7 × 5). We chosen the number of ntuples and size of them to make the number of weights in of
performance
0.95
0.90
0.85
0.80
0.75
nv
nv
nv
nv
nv
nv
2-i
3-i
4-i
4-i
5-i
3-i
x
x
x
l
l
l
l
l
l
8
7
0
a
a
a
ddd-1
ran
ran
ran
Figure 5: The comparison of all short straight n-tuple networks
(all-*) with random long snake-shaped n-tuple networks (rand*). The distribution of performances is presented as violin plots
(see Fig. 4 for explanation).
corresponding architectures are equal, or, if impossible, similar
(see Table II).
The results of the experiment are shown in Figure 5 as violin
plots. Statistical analysis of three pairs having equal or similar
number of weights reveals that:
• all-2 is better than rand-10 × 3,
• all-3 is better than rand-8 × 4, and
• all-4 is better than rand-7 × 5.
Let us notice that the differences in performance are substantial: for the pair all-2 vs. rand-10 × 3, where the difference in
performance is the lowest, the best result obtained by rand10 × 3 is still lower than the worst result obtained by all-2
(see Table IV for details).
All-* architectures are also more robust, due to lower
variances than rand-* architectures (cf. Fig. 5). This is because
the variance of rand-* architectures is attributed to both its
random initialization and non-deterministic learning process,
while the variance of all-* is only due to the latter.
D. 2-tuples are Long Enough
Intuitively, longer n-tuples should lead to higher network’s
performance, since they can ‘react’ to patterns that the shorter
ones cannot. However, the results presented in Fig. 5 show no
evidence that this is a case. Despite having two times more
weights, all-3 does not provide better performance than all-2
(no statistical difference). Furthermore, all-4 is significantly
worse than both than all-2 and all-3.
Figure 6 shows the pace of learning for each of six analyzed
architectures. It plots methods’ performance as a function of
computational effort, which is proportional to the number of
generations.
The figure suggests that all-2 is not only the best (together
with all-3) in the long run, but it is also the method that learns
the quickest. all-3 catches up all-2 eventually, but it does not
date
player name
encoding
n/a
2013-09-17
2011-01-30
2011-01-28
2011-01-25
2008-05-03
2008-05-03
2008-03-28
2007-09-14
all-2-inv
wj-1-2-3-tuples
epTDLmpx_12x6 [7]
prb_nt15_001
epTDLxover [7]
t15x6x8
x30x6x8
Stunner
MLP(...)312-ties0.FF
n-tuple network
288
n-tuple network
966
n-tuple network
3240
n-tuple network
6561
n-tuple network
4698
n-tuple network 10 935
n-tuple network 21 870
n-tuple network
7725
neural network
1915
weights performance
0.9592
0.9149
0.871
0.845
0.83
0.805
0.73
0.675
0.62
Table III: Selected milestones (improvements) in the on-line
Othello Position Evaluation Function League since September
2007. The table consists also all-2, not submitted to the
League since it uses board inversion. The performances of
all but the three best players come from the Othello League
website and have been estimated using 50 double games. The
performances of all-2 and wj-1-2-3-tuples players have been
estimated using 50 000 double games, and the performance of
epTDLmpx_12x6 has been reported in [7].
seem to be able to surpass it. all-4 learns even slower than
all-3. Although the gap between all-3 and all-4 decreases over
time, it is still noticeable after 5000 generations.
Thus, our results suggest that for Othello, all-2 with just
288 weights, the smallest among the six considered n-tuple
network architectures, is also the best one.
E. Othello League Results
The best player obtained in this research consists of all 2tuples; its performance is 0.9592 with 95% confidence delta
of ±0.0012. This result is significantly higher than the best
results reported to this date in the Othello League (see III).
Notice also how small it is (in terms of the number of weights)
compared to other players in the league. Unfortunately, the online Othello League accepts only players employing output
negation; it does not allow for board inversion. Thus, our
player could not be submitted to the Othello League.
To be accepted in the Othello League, we performed some
experiments also with output negation. The best output negation player we were able to evolve was submitted under the
name of wj-1-2-3-tuples. It consists of all straight 1-, 2-, and
3-tuples, thus having 966 weights in total.
wj-1-2-3-tuples took the lead in the league and is the first
player exceeding the performance of 0.9. It obtained 0.94 in
the league, but this result should be taken with care, since to
evaluate player’s performance Othello League plays just 100
games. We estimate its performance to 0.9149±0.0017 basing
on 50 000 double games.
We suspect that the performance of ca. 0.96 against Standard WPC Heuristic player that all-2 and all-3 converge to,
cannot be significantly improved at 1-ply. = 0.1 random
moves using in all games leads to the situation when even a
perfect-playing player cannot guarantee not losing a game.
1.00
0
1000
2000
generation
3000
4000
5000
performance
0.90
0.80
0.70
all-2-inv
all-3-inv
all-4-inv
0.60
0.50
0
200
400
600
games played
rand-10x3-inv
rand-8x4-inv
rand-7x5-inv
800
1000
(×106 )
Figure 6: Pace of learning of six analyzed n-tuple networks architectures. Each point on the plot denotes the average performance
of method’s fittest individual in a given generation.
Despite the first place obtained in the Othello League, the
evolved player is not good in ‘general’, against a variety of
opponents, because is was evolved specifically to play against
the Standard WPC Heuristic player. When evaluated against
random WPC players (the expected utility measure [14], [25]),
the best all-2 player obtains a score of only 0.9584 ± 0.0012.
This is not much, since with considerably less computational
effort that used in this paper, it is possible to evolve an n-tuple
player scoring > 0.99 [26], [7]. However, our goal here was
not to design good players in general, but to compare different
position evaluation functions.
The best all-2 player evolved in this paper is printed in
Fig. 7.
IV. D ISCUSSION : THE MORE WEIGHTS , THE WORSE FOR
EVOLUTION ?
We have shown that among all-* methods, the more weights
the worse results; the same applies to rand-* methods (see
Fig. 5). This finding confirms the one of Szubert et al. [7],
who found out that among the networks of rand-12 × 6 (8748
weights), rand-9 × 5 (2187 weights), and rand-7 × 4 (567
weights), it is the latter that allows (co)evolutionary algorithm
for obtaining best results. The authors stated that this effect it
due to the higher dimensionality of the search space, for which
“the weight mutation operator is not sufficiently efficient to
elaborate fast progress”.
Although we do not challenge this claim, our results suggest
that the number of weights in a network is not the only performance factor. all-4 has 1701 weights, thus, the dimensionality
of its search space is considerably higher than the one for
rand-10 × 3 and rand-8 × 4, which have 270 and 648 weights,
respectively. Nonetheless, among these three architectures, it
is the all-4 network that obtains the highest performance (see
Fig. 5). Therefore, the second performance factor in learning
an n-tuple network is its (proper or not) architecture.
Finally, let us notice that an alternative to a fixed n-tuple
network architecture is a self-adaptive one, which can change
in response to variation operators [7], such as mutation or
crossover. Although such architecture is, in principle, more
flexible, it adds another dimension to the search space, making
the learning problem even harder.
V. C ONCLUSIONS
In this paper, we have analyzed n-tuple network architectures for position evaluation function in board games. We have
shown that a network consisting of all possible, systematically
generated, short n-tuples leads to a significantly better play
than long random snake-shaped tuples originally used by
Lucas [11]. With a simple network consisting of all possible
straight 2-tuples (with just 288 weights) we were able to beat
the best result in the on-line Othello League (having usually
many times more weights).
Moreover, our results suggest that tuples longer than 2 give
no advantage, causing slower learning rate, at the same time.
This is surprising, since capturing opponent’s pieces in Othello
requires a line of at least three pieces (e.g. white, black, white).
Let us emphasize that our result has been obtained in an
intensive computational experiment involving 5000 generations, an order of magnitude more than other studies in this
domain. Nevertheless, it remains to be seen whether they hold
for different experimental settings. We used evolution against
an expert player in 1-ply -Othello. The interesting questions
are: i) whether our systematic short 2-tuple network is also
advantageous for reinforcement learning, such as temporal
difference learning, and ii) whether such networks are also
profitable for other board games, e.g. Connect Four.
ACKNOWLEDGMENT
This work has been supported by the Polish National
Science Centre grant no. DEC-2013/09/D/ST6/03932. The
{ 32
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
{ 2
{
}
8 { 6 7 } { 55 63 } { 7 15 } { 56 57 } { 62 63 } { 0 1 } { 48 56 } { 0 8 }
5 7 . 6 4 −91.70 1 1 1 . 0 5 −82.30 7 4 . 4 2 −96.30 2 1 1 . 4 7 −53.91 1 4 2 . 4 5 } }
4 { 49 56 } { 0 9 } { 7 14 } { 54 63 }
8 4 . 5 1 −33.37 −29.83 −72.21 −199.31 −18.72 −98.04 2 2 . 3 4 1 8 5 . 8 4 } }
8 { 1 2 } { 15 23 } { 8 16 } { 40 48 } { 57 58 } { 5 6 } { 47 55 } { 61 62 }
−42.97 1 0 9 . 7 6 −66.10 8 4 . 6 7 1 5 8 . 0 9 −148.21 −11.94 −94.92 1 1 1 . 5 2 } }
8 { 48 49 } { 54 55 } { 8 9 } { 54 62 } { 6 14 } { 14 15 } { 49 57 } { 1 9 }
−144.66 1 6 . 1 6 −78.97 −83.93 −7.37 −15.97 −102.98 −51.68 −3.99 } }
8 { 41 48 } { 8 17 } { 6 13 } { 50 57 } { 53 62 } { 1 10 } { 46 55 } { 15 22 }
−115.74 −43.49 −117.28 3 3 . 3 2 −14.29 1 2 . 8 5 −18.33 2 . 9 5 9 1 . 1 7 } }
8 { 4 5 } { 16 24 } { 39 47 } { 58 59 } { 60 61 } { 32 40 } { 2 3 } { 23 31 }
3 1 . 2 8 −32.65 3 7 . 4 6 −20.60 2 7 3 . 3 9 −90.25 6 0 . 1 6 −151.55 −30.45 } }
8 { 2 10 } { 16 17 } { 46 47 } { 22 23 } { 50 58 } { 40 41 } { 5 13 } { 53 61 }
9 3 . 1 8 −27.12 −16.53 −2.52 −80.43 6 2 . 2 3 3 6 . 4 6 3 9 . 8 6 −97.27 } }
8 { 52 61 } { 23 30 } { 51 58 } { 16 25 } { 5 12 } { 2 11 } { 38 47 } { 33 40 }
−42.75 3 6 . 9 1 4 4 . 6 5 −49.19 −19.01 −55.54 1 0 . 6 0 −36.50 7 4 . 3 7 } }
4 { 24 32 } { 59 60 } { 3 4 } { 31 39 }
4 3 . 7 7 −122.45 5 5 . 4 3 −5.38 2 8 4 . 3 9 −103.51 7 9 . 4 4 −94.92 1 5 7 . 0 2 } }
8 { 52 60 } { 24 25 } { 51 59 } { 32 33 } { 4 12 } { 30 31 } { 38 39 } { 3 11 }
9 9 . 4 5 2 0 . 8 9 −10.97 1 5 . 3 9 −28.77 1 6 . 6 5 −16.89 −12.47 −18.57 } }
8 { 30 39 } { 25 32 } { 3 12 } { 51 60 } { 24 33 } { 31 38 } { 52 59 } { 4 11 }
1 4 . 8 5 8 8 . 0 4 −3.28 2 6 . 8 8 3 3 . 4 6 −19.67 −5.65 2 8 . 8 5 −43.03 } }
8 { 39 46 } { 53 60 } { 4 13 } { 17 24 } { 50 59 } { 32 41 } { 3 10 } { 22 31 }
−32.77 2 2 . 2 7 −51.36 −2.01 −65.96 −33.23 1 6 . 3 9 8 . 5 9 −28.07 } }
8 { 47 54 } { 5 14 } { 40 49 } { 49 58 } { 2 9 } { 14 23 } { 9 16 } { 54 61 }
−82.49 −5.56 1 0 . 1 9 −68.91 2 9 . 8 4 −37.59 −69.56 −0.20 1 0 . 6 3 } }
4 { 6 15 } { 48 57 } { 55 62 } { 1 8 }
−34.15 2 0 . 8 9 −135.36 7 9 . 2 2 3 0 . 5 5 2 0 . 1 3 3 5 . 2 7 −11.57 1 6 . 4 3 } }
8 { 13 14 } { 9 10 } { 41 49 } { 9 17 } { 53 54 } { 46 54 } { 49 50 } { 14 22 }
2 0 . 9 8 −44.43 3 0 . 9 4 −64.79 −27.71 −37.59 1 7 . 0 5 4 . 3 4 −17.03 } }
4 { 45 54 } { 9 18 } { 14 21 } { 42 49 }
4 3 . 0 8 1 0 4 . 4 3 1 4 . 6 9 2 4 . 4 9 3 1 . 9 6 −14.64 −51.11 −22.12 1 4 . 4 8 } }
8 { 38 46 } { 12 13 } { 52 53 } { 22 30 } { 50 51 } { 10 11 } { 33 41 } { 17 25
1 9 . 7 5 3 9 . 6 3 −16.15 1 . 7 5 −38.84 9 . 2 1 6 . 7 7 1 4 . 8 5 1 9 . 9 9 } }
8 { 41 42 } { 13 21 } { 45 46 } { 45 53 } { 10 18 } { 21 22 } { 17 18 } { 42 50
5 . 2 8 −38.02 1 2 . 6 4 −90.60 6 0 . 6 0 5 . 9 6 6 0 . 3 8 2 7 . 6 1 3 . 0 0 } }
8 { 37 46 } { 10 19 } { 17 26 } { 43 50 } { 22 29 } { 44 53 } { 13 20 } { 34 41
−13.44 1 9 . 4 8 −13.49 0 . 7 2 −59.65 −3.23 4 5 . 2 7 4 5 . 3 1 3 0 . 3 9 } }
4 { 25 33 } { 30 38 } { 11 12 } { 51 52 }
8 . 6 6 1 4 . 8 3 1 5 . 7 3 −34.15 3 2 . 0 8 −9.15 1 5 . 1 5 4 1 . 6 1 6 6 . 0 3 } }
8 { 25 26 } { 12 20 } { 11 19 } { 33 34 } { 44 52 } { 43 51 } { 37 38 } { 29 30
−71.70 1 2 . 0 7 −54.50 1 8 . 1 2 8 6 . 3 6 2 2 . 2 7 −56.07 −4.46 −43.54 } }
8 { 30 37 } { 43 52 } { 29 38 } { 26 33 } { 44 51 } { 25 34 } { 11 20 } { 12 19
1 1 . 3 4 2 5 . 6 4 −28.34 4 1 . 8 2 7 3 . 3 3 −26.18 −0.64 −25.88 −29.12 } }
8 { 42 51 } { 21 30 } { 45 52 } { 11 18 } { 12 21 } { 18 25 } { 38 45 } { 33 42
−32.37 2 8 . 6 0 1 3 . 6 5 −48.41 −13.25 −63.15 −30.60 1 8 . 9 9 2 2 . 6 4 } }
4 { 10 17 } { 13 22 } { 46 53 } { 41 50 }
3 1 . 8 9 −78.88 −32.75 4 4 . 8 8 −42.65 3 9 . 9 1 2 6 . 4 8 −12.34 −46.59 } }
8 { 34 42 } { 21 29 } { 20 21 } { 18 26 } { 18 19 } { 42 43 } { 37 45 } { 44 45
0 . 6 6 −29.13 1 2 . 9 5 −17.71 −71.59 1 1 . 4 0 3 1 . 5 2 −4.66 7 9 . 8 9 } }
4 { 35 42 } { 18 27 } { 21 28 } { 36 45 }
7 6 . 5 3 1 4 1 . 0 5 −5.04 6 1 . 8 9 1 6 . 1 3 4 3 . 9 5 −4.87 1 8 2 . 8 5 −92.46 } }
4 { 26 34 } { 43 44 } { 29 37 } { 19 20 }
−10.60 1 1 . 6 0 −8.56 −6.06 −137.99 −8.81 −3.62 3 . 3 0 −39.96 } }
8 { 20 28 } { 34 35 } { 28 29 } { 36 37 } { 26 27 } { 35 43 } { 36 44 } { 19 27
3 8 . 0 3 5 6 . 9 0 −26.64 −61.84 2 1 . 6 9 −116.98 7 . 9 1 5 3 . 4 8 −58.83 } }
8 { 19 28 } { 36 43 } { 29 36 } { 20 27 } { 26 35 } { 27 34 } { 35 44 } { 28 37
−17.67 8 1 . 7 8 −63.71 1 5 . 6 2 2 . 1 6 1 6 . 6 3 −14.70 4 3 . 6 1 −24.67 } }
4 { 34 43 } { 37 44 } { 20 29 } { 19 26 }
−18.74 −52.99 −3.10 −31.68 −181.22 −24.62 3 2 . 6 3 5 8 . 1 9 4 0 . 7 9 } }
4 { 35 36 } { 27 28 } { 28 36 } { 27 35 }
−100.00 −28.48 −4.99 −63.27 −80.05 −55.95 2 2 . 6 3 2 2 . 5 7 1 3 3 . 6 5 } }
2 { 28 35 } { 27 36 }
2 0 . 0 0 −52.52 −12.54 4 0 . 6 4 1 9 2 . 6 1 1 4 . 1 4 4 . 7 0 −1.81 −42.70 } }
Figure 7: N-tuple network representing the best evolved all2 player in the online Othello League format. The player
contains 32 2-tuples. Each one has at most 8 symmetric
expansions (sometimes 4 or 2). The fields are numbered from
0 to 63 in row-wise fashion. The network has 32 × 9 = 288
weights. The player uses board inversion. Its Othello League
performance is 0.9584 ± 0.0012.
computations have been performed in Poznań Supercomputing
and Networking Center. The author would like to thank Marcin
Szubert for his helpful remarks on an earlier version of this
article.
R EFERENCES
[1] C. E. Shannon, “XXII. Programming a computer for playing chess,”
Philosophical magazine, vol. 41, no. 314, pp. 256–275, 1950.
[2] A. L. Samuel, “Some studies in machine learning using the game of
checkers,” IBM Journal of Research and Development, vol. 3, no. 3, pp.
211–229, 1959.
[3] L. V. Allis, A knowledge-based approach of connect-four.
Vrije
Universiteit, Subfaculteit Wiskunde en Informatica, 1988.
[4] M. Buro, “Experiments with Multi-ProbCut and a new high-quality
evaluation function for Othello,” Games in AI Research, pp. 77–96, 2000.
}
}
}
}
}
}
}
}
}
[5] ——, “An evaluation function for othello based on statistics,” NEC,
Princeton, NJ, NECI 31, Tech. Rep., 1997
[6] S. Lucas, “Learning to play Othello with n-tuple systems,” Australian
Journal of Intelligent Information . . . , no. 4, pp. 1–20, 2008
[7] M. Szubert, W. Jaśkowski, and K. Krawiec, “On Scalability, Generalization, and Hybridization of Coevolutionary Learning: a Case Study for
Othello,” IEEE Transactions on Computational Intelligence and AI in
Games, 2013.
[8] E. P. Manning and A. Othello, “Using Resource-Limited Nash Memory
to Improve an Othello Evaluation Function,” IEEE Transactions on
Computational Intelligence and AI in Games, vol. 2, no. 1, pp. 40–53,
2010.
[9] T. Runarsson and S. Lucas, “Preference Learning for Move Prediction
and Evaluation Function Approximation in Othello,” Computational
Intelligence and AI in Games, IEEE Transactions on, 2014.
[10] V. L. Allis, “Searching for solutions in games and artificial intelligence,”
Ph.D. dissertation, University of Limburg, Maastricht, The Netherlands,
1994.
[11] S. M. Lucas, “Learning to play Othello with N-tuple systems,” Australian Journal of Intelligent Information Processing Systems, Special
Issue on Game Technology, vol. 9, no. 4, pp. 01–20, 2007.
[12] Y. Osaki, K. Shibahara, Y. Tajima, and Y. Kotani, “An Othello
evaluation function based on Temporal Difference Learning using
probability of winning,” 2008 IEEE Symposium On Computational
Intelligence and Games, pp. 205–211, Dec. 2008
[13] E. P. Manning, “Using resource-limited nash memory to improve an
othello evaluation function,” Computational Intelligence and AI in
Games, IEEE Transactions on, vol. 2, no. 1, pp. 40 –53, march 2010.
[14] S. Y. Chong, P. Tino, D. C. Ku, and Y. Xin, “Improving Generalization
Performance in Co-Evolutionary Learning,” IEEE Transactions on
Evolutionary Computation, vol. 16, no. 1, pp. 70–85, 2012
[15] S. van den Dries and M. A. Wiering, “Neural-Fitted TD-Leaf
Learning for Playing Othello With Structured Neural Networks,” IEEE
Transactions on Neural Networks and Learning Systems, vol. 23,
no. 11, pp. 1701–1713, Nov. 2012
[16] S. Samothrakis, S. Lucas, T. Runarsson, and D. Robles, “Coevolving
Game-Playing Agents: Measuring Performance and Intransitivities,”
IEEE Transactions on Evolutionary Computation, no. 99, pp. 1–15,
2012
[17] W. Jaśkowski, M. Szubert, and P. Liskowski, “Multi-criteria comparison of coevolution and temporal difference learning on othello,” in
EvoGames, ser. Lectures Notes in Computer Science, 2014.
[18] S. M. Lucas and T. P. Runarsson, “Temporal difference learning versus
co-evolution for acquiring othello position evaluation,” in IEEE Symposium on Computational Intelligence and Games. IEEE, 2006, pp.
52–59.
[19] M. Szubert, W. Jaśkowski, and K. Krawiec, “Coevolutionary Temporal
Difference Learning for Othello,” in IEEE Symposium on Computational
Intelligence and Games, 2009, Conference proceedings (article), pp.
104–111
[20] T. Yoshioka, S. Ishii, and M. Ito, “Strategy acquisition for the game,”
Strategy Acquisition for the Game "Othello" Based on Reinforcement
Learning, vol. 82, no. 12, pp. 1618–1626, 1999.
[21] W. W. Bledsoe and I. Browning, “Pattern recognition and reading by
machine,” in Proc. Eastern Joint Comput. Conf., 1959, pp. 225–232.
[22] K. Krawiec and M. Szubert, “Learning n-tuple networks for othello by
coevolutionary gradient search,” in GECCO 2011 Proceedings, N. K.
et al, Ed., ACM. ACM, 2011, pp. 355–362.
[23] M. Thill, P. Koch, and W. Konen, “Reinforcement Learning with Ntuples on the Game Connect-4,” in Parallel Problem Solving from Nature
- PPSN XII, ser. Lecture Notes in Computer Science, C. A. C. Coello,
V. Cutello, K. Deb, S. Forrest, G. Nicosia, and M. Pavone, Eds., vol.
7491. Springer, 2012, pp. 184–194.
[24] H.-G. Beyer and H.-P. Schwefel, “Evolution strategies–a comprehensive
introduction,” Natural computing, vol. 1, no. 1, pp. 3–52, 2002.
[25] W. Jaśkowski, P. Liskowski, M. Szubert, and K. Krawiec, “Improving
coevolution by random sampling,” in GECCO’13: Proceedings of the
15th annual conference on Genetic and Evolutionary Computation,
C. Blum, Ed. Amsterdam, The Netherlands: ACM, July 2013, pp.
1141–1148.
[26] P. Liskowski, “Co-Evolution Versus Evolution with Random Sampling
for Acquiring Othello Position Evaluation,” Ph.D. dissertation, Poznan
University of Technology, 2012.
all-2-inv
all-3-inv
all-4-inv
rand-10x3-inv
rand-8x4-inv
rand-7x5-inv
rand-8x4-neg
all-1-inv
all-1-neg
mean
median
0
1
2
3
4
5
6
7
8
9
0.9401
0.9379
0.9122
0.8681
0.8454
0.8012
0.7698
0.7362
0.5946
0.9361
0.9385
0.9101
0.8770
0.8530
0.7974
0.7741
0.7374
0.5790
0.9529
0.9326
0.9108
0.8649
0.8519
0.8285
0.8104
0.7337
0.5799
0.9330
0.9314
0.8967
0.8145
0.7927
0.8565
0.6959
0.7395
0.5584
0.9270
0.9402
0.9171
0.8834
0.8050
0.7978
0.8429
0.7423
0.5821
0.9578
0.9282
0.9076
0.8920
0.8595
0.8559
0.7133
0.7355
0.5650
0.9327
0.9382
0.9094
0.8126
0.8366
0.7398
0.7453
0.7434
0.6588
0.9349
0.9387
0.9230
0.8398
0.8630
0.7310
0.7817
0.7217
0.5780
0.9225
0.9404
0.9051
0.9190
0.8728
0.7803
0.8085
0.7317
0.6678
0.9372
0.9472
0.9030
0.8958
0.8733
0.7810
0.7517
0.7395
0.6097
0.9495
0.9452
0.9269
0.8706
0.8540
0.7970
0.7668
0.7361
0.5713
0.9535
0.9370
0.9227
0.8880
0.8453
0.8440
0.7813
0.7387
0.5752
Table IV: Performances obtained in 10 evolutionary runs of all n-tuple network architectures considered in this study. Each
value is an average score in 50 000 double games against Standard WPC Heuristic in -Othello, where = 0.1.
| 9cs.NE
|
Forest-based methods and ensemble model output statistics for
rainfall ensemble forecasting
Maxime Taillardat1,2,3 , Anne-Laure Fougères2 , Philippe Naveau3 , and Olivier Mestre1
1
arXiv:1711.10937v1 [stat.ML] 29 Nov 2017
2
CNRM UMR 3589 and DirOP/COMPAS, Météo-France/CNRS, Toulouse, France
Univ Lyon, Université Claude Bernard Lyon 1, CNRS UMR 5208, Institut Camille Jordan, 43
blvd. du 11 novembre 1918, F-69622 Villeurbanne cedex, France
3
Laboratoire des Sciences du Climat et de l’Environnement (LSCE-CNRS-CEA-UVSQ-IPSL),
Gif-sur-Yvette, France
[email protected]
Abstract
Rainfall ensemble forecasts have to be skillful for both low precipitation and extreme events. We
present statistical post-processing methods based on Quantile Regression Forests (QRF) and Gradient
Forests (GF) with a parametric extension for heavy-tailed distributions. Our goal is to improve
ensemble quality for all types of precipitation events, heavy-tailed included, subject to a good overall
performance.
Our hybrid proposed methods are applied to daily 51-h forecasts of 6-h accumulated precipitation
from 2012 to 2015 over France using the Météo-France ensemble prediction system called PEARP.
They provide calibrated predictive distributions and compete favourably with state-of-the-art methods like Analogs method or Ensemble Model Output Statistics. In particular, hybrid forest-based
procedures appear to bring an added value to the forecast of heavy rainfall.
1
1.1
Introduction
Post-processing of ensemble forecasts
Accurately forecasting weather is paramount for a wide range of end-users, e.g. air traffic controllers,
emergency managers and energy providers (see, e.g. Pinson et al., 2007; Zamo et al., 2014). In meteorology, ensemble forecasts try to quantify forecast uncertainties due to observation errors and incomplete
physical representation of the atmosphere. Despite its recent developments in national meteorological
services, ensemble forecasts still suffer of bias and underdispersion (see, e.g. Hamill and Colucci, 1997).
Consequently, they need to be post-processed. At least two types of statistical methods have emerged in
the last decades: analogs method and ensemble model output statistics (EMOS) (see, e.g. Delle Monache
et al., 2013; Gneiting et al., 2005, respectively). The first one is fully non-parametric and consists in
finding similar atmospheric situations in the past and using them to improve the present forecast. In
contrast, EMOS belongs to the family of parametric regression schemes. If y represents the weather
variable of interest and (x1 , . . . , xm ) the corresponding m ensemble member forecasts, then the EMOS
predictive distribution is simply a distribution whose parameters depend on the values of (x1 , . . . , xm ).
Less conventional approaches have also been studied recently. For example, Van Schaeybroeck and Vannitsem (2015) investigated member-by-member post-processing techniques and Taillardat et al. (2016)
found that quantile regression forests (QRF) techniques performed well for temperatures and wind speed
data.
1
1.2
Forecasting and calibration of precipitation
Not all meteorological variables are equal in terms of forecast and calibration. In particular, Hemri et al.
(2014) highlighted that rainfall forecasting represents a steep hill. In this study, we will focus on 6-h
rainfall amounts in France because this is the unit of interest of the ensemble forecast system of MétéoFrance. For daily precipitation, extended logistic regression was frequently applied (see, e.g. Hamill
et al., 2008; Roulin and Vannitsem, 2012; Ben Bouallègue, 2013). Bayesian Model Averaging techniques
(Raftery et al., 2005; Sloughter et al., 2007) were also used in rainfall forecasting, but we will not cover
them here because a gamma fit is often applied to cube root transformed precipitation accumulations
and this complex transformation may not be adapted to 6h rainfall. Concerning analogs and EMOS
techniques, they have been applied to calibrate daily rainfall (see Hamill and Whitaker, 2006; Scheuerer,
2014; Scheuerer and Hamill, 2015). As the QRF method in Taillardat et al. (2016) performed better
than EMOS for temperatures and wind speeds, one may wonder if QRF could favourably compete with
EMOS and analogs techniques for rainfall calibration. This question is particularly relevant because
recent methodological advances have been made concerning random forests and quantile regressions. In
particular, Athey et al. (2016) proposed an innovative way, called gradient forests (GF), of using forests
to make quantile regression. In this context, we propose to implement and test this quantile regression
GF method for rainfall calibration and compare it with other approaches, see Section 2.
1.3
Parametric probability density functions (pdf ) of precipitation
Modeling precipitation distributions is a challenge by itself. It is a mixture of zeros (dry events) and
positive intensities, i.e. rainfall amounts for wet events. The latter have a skewed distribution. One
popular and flexible choice to model rainfall amounts is to use the gamma distribution or to built on
it. For example, Scheuerer and Hamill (2015) and Baran and Nemoda (2016) in a rainfall calibration
context employed the censored-shifted gamma (CSG) pdf defined by
(
κ−1
(1 − π) · (y+δ)
exp(−(y + δ)/θ), if y > 0
Γ(κ)
fCSG (y) =
(1)
π,
if y = 0,
where y ≥ 0, the positive constants (κ, θ) are the two gamma law parameters and the probability
π ∈ [0, 1] represents the mass of the gamma cumulative distribution function (cdf) below the level of
censoring δ ≥ 0. Hence, the probability of zero and positive precipitation are treated together. One
possible drawback of the CSG is that heavy daily and subdaily rainfall may not always have a nice upper
tail with an exponential decay like a gamma distribution, but rather a polynomial one, the latter point
being a key element in any weather risk analysis (see, e.g. Katz et al., 2002; De Haan and Ferreira, 2007).
To bring the necessary flexibility in modelling upper tail behavior in a rainfall EMOS context, Scheuerer
(2014) worked with a so-called censored generalized extreme value (CGEV) defined by
(1 − π) · g(y; µ, σ, ξ), if y > 0
fCGEV (y) =
(2)
π,
if y = 0,
where π = G(0; µ, σ, ξ) and the pdf g(y; µ, σ, ξ) which cumulative distribution function G is the classical
GEV
"
#
ξ(y − µ) −1/ξ
G(y; µ, σ, ξ) = exp − 1 +
for ξ 6= 0.
σ
+
Note that a+ = max(0, a) and that, if ξ = 0, then g(y; µ, σ, 0) represents the classical Gumbel pdf. To be
in compliance with extreme value theory (EVT) not only for heavy rainfall but also for low precipitation
amounts, Naveau et al. (2016) recently proposed a class of models referred as the extended generalized
Pareto (EGP) that allows a smooth transition between generalized Pareto (GP) type tails and the middle
part (bulk) of the distribution. It bypasses the complex thresholds selection step to define extremes.
Low precipitation can be shown to be gamma distributed, while heavy rainfall are Pareto distributed.
Mathematically, a cdf belonging to the EGP family has to be expressed as
T {Hξ (y/σ)} , for all y > 0,
2
where Hξ (y) = 1 − (1 + ξy)−1/ξ represents the GP cdf, while T denotes a continuous cdf on the unit
interval. To insure that the upper tail behavior of T is driven by the shape parameter ξ, the survival
function T̄ = 1 − T has to satisfy that lim T̄ (1−u)
is finite. To force low rainfall to follow a GPD for small
u
u↓0
values near zero, we need that
lim Tu(u)
s
u↓0
is finite for some real s > 0. Studies have already made this choice
(see, e.g. Vrac and Naveau, 2007; Naveau et al., 2016). In Naveau et al. (2016), different parametric
models of the cdf T satisfying the required constraints were compared. The special case where T (u) = uκ
with κ > 0 obeys these constraints and also corresponds to a model studied by Papastathopoulos and
Tawn (2013). In practice, this simple version of T appears to fit well daily and subdaily rainfall and
consequently, we will only focus on this case in this paper. In other words, our third model for the
precipitation pdf is
(1 − π) · σκ · {Hξ (x/σ)}κ−1 · hξ (y/σ), if y > 0
fEGP (y) =
(3)
π,
if y = 0,
where hξ (.) is the pdf associated with Hξ (.). In contrast to (1) and (2), the probability weight π is not
obtained by censoring, and it is just a parameter independent of (κ, σ, ξ)T .
At this stage, we have three parametric pdfs, see (1) and (2) and (3), to implement a EMOS approach
to 6-hour rainfall data, see Section 3. Besides comparing these three EMOS models, it is natural to
wonder if QRF and GF methods could take advantage of these three parametric forms.
1.4
Coupling parametric pdfs with random forest approaches
A drawback of data driven approaches like QRF and GF is that their intrinsic non parametric nature
make them useless to predict beyond the largest recorded rainfall. To circumvent this limit, we also
propose to combine random forest techniques with a EGP pdf defined by (3), see Section 2.3. Hence,
random forest-based post-processing techniques will be in compliance with EVT and this should be an
interesting path to improve prediction behind the largest values of the sample at hand.
1.5
Outline
This article is organized as follows. In Section 2, we recall the basic ingredients to create quantile
regression forests and gradient forests. In particular, we review the calibration process of the GF method
recently introduced by Athey et al. (2016) for quantile regression. Then, we explain how these trees are
combined with the EGP pdf defined by (3).
In Section 3, we propose to integrate the EGP pdf within a EMOS scheme.
The different approaches are implemented in Section 4 where the test bed dataset of 87 French
weather stations and the French ensemble forecast system of Météo-France called PEARP (Descamps
et al., 2014) is described. Then, we assess and compare each method with a special interest for heavy
rainfall, see Section 5. The paper closes with a discussion in Section 6.
2
2.1
Quantile regression forests and gradient forests
Quantile regression forests
Given a sample of predictors-response pairs, say (Xi , Yi ) for i = 1, . . . , n, classical regression techniques
connect the conditional mean of a response variable Y to a given set of predictors X. The quantile
regression forest (QRF) method introduced by Meinshausen (2006) also consists in building a link, but
between an empirical cdf and the outputs of a tree. Before explaining this particular cdf, we need to
recall how trees are constructed.
A random forest is an aggregation of randomized trees based on bootstrap aggregation on the one
hand, and on classification and regression trees (CART) (Breiman, 1996; Breiman et al., 1984) on the
other hand. These trees are built on a bootstrap copy of the samples by recursively maximizing a
3
splitting rule. Let D0 denote the group of observations to be divided into two subgroups, say D1 and
D2 . For each group, we can infer its homogeneity defined by
X
v(Dj ) =
[Y − Y (Dj )]2 ,
Y ∈Dj
where Y (Dj ) corresponds to the sample mean in Dj . To determine if this splitting choice is optimal,
the homogeneities v(D1 ) and v(D2 ) are compared to the one of D0 . For example, if wind speed is one
predictor in X and dividing low and large winds could better explain rainfall, then the cutting value,
say s, will be the one that maximizes
H(D1 , D2 ) = max∗ [v(D0 ) − v(D1 ) − v(D2 )]
s∈E
(4)
where E ∗ is a random subset of the predictors in the predictors’ space E. Each resulting group is itself
split into two, and so on until some stopping criterion is reached. As each tree is built on a random subset
of the predictors, the method is called “random forest” (Breiman, 2001). Binary regression trees can be
viewed as decision trees, each node being the criterion used to split the data and each final leaf giving
the predicted value. For example, if we observe a given wind speed x, we can find the final leaf that
corresponds to this value of x and the associated observations y, then we can compute the conditional
cumulative distribution function introduced by Meinshausen (2006)
Fb(y|x) =
n
X
ωi (x)1({Yi ≤ y}),
(5)
i=1
where the weights ωi (x) are deduced from the presence of Yi in a final leaf of each tree when one follows
the path determined by x. The interested reader is referred to Taillardat et al. (2016) for an application
of this approach to ensemble forecast of temperatures and winds.
2.2
Gradient forests
Meinshausen (2006) proposed splitting rule using CART regression splits. Arguing that this splitting
rule is not tailored to the quantile regression context, Athey et al. (2016) proposed another optimisation
scheme. Instead of maximizing the variance heterogeneity of the children nodes, one maximizes the
criterion
2
2
X
X
−1
∆(D1 , D2 ) =
ρi
(6)
|{i : Yi ∈ Dj }|
j=1
{i:Yi ∈Dj }
where the indicator function ρi = 1({Yi ≥ θ̂q,D0 }) is equal to one when Yi is greater than the q-th
quantile θ̂q,D0 of the observations of the parent node D0 . The terminology of gradient forests was
suggested because the choice of ρi is here linked with a gradient-based approximation of the quantile
function
Ψθ̂q,D (Yi ) = q1({Yi > q}) + (1 − q)1({Yi ≤ q}).
0
This technique using gradients is computationally feasible, an issue not to be omitted when dealing with
non-parametric techniques. Note here that for each split the order of the quantile is chosen among given
orders (0.1, 0.5, 0.9). In the special case of least-square regression, ρi becomes Yi −Y (D0 ), and H(D1 , D2 )
becomes equivalent to ∆(D1 , D2 ). In this special case, gradient trees are equivalent to build a standard
CART regression tree.
2.3
Fitting a parametric form to QRF and GF trees
As mentioned in Section 1.4, the predicted cdf defined by (5) cannot predict values which are not in the
learning sample. This can be a strong limitation if the learning sample sample is small or rare events
are of interest or both. The GF method has the same issue. To parametrically model rainfall, the EGP
4
pdf defined by (3) appears to be a good candidate. It allows more flexibility in the fitting than CSG or
CGEV. This distribution has four parameters, π, κ, σ and ξ, it is in compliance with EVT for low and
heavy rainfalls and works well in practice (see, e.g. Naveau et al., 2016). In terms of inference, a simple
and fast method-of-moment can be applied. Basically, probability weighted moments (PWM) of a given
random variable, say Y , with survival function F (y) = P(Y > y), can be expressed as (see, e.g. Hosking
and Wallis, 1987)
Z 1
r
µr = E([Y F (Y )]) =
F −1 (q)(1 − q)r dq.
(7)
0
If Y follows a EGP pdf defined by (3), then we have
ξ
ξ
1
µ0 = κB(κ, 1 − ξ) − 1 and µ1 = κ (B(κ, 1 − ξ) − B(2κ, 1 − ξ)) − ,
σ
σ
2
1
ξ
µ2 = κ (B(κ, 1 − ξ) − 2B(2κ, 1 − ξ) + B(3κ, 1 − ξ)) − ,
σ
3
where B(., .) represents the beta function. Knowing the PWM triplet (µ0 , µ1 , µ2 )T is equivalent to know
the parameter vector (κ, σ, ξ)T . Hence, we just need to estimate these three PWMs. For any given forest,
it is possible to estimate the distribution of [Y |X = x] by the empirical cdf Fb(y|X = x) defined by (5).
Then, we can plug it in (7) to get
Z
µ̂r (x) =
1
Fb−1 (q|X = x)(1 − q)r dq.
0
This leads to the estimates of (κ(x), σ(x), ξ(x))T and consequently of f (y|X = x) via Equation (3).
Note that the probability of no rain π(x) is just inferred by counting the number of dry events in the
corresponding trees. In the following, this technique is called ”EGP TAIL”, despite the fact that the
whole distribution is fitted from QRF and GF trees.
3
Ensemble model output statistics and EGP
In Section 1.3, three definitions of parametric pdfs were recalled. By regressing their parameters on the
ensemble values, different EMOS models have been proposed for the CSG and CGEV pdfs defined by
(1) and by (2), respectively. More precisely, Baran and Nemoda (2016) used the CSG pdf by letting
the mean µ = κθ and variance σ 2 = κθ2 depend linearly as functions of the raw ensemble values and
their mean, respectively. The coefficients of this regression were estimated by miminizing the continuous
ranked probability score (CRPS) (see, e.g. Scheuerer and Hamill, 2015; Hersbach, 2000). The same
strategy can be applied to fit the CGEV pdf (see, e.g. Hemri et al., 2014). Scheuerer (2014) modelled
the scale parameter σ in (2) as an affine function of the ensemble mean absolute deviation rather than
of the raw ensemble mean or variance. Another point to emphasize is that the shape parameter ξ was
considered invariant in space in Hemri et al. (2014).
In this section, we basically explain how an EMOS approach can be built with the EGP pdf defined
by (3) and we now highlight common features and differences between the two EMOS with CSG and
CGEV. The scale parameter σ 2 in (3) is estimated in the same way than for CGEV. The presence of the
parameter κ allows an additional degree of freedom. The expectation of our EGP is mainly driven by the
product κσ. Consequently, we model κ as an affine function of the predictors divided by σ. As France
has a diverse climate, it is not reasonable to assume a constant shape parameter among all locations,
see the map in Figure 1. In addition, minimizing the CRPS to infer different shape parameters may be
inefficient (see, e.g. Friederichs and Thorarinsdottir, 2012). To estimate ξ at each location, we simply
use the PWM inference scheme described in Section 2.3. To complete the estimation of the parameters
in (3), the probability π is modeled as an affine function on [0, 1] of the raw ensemble probability of
rain and affine function parameters are also estimated by CRPS minimization. The table 1 sums up the
optimal estimation strategies that we have found for each distribution.
5
shape parameter of EGPD3 derived from climatology
● ●
●
● ●
●
●
● ●
●
● ●●
●
● ●
● ● ●● ●
● ●
●
●
●
●
● ● ●
●
● ●●
●●
●
●
●
● ● ●
●
●
●
●
●
● ●
●
●
●
●
● ● ●
●
● ●● ● ● ● ●
●
● ● ● ●● ●
● ●
●
●● ● ● ● ● ● ●
● ●
0
0.025
0.05
0.075
0.1
0.125
0.15
0.175
0.2
0.225
0.25
0.275
0.3
0.325
0.35
0.375
0.4
0.425
0.45
0.475
0.5
0.525
0.55
0.575
0.6
●
●
Figure 1: Spatial values of ξ among locations.
4
4.1
Case study on the PEARP ensemble prediction system
Data description
Our rainfall dataset corresponds to 6-h rainfall amounts produced by 87 French weather stations and the
35-member ensemble forecast system called PEARP (Descamps et al., 2014) at a 51-h lead time forecast.
Our period of interest spans four years from 1 January 2012 to 31 December 2015.
4.2
Inferential details for EMOS and analogs
Verification has been made on this entire period. For a fair comparison each method has to be tuned
optimally. EMOS uses all the data available for each day (4 years less the forecast day as a training
period). The same strategy is used to fit the analogs method, see Appendix A for details on this approach.
6
Table 1: Optimal strategies for parameter estimation using CRPS minimization in the EMOS context.
Distribution
CSG
CGEV
EGP
Parameter
δ
µ
σ
κ
θ
µ
σ
ξ
θ
σ
µ
κ
ξ
π
Comments
free in R
affine function of covariates in C
affine function of raw ensemble mean
κ = µ2 /σ
θ = σ/µ
affine function of covariates in C
affine function of the mean absolute deviation of the raw ensemble
free in (−∞, 1)
θ = σ/µ
affine function of the mean absolute deviation of the raw ensemble
maximum between 0 and an affine function of covariates in C
κ = µ/σ
fixed, see Figure 1 for stations’ values
affine function of PR0 in C, bounded on [0, 1]
QRF and GF employ a cross-validation method: each month of the 4 years is kept as validation data
while the rest of the 4 years is used for learning. The tuning algorithm for EMOS is stopped after
few iterations in order to avoid overfitting, as suggested in Scheuerer (2014) concerning the parameter
estimations.
4.3
Sets of predictors used
We either use a subset of classical predictors (denoted by “C” in the rest of the paper) detailed in Table
2 or the whole set of available predictors as listed in Table 3.
Table 2: Subset “C” representing the most classical predictors.
Name
HRES
CTRL
MEAN
PR0
Description
high resolution member
control member
mean of raw ensemble
raw probability of rain
Note that we also considered for EMOS a third type of predictors set based on a variable selection
algorithm (see Appendix C). But this did not improve the results and we removed them from the analysis
(available upon request).
4.4
Zooming on extremes
Finding a way to assess the quality of ensembles for extreme and rare events is quite difficult, as seen
in Williams et al. (2014) in a comparison of ensemble calibration methods for extreme events. Weighted
scoring rules can be adopted as done in Gneiting and Ranjan (2011); Lerch et al. (2017) but there
are here two main issues. The ranking of compared methods depends on the weight function used, as
already suggested in Gneiting and Ranjan (2011). Besides, giving a weight to such rare events avoid
discriminant power of scoring rules, the same issue than for the Brier score (Brier, 1950). Moreover,
reliability is not sound here since there are not enough extreme cases (by definition) to measure it. We
have finally decided to focus on two ideas here, matching with forecasters’ desires: first, what is the
discriminant power of our forecasts for extreme events in terms of binary decisions ? Second, what is the
potential risk of our ensemble to mismatch an extreme event ? The choice done in our study is discussed
in Section 5.
7
Table 3: Set of all available predictors.
Name
HRES
CTRL
MEAN
MED
Q10
Q90
PR0
PR1
PR3
PR5
PR10
PR20
SIGMA
IQR
HU1500
UX
VX
FX
TCC
RR6CV
CAPE
Description
high resolution member
control member
mean of raw ensemble
median of raw ensemble
first decile of raw ensemble
ninth decile of raw ensemble
raw probability of rain
raw probability of rain > 1mm/6h
raw probability of rain > 3mm/6h
raw probability of rain > 5mm/6h
raw probability of rain > 10mm/6h
raw probability of rain > 20mm/6h
standard deviation of raw ensemble
IQR of raw ensemble
deterministic forecast of 6-h mean 1500m humidity
deterministic forecast of 6-h maximum of zonal wind gust
deterministic forecast of 6-h maximum of meridional wind gust
deterministic forecast of 6-h maximum of wind gust power
deterministic forecast of 6-h mean total cloud cover
deterministic forecast of 6-h convective rainfall amount
deterministic forecast of 6-h mean convective available potential energy
q10,50,90 are the first decile, the median and ninth decile of the raw ensemble for these variables:
HU q10,50,90
P q10,50,90
TCC q10,50,90
RR6CV q10,50,90
U10 q10,50,90
V10 q10,50,90
U500 q10,50,90
V500 q10,50,90
FF500 q10,50,90
TPW850 q10,50,90
FLIR6 q10,50,90
FLVIS6 q10,50,90
T q10,50,90
FF10 q10,50,90
5
6-h mean surface humidity
6-h mean sea level pressure
6-h mean total cloud cover
6-h convective rainfall amount
6-h mean surface zonal wind
6-h mean surface meridional wind
6-h mean 500m zonal wind
6-h mean 500m meridional wind
6-h mean 500m wind speed
6-h mean 850hPa potential wet-bulb temperature
6-h mean surface irradiation in infra-red wavelengths
6-h mean surface irradiation in visible wavelengths
6-h mean surface temperature
6-h mean surface wind speed
Results
Table 4 compares different metrics for all post-processing techniques which have been fitted to the 87
stations and averaged over 4 years of verification. Ten methods are competing: The raw ensemble, 4
analogs, 3 EMOS (3 different distributions using the set C), 2 forest-based methods (1 QRF and 1 GF)
and 2 tail-extended forest-based methods (1 QRF and 1 GF). Scores used concern respectively (i) global
performance (calibration and sharpness) measured by the CRPS; (ii) reliability performance, measured
by the mean, the normalized variance and the entropy of the PIT histograms, denoted by Ω in the sequel;
(iii) gain in CRPS compared to the raw ensemble, measured by the Skill of the CRPS using the raw
ensemble as baseline. A brief summary about these measures is done in D, where references are also
provided. And the boxplots showing rank histograms are in E.
According to Table 4, the raw ensemble is biased and underdispersive. The EMOS post-processed
ensembles share with QRF and GF a good CRPS. Moreover, we can consider them as unbiased and
mostly well-dispersed. The tail-extended methods get a lower CRPS, that can be explained by their
skill for extreme events. Finally, the four analog methods show a quite poor CRPS compared to the raw
ensemble, even if they exhibit reliability. Nevertheless we can notice that a weightning of the predictors,
especially with a non-linear variable selection algorithm (Analogs VSF), brings benefits to this method.
This phenomenon can be explained by Figure 2, where the ROC curves are given for the event of rain.
Consider a fixed threshold s and the contingency table associated to the predictor 1{rr6 > s}. Recall
that the ROC curve then plots the probability of detection (or hit rate) as a function of the probability
of false detection (or false alarm rate). A “good” prediction must maximize hit rates and minimize false
8
alarms (see, e.g. Jolliffe and Stephenson, 2012). Figure 2 explicitely shows the lack of resolution of the
analogs technique. Incidently, we can also notice that the rain event discrimination is not improved by
post-processed ensembles.
Table 4: Comparing performance statistics for different post-processing methods for 6-h rainfall forecasts
in France. The mean CRPS estimations come from bootstrap replicates, the estimation error is under
6.1 × 10−3 for all methods.
Types
Non-parametric
Parametric
with
covariates ∈ C
Hybrid
Methods
Raw ensemble
Analogs
Analogs C
Analogs COR
Analogs VSF
QRF
GF
EMOS
EMOS
EMOS
QRF
GF
pdf
CSG
GEV
EGP
EGP TAIL
EGP TAIL
CRPS
0.4694
0.5277
0.5376
0.5276
0.5247
0.4212
0.4134
0.4224
0.4228
0.4292
0.4138
0.4127
E(Z)
0.4164
0.5175
0.5050
0.5062
0.5060
0.5006
0.5070
0.4992
0.5000
0.4623
0.5095
0.5152
V(Z)
1.0612
1.0190
1.0051
1.0015
0.9986
0.9995
0.9771
1.0363
1.0073
1.0723
0.9558
0.9425
Ω
0.9809
0.9956
0.9964
0.9964
0.9961
0.9961
0.9957
0.9955
0.9961
0.9905
0.9957
0.9948
CRPSS
0%
-12.4%
-14.5%
-12.4%
-11.8%
10.3%
11.9%
10.0%
9.9%
8.6%
11.8%
12.1%
To sum up, the best improvement with respect to the raw ensemble is for the forest-based methods,
according to the CRPSS (which definition is in Appendix D). This improvement is however less significant
than for other weather variables (see Taillardat et al. (2016)). This corroborates Hemri et al. (2014)’s
conclusion that rainfall amounts are tricky to calibrate. If the analogs method looks less performant,
that might be imputable to the data depth of only 4 years. Indeed, this non-parametric technique is
data-driven (such as QRF and GF) and needs more data to be effective (see e.g. Van den Dool (1994)).
Concerning extreme events, Figure 3 shows the benefit of the tail extension for forest-based methods.
Note that we prefer to pay attention to the value of a forecast more than to its quality. According
to Murphy (1993), the value can be defined as the ability of the forecast to help users to take better
decisions. The quality of a forecast can be summarized by the area on the modelled ROC curve (classically
denoted by AUC), with some potential drawbacks exhibited by Lobo et al. (2008); Hand (2009). Zhu
et al. (2002) made a link between optimal decision thresholds, value and cost/loss ratios. In particular,
they show that the value of a forecast is maximized for the “climatological” threshold and equals the hit
rate minus the false alarm rate which is the maximum of the Peirce Skill Score (Manzato, 2007). This
value corresponds to the upper left corner of ROC curves, which is of main interest in terms of extremes
verification, as explained in Section 4.4. Several features already seen on Figure 2 can be observed on
Figure 3: analogs lack resolution and the other post-processed methods compete more or less favourably
with the raw ensemble. Nonetheless, the other post-processing techniques stay better than the raw
ensemble even for methods that cannot extrapolate observed values such as QRF and GF. Note that
QRF is rather surprisingly better than EMOS techniques. Tail extension methods show their gain in a
binary decision context.
6
Discussion
Throughout this study, we see that forest-based techniques compete favourably with EMOS techniques.
It is a good point to see that QRF and GF compared to EMOS exhibit nearly the same kind of improvement when focusing on rainfall amounts or on temperature and wind speed (see Taillardat et al. (2016)
Figures 6 and 13). It could be interesting to check these methods (especially GF) on smoother variables.
Tail extension of these non-parametric techniques generates ensembles more tailored for extremes
catchment. However, reliability as well as resolution remain quite stable when extending the tail, so that
our paradigm about verification (good extreme discrimination subject to satisfying overall performance)
9
0.4
Hit Rate
0.6
0.8
1.0
ROC Curve (rr6>0.2mm)
0.0
0.2
Raw ensemble
Analogs
Analogs_C
Analogs_COR
Analogs_VSF
EMOS CSG
EMOS EGP
EMOS GEV
QRF
QRF EGP TAIL
GF
GF EGP TAIL
0.0
0.2
0.4
0.6
0.8
1.0
False Alarm Rate
Figure 2: ROC Curves for the event of rain. A “good” prediction must maximize hit rate and minimize
false alarms. The analogs method lacks resolution. We can notice that there is no improvement of
post-processed methods compared to the raw ensemble.
remains.
One of the advantages of distribution-free calibration (analogs, QRF and GF) is that there is no
assumption on the parameters to calibrate. This benefit is emphasized for rainfall amounts for which
EMOS techniques have to be studied using different distributions. In this sense, the recent mixing method
of Baran and Lerch (2016) looks appealing. A brand new alternative solution consists in working with
(standardized) anomalies as done in Dabernig et al. (2016).
Another positive aspect of the forest-based methods is that there is no need of a predictor selection.
Concerning the analogs method, our results suggest that the work of Genuer et al. (2010) could be a
cheaper alternative to brute force algorithms like in Keller et al. (2017) for the weightning of predictors.
For analogs techniques, we can notice that the complete set of predictors gives the best results. In
10
0.4
Hit Rate
0.6
0.8
1.0
ROC Curve (rr6>15mm)
0.0
0.2
Raw ensemble
Analogs
Analogs_C
Analogs_COR
Analogs_VSF
EMOS CSG
EMOS EGP
EMOS GEV
QRF
QRF EGP TAIL
GF
GF EGP TAIL
0.0
0.2
0.4
0.6
0.8
1.0
False Alarm Rate
Figure 3: ROC Curves for the event of rain above 15mm. A “good” prediction must maximize hit rate
and minimize false alarms. The analogs method lacks resolution. Tail extension methods show their
gain in a binary decision context.
contrast, the choice of the set of predictors is still an ongoing issue for EMOS techniques regarding
precipitation. For easier variables to calibrate, Messner et al. (2017) shows that some variable selection
can be effective.
The tail extension can be viewed as a semi-parametric technique where the result of forest-based
methods is used to fit a distribution. This kind of procedure can be connected to the work of Junk
et al. (2015) who uses analogs on EMOS inputs. An interesting prospect would be to bring forest-based
methods in this context.
A natural perspective regarding spatial calibration and trajectory recovery could be to make use of
block regression techniques as done in Zamo et al. (2016), or of ensemble copula coupling, as suggested
by (Bremnes, 2007; Schefzik, 2016).
11
Finally, it appears that more and more weather services work on merging different forecasts from
different sources (multi-model ensembles). In this context, an attractive procedure could be to combine
raw ensembles and different methods of post-processing via sequential aggregation (Mallet, 2010; Thorey
et al., 2016), in order to get the best forecast according to the weather situations.
Acknowledgments
Part of the work of P. Naveau has been supported by the ANR-DADA, LEFE-INSU-Multirisk, AMERISKA,
A2C2, CHAVANA and Extremoscope projects. This work has been supported by Energy oriented Centre of Excellence (EoCoE), grant agreement number 676629, funded within the Horizon2020 framework
of the European Union. This work has been supported by the LABEX MILYON (ANR-10-LABX-0070)
of Université de Lyon, within the program ”Investissements d’Avenir” (ANR-11-IDEX-0007) operated
by the French National Research Agency (ANR). Thanks to Julie Tibshirani, Susan Athey and Stefan
Wager for providing gradient-forest source package.
Funding information
LABEX MILYON, Investissements d’Avenir and DADA, ANR, Grant/Award Numbers: ANR-10-LABX0070, ANR-11-IDEX-0007 and ANR-13-JS06-0007; EoCoE, Horizon2020, Grant/Award Number: 676629;
A2C2, ERC, Grant/Award Number: 338965
References
Akaike, H. (1998) Information theory and an extension of the maximum likelihood principle. In Selected
Papers of Hirotugu Akaike, 199–213. Springer.
Athey, S., Tibshirani, J. and Wager, S. (2016) Solving heterogeneous estimating equations with gradient
forests. arXiv preprint arXiv:1610.01271.
Baran, S. and Lerch, S. (2016) Mixture emos model for calibrating ensemble forecasts of wind speed.
Environmetrics.
Baran, S. and Nemoda, D. (2016) Censored and shifted gamma distribution based emos model for
probabilistic quantitative precipitation forecasting. Environmetrics, 27, 280–292.
Ben Bouallègue, Z. (2013) Calibrated short-range ensemble precipitation forecasts using extended logistic
regression with interaction terms. Weather and Forecasting, 28, 515–524.
Breiman, L. (1996) Bagging predictors. Machine learning, 24, 123–140.
— (2001) Random forests. Machine learning, 45, 5–32.
Breiman, L., Friedman, J., Stone, C. J. and Olshen, R. A. (1984) Classification and regression trees.
CRC press.
Bremnes, J. (2007) Improved calibration of precipitation forecasts using ensemble techniques. part 2:
Statistical calibration methods. met. Tech. rep., no research report 04.
Brier, G. W. (1950) Verification of forecasts expressed in terms of probability. Monthly weather review,
78, 1–3.
Bröcker, J. and Smith, L. A. (2007) Scoring probabilistic forecasts: The importance of being proper.
Weather and Forecasting, 22, 382–388.
Dabernig, M., Mayr, G. J., Messner, J. W. and Zeileis, A. (2016) Spatial ensemble post-processing with
standardized anomalies. Quarterly Journal of the Royal Meteorological Society.
12
De Haan, L. and Ferreira, A. (2007) Extreme value theory: an introduction. Springer Science & Business
Media.
Delle Monache, L., Eckel, F. A., Rife, D. L., Nagarajan, B. and Searight, K. (2013) Probabilistic weather
prediction with an analog ensemble. Monthly Weather Review, 141, 3498–3516.
Descamps, L., Labadie, C., Joly, A., Bazile, E., Arbogast, P. and Cébron, P. (2014) PEARP, the MétéoFrance short-range ensemble prediction system. Quarterly Journal of the Royal Meteorological Society.
Van den Dool, H. (1994) Searching for analogues, how long must we wait? Tellus A, 46, 314–324.
Ferro, C. (2014) Fair scores for ensemble forecasts. Quarterly Journal of the Royal Meteorological Society,
140, 1917–1923.
Friederichs, P. and Thorarinsdottir, T. L. (2012) Forecast verification for extreme value distributions
with an application to probabilistic peak wind prediction. Environmetrics, 23, 579–594.
Genuer, R., Poggi, J.-M. and Tuleau-Malot, C. (2010) Variable selection using random forests. Pattern
Recognition Letters, 31, 2225–2236.
Gneiting, T., Balabdaoui, F. and Raftery, A. E. (2007) Probabilistic forecasts, calibration and sharpness.
Journal of the Royal Statistical Society: Series B (Statistical Methodology), 69, 243–268.
Gneiting, T. and Katzfuss, M. (2014) Probabilistic forecasting. Annual Review of Statistics and Its
Application, 1, 125–151.
Gneiting, T. and Raftery, A. E. (2007) Strictly proper scoring rules, prediction, and estimation. Journal
of the American Statistical Association, 102, 359–378.
Gneiting, T., Raftery, A. E., Westveld III, A. H. and Goldman, T. (2005) Calibrated probabilistic
forecasting using ensemble model output statistics and minimum crps estimation. Monthly Weather
Review, 133, 1098–1118.
Gneiting, T. and Ranjan, R. (2011) Comparing density forecasts using threshold-and quantile-weighted
scoring rules. Journal of Business & Economic Statistics, 29, 411–422.
Hamill, T. M. and Colucci, S. J. (1997) Verification of eta-rsm short-range ensemble forecasts. Monthly
Weather Review, 125, 1312–1327.
Hamill, T. M., Hagedorn, R. and Whitaker, J. S. (2008) Probabilistic forecast calibration using ecmwf
and gfs ensemble reforecasts. part ii: Precipitation. Monthly weather review, 136, 2620–2632.
Hamill, T. M. and Whitaker, J. S. (2006) Probabilistic quantitative precipitation forecasts based on
reforecast analogs: Theory and application. Monthly Weather Review, 134, 3209–3229.
Hand, D. J. (2009) Measuring classifier performance: a coherent alternative to the area under the roc
curve. Machine learning, 77, 103–123.
Hemri, S., Scheuerer, M., Pappenberger, F., Bogner, K. and Haiden, T. (2014) Trends in the predictive
performance of raw ensemble weather forecasts. Geophysical Research Letters, 41, 9197–9205.
Hersbach, H. (2000) Decomposition of the continuous ranked probability score for ensemble prediction
systems. Weather and Forecasting, 15, 559–570.
Horton, P., Jaboyedoff, M. and Obled, C. (2017) Global optimization of an analog method by means of
genetic algorithms. Monthly Weather Review, 145, 1275–1294.
Hosking, J. R. and Wallis, J. R. (1987) Parameter and quantile estimation for the generalized pareto
distribution. Technometrics, 29, 339–349.
13
Hosking, J. R. M. (1989) Some theoretical results concerning L-moments. IBM Thomas J. Watson
Research Division.
Jolliffe, I. T. and Primo, C. (2008) Evaluating rank histograms using decompositions of the chi-square
test statistic. Monthly Weather Review, 136, 2133–2139.
Jolliffe, I. T. and Stephenson, D. B. (2012) Forecast verification: a practitioner’s guide in atmospheric
science. John Wiley & Sons.
Junk, C., Delle Monache, L. and Alessandrini, S. (2015) Analog-based ensemble model output statistics.
Monthly Weather Review, 143, 2909–2917.
Katz, R. W., Parlange, M. B. and Naveau, P. (2002) Statistics of extremes in hydrology. Advances in
water resources, 25, 1287–1304.
Keller, J. D., Delle Monache, L. and Alessandrini, S. (2017) Statistical downscaling of a high-resolution
precipitation reanalysis using the analog ensemble method. Journal of Applied Meteorology and Climatology.
Lerch, S., Thorarinsdottir, T. L., Ravazzolo, F., Gneiting, T. et al. (2017) Forecaster’s dilemma: extreme
events and forecast evaluation. Statistical Science, 32, 106–127.
Lobo, J. M., Jiménez-Valverde, A. and Real, R. (2008) Auc: a misleading measure of the performance
of predictive distribution models. Global ecology and Biogeography, 17, 145–151.
Mallet, V. (2010) Ensemble forecast of analyses: Coupling data assimilation and sequential aggregation.
Journal of Geophysical Research: Atmospheres, 115.
Manzato, A. (2007) A note on the maximum peirce skill score. Weather and Forecasting, 22, 1148–1154.
Matheson, J. E. and Winkler, R. L. (1976) Scoring rules for continuous probability distributions. Management science, 22, 1087–1096.
Meinshausen, N. (2006) Quantile regression forests. The Journal of Machine Learning Research, 7,
983–999.
Messner, J. W., Mayr, G. J. and Zeileis, A. (2017) Nonhomogeneous boosting for predictor selection in
ensemble postprocessing. Monthly Weather Review, 145, 137–147.
Murphy, A. H. (1993) What is a good forecast? an essay on the nature of goodness in weather forecasting.
Weather and forecasting, 8, 281–293.
Naveau, P., Huser, R., Ribereau, P. and Hannart, A. (2016) Modeling jointly low, moderate, and heavy
rainfall intensities without a threshold selection. Water Resources Research, 52, 2753–2769. URL:
http://dx.doi.org/10.1002/2015WR018552.
Papastathopoulos, I. and Tawn, J. A. (2013) Extended generalised pareto models for tail estimation.
Journal of Statistical Planning and Inference, 143, 131–143.
Pinson, P., Chevallier, C. and Kariniotakis, G. N. (2007) Trading wind generation from short-term
probabilistic forecasts of wind power. IEEE Transactions on Power Systems, 22, 1148–1156.
Raftery, A. E., Gneiting, T., Balabdaoui, F. and Polakowski, M. (2005) Using bayesian model averaging
to calibrate forecast ensembles. Monthly Weather Review, 133, 1155–1174.
Roulin, E. and Vannitsem, S. (2012) Postprocessing of ensemble precipitation predictions with extended
logistic regression based on hindcasts. Monthly weather review, 140, 874–888.
Roulston, M. S. and Smith, L. A. (2002) Evaluating probabilistic forecasts using information theory.
Monthly Weather Review, 130, 1653–1660.
14
Schefzik, R. (2016) Combining parametric low-dimensional ensemble postprocessing with reordering
methods. Quarterly Journal of the Royal Meteorological Society, 142, 2463–2477.
Scheuerer, M. (2014) Probabilistic quantitative precipitation forecasting using ensemble model output
statistics. Quarterly Journal of the Royal Meteorological Society, 140, 1086–1096.
Scheuerer, M. and Hamill, T. M. (2015) Statistical postprocessing of ensemble precipitation forecasts by
fitting censored, shifted gamma distributions. Monthly Weather Review, 143, 4578–4596.
Schwarz, G. et al. (1978) Estimating the dimension of a model. The annals of statistics, 6, 461–464.
Sloughter, J. M. L., Raftery, A. E., Gneiting, T. and Fraley, C. (2007) Probabilistic quantitative precipitation forecasting using bayesian model averaging. Monthly Weather Review, 135, 3209–3220.
Taillardat, M., Mestre, O., Zamo, M. and Naveau, P. (2016) Calibrated ensemble forecasts using quantile
regression forests and ensemble model output statistics. Monthly Weather Review, 144, 2375–2393.
Thorey, J., Mallet, V. and Baudin, P. (2016) Online learning with the continuous ranked probability
score for ensemble forecasting. Quarterly Journal of the Royal Meteorological Society.
Tribus, M. (1969) Rational Descriptions, Decisions and Designs. Pergamon Press, Elmsford, New York.
Van Schaeybroeck, B. and Vannitsem, S. (2015) Ensemble post-processing using member-by-member
approaches: theoretical aspects. Quarterly Journal of the Royal Meteorological Society, 141, 807–818.
Vrac, M. and Naveau, P. (2007) Stochastic downscaling of precipitation: From dry events to heavy
rainfalls. Water resources research, 43.
Weijs, S. V., Van Nooijen, R. and Van De Giesen, N. (2010) Kullback-leibler divergence as a forecast
skill score with classic reliability-resolution-uncertainty decomposition. Monthly Weather Review, 138,
3387–3399.
Williams, R., Ferro, C. and Kwasniok, F. (2014) A comparison of ensemble post-processing methods for
extreme events. Quarterly Journal of the Royal Meteorological Society, 140, 1112–1120.
Zamo, M. (2016) Statistical Post-processing of Deterministic and Ensemble Windspeed Forecasts on a
Grid. Ph.D. thesis, Université Paris-Saclay.
Zamo, M., Bel, L., Mestre, O. and Stein, J. (2016) Improved gridded wind speed forecasts by statistical
postprocessing of numerical models with block regression. Weather and Forecasting, 31, 1929–1945.
Zamo, M., Mestre, O., Arbogast, P. and Pannekoucke, O. (2014) A benchmark of statistical regression
methods for short-term forecasting of photovoltaic electricity production. part ii: Probabilistic forecast
of daily production. Solar Energy, 105, 804–816.
Zhou, B. and Zhai, P. (2016) A new forecast model based on the analog method for persistent extreme
precipitation. Weather and Forecasting, 31, 1325–1341.
Zhu, Y., Toth, Z., Wobus, R., Richardson, D. and Mylne, K. (2002) The economic value of ensemblebased weather forecasts. Bulletin of the American Meteorological Society, 83, 73–83.
A
Analogs method
Contrary to EMOS, this technique is data-driven. An analog for a given location and forecast lead
time is defined as a past prediction, from the same model, that has similar values for selected features
of the current model forecast. The method of analogs consists in finding these closest past forecasts
according to a given metric of the predictors’ space to build an analog-based ensemble (see e.g. Hamill
and Whitaker (2006)). We assume here that close forecasts leads to close observations. Making use of
15
analogs requires to choose both the set of predictors and the metric. Concerning the metric, several
have been tried like the Euclidean or the Mahalanobis distance but they have been outperformed by the
metric provided in Delle Monache et al. (2013):
v
u
N
t̃
v
X wj u
2
uX
t
Fj,t+i − Aj,t0 +i ,
(8)
σ fj
j=1
i=−t̃
where Ft represents the current forecast at time t for a given location. The analog for another time t0 at
this same location is At0 . The number of predictors is Nv and t̃ is half the time window used to search
analogs. We standardize the distance by the standard deviation of each predictor σfj calculated on the
learning sample for the considered location. In this study we take t̃ = 1 so the time window is ±24 hours
the forecast to calibrate. This distance has the advantages of being flow-dependent and thus defines a
real weather regime associated with the research of the analogs. Note that one could weight the different
predictors fj with wj and we fixed wj = 1 for all predictors in a first method (Analogs). We have
also tried two other weightning techniques using the absolute value of correlation coefficient between
predictors and the response variable (Analogs COR) like in Zhou and Zhai (2016), and a weightning
based on the frequency of predictors’ occurrences in variable selection algorithm described in Appendix
C (Analogs VSF). Note finally that other weightning techniques have been considered (Horton et al.,
2017; Keller et al., 2017) but we did not use them in this study because of their computational cost.
B
CRPS formula for EGP
The CRPS for the distribution F detailed in 3 is:
σ
CRP S(F, y) = y(2F (y) − 1) + (4π − 2F (y) − π 2 − 1)
ξ
"
!
#
1
2κσ(1 − π)
ξy − ξ
+
B
1+
; 1 − ξ, κ − (1 − π)B(1 − ξ, 2κ) − πB(1 − ξ, κ) ,
ξ
σ
where 0 < ξ < 1 and B( ; , ) and B( , ) denote respectively the incomplete beta and the beta functions.
C
Variable selection using random forests
We have seen that most parameters in EMOS and the distance used in analogs can be inferred using
different sets of predictors. Contrary to the QRF and GF methods where the add of a useless predictor
does not impact the predictive performance (since this predictor is never retained in the splitting rule),
it can be misguiding for EMOS and analogs. We have therefore investigated some methods that keep
the most informative meteorological variables and guarantee the best predictive performance. Our first
choice was to use the well-known Akaike information criterion and the Bayesian information criterion
(Akaike, 1998; Schwarz et al., 1978) but it resulted that the selection was not enough discriminant (too
many predictors kept in our initial set). The algorithm of Genuer et al. (2010) has then been considered.
Such an algorithm is appealing since it uses random forests (and we already have these objects from
the QRF method) and it permits to keep predictors without redundancy of information. For example
this algorithm eliminates correlated predictors even if they are informative. A reduced set of predictors
(mostly 3 or 4) is thus obtained, which avoids misestimation generated by multicolinearity. The method
of variable selection used here is one among plenty others. The interested reader in variable selection
using random forests can refer to Genuer et al. (2010) for detailed explanations.
The variable selection algorithm is used to keep the first predictors (max 4 of them) that form the set
of predictors for each location. Figure 4 shows the ranked frequency of each chosen predictor. Predictors
never retained are not on this figure. We can see here that only one third of the predictors in A are
retained at least in 10% of the cases. Moreover, predictors representing central and extreme tendencies
are preferred. Some predictors appear that differ from rainfall amounts ; see CAPE, FX or HU. It is not
16
Frequency of occurrence in variable selection algorithm on 86 stations
MEAN
Q90
IQR
PR1
HRES
PR3
CTRL
MED
PR0
RR6CV_q90
PR5
HU_q90
SIGMA
FLIR6_q90
FLIR6_q50
HU_q10
RR6CV_q50
U10_q90
CAPE
HU1500
RR6CV
FX
FLVIS6_q10
U500_q50
FF500_q90
FLVIS6_q50
TCC_q90
T_q50
FF500_q10
FLIR6_q10
HU_q50
P_q10
T_q10
T_q90
U500_q90
V10_q90
FF10_q90
FF500_q50
PR10
TCC_q10
TCC_q50
RR6CV_q10
U10_q10
FF10_q10
TCC
UX
VX
Q10
FLVIS6_q90
P_q50
P_q90
TPW850_q10
TPW850_q50
TPW850_q90
U10_q50
U500_q10
V500_q50
FF10_q50
0.0
0.2
0.4
0.6
0.8
Frequency
Figure 4: Frequency of predictors’ occurrence in variable selection algorithm. Variables representing
central and extreme tendencies are preferred. Some covariables like CAPE, FX or HU can be retained.
It is interesting to see that only one third of the predictors of the set is taken more than in 10% of the
cases.
surprising since these parameters are correlated with storms. It is not shown here but when the MEAN
variable is not selected, either MED or CTRL stands in the set. This shows that the algorithm mostly
selects just one information concerning central tendency and avoid potential correlations. So the results
concerning the variable algorithm selection seem to be sound. Last but not least, one notices that the
predictors of the set C are often chosen. This remark confirms both the robustness of the algorithm and
the relevance of previous studies on precipitation concerning the choice of the predictors.
17
D
Verification of ensembles
We recall here some facts about the scores used in this study.
D.1
Reliability
Reliability between observations and a predictive distribution can be checked by calculating Z 0 = F (Y )
where Y is the observation and F the cdf of the associated predictive distribution. Subject to calibration, the random variable Z 0 has a standard uniform distribution (Gneiting and Katzfuss, 2014) and we
can check ensemble bias by comparing E(Z 0 ) to 12 and ensemble dispersion by comparing the variance
1
Var(Z 0 ) to 12
. This approach is applied to a (K + 1) ranked ensemble forecast using the discrete random
variable Z = rank(y)−1
. Subject to calibration, Z has a discrete standard uniform distribution with
K
K
E(Z) = 21 and a normalized variance V(Z) = 12 K+2
Var(Z) = 1.
Another tool used to assess calibration is the entropy:
K+1
X
−1
Ω=
fi log(fi ).
log(K + 1)
i=1
For a calibrated system the entropy is maximum and equals 1. Tribus (1969) showed that the entropy is
an indicator of reliability linked to the Bayesian psi-test. It is also a proper measure of reliability used
in the divergence score described in Weijs et al. (2010); Roulston and Smith (2002).
These quantities are closely related to rank histograms which are discrete version of Probability
Integral Transform (PIT) histograms. However if one can assume the property of flatness of these
histograms, Jolliffe and Primo (2008) exhibit a test accounting for the slope and the shape of rank
histograms. In a recent work, Zamo (2016) extends this idea for accounting the presence of wave in
histograms as seen in Scheuerer and Hamill (2015); Taillardat et al. (2016). A more complete test can
thus be implemented that tests each histogram for flatness. Such a test is called the JPZ test (for
Jolliffe-Primo-Zamo). The results of the JPZ test is provided for each method in the E.
D.2
Scoring rules
Following Gneiting et al. (2007); Gneiting and Raftery (2007); Bröcker and Smith (2007), scoring rules
assign numerical scores to probabilistic forecasts and form attractive summary measures of predictive
performance, since they address calibration and sharpness simultaneously. These scores are generally
negatively oriented and we wish to minimize them. A proper scoring rule is designed such that the
expected value of the score is minimized by the perfect forecast, ie. when the observation is drawn
from the same distribution than the predictive distribution. The Continuous Ranked Probability Score
(CRPS) (Matheson and Winkler, 1976; Hersbach, 2000) is defined directly in terms of the predictive cdf,
F , as:
Z
∞
(F (x) − 1{x ≥ y})2 dx.
CRP S(F, y) =
−∞
Another representation (Gneiting and Raftery, 2007) shows that:
1
CRP S(F, y) = EF |X − y| − EF |X − X 0 |,
2
0
where X and X are independent copies of a random variable with distribution F and finite first moment.
An alternative representation for continuous distributions using L-moments (Hosking, 1989) is:
CRP S(F, y) = EF |X − y| + EF (X) − 2EF (XF (X)).
Throughout our study, if F is represented by an ensemble forecast with K members x1 , . . . , xK ∈ R,
we use a so-called fair estimator of the CRPS (Ferro, 2014) given by :
K
K X
K
X
1 X
1
\
CRP S(F, y) =
|xi − y| −
|xi − xj |.
K
2K(K − 1)
i=1
i=1 j=1
18
Notice that all CRPS have been computed following the recommendations of the Chapter 3 in Zamo
(2016).
We can also define the skill score in term of CRPS between an ensemble prediction system A and a
baseline B, in order to compare them directly:
CRP SS(A, B) = 1 −
CRP SA
CRP SB
The value of the CRPSS will be positive if and only if the system A is better than B for the CRPS
scoring rule.
E
Rank histograms boxplots
19
0.05
0.00
0.20
0.15
0.00
0.05
0.10
0.15
0.10
0.00
0.20
0.15
0.20
0.00
0.05
0.10
0.15
0.10
0.05
0.00
GF
GF EGP TAIL
JPZ test : 63 %
0.10
0.15
JPZ test : 98 %
0.05
0.05
0.10
0.15
JPZ test : 91 %
0.00
0.10
0.15
JPZ test : 100 %
QRF EGP TAIL
EMOS GEV
JPZ test : 98 %
0.00
QRF
EMOS EGP
JPZ test : 11 %
0.20
0.00
0.05
0.10
0.15
JPZ test : 89 %
0.20
0.20
0.00
0.05
0.10
0.15
JPZ test : 98 %
EMOS CSG
0.20
11
0.15
9
0.10
7
Analogs_COR
JPZ test : 99 %
0.05
5
JPZ test : 100 %
0.00
3
Analogs_VSF
Rank of observation
0.20
0.20
1
Analogs_C
0.05
0.10
0.05
0.00
0.00
0.05
0.10
0.15
JPZ test : 85 %
0.15
JPZ test : 0 %
Analogs
0.20
0.20
0.20
Raw ensemble
Figure 5: Boxplots of rank histograms for each technique according to the locations. The proportion
of rank histograms for which the JPZ test does not reject the flatness hypothesis is also provided. The
results confirm the Table 4.
20
| 10math.ST
|
Efficient Algorithms for Constructing Very Sparse Spanners and
Emulators∗
arXiv:1607.08337v2 [cs.DS] 6 Feb 2017
Michael Elkin†1 and Ofer Neiman‡1
1
Department of Computer Science, Ben-Gurion University of the Negev, Beer-Sheva, Israel.
Email: {elkinm,neimano}@cs.bgu.ac.il
Abstract
Miller et al. [MPVX15] devised a distributed1 algorithm in the CONGEST model, that given
a parameter k = 1, 2, . . ., constructs an O(k)-spanner of an input unweighted n-vertex graph with
O(n1+1/k ) expected edges in O(k) rounds of communication. In this paper we improve the result of
[MPVX15], by showing a k-round distributed algorithm in the same model, that constructs a (2k − 1)spanner with O(n1+1/k /ǫ) edges, with probability 1 − ǫ, for any ǫ > 0. Moreover, when k = ω(log n),
our algorithm produces (still in k rounds) ultra-sparse spanners, i.e., spanners of size n(1 + o(1)), with
probability 1 − o(1). To our knowledge, this is the first distributed algorithm in the CONGEST or in
the PRAM models that constructs spanners or skeletons (i.e., connected spanning subgraphs) that sparse.
Our algorithm can also be implemented in linear time in the standard centralized model, and for large k, it
provides spanners that are sparser than any other spanner given by a known (near-)linear time algorithm.
We also devise improved bounds (and algorithms realizing these bounds) for (1 + ǫ, β)-spanners and
emulators. In particular, we show that for any unweighted n-vertex graph and any ǫ > 0, there exists a
n log log n
)
)-emulator with O(n) edges. All previous constructions of (1 + ǫ, β)-spanners
(1 + ǫ, ( log log
ǫ
and emulators employ a superlinear number of edges, for all choices of parameters.
Finally, we provide some applications of our results to approximate shortest paths’ computation in
unweighted graphs.
1 Introduction
1.1 Setting, Definitions
We consider unweighted undirected n-vertex graphs G = (V, E). For a parameter α ≥ 1, a subgraph
H = (V, E ′ ), E ′ ⊆ E, is called an α-spanner of G, if for every pair u, v ∈ V of vertices, we have
dH (u, v) ≤ α · dG (u, v). Here dG (u, v) (respectively, dH (u, v)) stands for the distance between u and v
in G (resp., in H). The parameter α is called the stretch of the spanner H. More generally, if for a pair of
parameters α ≥ 1, β ≥ 0, for every pair u, v ∈ V of vertices, it holds that dH (u, v) ≤ α · dG (u, v) + β,
then the subgraph H is said to be an (α, β)-spanner of G. Particularly important is the case α = 1 + ǫ, for
∗
A preliminary version [EN17] of this paper appeared in SODA’17.
This research was supported by the ISF grant No. (724/15).
‡
Supported in part by ISF grant No. (523/12) and by BSF grant No. 2015813.
1
They actually showed a PRAM algorithm. The distributed algorithm with these properties is implicit in [MPVX15].
†
1
some small ǫ > 0. Such spanners are called near-additive. If H = (V, E ′′ , ω), where ω : E ′′ → R+ , is
not a subgraph of G, but nevertheless satisfies that for every pair u, v ∈ V of original vertices, dG (u, v) ≤
dH (u, v) ≤ (1 + ǫ)dG (u, v) + β, then H is called a near-additive β-emulator of G, or a (1 + ǫ, β)-emulator
of G.
Graph spanners have been introduced in [Awe85, PS89, PU89a], and have been intensively studied ever
since [ADD+ 93, ABCP93, Coh99, ACIM99, DHZ00, BS03, Elk04, Elk07a, EZ06, TZ06, Woo06, Elk07b,
Pet09, DGPV08, Pet10, BW15, MPVX15, AB16]. They were found useful for computing approximately
shortest paths [ABCP93, Coh99, Elk04, EZ06], routing [PU89b], distance oracles and labeling schemes
[Pel99, TZ05, EP15], synchronization [Awe85], and in other applications.
The simplest and most basic algorithm for computing a multiplicative α-spanner, for a parameter α ≥ 1,
is the greedy algorithm [ADD+ 93]. The algorithm starts with an empty spanner, and examines the edges of
the input graph G = (V, E) one after another. It tests if there is a path in H of length at most α between
the endpoints u and v of e. If it is not the case, the edge is inserted into the spanner. Otherwise the edge is
dropped.
It is obvious that the algorithm produces an α-spanner H. Moreover, the spanner H has no cycles of
length α + 1 or less, i.e., the girth of H, denoted g(H), satisfies g(H) ≥ α + 2. Denote m = m(n, g) the
maximum number of edges that a girth-g n-vertex graph may contain. It follows that |H| ≤ m(n, α + 2).
1+ 2
The function m(n, g) is known to be at most n g−2 , when g ≤ 2 log2 n, and for larger g (i.e., for m ≤ 2n),
it is given by m(n, g) ≤ n(1 + (1 + o(1)) ln(p+1)
), where p = m − n, [AHL02, BR10]. These bounds are
g
called “Moore’s bounds for irregular graphs”, or shortly, (generalized) Moore’s bounds.
Any construction of multiplicative α-spanners for n-vertex graphs with at most m′ (n, α + 2) edges
implies an upper bound m(n, α + 2) ≤ m′ (n, α + 2) for the function m(n, g). (As running the construction
on the extremal girth-(α + 2) n-vertex graph can eliminate no edge.) Hence the greedy algorithm produces
multiplicative spanners with optimal tradeoff between stretch and number of edges. (See also [FS16].)
However, the greedy algorithm is problematic from algorithmic perspective. In the centralized model of
computation, the best-known implementation of it [RZ04] requires O(α · n2+1/α ) time. Moreover, the
greedy algorithm is inherently sequential, and as such, it is generally hard2 to implement it in distributed
and parallel models of computation.
In the distributed model [Pel00] we have processors residing in vertices of the graph. The processors
communicate with their graph neighbors in synchronous rounds. In each round, messages of bounded length
can be sent. (This is the assumption of the CONGEST model. In the LOCAL model, messages’ size is
arbitrary.) The running time of an algorithm is this model is the number of rounds that it runs. By ”parallel”
model we mean here PRAM EREW model [Rei93], and we are interested in algorithms with small running
time (aka depth) and work complexities. (The latter measures the overall number of operations performed
by all processors.)
Dubhashi et al. [DMP+ 03] devised a distributed implementation of the greedy algorithm in the LOCAL
model of distributed computation. Their algorithm runs in O(α · log2 n) rounds, i.e., suboptimal by a factor
of log2 n. Moreover, it collects graph topology to depth O(α), and conducts heavy local computations.
To our knowledge, there is no distributed-CONGEST or PRAM implementation of the greedy algorithm
known. There is also no known efficient3 centralized, distributed-CONGEST, or PRAM algorithm that
constructs ultra-sparse spanners, i.e., spanners with n + o(n) edges.
2
In the sequel we discuss a distributed setting, specifically, the LOCAL model, in which a relatively efficient implementation of
the greedy is known.
3
By “efficient” centralized algorithm in this paper we mean an algorithm with running time close to O(|E|). By efficient
distributed or parallel algorithm we mean an algorithm that runs in polylogarithmic, or nearly-polylogarithmic, time.
2
In the distributed and parallel settings it is often enough to compute a sparse skeleton H of the input
graph G, where a skeleton is a connected subgraph that spans all the vertices of G, i.e., the stretch requirement is dropped. Dubhashi et al. [DMP+ 03] devised a distributed-LOCAL algorithm that computes ultrasparse skeletons of size m(n, α) ≤ n + O(n · logα n ) in O(α) rounds. Like their algorithm for constructing
spanners, this algorithm also collects topologies to depth O(α), and involves heavy local computations. To
our knowledge, no efficient distributed-CONGEST or PRAM algorithm for computing ultra-sparse skeletons is currently known. In this paper we devise the first such algorithms.
1.2 Prior Work and Our Results
In the centralized model of computation the best-known efficient algorithm for constructing multiplicative
spanners (for unweighted graphs) is due to Halperin and Zwick [HZ96]. Their deterministic algorithm, for
an integer parameter k ≥ 1, computes a (2k − 1)-spanner with n1+1/k + n edges in O(|E|) time. (Their
result improved previous pioneering work by [PS89, Coh99].) Note that their bound on the number of edges
is always at least 2n, i.e., in the range k = Ω(log n) it is very far from Moore’s bound.
Our centralized randomized algorithm computes (with probability close to 1), a (2k − 1)-spanner with
n · (1 + O( logk n )) edges in O(|E|) time, whenever k = Ω(log n). Note that when k = ω(log n), the number
of edges is n(1 + o(1)), i.e., in this range the algorithm computes an ultra-sparse spanner in O(|E|) time.
Moreover, whenever k ≤ n1−ǫ , for any constant ǫ > 0, up to a constant factor in the lower-order term, our
bound matches Moore’s bound. In fact, our algorithm and its analysis can be viewed as an alternative proof
of (a slightly weaker version of) the generalized Moore’s bound. Note that it is not the case for the greedy
algorithm and its implementations [ADD+ 93, RZ04, DMP+ 03]: the analysis of these algorithms relies on
Moore’s bounds, but these algorithms cannot be used to derive them.
Another variant of our algorithm, which works for any k ≥ 2, computes with high probability a (2k−1)2 ln n
spanner with n1+1/k (1 + O( logk k )) edges, in Õ(k|E|) time.4 In particular, for the range k ≥ ln
ln n the
1+1/k
number of edges in our spanner is n
+ o(n), improving the result of [HZ96] (albeit with a somewhat
2 ln n
1+1/k ) edges.
worse running time for ln
ln n ≤ k ≤ log n). Note that for any k ≥ 2 we have O(n
1+1/k
Yet another variant of our algorithm computes a (2k − 1)-spanner with O(n
) edges, in expected
O(|E|) time.
In the distributed-CONGEST and PRAM models, efficient algorithms for computing linear-size spanners were given in [Pet09, MPVX15]. Specifically, [MPVX15] devised an O(k)-round distributed-CONGEST
randomized algorithm for constructing O(k)-spanner (with high probability) with expected O(n1+1/k )
edges. In the PRAM model their algorithm has depth O(k log∗ n) and work O(|E|). There are also
k-round distributed-CONGEST randomized algorithms for constructing (2k − 1)-spanner with expected
O(k · n1+1/k ) edges [BS07, Elk07a]. It is known that at least k rounds are required for this task, under
Erdős’ girth conjecture [Elk07a, DGPV08].
Our randomized algorithm uses k rounds in the distributed-CONGEST model, and with probability at
least 1 − ǫ it constructs a (2k − 1)-spanner with O(n1+1/k /ǫ) edges (for any desired, possibly sub-constant,
ǫ > 0). In the PRAM model the depth and work complexities of our algorithm are the same as in [MPVX15].
n
Furthermore, when k ≥ log n we can bound the number of edges by n · (1 + O( log
ǫ·k )), again matching
Moore’s bound up to a constant factor in the lower-order term.
This result improves the previous state-of-the-art in the entire range of parameters. In particular, it is
also the first efficient algorithm in the distributed-CONGEST or PRAM models that constructs an ultrasparse skeleton. Specifically, in Õ(log n) time it computes an Õ(log n)-spanner with n(1 + o(1)) edges,
4
As usual, Õ(f (n)) stands for O(f (n)polylogf (n)).
3
with probability 1 − o(1).
We also use our algorithm for unweighted graphs to devise an improved algorithm for weighted graphs
as well. Specifically, our algorithm computes (2k − 1)(1 + ǫ)-spanner with O(n1+1/k · (log k)/ǫ) edges,
within expected O(|E|) time. See Theorem 2, and the discussion that follows it, for further details.
1.3 Near-Additive Spanners and Emulators
It was shown in [EP04] that for any ǫ > 0 and κ = 1, 2, . . ., and any (unweighted) n-vertex graph G =
(V, E), there exists a (1 + ǫ, β)-spanner with O(β · n1+1/κ ) edges, where β = β(κ, ǫ) ≤ O( logǫ κ )log κ . Additional algorithms for constructing such spanners were later given in [Elk04, EZ06, TZ06, DGPV08, Pet09,
Pet10]. Abboud and Bodwin [AB16] showed that multiplicative error of 1 + ǫ in [EP04]’s theorem cannot
be eliminated, while still keeping a constant (i.e., independent of n) additive error β, and more recently
log κ−1
1
.
[ABP17] showed that any such (1 + ǫ, β)-spanner of size O(n1+1/κ−δ ), δ > 0, has β = Ω ǫ·log
κ
In the regime of constant κ, the bound of [EP04] remains the state-of-the-art. Pettie [Pet09] showed that
n log log n
)
.
one can construct a (1 + ǫ, β)-spanner with O(n log log(ǫ−1 log log n)) edges and β = O( log log
ǫ
This result of [Pet09] is not efficient in the sense considered in this paper, i.e., no distributed or parallel
implementations of it are known, and also no efficient (that is, roughly O(|E|)-time) centralized algorithm
computing it is known. Also, this result does not extend ([Pet16]) to a general tradeoff between β and the
number of edges.
Improving upon previous results by [Elk05, EZ06], Pettie [Pet10] also devised an efficient distributedCONGEST algorithm, that for a parameter ρ > 0, constructs in Õ(nρ ) rounds a (1 + ǫ, β)-spanner with
logφ κ+1/ρ
√
O(n1+1/κ (ǫ−1 log κ)φ ) edges and β = O log κ+1/ρ
, for φ = 1+2 5 being the golden ratio. 5
ǫ
Independently and simultaneously to our work, [ABP17] showed that there exist (1 + ǫ, β)-spanners with
log κ−2
κ
< 3/4. This
, where h = (3/4)κ−1−log
O((ǫ−1 log κ)h · log κ · n1+1/κ ) edges and β = O logǫ κ
κ
spanner has improved dependence on ǫ in the number of edges (at the cost of worse dependence on κ).
In this paper we improve all of the tradeoffs [EP04, Pet10] in the entire range of parameters. Specifically,
log n
for any ǫ > 0, ρ > 0 and κ = 1, 2, . . . , log(1/ǫ)+log
log log n , our distributed-CONGEST algorithm constructs
in Õ(nρ ) rounds a (1 + ǫ, β)-spanner with O(n1+1/κ ) edges and
log κ + 1/ρ log κ+1/ρ
β≤O
.
ǫ
Our algorithm also admits efficient implementations in the streaming and standard models of computation,
see Section 3. Our spanners are sparser and have polynomially smaller β than the previous best efficient constructions. They are even sparser than the state-of-the-art existential ones (with essentially the same β), with
the following exceptions: whenever ǫ < 1/ log3 log n our result and that of [ABP17] are incomparable,6 and
the spanner from [Pet09] that has O(n log(4) n) edges, while ours never gets sparser than O(n log log n)/ǫ.
In the complementary range, ǫ > 1/ log3 log n, our result is strictly stronger than that of [ABP17].
Moreover, a variant of our algorithm efficiently constructs very sparse (1+ǫ, β)-emulators. In particular,
n log log n
we can obtain a linear-size (1 + ǫ, ( log log
)
)-emulator. (We stress that the number of edges does not
ǫ
depend even on ǫ.) All previous constructions of (1 + ǫ, β)-spanners or emulators employ a superlinear
number of edges, for all choices of parameters.
In the range of κ = o( logloglogn n ), the result of [Pet10] is incomparable with [EP04], as spanners of [EP04] provide smaller β,
while spanners of [Pet10] are slightly sparser.
6
The i-iterated logarithm is defined by log(i+1) n = log(log(i) n), for all i ≥ 0, and log(0) n = n.
5
4
We use our new algorithms for constructing near-additive spanners and emulators to improve approximate shortest paths’ algorithms, in the centralized and streaming models of computation. One notable result
in this context is a streaming algorithm that for any constant ǫ > 0 and any subset S ⊆ V with |S| = nΩ(1) ,
computes (1+ǫ)-approximate shortest paths for S ×V within O(|S|) passes over the stream, using O(n1+ǫ )
space. See Section 4 for more details, and additional applications of our spanners.
1.4 Technical Overview
Linial and Saks [LS93] were the first to employ exponential random variables to build network decompositions, i.e., partitions of graphs into clusters of small diameter, which possess some useful properties. This
technique was found useful for constructing padded partitions, hierarchically-separated trees, low-stretch
spanning trees [Bar96, Bar98, Bar04, EEST05, ABN11, AN12] and spanners [Coh99, BS03, Elk07b]. In
[LS93] every vertex v tosses a random variable rv from an exponential distribution, and broadcasts to all
vertices within distance rv from v. Every vertex v joins the cluster of a vertex u with largest identity number,
whose broadcast u heard.
Blelloch et al. [BGK+ 14] introduced a variant of this technique in which, roughly speaking, every vertex
v starts to broadcast at time −rv , and broadcasts indefinitely. A vertex x joins the cluster centered at a vertex
v, whose broadcast reaches x first. They called the resulting partition “exponential start time clustering”,
and it was demonstrated in [BGK+ 14, MPX13, EN16a] that this approach leads to very efficient distributed
and parallel algorithms for constructing padded partitions and network decompositions.
Miller et al. [MPVX15] used this approach to devise an efficient parallel and distributed-CONGEST
O(k)-time algorithm for constructing O(k)-spanner with O(n1+1/k ) edges. Specifically, they build the
exponential time clustering, add the spanning trees of the clusters into the spanner, and then every vertex x
adds into the spanner one edge (x, y) connecting x to every adjacent cluster Cy , y ∈ Cy .
The main property of the partition exploited by [MPVX15] in the analysis of their algorithm is that any
unit-radius ball in the input graph G intersects just O(n2/k ) clusters, in expectation. Note also that their
algorithm is doomed to use at least n1+1/k + (n − 1) edges, because it starts with inserting the spanning
trees of all clusters (amounting to up to n − 1 edges), and then inserts the O(n1+2/k ) edges crossing between
different clusters into the spanner. To get O(n1+1/k ) edges, one rescales k′ = 2k.
In our algorithm we do not explicitly construct the exponential start time clustering. Rather we run the
procedure that builds it, but every vertex x connects not just to the neighbor y through which x received its
first broadcast message at time, say, ty , but also to all neighbors z whose broadcast x received witin time
interval [ty , ty + 1]. We show that, in expectation, x connects to n1/k neighbors altogether, and not just to
that many adjacent clusters. As a result we obtain both a sparser spanner, a smaller stretch, and a smaller
running time. The stretch and running time are smaller roughly by a factor of 2 than in [MPVX15], because
we do not need to consider unit balls, that have diameter 2. Rather we tackle individual edges (of length 1).
In the context of weighted graphs, [MPVX15] showed how their efficient algorithm for constructing
sparse (4k − 2)-spanners for unweighted graphs can be converted into an efficient algorithm that constructs
(16k − 8)-spanners with O(n1+1/k log k) edges for weighted graphs. By using their scheme naively on top
of our algorithm for unweighted graphs, one gets an efficient algorithm for computing (4k − 2)(1 + ǫ)spanners with O(n1+1/k · (log k)/ǫ) edges. Roughly speaking, the overhead of 2 in the stretch is because
in the analysis of [MPVX15], every vertex contributes expected O(n1/k ) edges to the spanner on each of
roughly O(n1/k ) phases of the algorithm in which it participates. By employing a more delicate probabilistic
argument, we argue that in fact, the expected total contribution of every vertex in all phases altogether is
O(n1/k ), rather than O(n2/k ). This enables us to eliminate another factor of 2 from the stretch bound. See
Section 2.3 for details.
5
Our constructions of (1 + ǫ, β)-spanners and emulators follow the [EP04] superclustering and interconnection approach. One starts with a base partition P0 . In [EP04] this was the partition of [Awe85, PS89,
AP92], obtained via region-growing technique. Then every cluster C ∈ P0 that has “many” unclustered
clusters of P0 “nearby”, creates a supercluster around it. The “many” and the “nearby” are determined
by degree threshold deg 0 and distance threshold δ0 , respectively. Once the superclustering phase is over,
the remaining unclustered clusters enter an interconnection phase, i.e., every pair of participating nearby
clusters gets interconnected by a shortest path in the spanner. This completes one iteration of the process.
The resulted superclustering P1 is the input for the next iteration of this process, which runs with different,
carefully chosen thresholds deg 1 and δ1 . Such iterations continue until only very few clusters survive. The
latter are interconnected without further superclustering.
One bottleneck in devising efficient distributed algorithm based on this approach is the base partition. Known algorithms for constructing a region-growing partition of [Awe85] require almost linear distributed time [DMZ06]. We demonstrate that one can bypass it completely, and start from the base partition
P0 = {{v} | v ∈ V }. This requires some modfication of the algorithm, and a more nuanced analysis. In
addition, we show that the superclustering and interconnection steps themselves can be implemented efficiently. This part of the algorithm is based on our recent work on hopsets [EN16b], where we showed that
[EP04] approach is extremely useful in that context as well, and that it can be made efficient.
1.5 Related Work
Efficient algorithms for constructing (1 + ǫ, β)-spanners were also devised in [Elk05, EZ06, TZ06, Pet09,
Pet10]. These algorithms are based, however, on different approaches than that of the current paper. The
latter is based on [EP04]. Specifically, the approach of [Elk05, EZ06] is based on [Coh99, Coh00] construction of pairwise covers and hopsets, i.e., the algorithm works top-down. It recurses in small clusters, and
eliminates large ones. The approach of [TZ06, Pet09, Pet10] is based on [TZ05] collection of trees, used
originally for distance oracles.
Streaming algorithms for constructing multiplicative spanners were given in [FKM+ 05, Elk07b, Bas08],
and near-additive spanners in [Elk05, EZ06]. Spanners and emulators with sublinear error were given
in [TZ06, Pet09]. Spanners with purely additive error and lower bounds concerning them were given in
[ACIM99, EP04, BCE05, BKMP10, Che13, Woo06, BW15, AB16].
1.6 Organization
In Section 2 we present our algorithm for constructing multiplicative spanners and its analysis. In Section
2.3 we use this algorithm to provide improved spanners for weighted graphs as well. Our near-additive
spanners and emulators are presented in Section 3.
2 Sparse Multiplicative Spanners and Skeletons
Let G = (V, E) be a graph on n vertices, and let k ≥ 1 be an integer. Let c > 3 be a parameter governing
the success probability, and set β = ln(cn)/k. Recall the exponential distribution with parameter β, denoted
EX P(β), which has density
β · e−βx
x≥0
f (x) =
0
otherwise.
6
Construction. Each vertex u ∈ V samples a value ru from EX P(β), and broadcasts it to all vertices
within distance k. Each vertex x that received a message originated at u, stores mu (x) = ru − dG (x, u),
and also a neighbor pu (x) that lies on a shortest path from x to u (this neighbor sent x the message from u,
breaking ties arbitrarily if there is more than one). Let m(x) = maxu∈V {mu (x)}, then for every x ∈ V we
add to the spanner H the set of edges
C(x) = {(x, pu (x)) : mu (x) ≥ m(x) − 1} .
The following lemma is implicit in [MPVX15]. We provide a proof for completeness.
Lemma 1 ([MPVX15]). Let d1 ≤ . . . ≤ dn be arbitrary values and let δ1 , . . . , δn be independent random
variables sampled from EX P(β). Define the random variables M = maxi {δi − di } and I = {i : δi − di ≥
M − 1}. Then for any 1 ≤ t ≤ n,
Pr[|I| ≥ t] = (1 − e−β )t−1 .
Proof. Denote by X (t) the random variable which is the t-th largest among {δi − di }. Then for any value
a ∈ R, if we condition on X (t) = a, then the event |I| ≥ t is exactly the event that all the remaining t − 1
values X (1) , . . . , X (t−1) are at least a and at most a + 1. Using the memoryless property of the exponential
distribution and the independence of the {δi }, we have that
Pr[|I| ≥ t | X (t) = a] = (1 − e−β )t−1 .
Since this bound does not depend on the value of a, applying the law of total probability we conclude that
Pr[|I| ≥ t] = (1 − e−β )t−1 .
Using this lemma, we can bound the expected size of the spanner.
Lemma 2. The expected size of H is at most (cn)1/k · n.
Proof. Fix any x ∈ V , and we analyze E[|C(x)|]. Note that the event |C(x)| ≥ t happens when there are at
least t shifted random variables ru − dG (u, x) that are within 1 of the maximum. By Lemma 1 this happens
with probability at most (1 − e−β )t−1 (we remark that if x did not hear at least t messages, then trivially
Pr[|C(x)| ≥ t] = 0). We conclude that
E[|C(x)|] =
n
X
t=1
Pr[|C(x)| ≥ t] ≤
∞
X
t=0
(1 − e−β )t = eβ = (cn)1/k ,
and the lemma follows by linearity of expectation.
We now argue about the stretch of the spanner.
Claim 3. With probability at least 1 − 1/c, it holds that ru < k for all u ∈ V .
Proof. For any u ∈ V , Pr[ru ≥ k] = e−βk = 1/(cn). By the union bound, Pr[∃u, ru ≥ k] ≤ 1/c.
Assume for now that the event of Claim 3 holds, i.e., that ru < k for all u ∈ V .
Corollary 4. For any x ∈ V , if u ∈ V is the vertex maximizing mu (x), then dG (u, x) < k.
7
Proof. First note that m(x) ≥ mx (x) ≥ 0, and using Claim 3 we have ru < k. So 0 ≤ m(x) = mu (x) =
ru − dG (u, x) < k − dG (u, x).
Claim 5. For any u, x ∈ V , if x adds an edge to pu (x), then there is a shortest path P between u and x
that is fully contained in the spanner H.
Proof. We prove by induction on dG (u, x). In the base case dG (x, u) = 1, then pu (x) = u, so (x, u) is
in the spanner. Assume that every vertex y ∈ V with dG (u, y) = t − 1 which added an edge to pu (y) has
a shortest path to u in H, and we prove for x that has dG (u, x) = t. We know that x added an edge to
y = pu (x), which lies on a shortest path to u, and thus satisfies dG (u, y) = t − 1. It remains to show that
this y added an edge to pu (y). First we claim that
m(y) ≤ m(x) + 1 .
(1)
Seeking contradiction, assume that (1) does not hold, and let v ∈ V be the vertex maximizing mv (y). By
Corollary 4 we have dG (v, y) < k, and thus dG (v, x) ≤ k. Hence x will hear the message of v. This means
that mv (x) ≥ mv (y) − 1 = m(y) − 1 > m(x), which is a contradiction. This establishes (1). Now, since x
added an edge to y = pu (x), by construction
mu (x) ≥ m(x) − 1 .
We conclude that
(2)
(1)
(2)
mu (y) = mu (x) + 1 ≥ m(x) − 1 + 1 ≥ m(y) − 1 ,
so y indeed adds an edge to pu (y), and by the induction hypothesis we are done.
Lemma 6. The spanner H has stretch at most 2k − 1.
Proof. Since H is a subgraph of G, it suffices to prove for any (x, y) ∈ E, that dH (x, y) ≤ 2k − 1. Let
u be the vertex maximizing m(x) = mu (x), and w.l.o.g assume m(x) ≥ m(y). By Corollary 4 we have
dG (u, x) ≤ k − 1, so dG (u, y) ≤ k, and y heard the message of u (which was sent to distance k). This
implies that mu (y) ≥ mu (x) − 1 = m(x) − 1 ≥ m(y) − 1, so y adds the edge (y, pu (y)) to the spanner. By
applying Claim 5 on x and y, we see that both have shortest paths to u that are fully contained in H. Since
dG (x, u) ≤ k − 1 and dG (y, u) ≤ k, these two paths provide stretch 2k − 1 between x, y.
2.1 Main Theorem
We now state our main theorem, from which we will derive several interesting corollaries in various settings.
Theorem 1. For any unweighted graph G = (V, E) on n vertices, any integer k ≥ 1, c > 3 and δ > 0,
there is a randomized algorithm that with probability at least (1 − 1/c) · δ/(1 + δ) computes a spanner with
stretch 2k − 1 and number of edges at most
(1 + δ) ·
(cn)1+1/k
− δ(n − 1) .
c−1
Proof. Let Z be the event that {∀u ∈ V, ru < k}. By Claim 3 we have Pr[Z] ≥ 1 − 1/c. Note that
conditioning on Z, by Lemma 6 the algorithm produces a spanner H = (V, E ′ ) with stretch 2k − 1. In
particular, it must have at least n − 1 edges. Let X be the random variable |E ′ | − (n − 1), which conditioned
on Z takes only nonnegative values. By Lemma 2 we have E[X] = (cn)1/k · n − (n − 1). We now argue
8
that conditioning on Z will not affect this expectation by much. Indeed, by the law of total probability, for
any t, Pr[X = t] ≥ Pr[X = t | Z] · Pr[Z]. Thus
E[X | Z] ≤
By Markov inequality,
h
i
E[X]
c
≤
· (cn)1/k · n − (n − 1) .
Pr[Z]
c−1
Pr [X ≥ (1 + δ)E[X | Z] | Z] ≤
(3)
1
.
1+δ
We conclude that
Pr [(X < (1 + δ)E[X | Z]) ∧ Z] = Pr [X < (1 + δ)E[X | Z] | Z] · Pr[Z] ≥
1
1−
c
·
δ
.
1+δ
If this indeed happens, then
|E ′ | = X + n − 1
h
i
(3)
c
≤ (1 + δ) ·
· (cn)1/k · n − (n − 1) + n − 1
c−1
(cn)1+1/k − (n − 1)
= (1 + δ) ·
− δ(n − 1) .
c−1
2.1.1
Implementation Details
Distributed Model. It is straightforward to implement the algorithm in the LOCAL model of computation,
it will take k rounds to execute it – in each round, every vertex sends to its neighbors all the messages it
received so far. We claim that the algorithm can be implemented even when bandwidth is limited, i.e., in the
CONGEST model. This will require a small variation: in each round, every vertex v ∈ V will send to all its
neighbors the message (ru , dG (u, v)) for the vertex u that currently maximizes mu (v) = ru − dG (u, v). We
note that omitting all the other messages will not affect the algorithm, since if one such message would cause
some neighbor of v to add an edge to v, then the message about u will suffice, as the latter has the largest
mu (v) value. (Also recall that all vertices start their broadcast simultaneously, and do so for k rounds, so
any omitted message could not have been sent to further distance than the message from u, which implies
dropping it will have no effects on farther vertices as well.)
PRAM Model. In the parallel model of computation, we can use a variant of the construction that appeared
in [MPX13, MPVX15]. Roughly speaking, vertex u will start its broadcast at time k − ⌈ru ⌉, and every
vertex x will send only the first message that arrives to it (which realizes m(x)). As argued in [MPVX15],
the algorithm can be implemented in O(k log∗ n) depth and O(|E|) work.
Standard Centralized Model. Note that in the standard centralized model of computation, the running
time is at most the work of the PRAM algorithm, which is O(m). By taking constant c and δ, and repeating
the algorithm until the first success (we can easily check the number of edges of the spanner and that all
ru < k), we get a spanner with stretch 2k − 1 and O(n1+1/k ) edges in expected time O(|E|).
9
2.2 Implications of Theorem 1
2.2.1
Standard Centralized Model and PRAM
The currently sparsest spanners which can be constructed in linear time are those of Halperin and Zwick
[HZ96]. They provide for any k ≥ 1, a deterministic algorithm running in O(m) time, that produces a
spanner with n1+1/k + n edges. We can improve their result for a wide range of k, albeit with a randomized
algorithm. First we show a near-linear time algorithm (which can be also executed in parallel), that provides
a spanner sparser than Halperin and Zwick in the range k ≥ 2 ln n/ ln ln n.7
Corollary 7. For any unweighted graph G = (V, E) on n vertices and m edges, and any integer k ≥ 2,
there is a randomized
algorithm,
that with high probability8 computes a spanner for G with stretch 2k − 1
k)
edges. The algorithm has O(k2 ln n ln∗ n) depth and the running time (or work)
and n1+1/k · 1 + O(ln
k
is Õ(k|E|).
Proof. Apply Theorem 1 with parameters c = k and δ = 1/k. So with probability at least
we obtain a spanner whose number of edges is at most
(kn)1+1/k
O(ln k)
1+1/k
(1 + 1/k) ·
≤n
· 1+
.
k−1
k
k−1
k
·
1
k+1
≥
1
3k
(4)
Run the algorithm C · k ln n times for some constant C. We noted in Section 2.1.1 that each run takes
O(k ln∗ n) depth and O(|E|) work, so the time bounds are as promised. Now, with probability at least
1 − (1 − 1/(3k))C·k ln n ≥ 1 − n−C/3 , we achieved a spanner with number of edges as in (4) in one of the
executions.
√
Remark 1. Whenever k ≥ 2 ln n/ ln ln n, we have n1/k ≤ ln n, so the number of edges in Corollary 7 is
n1+1/k + o(n), and the running time is Õ(k|E|).
2.2.2
Distributed Model
In a distributed setting we have the following result.
Corollary 8. For any unweighted graph G = (V, E) on n vertices, any k ≥ 1 and 0 < ǫ < 1, there is a
randomized distributed algorithm that with probability at least 1 − ǫ computes a spanner with stretch 2k − 1
and O(n/ǫ)1+1/k edges, within k rounds.
Proof. Apply Theorem 1 with c = 3/ǫ and δ = 2/ǫ, so the success probability is at least
(1 − ǫ/3) · (1 − ǫ/(ǫ + 2)) > 1 − ǫ .
With these parameters, by Theorem 1, the number of edges in spanner will be bounded by O(n/ǫ)1+1/k .
7
8
In fact, the factor 2 can be replaced by any 1 + ǫ for constant ǫ > 0.
By high probability we mean probability at least 1 − n−C , for any desired constant C.
10
2.2.3
Ultra-Sparse Spanners and Skeletons
We now show that in the regime k ≥ ln n, our algorithm (that succeeds with probability close to 1) provides
a spanner whose number of edges is very close to n (as a function of k and the success probability). This
will hold in all computational models we considered. We note that for the centralized and PRAM models,
Corollary 7 gives high probability with roughly the same sparsity, albeit with larger depth and work.
Corollary 9. For any unweighted graph G = (V, E) on n vertices, and any integer k ≥ ln n and parameter
2/k < ǫ < 1, there is a randomized
algorithm,
that with probability at least 1 − ǫ computes a spanner for G
O(ln n)
with stretch 2k − 1 and n · 1 + ǫ·k
edges. The number of rounds in distributed model is k, in PRAM
∗
it is O(k log n) depth and O(|E|) work, and in the centralized model it is O(|E|) time.
Proof. Apply Theorem 1 with parameters c = k and δ = 2/ǫ, so with probability at least
≥ 1−ǫ
e(2 ln n)/k
≤ 1 + O(ln n)/k, so
(kn)1+1/k
O(ln n)
(1 + 2/ǫ) ·
− 2(n − 1)/ǫ ≤ n · 1 +
.
k−1
ǫ·k
(5)
we obtain a (2k − 1)-spanner. In the regime k ≥ ln n we have
the number of edges is at most
(kn)1/k
2/ǫ
k−1
k · 2/ǫ+1
≤
Remark 2. The spanner of Corollary 9 can be used as a skeleton. E.g., one can take ǫ = o(1) and
k = Õ(log n), to obtain with probability 1 − o(1), a skeleton with n(1 + o(1)) edges, which is computed in
Õ(log n) rounds.
2.3 Weighted Graphs
Miller et al. [MPVX15] used their efficient algorithm for constructing O(k)-spanners with O(n1+1/k ) edges
for unweighted graphs, to provide an efficient algorithm for constructing O(k)-spanners with O(n1+1/k log k)
edges for weighted graphs. In this section we argue that their scheme can be used to convert our algorithm
for constructing (2k − 1)-spanners with O(n1+1/k ) edges for unweighted graphs (Theorem 1; see also Section 2.1.1) into an efficient algorithm for constructing (2k − 1)(1 + ǫ)-spanners with O(n1+1/k · logǫ k ) edges
for weighted graphs.
The scheme of [MPVX15] works in the following way. It partitions all edges of G = (V, E) into
⌈log1+ǫ ωmax ⌉ = λ categories Et = {e ∈ E | ω(e) ∈ [(1 + ǫ)t−1 , (1 + ǫ)t )}, t = 1, 2, . . . , λ. (We assume
that the minimum weight is 1, and the maximum weight is ωmax . The last category Eλ should also contain
edges of weight exactly ωmax .) Now one defines ℓ = ⌈log1+ǫ kc ⌉, for a sufficiently large constant c, graphs
S
Gj = (V, Êj ), j = 0, 1, . . . , ℓ − 1, Êj = {Et | t ≡ j( mod ℓ)}.
Observe that the edge weights in (each) Gj are well-separated, i.e., Êj is a disjoint union of at most
q = ⌈λ/ℓ⌉ edge sets E (1) , . . . , E (q) , such that the edge weights within each set are within a factor of 1 + ǫ
from one another. Moreover, if edge weights in E (i) , i = 1, 2, . . . , q, are in the range [w(i) , (1 + ǫ)w(i) ),
then we have w(i) = w(1) (kc )i−1 , for every i = 1, 2, . . . , q.
For each graph Gj , the scheme of [MPVX15] constructs an O(k)-spanner with O(n1+1/k ) edges. It
then takes a union of ℓ = O( logǫ k ) such spanners as the ultimate O(k)-spanner of the original graphs. (In
fact, [MPVX15] used specifically ǫ = 1.) We will next outline the way in which [MPVX15] construct O(k)spanner with O(n1+1/k ) edges for each Gj , and show how to modify it to provide a (2k − 1)(1 + ǫ)-spanner.
11
The scheme starts with running a routine of [MPVX15] that constructs an O(k)-spanner H (1) with
O(n1+1/k ) edges for the unweighted graph (V (1) , E (1) ), V (1) = V , and constructing the exponential start
time partition P (1) for it. It then contracts each of the clusters of P (1) (which have unweighted radii at most
k − 1) into single vertices of V (2) , and runs the unweighted spanner routine on (V (2) , E (2) ). As a result, it
constructs an O(k)-spanner H (2) and a partition PS(2) of V (2) , contracts all clusters of P (2) to get V (3) , etc.
The final spanner returned by the scheme is H = qi=1 H (i) .
The scheme guarantees stretch (1 + ǫ)(1 + O(k−(c−1) ))O(k), because the blackbox routine for unweighted graphs provides stretch O(k) for each category of weights, but the weights are uniform only up to
a factor of 1 + ǫ. Also, the factor of 1 + O(k−(c−1) ) appears, because one contracts clusters of unweighted
diameter O(k) of lower scales, on which all edge weights are a factor of roughly k−c smaller than the edge
weights on the current scale.
In the analysis of |H|, [MPVX15] show that every vertex u is active (i.e., non-isolated vertex which
is not yet contracted into a larger super-vertex) for expected O(n1/k ) phases, and when it is active, it contributes expected O(n1/k ) edges to the spanner of the current phase. Hence the overall size of the spanner
is O(n1+2/k ), and by rescaling k′ = 2k, they ultimately get their result. (See the proof of Theorem 3.3 in
[MPVX15] for full details of this proof. We have sketched it for the sake of completeness.)
While the stretch analysis of [MPVX15] is sufficiently precise for our purposes, this is not the case with
the size analysis. Indeed, even when one plugs in stretch 2k − 1 of our unweighted spanner routine instead
of stretch O(k) of their routine, still one obtains a (2k − 1)(1 + ǫ)2 (1 + k−(c−1) )-spanner with O(n1+2/k )
edges, i.e., a (4k − 2)(1 + O(ǫ))-spanner with O(n1+1/k ) edges (for each Gj ).
In what follows we refine their size analysis, and show that, in fact, every vertex u contributes expected
O(n1/k ) edges in all phases of the algorithm altogether (for a single graph Gj with well-separated edge
(i)
weights). Denote by ru the radius that u tosses from EX P(β) in the ith phase, i = 1, 2, . . . , q, assuming
that it is active on that phase. We say that a vertex v (which is active on phase i) is a candidate vertex
(i)
of phase i if its broadcast message reaches u no later than within one time unit after the time −ru , i.e.,
(i)
(i)
−ru + 1 ≥ −rv + d(v, u), where d(v, u) is the unweighted distance between v and u in the graph on
which the unweighted spanner routine is invoked on phase i.
Let X (i) denote the random variable counting the number of such candidate vertices on phase i, for
i = 1, 2, . . . , q. (Recall that these are the vertices which might cause u to add an edge to the spanner.) Let
j denote the random variable which is the phase in which u was contracted (and j = q if there is no such
(i)
phase). That is, j indicates the level in which the broadcast of some candidate vertex v has −rv +d(v, u) <
(i)
−ru . Denote also by X̂ (i) , i = 1, 2, . . . , q, the total number of candidates u sees between the beginning
(j)
of phase i, and up until phase j, where a candidate vertex v reaches u before time −ru . On that phase
j, X̂ (i) counts the number of candidates with index not larger than that of the candidate vertex v (assume
every vertex has an arbitrary distinct index in {1, . . . n}). Note that for i > j, by definition, X̂ (i) = 0. Also,
in particular, X̂ = X̂ (1) is at least the total contribution of edges u adds to the spanner, except for up to
expected O(n1/k ) edges that it might contribute on phase j (as shown in Lemma 2).
We next argue that for any t = 0, 1, 2, . . .,
Pr(X̂ > t) ≤ (1 − e−β )t .
This implies that
E(X̂) =
∞
X
t=0
Pr(X̂ > t) ≤
∞
X
t=0
(1 − e−β )t = eβ = O(n1/k ) .
12
(6)
First, note that
Pr(X̂ > t) =
∞
X
t(1) =0
Pr(X (1) = t(1) ) · Pr(X̂ > t | X (1) = t(1) ) .
For t(1) > t, we have
Pr(X̂ > t | X (1) = t(1) ) = (1 − e−β )t .
To justify this equation, note that the left-hand side is exactly the probability that on the first phase, all the
(1)
(1)
t first candidate vertices v have −rv + d(v, u) ≥ −ru , conditioned on them being candidates, i.e., on
(1)
−rv + d(v, u) ≤ −ru + 1. Since these are independent shifted exponential random variables, the equation
follows from the memoryless property of the exponential distribution.
For t(1) ≤ t, we have
Pr(X̂ > t | X (1) = t(1) )
=
(1)
(1 − e−β )t
(1)
= (1 − e−β )t
· Pr(X̂ (2) > t − t(1) | X (1) = t(1) , X̂ (1) ≥ t(1) )
· Pr(X̂ (2) > t − t(1) | X̂ (1) ≥ X (1) ) .
(1)
(Again, (1 − e−β )t is the probability that no candidate vertex of the first phase reached u before time
(1)
−ru , and so u was not contracted away at this phase.)
We conduct an induction on the phase, where the induction base is the last phase i = q. On the last
phase, for any h,
Pr(X̂ (q) > h | X̂ (1) ≥ X (1) , X̂ (2) ≥ X (2) , . . . , X̂ (q−1) ≥ X (q−1) ) ≤ (1 − e−β )h ,
(q)
(1)
because it is just the probability that none of the first h candidates (that have −rv + d(v, u) ≤ −ru + 1)
(q)
reaches u before time −ru . (If there are fewer candidates or u was contracted in a previous phase, then
this probability is 0.)
Hence by the inductive hypothesis,
(1)
Pr(X̂ (2) > t − t(1) | X̂ (1) ≥ X (1) ) ≤ (1 − e−β )t−t
,
and so
for any t(1) ≤ t as well. Hence
Pr(X̂ (1) > t | X (1) = t(1) ) ≤ (1 − e−β )t ,
Pr(X̂ > t) =
≤
∞
X
t(1) =0
∞
X
t(1) =0
Pr(X (1) = t(1) ) · Pr(X̂ > t | X (1) = t(1) )
Pr(X (1) = t(1) ) · (1 − e−β )t = (1 − e−β )t ,
as required.
Hence the expected contribution of every vertex is O(n1/k ), and the overall spanner size for each Gj
is, in expectation, O(n1+1/k ). The running time of the algorithm is expected to be O(|E|), following the
analysis of [MPVX15]. Moreover, as in [MPVX15], our algorithm can be implemented in PRAM model,
in O(log n · log∗ n · log Λ) depth, and O(|E|) work, where Λ is the aspect ratio of the input graph. We
summarize the result below.
13
Theorem 2. Given a weighted n-vertex graph G, and a pair of parameters k ≥ 1, 0 < ǫ < 1, our algorithm
computes a (2k − 1)(1 + ǫ)-spanner of G with O(n1+1/k · (log k)/ǫ) edges, in expected O(|E|) centralized
time, or in O(log n · log∗ n · log Λ) depth and O(|E|) work.
The result of Theorem 2 can be used in conjunction with the scheme of [ES16] to devise an algorithm
that computes (2k − 1)(1 + ǫ)-spanners of size O(n1+1/k (log k/ǫ)1/ǫ), with lightness (i.e., weight of the
spanner divided by the weight of the MST of the input graph) O(k · n1/k (1/ǫ)2+1/k ), in expected time
O(|E| + min{n log n, |E|α(n)}), where α(·) is an inverse-Ackermann function. This improves a result
of [ES16] that provides spanners with the same stretch and lightness, but with more edges (specifically,
O((k + (1/ǫ)2+1/k )n1+1/k ), and using O(k · |E| + min{n log n, |E|α(n)}) time. Recently, consequently to
our work, Alstrup et al. [ADF+ 17] further improved these bounds. We thus omit the details of our argument
that provides the aforementioned bounds.
3 An Efficient Centralized Construction of Nearly-Additive Spanners and
Emulators
3.1 A Basic Variant of the Algorithm
In this section we present an algorithm for constructing (1 + ǫ, β)-spanners, which can be efficiently implemented in various settings. We start with the centralized setting. In this setting we present two variants of
our construction. The first variant presented in this section is somewhat simpler, while the second variant
presented in the next section provides better bounds.
Let G = (V, E) be an unweighted graph on n vertices, and let k ≥ 1, ǫ > 0 and 0 < ρ < 1/2 be
parameters. Unlike the algorithm of [EP04], our algorithm does not employ sparse partitions of [AP92].
The algorithm initializes the spanner H as an empty set, and proceeds in phases. It starts with setting
P̂0 = {{v} | v ∈ V } to be the partition of V into singleton clusters. The partition P̂0 is the input of phase 0
of our algorithm. More generally, P̂i is the input of phase i, for every index i in a certain appropriate range,
which we will specify in the sequel.
Throughout the algorithm, all clusters C that we will construct will be centered at designated centers rC . In particular, each singleton cluster C = {v} ∈ P̂0 is centered at v. We define Rad (C) =
max{dG(C) (rC , v) | v ∈ C}, and Rad (P̂i ) = maxC∈P̂i {Rad (C)}.
All phases of our algorithm except for the last one consist of two steps. Specifically, these are the
superclustering and the interconnection steps. The last phase contains only the interconnection step, and
the superclustering step is skipped. We also partition the phases into two stages. The first stage consists of
phases 0, l
1, . . . m, i0 = ⌊log(κρ)⌋, and the second stage consists of all the other phases i0 + 1, . . . , i1 where
i1 = i0 + κ+1
κρ − 2, except for the last phase ℓ = i1 + 1. The last phase will be referred to as the concluding
phase.
Each phase i accepts as input two parameters, the distance threshold parameter δi , and the degree parameter deg i . The difference between stage 1 and 2 is that in stage 1 the degree parameter grows exponentially,
while in stage 2 it is fixed. The distance threshold parameter grows in the same steady rate (increases by a
factor of 1/ǫ) all through the algorithm.
Next we describe the first stage of the algorithm. We start with describing its superclustering step. We
i
set deg i = n2 /κ , for all i = 0, 1, . . . , i0 . Let R0 = 0, and δi = (1/ǫ)i + 4 · Ri , where Ri is determined by
the following recursion: Ri+1 = δi + Ri = (1/ǫ)i + 5 · Ri . We will show that the inequality Rad (P̂i ) ≤ Ri
will hold for all i.
14
On phase i, each cluster C ∈ P̂i is sampled i.a.r. with probability 1/deg i . Let Si denote
S the set of
sampled clusters. We now conduct a BFS exploration to depth δi in G rooted at the set Si = C∈Si {rC }.
As a result, a forest Fi is constructed, rooted at vertices of Si . For a cluster center r ′ = rC ′ of a cluster
C ′ ∈ P̂i \Si such that r ′ is spanned by Fi , let rC be the root of the forest tree of Fi to which r ′ belongs. (The
vertex rC is by itself a cluster center of a cluster C ∈ Si .) The cluster C ′ becomes now superclustered in a
cluster Ĉ centered around the cluster C. (We also say that C ′ is associated with C. We will view association
as a transitive relation, i.e., if C ′ is associated with C and C ′′ is associated with C ′ , we will think of C ′′ as
associated with C as well.)
The cluster center rC of C becomes the new cluster center of Ĉ, i.e., rĈ = rC . The vertex set of the
new supercluster Ĉ is the union of the vertex set of C with the vertex sets of all clusters C ′ which are
superclustered into Ĉ. The edge set TĈ of the new cluster Ĉ contains the BFS spanning trees of all these
clusters, and, in addition, it contains shortest paths from the forest Fi between rC and each rC ′ as above. Ŝi
is the set of superclusters created by this process. We set P̂i+1 = Ŝi . All edges that belong to the edgeset of
one of these superclusters are now added to the spanner H.
For each supercluster Ĉ, we write Rad (Ĉ) = Rad (TĈ , rĈ ). Observe that Rad (P̂0 ) ≤ R0 = 0, and
Rad (Ŝ0 ) = max{Rad (Ĉ) | Ĉ ∈ Ŝ0 } ≤ δ0 + R0 = R1 = 1. More generally we have
Rad (Ŝi ) = max{Rad (Ĉ) | Ĉ ∈ Ŝi } ≤ δi + Rad (P̂i ) ≤ δi + Ri ≤ (1/ǫ)i + 5Ri = Ri+1 .
(7)
Denote by Ûi the set of clusters of P̂i which were not superclustered into clusters of Ŝi . In the interconnection step for i ≥ 1, every cluster center rC of a cluster C ∈ Ûi initiates a BFS exploration to depth 21 δi ,
i.e., half the depth of the exploration which took place in the superclustering step. For each cluster center
rC ′ for C ′ ∈ P̂i which is discovered by the exploration initiated in rC , the shortest path between rC and rC ′
is inserted into the spanner H. The first phase i = 0 is slightly different: the exploration depth is set to be
1, and we add an edge from {v} ∈ Û0 to all neighbors that are in Û0 . This completes the description of the
interconnection step.
Lemma 10. For any vertex v ∈ V , the expected number of explorations that visit v at the interconnection
step of phase i is at most deg i .
Proof. For i ≥ 1, assume that there are l clusters of P̂i whose centers are within distance δi /2 from v. If
at least one of them is sampled to Si , then no exploration will visit v (since in the superclustering phase the
sampled center will explore to distance δi , and thus will supercluster all these centers). The probability that
none of them is sampled is (1 − 1/deg i )l , in which case we get that l explorations visit v, so the expectation
is l · (1 − 1/deg i )l ≤ deg i (which holds for any l).
For i = 0, we note that we add an edge touching v iff none of its neighbors were sampled at phase 0
(as otherwise it would be clustered and thus not in Û0 ). The expected number of edges added is once again
l · (1 − 1/deg i )l ≤ deg i (here l is the number of neighbors).
We also note the following lemma for future use, its proof follows from a simple Chernoff bound.
Lemma 11. For any constant c > 1, with probability at least 1 − 1/nc−1 , for every vertex v ∈ V , at least
one among the deg i · c · ln n closest cluster centers rC ′ with C ′ ∈ P̂i to v is sampled, i.e., satisfies C ′ ∈ Si .
Observe that no vertex v ∈ V is explored by more than c · ln n · deg i explorations, with probability at
least 1 − n−(c−1) . Indeed, otherwise when i ≥ 1 there would be more than c · ln n · deg i cluster centers
rC of unsampled clusters C ∈ P̂i at pairwise distance at most δi . Applying Lemma 11 to any of them we
15
conclude that that the particular cluster C was superclustered by a nearby sampled cluster, i.e., C 6∈ Ûi ,
contradiction. In the case i = 0, we would have that v has at least c · ln n · deg i unsampled neighbors,
which occurs with probability at most n−c . Hence, by union-bound, every vertex v is explored by at most
c · ln n · deg i explorations, with probability at least 1 − n−(c−1) .
Lemma 10 suggests that the interconnection step of phase i can be carried out in expected O(|E| · deg i )
time. Clearly, the superclustering step can be carried out in just O(|E|) time, and thus the running time of
the interconnection step dominates the running time of phase i. In order to control the running time, we
terminate stage 1 and move on to stage 2 when i0 = ⌊log(κρ)⌋, so that deg i0 ≤ nρ .
Observe also that the superclustering step inserts into the spanner at most O(n) edges (because we
insert a subset of edges of Fi , and Fi is a forest), and by Lemma 10 the interconnection step inserts in
expectation at most O(|P̂i | · deg i · ((1/ǫ)i + Ri )) = O(|P̂i | · deg i · (1/ǫ)i ) edges. (We will soon show that
Ri = O((1/ǫ)i−1 ).) A more detailed argument providing an upper bound on the number of edges inserted
by the interconnection step will be given below.
Lemma 12. For all i, 1 ≤ i ≤ ℓ, for every pair of clusters C ∈ Ûi , C ′ ∈ P̂i at distance at most 21 (1/ǫ)i
from one another, a shortest path between the cluster centers of C and C ′ was inserted into the spanner H.
Moreover, for any pair {v} ∈ Û0 , {v ′ } ∈ P̂0 , such that e = (v, v ′ ) ∈ E, the edge e belongs to H.
Proof. We start with proving the first assertion of the lemma. For some index i, 1 ≤ i ≤ ℓ, and a pair
C ∈ Ûi , C ′ ∈ P̂i of clusters, let rC , rC ′ be the respective cluster centers. Then we have
dG (rC , rC ′ ) ≤ Rad (C) + dG (C, C ′ ) + Rad (C ′ )
≤ dG (C, C ′ ) + 2 · Ri
1
1
≤
(1/ǫ)i + 2 · Ri = δi ,
2
2
and so a shortest path between rC and rC ′ was inserted into the spanner H.
The second assertion of the lemma is guaranteed by the interconnection step of phase 0.
Next we analyze the radii of clusters’ collections P̂i , for i = 1, 2, . . ..
Lemma 13. For i = 0.1, . . . , ℓ, the value of Ri is given by
i−1
X
(1/ǫ)j · 5i−1−j .
Ri =
j=0
Proof. The proof is by induction of the index i. The basis (i = 0) is immediate as R0 = 0. For the induction
hypothesis, note that
Ri+1 = δi + Ri = (1/ǫ)i + 5 · Ri
i−1
X
= (1/ǫ)i + 5 · (1/ǫ)j · 5i−1−j
j=0
=
i
X
(1/ǫ)j · 5i−j ,
j=0
as required.
16
Observe that Lemma 13 implies that for ǫ < 1/10, we have Ri = 5i−1 ·
1/(5ǫ)i −1
1/(5ǫ)−1
≤
1
1−5ǫ
· (1/ǫ)i−1 ≤
2 · (1/ǫ)i−1 . Recall that P̂i = Ŝi−1 . Hence Rad (P̂i ) = Rad (Ŝi−1 ). By inequality (7), we have Rad (P̂i ) ≤
(1/ǫ)i−1 + 5 · Ri−1 = Ri , for all i = 0, 1, . . . , ℓ.
We analyze the number of clusters in collections P̂i in the following lemma.
Lemma 14. For i = 0, 1, . . . , i0 ,
|P̂i | ≤ 2 · n1−
with probability at least 1 − exp{−Ω(n
i
1− 2 κ−1
2i −1
κ
,
(8)
)}.
Proof. The probability that a vertex v ∈ V will be a center of a cluster in P̂i is
i
Thus the expected size of P̂i is n1−(2 −1)/κ , and by Chernoff bound,
Pr[|P̂i | ≥ 2E[|P̂i |]] ≤ exp{−Ω(E[|P̂i |])} = exp{−Ω(n1−
Since for ρ < 1/2 and i ≤ i0 = ⌊log(κρ)⌋, we have n1−
2i+1 −1
κ
i+1
1− 2 κ −1
Qi−1
2i −1
κ
j=0 1/deg j
i
= n−(2 −1)/κ .
)} .
≥ n1−2ρ = ω(log n), we conclude that
). Hence in particular, |P̂i0 +1 | = O(n1−ρ+1/κ ), whp.
whp for all 0 ≤ i ≤ i0 , |P̂i+1 | = |Ŝi | = O(n
The total expected running time of the first stage is at most
O(|E|)
i0
X
i=1
deg i = O(|E| · n
2i0
κ
) = O(|E| · nρ ) .
Since each superclustering step inserts at most O(n) edges into the spanner, the overall number of edges
inserted by the i0 superclustering steps of stage 1 is O(n · i0 ) = O(n · log(κρ)). The expected number of
edges added to the spanner by the interconnection step of phase i is at most
O(|P̂i | · deg i · (1/ǫ)i ) = O(n1−
2i −1
κ
2i
· (1/ǫ)log(κρ) · n κ ) = O(n1+1/κ · (1/ǫ)log(κρ) ) .
Next we describe stage 2 of the algorithm, i.e., phases i = i0 +1, i0 +2, . . . , i1 , where i1 = i0 +⌈ κ+1
κρ ⌉−2.
All these phases are executed with the same fixed degree parameter deg i = nρ . On the other hand, the
distance threshold keeps growing in the same steady rate as in stage 1, i.e., it is given by δi = (1/ǫ)i + 4Ri .
The sets P̂i0 +1 , P̂i0 +2 , . . . , P̂i1 on which phases i0 + 1, i0 + 2, . . . , i1 , respectively, operate are defined by
P̂i0 +i = Ŝi0 +i−1 , for 1 ≤ i ≤ i1 − i0 .
Lemma 13 keeps holding for these additional i1 − i0 phases, i.e.,
Rad (P̂i ) ≤ Ri ≤ 2 · (1/ǫ)i−1 .
(We assume all through that ǫ < 1/10.)
Also, for every pair of clusters C ∈ Ûi and C ′ ∈ P̂i which are at distance at most 12 (1/ǫ)i from one
another, their centers are interconnected in the spanner by a shortest path between them.
In addition, for every i ∈ [i0 , i1 ], the expected size of P̂i+1 is
E[|P̂i+1 |] = n ·
i
Y
j=0
1/deg j ≤ n1+1/κ−(i+1−i0 )ρ .
17
By Chernoff bound, for every such i, with probability at least 1 − exp{−Ω(nρ )}, we have
|P̂i+1 | = |Ŝi | ≤ 2 · n1+1/κ−ρ−(i−i0 )ρ .
Assuming that nρ = ω(1), we conclude that whp
|P̂i1 +1 | = |Ŝi1 | ≤ O(n1+1/κ−(i1 +1−i0 )ρ ) = O(n
⌉−1)ρ
1+1/κ−(⌈ κ+1
κρ
) = O(nρ ) .
(9)
log n
(For the assumption above to hold we will need to assume that ρ ≥ log
2 log n , say. We will show soon that
this assumption is valid in our setting.)
The time required to perform these ⌈ κ+1
κρ ⌉ − 2 ≤ 1/ρ additional phases is expected to be at most
ρ
O(|E| · deg i · (1/ρ)) = O(|E|n /ρ).
The final collection of clusters P̂i1 +1 is created by setting P̂i1 +1 = Ŝi1 .
We will next bound the expected number of edges inserted into the spanner during stage 2 of the algorithm. Each of the forests Fi , i ∈ [i0 + 1, i1 ], created during the superclustering steps contributes at
most n − 1 edges. The interconnection step of phase i + i0 contributes in expectation at most O(|P̂i+i0 | ·
deg i+i0 · (1/ǫ)i ) ≤ O(n1+1/κ−iρ · (1/ǫ)log(κρ)+i ) edges. Assuming that 1/ǫ < nρ /2, this becomes a
geometric progression, so the overall expected number of edges inserted into the spanner on stage 2 is
O(n1+1/κ · (1/ǫ)log(κρ) ). (We will show the validity of this assumption in the end of this section.)
Finally, we describe the concluding phase of the algorithm, i.e., phase ℓ = i1 + 1. In this phase we skip
the superclustering step (as the number of clusters is already sufficiently small), and proceed directly to the
interconnection step.
On this step each of the cluster centers rC for C ∈ P̂ℓ conducts a BFS exploration in G to depth
1
1
ℓ
δ
2 ℓ = 2 (1/ǫ) + 2Rℓ . (Essentially, we define Ûℓ = P̂ℓ , and perform the usual interconnection step of the
algorithm.) By (9), the number of edges inserted by this step into the spanner is whp only O(|P̂ℓ |2 ·(1/ǫ)ℓ ) =
O(|P̂i1 +1 |2 ·(1/ǫ)i1 +1 ) = O(n2ρ ·(1/ǫ)log(ρκ)+1/ρ ). Recall that we assume that ρ < 1/2. Hence this number
of edges is sublinear in n.
Hence the overall expected number of edges in the spanner is |H| = O(n1+1/κ · (1/ǫ)log(κρ) ). Observe
also that the running time of the last phase is O(|E| · nρ ). Hence the overall expected running time of the
ρ /ρ). It remains to analyze the stretch of the resulting spanner H.
algorithm is O(|E|n
Sℓ
Let Û = j=0 Ûj . Observe that every singleton cluster {v} ∈ P̂0 is associated with exactly one cluster
of Û , i.e., Û is a partition of V . Note that Rad (Û0 ) = 0, Rad (Û1 ) ≤ 1 = R1 , and for every j ∈ [ℓ], we have
Rad (Ûj ) ≤ Rad (P̂j ) ≤ Rj ≤ 2 · (1/ǫ)j−1 . Denote c = 2. Recall also (see Lemma 12) that for j ≥ 1, for
every pair of clusters C, C ′ ∈ Ûj at distance at most 21 (1/ǫ)j from one another, a shortest path between the
cluster centers of this pair of clusters in G was added to the spanner H. Moreover, neighboring clusters of
Û0 are also interconnected by a spanner edge.
Lemma 15. Consider a pair of indices 0 ≤ j < i ≤ ℓ, and a pair of neighboring (in G) clusters C ′ ∈ Ûj ,
C ∈ Ûi , and a vertex w′ ∈ C ′ and the center r of C. Then the spanner H contains a path of length at most
3Rad (Ûj ) + 1 + Rad (Ûi ) between w′ and r.
Proof. Let (z ′ , z) ∈ E ∩ (C ′ × C) be an edge connecting this pair of clusters. There exists a subcluster
C ′′ ⊆ C, C ′′ ∈ P̂j such that z ∈ C ′′ . Hence the interconnection step of phase j inserted a shortest path
π(r ′ , r ′′ ) in G between the cluster centers r ′ of C ′ and r ′′ of C ′′ into the spanner H. Note that the distance
between r ′ , r ′′ is at most 1 + 2Rad (P̂j ) ≤ 1 + 4(1/ǫ)j−1 < 1/2 · (1/ǫ)j , since we assume ǫ < 1/10.
Hence a path between w′ and r in H can be built by concatenating a path π(w′ , r ′ ) between w′ and r ′ in
18
the spanning tree T (C ′ ) of C ′ with the path π(r ′ , r ′′ ) in H, and with the path π(r ′′ , r) in the spanning tree
T (C) of C. (Note that both r ′′ and r belong to C.) Its length is at most
|π(w′ , r ′ )| + |π(r ′ , r ′′ )| + |π(r ′′ , r)| ≤ Rad (C ′ ) + (Rad (C ′ ) + 1 + Rad (C ′′ )) + Rad (C)
≤ 3Rad (Ûj ) + 1 + Rad (Ûi ) .
Now we are ready to analyze the stretch of our spanner.
Lemma 16. Suppose ǫ ≤ 1/10. Consider a pair of vertices u, v ∈ V . Fix a shortest path π(u, v) between
them in G, and suppose
that for some index i ∈ [0, ℓ], all vertices of π(u, v) are clustered in the set Û (i)
Si
(i)
defined by Û = j=0 Ûj . Then
dH (u, v) ≤ (1 + 16c · ǫ · i)dG (u, v) + 4
i
X
j=1
Rj · 2i−j .
Proof. The proof is by induction on i. For the induction basis i = 0, observe that all vertices of π(u, v) are
clustered in Û0 , and thus all edges of π(u, v) are inserted into the spanner on phase 0. Hence dH (u, v) =
dG (u, v).
For the induction step, consider first a pair of vertices x, y such that |π(x, y)| ≤ 21 (1/ǫ)i , and V (π(x, y)) ⊆
Û (i) . Let z1 and z2 be the leftmost and the rightmost Ûi -clustered vertices in π(x, y), if exist. (The case
when both these vertices exist is the one where the largest stretch is incurred; cf. [EP04].) Let C1 , C2 ∈ Ûi
be their respective clusters, i.e., z1 ∈ C1 , z2 ∈ C2 . Let w1 (respectively, w2 ) be the neighbor of z1 (resp.,
z2 ) on the subpath π(x, z1 ) (resp., π(z2 , y)) of π(x, y), and denote by C1′ and C2′ the respective clusters of
w1 and w2 . Observe that C1′ , C2′ ∈ Û (i−1) .
Denote r1 and r2 the cluster centers of C1 and C2 , respectively. The spanner H contains a path of length
at most dG (r1 , r2 ) between these cluster centers. Also, by Lemma 15, since C1′ and C1 are neighboring
clusters, the spanner H contains a path of length at most 3Rj + 1 + Ri ≤ 2Ri + 1 between w1 and r1 , and
a path of at most this length between r2 and w2 . (For ǫ < 1/10, 3Rj ≤ Ri , for all j < i.) Observe also
that the subpaths π(x, w1 ) and π(w2 , y) of π(x, y) have all their vertices clustered in Û (i−1) , and thus the
induction hypothesis is applicable to these subpaths.
Hence
dH (x, y) ≤ dH (x, w1 ) + dH (w1 , r1 ) + dH (r1 , r2 ) + dH (r2 , w2 ) + dH (w2 , y)
i−1
X
Rj · 2i−1−j + 2Ri + 1 + (dG (C1 , C2 ) + 2Ri )
≤ (1 + 16c · ǫ(i − 1))dG (x, w1 ) + 4
j=1
+2Ri + 1 + (1 + 16c · ǫ(i − 1)) · dG (w2 , y) + 4
i−1
X
j=1
Rj · 2i−1−j
= (1 + 16c · ǫ(i − 1)) · (dG (x, w1 ) + dG (w2 , y)) + dG (C1 , C2 ) + 4Ri + 2 + 8
i−1
X
j=1
Rj · 2i−1−j .
Note also that
dG (x, y) = dG (x, w1 ) + 1 + dG (z1 , z2 ) + 1 + dG (w2 , y) ≥ dG (x, w1 ) + dG (C1 , C2 ) + 2 + dG (w2 , y).
19
Hence
i−1
X
dH (x, y) ≤ (1 + 16c · ǫ(i − 1))dG (x, y) + 4Ri + 8
= (1 + 16c · ǫ(i − 1))dG (x, y) + 4
i
X
j=1
j=1
Rj · 2i−1−j
Rj · 2i−j .
Now consider a pair of vertices u, v such that all vertices of π(u, v) are clustered in Û (i) , without any
restriction on |π(u, v)|. We partition π(u, v) into segments π(x, y) of length exactly ⌊ 21 (1/ǫ)i ⌋, except
maybe one segment of possibly smaller length. Inequality (10) applies to all these segments. Hence
dH (u, v) ≤ (1 + 16c · ǫ(i − 1))dG (u, v) + 4
≤
1 + 16c · ǫ(i − 1) +
8
Pi
4
Pi
i
X
j=1
i
X
dG (u, v)
Rj · 2i−j
⌋
+
4
i
2 (1/ǫ) − 1
j=1
Rj · 2i−j ⌊ 1
i−j
j=1 Rj · 2
1
i
2 (1/ǫ) − 1
!
dG (u, v) + 4
i
X
j=1
Rj · 2i−j .
Rj 2i−j
j=1
≤ 16c · ǫ. Recall that for every j, we have Rj ≤ c · (1/ǫ)j−1 . Since
It remains to argue that (1/ǫ)
i −2
1/ǫ ≥ 10, the left-hand-side is at most
10c · ǫ
i
i
X
j−1
(1/ǫ)
j=1
i−j
·2
i−1
i
X
X
i−j
(2ǫ)h ≤ 16c · ǫ .
(2ǫ)
= 10c · ǫ
= 10c · ǫ
j=1
h=0
Observe that (as ǫ ≤ 1/10), we have
i
X
j=1
i−j
Rj · 2
i
i
X
X
1 j−1
j−1
i−j
i−1
(1/ǫ)
·2
= c·2
·
≤ c
2ǫ
j=1
j=1
= c · 2i−1
= c·ǫ·
1
2
i−1
X
j=0
1
2ǫ
j
(1/2ǫ)i − 1
(1/2ǫ) − 1
i−1 !
1
= O
.
ǫ
= c · 2i−1
· (1/ǫ)i − 2i−1
1
2 −ǫ
Note also that the condition of the last lemma holds with i = ℓ for every pair u, v ∈ V of vertices. Hence
Corollary 17. For every pair u, v ∈ V ,
dH (u, v) ≤ (1 + 16c · ℓ · ǫ)dG (u, v) + O((1/ǫ)ℓ−1 ) .
Recall that the spanner H contains, whp, |H| = O(n1+1/κ ·log n·(1/ǫ)log(κρ) +n·log n·(1/ǫ)log(κρ)+1/ρ )
ρ
edges, and the expected running time required to construct it is O(|E|
also that ℓ = i
1+1 ≤
· n /ρ). Recall
log κ+1/ρ
log
κ+1/ρ
. The
log(κρ) + 1/ρ + 1. Set now ǫ′ = 16c · ℓ · ǫ. We obtain stretch 1 + ǫ′ , O
ǫ′
condition ǫ < 1/10 translates now to ǫ′ ≤ 1.6c(log(κρ) + 1/ρ). We will replace it by a simpler stronger
condition ǫ ≤ 1.
20
Corollary 18. For any parameters 0 < ǫ ≤ 1, κ ≥ 2, and ρ > 0, and any n-vertex unweighted graph
log κ
·
G = (V, E), our algorithm computes a (1+ǫ, β)-spanner with expected number of edges O log κ+1/ρ
ǫ
n1+1/κ , in expected time O(|E| · nρ /ρ), where
O(log κ + 1/ρ) log κ+1/ρ
β=
.
ǫ
A particularly useful setting of parameters is ρ = 1/ log κ. Then we get a spanner with expected
log κ
2 log κ
1
O logǫ κ
· n1+1/κ edges, in time O(|E| · n log κ · log κ), and β = O logǫ κ
.
We remark that it makes no sense to set ρ < 1/κ, as the resulting parameters will be strictly worse than
when ρ = 1/κ. Also our assumptions that ρ > log log n/(2 log n) and 16c · ℓ/ǫ < nρ /2 are justified, as
otherwise we get β ≥ n, so a trivial spanner will do.
3.2 An Improved Variant of the Algorithm
In this section we show that the leading coefficient O((log κ + 1/ρ)/ǫ)log κ of n1+1/κ in the size of the
spanner can be almost completely eliminated at essentially no price. We also devise here yet sparser constructions of emulators.
For i = 0, 1, . . . , ℓ, denote by Ni = |P̂i | the expected number of clusters which take part in phase i.
′ i
Recall also that the interconnection step of the ith phase contributes Ni · deg i · cǫℓ edges in expectation,
where ℓ is the total number of steps, and c′ is a universal constant. Note that the contribution of the interconnection step dominates the contribution of the superclustering step in the current variant of the algorithm,
and it will still be the case after the modification that we will now introduce. Hence we will now focus on
decreasing the number of edges contributed by the interconnection steps.
We keep the structure of the algorithm intact, and have the values of distance thresholds δi unchanged.
The only change is in the degree sequence deg 0 , deg 1 , . . . of degree parameters used in phases 0, 1, . . .,
respectively. Next, we describe our new setting of these parameters for stage 1 of the algorithm (i.e., phases
i, 1 ≤ i ≤ i0 ). In the case that κ ≥ 16 let a = log log κ, otherwise, when κ < 16, let a = 2. Define
i
i0 = min{⌊log(aκρ)⌋, ⌊κρ⌋}, and for i = 0, 1, . . . , i0 let deg i = n(2 −1)/(aκ)+1/κ . We now have that for
i ≤ i0 + 1,
i−1
Y
2i −1−i
i
1/deg j = n1− aκ − κ ,
Ni = n
j=0
aκρ−1−(i0 +1)
i0 +1
− κ
aκ
and in particular, when i0 = ⌊log(aκρ)⌋ we have Ni0 +1 ≤ n1−
≤ n1−ρ (since a ≥ 2 and
i0 +1
i0 ≥ 1). Whenever i0 = ⌊κρ⌋ we also have Ni0 +1 ≤ n1− κ ≤ n1−ρ . Additionally, we always have
Ni · deg i = n1−
2i −1−i
− κi
aκ
·n
We restrict ourselves to the case that
2i −1
+ κ1
aκ
i
= n1+ aκ −
i−1
κ
.
1
c′ ℓ
(10)
≤ n 2κ /2 ,
ǫ
c0 ·log n
, for a sufficiently small constant c0 . Now the expected number of edges
which holds whenever κ ≤ log(ℓ/ǫ)
inserted at phase i ≤ i0 is at most
!i
′ i
1/(2κ)
i−1
i
1
cℓ
n
≤ n1+ 2κ − κ ·
Ni · deg i ·
= n1+ κ /2i .
(11)
ǫ
2
21
Thus the total expected number of edges inserted in the first stage is O(n1+1/κ ). The second stage proceeds
by setting deg i0 +1 = nρ/2 , and in all subsequent phases i0 + i, with i = 2, 3, . . . , i1 − i0 , we have deg i = nρ
as before. The ”price” for reducing the degree in the first phase of stage two is that the number of phases i1
may increase by an additive 1. It follows that Ni0 +1 · deg i0 +1 ≤ n1−ρ/2 . For i ≥ 2, at phase i0 + i we have
Ni0 +i ≤ n1−3ρ/2−(i−2)ρ . We set i1 = ⌊1/ρ⌋, so that Ni1 +1 ≤ nρ/2 , and whp we have that Ni1 +1 ≤ 2nρ/2 .
We calculate
Ni0 +i · deg i0 +i ≤ n1−ρ/2−(i−2)ρ .
Note that i0 /(2κ) ≤ ρ/2, which holds since i0 ≤ ⌊κρ⌋. Hence the condition (10) implies that
i0
≤ n 2κ ≤ nρ/2 . The total expected number of edges inserted at phase i0 + 1 is at most
(c′ ℓ/ǫ)i0
Ni0 +1 · deg i0 +1 ·
c′ ℓ
ǫ
i0 +1
≤ n1−ρ/2 · nρ/2 ·
c′ ℓ
≤ n1+1/κ .
ǫ
The expected contribution of phase i0 + i for i ≥ 2 is at most
Ni0 +i · deg i0 +i ·
c′ ℓ
ǫ
i0 +i
≤ n1−ρ/2−(i−2)ρ · nρ/2 · ni/(2κ) /2i ≤ n1+1/κ /2i ,
(12)
where the last inequality uses that ρ ≥ 1/κ (which we may assume w.l.o.g). This implies that the expected
number of edges in all these ⌊1/ρ⌋ phases is O(n1+1/κ ).
Ω(log n)
c0 ·log n
≥
The upper bound on κ under which this analysis was carried out is log(ℓ/ǫ)
.9
log(1/ǫ)+log(1/ρ)+log (3) n
We summarize this discussion with the following theorem.
c·log n
log(1/ǫ)+log(1/ρ)+log (3) n
= O( 1ǫ (log κ +
Theorem 3. For any unweighted graph G = (V, E) with n vertices , 0 < ǫ < 1/10, 2 ≤ κ ≤
for a constant c, and 1/κ ≤ ρ < 1/2, our algorithm computes a (1 + ǫ, β)-spanner with β
(3)
1/ρ))log κ+1/ρ+max{1,log κ} and expected number of edges O(n1+1/κ ). The expected running time is
O(|E| · nρ /ρ).
Note that the sparsest this spanner can be is O(n log log n), and at this level of sparsity its β =
O(log log n + 1/ρ)log log n+1/ρ . (To get this bound we set ǫ > 0 to be an arbitrary small constant, and
n
κ = c0 log
.)
log(3) n
This is sparser than the state-of-the-art efficiently-computable
sparsest (1+ ǫ, β)-spanner due to [Pet10],
√
1+ 5
φ
which has O(n(log log n) ) edges, where φ = 2 is the golden ratio. Moreover, this spanner has a
smaller β than the one of [Pet10] in its sparsest level. Denoting the latter as βP et , it holds that βP et ≈
O(log κ)1.44 log κ+1/ρ , i.e., for every setting of the time parameter ρ, the exponent of our β is smaller than
that of βP et .
3.2.1
Sparse Emulator
Finally, we note that if one allows an emulator instead of spanner, then we can decrease the size all the
way to O(n) when κ = log n. To achieve this, we insert single ”virtual” edges instead of every path (of
length (c′ ℓ/ǫ)i ) between every pair of cluster centers that we choose to interconnect on phase i, for every
i. Analogously, in the superclustering step we also form a supercluster around a center rC of a cluster C
by adding virtual edges (rC , rC ′ ) for each cluster C ′ associated with C. The weight of each such edge is
9
We denote log(k) n as the iterated logarithm function, e.g. log(3) n = log log log n.
22
defined by ω(rC , rC ′ ) = dG (rC , rC ′ ). The condition (10) was required to obtain converging sequences at
(11) and (12), but without the (c′ ℓ/ǫ)i terms, the number of edges already forms a converging sequence at
each stage.
Moreover, one can also use for emulators a shorter degree sequence than the one we used for spanners,
and as a result to save the additive term of log(3) κ in the exponent of β. Specifically, one can set deg i =
i
Q
2i
i
1− 2 κ−1
·
n κ /22 −1 , for each i = 0, 1, . . . , i0 = ⌊log(κρ)⌋. As a result we get Ni = n · i−1
j=0 1/deg j = n
i
22 −1−i , and thus the expected number of edges inserted at phase i ≤ i0 is at most
Ni · deg i = n1+1/κ /2i .
As before, when the first stage concludes, we run one phase with deg i0 +1 = nρ/2 , and all subsequent phases
with deg i = nρ . To bound the expected number of edges added at phase i0 + 1 we need to note that
i +1
22 0 ≤ 22κρ ≤ nρ/2 as long as κ ≤ (log n)/4. (The latter can be assumed without affecting any of the
2i0 +1 −1
i +1
parameters by more than a constant factor). It follows that Ni0 +1 · deg i0 +1 = n1− κ · 22 0 −1−(i0 +1) ·
nρ/2 ≤ n1+1/κ . In the remaining phases Ni0 +i ≤ n1+1/κ−(i−1)ρ for i ≥ 2, and the contribution of these
phases is a converging sequence. We conclude the discussion with the following theorem.
Theorem 4. For any unweighted graph G = (V, E) with n vertices, and for any parameters 0 < ǫ ≤
1, 2 ≤ κ ≤ (log n)/4, 1/κ ≤ ρ < 1/2, our algorithm computes a (1 + ǫ, β)-emulator with β =
log κ+1/ρ
and expected number of edges O(n1+1/κ ). The expected running time is O(|E| ·
O log κ+1/ρ
ǫ
nρ /ρ).
log log n+1/ρ
In particular, the algorithm produces a linear-size (1+ǫ, β)-emulator with β = O log logǫn+1/ρ
within this running time.
3.3 Distributed and Streaming Implementations
In this section we provide efficient distributed and streaming algorithms for constructing sparse (1 + ǫ, β)spanners. The distributed algorithm works in the CONGEST model.
To implement phase 0, each vertex selects itself into S0 with probability n−1/κ , i.a.r.. In distributed
model vertices of S0 send messages to their neighbors. Each vertex u that receives at least one message,
picks an origin v ∈ S0 of one of these messages, and joins the cluster centered at v. It also sends negative
acknowledgements to all its other neighbors from S0 . All unclustered vertices z insert all edges incident on
them into the spanner.
It is also straightforward to implement this in O(1) passes in the streaming model.
Each consecutive phase is now also implemented in a straightforward manner, i.e., BFS explorations
to depth δi in the superclustering steps are implemented via broadcasts and convergecasts in the distributed
model, and by δi passes in the streaming model. In the interconnection steps, however, we need to implement
many BFS explorations which may explore the same vertices. However, by Lemma 11 every vertex is
explored on phase i by up to O(degi · log n) explorations
P whp, it follows that in distributed setting this step
requires, whp, O(deg i · log n · δi ) time. Also note that i δi = O(β).
In the streaming model we have two possible tradeoffs. The first uses expected O(n · deg i ) space to
maintain for each vertex v the BFS parents and distance estimates, and requires just δi passes. To see that
such space suffices, recall that the expected number of explorations which visit any vertex v is at most deg i ,
by Lemma 10. Whp, the space is O(n · deg i · log n) = O(n1+ρ · log n).
23
The second option in the streaming algorithm is to divide the interconnection step of phase i to c ·
deg i · log n subphases, for a sufficiently large constant c. On each subphase each exploration source, which
was not sampled on previous subphases, samples itself i.a.r. with probability 1/deg i . Then the sampled
exploration sources conduct BFS explorations to depth δi /2. For every vertex v, the expected number of
explorations that traverse it on each subphase is O(log n). Moreover, by Chernoff’s inequality, whp, no
vertex v is ever traversed by more than O(log n) explorations. (Here we take a union-bound on all vertices,
all phases, and all subphases. The bad events are that some vertex is traversed by more than twice its
expectation explorations on some subphase.) Hence each subphase requires O(δi ) passes, and whp, the
space requirement is O(n log n), plus the size of the spanner. After c · deg i · log n subphases, whp, each
exploration source is sampled on at least one of the subphases, and so the algorithm performs all the required
explorations.
Finally, the stretch analysis of distributed and streaming variants of our algorithm remains the same as
in the centralized case.
Hence we obtain the following distributed and streaming analogues of Theorem 3.
Theorem 5. For any unweighted graph G = (V, E) with n vertices , 0 < ǫ < 1/10, 2 ≤ κ ≤
c·log n
log(1/ǫ)+log(1/ρ)+log (3) n
for a constant c, and 1/κ ≤ ρ < 1/2, our distributed algorithm (CONGEST model) computes a (1 + ǫ, β)(3)
spanner with β = O( 1ǫ (log κ + 1/ρ))log κ+1/ρ+max{1,log κ} and expected number of edges O(n1+1/κ ).
The required number of rounds is whp O(nρ /ρ · β · log n).
Our streaming algorithm computes a spanner with the above properties, in either: O(n log n + n1+1/κ )
expected space and O(nρ /ρ · log n · β) passes, or using O(n1+ρ · log n) space, whp, and O(β) passes.
The streaming algorithm described above can also be modified to provide a (1 + ǫ, β)-emulator as in
Theorem 4, within the same pass and space complexities.
4 Applications
In this section we describe applications of our improved constructions of spanners and emulators to computing approximate shortest paths for a set S × V of vertex pairs, for a subset S ⊆ V of designated sources.
4.1 Centralized Setting
We start with the centralized setting. Here our input graph G = (V, E) is unweighted, and we construct a
(1 + ǫ, β)-emulator H of G with O(n1+1/κ ) edges, in expected time O(|E| · nρ /ρ), where
β=O
log κ + 1/ρ
ǫ
log κ+1/ρ
.
(13)
Observe that all edge weights in H are integers in the range [1, β]. We round all edge weights up to the
closest power of 1 + ǫ. Let H ′ be the resulting emulator. Note that for any pair u, v of vertices, we have
dG (u, v) ≤ dH (u, v) ≤ dH ′ (u, v) ≤ (1 + ǫ)dH (u, v)
≤ (1 + ǫ)2 dG (u, v) + (1 + ǫ)β .
For a sufficiently small ǫ, (1 + ǫ)2 ≤ 3ǫ, and we rescale ǫ′ = 3ǫ. As a result the constant factor hidden by the
O-notation in the basis of β’s exponent grows, but other than that H ′ has all the properties of the emulator
24
H. Also, it employs only
log β
(log 1/ǫ + log(log κ + 1/ρ)) · (log κ + 1/ρ)
t=O
=O
ǫ
ǫ
different edge weights. Hence a single-source shortest path computation in H can be performed in O(|H| +
n log t) = O(n1+1/κ + n(log(1/ǫ) + log(log κ + 1/ρ))) time [OMSW10]. (See also [KMP11], Section 5.)
Hence computing S × V (1 + ǫ, β)-approximate shortest distances requires O(|E| · nρ /ρ + |S|(n1+1/κ +
n(log(1/ǫ) + log(log κ + 1/ρ)))) time.
Theorem 6. For any n, and for any parameters 0 < ǫ ≤ 1, 2 ≤ κ ≤ (log n)/4, 1/κ ≤ ρ ≤ 1/2, and any
n-vertex unweighted graph G = (V, E) with a set S ⊆ V , our algorithm computes (1 + ǫ, β)-approximate
S × V shortest distances in the centralized model in expected O(|E| · nρ /ρ + |S|(n1+1/κ + n(log(1/ǫ) +
log(log κ + 1/ρ)))) time, where β is given by (13).
If one is interested in actual paths rather than just in distances, then one can use our (1 + ǫ, β)-spanner
with
(3)
log κ + 1/ρ log κ+1/ρ+max{1,log κ}
β=O
,
(14)
ǫ
O(log n)
and O(n1+1/κ ) edges, but restricting κ ≤ log(1/ǫ)+log(1/ρ)+log
(3) . After computing the spanner H with
n
these properties, we conduct BFS explorations on H originated at each vertex of S. The overall running
time becomes O(|E| · nρ /ρ + |S| · n1+1/κ ).
Corollary 19. For any n, and for any parameters 0 < ǫ ≤ 1, κ ≤
O(log n)
,
log(1/ǫ)+log(1/ρ)+log (3) n
1/κ ≤ ρ ≤ 1/2,
and any n-vertex unweighted graph G = (V, E), our algorithm computes (1 + ǫ, β)-approximate S × V
shortest paths in the centralized model in expected O(|E| · nρ /ρ + |S| · n1+1/κ ) time, with β given by (14).
A useful setting of parameters is ρ = log1 κ . Then the running time of our algorithms from Theorem 6
and Corollary 19 become respectively O(|E| · n1/ log κ · log κ + |S|(n1+1/κ + n(log 1/ǫ + log log κ))) and
O(|E| · n1/ log κ · log κ + |S| · n1+1/κ ) (we note that the former has smaller β given by (13), while the latter
has slightly larger β given by (14)).
The algorithm of Corollary 19 always outperforms the algorithm which can be derived by using the
spanner of [Pet10] within the same scheme. Specifically, the running time of that algorithm is at least
Õ(|E|nρ ) + O(|S| · (n1+1/κ + n( logǫ κ )φ )), and the additive error βP et there is given by
βP et = O
log κ + 1/ρ
ǫ
√
logφ κ+1/ρ
,
where φ = 1+2 5 is the golden ratio. So βP et is typically polynomially larger than the additive error β in
our algorithm, e.g., for ρ = 1/ log κ we have βP et ≈ β 1.22 . (Setting ρ to be smaller than 1/ log κ makes less
sense, because then the additive errors β and βP et deteriorate. At any rate, as ρ tends to 0, the two estimates
approach each other.) Also, in the bound of [Pet10] there is a term of |S| · n · ( logǫ κ )φ )), which does not
occur in our construction.
25
4.2 Streaming Setting
In this section we show how efficient constructions of spanners and emulators for an unweighted graph G
in the streaming setting can be used for efficient computation of approximate shortest paths.
First, one can use O(nρ /ρ · log n · β) passes and expected space O(n1+1/κ + n · log n) to construct an
(1 + ǫ, β)-emulator H with β = (O(log κ + 1/ρ)/ǫ)log κ+1/ρ . One can now compute V × V (1 + ǫ, β)approximate shortest distances in G by computing exact V ×V shortest distances in H, using the same space
and without additional passes. (Observe that we do not store the output, as its size is larger than the size of
H.) In particular, one can use here space of O(n · log n), and have β = (O(log log n + 1/ρ)/ǫ)log log n+1/ρ .
O(
log n
)
1
log log n passes, β = O( log log n )2 log log n .
It is also possible to set here ρ = log log
n , and obtain 2
ǫ
Another option is to use space O(n1+ρ ·log n), whp, and O(β) passes for constructing the same emulator
H. Again given the emulator we can compute V × V approximate shortest distances in G by computing
shortest distances in H offline.
Corollary 20. For any n and any parameters 0 < ǫ ≤ 1, 2 ≤ κ ≤ (log n)/4, 1/κ ≤ ρ ≤ 1/2, and any
n-vertex unweighted graph G, our streaming algorithm computes (1 + ǫ, β)-approximate shortest distances
for V × V with β given by (13). It uses in expectation either O(nρ /ρ · log n · β) passes and expected space
O(n1+1/κ + n · log n) or O(β) passes and space O(n1+ρ · log n), whp.
If the actual paths rather than just distances are needed, then we compute a (1 + ǫ, β)-spanner H with
β given by (14) and with expected O(n1+1/κ ) edges (with the restriction on κ as above). Then we compute
V × V (1 + ǫ, β)-approximate shortest paths in H offline, using space O(H|).
Corollary 21. For any n, κ, ǫ, ρ and G as in Corollary 19, a variant of our streaming algorithm computes
(1 + ǫ, β)-approximate shortest paths for V × V with β given by (14). It uses either O(nρ /ρ · log n · β)
passes and expected space O(n1+1/κ + n · log n) or O(β) passes and space O(n1+ρ · log n), whp.
If one is interested only in S × V paths or distances, then it is possible to eliminate the additive term of
β by using O(|S| · β/ǫ) additional passes. These passes are used to compute exactly distances between pairs
(s, v) ∈ S × V , with dG (s, v) ≤ β/ǫ. The overall number of passes becomes O(|S| · β/ǫ + nρ /ρ · log n · β),
|S|
log log n
and space O(n1+1/κ + n · log n). Whenever |S| ≥ n1/κ log n, we can set ρ = log
log n − log n , and obtain
the following corollary.
Corollary 22. For any n, ǫ, ρ, κ and G as in Corollary 20, and any set S ⊆ V of size at least |S| ≥
n1/κ log n, a variant of our streaming algorithm computes (1 + ǫ)-approximate shortest distances for S × V
in expected
log n
log κ+
log |S|−log log n
1
|S|
log n
(15)
·O
log κ +
ǫ
ǫ
log |S|−log log n
passes, and space O(n1+1/κ + n · log n). To compute actual paths we have similar complexities10 , but one
needs to restrict κ as in Corollary 19.
Observe that for a constant ǫ > 0 and |S| = nΩ(1) , we can take a constant κ, so the number of passes
is O(|S|). One can also get for |S| = 2Ω(log n/ log log n) , a streaming algorithm for computing (1 + ǫ)approximate shortest paths for S × V by setting κ = c log n/ log log log n, and obtaining O(|S|/ǫ) ·
n 2 log log n
O( log log
)
passes and space O(n log n).
ǫ
10
Though there is an additional additive term of log(3) κ in the exponent in (15).
26
References
[AB16]
Amir Abboud and Greg Bodwin. The 4/3 additive spanner exponent is tight. In Proceedings of
the 48th Annual ACM SIGACT Symposium on Theory of Computing, STOC 2016, Cambridge,
MA, USA, June 18-21, 2016, pages 351–361, 2016.
[ABCP93] Baruch Awerbuch, Bonnie Berger, Lenore Cowen, and David Peleg. Near-linear cost sequential
and distribured constructions of sparse neighborhood covers. In 34th Annual Symposium on
Foundations of Computer Science, Palo Alto, California, USA, 3-5 November 1993, pages
638–647, 1993.
[ABN11]
Ittai Abraham, Yair Bartal, and Ofer Neiman. Advances in metric embedding theory. Advances
in Mathematics, 228(6):3026 – 3126, 2011.
[ABP17]
Amir Abboud, Greg Bodwin, and Seth Pettie. A hierarchy of lower bounds for sublinear
additive spanners. In Proceedings of the Twenty-Eighth Annual ACM-SIAM Symposium on
Discrete Algorithms, pages 568–576, 2017.
[ACIM99]
Donald Aingworth, Chandra Chekuri, Piotr Indyk, and Rajeev Motwani. Fast estimation of
diameter and shortest paths (without matrix multiplication). SIAM J. Comput., 28(4):1167–
1181, 1999.
[ADD+ 93] I. Althöfer, G. Das, D. Dobkin, D. Joseph, and J. Soares. On sparse spanners of weighted
graphs. Discrete Comput. Geom., 9:81–100, 1993.
[ADF+ 17] Stephen Alstrup, Soren Dahlgaard, Arnold Filtser, Morten Stockel, and Christian Wulff-Nilsen.
Personal communication, 2017.
[AHL02]
Noga Alon, Shlomo Hoory, and Nathan Linial. The moore bound for irregular graphs. Graphs
and Combinatorics, 18(1):53–57, 2002.
[AN12]
Ittai Abraham and Ofer Neiman. Using petal-decompositions to build a low stretch spanning
tree. In STOC, pages 395–406, 2012.
[AP92]
B. Awerbuch and D. Peleg. Routing with polynomial communication-space tradeoff. SIAM J.
Discrete Mathematics, 5:151–162, 1992.
[Awe85]
B. Awerbuch. Complexity of network synchronization. J. ACM, 4:804–823, 1985.
[Bar96]
Yair Bartal. Probabilistic approximations of metric spaces and its algorithmic applications. In
FOCS, pages 184–193, 1996.
[Bar98]
Yair Bartal. On approximating arbitrary metrices by tree metrics. In Proceedings of the Thirtieth Annual ACM Symposium on Theory of Computing, STOC ’98, pages 161–168, New York,
NY, USA, 1998. ACM.
[Bar04]
Yair Bartal. Graph decomposition lemmas and their role in metric embedding methods. In
Algorithms - ESA 2004, 12th Annual European Symposium, Bergen, Norway, September 1417, 2004, Proceedings, pages 89–97, 2004.
27
[Bas08]
S. Baswana. Streaming algorithm for graph spanners - single pass and constant processing
time per edge. Inf. Process. Lett., 106(3):110–114, 2008.
[BCE05]
Béla Bollobás, Don Coppersmith, and Michael Elkin. Sparse distance preservers and additive
spanners. SIAM J. Discrete Math., 19(4):1029–1055, 2005.
[BGK+ 14] Guy E. Blelloch, Anupam Gupta, Ioannis Koutis, Gary L. Miller, Richard Peng, and Kanat
Tangwongsan. Nearly-linear work parallel SDD solvers, low-diameter decomposition, and
low-stretch subgraphs. Theor. Comp. Sys., 55(3):521–554, October 2014.
[BKMP10] Surender Baswana, Telikepalli Kavitha, Kurt Mehlhorn, and Seth Pettie. Additive spanners
and (alpha, beta)-spanners. ACM Transactions on Algorithms, 7(1):5, 2010.
[BR10]
Ajesh Babu and Jaikumar Radhakrishnan. An entropy based proof of the moore bound for
irregular graphs. CoRR, abs/1011.1058, 2010.
[BS03]
S. Baswana and S. Sen. A simple linear time algorithm for computing a (2k − 1)-spanner
of O(n1+1/k ) size in weighted graphs. In Proceedings of the 30th International Colloquium
on Automata, Languages and Programming, volume 2719 of LNCS, pages 384–396. Springer,
2003.
[BS07]
Surender Baswana and Sandeep Sen. A simple and linear time randomized algorithm for
computing sparse spanners in weighted graphs. Random Struct. Algorithms, 30(4):532–563,
2007.
[BW15]
Gregory Bodwin and Virginia Vassilevska Williams. Very sparse additive spanners and emulators. In Proceedings of the 2015 Conference on Innovations in Theoretical Computer Science,
ITCS 2015, Rehovot, Israel, January 11-13, 2015, pages 377–382, 2015.
[Che13]
Shiri Chechik. New additive spanners. In Proceedings of the Twenty-Fourth Annual ACMSIAM Symposium on Discrete Algorithms, SODA ’13, pages 498–512, Philadelphia, PA, USA,
2013. Society for Industrial and Applied Mathematics.
[Coh99]
E. Cohen. Fast algorithms for t-spanners and stretch-t paths. SIAM J. Comput., 28:210–236,
1999.
[Coh00]
Edith Cohen. Polylog-time and near-linear work approximation scheme for undirected shortest
paths. J. ACM, 47(1):132–166, 2000.
[DGPV08] Bilel Derbel, Cyril Gavoille, David Peleg, and Laurent Viennot. On the locality of distributed
sparse spanner construction. In Proceedings of the Twenty-Seventh Annual ACM Symposium
on Principles of Distributed Computing, PODC 2008, Toronto, Canada, August 18-21, 2008,
pages 273–282, 2008.
[DHZ00]
D. Dor, S. Halperin, and U. Zwick. All-pairs almost shortest paths. SIAM J. Comput., 29:1740–
1759, 2000.
[DMP+ 03] D. Dubhashi, A. Mei, A. Panconesi, J. Radhakrishnan, and A. Srinivisan. Fast distributed
algorithm for (weakly) connected dominating sets and linear-size skeletons, 2003.
28
[DMZ06]
Bilel Derbel, Mohamed Mosbah, and Akka Zemmari. Fast distributed graph partition and
application. In 20th International Parallel and Distributed Processing Symposium (IPDPS
2006), Proceedings, 25-29 April 2006, Rhodes Island, Greece, 2006.
[EEST05]
Michael Elkin, Yuval Emek, Daniel A. Spielman, and Shang-Hua Teng. Lower-stretch spanning trees. In STOC, pages 494–503, 2005.
[Elk04]
M. Elkin. An unconditional lower bound on the time-approximation tradeoff of the minimum
spanning tree problem. In Proc. of the 36th ACM Symp. on Theory of Comput. (STOC 2004),
pages 331–340, 2004.
[Elk05]
Michael Elkin. Computing almost shortest paths. ACM Trans. Algorithms, 1(2):283–323,
2005.
[Elk07a]
Michael Elkin. A near-optimal distributed fully dynamic algorithm for maintaining sparse
spanners. In Proceedings of the Twenty-Sixth Annual ACM Symposium on Principles of Distributed Computing, PODC 2007, Portland, Oregon, USA, August 12-15, 2007, pages 185–194,
2007.
[Elk07b]
Michael Elkin. Streaming and fully dynamic centralized algorithms for constructing and maintaining sparse spanners. In Automata, Languages and Programming, 34th International Colloquium, ICALP 2007, Wroclaw, Poland, July 9-13, 2007, Proceedings, pages 716–727, 2007.
[EN16a]
Michael Elkin and Ofer Neiman. Distributed strong diameter network decomposition. In
Proceedings of the 2016 ACM Symposium on Principles of Distributed Computing, PODC
2016, Chicago, IL, USA, July 25-28, 2016, pages 211–216, 2016.
[EN16b]
Michael Elkin and Ofer Neiman. Hopsets with constant hopbound, and applications to approximate shortest paths. In IEEE 57th Annual Symposium on Foundations of Computer Science,
FOCS 2016, 9-11 October 2016, Hyatt Regency, New Brunswick, New Jersey, USA, pages
128–137, 2016.
[EN17]
Michael Elkin and Ofer Neiman. Efficient algorithms for constructing very sparse spanners and
emulators. In Proceedings of the Twenty-Eighth Annual ACM-SIAM Symposium on Discrete
Algorithms, pages 652–669, 2017.
[EP04]
Michael Elkin and David Peleg. (1+epsilon, beta)-spanner constructions for general graphs.
SIAM J. Comput., 33(3):608–631, 2004.
[EP15]
Michael Elkin and Seth Pettie. A linear-size logarithmic stretch path-reporting distance oracle
for general graphs. In Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on
Discrete Algorithms, SODA 2015, San Diego, CA, USA, January 4-6, 2015, pages 805–821,
2015.
[ES16]
Michael Elkin and Shay Solomon. Fast constructions of lightweight spanners for general
graphs. ACM Trans. Algorithms, 12(3):29:1–29:21, 2016.
[EZ06]
Michael Elkin and Jian Zhang. Efficient algorithms for constructing (1+epsilon, beta)-spanners
in the distributed and streaming models. Distributed Computing, 18(5):375–385, 2006.
29
[FKM+ 05] J. Feigenbaum, S. Kannan, A. McGregor, S. Suri, and J. Zhang. Graph distances in the streaming model: The value of space. In Proc. of the ACM-SIAM Symp. on Discrete Algorithms,
pages 745–754, 2005.
[FS16]
Arnold Filtser and Shay Solomon. The greedy spanner is existentially optimal. In Proceedings
of the 2016 ACM Symposium on Principles of Distributed Computing, PODC 2016, Chicago,
IL, USA, July 25-28, 2016, pages 9–17, 2016.
[HZ96]
S. Halperin and U. Zwick. Linear time deterministic algorithm for computing spanners for
unweighted graphs, 1996. manuscript.
[KMP11]
Ioannis Koutis, Gary L. Miller, and Richard Peng. A nearly-m log n time solver for sdd
linear systems. In Proceedings of the 2011 IEEE 52Nd Annual Symposium on Foundations of
Computer Science, FOCS ’11, pages 590–598, Washington, DC, USA, 2011. IEEE Computer
Society.
[LS93]
N. Linial and M. Saks. Decomposing graphs into regions of small diameter. Combinatorica,
13:441–454, 1993.
[MPVX15] Gary L. Miller, Richard Peng, Adrian Vladu, and Shen Chen Xu. Improved parallel algorithms for spanners and hopsets. In Proceedings of the 27th ACM Symposium on Parallelism in
Algorithms and Architectures, SPAA ’15, pages 192–201, New York, NY, USA, 2015. ACM.
[MPX13]
Gary L. Miller, Richard Peng, and Shen Chen Xu. Parallel graph decompositions using random
shifts. In 25th ACM Symposium on Parallelism in Algorithms and Architectures, SPAA ’13,
Montreal, QC, Canada - July 23 - 25, 2013, pages 196–203, 2013.
[OMSW10] James B. Orlin, Kamesh Madduri, K. Subramani, and M. Williamson. A faster algorithm
for the single source shortest path problem with few distinct positive lengths. J. of Discrete
Algorithms, 8:189–198, June 2010.
[Pel99]
David Peleg. Proximity-preserving labeling schemes and their applications. In GraphTheoretic Concepts in Computer Science, 25th International Workshop, WG ’99, Ascona,
Switzerland, June 17-19, 1999, Proceedings, pages 30–41, 1999.
[Pel00]
David Peleg. Distributed Computing: A Locality-sensitive Approach. Society for Industrial
and Applied Mathematics, Philadelphia, PA, USA, 2000.
[Pet09]
Seth Pettie. Low distortion spanners. ACM Transactions on Algorithms, 6(1), 2009.
[Pet10]
Seth Pettie. Distributed algorithms for ultrasparse spanners and linear size skeletons. Distributed Computing, 22(3):147–166, 2010.
[Pet16]
Seth Pettie. Personal communication, 2016.
[PS89]
D. Peleg and A. Schäffer. Graph spanners. J. Graph Theory, 13:99–116, 1989.
[PU89a]
D. Peleg and J. D. Ullman. An optimal synchronizer for the hypercube. SIAM J. on Comput.,
18:740–747, 1989.
30
[PU89b]
D. Peleg and E. Upfal. A tradeoff between size and efficiency for routing tables. J. of the ACM,
36:510–530, 1989.
[Rei93]
John H. Reif. Synthesis of Parallel Algorithms. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 1st edition, 1993.
[RZ04]
Liam Roditty and Uri Zwick. On dynamic shortest paths problems. In Algorithms - ESA 2004,
12th Annual European Symposium, Bergen, Norway, September 14-17, 2004, Proceedings,
pages 580–591, 2004.
[TZ05]
Mikkel Thorup and Uri Zwick. Approximate distance oracles. J. ACM, 52(1):1–24, 2005.
[TZ06]
M. Thorup and U. Zwick. Spanners and emulators with sublinear distance errors. In Proc. of
Symp. on Discr. Algorithms, pages 802–809, 2006.
[Woo06]
David P. Woodruff. Lower bounds for additive spanners, emulators, and more. In Proceedings
of the 47th Annual IEEE Symposium on Foundations of Computer Science, FOCS ’06, pages
389–398, Washington, DC, USA, 2006. IEEE Computer Society.
31
| 8cs.DS
|
Draft of 2016-09-30
Type checking through unification
Francesco Mazzoli1 and Andreas Abel2
1
FP Complete∗
<[email protected]>
Gothenburg University
<[email protected]>
arXiv:1609.09709v1 [cs.PL] 30 Sep 2016
2
Abstract
In this paper we describe how to leverage higher-order unification to type check a dependently
typed language with meta-variables. The literature usually presents the unification algorithm
as a standalone component, however the need to check definitional equality of terms while type
checking gives rise to a tight interplay between type checking and unification. This interplay
is a major source of complexity in the type-checking algorithm for existing dependently typed
programming languages. We propose an algorithm that encodes a type-checking problem entirely
in the form of unification constraints, reducing the complexity of the type-checking code by
taking advantage of higher order unification, which is already part of the implementation of
many dependently typed languages.
Keywords and phrases Dependent types, type checking, higher order unification, type reconstruction
1
Introduction
Theories with dependent types have been successfully exploited to design programming
languages and theorem provers, such as Agda [9], Idris [3], or Coq [2]. To make these
systems practical, the user is presented with a language much richer than the underlying
type theory, which will hopefully be small enough to gain confidence in the correctness of
the code that type checks it.
One common way to make a type theory palatable is extending it with meta-variables,
standing for yet to be determined terms, and solved by unification. Their usage in traditional
programming languages is confined to type inference, and thus traditionally they can stand
for types only. In dependently typed languages types can contain terms, and thus metavariables are usually extended to stand for any term in our language. A typical use case
for meta-variables is implicit arguments as introduced by Pollack [11], relieving the user of
having to write easily inferrable arguments to functions. For example, in Agda we can write
a safe head function which extracts the first element of a list, inferring both the type of the
elements and the length of the list:
head : {A : Set} → {n : Nat} → Vec A (1 + n) → A
head (x :: xs) = x
∗
The work presented in this paper was performed at the Chalmers University of Technology.
2
Type checking through unification
Here, Vec A n denotes a list of length n with elements of type A, and Set is the type of types.
The expression {A : Set} → {n : Nat} → ... binds two implicit arguments. When invoking
head, the type checker will insert two meta-variables standing for A and n and attempt to
solve them by inspecting the Vec argument that follows. Note that n is a value, while in
languages such ML and Haskell only types can be implicit.
The task of integrating meta-variables in a type-checking algorithm for dependent types
gives rise to complications. For example, consider the task of type checking
true : if α 6 2 then Bool else Nat,
where α is a yet to be determined (uninstantiated) meta-variable of type Nat. We want the
type of true to be Bool, but reduction is impeded by α. Thus, we cannot complete type
checking until α is instantiated.1 The problem lies in the fact that type checking dependent
types involves reducing terms to their normal forms, something that can be affected by
meta-variables, like in this case.
To solve issues like the one above, the only viable option—apart from refusing to solve
them—is to wait for the meta-variables that are affecting type checking to be instantiated,
and then resume. This gives rise to a sort of concurrency that makes reasoning about the
type checking algorithm arduous. In this paper, expanding on ideas developed in Agda [9]
and Epigram [7], we propose an algorithm that encodes a type-checking problem in a set
of unification constraints with a single traversal of the term to be checked. The generated
constraints can be solved effectively by the unification procedure already employed by Agda,
but our elaboration procedure is considerably simpler and shorter than Agda’s type-checking
code. This highlights an overlap in functionality between the type checker, which needs to
check that types and terms are of a certain shape; and the unifier, which checks the equality
of terms. Moreover, our algorithm lets us clearly separate concerns between type checking
and unification, making it easier to gain confidence on the elaboration procedure and then
experiment with various unification “backends”.
In the rest of the paper, we will explain the problem more clearly (Section 2). Then we
will introduce a simple type theory (Section 3) that will serve as a vector to explain our
algorithm in detail. In Section 4 we will give a specification to the unification procedure.
The algorithm itself is presented in Section 5, along with some of its properties. We will then
briefly discuss the performance and how the algorithm can be extended to support certain
popular language features (Section 6).
We have implemented the presented algorithm in a prototype, tog, which covers a subset
of Agda—every tog program is also a valid Agda program.2
2
The problem
In this section we will explain the challenges faced when type checking dependent types with
meta-variables. An Agda-like syntax will be used throughout the examples, please refer to
Appendix A for clarifications.
Coming back to the problem of type checking
true : BoolOrNat α,
given unistantiated meta-variable α and definition
1
2
Note that we cannot instantiate α without loss of generality, since both 0 and 1 are acceptable solutions.
The source code for tog is available at https://github.com/bitonic/tog.
F. Mazzoli and A. Abel
BoolOrNat : Nat → Set
BoolOrNat = λx → if x 6 2 then Bool else Nat
there are various tempting ways to approach the problem. The most conservative approach
is to stop type checking when faced with blocked terms (terms whose normalization is impeded by some meta-variables). However, this approach is unsatisfactory in many instances.
Consider
(true, refl) : BoolOrNat α × (α ≡ 0)
Where x ≡ y is the type inhabited by proofs that x is equal to y (propositional equality), and
refl is of type t ≡ t for any t (reflexivity). Type checking this pair will involve type checking
true : BoolOrNat α and then refl : α ≡ 0. If we give up on the first type-checking problem, we
will not examine the second, which will give us a solution for α (α := 0). After instantiating
α we can easily go back and successfully type check the first part. In general, we want to
attempt to type check as much as possible, and to instantiate as many meta-variables as
possible—as long as we do so without loss of generality, like in this case.
Another approach is to assume that blocked type-checking problems will eventually be
solved, and continue type checking. However, this road is dangerous since we need to be
careful not to generate ill-typed terms or invalid type-checking contexts, as noted by Norell
and Coquand [10]. Consider
test : (α ≡ 0) × (((x : BoolOrNat α) → BoolOrNat (1 + x)) → Nat)
test = (refl, λg → g true)
Type checking the definition test will involve checking that its type is a valid type, and that
its body is well typed. Checking the former will involve making sure that
BoolOrNat α = Nat
since we know that the type of x must be Nat, given that x is used as an argument of
(1+) : Nat → Nat.3
If we assume that the type is valid, we will proceed and type check the body pairwise.
Type checking the first element—a proof by reflexivity that α is equal to 0—will instantiate
α to 0, and then we will be faced with
(λg → g true) : ((x : Bool) → BoolOrNat (1 + x)) → Nat
Note that the type is ill-typed,4 violating the usual invariants present when type checking—
namely the fact that when we make progress we always generate well-typed terms. Worse,
to type check we will instantiate x with 0, ending up with BoolOrNat (1 + true). With some
effort we can exploit this problem to make the type checker loop, and thus type checking
will be undecidable.
As mentioned in the introduction, at the heart of the problem lies the fact that to type
check we need to reduce terms to their weak head normal form. If reduction is impeded
3
4
Note that checking that an equality type is a well-formed type does not involved checking that the
equated things are equal—4 ≡ 5 is a valid type. In this instance while α ≡ 0 appears in the type for
test, this does not mean that α will be unified with 0 when type checking the type. However, type
checking its proof refl : α ≡ 0 will.
x, of type Bool, appears as an argument to the function (1+).
3
4
Type checking through unification
by meta-variables, we cannot proceed. To overcome this problem, Norell proposed to define
type checking as an elaboration procedure:5 given the problem of type checking t against A
in context Γ, type checking will produce a term u that approximates t:
JΓ ⊢ t : AK
u
u is an approximation of t in the sense that it it can be turned into t by instantiating certain
meta-variables—if a subterm of t cannot be type checked a placeholder meta-variable will be
put in its place, an type checking that subterm will be postponed. Type checking will also
consist in making sure that, once the postponed type-checking problems can be solved, the
placeholder meta-variables will be instantiated accordingly with the corresponding omitted
subterm of t (possibly instantiated further).
For instance, when type checking the type of test, we’ll have
J· ⊢ ((x : BoolOrNat α) → BoolOrNat (1 + x)) → Nat : SetK
((x : BoolOrNat α) → BoolOrNat β) → Nat
Since we cannot type check
x : BoolOrNat α ⊢ 1 + x : Nat
a fresh meta-variable β of type Nat in context x : BoolOrNat α replaces 1 + x. Then, when
checking the body of test, we will check it against the approximated type generated above.
When α is instantiated, we can resume checking that BoolOrNat α = Nat, and if we are
successful, instantiate β := 1 + x. This will prevent us from running into problems when type
checking the body, since when we do instantiate α to 0, we do not have 1 + x later: instead,
β is in its place, preserving the well-typedness of the type.
The Agda system, as described in Norell’s thesis, currently implements this elaboration
interleaving type checking and unification, using some fairly delicate machinery. Our contribution is to describe a type-checking problem entirely in terms of unification constraints,
thus simplifying the algorithm. This highlights an overlap in functionality between the type
checker, which needs to check that types and terms are of a certain shape, and the unifier,
which checks the equality of terms: we are using the unifier as an engine to pattern match
on types, with taking meta-variables into account. Moreover, separating the unifier from the
type checker makes it easy easy to experiment with different unification “backends” used by
the same type checking “frontend”.
3
The type theory
To illustrate the type-checking algorithm we will make use of a dependent type theory with
booleans and one universe. Its syntax is shown in Figure 1.
3.1
Terms and types
Terms and types inhabit the same syntactic class. We usually denote terms with t, u, and
v; and types with A, B, and C . The theory is designed to be the simplest fragment that
presents the problems described in Section 2. For this reason we include a universe Set
5
Norell was following a long tradition of elaborating user syntax when working with type theories, see
Section 7.1 for more details.
F. Mazzoli and A. Abel
5
x, y, z
α, β, γ
-- Variables
-- Meta-variables
-- Types/terms
A, B, C , t, u, v
::= Set
| Bool | true | false
| (x : A) → B | λx → t
| n
-----
n
-- Neutral terms
::= h
| nt
| if n /x.A then t else u
::= x | α
h
Type of types
Booleans
Dependent functions
Neutral term
-- Head
-- Function application
-- Bool elimination
-- Neutral term heads
Γ, ∆ ::= · | Γ, x : A
-- Contexts
Σ, Ξ ::= · | Σ, α : A
-- Signatures
θ, η ::= · | θ, α := t
-- Meta-variable substitutions
Figure 1 Syntax
and means of computing with booleans, so that we can write functions from booleans to
types—otherwise meta-variables can never prevent us from knowing how a type looks like.
The typing rules and algorithms presented in this paper can be extended to a richer theory,
as we have done for our implementation, which includes implicit arguments, user defined
inductive data types and records, and propositional equality.
3.2
Neutral terms, substitutions, and term application
Terms are always kept in β-normal form. Terms headed by a variable or meta-variable are
called neutral, the others canonical. Variable and meta-variable substitution immediately
restore the β-normal form as soon as the substitution is performed, a technique known as
hereditary substitution [13]. Substituting a term u for a head h, i.e., a variable x or metavariable α, is written t[h := u ], reading “substitute u for h in t”. The rules for substitution
are not relevant to the paper and are shown in Appendix B. The appendix also defines rules
to eliminate redexes that substitution might generate, to restore the β-normal form.
3.3
Contexts and signatures
Most operations are done under a context (denoted by Γ or ∆), that stores the types of free
variables; and a signature (denoted by Σ or Ξ), that stores the type of meta-variables. We
tacitly assume that no duplicate names are present in contexts and signatures. We often
make use of a global signature Σ throghout the rules, if there is no need for the rules to carry
it explicitely since it is never changed. Note that a signature contains only closed types for
the meta-variables—we do not make use of an explicit representation of meta-variables in
context. This is for the sake of simplicity, since we do not present our unification algorithm
in detail, where the contextual representation would be most convenient. Throughout the
paper, we will use Γ → A to indicate the function type formed by all the types in Γ as the
6
Type checking through unification
Γ ⊢ Set : Set
Γ ⊢ Bool : Set
Γ ⊢ true : Bool
Γ ⊢ false : Bool
Γ ⊢ A : Set Γ, x : A ⊢ B : Set
Γ, x : A ⊢ t : B
Γ ⊢ n ⇒ A Γ ⊢ A ≡ B : Set
Γ⊢n:B
Γ ⊢ (x : A) → B : Set
Γ ⊢ λx → t : (x : A) → B
Figure 2 Σ; Γ ⊢ t : A Type checking canonical terms
x :A∈Γ
Γ⊢x⇒A
Γ ⊢ n ⇒ Bool
α:A ∈ Σ
Γ⊢α⇒A
Γ ⊢ n ⇒ (x : A) → B
Γ⊢t :A
Γ ⊢ n t ⇒ B[x := t]
Γ, x : Bool ⊢ A : Set
Γ ⊢ t : A[x := true]
Γ ⊢ u : A[x := false]
Γ ⊢ if n /x.A then t else u ⇒ A[x := n]
Figure 3 Σ; Γ ⊢ n ⇒ A Type inference for neutral terms
domains and terminating with A, and t Γ to indicate the term formed by t applied to all
the variables in Γ. Moreover, every mention of a context and a signature is assumed to be
valid according to the rules in Figure 5 and 4.
Throughout the paper, we will also concatenate whole contexts and signatures, e.g. Σ, Ξ.
3.4
Well-typed terms
The bidirectional typing checking rules are shown in figures 2 and 3. The type of neutral
terms can be inferred, while canonical terms are checked. Our type theory includes a universe
Set equipped with an inconsistent typing rule Set : Set for the sake of simplicity, but our
presentation can be extended with stratified universes.
Finally, definitional equality of terms (needed to define the typing rules) is specified in
figures 7 and 8. The conversion rules are performed in a type-directed way, so that it can
respect the η-laws of functions.
3.5
Meta-variable substitutions
We specify typed meta-variable substitutions, a construct that will be useful to give a specification to our unification algorithm. A meta-variable substitution θ from Σ to Ξ gives
an instantiation to each meta-variable in Σ. The meta-variables in the instantiations are
scoped over a new signature Ξ. The validity rule shown in figure 6 makes sure that the
meta-variables instantiations are well-typed with respect to their type.
Applying a substitution θ to a term t amounts to substitute each meta-variable for its
instantiation in θ, as specified in appendix B. Moreover, if the substitution is valid, if we
have Σ; Γ ⊢ t : A and Ξ ⊢ θ : Σ, we will have that
Ξ; θΓ ⊢ θt : θA,
where applying a substitution to a context amounts to applying it to all the types it contains.
Like with contexts and signatures, we will use , to also concatenate meta-variable substitutions. So for example if we have Ξ ⊢ θ : Σ and Ξ ⊢ η : Σ′ , we have that Ξ ⊢ (θ, η) : (Σ, Σ′ ).
F. Mazzoli and A. Abel
⊢Σ
⊢·
7
Σ; · ⊢ A : Set
⊢ Σ, α : A
Σ⊢Γ
Σ; Γ ⊢ A : Set
Σ ⊢ Γ, x : A
Σ⊢·
Figure 4 ⊢ Σ Well-formed signatures
Figure 5 Σ ⊢ Γ Well-formed contexts
for all (α : A) ∈ Σ, α := t ∈ θ and Ξ; · ⊢ t : θA
Ξ⊢θ:Σ
Figure 6 Ξ ⊢ θ : Σ Well-formed meta-variable substitutions
Γ ⊢ Set ≡ Set : Set
Γ ⊢ Bool ≡ Bool : Set
Γ ⊢ false ≡ false : Bool
Γ, x : A ⊢ f x ≡ g x : B
Γ ⊢ f ≡ g : (x : A) → B
Γ ⊢ true ≡ true : Bool
Γ ⊢ A1 ≡ A2 : Set Γ, x : A1 ⊢ B1 ≡ B2 : Set
Γ ⊢ (x : A1 ) → B1 ≡ (x : A2 ) → B2 : Set
Γ ⊢ n ≡ n′ ⇒ A
Γ ⊢ A ≡ B : Set
Γ ⊢ n ≡ n′ : B
Figure 7 Σ; Γ ⊢ t ≡ u : A Definitional equality of canonical terms
Γ ⊢ n ≡ n ′ ⇒ (x : A) → B
Γ ⊢ t ≡ t′ : A
Γ ⊢ n t ≡ n ′ t ′ ⇒ B[x := n]
Γ⊢h⇒A
Γ⊢h≡h⇒A
Γ ⊢ n ≡ n ′ ⇒ Bool
Γ, x : Bool ⊢ A ≡ A : Set
Γ ⊢ t ≡ t ′ : A[x := true]
Γ ⊢ u ≡ u ′ : A[x := false]
Γ ⊢ if n /x.A then t else u ≡ if n ′ /x.A′ then t ′ else u ′ ⇒ A[x := n]
′
Figure 8 Σ; Γ ⊢ n ≡ n ′ ⇒ A Definitional equality of neutral terms
8
Type checking through unification
4
Unification
In this section we will give a specification for the unification algorithm that we will need to
solve the constraints generated by elaboration.
4.1
Unification constraints
The input for the unification algorithm are heterogeneous constraints of the form
Γ ⊢ t : A = u : B.
Such a constraint is well formed if we have that Γ ⊢ t : A and Γ ⊢ u : B. As we will see in
Section 5, it is crucial for the constraints to be heterogeneous for the elaboration procedure
to work as we intend to.
A constraint is solved if we have that Γ ⊢ A ≡ B : Set and Γ ⊢ t ≡ u : A. This means
that to solve a constraint the unifier will have to establish definitional equality of both the
types and the terms.
4.2
Unification algorithm specification
A unification algorithm takes a signature and set of constraints, and attempts to solve them
by instantiating meta-variables. Thus, unification rules will be of the form
Σ, C
Ξ, θ
Where Σ and C are the input signature and constraints, and Ξ and θ are the output signature
and substitution, such that Ξ ⊢ θ : Σ. If the substitution solves all the constraints we have
that
Ξ
θC.
Note that unification might make no progress at all, and just return Σ and the identity
substitution.
4.3
A suitable unification algorithm
Higher order unification in the context of dependently typed languages is still a poorly
understood topic, and we do not have the space to discuss it in depth here. The basis
for most of the unification algorithms employed in such languages is pattern unification, as
introduced by Miller [8]. However in modern languages such as Agda or Coq there are factors
complicating this process. The main inconvenience, as already mentioned in Section 1, is the
fact that we cannot always know if a constraint is solvable, since solving other constraints
might enable us to solve the current one. A algorithm that takes care of handling constraints
under these assumptions is usually called dynamic.
However, another issue when unifying terms in these type theories is that definitional
equality (specified by conversion rules such as the ones presented in this paper) often includes η-laws for functions and product types, and possibly other type-directed conversion
rules. For this reason our constraints equate typed expressions.6 Note that the types in the
constraints are only needed so that we can abide by the η laws.
6
We also have the constraints to equate the types too, but this is only for brevity: if that wasn’t the case,
all we would have to do is to convert each Γ ⊢ t : A = u : B into {Γ ⊢ A : Set = B : Set, Γ ⊢ t : A = u : B }.
F. Mazzoli and A. Abel
Ideally, to solve the heterogeneous constraints, a heterogeneous pattern unification algorithm is needed, such as the one described in chapter 4 of Adam Gundry’s thesis [5].
However, in our prototype, we have found that implementing such an algorithm is impractical for performance reasons.7 In practice, a homogeneous pattern unification algorithm
like the one employed in the Agda system works well enough in many of the examples that
we have analysed. In this context, a heterogeneous constraint Γ ⊢ t :A = u :B is converted to
two homogeneous constraints, Γ ⊢ A = B : Set and Γ ⊢ t = u : A. Some bookkeeping will be
needed to ensure that the constraint equating the types is solved before attempting the one
equating the terms, so that Γ ⊢ t = u : A is considered only when it is well-formed—when
we know that A ≡ B.
5
Type checking through unification
As mentioned in Section 2, our algorithm will elaborate a type-checking problem into a
well-typed term and a set of unification constraints. Given some type-checking problem
Σ; Γ ⊢ t : A, the algorithm will elaborate it into a term u and constraints C, along with an
extended signature Ξ—since the elaboration process will add new meta-variables.
The algorithm is specified using rules of the form
JΣ; Γ ⊢ t : AK
Ξ, u, C,
such that:
Ξ is an extension of Σ;
u is well typed: Ξ; Γ ⊢ u : A;
C is a set of well formed unification constraints;
if unification produces a signature Ξ′ and substitution Ξ′ ⊢ θ : Ξ such that Ξ′ θC, we
have that Ξ′ ; θΓ ⊢ θt ≡ θu : θA. In other words, solving all the constraints restores
definitional equality between the original term t and the original term u.
The main idea is to infer what the type must look like based on the term, and generate
constraints that make sure that, if the constraints can be solved, that will be the case. For
example, if faced with problem
Σ; Γ ⊢ true : A,
we know that A must be Bool. However, A might be a type stuck on a meta-variable, as
discussed in Section 2. At the same time, we want the elaboration procedure to immediately
return a well-typed term. The heterogeneous constraints let us do just that: we will create
a new meta-variable of type A in Γ, and use that as the elaborated term. Moreover we will
return a constraint equating the newly created meta-variable to true:
JΣ; Γ ⊢ true : AK
(Σ, α : Γ → A), α Γ, {Γ ⊢ true : Bool = α Γ : A}
Note how we respect the contract of elaboration—the elaborated term is well-typed, the
constraint is well-formed—without making any commitment on the shape of A.
For a more complicated example, consider the type-checking problem
7
Specifically, the need to typecheck meta-variable instantiations and retry if they are ill-typed is particularly taxing.
9
10
Type checking through unification
Σ; Γ ⊢ λx → t : A.
We know that A needs to be a function type, but we do not know what the domain and
codomain types are yet. To get around this problem we will add new meta-variables to the
signature acting as the domain and codomain, then elaborate the body using those, and
follow the same technique illustrated above to return a well-typed term by adding a new
meta-variable:
-- Add meta-variables for the domain (β) and codomain (γ):
Σ1 := Σ, β : Γ → Set, γ : Γ → (x : β) → Set
-- Elaborate the body of the abstraction:
JΣ1 ; Γ, x : β ⊢ t : γ Γ xK
Σ2 , u, C
-- Add meta-variable that we will return as the elaborated term:
Σ3 := Σ2 , α : Γ → A
-- Return the appropriate constraint equating the abstracted elaborated body to the new
-- meta-variable, together with the constraints generated from elaborating the body:
JΣ; Γ ⊢ λx → t : AK
Σ4 ,
α Γ,
{Γ ⊢ (λx → u) : (x : β Γ) → γ Γ x = α Γ : A} ∪ C
Note how the use of heterogeneous equality is crucial if we want to avoid to ever commit
to the types having a particular shape, while having all the constraints to be well-formed
immediately.
Every rule in full algorithm is going to follow the general pattern that emerged above:
(a) Elaborate sub-terms, adding meta-variables to have types to work with;
(b) Add a meta-variable serving as the elaborated term, say α;
(c) Return a constraint equating α to a term properly constructed using the elaborated
subterms; together with the constraints returned by elaborating said subterms.
For brevity we present the full algorithm using an abbreviated notation that implicitely
threads the signatures (Σ, Σ1 , Σ2 , and Σ3 in the example above) across rules. We will also use
the macro Fresh(_, _) to add new meta-variables in a context, such that α := Fresh(Γ, A)
is equivalent to Ξ := Σ, α : Γ → A, and successive appearances of α are automatically applied
to Γ, and where Σ and Ξ are the old and new signature—which are, as said, implicitly
threaded. Finally, we implicitly collect the constraints generated when elaborating the
subterms, and implicitly add a meta-variable that stands for the elaborated term, together
with an appropriate constraint—steps (b) and (c) in the process described above.8 For
example, the rule to elaborate abstractions, explained before, will be shown as
β := Fresh(Γ, Set)
γ := Fresh(Γ, x : β, Set)
JΓ, x : β ⊢ t : γK
u
JΓ ⊢ λx → t : AK
(λx → u) : (x : β) → γ
The complete algorithm is shown in figure 9. They are remarkably similar to the typing rules, however instead of matching directly on the type we expect we match through
constraints.
8
In our Haskell implementation, the elaboration is a state and writer monad over the signature and the
constraint set, respectively.
F. Mazzoli and A. Abel
JΓ ⊢ Set : AK
JΓ ⊢ false : AK
11
Set : Set
JΓ ⊢ Bool : AK
Bool : Set
JΓ ⊢ true : AK
true : Bool
JΓ ⊢ A : SetK
A′
Γ, x : A′ ⊢ B : Set
B′
JΓ ⊢ (x : A) → B : S K
(x : A′ ) → B ′ : Set
false : Bool
β := Fresh(Γ, Set)
γ := Fresh(Γ, x : β, Set)
JΓ, x : β ⊢ t : γK
u
JΓ ⊢ λx → t : AK
(λx → u) : (x : β) → γ
x :A∈Γ
JΓ ⊢ xK
x :A
β := Fresh(Γ, Set)
γ := Fresh(Γ, x : β, Set)
JΓ ⊢ n : (x : β) → γK
t
JΓ ⊢ u : βK
v
JΓ ⊢ n u : AK
t v : γ[x := v]
α:A ∈Σ
JΓ ⊢ αK
α:A
JΓ, x : Bool ⊢ B : SetK
B′
JΓ ⊢ n : BoolK
t
′
′
JΓ ⊢ u : B [x := true]K
u
JΓ ⊢ v : B ′ [x := false]K
v′
JΓ ⊢ if n /x.B then u else v : AK
if t /x.B ′ then u ′ else v ′ : B ′ [x := t]
Figure 9 JΣ; Γ ⊢ t : AK
Ξ, u, C Elaboration
The rules can easily be turned into an algorithm which pattern matches on the term and
decides how to proceed.9 Naturally the algorithm can still fail if it encounters an out of
scope meta-variable or variable, although in real systems scope checking is usually performed
beforehand.
5.1
Examples
We will explore how the algorithm works by going through various common situations. The
reader can experiment using the mentioned tog tool, passing the -d ’elaborate’ commandline flag to have it to print out the generated constraints. A wealth of examples are present
in the repository. We will assume the usage of pattern unification to solve constraints when
examining the examples.
5.1.1
A simple problem
Let’s take type-checking problem
·; · ⊢ true : Bool.
The algorithm will return the triple
α : Bool, α, {· ⊢ true : Bool = α : Bool }
The constraint is immediately solvable, yielding the substitution α := true, which will restore
definitional equality between the elaborated term and true.
5.1.2
An ill-typed problem
Now for something that should fail. Given
9
Our implementation can be found at https://github.com/bitonic/tog/blob/master/src/Tog/Elaborate.hs.
12
Type checking through unification
add : Nat → Nat → Nat,
we want to solve
·; x : Nat ⊢ add x : Nat.
The algorithm will return the triple
Σ, ζ x, C where
Σ = α : Nat → Nat, β : (x : Nat) → Nat → Set, γ : (x : Nat) → α x,
δ : (x : Nat) → (y : α x) → β x y, ζ : Nat → Nat
C = {x : Nat ⊢ δ x (γ x) : β x (γ x) = ζ x : Nat,
x : Nat ⊢ add : Nat → Nat → Nat = δ x : (y : α x) → β x y,
x : Nat ⊢ x : Nat = γ x : α x }
While looking scary at first, the meaning of the meta-variables and constraints is easy to
interpret, keeping in mind that we generate one constraint per subterm (including the top
level term).
At the top level, we elaborate the two subterms add and x. We know that add must be a
function type, since it is applied to something; and that the type of x must match the type
of the domain of said function type. Thus, two meta-variables are created to represent the
domain and codomain—α and β. Then, add and x are elaborated with said types, which in
turn requires the addition of γ and δ, serving as elaborated terms. These are the ingredients
for the second and third constraint. Finally, the elaborated δ (representing add) is applied to
γ (representing x), and equated to a new meta-variable ζ, which is the result of the top-level
elaboration.
The constraints reflect the fact that the term is ill-typed: β is equated both to Nat (in
the first constraint), and to Nat → Nat, in the second constraint. Thus, unification will fail.
5.1.3
An unsolvable problem
Finally, let’s go back to an example which cannot immediately be solved in its entirety:
α : Nat; · ⊢ (true, 0) : BoolOrNat α × Nat.
The desired outcome for this problem is to type check the second element of the pair, 0;
but “suspend” the type checking of the first element by replacing refl with a meta-variable.
Running the type-checking problem through the elaboration procedure yields
Σ, ζ x, C where
Σ = β : Set, γ : Set, δ : β, ζ : γ, ι : BoolOrNat α × Nat
C = {· ⊢ (δ, ζ) : β × γ = ι : BoolOrNat α × Nat,
· ⊢ 0 : Nat = ζ : γ,
· ⊢ true : Bool = δ : β }
The second and third constraints regard the two sub-problems individually, and are solvable
yielding the substitution
β := Bool, δ := true, γ := Nat, ζ := 0,
and leaving us with
F. Mazzoli and A. Abel
13
· ⊢ (true, 0) : Bool × Nat = ι : BoolOrNat α × Nat.
At this point the unifier will be able to substitute ι with a pair
· ⊢ (true, 0) : Bool × Nat = (ξ, 0) : BoolOrNat α × Nat.
Now solving the constraint amounts to solve
· ⊢ true : Bool = ξ : BoolOrNat α,
but the unifier is not going to be able to make progress, since substituting ξ for true will
leave the constraint ill-formed. Thus, the final result of the elaboration and unification will
be (ξ, 0), which is all we could hope for.
5.2
Some properties
◮ Lemma 1 (Well-formedness and functionality of elaboration). Let ⊢ Σ and Σ ⊢ Γ.
1. There uniquely exist a signature Ξ, constraints C, a term t ′ , and a type A such that
JΣ; Γ ⊢ tK
Ξ | C ⊢ t′ ⇒ A
and Ξ; Γ ⊢ A : Set.
2. If Σ; Γ ⊢ A : Set, then there uniquely exist a signature Ξ, constraints C, and a term t ′
such that
JΣ; Γ ⊢ t ⇐ AK
Ξ | C ⊢ t′.
In both cases, Ξ is necessarily an extension of Σ and all the outputs are well-formed, meaning
⊢ Ξ and Ξ; Γ ⊢ t ′ : A and Ξ; Γ ⊢ C.
Proof. By induction on t.
◭
◮ Lemma 2 (Soundness of elaboration).
If one of
JΣ; Γ ⊢ tK
Ξ | C ⊢ t′ ⇒ A
JΣ; Γ ⊢ t ⇐ AK
Ξ | C ⊢ t′
and there is some well-typed meta substitution Σ′ ⊢ θ : Ξ that solves the constraints, i.e.,
Σ′ θC, then
Σ′ , θΓ ⊢ θt ≡ θt ′ : θA.
Proof. By induction on term t.
◭
◮ Lemma 3 (Strong soundness of elaboration).
1. If
JΣ; Γ ⊢ tK
Ξ | C ⊢ t′ ⇒ A
and there is some closing untyped substitution σ such that dom(σ) ⊇ dom(Ξ) and ⊢ σ : Σ
and σΓ ⊢ σA : Set and σ solves the untyped constraints, then
σΓ ⊢ σt ≡ σt ′ : σA.
14
Type checking through unification
2. If
JΣ; Γ ⊢ t ⇐ AK
Ξ | C ⊢ t′
and there is some closing untyped substitution σ : Ξ that is well-typed for Σ, i.e, ⊢ σ : Σ
and solves the untyped constraints, i.e., σC, then
σΓ ⊢ σt ≡ σt ′ : σA.
Proof. By induction on term t.
Case Abstraction λx → t
[[(Σ, α : (Γ → Set)); (Γ, x : (α Γ)) ⊢ t]]
Ξ | C ⊢ t′ ⇒ B
[[Σ; Γ ⊢ (λx → t)]]
Ξ | C ⊢ (λx → t′ ) ⇒ ((x : α Γ) → B)
By assumption ⊢ σ : Σ and σΓ ⊢ σ((x : α Γ) → B) : Set, which by inversion gives
us σΓ ⊢ (σα) Γ : Set and σ(Γ, x : α Γ) ⊢ σB : Set. Thus ⊢ σα : (σΓ → Set) and
⊢ σ : (Σ, α : (Γ → Set)). By induction hypothesis, σΓ, x : σ(α Γ) ⊢ σt = σt′ : σB, thus by
the ξ rule for equality, σΓ ⊢ σ(λx → t) = σ(λx → t′ ) : σ((x:α Γ) → B).
Case Application t u
[[Σ; Γ ⊢ u]]
Σ1 | C 1 ⊢ u ′ ⇒ A
[[Σ1 , β : ((Γ, x:A) → Set); Γ, x:A ⊢ t ⇐ ((x:A) → β Γ x)]]
[[Σ; Γ ⊢ t u]]
Σ 2 | C 1 ∪ C 2 ⊢ t′ u ′ ⇒ β Γ u ′
Σ 2 | C 2 ⊢ t′
By the first induction hypothesis, σΓ ⊢ σu = σu′ : σA. By the second induction hypothesis
◭
◮ Lemma 4 (Completeness of elaboration). Let us assume a term t and a wellformed signature Σ and Σ ⊢ Γ and a second signature Ξ and a substitution Ξ ⊢ θ : Σ such that
Ξ; θΓ ⊢ θt : C .
1. Then
JΣ; Γ ⊢ tK
(Σ, Σ′ ) | C ⊢ t ′ ⇒ B
and there exists a substitution θ′ such that
Ξ ⊢ (θ, θ′ ) : (Σ, Σ′ )
Ξ (θ, θ′ )C
Ξ; θΓ ⊢ C = (θ, θ′ )B : Set
Ξ; θΓ ⊢ θt = (θ, θ′ )t ′ : C .
2. Further, if Σ; Γ ⊢ A : Set and Ξ; θΓ ⊢ θA = C : Set, then
JΣ; Γ ⊢ t ⇐ AK
(Σ, Σ′ ) | C ⊢ t ′
and there exists a substitution θ′ such that
Ξ ⊢ (θ, θ′ ) : (Σ, Σ′ )
Ξ (θ, θ′ )C
Ξ; θΓ ⊢ θt = (θ, θ′ )t ′ : C .
F. Mazzoli and A. Abel
15
In other words, if it’s possible to instantiate some meta-variables to make t well-typed, then
all the constraints generated by the elaboration procedure are solvable.
Proof. By induction on t.
◭
(Σ, Σ′ ) | C ⊢ t ′ ⇒ A′ , then
◮ Corollary 5 (Simple inference). If Σ; Γ ⊢ t : A and JΣ; Γ ⊢ tK
all the constraints in C are solvable by some substitution Σ ⊢ θ′ : Σ′ and Σ; Γ ⊢ A = θ′ A′ : Set.
Proof. From Completeness with Ξ = Σ and identity substitution θ.
◭
◮ Corollary 6 (Simple checking). If Σ; Γ ⊢ t : A and JΣ; Γ ⊢ t ⇐ AK
all the constraints in C are solvable.
(Σ, Σ′ ) | C ⊢ t ′ , then
Proof. From Completeness with Ξ = Σ and identity substitution θ.
◭
5.2.1
Effectiveness
While the properties above guarantee establish some basic results about the algorithm, they
are all also satisfied by the very useless elaboration algorithm
JΣ; Γ ⊢ t : AK
(Σ, α : Γ → A, β : Γ → Set), α Γ, {Γ ⊢ α Γ : A = t : β Γ}.
However, the intent of the developed algorithm is to be as effective, when used with pattern
unification, as the current techniques to type check dependent types with meta-variables.
To achieve this, it has been designed so that the generated constraints fall in the pattern
fragment when existing type checkers would be able to make progress, and by testing the
algorithm “in the wild” with existing Agda programs we have found it effective in practice.
For instance, if we did not keep our terms in β-normal form, we could not elaborate applications into constraints falling into the pattern fragment, due to the fact that we cannot
reliably infer the type of the function.
However, future work should involve a formal characterization of completeness in the
context of type checkers for dependent types with meta-variables, and a proof that our
algorithm is indeed complete.
6
Additional remarks
6.1
Additional features
In this section we will briefly sketch how to fit popular features into the framework described.
The general idea is to do everything which doesn’t require normalization in the elaboration
procedure, and the rest into the unifier.
6.1.1
Implicit arguments
We have already implemented a restricted form of implicit arguments, which we call “type
schemes” in line with ML terminology. Type schemes allow the user to define a number of
implicit arguments in top-level definitions, for example the already mentioned
head : {A : Set} → {n : Nat} → Vec A (1 + n) → A.
, before even
Under this scheme, every occurrence of head is statically replaced with head
elaborating—we do it while scope checking. However this is obviously limited, since we can
only have implicit arguments before any explicit ones and only for top-level definitions.
16
Type checking through unification
We want to enable a more liberal placement of implicit arguments. This is achieved in
Agda by allowing implicit arguments in all types, and in any many positions. The details of
the implementation are still in flux in Agda itself, but the core idea is to type-check function
application and abstractions bidirectionally, by first looking at the type and inserting implicit
arguments if needed. So elaborating t u where t : {x : A} → (y : B) → C will result in t u.
Similarly, elaborating t : {x : A} → B will result in λ{x } → t, where {x } binds an implicit
arguments.
We are exploring two ways to integrate this kind of mechanism in our framework:
Mimic the bidirectional type-checking performed in Agda and similar systems closely, by
adding a new kind of constraint for function application and abstraction which waits on
the type to have a rigid head, that is to say a type not blocked on a meta-variable.
Alternatively, force all implicit arguments to appear before an explicit one (with the
exception of type schemes), and always include an implicit arguments in →-types.10
Multiple implicit arguments will then be handled by one implicit argument carrying a
tuple.
We are currently implementing the latter proposal, since it is simpler to describe and to
implement, although only using it will tell if it is practical.11
6.1.2
Type classes
Type classes, as employed by Haskell, were introduced to handle overloaded operators in
programming languages [12]. Other similar structures include canonical structures in Coq.
In short, type classes let the user specify a collection of methods that can be implemented
for a specific type. For example, the type class identifying monoid is defined as
class Monoid a where
mempty : a
mappend : a → a → a
-- Monoid instance for list
instance Monoid [a ] where
mempty = [ ]
-- Empty list
mappend = (+
+) -- Concatenation
-- Monoid instance for pairs of monoids
instance (Monoid a, Monoid b) ⇒ Monoid (a, b) where
mempty = (mempty, mempty)
mappend (x1 , y1 ) (x2 , y2 ) = (x1 ‘mappend‘ x2 , y1 ‘mappend‘ y2 )
Type classes are a form of type-directed name resolution, and thus cannot be resolved at
elaboration time—we might need to instantiate meta-variables before being able to resolve
the right method. To integrate it in our framework we have to include type-classes into
our unification procedure. Luckily, this is exactly what the authors of the theorem prover
Matita accomplished using what they dubbed unification hints [1]. Briefly, the unifier is
given “hints” on how to solve problems that it cannot resolve by itself, and such hints are
repeatedly tried if unification fails.
10
Note that the restriction of having implicit arguments always paired with explicit ones is being planned
in Agda too, to make the automatic insertion of implicit arguments more predictable.
11
A more thorough proposal is available at https://github.com/bitonic/tog/wiki/Implicit-Arguments.
F. Mazzoli and A. Abel
Similar to type-classes, overloaded constructors are a feature introduced by Agda, that
lets the user define multiple data constructors with the same name. When such an overloaded constructor is used, its type must be determined by the type we are type checking
it against. It is easy to see how this problem is essentially the same as resolving the right
type class instance when encountering one of its methods—the type-class being “data types
with the same constructors”, the instances being the data-types, and the methods being the
constructors—, and thus we plan to implement this feature using unification hints as well.
6.2
Performance
One reason for concern is that our algorithm generates more constraints than ordinary typechecking algorithms. However, as already remarked, our algorithm generates a number of
constraints and meta-variables which is linear in the size of the input term. Moreover, we
would expect the unifier to spend little time solving the trivial constraints which are normally
handled by the type-checker directly.
In the examples we have collected, we have found that this is the case, and the run time
is dominated by unification filling in implicit arguments which can become very large. More
specifically, most of the time is spent dealing with η-expansion, which we plan to improve
in the future.
7
Conclusion
We have presented an algorithm that leverages the unifier, an important part of already
existing dependently typed programming languages and theorem provers, to greatly simplify
the process of type checking. The expressivity of higher-order unification lets us specify the
type-rules concisely. Moreover, we have clearly separated type checking from unification,
allowing for greater modularity.
We have implemented the ideas presented in the tog, covering a large subset of the Agda
languages. We are currently in the process of improving tog to get narrow the gap between
its capabilities and Agda.
7.1
Acknowledgements and related work
This work is a continuation of the work by Norell & Coquand [10], which describes how
Agda deals with issues that we presented. In fact, the algorithm described here is a simpler
re-implementation of what they specified.
The other main inspiration came from a discussion with Adam Gundry about how Epigram 2 deals with type checking in the presence of meta-variables. The propositional equality
Epigram 2 is powerful enough to represent constraints stuck on some meta-variable as an
unfinished equality proof. Thus, the unifier, when given a constraint Γ ⊢ t : A = u : B
produces a—possibly unfinished—proof that Γ ⊢ t : A ≡ u : B, where here ≡ denotes Epigram’s heterogeneous propositional equality. This proof can be used to “transport” terms
on one side of the equation to the other. From this discussion I realised that such a powerful
equality was not needed to implement a similar system.
The already cited work on Matita [1] convinced us further that isolating “dynamic”
procedures in the unifier is a good idea.
Finally, our work follows a long tradition of separating elaboration of the user syntax into
a core type theory, and the type-checking of such theory. As far as we know this line of work
17
18
Type checking through unification
Set[h := t]
Set
Bool[h := t]
Bool
true[h := t]
A[h := t]
A′
B[h := t]
B′
((x : A) → B)[h := t]
(x : A′ ) → B ′
h[h := t]
t
n[h := t]
w
h 6≡ h ′
h [h := t]
h′
true
u[h := t]
(λx → u)[h := t]
n[h := t]
′
v
u[h := t]
(n u)[h := t]
A[h := t]
A′
u[h := t]
u′
v[h := t]
′
′
′
if w /x.A then u else v
w′
(if n /x.A then u else v)[h := t]
w′
Figure 10 u[h := t]
terms
u ′ and n[h := t]
false[h := t]
u′
v′
false
u′
λx → u ′
v u′
v′
v′
u Hereditary substitution into canonical and neutral
goes back at least to Coquand and Huet’s “Constructive Engine” [6], with its separation
between the “concrete” user syntax and the internal “abstract” syntax.
Moreover, we’d like to thank Andrea Vezzosi, Nils Anders Danielsson, Conor McBride,
and Bas Spitters for the useful input.
A
Syntax and types used in examples
Throughout the examples we use a syntax close to the one in Agda and Haskell. Some
details regarding the syntax:
type and data constructors are shown in a sans serif font;
new meta-variables are introduced using underscores ( );
binary operators are referred to with parentheses, e.g. (+);
implicit arguments are indicated by curly braces;
definitions can be defined by pattern matching.
We use some standard types, namely:
Set : Set—Set is the type of all types, and its type is Set itself;
Bool : Set, inhabited by true and false;
Nat : Set, representing the natural numbers, introduced by number literals (1, 2, etc.);
(≡):{A:Set} → A → A → Set, the identity type; inhabited by refl:{A:Set} → {x :A} →
x ≡ x.
B
Substitution and term elimination
Figure 10 show the rules to substitute a term u for a head h, which can be a variable x or a
meta-variable α. Substituting (n u)[h := t] into proper neutral terms n t might (in case the
head of n is h) generate redexes, which are eliminated by invocations of the term elimination
judgment t u
v and if t /x.A then u else v
t, whose rules are given respectively in
Figure 11 and Figure 12. Term elimination in turn invokes substitution when the eliminated
term is a λ-abstraction.
While we use explicit names in our rules, we do not address issues related to α-renaming
here. In our prototype substitution is implemented using de Bruijn indices [4].
F. Mazzoli and A. Abel
19
t[x := u]
(λx → t) u
Figure 11 t u
t′
t′
v Application elimination
(if true /x.A then t else u)
Figure 12 if t /x.A then u else v
t
(if false /x.A then t else u)
u
t Bool elimination
References
1
2
3
4
5
6
7
8
9
10
11
12
13
Andrea Asperti, Wilmer Ricciotti, Claudio Sacerdoti Coen, and Enrico Tassi. Hints in
unification. In Theorem Proving in Higher Order Logics, pages 84–98. Springer, 2009.
Yves Bertot and Pierre Castéran. Interactive Theorem Proving and Program Development,
Coq’Art:the Calculus of Inductive Constructions. Springer-Verlag, 2004.
Edwin Brady. Idris, a general-purpose dependently typed programming language: Design
and implementation. Journal of Functional Programming, 23(05):552–593, 2013.
Nicolaas Govert De Bruijn. Lambda calculus notation with nameless dummies, a tool
for automatic formula manipulation, with application to the church-rosser theorem. In
Indagationes Mathematicae (Proceedings), volume 75, pages 381–392. Elsevier, 1972.
Adam Gundry. Type Inference, Haskell and Dependent Types. PhD thesis, Department of
Computer and Information Sciences, University of Strathclyde, 2013.
Gérard Huet. The constructive engine. A Perspective in Theoreticak Computer Science,
pages 38–69, 1989.
Conor McBride and James McKinna. The view from the left. Journal of Functional
Programming, 14(1):69–111, 2004.
Dale Miller. Unification under a mixed prefix. J. Symb. Comput., 14(4):321–358, October
1992.
Ulf Norell. Towards a practical programming language based on dependent type theory.
PhD thesis, Department of Computer Science and Engineering, Chalmers University of
Technology, SE-412 96 Göteborg, Sweden, 2007.
Ulf Norell and Catarina Coquand. Type checking in the presence of meta-variables. Unpublished draft, 2007.
Robert Pollack. Implicit syntax. Informal Proceedings of First Workshop on Logical Frameworks, Antibes, May 1990.
Philip Wadler and Stephen Blott. How to make ad-hoc polymorphism less ad hoc. In Proceedings of the 16th ACM SIGPLAN-SIGACT Symposium on Principles of Programming
Languages, POPL ’89, pages 60–76, New York, NY, USA, 1989. ACM.
Kevin Watkins, Iliano Cervesato, Frank Pfenning, and David Walker. A concurrent logical
framework i: Judgments and properties. Technical report, DTIC Document, 2003.
| 6cs.PL
|
Optimal Mission Planner with Timed Temporal Logic Constraints
arXiv:1510.01261v1 [cs.SY] 5 Oct 2015
Yuchen Zhou, Dipankar Maity and John S. Baras
Abstract— In this paper, we present an optimization based
method for path planning of a mobile robot subject to time
bounded temporal constraints, in a dynamic environment.
Temporal logic (TL) can address very complex task specification such as safety, coverage, motion sequencing etc. We use
metric temporal logic (MTL) to encode the task specifications
with timing constraints. We then translate the MTL formulae
into mixed integer linear constraints and solve the associated
optimization problem using a mixed integer linear program
solver. This approach is different from the automata based
methods which generate a finite abstraction of the environment
and dynamics, and use an automata theoretic approach to
formally generate a path that satisfies the TL. We have applied
our approach on several case studies in complex dynamical
environments subjected to timed temporal specifications.
I. I NTRODUCTION
Autonomous aircraft have been deployed for agriculture
research and management, surveillance and sensor coverage
for threat detection and disaster search and rescue operations.
All these applications require the tasks to be performed
in an optimal manner with specific timing constraints. The
high level task specifications for these applications generally
consist of temporal ordering of subtasks, motion sequencing
and synchronization etc. Given such specifications, it is
desirable to synthesize a reference trajectory that is both
optimal considering the dynamics of the vehicle and satisfies
the temporal constraints. Motion planning [1], [2], at its
early stage, considered optimal planning to reach a goal
from an initial position while avoiding obstacles [3]. New
techniques such as artificial potential functions [3], [4], cell
decomposition and probabilistic roadmaps [2] are introduced
for efficient planning in complex environment [5], [6] and
high dimensional state-space. However, these approaches
failed when task specifications have multiple goals or specific
ordering of goals, for example surveying some areas in
particular sequence.
Temporal logic [7]–[9] provides a compact mathematical
formulation for specifying such complex mission specifications. Previous approaches mainly focus on the usage of
linear temporal logic (LTL), which can specify tasks such
as visiting goals, periodically surveying areas, staying stable
and safe. The main drawback of the LTL formulation is that
it cannot specify time distance between tasks. In surveillance
examples, a simple task may be to individually monitor
*This work is supported by US AFOSR MURI grant FA9550-09-1-0538,
NSF grant CNS-1035655, NIST grant 70NANB11H148 and by DARPA
(through ARO) grant W911NF1410384.
The authors are with the Department of Electrical and Computer
Engineering, and the Institute for Systems Research, University of
Maryland, College Park, Maryland, USA. email: {yzh89, dmaity,
baras}@umd.edu
multiple areas for at least x amount of time. Additionally, the
LTL formulation commonly assumes the environment to be
static. Traditional approaches commonly start with creating a
finite abstraction of the environment including the dynamics,
then combine it with the automata that is generated from the
LTL specification [10]. The cell decomposition performed in
the abstraction process requires the environment to be static;
but in most situations this is not the case. For example, the
use of Unmanned Aerial Vehicles (UAVs) for surveillance in
the commercial airspace needs to consider motion of other
aircraft. The other weakness of the automata-based approach
is that it is computationally expensive. In this work, we are
interested in motion planning for surveillance in an airspace
with finite time task constraints and safety guarantees. For
this work, we only consider the other aircraft in the target
area as dynamic obstacles to the UAV. Further, we assume
that the motions of these dynamic obstacles can be either
predicted during the planning or are known a priori.
Due to the limitations of the previous approaches, we
instead present the method based on metric temporal logic
(MTL) [11], [12] and an optimization problem formulation
to solve the planning problem. MTL extends the LTL [7]
temporal operators so that it can express the requirements
on time distance between events and event durations. This
allows us to describe the dynamic obstacle and survey
durations in our case study.
An optimization based method for LTL was previously
proposed and extended by [13], [14]. In [13], the authors
propose to transform LTL specifications to mixed integer
constraints and solve the planning problem for finite horizon
using either a mixed integer linear program (MILP) or a
mixed integer quadratic program (MIQP) solver. In [14], the
algorithm is extended to infinite horizon so that trajectories
can contain loops. However, none of the methods consider
dynamic environment or moving obstacles, time varying
constraints, or duration of the tasks. MTL was used as a
temporal constraint to a routing planning problem in [15].
Their formulation unfortunately does not allow users to
incorporate dynamics of the vehicle.
In this paper, we consider a path planning problem for
surveillance under survey durations constraints for each
region and overall temporal constraint to visit each region
within given times. Our problem is considerably different
from the problems formulated in the existing literature in
the sense that we not only consider dynamic environment
but also associate each subtask with a duration constraint.
We generate a path that guarantee safety by avoiding static
and moving obstacles in the workspace and the path is
optimal in the sense that it minimizes a predefined cost
function. We do not adopt the linear encoding from [14],
since moving obstacles is not periodic in nature. Similar
to their approaches, we adopt the usage of mixed logic
dynamic (MLD) to model vehicle dynamics, so that the
overall problem is a MILP (considering linear cost function).
Our main contribution is usage of MTL to specify time
bounded tasks for the mission planner and reformulation of
the problem into a MILP. We also demonstrate the methods
on the case studies of using a quadrotor and ground vehicle
to survey multiple areas given the MTL specifications. The
rest of the paper is organized as follows. In section II we
present the fundamentals of MTL and define the overall
motion planning problem. We then formulate it into a mixed
integer linear optimization problem incorporating the temporal constraints in section III. Afterward, we demonstrate our
approach on motion planning for different simulation setups.
II. P RELIMINARIES
In this paper we consider a surveying task in an area by a
robot whose dynamics are given by the nonlinear model (1).
x(t + 1) = f (t, x(t), u(t))
(1)
where x(t) ∈ X , x(0) ∈ X0 ⊆ X , and u(t) ∈ U for all
t = 0, 1, 2, · · · . Let us denote the trajectory of the system
(1) starting at t0 with initial condition x0 and input u(t) as
xxt00 ,u = {x(s) |s ≥ t0 x(t + 1) = f (t, x(t), u(t)), x(t0 ) =
x0 }. For brevity, we will use xt0 instead of xxt00 ,u whenever
we do not need the explicit information about u(t) and x0 .
Definition 2.1: An atomic proposition is a statement
about the system variables (x) that is either True(>) or
False(⊥) for some given values of the state variables. [13]
Let Π = {π1 , π2 , · · · πn } be the set of atomic propositions
which labels X as a collection of survey areas, free space,
obstacles etc. The moving obstacles make the free space to
change from time to time, and hence labeling the environment is time dependent. We define a map that labels the time
varying environment as follows
L : X × I → 2Π
(2)
where I = {[a, b]| b > a ≥ 0} and 2Π is denoted as the
power set of Π. In general, I is used to denote a time interval
but it can be used to denote a time instance as well.
The trajectory of the system is a sequence of states such
that each state x(t) stays in X for all t and there exists u(t) ∈
U for all t such that x(t + 1) = f (t, x(t), u(t)). Corresponding to each trajectory x0 , the sequence of atomic proposition
satisfied is given by L(x0 ) = L(x(0), 0)L(x(1), 1)...
The high level specification of the surveying task will be
expressed formally using MTL which can incorporate timing
specifications.
A. Metric Temporal Logic (MTL)
Metric temporal logic, a member of temporal logic family,
deals with model checking under timing constraints. The
formulas for MTL are build on atomic propositions by
obeying some grammar.
Definition 2.2: The syntax of MTL formulas are defined
according to the following grammar rules:
φ ::= > | π | ¬φ | φ ∨ φ | φUI φ
where I ⊆ [0, ∞] is an interval with end points in N ∪ {∞}.
π ∈ Π, > and ⊥(= ¬>) are the Boolean constants true and
f alse respectively. ∨ denotes the disjunction operator and
¬ denotes the negation operator. UI symbolizes the timed
Until operator. Sometimes we will represent U[0,∞] by U.
Other Boolean and temporal operators such as conjunction
(∧), eventually within I (♦I ), always on I (I ) etc. can
be represented using the grammar desired in definition 2.2.
For example, we can express time constrained eventually
operator ♦I φ ≡ >UI φ and so on.
Definition 2.3: The semantics of any MTL formula φ is
recursively defined over a trajectory xt as:
xt |= π iff π ∈ L(x(t), t)
xt |= ¬π iff π ∈
/ L(x(t), t)
xt |= φ1 ∨ φ2 iff xt |= φ1 or xt |= φ2
xt |= φ1 ∧ φ2 iff xt |= φ1 and xt |= φ2
xt |= φ iff xt+1 |= φ
xt |= φ1 UI φ2 iff ∃s ∈ I s.t. xt+s |= φ2 and ∀ s0 ≤
s, xt+s0 |= φ1 .
Thus, the expression φ1 UI φ2 means that φ2 will be true
within time interval I and until φ2 becomes true φ1 must
be true. Similarly, the release operator φ1 Rφ2 denotes the
specification that φ2 should always hold and it can be
released only when φ1 is true. The MTL operator φ means
that the specification φ is true at next time instance, I φ
means that φ is always true for the time duration I, ♦I φ
means that φ will eventually become true within the time
interval I. Composition of two or more MTL operators
can express very sophisticated specifications; for example
♦I1 I2 φ means that within time interval I1 , φ will be true
and from that instance it will hold true always for a duration
of I2 . Other Boolean operators such as implication (⇒)
and equivalence (⇔) can be expressed using the grammar
rules and semantics given in definitions 2.2 and 2.3. More
details on MTL grammar and semantics can be found in [11].
Satisfaction of a temporal specification φ by a trajectory xt0
will be denoted as xt0 |= φ.
III. P ROBLEM F ORMULATION AND S OLUTION
We consider a planning problem to periodically survey
some selected areas in a given workspace. There are specific
time bounds, associated with the regions, by which the
surveillance has to be finished. Given the system dynamics
(1), the objective is to find a suitable control law that
will steer the robot in the survey area so that all regions
are surveyed within the time bound and that control will
optimize some cost function as well. The surveying task and
its associated timing constraints and safety constraints can
be expressed formally by Metric Temporal Logic (MTL).
Let φ denote the MTL formula for the surveying task, and
J(x(t, u), u) be a cost function to make the path optimal in
some sense. We formally present our planning objective as
an optimization problem given in Problem (3.1)
Problem 3.1:
min
J(x(t, u), u(t))
u
x(t + 1) = f (t, x(t), u(t))
xt0 |= φ
In the following section we are going to discuss the linearization techniques for the dynamics of the robot, and our
approach to translate an MTL constraint as linear constraints.
subject to
A. Linearized Dynamics of the Robot
Since we are interested in solving the planning problem
as an optimization problem with mixed integer and linear
constraints, we need to represent the dynamics of the robot
as a linear constraint to the optimization problem (3.1). We
will consider surveying some given areas by ground robots as
well as aerial robots. The dynamics of the autonomous aerial
robot and ground robot are given in (3) and (4) respectively.
1) Quadrotor Model: To capture the dynamics of the
quadrotor properly, we need two coordinate frames. One of
them is a fixed frame and will be named as earth frame,
and the second one is body frame which moves with the
quadrotor. The transformation matrix from body frame to
earth frame is R(t). The quadrotor dynamics has twelve
state variables (x, y, z, vx , vy , vz , φ, θ, ψ, p, q, r), where ξ =
[x, y, z]T and v = [vx , vy , vz ]T represent the position and
velocity of the quadrotor w.r.t the body frame. (φ, θ, ψ) are
the roll, pitch and yaw angle, and Ω = [p, q, r]T are the rates
of change of roll, pitch and yaw respectively.
The Newton-Euler formalism for the quadrotor rigid body
dynamics in earth fixed frame is given by:
ξ˙ = v
v̇ = −ge3 +
F
Re3
m
(3)
Ṙ = RΩ̂
Ω̇ = J −1 (−Ω × JΩ + u)
where g is the acceleration due to gravity, e3 = [0, 0, 1]T , F
is the total lift force and u = [u1 , u2 , u3 ]T are the torques
applied. F and u are the control inputs. More details on
the quadrotor dynamics can be found in [16], [17]. For this
work, we linearize the dynamics (3) about the hover with
yaw constraint to be zero, as it has been done in [14]. Since
ψ is constrained to be zero, we remove ψ and r from our
system and make the system ten dimensional. Consequently,
we only need three control inputs, F, u1 , and u2 for the
system. The linearized model is the same as what is done in
[14],
[18]. The system matrices
for thelinearized model are:
0 I 0 0
0
0
0
0 g
0 0 −g 0 0
0
0
; B =
A=
0 0
1/m
0 0
0
I
0
0
−1
0
I2×3 J
0 0
0
0
1 0 0
I2,3 =
0 1 0
All zero and identity matrices in A and B are of proper
dimensions.
2) Car-Like Model: We also investigate our approach on
a car-like dynamical system (4). The system has three state
variables: positions (x, y), heading angle θ.
ẋ
cos(θ) 0
ẏ = sin(θ) 0 u1
(4)
u2
0
1
θ̇
where u1 and u2 are the control inputs. We linearize this
nonlinear model about several values for θ and depending
on the value of θ, the closest linearization was used to drive
the linearized system. The linearization is similar to what is
suggested in [14]. The linearized system matrices at θ̂ are
given by:
0 0 −û1 sin(θ̂)
cos(θ̂) 0
A = 0 0 û1 cos(θ̂) ; B = sin(θ̂) 0
0 0
0
0
1
B. Mixed Integer Linear Constraints
In this section, we demonstrate our approach to translate a
time-bounded temporal logic formula to linear constraints on
state variables and inputs. The easiest example of it would be
how to express the temporal constraint that x(t) lies within
a convex polygon P at time t. This simple example will
serve as a building block for other complicated temporal
operators. Any convex polygon P can be represented as an
intersection of several halfspaces. A halfspace is expressed
by a set of points, Hi = {x : hTi x ≤ ki }. Thus, x(t) ∈ P
is equivalent to x(t) ∈ ∩ni=1 Hi = ∩ni=1 {x : hTi x ≤ ki }.
The temporal constraint that x(t) will be inside P for all
t ∈ {t1 , t1 + 1, · · · t1 + n} can be represented by the set of
linear constraints {hTi x(t) ≤ ki } for all i = {1, 2, · · · , n}
and ∀t ∈ {t1 , t1 + 1, · · · t1 + n}.
We adopted a method similar to the method used in [13]
to translate temporal constraint φ into mixed integer linear
constraints. We extend it to incorporate duration for task
completion and loop constraints of the trajectory. Comparing
to [14], we only enforce the loop constraints at the trajectory
level instead of the temporal logic level, since the moving
obstacles is not periodic in nature. The planning process will
be repeated when the vehicle returns to the initial point.
In a polygonal environment, atomic propositions (AP), p ∈
Π, can be related to states of the system using disjunction and
conjunction of halfspaces. In other words, the relationship
between measured outputs such as location of the vehicle and
the halfspaces defines the proposition used in the temporal
logics. Consider the convex polygon case and let zit ∈ {0, 1}
be the binary variables associated with halfspaces {x(t) :
hTi x(t) ≤ ki } at time t = 0, ..., N . We enforce the following
constraint zit = 1 if and only if hTi x(t) ≤ ki by adding the
linear constraints,
hTi x(t) ≤ ki + M (1 − zit )
(5)
hTi x(t) ≥ ki − M zit +
where M is a large positive number and is a small positive
number. If we denote PtP = ∧ni=1 zit , then PtP = 1 if and
only if x(t) ∈ P. This can be extended to the nonconvex
case by decomposing the polygon to convex ones and linking
them using disjunction operators. As discussed later in this
section, the disjunction operator can also be translated to
mixed integer linear constraints.
We will use Fφ (x, z, u, t) to denote the set of all mixed
integer linear constraints corresponding to the temporal logic
formula φ. Once we have formulated Fp (x, z, u, t) for atomic
propositions p, we can find Fφ (x, z, u, t) for any MTL
formula φ. The next essential part of the semantics of MTL
is the Boolean operations, such as ¬, ∧, ∨. Similarly these
operators can be translated into linear constraints. Let t ∈
{0, 1, ..., N }, Ptφ be the continuous variables within [0,1]
associated with formula φ made up with propositions p ∈ Π
at time t.
• φ = ¬p is the negation of an atomic proposition, and it
can be modeled as
Ptφ = 1 − Ptp .
•
(6)
The conjunction operation, φ = ∧m
i=1 pi , is modeled as
Ptφ ≤ Ptpi ,
i = 1, ..., m,
m
X
Ptpi ,
Ptφ ≥ 1 − m +
(7)
(8)
i=1
Similarly, the temporal operators can be modeled using
linear constraints as well. Let t ∈ {0, 1, ..., N − t2 }, where
[t1 , t2 ] is the time interval used in the MTL.
• Eventually: φ = ♦[t1 ,t2 ] p is equivalent to
Ptφ ≥ Pτp ,
Ptφ ≤
t+t
X2
τ ∈ {t + t1 , ..., t + t2 }
(9)
Always: φ = [t1 ,t2 ] p is equivalent to
Ptφ ≤ Pτp ,
Ptφ ≥
t+t
X2
τ ∈ {t + t1 , ..., t + t2 }
(10)
Pτp − (t2 − t1 ),
τ =t+t1
•
Until: φ = pU[t1 ,t2 ] q is equivalent to
atj ≤ Pqj
atj ≤
Ppk
(∧k=j−1
Ppk ) ∧ Pqj .
k=t
Other combinations for different temporal operators are also
straight forward and we are not enumerating them for the
sake of space, but one can easily derive them by using (6)
through (11).
Using this approach, we translate the given high level specification in MTL (xt0 |= φ) to a set of mixed integer linear
constraints Fφ (x, z, u, t). At the end, we add the constraint
P0φ = 1, i.e. the overall specification φ is satisfied. Since
Boolean variables are only introduced when halfspaces are
defined, the computation cost of MILP is at most exponential
to the number of halfspaces times the discrete steps N .
IV. C ASE S TUDY AND D ISCUSSION
Pτp
τ =t+t1
•
=
t+t
_2
j=t+t1
The disjunction operator, φ = ∨m
i=1 pi , is modeled as
Ptφ ≥ Ptpi , i = 1, ..., m,
m
X
Ptφ ≤
Ptpi ,
The until formulation (11) is obtained similarly to [13]. For
MTL, it is modified by noticing the following equality,
Ptφ
i=1
•
Fig. 1. Workspace setup of the first test case. Blue area represents an
obstacle moving from right to left. Grey areas represent static obstacles.
Green areas represent survey areas used in the temporal logic. The shaded
areas around obstacles represent boundary of the obstacles for discrete
planning, so that the obstacles can be avoided even for continuous dynamics.
The resulting 2D trajectory of the quadrotor for φ1 is shown as blue dots.
The resulting motion is counter clockwise.
j ∈ {t + t1 , · · · , t + t2 },
k ∈ {t, · · · , j − 1}, j ∈ {t + t1 , · · · , t + t2 },
We apply our method for solving mission planning with
finite time constraints on two different workspaces. Both
workspaces contain static and moving obstacles.The experiments are run through YALMIP-CPLEX on a computer
with 3.4GHz processor and 8GB memory. We performed the
simulation for both a quadrotor model and a car model.
The first environment is the one shown in Fig. 1, where
blue and gray areas represent moving and static obstacles
respectively, and green areas represent survey targets.
Let the temporal specification be the following
φ1 = ♦[0,2] A ∧ ♦[0,2] B ∧ ♦[0,2] C ∧ ¬O
φ2 = ♦[0,2] A ∧ ♦[0,2] B ∧ ♦[0,2] C ∧ ¬O ∧ ¬B U[0,N ] A
Ppk − (j − t) j ∈ {t + t1 , · · · , t + t2 }, where O represents the static and moving obstacles, N is the
k=t
horizon of the planning trajectory. Such N can be generally
t+t
2
X
obtained by performing a feasibility test using MILP solver
Ptφ ≤
atj ,
(11)
starting from N = 2, and increasing N until finding a
j=t+t1
feasible one. For the purpose of comparing different temporal
Ptφ ≥ atj j ∈ {t + t1 , · · · , t + t2 }.
constraints, we choose a feasible horizon N = 50 for both
atj ≥ Pqj +
j−1
X
Fig. 2. Time-state-space representation of the environment and resulting
trajectory for the quadrotor with temporal constraints φ1 . The vehicle starts
from (0.5,0.5) and surveys area b, c and a sequentially.
Fig. 3. Time-state-space representation of the environment and resulting
trajectory for the quadrotor with temporal constraints φ2 . Because of the
additional ordering requirement, the vehicle covers area a first before visiting
area b.
MTL specifications φ1 and φ2 . The specification φ1 requires
the vehicle to visit areas A, B and C eventually and stay
there for at least 2 time units, while avoiding obstacles O.
φ2 adds an additional requirement on the ordering between A
and B, so that region A has to be visited first.
PN We consider
the cost function, J, to be minimized is t=0 |u(t)|. The
dynamics of the vehicles are discretized at a rate of 2Hz.
The resulting trajectory for the quadrotor with temporal
specification φ1 is plotted in time-state-space in Fig. 2. The
projection of the trajectory on the workspace is shown in
Fig. 1. The motion of the quadrotor in Fig. 1 is counter
clockwise. The quadrotor safely avoids the moving obstacle
by navigating through the area after the obstacle passes.
The survey duration in individual area also satisfies the
requirements as shown in Fig. 1. The quadrotor stays within
each survey area for 5 discrete time units which is equivalent
to 2s duration. These plots show a better realization of
visiting individual areas comparing to [13] and [14]. In their
results, trajectory visits targeted areas for only 1 discrete
time unit which is not desired for most surveillance tasks.
Additional local planning is possible to generate a sweeping
pattern so that the target areas can be covered by onboard
sensors in designed durations.
The resulting trajectory for the vehicle with specification
φ2 is shown in time-state-space in Fig. 3. As can be verified
from the trajectory, the vehicle travels first to area A before
Fig. 4. Time-state-space representation of the environment and resulting
trajectory for the car model with temporal constraints φ1 . The result for φ2
is similar since the optimal solution for φ1 already goes through area A
first.
Fig. 5. Workspace setup of the second test case. The moving obstacle
moves from the the bottom left corner to upper right corner. 2D trajectory
of the vehicle is shown as blue dots. The resulting motion is clockwise.
visiting B as specified. The CPLEX solver returns the
solution for the first trajectory in 20.8 sec while the second
one takes 34.7 sec. The additional continuous variables used
in the until encoding have large influence in the performance.
One of the possible future research directions would be
finding different encoding of the until operator to improve
the speed of the algorithm.
The result of specification φ1 for the car model is shown
in Fig. 4, where the small blue arrow associated with each
blue dot indicates the instantaneous heading of the vehicle.
The computation time is 150s. The longer computation time
is caused by the additional binary variables introduced by
the linearization of the dynamics at various heading angles.
The second environment considers a fast moving obstacle
that moves across the workspace diagonally, and hence the
vehicle has to adjust its motion accordingly. The environment
is shown in Fig 5. Similar to the previous example, it shows
the motion of the moving obstacle, static obstacles and
survey areas.
The temporal logic specifications are similar to the previous one but have an additional area to be visited. We also
tested the case when certain area has to be visited first. The
result is similar to previous cases, so we only show the plots
Fig. 6. Time-state-space representation of the environment and resulting
trajectory for the quadrotor with temporal constraints φ3 . The planned
trajectory is very close to the dynamic obstacles.
problem as a Linear Programming problem. We have considered polygonal environments for our case studies, but if the
environment is not polygonal, one can approximate it with a
polygonal environment. We have used a binary variable (z)
with each halfspace, so if the polygonal approximation of
the environment contains too many halfspaces, the problem
would be complex.
Due to space constraints, we have reported the case studies
only for quadrotor and car dynamics. The simulation results
show promising performance of our approach to find an
optimal solution. We consider dynamic but deterministic environments and uncertainties in the dynamics of the robot are
also not considered. There are many possible directions such
as planning in uncertain environment, stochastic dynamics of
the robots or multi-robot cooperative planning that one might
consider as an extension of this framework.
R EFERENCES
Fig. 7. Time-state-space representation of the environment and resulting
trajectory for the car model with temporal constraints φ3 . The result for φ4
is the same since the optimal solution already goes through A first.
for φ3 .
φ3 = ♦[0,2] A ∧ ♦[0,2] B ∧ ♦[0,2] C ∧ ♦[0,2] D ∧ ¬O
The resulting trajectory in time-state-space for the quadrotor with temporal specification φ3 is plotted in Fig. 6. The
projected trajectory on the workspace of the robot is shown
in Fig. 5. The motion of the vehicle is clockwise in Fig.
5. As can be seen from Fig. 6, the quadrotor safely avoids
the moving obstacle and the static ones nearby. The survey
duration in individual area also satisfies the requirements
as shown in Fig. 5. The computation time is 138.8s. The
increase in computation time is because the complex environment introduces more binary variables.
The result of specification φ3 for the car model is shown
in Fig. 7, where the blue arrows indicate the heading of the
vehicle. The computation time is 500s.
V. C ONCLUSION
In this paper, we have presented an optimization based
approach to plan the trajectory of a robot in a dynamic
environment to perform some temporal task with finite time
constraints. Our approach is simple in the sense that it
translates the time constraints and the temporal specifications
as linear constraints on the state and input variables. We have
linearized the dynamics of the robot in order to formulate the
[1] J. Latombe, Robot motion planning, ser. Kluwer international series
in engineering and computer science. Boston: Kluwer Academic
Publishers, 1991.
[2] S. M. LaValle, Planning algorithms. Cambridge university press,
2006.
[3] H. M. Choset, Principles of robot motion: theory, algorithms, and
implementation. MIT press, 2005.
[4] W. Xi, X. Tan, and J. S. Baras, “A hybrid scheme for distributed control
of autonomous swarms,” in Proceedings of the 2005 American Control
Conference. IEEE, 2005, pp. 3486–3491.
[5] R. Sharma, “Locally efficient path planning in an uncertain, dynamic
environment using a probabilistic model,” Robotics and Automation,
IEEE Transactions on, vol. 8, no. 1, pp. 105–110, 1992.
[6] S. M. LaValle and R. Sharma, “On motion planning in changing, partially predictable environments,” The International Journal of Robotics
Research, vol. 16, no. 6, pp. 775–805, 1997.
[7] C. Baier, J.-P. Katoen et al., Principles of model checking. MIT press
Cambridge, 2008, vol. 26202649.
[8] E. M. Clarke, O. Grumberg, and D. Peled, Model checking. MIT
press, 1999.
[9] M. M. Quottrup, T. Bak, and R. Zamanabadi, “Multi-robot planning:
A timed automata approach,” in Robotics and Automation, 2004.
Proceedings. ICRA’04. 2004 IEEE International Conference on, vol. 5.
IEEE, 2004, pp. 4417–4422.
[10] H. Kress-Gazit, G. E. Fainekos, and G. J. Pappas, “Temporal-logicbased reactive mission and motion planning,” IEEE Transactions on
Robotics, vol. 25, no. 6, pp. 1370–1381, 2009.
[11] R. Koymans, “Specifying real-time properties with metric temporal
logic,” Real-time systems, vol. 2, no. 4, pp. 255–299, 1990.
[12] J. Ouaknine and J. Worrell, “Some recent results in metric temporal
logic,” in Formal Modeling and Analysis of Timed Systems. Springer,
2008, pp. 1–13.
[13] S. Karaman, R. G. Sanfelice, and E. Frazzoli, “Optimal control of
mixed logical dynamical systems with linear temporal logic specifications,” in 47th IEEE Conference on Decision and Control. IEEE,
2008, pp. 2117–2122.
[14] E. M. Wolff, U. Topcu, and R. M. Murray, “Optimization-based
trajectory generation with linear temporal logic specifications,” in Int.
Conf. on Robotics and Automation, 2014.
[15] S. Karaman and E. Frazzoli, “Vehicle routing problem with metric
temporal logic specifications,” in Decision and Control, 2008. CDC
2008. 47th IEEE Conference on. IEEE, 2008, pp. 3953–3958.
[16] N. Michael, D. Mellinger, Q. Lindsey, and V. Kumar, “The grasp
multiple micro-uav testbed,” IEEE Robotics & Automation Magazine,
vol. 17, no. 3, pp. 56–65, 2010.
[17] L. R. Garcı́a Carrillo, A. E. Dzul López, R. Lozano, and C. Pégard,
“Modeling the quad-rotor mini-rotorcraft,” Quad Rotorcraft Control,
pp. 23–34, 2013.
[18] D. J. Webb and J. van den Berg, “Kinodynamic rrt*: Asymptotically
optimal motion planning for robots with linear dynamics,” in 2013
IEEE International Conference on Robotics and Automation (ICRA).
IEEE, 2013, pp. 5054–5061.
| 3cs.SY
|
arXiv:1710.11087v1 [cs.CV] 30 Oct 2017
An Integrated Approach to Crowd Video Analysis: From Tracking to Multi-level
Activity Recognition
Neha Bhargava
Indian Institute of Technology Bombay
India
Subhasis Chaudhuri
Indian Institute of Technology Bombay
India
[email protected]
[email protected]
Abstract
priate to estimate them together instead of sequentially. See
Figure 1 as an example of this hierarchical structure where
the atomic actions of the individuals are all standing, there
are two groups each with group activity as talking and thus
leading to the collective activity also as talking. These inherent dependencies among the various components motivate us to explore this idea of simultaneous recognition of
tracks, groups and activities. We propose a novel approach
to build on the detections to obtain the tracks, groups and
activities in a causal framework, i.e. without considering
future frames into estimation process. We solve this unified
problem in two passes. The first pass consists of finding the
graph structure that corresponds to the track association and
group detection. We propose an linear programming based
formulation for the same. The second pass involves activity
recognition at various levels of granularity. We formulate
this problem under the structured SVM formulation [29].
We present an integrated framework for simultaneous
tracking, group detection and multi-level activity recognition in crowd videos. Instead of solving these problems independently and sequentially, we solve them together in a
unified framework to utilize the strong correlation that exists among individual motion, groups and activities. We explore the hierarchical structure hidden in the video that connects individuals over time to produce tracks, connects individuals to form groups and also connects groups together to
form a crowd. We show that estimation of this hidden structure corresponds to track association and group detection.
We estimate this hidden structure under a linear programming formulation. The obtained graphical representation
is further explored to recognize the node values that corresponds to multi-level activity recognition. This problem
is solved under a structured SVM framework. The results
on publicly available dataset show very competitive performance at all levels of granularity with the state-of-the-art
batch processing methods despite the proposed technique
being an online (causal) one.
1. Introduction
A crowd video analysis system first detects the individuals and then tracks them over time. These tracks are used
for higher level applications such as group detection and
activity recognition. This approach is sequential in nature
whereas the various components of the system are highly
correlated and influence each other. For example, a particular group activity motivates its group member for a particular action and all the groups together define the crowd
activity. On the other hand, group behavior is influenced
by its members and the overall crowd behavior. Effectively,
these components - individual’s motion, groups, group activity and collective activity are correlated and can be expressed in a hierarchical structure. Hence it is more appro-
Figure 1. Illustration of hierarchical structure present in a video.
It represents video in terms of atomic actions, groups, group activities and collective activity. There are 6 individuals who are
standing and forming two groups with group activities as talking
and hence the collective activity is also talking.
In this paper, the term action refers to an atomic movement performed by a single person, the term group activity denotes an activity performed by a group and collective
activity refers to the activity performed by all the groups
collectively. The paper is organized as follows. The next
section discusses the related work followed by our contri1
(a)
(b)
Figure 2. (a). Graphical representation showing the hierarchical structure present in a video. The first layer Tk−1 and second layer dk are
fully connected since the track association is unknown until estimated. The third layer corresponds to the actions hk associated with the
detections. The third and fourth layers are again fully connected since the group association is unknown. The final layer corresponds to the
connection between the overall activity and the group activities. (b). The figure shows a possible graphical structure obtained after track
association and group detection. xi , xg and x0 are the respective features for a person, group and collectively that are derived from the
video frames and to be used for activity recognition.
butions. The proposed model is described in Section 3. The
subsequent Sections 4, 5 and 6 elaborate on frameworks for
multi-target tracking with group detection followed by activity recognition. Experimentation details are provided in
Section 7 and the paper concludes in Section 8.
[2, 3, 7, 13, 14]. Amer et al. in [2] proposed a hierarchical
random field based modeling of higher order temporal dependencies of video features. Lan et al. in [14] jointly estimate actions, pairwise interactions and group activity. However, they assume the availability of action labels and they
do not handle track association. Choi and Savarese in [7]
proposed a hierarchical model and combine the problems
of tracklet association and multi-level activity recognition
(action, pairwise interaction and collective). All these methods either assumed availability of action label or trackelets
whereas our proposed framework requires only detections.
2. Related work
The task of multi-target tracking (MTT) is to correctly
associate all the detections (or tracklets) corresponding to
each individual. Linear programming (LP) based global
optimization for MTT is a popular approach. Many approaches formulate MTT either as min-cost flow optimization problem or MAP and use LP to find the global optimum. [4, 6, 9, 20, 35]. Recently, the approaches of utilizing social behavior to improve tracking are gaining attention [5, 15, 19, 22, 34]. The idea is to simultaneously associate a detection to a track and to a group. Our approach for
obtaining groups is similar to that of [22] where they combine track association with grouping under a LP framework.
Since the number of groups K is unknown, they run the algorithm with all possible values of K with a linear penalty.
Our proposed method exploits the group information from
the previous time instant and does not require to run for all
values of K resulting in a fast convergence.
Due to its various applications in video surveillance, activity recognition has been an active area of research. The
survey on action and activity recognition can be found in
[21, 30, 32]. There are many works dealing with single
person action [12, 17, 18, 26, 31] and with single group activity recognition [8, 11, 16, 24, 25]. Recently, researchers
have started exploring the problem of joint recognition of
these actions and activities under a hierarchical framework
Our work in this paper advances the existing approaches
and add one more intermediate layer (i.e. grouping layer) in
the hierarchy as shown in Figure 2a and explained in Section 3. The main contributions of this work are as follows:
1. We propose a hierarchical graphical structure that
combines multi-target tracking, group detection and
activity recognition under an unified framework.
2. We built a causal framework that takes only human
detections as an input and outputs tracks, groups and
activities at each time step.
3. We propose an iterative linear programming based
method for simultaneous track association and group
detection.
4. We propose an approach for simultaneous recognition
of activities at various levels of granularity under a
structured SVM framework.
5. To make it suitable for real-time applications, we provide a fast algorithm for both training and inferencing.
2
3. The Proposed Model
nition (Eq. 3). i.e.
In this section, we discuss the proposed model. Let
yk ∈ Y be the collective activity at time k with group activity vector gk = [g1k , g2k , ..., gmk ] where gik ∈ G is the
activity of ith group and m be the number of groups present
at time instant k. The atomic activity vector is denoted as
hk = [h1k , h2k , ..., hkN ] with hik ∈ H as the atomic activity of the ith person and N is the total number of persons
present at time k. Let Tk−1 denotes the estimated tracks
available till time k − 1 and Gk be the group label vector
of length N where its ith entry denotes the group label for
the ith detection at time k. Let dk denotes the detections
at time k. By a detection, we mean a person’s location in
the form of a bounding box. Now the problem statement
is as follows: Given the detections dk and tracks Tk−1 at
time k, the goals are (a) to associate these detections dk to
the appropriate tracks in Tk−1 to get Tk , (b) identify the
group label vector Gk and (c) recognize the atomic, group
and collective activities (hk , gk , yk ). Let zk = [yk , gk , hk ]
be the activity vector for notational convenience. The problem is formulated under the score maximization framework
with a linear function as,
z∗k , G∗k , T∗k
= arg
G∗k , T∗k = arg max w1 T Φ1 (Gk , Tk ; dk , Tk−1 , zk−1 )
Gk ,Tk
(2)
z∗k = arg max w2 T Φ2 (zk ; Tk , Gk , zk−1 ).
zk
(3)
The next two sections discuss these two sub problems in
detail.
4. Multi target tracking (MTT) and group detection
We estimate the tracks and groups together under a linear programming framework. Let N number of detections
and Ng (unknown) number of groups be present at time k.
Let Tk−1 be a set of T trajectories present at k − 1. We define Ψ ∈ {0, 1}N ×T as the track association matrix where
Ψij = 1 indicates association of the ith detection with the
j th track. We also define Ω ∈ {0, 1}N ×Ng as the group
association matrix where Ωil = 1 indicates that the ith
detection belongs to the lth group. Then the optimization
equation to estimate Ψ and Ω is as follows:
Ψ∗ , Ω∗ = arg max
T
max w Φ(zk , Gk , Tk ; dk , Tk−1 ).
Ψ,Ω
zk ,Gk ,Tk
(1)
N X
T
X
Ψij Mij +λ
i=1 j=1
Ng
N X
X
Ωil Cil (4)
i=1 l=1
s.t.
The problem is illustrated as a graphical model in Figure 2a. There are N detections with an unknown number
Ng of groups at time k. Tik−1 denotes the ith track available at time k − 1. The root node denotes the collective
activity which is connected to the group activity nodes. The
group activity nodes are also connected to the atomic activity nodes of the group members. The graph emphasizes that
collective activity is related to the group activities while a
group activity is correlated both with its members’ actions
and the collective activity of the scene. Since the track association (Tk−1 ↔ dk ) and group information (hi ↔ gj )
are unknown - (a) every node Tik−1 is connected to all the
detection nodes and (b) each node of the action layer is
connected to all the group activity nodes as shown in Figure 2a. Once we know the track association and group labels, the corresponding graph structure is known. One possible graph structure corresponding to Figure 2a is shown
in Figure 2b. Here xi , xg and x0 are the respective observations for a person, group and collective entity defining
the video. The procedure to obtain these observations are
discussed later.
Ψij , Ωij ∈ {0, 1},
Ng
X
Ωil ≤ 1 ∀i,
j=1
N
X
Ψij ≤ 1 ∀j,
i=1
T
X
Ψij ≤ 1 ∀i,
(5)
j=1
where λ ∈ R+ is a weighing factor that decides the balancing between the group association and track association
scores. Mij ∈ [−1, 1] is the compatibility score of the ith
detection with the j th track based on motion and visual similarity, and Cil ∈ [−1, 1] is the compatibility score for the
ith detection with the lth group based on motion, spatial
and pose compatibility. The constraints in Eq. 5 ensure that
each detection is assigned to at most one track and to one
group. It also ensures that each track gets at most one detection while there is no such constraint for the group. The
next sub-sections discuss the construction of compatibility
matrices M and C.
4.1. Construction of M
We break this complete problem in two sub problems (a) Graph structure estimation: This corresponds to track
association and group detection (Eq. 2), and (b) Node value
estimation: This corresponds to multi-level activity recog-
M ∈ RN ×T is a score matrix for track association where
Mij is the score of assigning ith track to the j th detection.
It is calculated based on visual similarity, spatial proximity
3
and the motion compatibility between the ith detection with
the j th track.
The visual similarity score is based on color histogram
matching of the j th detection with that at the last location
of the ith track. The spatial proximity is the measure of
closeness of the j th detection from the ith track. Lastly, the
motion compatibility is based on the velocity consistency
when the j th detection is added to the ith track. By combining these three scores, we obtain
Mij =
3
X
(n)
αn (2e−βn ||xi
(n)
−xj
||22
− 1),
Figure 3. (a) Illustration of score calculation for pose compatibility between a candidate group and a detection. In the figure, (p1, p2, p3) form a group with group’s FoV as ABCDE.
F IDE and GHIF are FoVs of detections q1 and q2 respecIDE
tively. Therefore S(q1, (p1, p2, p3)) = ABCDE∩F
= 1
ABCDE
while S(q2, (p1, p2, p3)) = ABCDE∩GHIF
=
0.
(b)
Illustration
ABCDE
of field of view (FoV) for a person. The arrow signifies the pose
direction. The boundary rectangle corresponds to the observed
image. Best viewd in color.
(6)
n=1
where α and β are the weight and normalizing vectors respectively. x(1) represents color histogram, x(2) is the location and x(3) is the velocity estimate. We keep αn = 31 ,
β1 = 1 and β3 = 1 in the experiments. β2 is chosen as the
inverse of height of the bounding box.
4.2. Construction of C
from the previous time instant k − 1 to get an initial estimate of C for the present detections. Then we solve Eq. 4
to get the optimal Ω. If any row (say ith ) of Ω consists of
all zeros, it indicates that ith detection does not belong to
any of the groups. In such a case, we add one more group
to the list with ith detection as its founding member and
again solve for Eq. 4. We iteratively do this until we get
group assignment for all the detections. Also, to discourage
formation of singleton groups (with one member) , we remove such groups before the start of the iterative algorithm
at each frame. To initialize in the first frame of the video,
we consider Ng = 1 i.e. all the detections belong to a single group. This iterative method is detailed in Algorithm 1.
The algorithm is found to converge within a few iterations
only. In the worst case when all the detections form singleton groups and different from the groups present at previous
time instant, the algorithm takes N number of iterations.
C ∈ RN ×Ng is a score matrix for group association
where Cil is the score of assigning the ith detection to the
lth group. It is calculated based on the motion similarity,
spatial closeness and the pose compatibility between the ith
detection and the lth group. The group location and group
velocity are defined as the averages over the locations and
velocities of the members, respectively. To compute motion
similarity between ith detection and lth group, we first find
the track associated with the ith detection from Ψ. We then
compute the velocity compatibility between the obtained
track with the lth group. To obtain pose compatibility, we
first calculate the interacting zone of the group formed by
the members. The normalized intersection of the field of vision of the detection with the group’s interacting zone gives
the score for the pose compatibility. This is illustrated in
Figure 3. Let p1, p2 and p3 form a group and q1, q2 are the
detections. We define field of view (FoV) for a person as
the complete area in the pose direction as illustrated in Figure 3(b). The pose compatibility between a detection d and
oV (d)
a group ḡ is defined as S(d, ḡ) = F oV F(ḡ)∩F
, where
oV (ḡ)
F oV (ḡ) is the intersection of FoVs of the group members.
In Figure 3, q1 has high compatibility score while q2 has
zero score since it has no intersection. The pose compatibility is added to discourage the non-facing persons forming a
group. Finally, we combine the three scores obtained from
motion, spatial and pose compatibilities to construct C as
done previously for M .
Algorithm 1 Algorithm to obtain Ψ and Ω
1: procedure
2:
t=0, Gtk = Gk−1
3:
Solve Eq. 4 to get Ψ∗ and Ω∗
4:
d = set of detections without any group assignment
5:
while d is non-empty do
• Add one column to Ω∗ with one of the detections
from d
• Solve Eq. 4 to get Ψ∗ and Ω∗
• Update Gtk and id
• t←t+1
6:
return Gtk , Ψ∗ and Ω∗
end
4.3. Iterative algorithm to obtain Ψ and Ω
Since the group information is initially unknown at time
k, we do not know the score matrix for group association
i.e. C. Hence, we propose an iterative algorithm to construct C and to solve Eq. 4. We use the group information
4
5. Activity recognition
4. Group Activity Potential:
wT3 φ3 (g, H(hg )) defines the compatibility of atomic
activities of the group members with the group activity.
It is modeled as
X
wT3 φ2 (g, H(hg )) =
wT3b 1(g = b)H(hg ), (12)
Solution of Eq. 4 gives an estimate of the latent graph
structure (e.g. Figure 2b). The next problem is to estimate
the optimal node values of this graph structure at all time
instants causally. In other words, the aim is to recognize the
activities at individual, group and collective levels.
The problem is cast under a linear energy function
framework as
Φ(y, g, h, x) = wT φ(y, g, h, x),
b∈G
where H(hg ) is the histogram of atomic activities of
the group members.
(7)
5. Atomic Action - Image Potential:
wT4 φ4 (hi , xi ) defines the compatibility of the individual’s observation with the atomic activity and modeled
as
X
wT4 φ4 (hi , xi ) =
wT4c 1(hi = c)xi .
(13)
where φ calculates the compatibility of activities (y, h, g)
and the observations x = {x0 , xg , xi }. We follow the motivation of [7] to solve the problem. As said before, x contains individual, group and collective features which are obtained once Tk and Gk are known. We take advantage of
hierarchical structure and decompose Φ(y, g, h, x) according to the graph Figure 2b as follows:
Φ(y, g, h, x) =
+
Ng
P
wT0 φ0 (y, x0 )
Ng
P
wT2 φ2 (gi , xgi ) +
i=1
+
c∈H
6. Inference and Learning for activity recognition
wT1 φ1 (y, H(g))
In this section, we discuss the learning and inference algorithms. Given a graph structure at any time instant k (i.e
Tk and Gk ), we need to recognize the activities at all levels.
Solution of Eq. 7 provides the inference about the unknown
node variables (y, g, h). We use the structured SVM framework [10] to learn w, and an iterative alternate optimization
method for the inference. The next two subsections discuss
both these algorithms in detail.
wT3 φ3 (gi , H(hgi ))
i=1
+wT4
N
P
φ4 (hi , xi ),
(8)
i=1
where w
=
[w0 , w1 , w2 , w3 , w4 ] and φ
[φ0 , φ1 , φ2 , φ3 , φ4 ]. Each term is described as follows:
=
1. Collective Activity - Image Potential:
It is the compatibility score of collective activity y ∈ Y
with the collective observation x0 . It is modeled as
X
wT0 φ0 (y, x0 ) =
wT0a 1(y = a)x0 ,
(9)
6.1. Inference
Given the learned model parameters w, the inference
problem is to find the optimal collective activity y ∗ , group
activity vector g∗ and atomic activity vector h∗ for the input
x. i.e.
a∈Y
where w0 = [w01 , w02 , ..., w0|Y| ] and 1(:) is an indicator function.
y ∗ , g∗ , h∗ = arg max wT φ(y, g, h, x).
y,g,h
2. Collective - Group Activity Potential:
wT1 φ1 (y, g) is the compatibility of group activities g
with the collective activity y and defined as
X
wT1 φ1 (y, H(g)) =
wT1a 1(y = a)H(g), (10)
We use an iterative method to solve Eq. 14. We initialize y, g and h with the values in the previous time step if
available or random otherwise. The method is detailed in
Algorithm 2.
6.2. Learning
a∈Y
Given a training data D = {xi , Gi , hi , gi , y i } ∀{i =
1, 2, ..., S} where S is the total number of training samples and Gi is the group label vector, the goal is to learn
the optimal weight vector w∗ . We use 1-Slack structured
SVM with margin-rescaling [10] where there is only a single slack variable ξ for all the constraints. Let us define
zi = [hi , gi , y i ] to simplify the notations. The optimization
equation is as follows:
where H(g) is the histogram of group activities.
3. Group Activity - Image Potential:
wT2 φ2 (g, xg ) defines the compatibility of group activity g ∈ G with the group observation xg as
wT2 φ2 (g, xg ) =
X
wT2b 1(g = b)xg .
(14)
(11)
b∈G
5
Algorithm 2 Inference algorithm
1: procedure I NFERENCE
2:
Initialize y 0 , g0 , h0
3:
t=0, err=1000 , = 0.01
4:
while err > do
5:
y t+1 ← arg max{wT0 φ0 (y, x0 ) + wT1 φ1 (y, H(gt ))}
y
6:
git+1 ← arg max{wT1 φ1 (y t+1 , H(gt \git , g)) + wT2 φ2 (gi , xg ) + wT3 φ3 (g, H(htg )}, ∀i = 1 : Ng
7:
ht+1
← arg max{wT3 φ3 (gjt+1 , H(ht \hti , h)) + wT4 φ4 (xi , h)}, ∀i = 1 : N , j: group index of ith person
i
g
h
err ← 1+N1+Ng {1(y t 6= y t+1 ) + 1(gt 6= gt+1 ) + 1(ht 6= ht+1 )}
t←t+1
10:
return y t , gt and ht
end
8:
9:
w∗ = arg min 12 ||w||22 + Dξ
w
talking) and 8 poses (right, right-front, ..., etc.) are provided. Additionally, the authors of [7] have provided annotations for target correspondence, atomic action labels
(standing, walking) and 8 pairwise interaction labels. Since
we are interested in finding groups and group activities instead of pair-wise interactions, we provide annotations for
group labels and group activities (walking, waiting, queuing
and talking) after every 10 frames. We consider collective
activity as the major activity happening at a time. For example - if out of 5 groups, 3 or 4 groups are talking and one is
walking, we consider the overall activity as talking. Moreover, we differ in the definition of crossing from that mentioned in [7]. In this paper, we consider crossing happens
when two or more groups cross each other on the contrary
to road crossing used in [7]. We have re-annotated crossing
videos accordingly.
As is common in most feature tracking methods, we preprocess the videos for image stabilization. To do this, we
use a time window of 20 frames where the 1st frame acts
as the reference frame. The camera motion is compensated
in the subsequent frames with respect to it by estimating an
affine transformation between the reference frame and the
k th frame.
(15)
s.t. ∀zi :
S
S
1X
1 TX
∆(zi , z̄i ) − ξ. (16)
w
[φ(xi , zi ) − φ(xi , z̄i )] ≥
S
S
i=1
i=1
The loss function ∆(z̄, zi ) is defined as
|z|
∆(z̄, zi ) =
1 X
1(z̄j 6= zji ),
|z| j=1
(17)
where z̄ is any possible combination and zi is the actual
output corresponding to the ith input.
Since the number of constraints grows exponentially
with S, the cutting plane algorithm [10] constructs a set
of working constraints and optimize the function over this
set. This set is constructed by identifying the most violated
constraint for each data sample (xi , zi ) at each iteration.
Finding the most violated constraint for (xi , zi ) is again an
optimization problem and is as follows:
ẑi = arg max wT φ(z, xi ) + ∆(z, zi ).
z∈Z
7.2. Observations
(18)
The observations x consist of individual related features
xi , group level features xg and collective features x0 . The
individual observations xi ∈ R|P|×|H| include pose ∈ P
and action ∈ H. xg is the mean of the feature vectors of
the group members while x0 is the mean of feature vectors of all the individuals. Note that only pose and action
are not enough to discriminate between waiting and queue
since all the members possess the same pose and action. To
incorporate some discrimination, we additionally include a
pose-position compatibility to xg . The score is calculated
for all the pairs (i, j) of the group members and is defined
as |pT (di − dj )| where p is the pose vector corresponding
to the statistical mode of the member poses and di is the
This is same as our inference problem with an additional
term of ∆(z, zi ). We use the same method to solve this.
7. Discussions and Experiments
7.1. Dataset
We demonstrate the performance of the proposed
method on the commonly used collective activity dataset
provided in [7]. The dataset has 44 video clips composed
of different challenging videos. The annotations for 5 collective activities (crossing, waiting, queuing, walking, and
6
position of the ith member. Higher value of the score corresponds to queue since both the vectors are aligned in the
same direction while waiting will have a low value because
both the vectors are orthogonal to each other. This is illustrated in Figure 4. We append the mean value of the score
values obtained for all the pairs of the group to xg .
model indicates the effectiveness of combined estimation of
groups and tracks over independent track association.
Table 1. Table showing tracking performance
Baseline model Full model
ID Switches
22 (4.5%)
17 (3.7%)
7.4. Group detection performance
(a) Waiting
To evaluate the group detection performance, we use the
following clustering measures which are commonly used:
Purity [1], Rand Index [23] and Normalized mutual information (NMI) [33]. We compare with a baseline case
present within our framework which corresponds to group
association (second part of Eq. 4). The full model incorporates both track association and group association. The
quantitative results are given in Table 2.
(b) Queue
Figure 4. Illustration of pose-position compatibility score. The
arrows for p1 and p2 indicate their pose directions. Basic setup in
case of waiting and queue to be utilized to discriminate between
them. (a) In case of waiting, the persons p1 and p2 are standing
side by side, thereby creating a right angle between position vector
(p1-p2) with pose vector. (b) In case of queue, the persons p1 and
p2 are one after another and hence the position vector is aligned
with pose vector.
Table 2. Table showing group detection performance
Framework Purity Rand Index NMI
Baseline
0.82
0.75
0.65
Full
0.89
0.81
0.72
To learn a pose classifier, we fine-tune all the 19 layers
of the VGG [27] network on PARSE-27k [28] pedestrian
attribute dataset comprised of 27 thousand labeled training
images. To account for inherent order in poses, we modify cross entropy loss by penalizing misclassification. The
penalty is less for predicting nearby pose and high otherwise; For example, the penalty is less if the classifier predicts Right-Front for the true pose of Right while the penalty
is high if the prediction is Left.
We employ the following procedure to estimate action.
We fit lines separately on the x and y coordinates of the topleft and bottom right of the bounding box as a function of
time over 20 frames and use the estimated slopes to learn a
SVM classifier for atomic action classification. The reason
for considering both top-left and bottom right coordinates
of the bounding box is to capture the possible movement
along the viewing direction of the camera (i.e. effect of approaching and receding).
Again, the higher values of the clustering measures in the
full model indicates the effectiveness of combined estimation of groups and tracks over independent group detection.
7.5. Activity recognition performance
We compare the collective activity results with [2], [14],
[8] and [7] in the Table 3. To ensure a fair comparison with
[2] and [14], we divide the dataset into separate training and
testing sets as suggested by them. We use leave-one-videoout method to compare with [8] as suggested. To compare
with [7], we use four fold setup with the splits mentioned
by [7]. The Figure 5 compares the confusion table of the
proposed framework with that of [7]. To find the accuracy
for the group activity, we first identify the correctly detected
groups and estimate accuracy for group activity on these
groups. The confusion tables for group activity and atomic
action are also given in Figure 5.
7.3. Tracking performance
We assume that the detections per frame are available to
us. We do not handle occlusion in this paper. Whenever
any target returns back to the scene after occlusion, a new
id is assigned to it. To evaluate the tracking performance,
we consider the number of identity switches. We compare
the tracking results with a baseline model present in our
framework. It corresponds to the track association based
on visual, spatial and velocity compatibility (first part of
Eq. 4) only. The full model incorporates both track association and group association. The number of ID switches are
given in Table 1. The total number of tracks in the dataset is
466. The decrease in the number of ID switches in the full
Table 3. Comparison of overall and mean accuracies
Accuracy [2] / [14] / Ours
[8] / Ours
[7] / Ours
Overall
79.7 81.1 74.4 79.1 76.3
Mean
92.0 78.4 80.5 70.9 75.7 79.9 76.2
Class
Form the Table 3, we notice that the proposed method
offers a better accuracy than the methods [14] and [8], and
is marginally inferior to [7]. However, all these methods assume availability of either tracklets or action labels whereas
our method only needs the detections. Further, all these
7
Cross 61.30% 9.50% 2.80% 2.00% 24.50%
Cross 78.63% 5.34% 0.00% 0.00% 16.03%
Wait 2.40% 82.90% 4.40% 2.40% 7.80%
Wait 7.67% 76.99% 13.70% 1.37% 0.27%
Queue 4.60% 0.00% 95.40% 0.00% 0.00%
Queue 8.21% 7.73% 77.54% 5.56% 0.97%
Talk 0.00% 0.00% 0.00% 94.90% 5.10%
Talk 5.01% 4.33% 18.22% 72.21% 0.23%
Walk 29.00% 4.80% 1.20% 0.00% 65.10%
Walk 23.14% 1.33% 0.00% 0.00% 75.53%
Cross
Wait
Queue
Talk
Walk
Cross
(a)
Wait
Queue
Talk
Walk
Walk 85.21%
14.43%
0.31%
0.05%
Wait 14.06%
73.48%
9.27%
3.19%
Queue 8.08%
18.46%
66.54%
6.92%
Talk 9.14%
14.52%
11.29%
65.05%
Wait
Queue
Talk
Walk
(b)
(c)
Walk
90.12%
9.88%
Stand
14.72%
85.28%
Walk
Stand
(d)
Figure 5. (a) Confusion matrix for collective activity y from [7]. (b), (c), (d) Confusion matrices for collective activity, group activity and
atomic action respectively form the proposed method.
(a) Two crossing groups
(b) Two waiting groups
(c) A group in a queue
(d) Two talking groups
(e) Two walking groups
(f) Two crossing groups
(g) A waiting group
(h) A group in a queue
(i) A talking group
(j) Three walking groups
Figure 6. Qualitative results showing various collective and group activities. Collective activities column-wise: ’cross’, ’wait’, ’queue’,
’talk’, and ’walk’. A group is represented by a same color. Best viewed in color and when zoomed.
methods are non-causal in nature and involve batch processing of data unlike the proposed method. Additionally,
we provide results at all levels of granularity (individual,
group and collective). Figure 6 shows some qualitative results for group detection, group activity and collective activity. The members forming a group are represented by
the same color. For example, Figure 6(a) has two groups
which are correctly identified as crossing each other. Hence
the group activity for both the groups is walking while the
collective activity is crossing.
8. Conclusions
7.6. Computational Performance
References
In this paper, we have proposed a novel approach for
video understanding at various levels of granularity. We
have presented a linear programming based method for joint
estimation of tracks and groups. We have also proposed a
method to recognize activities at various levels - individual,
group and collective. The framework being causal in nature
and computationally efficient, it is amenable for real-time
implementation in video surveillance applications. The experiments show that the proposed method is very competitive with the state-of-the-art algorithms.
[1] C. C. Aggarwal. A human-computer interactive method for
projected clustering. IEEE Transactions on Knowledge and
Data Engineering, 16(4):448–460, 2004.
[2] M. R. Amer, P. Lei, and S. Todorovic. Hirf: Hierarchical
random field for collective activity recognition in videos. In
European Conference on Computer Vision, pages 572–585.
Springer, 2014.
[3] M. R. Amer, D. Xie, M. Zhao, S. Todorovic, and S.-C. Zhu.
Cost-sensitive top-down/bottom-up inference for multiscale
activity recognition. In European Conference on Computer
Vision, pages 187–200. Springer, 2012.
Towards our main aim of developing a real-time system capable of simultaneous tracking, group detection
and multi-level activity recognition, currently we achieve
around 3 fps with our unoptimized MATLAB code on a i7
machine with 3.50 GHz processor. With a proper implementation in GPU, we expect the frame rate to go up to 25
fps. To compare the computation time with one of the stateof-the-art algorithms, the method proposed in [2] takes 6
hours of training and 120 s per inference whereas our proposed method takes around 90 s and 0.3 s respectively.
8
[4] A. Andriyenko and K. Schindler. Globally optimal multitarget tracking on a hexagonal lattice. In European Conference on Computer Vision, pages 466–479. Springer, 2010.
[5] L. Bazzani, M. Zanotto, M. Cristani, and V. Murino. Joint
individual-group modeling for tracking. IEEE transactions
on pattern analysis and machine intelligence, 37(4):746–
759, 2015.
[6] J. Berclaz, F. Fleuret, E. Turetken, and P. Fua. Multiple
object tracking using k-shortest paths optimization. IEEE
transactions on pattern analysis and machine intelligence,
33(9):1806–1819, 2011.
[7] W. Choi and S. Savarese. A unified framework for multitarget tracking and collective activity recognition. In European Conference on Computer Vision, pages 215–230.
Springer, 2012.
[8] W. Choi, K. Shahid, and S. Savarese. Learning context for
collective activity recognition. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, pages
3273–3280. IEEE, 2011.
[9] H. Jiang, S. Fels, and J. J. Little. A linear programming
approach for multiple object tracking. In Computer Vision
and Pattern Recognition, 2007. CVPR’07. IEEE Conference
on, pages 1–8. IEEE, 2007.
[10] T. Joachims, T. Finley, and C.-N. J. Yu. Cutting-plane training of structural svms. Machine Learning, 77(1):27–59,
2009.
[11] S. M. Khan and M. Shah. Detecting group activities using
rigidity of formation. In Proceedings of the 13th annual
ACM international conference on Multimedia, pages 403–
406. ACM, 2005.
[12] T. Lan, L. Sigal, and G. Mori. Social roles in hierarchical
models for human activity recognition. In Computer Vision
and Pattern Recognition (CVPR), 2012 IEEE Conference on,
pages 1354–1361. IEEE, 2012.
[13] T. Lan, Y. Wang, W. Yang, and G. Mori. Beyond actions:
Discriminative models for contextual group activities. In
Advances in neural information processing systems, pages
1216–1224, 2010.
[14] T. Lan, Y. Wang, W. Yang, S. N. Robinovitch, and G. Mori.
Discriminative latent models for recognizing contextual
group activities. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 34(8):1549–1562, 2012.
[15] L. Leal-Taixé, G. Pons-Moll, and B. Rosenhahn. Everybody
needs somebody: Modeling social and grouping behavior on
a linear programming multiple people tracker. In Computer
Vision Workshops (ICCV Workshops), 2011 IEEE International Conference on, pages 120–127. IEEE, 2011.
[16] R. Li, R. Chellappa, and S. K. Zhou. Learning multi-modal
densities on discriminative temporal interaction manifold for
group activity recognition. In Computer Vision and Pattern
Recognition, 2009. CVPR 2009. IEEE Conference on, pages
2450–2457. IEEE, 2009.
[17] J. Liu, J. Luo, and M. Shah. Recognizing realistic actions
from videos in the wild. In Computer vision and pattern
recognition, 2009. CVPR 2009. IEEE conference on, pages
1996–2003. IEEE, 2009.
[18] J. C. Niebles, H. Wang, and L. Fei-Fei. Unsupervised learning of human action categories using spatial-temporal words.
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
9
International journal of computer vision, 79(3):299–318,
2008.
S. Pellegrini, A. Ess, and L. Van Gool. Improving data association by joint modeling of pedestrian trajectories and
groupings. In European Conference on Computer Vision,
pages 452–465. Springer, 2010.
H. Pirsiavash, D. Ramanan, and C. C. Fowlkes. Globallyoptimal greedy algorithms for tracking a variable number of objects. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, pages 1201–1208.
IEEE, 2011.
R. Poppe. A survey on vision-based human action recognition. Image and vision computing, 28(6):976–990, 2010.
Z. Qin and C. R. Shelton. Improving multi-target tracking
via social grouping. In Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, pages 1972–1978.
IEEE, 2012.
W. M. Rand. Objective criteria for the evaluation of clustering methods. Journal of the American Statistical association,
66(336):846–850, 1971.
M. Ryoo and J. Aggarwal. Stochastic representation and
recognition of high-level group activities. International journal of computer Vision, 93(2):183–200, 2011.
M. S. Ryoo and J. K. Aggarwal. Spatio-temporal relationship
match: Video structure comparison for recognition of complex human activities. In Computer vision, 2009 ieee 12th
international conference on, pages 1593–1600. IEEE, 2009.
S. Savarese, A. DelPozo, J. C. Niebles, and L. Fei-Fei.
Spatial-temporal correlatons for unsupervised action classification. In Motion and video Computing, 2008. WMVC 2008.
IEEE Workshop on, pages 1–8. IEEE, 2008.
K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR,
abs/1409.1556, 2014.
P. Sudowe, H. Spitzer, and B. Leibe. Person Attribute
Recognition with a Jointly-trained Holistic CNN Model. In
ICCV’15 ChaLearn Looking at People Workshop, 2015.
I. Tsochantaridis, T. Hofmann, T. Joachims, and Y. Altun. Support vector machine learning for interdependent and
structured output spaces. In Proceedings of the twenty-first
international conference on Machine learning, page 104.
ACM, 2004.
P. Turaga, R. Chellappa, V. S. Subrahmanian, and O. Udrea.
Machine recognition of human activities: A survey. IEEE
Transactions on Circuits and Systems for Video Technology,
18(11):1473–1488, 2008.
L. Wang, Y. Qiao, and X. Tang. Latent hierarchical model of
temporal structure for complex activity classification. IEEE
Transactions on Image Processing, 23(2):810–822, 2014.
D. Weinland, R. Ronfard, and E. Boyer. A survey of visionbased methods for action representation, segmentation and
recognition. Computer vision and image understanding,
115(2):224–241, 2011.
M. Wu and B. Schölkopf. A local learning approach for clustering. In Advances in neural information processing systems, pages 1529–1536, 2006.
[34] K. Yamaguchi, A. C. Berg, L. E. Ortiz, and T. L. Berg. Who
are you with and where are you going? In Computer Vision
and Pattern Recognition (CVPR), 2011 IEEE Conference on,
pages 1345–1352. IEEE, 2011.
[35] L. Zhang, Y. Li, and R. Nevatia. Global data association for
multi-object tracking using network flows. In Computer Vision and Pattern Recognition, 2008. CVPR 2008. IEEE Conference on, pages 1–8. IEEE, 2008.
10
| 7cs.IT
|
Warmstarting of Model-based Algorithm Configuration
Marius Lindauer and Frank Hutter
arXiv:1709.04636v3 [cs.AI] 28 Nov 2017
University of Freiburg
{lindauer,fh}@cs.uni-freiburg.de
Abstract
The performance of many hard combinatorial problem
solvers depends strongly on their parameter settings, and
since manual parameter tuning is both tedious and suboptimal the AI community has recently developed several algorithm configuration (AC) methods to automatically address
this problem. While all existing AC methods start the configuration process of an algorithm A from scratch for each new
type of benchmark instances, here we propose to exploit information about A’s performance on previous benchmarks in
order to warmstart its configuration on new types of benchmarks. We introduce two complementary ways in which we
can exploit this information to warmstart AC methods based
on a predictive model. Experiments for optimizing a flexible
modern SAT solver on twelve different instance sets show
that our methods often yield substantial speedups over existing AC methods (up to 165-fold) and can also find substantially better configurations given the same compute budget.
Introduction
Many algorithms in the field of artificial intelligence rely
crucially on good parameter settings to yield strong performance; prominent examples include solvers for many hard
combinatorial problems (e.g., the propositional satisfiability
problem SAT (Hutter et al. 2017) or AI planning (Fawcett
et al. 2011)) as well as a wide range of machine learning algorithms (in particular deep neural networks (Snoek,
Larochelle, and Adams 2012) and automated machine learning frameworks (Feurer et al. 2015)). To overcome the tedious and error-prone task of manual parameter tuning for a
given algorithm A, algorithm configuration (AC) procedures
automatically determine a parameter configuration of A with
low cost (e.g., runtime) on a given benchmark set. General
algorithm configuration procedures fall into two categories:
model-free approaches, such as ParamILS (Hutter et al.
2009), irace (López-Ibáñez et al. 2016) or GGA (Ansótegui,
Sellmann, and Tierney 2009), and model-based approaches,
such as SMAC (Hutter, Hoos, and Leyton-Brown 2011) or
GGA++ (Ansótegui et al. 2015).
Even though model-based approaches learn to predict the
cost of different configurations on the benchmark instances
at hand, so far all AC procedures start their configuration
Copyright c 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
process from scratch when presented with a new set of
benchmark instances. Compared with the way humans exploit information from past benchmark sets, this is obviously
suboptimal. Inspired by the human ability to learn across different tasks, we propose to use performance measurements
for an algorithm on previous benchmark sets in order to
warmstart its configuration on a new benchmark set. As we
will show in the experiments, our new warmstarting methods can substantially speed up AC procedures, by up to a
factor of 165. In our experiments, this amounts to spending
less than 20 minutes to obtain comparable performance as
could previously be obtained within two days.
Preliminaries
Algorithm configuration (AC). Formally, given a target
algorithm with configuration space Θ, a probability distribution D across problem instances, as well as a cost metric c
to be minimized, the algorithm configuration (AC) problem
is to determine a parameter configuration θ∗ ∈ Θ with low
expected cost on instances drawn from D:
θ∗ ∈ arg min Eπ ∼D [c(θ, π)].
(1)
θ∈Θ
In practice, π ∼ D is typically approximated by a finite
set of instances Π drawn from D. An example AC problem is to set a SAT solver’s parameters to minimize its average runtime on a given benchmark set of formal verification instances. Througout the paper, we refer to algorithms
for solving the AC problem as AC procedures. They execute
the target algorithm with different parameter configurations
θ ∈ Θ on different instances π ∈ Π and measure the resulting costs c(θ, π).
Empirical performance models (EPMs). A core ingredient in model-based approaches for AC is a probabilistic
regression model ĉ : Θ × Π → R that is trained based on
the cost values hx = [θ, π], y = c(θ, π)i observed thus far
and can be used to predict the cost of new parameter configurations θ ∈ Θ on new problem instances π ∈ Π. Since this
regression model predicts empirical algorithm performance
(i.e., its cost), it is known as an empirical performance model
(EPM; Leyton-Brown, Nudelman, and Shoham; Hutter et
al. 2009; 2014b). Random forests have been established as
Algorithm 1: Model-based Algorithm Configuration
Input : Configuration Space Θ, Instances Π,
Configuration Budget B
1
2
3
4
5
6
θinc , H ← initial design(Θ, Π);
while B not exhausted do
ĉ ← fit EPM based on H;
Θchall ← select challengers based on ĉ and H;
θinc , H ← race(Θchall ∪ {θinc }, Π, H);
return θinc
θ, π
Configurator
Π1
c(θ, π)
1
1
ΘΠ
inc H := hθ, π, c(θ, π)i
θ, π
Configurator
Π2
Model-based algorithm configuration. The core idea of
sequential model-based algorithm configuration is to iteratively fit an EPM based on the cost data observed so far
and use it to guide the search for well-performing parameter configurations. Algorithm 1 outlines the model-based
algorithm configuration framework, similarly as introduced
by Hutter, Hoos, and Leyton-Brown (2011) for the AC procedure SMAC, but also encompassing the GGA++ approach
by Ansótegui et al. (2015). We now discuss this algorithm
framework in detail since our warmstarting extensions will
adapt its various elements.
First, in Line 1 a model-based AC procedure runs the algorithm to be optimized with configurations in a so-called
initial design, keeping track of their costs and of the best
configuration θinc seen so far (the so-called incumbent). It
also keeps track of a runhistory H, which contains tuples
hθ, π, c(θ, π)i of the cost c(θ, π) obtained when evaluating configuration θ on instance π. To obtain good anytime
performance, by default SMAC only executes a single run
of a user-defined default configuration θdef on a randomlychosen instance as its initial design and uses θdef as its initial incumbent θinc . GGA++ samples a set of configurations
as initial generation and races them against each other on a
subset of the instances.
In Lines 2-5, the AC procedure performs the model-based
search. While a user-specified configuration budget B (e.g.,
number of algorithm runs or wall-clock time) is not exhausted, it fits a random-forest-based EPM on the existing
cost data in H (Line 3), aggregates the EPM’s predictions
over the instances Π in order to obtain marginal cost predictions ĉ(θ) for each configuration θ ∈ Θ and then uses these
predictions in order to select a set of promising configurations Θchall to challenge the incumbent θinc (Line 4) (SMAC)
Algorithm
c(θ, π)
1
ΘΠ
inc
the best-performing type of EPM and are thus used in all
current model-based AC approaches.
For the purposes of this regression model, the instances π
are characterized by instance features. These features reach
from simple ones (such as the number of clauses and variables of a SAT formula) to more complex ones (such as
statistics gathered by briefly running a probing algorithm).
Nowadays, informative instance features are available for
most hard combinatorial problems (e.g., SAT (Nudelman et
al. 2004), mixed integer programming (Hutter et al. 2014b),
AI planning (Fawcett et al. 2014), and answer set programming (Hoos, Lindauer, and Schaub 2014)).
Algorithm
∪
2
ΘΠ
inc
1
H ∪H
2
θ, π
Π
3
Configurator
Algorithm
c(θ, π)
Figure 1: Control flow of warmstarting information
or to generate well-performing offsprings (GGA++). For this
step, a so-called acquisition function trades off exploitation
of promising areas of the configuration space versus exploration of areas for which the model is still uncertain; common choices are expected improvement (Jones, Schonlau,
and Welch 1998), upper confidence bounds (Srinivas et al.
2010) or entropy search (Hennig and Schuler 2012).
To determine a new incumbent configuration θinc , in
Line 5 the AC procedure races these challengers and the current incumbent by evaluating them on individual instances
π ∈ Π and adding the observed data to H. Since these evaluations can be computationally costly the race only evaluates
as many instances as needed per configuration and terminates slow runs early (Hutter et al. 2009).
Warmstarting Approaches for AC
In this section, we discuss how the efficiency of model-based
AC procedures (as described in the previous section) can be
improved by warmstarting the search from data generated
in previous AC runs. We assume that the algorithm to be
optimized and its configuration space Θ is the same in all
runs, but the set of instances Π can change between the runs.
To warmstart a new AC run, we consider the following data
from previous AC runs on previous instance sets Πi :
i
• Sets of optimized configurations ΘΠ
inc found in previous AC runs on Πi —potentially, multiple runs were performed on the same instance set to return
the result with
i
contains
the final
best training performance such that ΘΠ
inc
incumbents from each of these runs;
• We denote
the union of previous instances as
S
Π0 := i∈I Πi for set superscripts i ∈ I.
S
• Runhistory data H0 := i∈I Hi of all AC runs on previous instance sets Πi . 1
1
If the set of instances Π and the runhistory H are not indexed,
we always refer to the ones of the current AC run.
To design warmstarting approaches, we consider the following desired properties:
1. When the performance data gathered on previous instance
sets is informative about performance on the current instance set, it should speed up our method.
2. When said performance data is misleading, our method
should stop using it and should not be much slower than
without it.
3. The runtime overhead generated by using the prior data
should be fairly small.
In the following subsections, we describe different warmstarting approaches that satisfy these properties.
Warmstarting Initial Design (INIT)
The first approach we consider for warmstarting our modelbased AC procedure is to adapt its initial design (Line 1
of Algorithm 1) to start from configurations that performed
well in the past. Specifically, we include the incumbent coni
figurations ΘΠ
inc from all previous AC runs as well as the
user-specified default θdef .
i
Evaluating all previous incumbents ΘΠ
inc in the initial design can be inefficient (contradicting Property 3), particularly if they are very similar. This can happen when the previous instance sets are quite similar, or when multiple runs
were performed on a single instance set.
To obtain a complementary set of configurations that covers all previously optimized instances well but is not redundant, we propose to use a two step approach. First, we determine the best configuration for each previous Πi .
[
X
Θinc :=
arg min
c(θ, π)
(2)
i
Π
i∈I θ∈Θinc
π∈Πi
Secondly, we use an iterative, greedy forward search
to select a complementary set of configurations across all
previous instance sets—inspired by the per-instance selection procedure Hydra (Xu, Hoos, and Leyton-Brown 2010).
Specifically, for the second step we define the mincost c̃(Θj )
of a set of configurations Θj on the union of all previous instances Π0 as
1 X
c̃(Θj ) := 0
min c(θ, π),
(3)
θ∈Θj
|Π |
0
π∈Π
start with Θ1 := {θdef }, and at each iteration, add the configuration θ0 ∈ Θinc to Θj that minimizes c̃(Θj ∪{θ0 }). Because
c̃(·) is a supermodular set function this greedy algorithm is
guaranteed to select a set of configurations whose mincost is
within a factor of (1 − 1/e) ≈ 0.63 of optimal among sets
of the same size (Krause and Golovin 2012).
Since we do not necessarily know the empirical cost of
all θ0 ∈ Θinc on all π ∈ Π0 , we use an EPM ĉ : Θ × Π → R
as a plug-in estimator to predict these costs. We train this
EPM on all previous runhistory data H0 . In order to enable
this, the benchmark sets for all previous AC runs have to be
characterized with the same set of instance features.
In SMAC, we use this set of complementary configurations in the initial design using the same racing function as
in comparing challengers to the incumbent (Line 5) to obtain the initial incumbent; to avoid rejecting challengers too
quickly, a challenger is compared on at least 3 instances before it can be rejected. In GGA++, these configurations can
be included in the first generation of configurations.
Data-Driven Model-Warmstarting (DMW)
Since model-based AC procedures are guided by their EPM,
we considered to warmstart this EPM by including all cost
data H0 gathered in previous AC runs as part of its training
data. In the beginning, the predictions of this EPM would
mostly rely on H0 , and as more data is acquired on the current benchmark this would increasingly affect the model.
However, this approach has two disadvantages:
1. When a lot of warmstarting data is available it requires
many evaluations on the current instance set to affect
model predictions. If the previous data is misleading, this
would violate our desired Property 2.
2. Fitting the EPM on H ∪ H0 will be expensive even in
early iterations, because H0 will typically contain many
observations. Even by using SMAC’s mechanism to invest at least the same amount of time in Lines 3 and 4 as
in Line 5, in preliminary experiments this slowed down
SMAC substantially (violating Property 3).
For these two reasons, we do not use this approach for warmstarting but propose an alternative. Specifically, to avoid the
computational overhead of refitting a very large EPM in each
iteration, and to allow our model to discard misleading previous data, we propose to fit individual EPMs ĉi for each
Hi once and to combine their predictions with those of an
EPM ĉ fitted on the newly gathered cost data H. This relates
to stacking in ensemble learning (Wolpert 1992); however
in our case, each constituent EPM is trained on a different
dataset. Hence, in principle we could even use different instance features for each instance set.
To aggregate predictions of the individual EPMs, we propose to use a linear combination:
X
ĉDMW (θ, π) := w0 + wĉ · ĉ(θ, π) +
wi · ĉi (θ, π) (4)
i∈I
where w are weights fitted with stochastic gradient descent (SGD) to minimize the combined model’s root mean
squared error (RMSE). To avoid overfitting of the weights,
we randomly split the current H into a training and validation set (2 : 1), use the training set to fit ĉ, and then compute
predictions of ĉ and each ĉi on the validation set, which are
used to fit the weights w. Finally, we re-fit the EPM ĉ on all
data in H to obtain a maximally informed model.
In the beginning of a new AC run, with few data in H,
ĉ will not be very accurate, causing its weight wĉ to be low,
such that the previous models ĉi will substantially influence
the cost predictions. As more data is gathered in H, the predictive accuracy of ĉ will improve and the predictions of the
previous models ĉi will become less important.
Besides weighting based on the accuracy of the individual models, the weights have the second purpose of scaling the individual model’s predictions appropriately: these
scales reflect the different hardnesses of the instance sets
they were trained on and by setting the weights to minimize
RMSE of the combined model on the current instances Π,
they will automatically normalize for scale.
The performance predictions of DMW can be used in any
model-based AC procedure, such as SMAC and GGA++.
Combining INIT and DMW (IDMW)
Importantly, the two methods we propose are complementary. A warmstarted initial design (INIT) can be easily
combined with data-driven model-warmstarting (DMW) because both approaches affect different parts of model-based
algorithm configuration: where to start from and how to integrate the full performance data from the current and the
previous benchmarks to decide where to sample next. In
fact, the two warmstarting methods can even synergize to
yield more than the sum of their pieces: by evaluating strong
configurations from previous AC runs in the initial design
through INIT, the weights of the stacked model in DMW can
be fitted on these important observations early on, improving
the accuracy of its predictions even in early iterations.
Experiments
We evaluated how our three warmstarting approaches improve the state-of-the-art AC procedure SMAC.2 In particular, we were interested in the following research questions:
Q1 Can warmstarted SMAC find better performing configurations within the same configuration budget?
Q2 Can warmstarted SMAC find well-performing configurations faster than default SMAC?
Q3 What is the effect of using warmstarting data Hi from
related and unrelated benchmarks?
Experimental Setup To answer these questions, we ran
SMAC (0.5.0) and our warmstarting variants3 on twelve
well-studied AC tasks from the configurable SAT solver
challenge (Hutter et al. 2017), which are publicly available
in the algorithm configuration library (Hutter et al. 2014a).
Since our warmstarting approaches have to generalize across
different instance sets and not across algorithms, we considered AC tasks of the highly flexible and robust SAT solver
SparrowToRiss across 12 instance sets. SparrowToRiss is a
combination of two well-performing solvers: Riss (Manthey
2014) is a tree-based solver that performs well on industrial and hand-crafted instances; Sparrow (Balint et al. 2011)
is a local-search solver that performs well on random, satisfiable instances. SparrowToRiss first runs Sparrow for a
parametrized amount of time and then runs Riss if Sparrow
could not find a satisfying assignment. Thus, SparrowToRiss
can be applied to a large variety of different SAT instances.
Riss, Sparrow and SparrowToRiss also won several medals
in the international SAT competition. Furthermore, configuring SparrowToRiss is a challenging task because it has a
2
The source code of GGA++ is not publicly available and thus,
we could not run experiments on GGA++.
3
Code and data is publicly available at: http://www.
ml4aad.org/smac/.
very large configuration space with 222 parameters and 176
conditional dependencies.
To study warmstarting on different categories of instances, the AC tasks consider SAT instances from applications with a lot of internal structure, hand-crafted instances
with some internal structure, and randomly-generated SAT
instances with little structure. We ran SparrowToRiss on
• application instances from bounded-model checking (BMC), hardware verification (IBM) and fuzz testing
based on circuits (CF);
• hand-crafted instances from graph-isomorphism (GI),
low autocorrelation binary sequence (LABS) and n-rooks
instances (N-Rooks);
• randomly generated instances, specifically, 3-SAT instances at the phase transition from the ToughSAT instance generator (3cnf ), a mix of satisfiable and unsatisfiable 3-SAT instances at the phase transition (K3), and
unsatisfiable 5-SAT instances from a generator used in
the SAT Challenge 2012 and SAT Competition 2013
(UNSAT-k5); and on
• randomly generated satisfiable instances, specifically, instances with 3 literals per clause and 1000 clauses
(3SAT1k), instances with 5 literals per clause and 500
clauses (5SAT500) and instances with 7 literals per clause
and 90 clauses (7SAT90).
Further details on these instances are given in the description of the configurable SAT solver challenge (Hutter et al.
2017). The instances were split into a training set for configuration and a test set to validate the performance of the
configured SparrowToRiss on unseen instances.
For each configuration run on a benchmark set in one of
the categories, our warmstarting methods had access to observations on the other two benchmark sets in the category.
For example, warmstarted SMAC optimizing SparrowToRiss
on IBM had access to the observations and final incumbents
of SparrowToRiss on CF and BMC.
As a cost metric, we chose the commonly-used penalized
average runtime metric (PAR10, i.e., counting each timeout
as 10 times the runtime cutoff) with a cutoff of 300 CPU
seconds. To avoid a constant inflation of the PAR10 values,
we removed all test instances post hoc that were never solved
by any configuration in our experiments (11 CF instances,
69 IBM instances, 17 BMC instances, 21 GI instances, 72
LABS instances and 73 3cnf instances).
On each AC task, we ran 10 independent SMAC runs with
a configuration budget of 2 days each. All runs were run on
a compute cluster with nodes equipped with two Intel Xeon
E5-2630v4 and 128GB memory running CentOS 7.
Baselines
As baselines, we ran (I) the user-specified default configuration θdef to show the effect of algorithm configuration, (II) SMAC without warmstarting, and (III) a stateof-the-art warmstarting approach for hyperparameter optimizers proposed by Wistuba, Schilling, and SchmidtThieme (2016), which we abbreviate as “adapted acquisition
function” (AAF). The goal of AAF is to bias the acquisition
θdef
SMAC
PAR10 scores
AAF
INIT
DMW
IDMW
Speedup over default SMAC
AAF INIT DMW IDMW
CF
IBM
BMC
326.5
150.6
421.5
125.8
50.6
209.6
140.0
49.0
230.7
126.6
47.8
203.1
122.0
47.5
155.9
116.4
48.8
137.4
0.1
3.9
1.2
0.5
16.2
1
0.7
1.4
11
2.7
9
29.3
GI
LABS
N-Rooks
314.1
330.1
116.7
165.0
232.9
8.6
165.6
291.2
18.1
165.6
271.2
27.3
165.4
285.9
27.5
152.7
286.7
12.7
25.6
0.8
0.4
0.6
0.8
0.4
7.1
0.8
0.4
19.4
0.8
0.5
3cnf
K3
UNSAT-k5
890.5
152.8
151.9
890.5
30.1
1.1
822.8
53.9
1.2
877.5
42.9
1.1
890.3
39.9
1.3
812.8
29.7
1.2
10.7
0.9
1
1
0.9
1
1
1.8
1
8.4
1.8
1
3SAT1k
5SAT500
7SAT90
104.4
3000
52.3
76.6
20.5
38.0
75.2
14.6
31.7
75.2
14.7
20.2
82.6
8.3
19.7
69.8
9.6
31.2
3.1
6
53.5
2.1
0.7
2.3
2.1
0.7
0.5
3.8
0.8
165.3
2.4
1.1
1.3
4.3
∅
Table 1: Left: PAR10 score (sec) of θdef , i.e., the default configuation of SparrowToRiss, and the final SparrowToRiss configurations returned by the different SMAC variants; median across 10 SMAC runs. The “SMAC” column shows the performance
of default SMAC without warmstarting. Best PAR10 is underlined and we highlighted runs in bold face for which there is no
statistical evidence according to a (one-sided) Mann-Whitney U test (α = 0.05) that they performed worse than the best configurator. Right: Speedup of warmstarted SMAC compared to default SMAC. This is computed by comparing the time points
of SMAC with and without warmstarting after which they do not perform significantly worse (according to a permutation test)
than SMAC with the full budget. Speedups > 1 indicate that warmstarted SMAC reached the final performance of default SMAC
faster, speedups < 1 indicate that default SMAC was faster. We marked the best speedup (> 1) in bold-face. The last row shows
the geometric average across all speedups.
function (Line 4 in Algorithm 1) towards previously wellperforming regions in the configuration space.4 To generalize AAF to algorithm configuration, we P
use marginalized
1
prediction across all instances ĉ(θ) := |Π|
π∈Π ĉ(θ, π).
Q1: Same configuration Budget
The left part of Table 1 shows the median PAR10 test
scores of the finally-returned configurations θinc across the
10 SMAC runs. Default SMAC nearly always improved
the PAR10 scores of SparrowToRiss substantially compared
to the SparrowToRiss default, yielding up to a 138-fold
speedup (on UNSAT-k5). Warmstarted SMAC performed significantly better yet on 4 of the AC tasks (BMC, 3cnf,
5SAT500 and 7SAT90), with additional speedups up to 2.1fold (on 5SAT500). On two of the crafted instance sets
(LABS and N-Rooks), the warmstarting approaches performed worse than default SMAC—details discussed later.
Overall, the best results were achieved by the combination of our approaches, IDMW. This yielded the best performance of all approaches in 6 of the 12 scenarios (with
sometimes substantial improvements over default SMAC)
and statistically insignificantly different results than the best
approach in 3 of the scenarios. Notably, IDMW performed
better on average than its individual components INIT and
DMW and clearly outperformed AAF.
4
We note that combining AAF and INIT is not effective because
evaluating the incumbents of INIT would nullify the acquisition
function bias of AAF.
Q2: Speedup
The right part of Table 1 shows how much faster our warmstarted SMAC reached the PAR10 performance default that
SMAC reached with the full configuration budget.5 The
warmstarting methods outperformed default SMAC in almost all cases (again except LABS and N-Rooks), with
up to 165-fold speedups. The most consistent speedups
were achieved by the combination of our warmstarting
approaches, IDMW, with a geometric-average 4.3-fold
speedup. We note that our baseline AAF also yielded good
speedups (geometric average of 2.4), but its final performance was often quite poor (see left part of Table 1).
Figure 2 illustrates the anytime test performance of all
SMAC variants.6 In Figure 2a, AAF, INIT and IDMW improved the performance of SparrowToRiss very early (after
roughly 700-1000 seconds), but only the DMW variants per5
A priori it is not clear how to define a speedup metric comparing algorithm configurators across several runs. To take noise into
account across our 10 runs, we performed a permutation test (with
α = 0.05 with 10 000 permutations) to determine the first time
point from which onwards there was no statistical evidence that default SMAC with a full budget would perform better. To take early
convergence/stagnation of default SMAC into account, we compute
the speedup of default SMAC to itself and divide the speedups by
default SMAC’s speedup.
6
Since Figure 2 shows test performance on unseen test instances, performance is not guaranteed to improve monotonically
(a new best configuration on the training instances might not generalize well to the test instances).
(a) BMC
(b) N-Rooks
(c) LABS
Figure 2: Median PAR10 of SparrowToRiss over configuration time with 25% and 75% percentiles as uncertainties.
(a) red: IBM, blue: CF,
green: BMC
(b) red: UNSAT-k5, blue: K3,
green: 3cnf
Figure 3: Weights over time of all 10 runs SMAC+DMW.
The red curve is the weight on EPM ĉ on the current instances; the blue and green curves corresponds to weights
on EPMs based on previously optimized instances.
formed well in the long run.
To study the effect of our worst results, Figure 2b and
2c show the anytime performance on N-Rooks and LABS,
respectively. Figure 2b shows that warmstarted SMAC performed better in the beginning, but that default SMAC performed slightly better in the end. The better initial performance is not captured in our quantitative analysis in Table 1.
In contrast, Figure 2c shows that for LABS, warmstarted
SMAC was initially mislead and then started improving like
default SMAC, but with a time lag; we note that we only observed this pattern on LABS and conclude that configurations
found on N-Rooks and GI do not generalize to LABS.
Q3: Warmstarting Influence
To study how our warmstarting methods learn from previous data, in Figure 3 we show how the weights of the DMW
approach changed over time. Figure 3a shows a representative plot: the weights were similar in the beginning (i.e.,
all EPMs contributed similarly to cost predictions) and over
time, the weights of the previous models decreased, with the
weight of the current EPM dominating. When optimizing
on IBM, the EPM trained on observations from CF was the
most important EPM in the beginning.
In contrast, Figure 3b shows a case in which the previous performance data acquired for benchmarks K3 and 3cnf
do not help for cost predictions on UNSAT-k5. (This was
to be expected, because 3cnf comprises only satisfiable instances, K3 a mix of satisfiable and unsatisfiable instances,
and UNSAT-k5 only unsatisfiable instances.) As the figure
shows, our DMW approach briefly used the data from the
mixed K3 benchmark (blue curves), but quickly focused
only on data from the current benchmark. These two examples illustrate that our DMW approach indeed successfully
used data from related benchmarks and quickly ignored data
from unrelated ones.
Related Work
The most related work comes from the field of hyperparameter optimization (HPO) of machine learning algorithms.
HPO, when cast as the optimization of (cross-)validation
error, is a special case of AC. This special case does not
require the concept of problem instances, does not require
the modelling of runtimes of randomized algorithms, does
not need to adaptively terminate slow algorithm runs and
handle the resulting censored algorithm runtimes, and typically deals with fairly low-dimensional and all-continuous
(hyper-)parameter configuration spaces. These works therefore do not directly transfer to the general AC problem.
Several warmstarting approaches exist for HPO. A
prominent approach is to learn surrogate models across
datasets (Swersky, Snoek, and Adams 2013; Bardenet et
al. 2014; Yogatama and Mann 2014). All of these works
are based on Gaussian process models whose computational
complexity scales cubically in the number of data points,
and therefore, all of them were limited to hundreds or at
most thousands of data points. We generalize them to the
AC setting (which, on top of the differences to HPO stated
above, also needs to handle up to a million cost measurements for an algorithm) in our DMW approach.
Another approach for warmstarting HPO is by adapting
the initial design. Feurer, Springenberg, and Hutter (2015)
proposed to initialize HPO in the automatic machine learning framework Auto-Sklearn with well-performing configurations from previous datasets. They had optimized configurations from 57 different machine learning data sets available as warmstarting data and chose which of these to use for
a new dataset based on its characteristics; specifically, they
used the optimized configurations from the k most similar
datasets. This approach could be adapted to AC warmstarting in cases where we have many AC benchmarks. However, one disadvantage of the approach is that – unlike our
INIT approach – it does not aim for complementarity in the
selected configurations. Wistuba, Schilling, and SchmidtThieme (2015) proposed another approach for warmstarting
the initial design which does not depend on instance features
and is not limited to configurations returned in previous optimization experiments. They combined surrogate predictions
from previous runs and used gradient descent to determine
promising configurations. This approach is limited to continuous (hyper-)parameters and thus does not apply to the
general AC setting.
One related variant of algorithm configuration is the problem of configuring on a stream of problem instances that
changes over time. The ReACT approach (Fitzgerald et al.
2014) targets this problem setting, keeping track of configurations that worked well on previous instances. If the characteristics of the instances change over time, it also adapts
the current configuration by combining observations on previous instances and on new instances. In contrast to our setting, ReACT does not return a single configuration for an
instance set and requires parallel compute resources to run a
parallel portfolio all the time.
Discussion & Conclusion
In this paper, we introduced several methods to warmstart
model-based algorithm configuration (AC) using observations from previous AC experiments on different benchmark
instance sets. As we showed in our experiments, warmstarting can speed up the configuration process up to 165-fold
and can also improve the configurations finally returned.
While we focused on the state-of-the-art configurator
SMAC in our experiments, our methods are also applicable to other model-based configurators, such as GGA++, and
our warmstarted initial design approach is even applicable to
model-free configurators, such as ParamILS and irace. We
expect that our results would similarly generalize to these.
A practical limitation of our DMW approach (and thus
also of IDMW) is that the memory consumption grows substantially with each additional EPM (at least when using
random forests fitted on hundreds of thousands of observations). We also tried to study warmstarting SMAC for optimizing SparrowToRiss on all instance sets except the one at
hand, but unfortunately, the memory consumption exceeded
12GB RAM. Therefore, one possible approach would be to
reduce memory consumption and to use instance features to
select a subset of EPMs constructed on similar instances.
Another direction for future work is to combine warmstarting with parameter importance analysis (Hutter, Hoos,
and Leyton-Brown 2014; Biedenkapp et al. 2017), e.g., for
determining important parameters on previous instance sets
and focusing the search on these parameters for a new instance set. Finally, a promising future direction is to integrate warmstarting into iterative configuration procedures,
such as Hydra (Xu, Hoos, and Leyton-Brown 2010), ParHydra (Lindauer et al. 2017), or Cedalion (Seipp et al. 2015),
which construct portfolios of complementary configurations
in an iterative fashion using multiple AC runs.
Acknowledgements
The authors acknowledge funding by the DFG (German Research Foundation) under Emmy Noether grant HU 1900/21 and support by the state of Baden-Württemberg through
bwHPC and the DFG through grant no INST 39/963-1
FUGG.
References
[Ansótegui et al. 2015] Ansótegui, C.; Malitsky, Y.; Sellmann, M.;
and Tierney, K. 2015. Model-based genetic algorithms for algorithm configuration. In Yang, Q., and Wooldridge, M., eds., Proceedings of the 25th International Joint Conference on Artificial
Intelligence (IJCAI’15), 733–739.
[Ansótegui, Sellmann, and Tierney 2009] Ansótegui, C.; Sellmann,
M.; and Tierney, K. 2009. A gender-based genetic algorithm for
the automatic configuration of algorithms. In Gent, I., ed., Proceedings of the Fifteenth International Conference on Principles
and Practice of Constraint Programming (CP’09), volume 5732 of
Lecture Notes in Computer Science, 142–157. Springer-Verlag.
[Balint et al. 2011] Balint, A.; Frohlich, A.; Tompkins, D.; and
Hoos, H. 2011. Sparrow2011. In Proceedings of SAT Competition 2011.
[Bardenet et al. 2014] Bardenet, R.; Brendel, M.; Kégl, B.; and Sebag, M. 2014. Collaborative hyperparameter tuning. In Dasgupta,
S., and McAllester, D., eds., Proceedings of the 30th International
Conference on Machine Learning (ICML’13), 199–207. Omnipress.
[Biedenkapp et al. 2017] Biedenkapp,
A.;
Lindauer,
M.;
Eggensperger, K.; Fawcett, C.; Hoos, H.; and Hutter, F. 2017.
Efficient parameter importance analysis via ablation with surrogates. In Proceedings of the Thirty-First Conference on Artificial
Intelligence (AAAI’17), 773–779.
[Bonet and Koenig 2015] Bonet, B., and Koenig, S., eds. 2015.
Proceedings of the Twenty-nineth Conference on Artificial Intelligence (AAAI’15). AAAI Press.
[Fawcett et al. 2011] Fawcett, C.; Helmert, M.; Hoos, H.; Karpas,
E.; Roger, G.; and Seipp, J. 2011. Fd-autotune: Domain-specific
configuration using fast-downward. In Helmert, M., and Edelkamp,
S., eds., Working notes of the Twenty-first International Conference
on Automated Planning and Scheduling (ICAPS-11), Workshop on
Planning and Learning.
[Fawcett et al. 2014] Fawcett, C.; Vallati, M.; Hutter, F.; Hoffmann,
J.; Hoos, H.; and Leyton-Brown, K. 2014. Improved features
for runtime prediction of domain-independent planners. In Chien,
S.; Minh, D.; Fern, A.; and Ruml, W., eds., Proceedings of the
Twenty-Fourth International Conference on Automated Planning
and Scheduling (ICAPS-14). AAAI.
[Feurer et al. 2015] Feurer, M.; Klein, A.; Eggensperger, K.; Springenberg, J. T.; Blum, M.; and Hutter, F. 2015. Efficient and robust
automated machine learning. In Cortes, C.; Lawrence, N.; Lee,
D.; Sugiyama, M.; and Garnett, R., eds., Proceedings of the 29th
International Conference on Advances in Neural Information Processing Systems (NIPS’15).
[Feurer, Springenberg, and Hutter 2015] Feurer, M.; Springenberg,
T.; and Hutter, F. 2015. Initializing Bayesian hyperparameter optimization via meta-learning. In Bonet and Koenig (2015), 1128–
1135.
[2014] Fitzgerald, T.; O’Sullivan, B.; Malitsky, Y.; and Tierney, K.
2014. React: Real-time algorithm configuration through tournaments. In Edelkamp, S., and Barták, R., eds., Proceedings of the
Seventh Annual Symposium on Combinatorial Search (SOCS’14).
AAAI Press.
[2012] Hennig, P., and Schuler, C. 2012. Entropy search for
information-efficient global optimization. Journal of Machine
Learning Research 98888(1):1809–1837.
[2014] Hoos, H.; Lindauer, M.; and Schaub, T. 2014. claspfolio
2: Advances in algorithm selection for answer set programming.
Theory and Practice of Logic Programming 14:569–585.
[2009] Hutter, F.; Hoos, H.; Leyton-Brown, K.; and Stützle, T.
2009. ParamILS: An automatic algorithm configuration framework. Journal of Artificial Intelligence Research 36:267–306.
[2014a] Hutter, F.; López-Ibánez, M.; Fawcett, C.; Lindauer, M.;
Hoos, H.; Leyton-Brown, K.; and Stützle, T. 2014a. Aclib: a
benchmark library for algorithm configuration. In Pardalos, P., and
Resende, M., eds., Proceedings of the Eighth International Conference on Learning and Intelligent Optimization (LION’14), Lecture
Notes in Computer Science, 36–40. Springer-Verlag.
[2014b] Hutter, F.; Xu, L.; Hoos, H.; and Leyton-Brown, K. 2014b.
Algorithm runtime prediction: Methods and evaluation. Artificial
Intelligence 206:79–111.
[2017] Hutter, F.; Lindauer, M.; Balint, A.; Bayless, S.; Hoos, H.;
and Leyton-Brown, K. 2017. The configurable SAT solver challenge (CSSC). Artificial Intelligence Journal (AIJ) 243:1–25.
[2011] Hutter, F.; Hoos, H.; and Leyton-Brown, K. 2011. Sequential model-based optimization for general algorithm configuration.
In Coello, C., ed., Proceedings of the Fifth International Conference on Learning and Intelligent Optimization (LION’11), volume
6683 of Lecture Notes in Computer Science, 507–523. SpringerVerlag.
[2014] Hutter, F.; Hoos, H.; and Leyton-Brown, K. 2014. An efficient approach for assessing hyperparameter importance. In Xing,
E., and Jebara, T., eds., Proceedings of the 31th International Conference on Machine Learning, (ICML’14), 754–762. Omnipress.
[1998] Jones, D.; Schonlau, M.; and Welch, W. 1998. Efficient
global optimization of expensive black box functions. Journal of
Global Optimization 13:455–492.
[2012] Krause, A., and Golovin, D. 2012. Submodular function
maximization. Tractability: Practical Approaches to Hard Problems 3(19):8.
[2009] Leyton-Brown, K.; Nudelman, E.; and Shoham, Y. 2009.
Empirical hardness models: Methodology and a case study on combinatorial auctions. Journal of the ACM 56(4).
[2017] Lindauer, M.; Hoos, H.; Leyton-Brown, K.; and Schaub, T.
2017. Automatic construction of parallel portfolios via algorithm
configuration. Artificial Intelligence 244:272–290.
[2016] López-Ibáñez, M.; Dubois-Lacoste, J.; Caceres, L. P.; Birattari, M.; and Stützle, T. 2016. The irace package: Iterated racing
for automatic algorithm configuration. Operations Research Perspectives 3:43–58.
[2014] Manthey, N. 2014. Riss 4.27. In Belov, A.; Diepold, D.;
Heule, M.; and Järvisalo, M., eds., Proceedings of SAT Competition
2014: Solver and Benchmark Descriptions, volume B-2014-2 of
Department of Computer Science Series of Publications B, 65–67.
University of Helsinki.
[2004] Nudelman, E.; Leyton-Brown, K.; Devkar, A.; Shoham, Y.;
and Hoos, H. 2004. Understanding random SAT: Beyond the
clauses-to-variables ratio. In Wallace, M., ed., Proceedings of the
10th International Conference on Principles and Practice of Constraint Programming (CP’04), volume 3258 of Lecture Notes in
Computer Science, 438–452. Springer-Verlag.
[2015] Seipp, J.; Sievers, S.; Helmert, M.; and Hutter, F. 2015. Automatic configuration of sequential planning portfolios. In Bonet
and Koenig (2015), 3364–3370.
[2012] Snoek, J.; Larochelle, H.; and Adams, R. P. 2012. Practical
Bayesian optimization of machine learning algorithms. In Bartlett,
P.; Pereira, F.; Burges, C.; Bottou, L.; and Weinberger, K., eds.,
Proceedings of the 26th International Conference on Advances in
Neural Information Processing Systems (NIPS’12), 2960–2968.
[2010] Srinivas, N.; Krause, A.; Kakade, S.; and Seeger, M. 2010.
Gaussian process optimization in the bandit setting: No regret and
experimental design. In Fürnkranz, J., and Joachims, T., eds., Proceedings of the 27th International Conference on Machine Learning (ICML’10), 1015–1022. Omnipress.
[2013] Swersky, K.; Snoek, J.; and Adams, R. 2013. Multi-task
Bayesian optimization. In Burges, C.; Bottou, L.; Welling, M.;
Ghahramani, Z.; and Weinberger, K., eds., Proceedings of the 27th
International Conference on Advances in Neural Information Processing Systems (NIPS’13), 2004–2012.
[2015] Wistuba, M.; Schilling, N.; and Schmidt-Thieme, L. 2015.
Learning hyperparameter optimization initializations. In Proceedings of the International Conference on Data Science and Advanced Analytics (DSAA), 1–10. IEEE.
[2016] Wistuba, M.; Schilling, N.; and Schmidt-Thieme, L. 2016.
Hyperparameter optimization machines. In Proceedings of the International Conference on Data Science and Advanced Analytics
(DSAA), 41–50. IEEE.
[1992] Wolpert, D. 1992. Stacked generalization. Neural Networks
5(2):241–259.
[2010] Xu, L.; Hoos, H.; and Leyton-Brown, K. 2010. Hydra: Automatically configuring algorithms for portfolio-based selection. In
Fox, M., and Poole, D., eds., Proceedings of the Twenty-fourth National Conference on Artificial Intelligence (AAAI’10), 210–216.
AAAI Press.
[2014] Yogatama, D., and Mann, G. 2014. Efficient transfer learning method for automatic hyperparameter tuning. In Kaski, S., and
Corander, J., eds., Proceedings of the Seventeenth International
Conference on Artificial Intelligence and Statistics (AISTATS), volume 33 of JMLR Workshop and Conference Proceedings, 1077–
1085.
| 2cs.AI
|
A Comparison of Hybridized and
Standard DG Methods for
Target-Based hp-Adaptive
Simulation of Compressible Flow
Michael Woopen, Aravind Balan, Georg May, and Jochen Schütz
Aachen Institute
for Advanced Study in
Computational Engineering Science
Financial support from the
Deutsche Forschungsgemeinschaft (German Research Foundation)
through grant GSC 111 is gratefully acknowledged.
A COMPARISON OF HYBRIDIZED AND STANDARD DG
METHODS FOR TARGET-BASED hp-ADAPTIVE SIMULATION
OF COMPRESSIBLE FLOW
MICHAEL WOOPEN, ARAVIND BALAN, GEORG MAY
Aachen Institute for Advanced Study in Computational Engineering Science,
RWTH Aachen University, Aachen, Germany
JOCHEN SCHÜTZ
Institut für Geometrie und Praktische Mathematik, RWTH Aachen University,
Aachen, Germany
Abstract. We present a comparison between hybridized and non-hybridized
discontinuous Galerkin methods in the context of target-based hp-adaptation
for compressible flow problems. The aim is to provide a critical assessment of
the computational efficiency of hybridized DG methods.
Hybridization of finite element discretizations has the main advantage, that
the resulting set of algebraic equations has globally coupled degrees of freedom only on the skeleton of the computational mesh. Consequently, solving for
these degrees of freedom involves the solution of a potentially much smaller system. This not only reduces storage requirements, but also allows for a faster
solution with iterative solvers. Using a discrete-adjoint approach, sensitivities with respect to output functionals are computed to drive the adaptation.
From the error distribution given by the adjoint-based error estimator, h- or
p-refinement is chosen based on the smoothness of the solution which can be
quantified by properly-chosen smoothness indicators.
Numerical results are shown for subsonic, transonic, and supersonic flow
around the NACA0012 airfoil. hp-adaptation proves to be superior to pure
h-adaptation if discontinuous or singular flow features are involved. In all
cases, a higher polynomial degree turns out to be beneficial. We show that for
polynomial degree of approximation p = 2 and higher, and for a broad range
of test cases, HDG performs better than DG in terms of runtime and memory
requirements.
1. Introduction
During the last years, discontinuous Galerkin (DG) methods (see, e.g., [1, 2, 3])
have become increasingly popular. This is indisputable due to their advantages
— high-order accuracy on unstructured meshes, a variational setting, and local
conservation, just to name a few.
However, the use of discontinuous function spaces is at the same time the reason
for a major disadvantage: unlike in continuous Galerkin (CG) methods, degrees
1
2
WOOPEN, BALAN, MAY, SCHÜTZ
of freedom are not shared between elements. As a consequence, the number of
unknowns is substantially higher compared to a CG discretization. Especially for
implicit time discretization this imposes large memory requirements, and potentially leads to increased time-to-solution.
In order to avoid these disadvantages, a technique called hybridization may be
utilized (see [4, 5, 6, 7, 8, 9, 10]), resulting in hybridized discontinuous Galerkin
(HDG) methods. Here, the globally coupled unknowns have support on the mesh
skeleton, i.e. the element interfaces, only. This reduces the size of the global system
and coincidentally improves the sparsity pattern.
However, aiming at industry applications, e.g. turbulent flow around a complete airplane or within an aircraft engine, hybridization alone does most likely
not provide a sufficiently successful overall algorithm. In these applications one
is usually interested in certain quantities only, for example lift or drag coefficients
in aerospace, instead of the solution quality per se. Thus, it might be beneficial
to distribute the degrees of freedom within the computational domain in such a
way that the solution to the discretized problem is close to optimal with respect
to the accuracy of these quantities. To achieve this goal, target-based error control
methods have been developed (see [11, 12, 13, 14, 15]). One such method is based
on the adjoint solution of the original governing equations with respect to the target functional. In this method, an additional linear system of partial differential
equations is solved which then gives an estimate on the spatial error distribution
contributing to the error in the target functional. This estimate can be used as
a criterion for local adaptation. Within the context of low order schemes, mesh
refinement is used for adaptation [12, 13]. Using DG (or HDG), however, offers the
additional possibility of varying the polynomial degree within each element. For
smooth solutions, this is more efficient compared to mesh refinement, as it yields
exponential convergence. In the context of wave problems, Giorgiani et al. [16]
showed the benefit of using p-adaptation within an HDG-framework. Combining
both mesh- and order-refinement results in so-called hp-adaptation.
In [17], we presented a discretization method for nonlinear convection-diffusion
equations. The method is based on a discontinuous Galerkin discretization for
convection terms, and a mixed method using H(div) spaces for the diffusive terms.
Furthermore, hybridization is used to reduce the number of globally coupled degrees
of freedom. Adjoint consistency was shown in [18]. In [19, 20], we extended our
computational framework to include HDG schemes, as well as adjoint-based hand hp-adaptation. In the current paper, we compare our HDG method with a
standard DG method in the context of hp-adaptation for stationary compressible
flow, mainly with the aim to assess the efficiency of both methods.
This paper is structured as follows. We briefly cover the governing equations,
namely the compressible Euler and Navier-Stokes equations, in Sec. 2. After that
we introduce our discretization and describe the concept of hybridization in Sec. 3.
In Sec. 4 we establish the adjoint formulation and show how hybridization can be
applied to the dual problem. Then we show its efficiency and robustness with
examples from compressible flow, including the subsonic, transonic, and supersonic
regime, in Sec. 5. Finally, we offer conclusions and outlook on future work in Sec. 6.
A COMPARISON OF HDG AND DG FOR TARGET-BASED hp-ADAPTATION
3
2. Governing Equations
We consider systems of partial differential equations
(1)
∇ · (fc (w) − fv (w, ∇w)) = s (w, ∇w)
with convective and diffusive fluxes
(2)
fc : Rm → Rm×d
and fv : Rm × Rm×d → Rm×d ,
respectively, and a state-dependent source term
(3)
s : Rm × Rm×d → Rm
on domain Ω ⊂ Rd . Potentially, some of these quantities could be zero. We
denote the spatial dimension by d and the number of conservative variables by m.
Boundary conditions can be applied either to the conservative variables w ∈ Rm
and their gradient ∇w ∈ Rm×d or directly to the fluxes fc and fv .
2.1. Two-Dimensional Euler Equations. The Euler equations are comprised
of the inviscid compressible continuity, momentum and energy equations. They are
given in conservative form as
(4)
∇ · fc (w) = 0
with the vector of conserved variables
(5)
w = (ρ, ρv, E)T
where ρ is the density, v is the velocity vector v := (vx , vy )T , and E the total
energy. The convective flux is given by
(6)
T
fc = (ρv, p Id + v ⊗ v, v(E + p)) .
Pressure is related to the conservative flow variables w by the equation of state
1
(7)
p = (γ − 1) E − ρv · v
2
where γ = cp /cv is the ratio of specific heats, generally taken as 1.4 for air.
Along wall boundaries we apply the slip boundary condition
(8)
vn (w) := v · n = 0.
We also define a boundary function which satisfies vn (w∂Ω (w)) = 0 as
1
0T
0
(9)
w∂Ω (w) = 0 Id − n ⊗ n 0 w.
0
0T
1
Prescribing boundary conditions at the far-field can be realized with the aid of
characteristic upwinding [21]. Here, the normal convective flux Jacobian is decomposed as
(10)
fc0 (w) · n = Q(w, n) · Λ(w, n) · Q−1 (w, n)
with Λ(w, n) being a diagonal matrix containing the eigenvalues of fc0 (w) · n. The
corresponding right eigenvectors can be found in the columns of Q(w, n). The interior and far-field states in characteristic variables are then given by wc = Q(w, n)w
4
WOOPEN, BALAN, MAY, SCHÜTZ
and wc,∞ = Q(w, n)w∞ , respectively. Finally, depending on the sign of (Λ(w, n))i,i ,
we can construct a boundary state
(
(Q(w, n)wc )i ,
(Λ(w, n))i,i ≥ 0
(11)
(w∂Ω (w))i =
(Q(w, n)wc,∞ )i , (Λ(w, n))i,i < 0.
2.2. Two-Dimensional Navier-Stokes Equations. The Navier-Stokes equations
in conservative form are given by
∇ · (fc (w) − fv (w, ∇w)) = 0.
(12)
The convective part fc of the Navier-Stokes equations coincides with the Euler
equations. The viscous flux is given by
T
(13)
fv = (0, τ , τ v + k∇T ) .
The temperature is defined via the ideal gas law
µγ
E
1
1
p
(14)
T =
− v·v =
k · Pr ρ
2
(γ − 1)cv ρ
µc
where Pr = kp is the Prandtl number, which for air at moderate conditions can be
taken as a constant with a value of Pr ≈ 0.72. k denotes the thermal conductivity
coefficient. For a Newtonian fluid, the stress tensor is defined as
2
T
(15)
τ = µ ∇v + (∇v) − (∇ · v) Id .
3
The variation of the molecular viscosity µ as a function of temperature is determined by Sutherland’s law as
(16)
µ=
C1 T 3/2
T + C2
√
with C1 = 1.458 × 10−6 mskg
and C2 = 110.4 K.
K
Along wall boundaries, we apply the no-slip boundary condition, i.e.
(17)
v=0
with corresponding boundary function
(18)
T
w∂Ω (w) = (ρ, 0, E) .
Furthermore, one has to give boundary conditions for the temperature. In the
present work we use the adiabatic wall condition, i.e.
(19)
∇T · n = 0.
Combining both no-slip and adiabatic wall boundary conditions, yields a condition
for the viscous flux, namely
T
(20)
fv (w∂Ω , q∂Ω ) = 0 τ 0 .
A COMPARISON OF HDG AND DG FOR TARGET-BASED hp-ADAPTATION
5
3. Discretization
3.1. Notation. We tesselate the domain
S Ω into a collection of non-overlapping
elements, denoted by Th , such that K∈Th K = Ω. For the element edges we
consider two different kinds of sets, ∂Th and Γh , which are element-oriented and
edge-oriented, respectively.
(21)
(22)
∂Th := { ∂K\∂Ω : K ∈ Th },
Γh := { e : e = K ∩ K 0 for K, K 0 ∈ Th ; measd−1 (e) 6= 0 }.
The first is the collection of all element boundaries, which means that every edge
appears twice. The latter, however, includes every edge just once. The reason for
this distinction will become clear later. Please note that neither of these sets shall
include edges lying on the domain boundary; the set of boundary edges is denoted
by Γbh .
We denote by Πp (D) the set of polynomials of degree at most p on some domain D. We will need discontinuous function spaces for the domain and the mesh
skeleton:
(23)
(24)
(25)
Vh = {v ∈ L2 (Ω) : v|K ∈ ΠpK (K),
2
Wh = {w ∈ L (Ω) : w|K ∈ Π
pK
(K),
Mh = {µ ∈ L2 (Γh ) : µ|e ∈ Πpe (e),
K ∈ Th }m×d
K ∈ Th }m
e ∈ Γh }m .
Thus, v ∈ Vh , w ∈ Wh and µ ∈ Mh are piecewise polynomials of degree p which
can be discontinuous across edges (for v, w) or vertices (for µ), respectively.
Usually, the polynomial degree between elements and interfaces does not vary.
In the case of varying polynomial degrees pK − and pK + , we follow Cockburn
et al. [6] and choose the polynomial degree for the interface e = K − ∩ K + as
pe = max {pK − , pK + }. This approach is commonly applied for p-adaptive HDG
discretizations (see Chen and Cockburn [22], Cockburn and Gopalakrishnan [23]
and Giorgiani et al. [16]).
We will distinguish between element-oriented inner products (defined with Th ),
i.e.
X Z
(v, w)Th :=
(26)
vw dx,
K∈Th
(27)
(v, w)Th :=
K∈Th
(28)
hv, wi∂Th :=
K
X Z
v · w dx,
K
X Z
K∈Th
vw dσ,
∂K
and edge-oriented inner products (defined with Γh ), i.e.
XZ
(29)
hv, wiΓh :=
vw dσ.
e∈Γh
e
3.2. Weak Formulation. We can rewrite general convection-diffusion equations
as a first-order system by introducing an additional unknown representing the gradient of the solution
q = ∇w
(30)
∇ · (fc (w) − fv (w, q)) = s (w, q) .
6
WOOPEN, BALAN, MAY, SCHÜTZ
By multiplying the strong, mixed form (30) with appropriate test functions
(τ h , ϕh ) ∈ Vh × Wh and integrating by parts, we obtain a standard DG discretization in mixed formulation of the problem, i.e.:
Find (qh , wh ) ∈ Vh × Wh s.t. ∀(τ h , ϕh ) ∈ Vh × Wh
(31)
(32)
(33)
(34)
0 = NhDG (qh , wh ; τ h , ϕh )
:= (τ h , qh )Th + (∇ · τ h , wh )Th − hτ h · n, wi
b ∂Th
− (∇ϕh , fc (wh ) − fv (wh , qh ))Th
D
E
+ ϕh , fbc − fbv
− (ϕh , s(wh , qh ))Th
∂Th
(35)
+
DG
Nh,∂Ω
DG
(qh , wh ; τ h , ϕh ) + Nh,sc
(qh , wh ; ϕh ) .
Here the numerical trace w
b and the numerical fluxes fbc , fbv have to be chosen
appropriately to define a stable and consistent method. Furthermore, boundary
conditions and shock-capturing are incorporated using appropriate operators, here
DG
DG
denoted by Nh,∂Ω
(qh , wh ; τ h , ϕh ) and Nh,sc
(qh , wh ; ϕh ).
In contrast to a DG discretization, where the numerical trace w
b is defined explicitly in terms of wh and qh , it is treated as an additional unknown in an HDG
method. This additional unknown is called λh and has support on the skeleton of
the mesh only. In order to close the system the continuity of the numerical fluxes
across edges is required in a weak sense, resulting in an additional equation.
The weak formulation of the hybrid system, comprised of equations for the gradient qh , the solution wh itself and its trace on the mesh skeleton λh , is then given
by:
Find (qh , wh , λh ) ∈ Xh := Vh × Wh × Mh s.t. ∀(τ h , ϕh , µh ) ∈ Xh
(36)
(37)
(38)
(39)
0 = Nh (qh , wh , λh ; τ h , ϕh , µh )
:= (τ h , qh )Th + (∇ · τ h , wh )Th − hτ h · n, λh i∂Th
− (∇ϕh , fc (wh ) − fv (wh , qh ))Th
D
E
+ ϕh , fbc − fbv
− (ϕh , s(wh , qh ))Th
∂Th
(40)
(41)
+ Nh,∂Ω (qh , wh ; τ h , ϕh ) + Nh,sc (qh , wh ; ϕh )
D
r
zE
+ µh , fbc − fbv
.
Γh
Please note the use of ∂Th in the weak formulation of the mixed form (see
Eq. (37)–(40)) and Γh in Eq. (41) defining λh . This perfectly resembles the character of these equations, being element- and edge-oriented, respectively. The terms
tested against τ h and ϕh are called local solvers, meaning they do not depend
on the solution within neighboring elements but only on the trace of the solution
which is approximated by λh . The coupling between elements is then introduced
by weakly enforcing the normal continuity of the numerical fluxes across interfaces
(see Eq. (41)).
We choose numerical fluxes similar to the Lax-Friedrich flux and to the LDG
flux for the convective and diffusive flux, respectively, i.e.
(42)
fbc (λh , wh ) = fc (λh ) · n − αc (λh − wh )
(43)
fbv (λh , wh , qh ) = fv (λh , qh ) · n + αv (λh − wh )
A COMPARISON OF HDG AND DG FOR TARGET-BASED hp-ADAPTATION
7
which can be combined into
(44)
fbc − fbv = (fc (λh ) − fv (λh , qh )) · n − α (λh − wh )
where α = αc + αv . The stabilization introduced can be given by a tensor; in
our work, however, we restrict ourselves to a constant scalar α which seems to be
sufficient for a wide range of test cases.
3.2.1. Boundary Conditions. In order to retrieve an adjoint-consistent scheme, special care has to be taken when discretizing the boundary conditions (see Schütz and
May [18]). The boundary conditions have to be incorporated by using the boundary
states w∂Ω (wh ) and gradients q∂Ω (wh , qh ), i.e.
(45)
Nh,∂Ω (qh , wh ; τ h , ϕh )
:= hτ h · n, w∂Ω iΓb + hϕh , (fc (w∂Ω ) − fv (w∂Ω , q∂Ω )) · niΓb .
h
h
We would like to emphasize that λh does not occur in this boundary term.
3.2.2. Shock-Capturing. In non-smooth parts of the solution, for example shocks in
compressible flows, a stabilization term has to be introduced. We adopt the shockcapturing approach by Hartmann and Houston [11] where an artificial viscosity
term, given by ∇ · ( (w, ∇w) ∇w), is used. The viscosity is given by the L1 norm of the strong residual ∇ · fc (w) in every element. In order to accelerate the
convergence of this term to zero with mesh refinement, it is premultiplied with an
K
. The latter resembles the actual resolution within an
effective mesh size h̃K := hpK
element. Furthermore, a user-defined factor 0 is introduced which can be reliably
tuned for a rather large range of test cases. Finally, the artificial viscosity is given
by
Z
0 h̃2−β
K
(46)
|K :=
d(w) dx
|K|
K
where the strong residual is given by
(47)
d(w) :=
m
X
|(∇ · fc (w))i | .
i=1
In the discretization of this shock-capturing term the interface integral is neglected so that only the volume contribution is considered, i.e.
(48)
Nh,sc (wh ; ϕh ) := (∇ϕh , (wh , ∇wh ) ∇wh )Th .
This obviates the need for introducing qh in purely convective problems (e.g. the
compressible Euler equations). In the viscous case, where the gradient is explicitly
given, ∇wh can be replaced by qh yielding
(49)
Nh,sc (qh , wh ; ϕh ) := (∇ϕh , (wh , qh ) qh )Th .
Please note, that this term enters only the local part of the discretization.
Furthermore, it is noteworthy that this shock-capturing term is not only consistent and conservative but also yields an asymptotically adjoint-consistent scheme.
8
WOOPEN, BALAN, MAY, SCHÜTZ
3.3. Relaxation. In order to solve the nonlinear system of equations that defines
the HDG method, the Newton-Raphson
method is applied. Beginning with an
initial guess x0h := q0h , wh0 , λ0h , one iteratively solves the linear system
Nh0 [xnh ] (δ xnh ; yh ) = −Nh (xnh ; yh )
(50)
∀yh ∈ Xh
and updates the solution as
xn+1
= xnh + δ xnh
h
(51)
until the residual Nh (xnh ; yh ) has reached a certain threshold. Please note, that
we have grouped qh , wh and λh , and the test functions into xh := (qh , wh , λh )
and yh := (τ h , ϕh , µh ), respectively. Nh0 denotes the Fréchet derivative of Nh with
respect to xh .
This routine can, however, lead to stability problems if the starting value x0h is
too far away from the solution xh . Therefore, an artificial time is introduced and a
backward Euler method is applied, which yields a slight modification of the linear
system given in Eq. (50), namely ∀yh ∈ Xh
(52)
ϕh ,
1
δwhn
∆tn
n
n
+ Nh0 [xn
h ] (δ xh ; yh ) = −Nh (xh ; yh ) .
Th
Please note that by choosing ∆tn → ∞, a pure Newton-Raphson method is obtained. Usually the time step is kept finite for a few initial steps to ensure stability.
As soon as the residual is lower than a certain threshold, i.e. the current approximation xnh is thought to be sufficiently close to the solution xh , we let the time
step go towards infinity.
As we are interested in steady-state problems, we can use local time-stepping in
order to accelerate the computation. For each element, we apply a time step based
on a global CFL number, the element volume, and an approximation of the flux
Jacobian’s spectral radius, i.e.
(53)
∆tnK = CFLn
|K|
.
λc + 4λv
Here, λc and λv represent approximations to the maximum eigenvalue of the convective and diffusive flux, respectively (see Mavriplis and Jameson [24]).
The next question is, how to choose CFLn in each iteration such that for xnh →
xh , CFLn → ∞. Therefore, we choose a sequence (CFLn )n to consist of a ramping
phase followed by a very fast increase, i.e.
(54)
3
2
− 2 nn
c0 3 nn0
0
n−1
CFLn :=
kNh
k2
n−1
1 + c1 max 0, log
CFL
kN n k
h
2
n ≤ n0
otherwise
Strictly speaking, the parameters c0 , c1 and n0 depend on the actual problem;
experiments, however, showed that they can be reliably tuned for a wide range of
problems.
3.4. Hybridization. Using an appropriate polynomial expansion for δqh , δwh and
δλh , the linearized global system Eq. (50) is given in matrix form as
A B R
δQ
F
C D S δW = G
(55)
L M N
δΛ
H
A COMPARISON OF HDG AND DG FOR TARGET-BASED hp-ADAPTATION
9
T
where the vector [δQ, δW, δΛ] contains the expansion coefficients of δxh with respect to the chosen basis.
In order to carry on with the derivation of the hybridized method, we want to
formulate that system in terms of δΛ only. Therefore we split it into
A B
δQ
F
R
(56)
=
−
δΛ
C D
δW
G
S
and
(57)
L
M
δQ
δW
+ N δΛ = H.
Substituting Eq. (56) into Eq. (57) yields the hybridized system
(58)
K δΛ = E.
where the system matrix and right-hand side vector are given by
!
A B −1 R
(59)
K= N− L M
C D
S
−1
A B
F
E=H− L M
(60)
.
C D
G
The workflow is as follows: First, the hybridized system is assembled and then
being solved for δΛ. Subsequently, δQ and δW can be reconstructed inside the
elements via Eq. (56). It is very important to note that it is not necessary to solve
the large system given by Eq. (56). In fact, the matrix in Eq. (56) can be reordered
to be block diagonal. Each of these blocks is associated to one element. Thus, both
the assembly of the hybridized matrix in Eq. (58) and the reconstruction of δQ and
δW can be done in an element-wise fashion. In order to save computational time,
the solutions to Eq. (56) can be saved after the assembly of the hybridized system
and reused during the reconstruction of δQ and δW .
The hybridized matrix is an nf × nf block matrix, where nf = |Γh | is the
number of interior edges. In each block row there is one block on the diagonal
and 2d off-diagonal blocks in the case of simplex elements. These blocks represent
the edges of the
neighboring elements of one edge. Each block is dense and has
O m2 · p2(d−1) entries. Please recall that p is the polynomial degree of the ansatz
functions, d is the spatial dimension of the domain Ω and m is the number of
partial differential equations (m = 4 for the 2-dimensional Euler or Navier-Stokes
equations). This structure is very similar to that
of a normal DG discretization,
whereas the blocks in the latter have O m2 · p2d entries and thus are considerably
bigger for higher polynomial order p. The size of the system matrix does not only
play a big role in terms of memory consumption but also for the iterative solver.
Here, a major portion of the overall workload goes into matrix-vector products
which can obviously be performed faster, if the problem dimensions are smaller. In
our code we use an ILU(n)-preconditioned GMRES which is available through the
PETSc library [25, 26].
4. Adaptation Procedure
In the context of adjoint-based (also referred to as target- or output-based) error
estimation, one is interested in quantifying the error of a specific target functional
10
WOOPEN, BALAN, MAY, SCHÜTZ
Jh : Xh → R, i.e.
eh := Jh (x) − Jh (xh ) ,
(61)
where xh is the approximation to x in Xh . This target functional can for example
represent lift or drag coefficients in aerospace applications. In general, the target
functional is an integrated value, where integration can be both on a volume or
along the boundary. For the derivation of the adjoint-based error estimate we
expand the target functional in a Taylor series as follows
(62)
Jh (x) − Jh (xh ) = Jh0 [xh ] δ xh + O kδ xh k2
where we defined δ xh := x − xh .
We proceed in a similar manner with the error in the residual, i.e.
Nh (x; yh ) − Nh (xh ; yh ) = Nh0 [xh ] (δ xh ; yh ) + O kδ xh k2 .
(63)
As our discretization is consistent the first term Nh (x; yh ) vanishes.
Substituting Eq. (63) into Eq. (62) and neglecting the quadratic terms yields
eh ≈ η := −Nh (xh ; zh )
(64)
where
zh
is defined by the so-called adjoint equation
e h.
Nh0 [xh ] (yh ; zh ) = Jh0 [xh ] (yh )
∀yh ∈ X
eh ∈ X
e h represents the link between variaThe adjoint solution zh = qeh , w
eh , λ
tions in the residual and in the target functional.
The global error estimate η can then be restricted to a single element to yield a
local indicator to drive an adaptation procedure, i.e.
(65)
(66)
eh , w
eh , 0)
ηK := Nh (qh , wh , λh ; q
K
.
Here, we want to emphasize that, in contrast to the global error estimate, we ignore
eh . We found that by taking the
the contribution of the hybrid adjoint variable λ
whole adjoint into account, jumps across element interfaces were overly penalized.
This deserves a more in-depth analysis.
Please note, that the functionals Nh and Jh and their jacobians have to be
e h ⊃ Xh . Otherwise the
evaluated in a somewhat richer space than Xh , namely X
weighted residual Nh (xh ; zh ) would be identical zero as
(67)
Nh (xh ; yh ) = 0
∀yh ∈ Xh .
This can be achieved by either mesh refinement or a higher polynomial degree of
the ansatz functions. In our setting, especially when using a hierarchical basis, the
latter is advantageous with respect to implementational effort and efficiency.
4.1. Hybridization. In matrix form, the
follows
T
A B R
C D S
(68)
L M N
adjoint system (see Eq. (65)) reads as
e
Q
f
W
=
e
Λ
Fe
e
G
e
H
e = 0 as λh is not defined on the boundary
Please note, that in our formulation H
and thus the target functional depends only on wh and qh .
A COMPARISON OF HDG AND DG FOR TARGET-BASED hp-ADAPTATION
11
As the overall structure of the adjoint equation is similar to the primal system
(see Eq. (58)), one can also apply static condensation to the adjoint system which
then yields its hybridized form:
e=E
e
KT Λ
(69)
where the right-hand side is given by
(70)
e=−
E
RT
ST
A
C
B
D
−T "
Fe
e
G
#
.
It is interesting to note that the hybridized adjoint system matrix is also the
transpose of the hybridized primal system matrix (for a higher polynomial order).
This is very beneficial for the implementation as the routines for the assembly of
this matrix are already available.
The adjoint solution within each element can then be computed with the aid of
the adjoint local system, given by
#
# "
T "
e
T
Q
A B
Fe
e
− L M
Λ
(71)
=
e
f
C D
G
W
where the matrix is also the transpose of the primal local matrix (see Eq. (56)).
4.2. Marking Elements for Refinement. After having obtained a localized error estimate, we have to choose a set of elements to be refined. This can be done
in many different ways. We adopt a strategy proposed by Dörfler [27]. The aim
of this marking strategy is to find the smallest set M ⊆ Th such that the error
contributed by this set represents a certain fraction of the total error, i.e.
(72)
ηM ≥ (1 − θ) ηTh .
The user-defined parameter θ is of course problem dependent. It can, however, be
tuned for a large rangePof test cases. Please note, that we define the error of any
2
2
subset of Th as ηM
:= K∈M ηK
.
4.3. Choosing between h- and p-Adaptation. The final step in the adaptation
procedure is the decision between mesh refinement and order enrichment. There
exist several ways to make this decision. Ceze and Fidkowski [28] solve local adjoint
problems for both options and then decide which one is more efficient with respect
to degrees of freedom or the non-zero entries in the system matrix. We, however,
adopt the strategy by Wang and Mavriplis [29]. They used a smoothness sensor
originally devised by Persson and Peraire [30] for an artificial viscosity approach.
This sensor exploits the fact that the decay of the expansion coefficients of the
solution is closely linked to its smoothness. For smoother solutions, decay is faster.
This can be exploited to check the regularity of the solution. On each element, the
smoothness sensor is defined as
(w − w? , w − w? )K
(73)
SK :=
(w, w)K
where w? is the element-wise projection of w to the next smaller polynomial space,
given by
(74)
(ϕh , w? )K = (ϕh , w)K
∀ϕh ∈ ΠpK −1 (K).
12
WOOPEN, BALAN, MAY, SCHÜTZ
(a) Coarse mesh
(b) Adapted mesh
Figure 1. Smoothness sensor for a transonic test case
Hence, w − w? represents the higher order components of the solution (see Fig. 1).
As we use a hierarchical basis [31], this projection is very cheap. By introducing a
threshold S , a decision between mesh-refinement and p-enrichment can be made,
i.e.
(
(75)
SK
< S
≥ S
p-enrichment
mesh-refinement
First promising results for this approach are given in [20].
A COMPARISON OF HDG AND DG FOR TARGET-BASED hp-ADAPTATION
13
5. Results
In the following we compare our in-house HDG and DG solvers in terms of degrees of freedom and runtime. The DG discretization is based on the Lax-Friedrich
and the BR2 [32] fluxes for convective and viscous terms, respectively. Boundary
conditions and target functionals are evaluated in an adjoint-consistent manner
[33]. Both solvers share the same computational framework, so we believe that our
comparison is meaningful.
We apply both solvers to compressible flow problems, including inviscid subsonic,
transonic and supersonic, as well as subsonic laminar flow. In all cases we show
results for pure mesh-adaptation (p = 1 . . . 4) and hp-adaptation (p = 2 . . . 5).
Please note, that we choose the same relaxation parameters for all test cases,
namely, c0 = 106 , c1 = 103 , n0 = 4 (see Eq. (54)). If artificial viscosity is necessary,
we use 0 = 0.2 and β = 0. The parameters for the adaptation procedure are chosen
as θ = 0.05 and S = 10−6 . This set of parameters seems to be robust for both
DG and HDG for a broad range of test cases. In order to approximate the error
in drag, reference solutions on hp-adapted meshes with more than 2 · 106 degrees
of freedom are used. (Please note, that we refer to ndof w whenever we speak of
degrees of freedom in the following as this is a good measure for the resolution.)
Before we turn our attention to the adaptive computations, we compare runtimes for both methods on a fixed mesh for several polynomial orders. This way,
we can learn which improvement can be expected. In Fig. 2 timings for the assembly
procedure and the iterative solver are given. We show Euler and a Navier-Stokes
computation on a mesh with 2396 elements and 3544 interior faces. We used polynomial orders from p = 0 to p = 6. We want to emphasize that the necessary
Newton and GMRES iterations for both HDG and DG were comparable. As there
are more faces than elements, DG is faster than HDG for p = 0 and p = 1. However, already for p = 2 HDG outruns DG. At p = 6 there is a ratio of 2.5 for the
Euler test case and 2.1 for the Navier-Stokes test case. For the Euler test case it
is interesting to note, that the iterative solver dominates the computational time
for DG. For HDG it is the other way around. This has two reasons: firstly, the
assembly is more involved due to the local solves; and secondly, the global system
is considerably smaller for higher polynomial degree. In the case of a Navier-Stokes
computation, the DG assembly takes over the dominating part as the lifting operators are very expensive to compute. The time for the HDG assembly is also
increased which is among other things due to the introduction of the gradient. For
both HDG and DG, the time spent in the iterative solver is comparable for both
Euler and Navier-Stokes.
5.1. Subsonic Inviscid Flow over the NACA 0012 Airfoil. In the first test
case, we consider subsonic inviscid flow over the NACA 0012 airfoil which is defined
by
√
y = ±0.6 0.2969 x − 0.1260x − 0.3516x2 + 0.2843x3 − 0.1015x4
with x ∈ [0, 1]. Using this definition, the airfoil would have a finite trailing edge
thickness of .252 %. In order to obtain a sharp trailing edge we modify the x4
coefficient, i.e.
√
y = ±0.6 0.2969 x − 0.1260x − 0.3516x2 + 0.2843x3 − 0.1036x4 .
The flow is characterized by a free stream Mach number of Ma∞ = 0.5 and an
angle of attack of α = 2◦ . In Fig. 3 the baseline mesh for the Euler test cases
14
WOOPEN, BALAN, MAY, SCHÜTZ
Assembly
40
Assembly
200
Linear Solve
Linear Solve
t
t
150
100
20
50
0
0
1
2
3
p
4
5
6
(a) Euler
0
0
1
2
3
p
4
5
6
(b) Navier-Stokes
Figure 2. Runtime comparison of the hybridized and nonhybridized DG method for a fixed mesh and varying polynomial
degree (Solid: HDG, Shaded: DG).
(subsonic and transonic) can be seen. It consists of 719 elements and its far field is
over a 1000 chords away.
Admissible target functionals defined on the boundary for the Euler equations
are given by the weighted pressure along wall boundaries, i.e.
Z
(76)
J(w) =
ψ · (pn) dσ
∂Ω
where n is the outward pointing normal. By using ψ =
1
C∞
T
1
C∞
T
(cos α, sin α)
or
ψ =
(− sin α, cos α) along wall boundaries and 0 otherwise, the functional
represents the pressure drag coefficient cD or the pressure lift coefficient cL , respectively. C∞ is a normalized reference value defined by C∞ = 12 γMa2∞ p∞ l. Here, l is
the chord length of the airfoil.
In Fig. 4a, a purely h-adapted mesh can be seen. The most refined regions are
the leading and trailing edge. The former is of importance as the flow experiences
high gradients towards the stagnation point. Refinement of the latter is necessary
due to the sharp trailing edge and the slip-wall boundary conditions. As soon as
the error in these two regions is sufficiently low, other elements close to the airfoil
get refined as well. For the hp-adapted mesh (see Fig. 4b) the leading and trailing
edge are refined as well. All other regions, however, undergo mostly p-enrichment.
In terms of degrees of freedom, both HDG and DG show similar results. For
all computations it takes some adaptations until the critical regions, leading and
trailing edge, are resolved. From this point on, one can see the benefit of a higher
order discretization: the error drops significantly faster with respect to degrees of
freedom and computational time (see Fig. 5 and 6).
In Tbl. 1, we give the runtime ratios for a fixed error level (we always choose
the minimum level attained). For p = 1, HDG and DG are comparable in both
runtime and nonzero entries. From p = 2 on, HDG is faster than DG and needs
less memories for the global system. Here, it is important to note that the adjoint is approximated with p + 1, so that already for lower orders HDG is faster.
A COMPARISON OF HDG AND DG FOR TARGET-BASED hp-ADAPTATION
15
Figure 3. Baseline mesh with 719 elements for inviscid computations
p
1
2
3
4
tDG /tHDG
1.293 3.799 3.551 3.092
nnz,DG /nnz,HDG 1.123 2.150 3.248 4.281
Table 1. Runtime and nonzero ratios for a fixed
(Ma∞ = 0.5, α = 2◦ )
hp
2.209
4.601
error level
Furthermore, we list the ratio of nonzero entries in the system matrix in order to
compare the memory requirements of both methods. As expected, the ratio grows
with increasing polynomial degree in favor of HDG.
5.2. Transonic Inviscid Flow over the NACA 0012 Airfoil. Next, we turn
our attention to transonic flow which develops more complex features (e.g. compression shocks) compared to the subsonic regime. The flow is characterized by a
free stream Mach number of Ma∞ = 0.8 and an angle of attack of α = 1.25◦ .
In Fig. 7a a purely h-adapted mesh can be seen. The adjoint sensor detects all
regions of relevance for the drag: the upper shock, the leading and trailing edge,
and the lower weak shock. Further refinement is added upstream of the shock,
where the adjoint has steep gradients and thus needs higher resolution. In the case
of hp-adaptation, the mesh-refinement is stronger confined to the shock region and
the trailing edge. The other features undergo p-enrichment.
As expected, both methods show a similar accuracy for a given number of degrees
of freedom. The computations with p = 2 . . . 4 outperform p = 1 but are comparable
to each other. hp-adaptation shows very good results which is due to the accurate
prediction of the solution smoothness (see Fig. 8 and 9).
For this test case, HDG is faster than DG from p = 1 on (see Tbl. 2). The hpadaptive run is more than 2.5 times faster. The ratio of necessary nonzero entries
attains its highest value for p = 4. In the case of hp-adaptation, this ratio is not as
high, as in the shock region a lot of elements with p = 2 exist.
16
WOOPEN, BALAN, MAY, SCHÜTZ
(a) Pure h-adapation (p = 2)
(b) hp-adapation (p = 2 . . . 5)
Figure 4. Adapted meshes for the subsonic Euler test case
(Ma∞ = 0.5, α = 2◦ )
p
1
2
3
4
tDG /tHDG
2.636 1.607 2.154 2.362
nnz,DG /nnz,HDG 1.237 1.807 3.098 4.437
Table 2. Runtime and nonzero ratios for a fixed
(Ma∞ = 0.8, α = 1.25◦ )
hp
2.628
2.411
error level
A COMPARISON OF HDG AND DG FOR TARGET-BASED hp-ADAPTATION
10−3
10−5
∆cD
∆cD
10−3
17
p=1
p=2
p=3
p=4
hp
10−7
10−5
10−7
103
104
ndof w
105
p=1
p=2
p=3
p=4
hp
103
104
ndof w
(a) HDG
105
(b) DG
Figure 5. Drag convergence with respect to degrees of freedom
(Ma∞ = 0.5, α = 2◦ )
10−5
10−7
10−2
10−3
∆cD
∆cD
10−3
p=1
p=2
p=3
p=4
hp
10−5
10−7
10−1
100
t/t0
(a) HDG
101
102
10−2
p=1
p=2
p=3
p=4
hp
10−1
100
t/t0
101
102
(b) DG
Figure 6. Drag convergence with respect to time (Ma∞ = 0.5, α = 2◦ )
5.3. Supersonic Inviscid Flow over the NACA 0012 Airfoil. Now, we consider supersonic, inviscid flow with a free stream Mach number of Ma = 1.4. We
choose the angle of attack to be α = 0◦ . In contrast to the transonic test case, a
so-called bow shock develops in front of the airfoil.
A purely h-adapted mesh is given in Fig. 10a. Several refinement regions can
be identified: the bow shock, leading and trailing edge, and a feature in the adjoint solution emanating from the trailing edge in the upstream direction. During
hp-adaptation, the latter feature and the leading edge are being treated with penrichment.
Both HDG and DG show a similar behavior in terms of error reduction with
respect to degrees of freedom (see Fig. 11). A comparison between different polynomial orders is rather inconclusive which might be due to the lack of smoothness
of this test case. Furthermore, we note that the convergence behavior is not monotonic (see for example p = 3 in Fig. 11). The error reduction with respect to
18
WOOPEN, BALAN, MAY, SCHÜTZ
(a) Pure h-adaptation (p = 2)
(b) hp-adaptation (p = 2 . . . 5)
Figure 7. Adapted meshes for the transonic Euler test case
(Ma∞ = 0.8, α = 1.25◦ )
computational time is similar (see Fig. 12). In both measures, the hp-adaptive run
gives the overall best solution.
Considering the time to solution, HDG is approximately 50% faster than DG
(see Tbl. 3). The hp-adaptive run is more than 3.5 times faster. Similar to the
transonic test case, the ratio of necessary nonzero entries attains its highest value
for p = 4. In the case of hp-adaptation, this ratio is not as high, as in the shock
region a lot of elements with p = 2 exist.
5.4. Subsonic Laminar Flow over the NACA 0012 Airfoil. Finally, we consider viscous flow in the subsonic regime. The free stream Mach number is Ma∞ =
10−1
10−1
10−2
10−2
10−3
10−4
10−5
103
∆cD
∆cD
A COMPARISON OF HDG AND DG FOR TARGET-BASED hp-ADAPTATION
p=1
p=2
p=3
p=4
hp
10−3
10−4
104
ndof w
10−5
103
105
19
p=1
p=2
p=3
p=4
hp
104
ndof w
(a) HDG
105
(b) DG
10−1
10−1
10−2
10−2
10−3
10−4
10−5
10−1
∆cD
∆cD
Figure 8. Drag convergence with respect to degrees of freedom
(Ma∞ = 0.8, α = 1.25◦ )
p=1
p=2
p=3
p=4
hp
10−3
10−4
101
100
t/t0
(a) HDG
102
10−5
10−1
p=1
p=2
p=3
p=4
hp
101
100
102
t/t0
(b) DG
Figure 9. Drag convergence with respect to time (Ma∞ = 0.8,
α = 1.25◦ )
p
1
2
3
4
tDG /tHDG
2.817 1.335 1.751 1.441
nnz,DG /nnz,HDG 1.212 2.132 3.429 4.030
Table 3. Runtime and nonzero ratios for a fixed
(Ma∞ = 1.4, α = 0◦ )
hp
3.554
3.806
error level
0.5, the angle of attack α = 1◦ and the Reynolds number Re = 5000. Due to the
latter, a thin boundary layer develops around the airfoil.
The baseline mesh for the Navier-Stokes test case is more refined around the
airfoil such that the boundary layer is correctly captured (see Fig. 13). It consists
of 1781 elements and its far field is over a 1000 chords away.
20
WOOPEN, BALAN, MAY, SCHÜTZ
(a) Pure h-adapation (p = 2)
(b) hp-adapation (p = 2 . . . 5)
Figure 10. Adapted meshes for the supersonic Euler test case
(Ma∞ = 1.4, α = 0◦ )
Admissible target functionals defined on the boundary for the Navier-Stokes
equations are given by the weighted boundary flux (including pressure and shear
forces) along wall boundaries, i.e.
Z
(77)
J (w, ∇w) =
ψ · (pn − τ n) dσ
∂Ω
where n is the outward pointing normal. Here, ψ is nonzero only on wall boundaries.
T
T
By using ψ = C1∞ (cos α, sin α) or ψ = C1∞ (− sin α, cos α) along wall boundaries
and otherwise 0, the functional represents the viscous drag coefficient cD or the
viscous lift coefficient cL , respectively.
10−2
10−2
10−3
10−3
10−4
10−5
10−6
103
∆cD
∆cD
A COMPARISON OF HDG AND DG FOR TARGET-BASED hp-ADAPTATION
p=1
p=2
p=3
p=4
hp
10−4
10−5
104
ndof w
10−6
103
105
(a) HDG
21
p=1
p=2
p=3
p=4
hp
104
ndof w
105
(b) DG
Figure 11. Drag convergence with respect to degrees of freedom
(Ma∞ = 1.4, α = 0◦ )
p=1
p=2
p=3
p=4
hp
∆cD
10−3
10−4
10−5
10−6
10−1
10−2
p=1
p=2
p=3
p=4
hp
10−3
∆cD
10−2
10−4
10−5
100
101
t/t0
(a) HDG
102
10−6
10−1
100
101
102
t/t0
(b) DG
Figure 12. Drag convergence with respect to time (Ma∞ = 1.4, α = 0◦ )
Both the h-adapted mesh (see Fig. 14a) and the hp-adapted mesh (see Fig. 14b)
undergo refinement within the boundary layer and the wake region. The mesh
refinement for the hp-adaptive run is however more restricted to the leading edge
region where the boundary layer develops. Further downstream, p-enrichment is
used as soon as the necessary mesh-resolution is reached.
In terms of accuracy versus degrees of freedom, HDG and DG perform comparably well. The higher the polynomial degree the more accurate and efficient the
computations are for both HDG and DG (see Fig. 15 and 16). The difference between hp, p = 3 and p = 4 is not as big, though. This might lead to the conclusion
that isotropic mesh refinement is not longer efficiently applicable in cases involving
strong gradients. Anisotropic refinement might be able to remedy this problem so
that the full potential of hp-adaptation can be attained.
Concerning the timings, we can see a similar trend as in the previous test cases
(see Tbl. 4). For p = 1 . . . 4, HDG is more than twice as fast. The hp-adaptive
22
WOOPEN, BALAN, MAY, SCHÜTZ
Figure 13. Baseline mesh with 1781 elements for viscous computations
p
1
2
3
4
tDG /tHDG
2.244 2.572 2.397 2.288
nnz,DG /nnz,HDG 1.196 2.235 3.472 4.531
Table 4. Runtime and nonzero ratios for a fixed
(Ma∞ = 0.5, α = 1◦ , Re = 5000)
hp
3.122
3.463
error level
HDG computation is even three times as fast compared to the DG run. For p ≥ 2
the savings in nonzero entries for HDG become significant.
6. Conclusion and Outlook
We presented an adjoint-based hp-adaptation methodology and compared it for
hybridized and non-hybridized discontinuous Galerkin methods. hp-adaptation
proved to be superior to pure h-adaptation if discontinuous or singular flow features
were involved. In all cases, a higher polynomial degree turned out to be beneficial.
We showed that one can expect HDG to be better than DG in terms of runtime and
memory requirements from p = 2 on for a broad range of test cases. The results
are promising enough to pursue further comparisons with more complex test cases.
We plan to extend our computational framework to three dimensional problems
[34]. Then, adaptivity will play an even more crucial role, as the problem size
increases drastically compared to the two dimensional case.
Furthermore, compressible flows are often dominated by anisotropic features,
such as shocks or very thin boundary layers. Thus, taking this anisotropy into
account during adaptation is crucial. We plan to incorporate anisotropic adaptation
within future projects.
Acknowledgments
Financial support from the Deutsche Forschungsgemeinschaft (German Research
Association) through grant GSC 111 is gratefully acknowledged.
A COMPARISON OF HDG AND DG FOR TARGET-BASED hp-ADAPTATION
23
(a) Pure h-adaptation (p = 2)
(b) hp-adaptation (p = 2 . . . 5)
Figure 14. Adapted meshes for the Navier-Stokes test case
(Ma∞ = 0.5, α = 1◦ , Re = 5000)
References
[1] F. Bassi, S. Rebay, A High-Order Accurate Discontinuous Finite-Element
Method for the Numerical Solution of the Compressible Navier-Stokes Equations, Journal of Computational Physics 131 (1997) 267–279.
[2] B. Cockburn, C.-W. Shu, TVB Runge-Kutta Local Projection Discontinuous
Galerkin Finite Element Method for Conservation Laws II: General Framework, Mathematics of Computation 52 (1988) 411–435.
[3] D. Arnold, F. Brezzi, B. Cockburn, L. Marini, Unified Analysis of Discontinuous Galerkin Methods for Elliptic Problems, SIAM Journal on Numerical
Analysis 39 (2002) 1749–1779.
WOOPEN, BALAN, MAY, SCHÜTZ
10−2
10−2
10−3
10−3
10−4
∆cD
∆cD
24
p=1
p=2
p=3
p=4
hp
10−5
10−6
10−4
10−5
104
10−6
105
ndof w
p=1
p=2
p=3
p=4
hp
104
(a) HDG
105
ndof w
(b) DG
10−2
10−2
10−3
10−3
10−4
10−5
10−6
10−1
∆cD
∆cD
Figure 15. Drag convergence with respect to degrees of freedom
(Ma∞ = 0.5, α = 1◦ , Re = 5000)
p=1
p=2
p=3
p=4
hp
10−4
10−5
100
101
t/t0
(a) HDG
102
103
10−6
10−1
p=1
p=2
p=3
p=4
hp
100
101
t/t0
102
103
(b) DG
Figure 16. Drag convergence with respect to time (Ma∞ = 0.5,
α = 1◦ , Re = 5000)
[4] B. F. de Veubeke, Displacement and Equilibrium Models in the Finite Element
Method, chap. 9, Wiley New York, 145–197, 1965.
[5] B. Cockburn, J. Gopalakrishnan, A Characterization of Hybridized Mixed
Methods for Second Order Elliptic Problems, SIAM Journal on Numerical
Analysis 42 (1) (2004) 283–301.
[6] B. Cockburn, J. Gopalakrishnan, R. Lazarov, Unified Hybridization of Discontinuous Galerkin, Mixed, and Continuous Galerkin Methods for Second
Order Elliptic Problems, SIAM Journal on Numerical Analysis 47 (2) (2009)
1319–1365.
[7] N. Nguyen, J. Peraire, B. Cockburn, An Implicit High-Order Hybridizable
Discontinuous Galerkin Method for Linear Convection-Diffusion Equations,
Journal of Computational Physics 228 (9) (2009) 3232–3254.
A COMPARISON OF HDG AND DG FOR TARGET-BASED hp-ADAPTATION
25
[8] N. Nguyen, J. Peraire, B. Cockburn, An Implicit High-Order Hybridizable
Discontinuous Galerkin Method for Nonlinear Convection-Diffusion Equations,
Journal of Computational Physics 228 (2009) 8841–8855.
[9] J. Peraire, N. Nguyen, B. Cockburn, A Hybridizable Discontinuous Galerkin
Method for the Compressible Euler and Navier-Stokes Equations, AIAA Paper
2010-0363, 2010.
[10] A. Huerta, A. Angeloski, X. Roca, J. Peraire, Efficiency of High-Order Elements for Continuous and Discontinuous Galerkin Methods, Int. J. Numer.
Methods Eng. 96 (9) (2013) 529–560.
[11] R. Hartmann, P. Houston, Adaptive Discontinuous Galerkin Finite Element
Methods for the Compressible Euler Equations, Journal of Computational
Physics 183 (2) (2002) 508–532.
[12] D. Venditti, D. Darmofal, Grid Adaptation for Functional Outputs: Application to Two-Dimensional Inviscid Flows, Journal of Computational Physics
176 (1) (2002) 40–69.
[13] D. Venditti, D. Darmofal, Anisotropic Grid Adaptation for Functional Outputs: Application to Two-Dimensional Viscous Flows, Journal of Computational Physics 187 (1) (2003) 22–46.
[14] R. Hartmann, Adaptive Discontinuous Galerkin Methods with ShockCapturing for the Compressible Navier–Stokes Equations, International Journal for Numerical Methods in Fluids 51 (9-10) (2006) 1131–1156.
[15] K. Fidkowski, D. Darmofal, Review of Output-Based Error Estimation and
Mesh Adaptation in Computational Fluid Dynamics, AIAA Journal 49 (4)
(2011) 673–694.
[16] G. Giorgiani, S. Fernández-Méndez, A. Huerta, Hybridizable Discontinuous
Galerkin p-Adaptivity for Wave Propagation Problems, International Journal
for Numerical Methods in Fluids .
[17] J. Schütz, G. May, A Hybrid Mixed Method for the Compressible Navier–
Stokes Equations, Journal of Computational Physics 240 (2013) 58–75.
[18] J. Schütz, G. May, An Adjoint Consistency Analysis for a Class of Hybrid
Mixed Methods, IMA Journal of Numerical Analysis .
[19] M. Woopen, G. May, J. Schütz, Adjoint-Based Error Estimation and Mesh
Adaptation for Hybridized Discontinuous Galerkin Methods, arXiv preprint
arXiv:1309.3649 .
[20] A. Balan, M. Woopen, G. May, Adjoint-Based hp-Adaptation for a Class of
High-Order Hybridized Finite Element Schemes for Compressible Flows, AIAA
Paper 2013-2938, 2013.
[21] E. Godlewski, P. Raviart, Numerical Approximation of Hyperbolic Systems of
Conservation Laws, no. Nr. 118 in Applied Mathematical Sciences, Springer,
ISBN 9780387945293, 1996.
[22] Y. Chen, B. Cockburn, Analysis of Variable-Degree HDG Methods for
Convection–Diffusion Equations. Part I: General Nonconforming Meshes, IMA
Journal of Numerical Analysis 32 (4) (2012) 1267–1293.
[23] B. Cockburn, J. Gopalakrishnan, Error Analysis of Variable Degree Mixed
Methods for Elliptic Problems via Hybridization, Mathematics of Computation
74 (252) (2005) 1653–1677.
[24] D. Mavriplis, A. Jameson, Multigrid Solution of the Navier-Stokes Equations
on Triangular Meshes, AIAA Journal 28 (8) (1990) 1415–1425.
26
WOOPEN, BALAN, MAY, SCHÜTZ
[25] S. Balay, J. Brown, K. Buschelman, W. Gropp, D. Kaushik,
M. Knepley, L. McInnes, B. Smith, H. Zhang, PETSc Web Page,
http://www.mcs.anl.gov/petsc, 2013.
[26] S. Balay, J. Brown, K. Buschelman, V. Eijkhout, W. Gropp, D. Kaushik,
M. Knepley, L. McInnes, B. Smith, H. Zhang, PETSc Users Manual, Tech.
Rep. ANL-95/11 - Revision 3.4, Argonne National Laboratory, 2013.
[27] W. Dörfler, A Convergent Adaptive Algorithm for Poisson’s Equation, SIAM
Journal on Numerical Analysis 33 (3) (1996) 1106–1124.
[28] M. Ceze, K. Fidkowski, Output-Driven Anisotropic Mesh Adaptation for Viscous Flows using Discrete Choice Optimization, AIAA Paper 2010-0170, 2010.
[29] L. Wang, D. J. Mavriplis, Adjoint-Based h-p Adaptive Discontinuous Galerkin
Methods for the 2D Compressible Euler Equations, Journal of Computational
Physics 228 (2009) 7643–7661.
[30] P. Persson, J. Peraire, Sub-cell Shock Capturing for Discontinuous Galerkin
Methods, AIAA Paper 2006-0112, 2006.
[31] M. Dubiner, Spectral Methods on Triangles and Other Domains, J. Sci. Comput. 6 (4) (1991) 345–390.
[32] F. Bassi, A. Crivellini, S. Rebay, M. Savini, Discontinuous Galerkin Solution of
the Reynolds-Averaged Navier-Stokes and k-ω Turbulence Model Equations,
Computers & Fluids 34 (4) (2005) 507–540.
[33] R. Hartmann, Adjoint Consistency Analysis of Discontinuous Galerkin Discretizations, SIAM Journal on Numerical Analysis 45 (6) (2007) 2671–2696.
[34] M. Woopen, A. Balan, G. May, A Hybridized Discontinuous Galerkin Method
for Three-Dimensional Compressible Flow Problems, AIAA Paper 2014-0938,
2014.
| 5cs.CE
|
arXiv:1611.04196v3 [cs.IT] 21 Nov 2017
Self-Calibration and Bilinear Inverse Problems
via Linear Least Squares∗
Shuyang Ling†, Thomas Strohmer‡
November 22, 2017
Abstract
Whenever we use devices to take measurements, calibration is indispensable. While the
purpose of calibration is to reduce bias and uncertainty in the measurements, it can be quite
difficult, expensive, and sometimes even impossible to implement. We study a challenging
problem called self-calibration, i.e., the task of designing an algorithm for devices so that the
algorithm is able to perform calibration automatically. More precisely, we consider the setup
y = A(d)x+ε where only partial information about the sensing matrix A(d) is known and where
A(d) linearly depends on d. The goal is to estimate the calibration parameter d (resolve the
uncertainty in the sensing process) and the signal/object of interests x simultaneously. For three
different models of practical relevance, we show how such a bilinear inverse problem, including
blind deconvolution as an important example, can be solved via a simple linear least squares
approach. As a consequence, the proposed algorithms are numerically extremely efficient, thus
potentially allowing for real-time deployment. We also present a variation of the least squares
approach, which leads to a spectral method, where the solution to the bilinear inverse problem
can be found by computing the singular vector associated with the smallest singular value of a
certain matrix derived from the bilinear system. Explicit theoretical guarantees and stability
theory are derived for both techniques; and the number of sampling complexity is nearly optimal
(up to a poly-log factor). Applications in imaging sciences and signal processing are discussed
and numerical simulations are presented to demonstrate the effectiveness and efficiency of our
approach.
1
Introduction
Calibration is ubiquitous in all fields of science and engineering. It is an essential step to guarantee
that the devices measure accurately what scientists and engineers want. If sensor devices are not
properly calibrated, their measurements are likely of little use to the application. While calibration
is mostly done by specialists, it often can be expensive, time-consuming and sometimes even impossible to do in practice. Hence, one may wonder whether it is possible to enable machines to calibrate
themselves automatically with a smart algorithm and give the desired measurements. This leads
to the challenging field of self-calibration (or blind calibration). It has a long history in imaging
sciences, such as camera self-calibration [33, 22], blind image deconvolution [12], self-calibration in
∗
This research was supported by the NSF via Award Nr. dtra-dms 1322393 and Award Nr. DMS 1620455.
Courant Institute of Mathematical Sciences and the Center for Data Science, New York University, NY 10003
(Email: [email protected])
‡
Department of Mathematics, University of California Davis, CA 95616 (Email: [email protected]).
†
1
medical imaging [35], and the well-known phase retrieval problem (phase calibration) [16]. It also
plays an important role in signal processing [18] and wireless communications [38, 32].
Self-calibration is not only a challenging problem for engineers, but also for mathematicians. It
means that one needs to estimate the calibration parameter of the devices to adjust the measurements as well as recover the signal of interests. More precisely, many self-calibration problems are
expressed in the following mathematical form,
y = A(d)x + ε,
(1.1)
where y is the observation, A(d) is a partially unknown sensing matrix, which depends on an unknown parameter d and x is the desired signal. An uncalibrated sensor/device directly corresponds
to “imperfect sensing”, i.e., uncertainty exists within the sensing procedure and we do not know
everything about A(d) due to the lack of calibration. The purpose of self-calibration is to resolve
the uncertainty i.e., to estimate d in A(d) and to recover the signal x at the same time.
The general model (1.1) is too hard to get meaningful solutions without any further assumption
since there are many variants of the general model under different settings. In (1.1), A(d) may
depend on d in a nonlinear way, e.g., d can be the unknown orientation of a protein molecule
and x is the desired object [44]; in phase retrieval, d is the unknown phase information of the
Fourier transform of the object [16]; in direction-of-arrival estimation d represents unknown offset,
gain, and phase of the sensors [47]. Hence, it is impossible to resolve every issue in this field,
but we want to understand several scenarios of self-calibration which have great potential in real
world applications. Among all the cases of interest, we assume that A(d) linearly depends on the
unknown d and will explore three different types of self-calibration models that are of considerable
practical relevance. However, even for linear dependence, the problem is already quite challenging,
since in fact we are dealing with bilinear (nonlinear) inverse problems. All those three models have
wide applications in imaging sciences, signal processing, wireless communications, etc., which will
be addressed later. Common to these applications is the desire or need for fast algorithms, which
ideally should be accompanied by theoretical performance guarantees. We will show under certain
cases, these bilinear problems can be solved by linear least squares exactly and efficiently if no
noise exists, which is guaranteed by rigorous mathematical proofs. Moreover, we prove that the
solution is also robust to noise with tools from random matrix theory. Furthermore, we show that
a variation of our approach leads to a spectral method, where the solution to the bilinear problem
can be found by computing the singular vector associated with the smallest singular matrix of a
certain matrix derived from the bilinear system.
1.1
State of the art
By assuming that A(d) linearly depends on d, (1.1) becomes a bilinear inverse problem, i.e., we
want to estimate d and x from y, where y is the output of a bilinear map from (d, x). Bilinear
inverse problems, due to its importance, are getting more and more attentions over the last few
years. On the other hand, they are also notoriously difficult to solve in general. Bilinear inverse
problems are closely related to low-rank matrix recovery, see [15] for a comprehensive review. There
exists extensive literature on this topic and it is not possible do justice to all these contributions.
Instead we will only highlight some of the works which have inspired us.
Blind deconvolution might be one of the most important examples of bilinear inverse problems [12], i.e., recovering f and g from y = f ∗ g, where “∗” stands for convolution. If both
f and g are inside known low-dimensional subspaces, the blind deconvolution can be rewritten
2
as F(y) = diag(Bd)Ax, where F(f ) = Bd, F(g) = Ax and “F” denotes the Fourier transform. In the inspiring work [4], Ahmed, Romberg and Recht apply the “lifting” techniques [13]
and convert the problem into estimation of the rank-one matrix dx∗ . It is shown that solving a
convex relaxation enables recovery of dx∗ under certain choices of B and A. Following a similar
spirit, [30] uses “lifting” combined with a convex approach to solve the scenarios with sparse x
and [29] studies the so called “blind deconvolution and blind demixing” problem. The other line of
blind deconvolution follows a nonconvex optimization approach [3, 26, 25]. In [3], Ahmed, Romberg
and Krahmer, using tools from generic chaining, obtain local convergence of a sparse power factorization algorithm to solve this blind deconvolution problem when h and x are sparse and B and A
are Gaussian random matrices. Under the same setting as [3], Lee et al. [25] propose a projected
gradient descent algorithm based on matrix factorizations and provide a convergence analysis to
recover sparse signals from subsampled convolution. However, this projection step can be hard to
implement. As an alternative, the expensive projection step is replaced by a heuristic approximate
projection, but then the global convergence is not fully guaranteed. Both [3, 25] achieve nearly
optimal sampling complexity. [26] proves global convergence of a gradient descent type algorithm
when B is a deterministic Fourier type matrix and A is Gaussian. Results about identifiability
issue of bilinear inverse problems can be found in [27, 23, 28].
Another example of self-calibration focuses on the setup yl = DAl x, where D = diag(d).
The difference from the previous model consists in replacing the subspace assumption by multiple
measurements. There are two main applications of this model. One application deals with blind
deconvolution in an imaging system which uses randomly coded masks [5, 37]. The measurements
are obtained by (subsampled) convolution of an unknown blurring function D with several random
binary modulations of one image. Both [5] and [2] developed convex relaxing approaches (nuclear
norm minimization) to achieve exact recovery of the signals and the blurring function. The other
application is concerned with calibration of the unknown gains and phases D and recovery of the
signal x, see e.g. [10, 9]. Cambareri and Jacques propose a gradient descent type algorithm in [10, 11]
and show convergence of the iterates by first constructing a proper initial guess. An empirical
study is given in [9] when x is sparse by applying an alternating hard thresholding algorithm.
Recently, [1, 2] study the blind deconvolution when inputs are changing. More precisely, the authors
consider yl = f ∗ gl where each gl belongs to a different known subspace, i.e., gl = Cl xl . They
employ a similar convex approach as in [4] to achieve exact recovery with number of measurements
close to the information theoretic limit.
An even more difficult, and from a practical viewpoint highly relevant, scenario focuses on selfcalibration from multiple snapshots [47]. Here, one wishes to recover the unknown gains/phases
D = diag(d) and a signal matrix X = [x1 , · · · , xp ] from Y = DAX. For this model, the
sensing matrix A is fixed throughout the sensing process and one measures output under different
snapshots {xl }pl=1 . One wants to understand under what conditions we can identify D and {xl }pl=1
jointly. If A is a Fourier type matrix, this model has applications in both image restoration from
multiple filters [21] and also network calibration [6, 31]. We especially benefitted from work by
Gribonval and coauthors [20, 8], as well as by Balzano and Nowak [6, 7]. The papers [6, 7] study
the noiseless version of the problem by solving a linear system and [31] takes a total least squares
approach in order to obtain empirically robust recovery in the presence of noise. If each xl is
sparse, this model becomes more difficult and Gribonval et al. [8, 20] give a thorough numerical
study. Very recently, [43] gave a theoretic result under certain conditions. This calibration problem
is viewed as a special case of the dictionary learning problem where the underlying dictionary DA
3
possesses some additional structure. The idea of transforming a blind deconvolution problem into a
linear problem can also be found in [32], where the authors analyze a certain non-coherent wireless
communication scenario.
1.2
Our contributions
In our work, we consider three different models of self-calibration, namely, yl = DAl x, yl = DAl xl
and yl = DAxl . Detailed descriptions of these models are given in the next Section. We do not
impose any sparsity constraints on x or xl . We want to find out xl (or x) and D when yl (or y) and
Al (or A) are given. Roughly, they correspond to the models in [5, 1, 8] respectively. Though all of
the three models belong to the class of bilinear inverse problems, we will prove that simply solving
linear least squares will give solutions to all those models exactly and robustly for invertible D and
for several useful choices of A and Al . Moreover, the sampling complexity is nearly optimal (up to
poly-log factors) with respect to the information theoretic limit (degree of freedom of unknowns).
As mentioned before, our approach is largely inspired by [8] and [6, 7]; there the authors convert
a bilinear inverse problem into a linear problem via a proper transformation. We follow a similar
approach in our paper. The paper [8] provides an extensive empirical study, but no theoretical analysis. Nowak and Balzano, in [6, 7] provide numerical simulations as well as theoretical conditions
on the number of measurements required to solve the noiseless case.
Our paper goes an important step further: On the one hand we consider more general selfcalibration settings. And on the other hand we provide a rigorous theoretical analysis for recoverability and, perhaps most importantly, stability theory in the presence of measurement errors.
Owing to the simplicity of our approach and the structural properties of the underlying matrices, our framework yields self-calibration algorithms that are numerically extremely efficient, thus
potentially allowing for deployment in applications where real-time self-calibration is needed.
1.3
Notation and Outline
We introduce notation which will be used throughout the paper. Matrices are denoted in boldface
or a calligraphic font such as Z and Z; vectors are denoted by boldface lower case letters, e.g. z.
The individual entries of a matrix or a vector are denoted in normal font such as Zij or zi . For any
matrix Z, kZk denotes its operator
qP norm, i.e., the largest singular value, and kZkF denotes its the
2
Frobenius norm, i.e., kZkF =
ij |Zij | . For any vector z, kzk denotes its Euclidean norm. For
both matrices and vectors, Z T and z T stand for the transpose of Z and z respectively while Z ∗
and z ∗ denote their complex conjugate transpose. For any real number z, we let z+ = 12 (z + |z|).
We equip the matrix space CK×N with the inner product defined by hU , V i := Tr(U ∗ V ). A special
case is the inner product of two vectors, i.e., hu, vi = Tr(u∗ v) = u∗ v. We define the correlation
u∗ v
. For a given vector v, diag(v) represents the
between two vectors u and v as Corr(u, v) = kukkvk
diagonal matrix whose diagonal entries are given by the vector v.
C is an absolute constant and Cγ is a constant which depends linearly on γ, but on no other
parameters. In and 1n always denote the n × n identity matrix and a column vector of “1” in
Rn respectively. And {ei }m
el }pl=1 stand for the standard orthonormal basis in Rm and Rp
i=1 and {e
respectively. “∗” is the circular convolution and “⊗” is the Kronecker product.
The paper is organized as follows. The more detailed discussion of the models under consideration and the proposed method will be given in Section 2. Section 3 presents the main results
4
of our paper and we will give numerical simulations in Section 4. Section 5 contains the proof for
each scenario. We collect some useful auxiliary results in the Appendix.
2
Problem setup: Three self-calibration models
This section is devoted to describing three different models for self-calibration in detail. We will also
explain how those bilinear inverse problems are reformulated and solved via linear least squares.
2.1
Three special models of self-calibration
Self-calibration via repeated measurements Suppose we are seeking for information with
respect to an unknown signal x0 with several randomized linear sensing designs. Throughout this
procedure, the calibration parameter D remains the same for each sensing procedure. How can
we recover the signal x0 and D simultaneously? Let us make it more concrete by introducing the
following model,
yl = DAl x0 + εl , 1 ≤ l ≤ p
(2.1)
where D = diag(d) ∈ Cm×m is a diagonal matrix and each Al ∈ Cm×n is a measurement matrix.
Here yl and Al are given while D and x0 are unknown. For simplicity, we refer to the setup (2.1)
as “self-calibration from repeated measurements”. This model has various applications in selfcalibration for imaging systems [10, 11], networks [6], as well as in blind deconvolution from random
masks [5, 37].
Blind deconvolution via diverse inputs Suppose that one sends several different signals
through the same unknown channel, and each signal is encoded differently. Namely, we are considering
yl = f ∗ Cl xl + εl , 1 ≤ l ≤ p.
How can one estimate the channel and each signal jointly? In the frequency domain, this “blind
deconvolution via diverse inputs” [1, 2] problem can be written as (with a bit abuse of notation),
yl = DAl xl + εl ,
1≤l≤p
(2.2)
where D = diag(d) ∈ Cm×m and Al ∈ Cm×n are the Fourier transform of f and Cl respectively.
We aim to recover {xl }pl=1 and D from {yl , Al }pl=1 .
Self-calibration from multiple snapshots Suppose we take measurements of several signals
{xl }pl=1 with the same set of design matrix DA (i.e., each sensor corresponds one row of A and has
an unknown complex-valued calibration term di ). When and how can we recover D and {xl }pl=1
simultaneously? More precisely, we consider the following model of self-calibration model from
multiple snapshots:
yl = DAxl + εl , 1 ≤ l ≤ p.
(2.3)
Here D = diag(d) is an unknown diagonal matrix, A ∈ Cm×n is a sensing matrix, {xl }pl=1 are
n × 1 unknown signals and {yl }pl=1 are their corresponding observations. This multiple snapshots
model has been used in image restoration from multiple filters [21] and self-calibration model for
sensors [47, 6, 31, 20, 8].
5
2.2
Linear least squares approach
Throughout our discussion, we assume that D is invertible, and we let S := diag(s) = D−1 . Here,
D = diag(d) stands for the calibration factors of the sensor(s) [6, 8] and hence it is reasonable to
assume invertibility of D. For, if a sensor’s gain were equal to zero, then it would not contribute
any measurements to the observable y, in which case the associated entry of y would be zero.
But then we could simply discard that entry and consider the correspondingly reduced system of
equations, for which the associated D is now invertible.
One simple solution is to minimize a nonlinear least squares objective function. Let us take (2.1)
as an example (the others (2.2) and (2.3) have quite similar formulations),
min
D,x
p
X
l=1
kDAl x − yl k2 .
(2.4)
The obvious difficulty lies in the biconvexity of (2.4), i.e., if either D or x is fixed, minimizing over
the other variable is a convex program. In general, there is no way to guarantee that any gradient
descent algorithm/alternating minimization will give the global minimum. However, for the three
models described above, there is one shortcut towards the exact and robust recovery of the solution
via linear least squares if D is invertible.
We continue with (2.1) when εl = 0, i.e.,
diag(yl )s = Al x,
1≤l≤p
(2.5)
where Syl = diag(yl )s with si = d−1
and S is defined as diag(s). The original measurement
i
equation turns out to be a linear system with unknown s and x. The same idea of linearization
can be also found in [20, 8, 43, 6]. In this way, the ground truth z0 := (s0 , x0 ) lies actually inside
the null space of this linear system.
Two issues arise immediately: One the one hand, we need to make sure that (s0 , x0 ) spans
the whole null space of this linear system. This is equivalent to the identifiability issue of bilinear
problems of the form (2.4), because if the pair (α−1 d0 , αx0 ) for some α 6= 0 is (up to the scalar
α) unique solution to (2.1), then (αs0 , αx0 ) spans the null space of (2.5), see also [27, 23, 28]. On
the other hand, we also need to avoid the trivial scenario (s0 , x0 ) = (0, 0), since it has no physical
meaning. To resolve the latter issue, we add the extra linear constraint (see also [8, 6])
s
w,
= c,
(2.6)
x
where the scalar c can be any nonzero number (we note that w should of course not be orthogonal
to the solution). Therefore, we hope that in the noiseless case it suffices to solve the following linear
system to recover (d, x0 ) up to a scalar, i.e.,
diag(y1 ) −A1
..
..
0
s
.
.
(2.7)
=
c (mp+1)×1
x (m+n)×1
diag(yp ) −Ap
{z
} |
{z
}
|
w∗
z
(mp+1)×(m+n)
b
{z
}
|
Aw
6
In the presence of additive noise, we replace the linear system above by a linear least squares
problem
p
X
minimize
k diag(yl )s − Al xk2 + |w ∗ z − c|2
l=1
with respect to s and x, or equivalently,
minimize kAw z − bk2
z
(2.8)
s
0
where z =
,b=
, and Aw is the matrix on the left hand side of (2.7). Following the same
x
c
idea, (2.2) and (2.3) can also be reformulated into linear systems and be solved via linear least
squares. The matrix Aw and the vector z take a slightly different form for those cases, see (3.1)
and (3.3), respectively.
Remark 2.1. Note that solving (2.8) may not be the optimal choice to recover the unknowns
from the perspective of statistics since the noisy perturbation actually enters into Aw instead of b.
More precisely, the noisy perturbation δA to the left hand side of the corresponding linear system
for (2.1), (2.2) and (2.3), is always in the form of
diag(ε1 ) 0
..
..
.
.
(2.9)
δA :=
.
diag(εp ) 0
0
0
The size of δA depends on the models. Hence total least squares [31] could be a better alternative
while it is more difficult to analyze and significantly more costly to compute. Since computational
efficiency is essential for many practical applications, a straightforward implementation of total least
squares is of limited use. Instead one should keep in mind that the actual perturbation enters only
into diag(yl ), while the other matrix blocks remain unperturbed. Constructing a total least squares
solution that obeys these constraints, doing so in a numerically efficient manner and providing
theoretical error bounds for it, is a rather challenging task, which we plan to address in our future
work.
Remark 2.2. Numerical simulations imply that the performance under noisy measurements depends on the choice of w, especially how much w and z0 are correlated. One extreme case is that
hw, z0 i = 0, in which case we cannot avoid the solution z = 0. It might be better to add a constraint
like kwk = 1. However, this will lead to a nonlinear problem which may not be solved efficiently
and not come with rigorous recovery guarantees. Therefore, we present an alternative approach in
the next subsection.
2.3
Spectral method
In this subsection, we discuss a method for solving the self-calibration problem, whose performance
does not depend on the choice of w as it avoids the need of w in the first place. Let S be the
7
matrix Aw excluding the last row (the one which contains w). We decompose S into S = S0 + δS
where S0 is the noiseless part of S and δS is the noisy part1 .
We start with the noise-free scenario: if δS = 0, there holds S = S0 and the right singular
vector of S corresponding to the smallest singular value is actually z0 . Therefore we can recover z0
by solving the following optimization problem:
min kSzk.
kzk=1
(2.10)
Obviously, its solution is equivalent to the smallest singular value of S and its corresponding singular
vector.
If noise exists, the performance will depend on how large the second smallest singular value of
S0 is and on the amount of noise, given by kδSk. We will discuss the corresponding theory and
algorithms in Section 3.4, and the proof in Section 5.4.
3
Theoretical results
We present our theoretical findings for the three models (2.1), (2.2) and (2.3) respectively for
different choices of Al or A. In one of our choices the Al are Gaussian random matrices. The
rationale for this choice is that, while a Gaussian random matrix is not useful or feasible in most
applications, it often provides a benchmark for theoretical guarantees and numerical performance.
Our other choices for the sensing matrices are structured random matrices, such as e.g. the product
of a deterministic partial (or a randomly subsampled) Fourier matrix or a Hadamard matrix2 with a
diagonal binary random matrix3 . These matrices bring us closer to what we encounter in real world
applications. Indeed, structured random matrices of this type have been deployed for instance in
imaging and wireless communications, see e.g. [19, 41].
By solving simple variations (for different models) of (2.8), we can guarantee that the ground
truth is recovered exactly up to a scalar if no noise exists and robustly if noise is present. The
number of measurements required for exact and robust recovery is nearly optimal, i.e., close to the
information-theoretic limit up to a poly-log factor. However, the error bound for robust recovery
is not optimal. It is worth mentioning that once the signals and calibration parameter D are
identifiable, we are able to recover both of them exactly in absence of noise by simply solving a
linear system. However, identifiability alone cannot guarantee robustness.
Throughout this section, we let dmax := max1≤i≤m |di,0 | and dmin := min1≤i≤m |di,0 | where
{di,0 }m
i=1 are the entries of the ground truth d0 . We also define Aw,0 as the noiseless part of Aw
for each individual model.
3.1
Self-calibration via repeated measurements
For model (2.1) we will focus on three cases:
1
This is a slight abuse of notation, since the δA defined in (2.9) has an additional row with zeros. However, it will
be clear from the context which δA we refer to, and more importantly, in our estimates we mainly care about kδAk
which coincides for both choices of δA.
2
A randomly subsampled Fourier matrix is one, where we randomly choose a certain number of rows or columns
of the Discrete Fourier Transform matrix, and analogously for a randomly subsampled Hadamard matrix
3
At this point, we are not able to prove competitive results for fully deterministic sensing matrices.
8
(a) Al is an m × n complex Gaussian random matrix, i.e., each entry in Al is given by
√i N (0, 1).
2
√1 N (0, 1) +
2
(b) Al is an m × n “tall” random DFT/Hadamard matrix with m ≥ n, i.e., Al := HMl where H
consists of the first n columns of an m × m DFT/Hadamard matrix and each Ml := diag(ml )
is a diagonal matrix with entries taking on the value ±1 with equal probability. In particular,
there holds,
A∗l Al = Ml∗ H ∗ HMl = mIn .
(c) Al is an m × n “fat” random partial DFT matrix with m < n, i.e., Al := HMl where H
consists of m columns of an n × n DFT/Hadamard matrix and each Ml := diag(ml ) is a
diagonal matrix, which is defined the same as case (b),
Al A∗l = HH ∗ = nIm .
Our main findings are summarized as follows:
Theorem 3.1. Consider the self-calibration model given in (2.1), where Aw is as in (2.7) and
Aw,0 is the noiseless part of Aw . Then, for the solution ẑ of (2.8) and α = w∗cz0 , there holds
2
kẑ − αz0 k
≤ κ(Aw,0 )η 1 +
kαz0 k
1 − κ(Aw,0 )η
if κ(Aw,0 )η < 1 where η =
2kδAk
√
mp .
κ(Aw,0 ) ≤
s
The condition number of Aw satisfies
max{d2max kxk2 , m}
6(mp + kwk2 )
,
min{mp, kwk2 | Corr(w, z0 )|2 } min{d2min kxk2 , m}
√
mp
s
√
2 3
max{d2max kxk2 , m}
κ(Aw,0 ) ≤
,
| Corr(w, z0 )| min{d2min kxk2 , m}
n
log2 (m + n);
(a) with probability 1 − (m + n)−γ if Al is Gaussian and p ≥ c0 γ max 1, m
where Corr(w, z0 ) =
w ∗ z0
kwkkz0 k ,
and for kwk =
(b) with probability 1 − (m + n)−γ − 2(mp)−γ+1 if each Al is a “tall” (m × n, m ≥ n) random
Hadamard/DFT matrix and p ≥ c0 γ 2 log(m + n) log(mp);
(c) with probability 1 − (m + n)−γ − 2(mp)−γ+1 if each Al is a “fat” (m × n, m ≤ n) random
Hadamard/DFT matrix and mp ≥ c0 γ 2 n log(m + n) log(mp).
Remark 3.2. Our result is nearly optimal in terms of required number of measurements, because
the number of constraints mp is required to be slightly greater than n + m, the number of unknowns.
Theorem 3.1 can be regarded as a generalized result of [10, 11], in which D is assumed to be positive
and Al is Gaussian. In our result, we only need D to be an invertible complex diagonal matrix and
Al can be a Gaussian or random Fourier type matrix. The approaches are quite different, i.e., [10]
essentially uses nonconvex optimization by first constructing a good initial guess and then applying
gradient descent to recover D and x. Our result also provides a provable fast alternative algorithm
to “the blind deconvolution via random masks” in [5, 37] where a SDP-based approach is proposed.
9
3.2
Blind deconvolution via diverse inputs
We now analyze model (2.2). Following similar steps that led us from (2.1) to (2.7), it is easy to
see that the linear system associated with (2.2) is given by
diag(y1 ) −A1
0
···
0
s
diag(y2 )
0
−A2 · · ·
0
x1
0
..
..
..
.
.
x
..
..
=
.
(3.1)
2
.
.
.
c (mp+1)×1
..
diag(yp )
.
0
0
· · · −Ap
x
w∗
p (m+n)×1
(mp+1)×(np+m)
{z
}|
{z
}
|
z
Aw
We consider two scenarios:
(a) Al is an m × n complex Gaussian random matrix, i.e., each entry in Al yields
√i N (0, 1).
2
√1 N (0, 1)
2
+
(b) Al is of the form
Al = Hl Ml ,
(3.2)
where Hl ∈ Cm×n is a random partial Hadamard/Fourier matrix, i.e., the columns of Hl
are uniformly sampled without replacement from an m × m DFT/Hadamard matrix; Ml :=
diag(ml ) = diag(ml,1 , · · · , ml,n ) is a diagonal matrix with {ml,i }ni=1 being i.i.d. Bernoulli
random variables.
Theorem 3.3. Consider the self-calibration model given in (2.2), where Aw is as in (3.1). Let
xmin := min1≤l≤p kxl k and xmax := max1≤l≤p kxl k. Then, for the solution ẑ of (2.8) and α = w∗cz0 ,
there holds
2
kẑ − αz0 k
≤ κ(Aw,0 )η 1 +
kαz0 k
1 − κ(Aw,0 )η
if κ(Aw,0 )η < 1 where η =
κ(Aw,0
and for kwk =
√
2kδAk
√
.
m
v
u
u
)≤t
The condition number of Aw,0 obeys
max{pd2max , x2m }
6x2max (m + kwk2 )
min
,
x2min min{m, kwk2 | Corr(w, z0 )|2 } min{pd2min , x2m }
max
m, there holds
v
u
√
u max{pd2max , x2m }
2 3xmax
min
t
κ(Aw,0 ) ≤
,
xmin | Corr(w, z0 )| min{pd2min , x2m }
max
(a) with probability at least 1 − (np + m)−γ if Al is an m × n (m > n) complex Gaussian random
matrix and
n
1
1
+
(γ + 1) log 2 (np + m) ≤ .
C0
p m
4
10
(b) with probability at least 1 − (np + m)−γ if Al yields (3.2) and
n−1
1
1
+
γ 3 log4 (np + m) ≤ ,
C0
p m−1
4
m ≥ 2.
Remark 3.4. Note that if δA = 0, i.e., in the noiseless case, we have z = αz0 if mp ≥ (np +
m)poly(log(np + m)). Here mp is the number of constraints and np + m is the degree of freedom.
Therefore, our result is nearly optimal in terms of information theoretic limit. Compared with a
similar setup in [1], we have a more efficient algorithm since [1] uses nuclear norm minimization
to achieve exact recovery. However, the assumptions are slightly different, i.e., we assume that D
is invertible and hence the result depends on D while [1] imposes “incoherence” on d by requiring
kF (d)k∞
relatively small, where F denotes Fourier transform.
kdk
3.3
Self-calibration from multiple snapshots
We again follow a by now familiar procedure to derive the linear system associated with (2.3),
which turns out to be
s
diag(y1 ) −A 0 · · ·
0
x1
diag(y2 ) 0 −A · · ·
0
0
x2
..
..
..
..
.
.
=
.
(3.3)
.
.
.
.
.
c (mp+1)×1
..
diag(yp ) 0
.
0 · · · −A
x
w∗
p (m+np)×1
(mp+1)×(m+np)
{z
}|
{z
}
|
z
Aw
For this scenario we only consider the case when A is a complex Gaussian random matrix.
Theorem 3.5. Consider the self-calibration model given in (2.3), where Aw is as in (3.3) and Aw,0
corresponds to the noiseless part of Aw . Let xmin := min1≤l≤p kxl k and xmax := max1≤l≤p kxl k.
Then, for the solution ẑ of (2.8) and α = w∗cz0 , there holds
2
kẑ − αz0 k
≤ κ(Aw,0 )η 1 +
kαz0 k
1 − κ(Aw,0 )η
if κ(Aw,0 )η < 1 where η =
κ(Aw,0
and for kwk =
√
2kδAk
√
.
m
v
u
u
)≤t
Here the upper bound of κ(Aw,0 ) obeys
max{pd2max , x2m }
6x2max (m + kwk2 )
min
.,
x2min min{m, kwk2 | Corr(w, z0 )|2 } min{pd2min , x2m }
max
m, there holds
v
u
√
u max{pd2max , x2m }
2 3xmax
min
t
κ(Aw,0 ) ≤
,
xmin | Corr(w, z0 )| min{pd2min , x2m }
max
with probability at least 1 − 2m(np + m)−γ if A is a complex Gaussian random matrix and
n
1
kGk kGk2F
,
(3.4)
+
log2 (np + m) ≤
C0 max
2
p
p
m
16(γ + 1)
11
where G is the Gram matrix of
n
C0
xl
kxl k
op
l=1
n
1
+
p m
. In particular, if G = Ip and kGkF =
log2 (np + m) ≤
√
p, (3.4) becomes
1
.
16(γ + 1)
Remark 3.6. When kδAk = 0, Theorem 3.5qsays that the solution to (2.3) is uniquely determined
up to a scalar if mp = O(max{np + mkGk, mnp2 + m2 kGk2F }), which involves the norm of the
Gram matrix G. This makes sense if we consider two extreme cases: if {vl }pl=1 are all identical,
then we have kGk = p and kGkF = p and if the {vl }pl=1 are mutually orthogonal, then G = Ip .
Remark 3.7. Balzano and Nowak [6] show exact recovery of this model when A is a deterministic
DFT matrix (discrete Fourier matrix) and {xl }pl=1 are generic signals drawn from a probability
distribution, but their results do not include stability theory in the presence of noise.
Remark 3.8. For Theorem 3.3 and Theorem 3.5 it does not come as a surprise that the error
bound depends on the norm of {xl }pl=1 and D as well as on how much z0 and w are correlated.
We cannot expect a relatively good condition number for Aw if kxl k varies greatly over 1 ≤ l ≤ p.
Concerning the correlation between z0 and w, one extreme case is hz0 , wi = 0, which does not rule
out the possibility of z = 0. Hence, the quantity hz0 , wi affects the condition number.
3.4
Theoretical results for the spectral method
Let S be the matrix Aw excluding the last row and consider S = S0 + δS where S0 is the noiseless
part of S and δS is the noise term. The performance under noise depends on the second smallest
singular value of S0 and the noise strength kδSk.
Theorem 3.9. Denote ẑ as the solution to (2.10), i.e., the right singular vector of S w.r.t. the
smallest singular value and kẑk = 1. Then there holds,
kδSk
(I − ẑ ẑ ∗ )z0
kα0 ẑ − z0 k
≤
=
kz0 k
kz0 k
[σ2 (S0 ) − kδSk]+
α0 ∈C
min
where σ2 (S0 ) is the second smallest singular value of S0 , z0 satisfies S0 z0 = 0, and ẑ is the right
singular vector with respect to the smallest singular value of S, i.e., the solution to (2.10).
Moreover, the lower bound of σ2 (S0 ) satisfies
q
√
(a) σ2 (S0 ) ≥ p2 min{ m, dmin kxk} for model (2.1) under the assumption of Theorem 3.1;
(b) σ2 (S0 ) ≥
√1 xmin min
2
(c) σ2 (S0 ) ≥
√1 xmin min
2
n√
√ o
m
pdmin , xmax
for model (2.2) under the assumption of Theorem 3.3;
n√
√ o
m
for model (2.3) under the assumption of Theorem 3.5.
pdmin , xmax
Remark 3.10. Note that finding z which minimizes (2.10) is equivalent to finding the eigenvector
with respect to the smallest eigenvalue of S ∗ S. If we have a good approximate upper bound λ of
kS ∗ Sk, then it suffices to find the leading eigenvector of λI − S ∗ S, which can be done efficiently by
using power iteration.
How to choose λ properly in practice? We do not want to choose λ too large since this would
imply slow convergence of the power iteration. For each case, it is easy to get a good upper bound
of kSk2 based on the measurements and sensing matrices,
12
(a) for (2.1), λ = k
Pp
∗
l=1 diag(yl ) diag(yl )k
+k
Pp
∗
l=1 Al Al k;
(b) for (2.2), λ = k diag(yl ) diag(yl∗ )k + max1≤l≤p kAl k2 ;
(c) for (2.3), λ = k diag(yl ) diag(yl∗ )k + kAk2 .
Those choices of λ are used in our numerical simulations.
4
Numerical simulations
This section is devoted to numerical simulations. Four experiments for both synthetic and real data
will be presented to address the effectiveness, efficiency and robustness of the proposed approach.
For all three models presented (2.1), (2.2) and (2.3), the corresponding linear systems have
simple block structures which allow for fast implementation via the conjugate gradient method for
non-hermitian matrices [34]. In our simulations, we do not need to set up Aw explicitly to carry
out the matrix-vector multiplications arising in the conjugate gradient method. Moreover, applying
preconditioning via rescaling all the columns to be of similar norms can give rise to an even faster
convergence rate. Therefore, we are able to deal with medium- or large-scale problems from image
processing in a computationally efficient manner.
In our simulations the iteration stops if either the number of iterations reaches at most 2000 or
the residual of the corresponding normal equation is smaller than 10−8 . Throughout our discussion,
the SNR (signal-to-noise ratio) in the scale of dB is defined as
Pp
2
l=1 kyl k
P
.
SNR := 10 log10
p
2
l=1 kεl k
We measure the performance with RelError (in dB) := 20 log 10 RelError where
(
)
kα1 x̂ − x0 k
kα2 dˆ − d0 k
RelError = max min
, min
.
kx0 k
kd0 k
α1 ∈C
α2 ∈C
Here d0 and x0 are the ground truth. Although RelError does not match the error bound in our
theoretic analysis, certain equivalent relations hold if one assumes all |di,0 | are bounded away from
0 because there holds ŝ ≈ s0 if and only if dˆ ≈ d0 .
For the imaging examples, we only measure the relative error with respect to the recovered
x̂−x0 k
.
image x̂, i.e., minα1 ∈C kα1kx
0k
4.1
Self-calibration from repeated measurements
Suppose we have a target image x and try to estimate x through multiple measurements. However,
the sensing process is not perfect because of the missing calibration of the sensors. In order to
estimate both the unknown gains and phases as well as the target signal, a randomized sensing
procedure is used by employing several random binary masks.
We assume that yl = DHMl x0 + εl where H is a “tall” low-frequency DFT matrix, Ml is a
diagonal ±1-random matrix and x0 is an image of size 512×512. We set m = 10242 , n = 5122 , p = 8
2
pm
= 6.4. We
and D = diag(d) ∈ 10242 × 10242 with d ∈ C1024 ; the oversampling ratio is m+n
compare two cases: (i) d is a sequence distributed uniformly over [0.5, 1.5] with w = 1m+n , and (ii)
13
0m
. We
d is a Steinhaus sequence (uniformly distributed over the complex unit circle) with w =
1n
pick those choices of w because we know that the image we try to reconstruct has only non-negative
values. Thus, by choosing w to be non-negative, there are fewer cancellation in the expression w ∗ z0 ,
which in turn leads to a smaller condition number and better robustness. The corresponding results
of our simulations are shown in Figure 1 and Figure 2, respectively. In both cases, we only measure
the relative error of the recovered image.
Figure 1: Here m = 10242 , n = 5122 , p = 8, SNR=5dB, D = diag(d) where di is uniformly
distributed over [0.5, 1.5]. Left: original image; Middle: uncalibrated image, RelError = −13.85dB;
Right: calibrated image, RelError = −20.23dB
Figure 2: Here m = 10242 , n = 5122 , p = 8, SNR=5dB, D = diag(d) where d is a Steinhauss
sequence. Left: original image; Middle: uncalibrated image; Right: calibrated image, RelError =
−10.02dB
In Figure 1, we can see that both, the uncalibrated and the calibrated image are quite good.
Here the uncalibrated image is obtained by first applying the inverse Fourier transform and the
inverse of the mask to each yi and then taking the average of p samples. We explain briefly why
the uncalibrated image still looks good. Note that
x̂uncali
p
p
l=1
l=1
1 X −1 †
1 X −1 †
Ml H (DHMl x0 ) = x0 +
Ml H (D − I)HMl x0
=
p
p
1
H ∗ is the pseudo inverse of H. Here D−I is actually a diagonal matrix with random
where H † = m
1
entries ± 2 . As a result, each Ml−1 H † (D −I)HMl x0 is the sum of m rank-1 matrices with random
± 12 coefficients and is relatively small due to many cancellations. Moreover, [14] showed that most
2-D signals can be reconstructed within a scale factor from only knowing the phase of its Fourier
transform, which applies to the case when d is positive.
However, when the unknown calibration parameters are complex variables (i.e., we do not know
much about the phase information), Figure 2 shows that the uncalibrated recovered image is totally
14
meaningless. Our approach still gives a quite satisfactory result even at a relatively low SNR of
5dB.
4.2
Blind deconvolution in random mask imaging
The second experiment is about blind deconvolution in random mask imaging [5, 37]. Suppose we
observe the convolution of two components,
y l = h ∗ M l x 0 + εl ,
1≤l≤p
where both, the filter h and the signal of interests x0 are unknown. Each Ml is a random ±1-mask.
The blind deconvolution problem is to recover (h, x0 ). Moreover, here we assume that the filter is
actually a low-pass filter, i.e., F(h) is compactly supported in an interval around the origin, where
F is the Fourier transform. After taking the Fourier transform on both sides, the model actually
ends up being of the form (2.1) with Al = HMl where H is a “fat” partial DFT matrix and d is
the nonzero part of F(h). In our experiment, we let x0 be a 128 × 128 image and d = F(h) be a
2-D Gaussian filter of size 45 × 45 as shown in Figure 3.
In those experiments, we choose w = 1m+n since both d and x0 are nonnegative. Figure 4
shows the recovered image from p = 32 sets of noiseless measurements and the performance is quite
pm
≈ 3.52. We can see from Figure 5 that the blurring
satisfactory. Here the oversampling ratio is m+n
effect has been removed while the noise still exists. That is partially because we did not impose any
denoising procedure after the deconvolution. A natural way to improve this reconstruction further
would be to combine the blind deconvolution method with a total variation minimization denoising
step.
4.3
Blind deconvolution via diverse inputs
We choose Al to be random Hadamard matrices with m = 256 and n = 64 and D = diag(d0 )
with d0 being a positive/Steinhaus sequence, as we did previously. Each xl is sampled from
1m
standard Gaussian distribution. We choose w =
if d0 is uniformly distributed over [0.5, 1.5]
0np×1
√
me1
for Steinhaus d0 . 10 simulations are performed for each level of SNR. The
and w =
0np×1
pm
is 2, 2.67 and 3 for
test is also given under different choices of p. The oversampling ratio pn+m
p = 4, 8, 12 respectively. From Figure 6, we can see that the error scales linearly with SNR in
dB. The performance of Steinhaus d0 is not as good as that of positive d0 for SNR ≤ 10 when
∗
we use
squares method. That is because w z0 is quite small when d0 is complex and
√the least
me1
. Note that the error between ẑ and z0 is bounded by κ(Aw,0 )η 1 + 1−κ(A2 w,0 )η .
w =
0np×1
Therefore, the error bound does not depend on kδAk linearly if κ(Aw,0 )η is close to 1. This may
explain the nonlinear behavior of the relative error in the low SNR regime.
We also apply spectral method to this model with complex gains d0 and Al chosen as either a
random Hadamard matrix or a Gaussian matrix. Compared to the linear least squares approach,
the spectral method is much more robust to noise, as suggested in Figure 7, especially in the low
SNR regime.
15
Figure 3: Left: Original image; Right: Gaussian filter in Fourier domain. The support of the filter
is 45 × 45 and hence m = 452 = 2025, n = 1282 , p = 32.
Figure 4: Left: Blurred image without noise, Right: Recovered image, RelError = -45.47dB
Figure 5: Left: Blurred image with SNR = 5dB, Right: Recovered image, RelError = -5.84dB
4.4
Multiple snapshots
We make a comparison of performances between the linear least squares approach and the spectral
method when d0 is a Steinhaus sequence and A is a Gaussian random matrix A. Each xl is sampled
from the standard Gaussian distribution and hence the underlying Gram matrix G is quite close
to Ip (this closeness could be easily made more precise, but we refrain doing so here). The choice
of w and oversampling ratio are the same as those in Section 4.3. From Figure 8, we see that
the performance is not satisfactory for the Steinhaus case using the linear least squares approach,
especially in the lower SNR regime
√ (SNR
≤ 10). The reason is the low correlation between w and
me1
z0 if d0 is Steinhaus and w =
. The difficulty of choosing w is avoided by the spectral
0np×1
method. As we can see in Figure 8, the relative error given by spectral method is approximately
7dB smaller than that given by linear least squares approach when d0 is a complex vector and the
SNR is smaller than 20dB.
16
m = 256, n = 64, d:Positive, LS method, Hadamard
m = 256, n = 64, d:Steinhaus, LS method, Hadamard
0
0
p=4
p=8
p = 12
−20
−30
−40
−50
−60
−70
−80
p=4
p=8
p = 12
−10
Average of sin(θ) of 10 Samples(dB)
Average of sin(θ) of 10 Samples(dB)
−10
−20
−30
−40
−50
−60
−70
0
10
20
30
40
50
60
−80
70
0
10
20
30
SNR(dB)
40
50
60
70
SNR(dB)
Figure 6: Performance of linear least squares approach: RelError (in dB) vs. SNR for yl =
DAl xl + εl , 1 ≤ l ≤ p where m = 256, n = 64, D = diag(d0 ) and each Al is a random Hadamard
matrix. d0 is a Steinhaus sequence.
m = 256, n = 64, d:Steinhaus, spectral method, Gaussian
m = 256, n = 64, d:Steinhaus, spectral method, Hadamard
0
0
p=4
p=8
p = 12
−20
−30
−40
−50
−60
−70
−80
p=4
p=8
p = 12
−10
Average of sin(θ) of 10 Samples(dB)
Average of sin(θ) of 10 Samples(dB)
−10
−20
−30
−40
−50
−60
−70
0
10
20
30
40
50
60
−80
70
0
10
20
30
SNR(dB)
40
50
60
70
SNR(dB)
Figure 7: Performance of spectral method: RelError (in dB) vs. SNR for yl = DAl xl +εl , 1 ≤ l ≤ p
where m = 256, n = 64, D = diag(d0 ). Here Al is either a random Hadamard matrix or a Gaussian
matrix. d0 is a Steinhaus sequence.
5
Proofs
For each subsection, we will first give the result of noiseless measurements. We then prove the
stability theory by using the result below. The proof of spectral method can be found in Section 5.4.
Proposition 5.1. [46] Suppose that Au0 = b is a consistent and overdetermined system. Denote
û as the least squares solution to k(A + δA)u − bk2 with kδAk ≤ ηkAk. If κ(A)η < 1, there holds,
kû − u0 k
2
≤ κ(A)η 1 +
.
ku0 k
1 − κ(A)η
To apply the proposition above, it suffices to bound κ(A) and η.
17
m = 256, n = 64, d:Steinhaus, LS method
m = 256, n = 64, d:Steinhaus, spectral method
0
0
p=4
p=8
p = 12
−20
−30
−40
−50
−60
−70
−80
p=4
p=8
p = 12
−10
Average of sin(θ) of 10 Samples(dB)
Average of sin(θ) of 10 Samples(dB)
−10
−20
−30
−40
−50
−60
−70
0
10
20
30
40
50
60
−80
70
0
10
SNR(dB)
20
30
40
50
60
70
SNR(dB)
Figure 8: Comparison between linear least squares approach and spectral method: RelError (in
dB) vs. SNR for yl = DAxl + εl , 1 ≤ l ≤ p where m = 256, n = 64, D = diag(d0 ) and A is a
Gaussian random matrix. The gain d0 is a random vector with each entry uniformly distributed
over unit circle.
5.1
Self-calibration from repeated measurements
Let us start with (2.7) when εl = 0 and denote Λl := diag(Al v) and v :=
x
kxk
∈ Cn ,
∗
Λ1 − √1m A1
diag(y1 ) −A1
Dkxk √ 0
..
.. = ..
..
A0 :=
.
.
. .
.
0
mIn
1
∗
Λp − √m Ap
diag(yp ) −Ap
Then we rewrite A∗w Aw as
A∗w Aw = A∗0 A0 + ww ∗ =
where
Zl :=
"
Λl
− √1m A∗l
By definition,
#
∈C
(m+n)×m
,
p
X
P Zl Zl∗ P ∗ + ww ∗
l=1
D∗ kxk
0
√
P :=
∈ C(m+n)×(m+n) .
0
mIn
(5.1)
#
∗
√1 Λl Al
−
Λ
Λ
l
l
m
∈ C(m+n)×(m+n) .
Zl Zl∗ =
1
∗A
− √1m A∗l Λ∗l
A
l
m l
"
Our goal is to find out P
the smallest and the largest eigenvalue of Aw . Actually it suffices to under∗
stand the spectrum of " pl=1 Z
# l Zl . Obviously, its smallest eigenvalue is zero and the corresponding
√1
2
1m
√
m
. Let al,i be the i-th column of A∗l and we have E(al,i a∗l,i ) = In under
"
#
√1 1m v ∗
I
−
m
m
all the three settings in Section 3.1. Hence, C := E(Zl Zl∗ ) =
.
− √1m v1∗m
In
It is easy to see that rank(C) = m + n − 1 and the null space of C is spanned by u1 . C has an
eigenvalue with value 1 of multiplicity m + n − 2 and an eigenvalue with value 2 of multiplicity 1.
eigenvector is u1 :=
v
18
More importantly, the following proposition holds and combined with Proposition 5.1, we are able
to prove Theorem 3.1.
Proposition 5.2. There holds
p
X
l=1
(a) with probability 1 − (m +
n)−γ
Zl Zl∗ − pC ≤
p
2
n
if Al is Gaussian and p ≥ c0 γ max 1, m
log2 (m + n);
(b) with probability 1 − (m + n)−γ − 2(mp)−γ+1 if each Al is a “tall” (m × n, m ≥ n) random
Hadamard/DFT matrix and p ≥ c0 γ 2 log(m + n) log(mp);
(c) with probability 1 − (m + n)−γ − 2(mp)−γ+1 if each Al is a “fat” (m × n, m ≤ n) random
Hadamard/DFT matrix and mp ≥ c0 γ 2 n log(m + n) log(mp).
Remark 5.3. Proposition 5.2 actually addresses the identifiability issue of the model (2.1) in
absence of noise. More precisely, the invertibility of P is guaranteed by that
P of D. By Weyl’s
theorem for singular value perturbation in [36], mP
+ n − 1 eigenvalues of pl=1 Zl Zl∗ are greater
than p2 . Hence, the rank of A0 is equal to rank( pl=1 Zl Zl∗ ) = m + n − 1 if p is close to the
n
). In other words, the
information theoretic limit under the conditions given above, i.e., p ≥ O( m
null space of A0 is completely spanned by z0 := (s0 , x0 ).
5.1.1
Proof of Theorem 3.1
Note that Proposition 5.2 gives the result if εl = 0. The noisy counterpart is obtained by applying
perturbation theory for linear least squares.
Let Aw := Aw,0 + δA where Aw,0 is the noiseless part and δA is defined in (2.9).
Note
0
by the
that αz0 with α = w∗cz0 is actually a solution to the overdetermined system Aw,0 z =
c
definition of Aw,0 . Proposition 5.1 implies that it suffices to estimate the condition number κ(Aw,0 )
of Aw,0 and η such that kδAk ≤ ηkAw,0 k holds. Note that
!
p
X
A∗w,0 Aw,0 = P
(5.2)
Zl Zl∗ P ∗ + ww ∗
Proof:
= P
l=1
p
X
Zl Zl∗
l=1
e :=
where w
P −1 w.
ew
e
+w
∗
!
e ∗
P ∗ =: P CP
(5.3)
From Proposition 5.2 and Theorem 1 in [36], we know that
!
!
p
p
X
X
p
5p
Zl Zl∗ ≥ , λn+m
λ2
Zl Zl∗ ≤
2
2
l=1
l=1
Pp
where λ1 ≤ · · · ≤ λn+m and λ1 ( l=1 Zl Zl∗ ) = 0. Following from (5.2), we have
kA∗w,0 Aw,0 k
2
≤ kP k
p
X
l=1
Zl Zl∗ + kwk2 ≤ 3p max{d2max kxk2 , m} + kwk2
kwk2
≤ 3 p+
m
λ2max (P ).
19
(5.4)
On the other hand,
kA∗w,0 Aw,0 k ≥
p
X
l=1
A∗l Al ≥
mp
2
(5.5)
follows from Proposition 5.2. In other words, we have found the lower and upper bounds for
kAw,0 k or equivalently, λmax (A∗w,0 Aw,0 ). Now we proceed to the estimation of λmin (A∗w,0 Aw,0 ).
#
"
1m
√
P
Pm+n
m+n
1
2
m
with m+n
Let u := j=1 αj uj be a unit vector, where u1 = √2
j=1 |αj | = 1 and {uj }j=1
v
P
e defined in (5.3) has a lower
are the eigenvectors of pl=1 Zl Zl∗ . Then the smallest eigenvalue of C
bound as follows:
!
p
X
e
ew
e ∗u
Zl Z ∗ u + u∗ w
u∗ Cu
= u∗
l
l=1
≥
≥
e ≥
which implies λmin (C)
e ∗,
P CP
1
2
m+n
X
j=2
m+n
X
j=2
ew
e ∗ u1 u∗1 u
λj |αj |2 + u∗ u1 u∗1 w
λj |αj |2 +
o
n
∗ z |2
0
≥
min p, |w
mkxk2
1
2
|w ∗ z0 |2
|α1 |2
2mkxk2
(5.6)
n
o
2
2
0 )|
min p, kwk | Corr(w,z
. Combined with A∗w,0 Aw,0 =
m
e 2 (P ) ≥ 1 min mp, kwk2 | Corr(w, z0 )|2 λ2 (P ).
λmin (A∗w,0 Aw,0 ) ≥ λmin (C)λ
min
min
2m
Therefore, with (5.4), the condition number of A∗w,0 Aw,0 is bounded by
κ(A∗w,0 Aw,0 ) ≤
From (5.5), we set η =
2kδAk
√
mp
≥
6(mp + kwk2 )
κ2 (P ).
min{mp, kwk2 | Corr(w, z0 )|2 }
kδAk
kAw,0 k .
Applying Proposition 5.1 gives the following upper bound of the estimation error
κ(Aw,0 )η 1 + 1−κ(A2 w,0 )η where α = w∗cz0 and
κ(Aw,0 ) ≤
If kwk =
√
s
6(mp + kwk2 )
κ(P ).
min{mp, kwk2 | Corr(w, z0 )|2 }
mp, the upper bound of κ(Aw,0 ) satisfies
κ(Aw,0 ) ≤
s
s
√
2 3
12
max{d2max kxk2 , m}
κ(P
)
≤
.
| Corr(w, z0 )|2
|Corr(w, z0 )| min{d2min kxk2 , m}
20
kẑ−αz0 k
kαz0 k
≤
5.1.2
Proof of Proposition 5.2(a)
Proof: [Proof of Proposition 5.2(a)] From now on, we assume al,i ∈ Cn , i.e., the i-th column
of A∗l , obeys a complex Gaussian distribution, √12 N (0, In )+ √i2 N (0, In ). Let zl,i be the i-th column
of Zl ; it can be written in explicit form as
"
#
"
#
| hal,i , vi |2 ei e∗i − √1m ei v ∗ al,i a∗l,i
hal,i , viei
∗
, zl,i zl,i =
zl,i =
1
∗
− √1m al,i
− √1m al,i a∗l,i ve∗i
m al,i al,i .
∗ − E(z z ∗ ), we obtain
Denoting Zl,i := zl,i zl,i
l,i l,i
"
#
(| hal,i , vi |2 − 1)ei e∗i
− √1m ei v ∗ (al,i a∗l,i − In )
Zl,i :=
.
1
∗
− √1m (al,i a∗l,i − In )ve∗i
m (al,i al,i − In )
P
Obviously each Zl,i is independent. In order to apply Theorem 5.12 to estimate k l,i Zl,i k, we
Pp Pm
∗
need to bound maxl,i kZl,i kψ1 and
i=1 E(Zl,i Zl,i ) . Due to the semi-definite positivity of
l=1
∗ , we have kZ k ≤ max{kz k2 , k E(z z ∗ )k} ≤ max | ha , vi |2 + 1 ka k2 , 2 and hence
zl,i zl,i
l,i
l,i
l,i
l,i l,i
l,i
m
n
1
kZl,i kψ1 ≤ (| hal,i , vi |2 )ψ1 + (kal,i k2 )ψ1 ≤ C 1 +
m
m
n
which follows from Lemm 5.14 and k·kψ1 is a norm. This implies R := maxl,i kZl,i kψ1 ≤ C 1 + m
.
P
P
p
m
∗
∗
∗
Now we consider σ02 =
i=1 E(Zl,i Zl,i ) by computing (Zl,i Zl,i )1,1 and (Zl,i Zl,i )2,2 , i.e.,
l=1
∗ ,
the (1, 1)-th and (2, 2)-th block of Zl,i Zl,i
1 ∗
2
2
2
∗
∗
(Zl,i Zl,i )1,1 = (| hal,i , vi | − 1) + v (al,i al,i − In ) v ei e∗i ,
m
1
1
∗
(al,i a∗l,i − In )vv ∗ (al,i a∗l,i − In ) + 2 (al,i a∗l,i − In )2 .
(Zl,i Zl,i
)2,2 =
m
m
Following from (5.36), (5.37), (5.38) and Lemma 5.10, there holds
p X
p X
m
m
∗ )
X
X
0
E(Zl,i Zl,i
1,1
∗
2
E(Zl,i Zl,i ) ≤ 2
σ0 =
∗ )
0
E(Zl,i Zl,i
2,2
l=1 i=1
l=1 i=1
p X
m
n
X
1+ m
ei e∗i
0
= 2
n
1
0
m + m2 In
l=1 i=1
n
n
Im
0
1+ m
.
= 2p 1 +
= 2p
n
0
1 + m In
m
By applying the matrix Bernstein inequality (see Theorem 5.13) we obtain
p X
m
nr
X
n p
≤ C0 max
Zl,i
p 1+
t + log(m + n),
m
l=1 i=1
o p
n
(t + log(m + n)) log(m + n) ≤
1+
m
2
n
log2 (m+
with probability 1−e−t . In particular, by choosing t = γ log(m+n), i.e, p ≥ c0 γ max 1, m
n), the inequality above holds with probability 1 − (m + n)−γ .
21
5.1.3
Proof of Proposition 5.2(b)
Proof: [Proof of Proposition 5.2(b)] Each Zl is independent by its definition in (5.1) if
Al := HMl where H is an m × n partial DFT/Hadamard matrix with m ≥ n and H ∗ H = mIn
and Ml = diag(ml ) is a diagonal random binary ±1 matrix. Let Zl := Zl Zl∗ − C ∈ C(m+n)×(m+n) ;
in explicit form
#
"
− √1m (Λl Al − 1m v ∗ )
Λl Λ∗l − Im
.
Zl =
− √1m (A∗l Λ∗l − v1∗m )
0
where A∗l Al = mIn follows from the assumption. First we take a look at the upper bound of kZl k.
It suffices to bound kZl Zl∗ k since kZl k = kZl Zl∗ − Ck ≤ max {kZl Zl∗ k, kCk} and C is positive
semi-definite. On the other hand, due to Lemma 5.10, we have kZl Zl∗ k ≤ 2 max{kΛl Λ∗l k, 1} and
hence we only need to bound kΛl k. For kΛl k, there holds
max kΛl k = max kAl vk∞ =
1≤l≤p
1≤l≤p
max
1≤l≤p,1≤i≤m
| hal,i , vi |.
Also for any pair of (l, i), hal,i , vi can be rewritten as
| hal,i , vi | = | hdiag(ml )hi , vi | = | ml , diag(h̄i )v |
where hi is the i-th column of H ∗ and k diag(h̄i )vk = kvk = 1. Then there holds
P
max kΛl k ≥
1≤l≤p
p
p X
m
X
p
2γ log(mp)
≤
P | hal,i , vi | ≥ 2γ log(mp)
l=1 i=1
p
≤ mpP | hal,i , vi | ≥ 2γ log(mp)
≤ 2mp · e−γ log(mp) ≤ 2(mp)−γ+1 ,
(5.7)
where the third inequality follows from Lemma 5.11. Applying Lemma 5.10 to Zl ,
R := max kZl k ≤ 2 max {kΛl Λ∗l k, 1} ≤ 4γ log(mp),
1≤l≤p
1≤l≤p
with probability at least 1 − 2(mp)−γ+1
P. Denote the event {max1≤l≤p kZl k ≤ 4γ log(mp)} by E1 .
Now we try to understand σ02 = k pl=1 E(Zl Zl∗ )k. The (1, 1)-th and (2, 2)-th block of Zl Zl∗ are
given by
(Zl Zl∗ )1,1 = (Λl Λ∗l − Im )2 +
(Zl Zl∗ )2,2 =
1
(Λl Al − 1m v ∗ )(Λl Al − 1m v ∗ )∗ ,
m
1
(A∗ Λ∗ − v1∗m )(Λl Al − 1m v ∗ ).
m l l
By using (5.41), (5.42) and Al A∗l = HH ∗ mIm , we have
∗
E((Λl Λ∗l − Im )2 ) =
∗ ∗
E(Λl Al − 1m v )(Λl Al − 1m v )
E(A∗l Λ∗l
−
v1∗m )(ΛAl
=
∗
− 1m v ) =
=
22
E(Λl Λ∗l )2 − Im 2Im ,
E(Λl Al A∗l Λ∗l ) − 1m 1∗m mIm ,
E(A∗l Λ∗l Λl Al ) − mvv ∗
m
X
| hal,i , vi |2 al,i a∗l,i − mvv ∗
i=1
(5.8)
(5.9)
3mIn .
(5.10)
Combining (5.8), (5.9), (5.10) and Lemma 5.10,
p
X
0
E(Zl Zl∗ )1,1
2Im + Im 0
2
≤ 2p
≤ 6p.
σ0 ≤ 2
0
E(Zl Zl∗ )2,2
0
3In
l=1
By applying (5.32) with t = γ log(m + n) and R ≤ 4γ log(mp) over event E1 , we have
p
X
p
√ p
(Zl Zl∗ − C) ≤ C0 max{ p (γ + 1) log(m + n), γ(γ + 1) log(mp) log(m + n)} ≤
2
l=1
with probability 1 − (m + n)−γ − 2(mp)−γ+1 if p ≥ c0 γ 2 log(m + n) log(mp).
5.1.4
Proof of Proposition 5.2(c)
Proof: [Proof of Proposition 5.2(c)] Each Zl is independent due to (5.1). Let Zl := Zl Zl∗ −
C ∈ C(m+n)×(m+n) ; in explicit form
#
"
Λl Λ∗l − Im
− √1m (Λl Al − 1m v ∗ )
.
Zl =
1
∗
− √1m (A∗l Λl − v1∗m )
m Al Al − In
Here Al = HMl where H is a “fat” m × n, (m ≤ n) partial DFT/Hadamard matrix satisfying
HH ∗ = nIm and Ml is a diagonal ±1-random matrix. There holds A∗l Al = Ml∗ H ∗ HMl ∈ Cn×n
where H ∈ Cm×n and E(A∗l Al ) = mIn . For each l, kA∗l Al k = kH ∗ Hk = kHH ∗ k = n. Hence,
there holds,
nn
o
1
kA∗l Al k, kΛl k2 , 1 ≤ 2 max
, γ log(mp)
kZl k ≤ max {kZl Zl∗ k, kCkk ≤ 2 max
m
m
with probability at least 1 − 2(mp)−γ+1 , which
Pfollows exactly from (5.7) and Lemma 5.10.
Now we give an upper bound for σ02 := k pl=1 E(Zl Zl∗ )k. The (1, 1)-th and (2, 2)-th block of
Zl Zl∗ are given by
1
(Λl Al − 1m v ∗ )(ΛAl − 1m v ∗ )∗ ,
m
2
1 ∗
1
(A∗l Λ∗l − v1∗m )(Λl Al − 1m v ∗ ) +
Al Al − In .
m
m
(Zl Zl∗ )1,1 = (Λl Λ∗l − Im )2 +
(Zl Zl∗ )2,2 =
By using (5.41), (5.42) and Al A∗l = HH ∗ = nIm , we have
∗
E((Λl Λ∗l − Im )2 ) =
∗ ∗
E(Λl Al − 1m v )(Λl Al − 1m v )
E(A∗l Λ∗l
−
v1∗m )(Λl Al
=
∗
− 1m v ) =
=
For E
E(Λl Λ∗l )2 − Im 2Im ,
E(Λl Al A∗l Λ∗l ) − 1m 1∗m nIm ,
E(A∗l Λ∗l Λl Al ) − mvv ∗
m
X
| hal,i , vi |2 al,i a∗l,i − mvv ∗
i=1
(5.11)
(5.12)
3mIn . (5.13)
2
− In , we have
2
1 ∗
n − 2m ∗
2
1
Al Al − In = 2 A∗l Al A∗l Al − A∗l Al + In =
Al Al + In
m
m
m
m2
1
∗
m Al Al
23
where Al A∗l = nIm . Note that E(A∗l Al ) = mIn , and there holds,
2
1 ∗
n−m
A Al − In =
In .
E
m l
m
(5.14)
Combining (5.11), (5.12), (5.13), (5.14) and Lemma 5.10,
n
6np
Im
0
2+ m
2
σ0 ≤ 2p
≤
.
n
0
2 + m In
m
By applying (5.32) with t = γ log(m + n), we have
r
p
X
p
np p
n
∗
log(m + n)} ≤
(γ + 1) log(m + n), (γ + 1) γ log(mp) +
(Zl Zl − C) ≤ C0 max{
m
m
2
l=1
with probability 1 − (m + n)−γ − 2(mp)−γ+1 if mp ≥ c0 γ 2 n log(m + n) log(mp).
5.2
Blind deconvolution via diverse inputs
We start with (3.1) by setting εl = 0. In this way, we can factorize the matrix Aw
last row) into
√
pD √ 0
diag(A
A1
1 v1 )
√
√
−
·
·
·
0
mIn
kx1 kIm · · ·
0
p
m
0
kx1 k
.
.
.
.
.
.
.
..
..
..
..
..
..
..
A0 :=
..
...
.
Ap
diag(Ap vp )
0
· · · kxp kIm
√
0
· · · − √m
p
|
{z
}|
0
0
{z
}
Q
{z
|
mp×(np+m)
Z∈C
P
(excluding the
···
···
..
.
···
0
0
..
.
√
mIn
kxp k
}
(5.15)
where vl =
is the normalized xl , 1 ≤ l ≤ p. We will show that the matrix Z is of rank
v1
..
np + m − 1 to guarantee that the solution is unique (up to a scalar). Denote v := . ∈ Cnp×1
vp
P
√
and pl=1 eel ⊗ vl = v with kvk = p where {e
el }pl=1 is a standard orthonormal basis in Rp .
xl
kxl k
5.2.1
Proof of Theorem 3.3
The proof of Theorem 3.3 relies on the following proposition. We defer the proof of Proposition 5.4
to the Sections 5.2.2 and 5.2.3.
Proposition 5.4. There holds,
1
kZ ∗ Z − Ck ≤ ,
2
"
1
Im
− √mp
1m v ∗
C := E(Z ∗ Z) =
1
Inp
v1∗m
− √mp
#
(5.16)
(a) with probability at least 1 − (np + m)−γ with γ ≥ 1 if Al is an m × n (m > n) complex Gaussian
random matrix and
n
1
1
+
(γ + 1) log 2 (np + m) ≤ ;
C0
p m
4
24
(b) with probability at least 1 − (np + m)−γ − 2(mp)−γ+1 with γ ≥ 1 if Al yields (3.2) and
n−1
1
1
+
γ 3 log4 (np + m) ≤ .
C0
p m−1
4
Remark 5.5. Note that C has one eigenvalue equal to 0 and all the other eigenvalues are at
least 1. Hence the rank of C is np + m − 1. Similar to Remark 5.3, Proposition 5.4 shows
that the solution (s0 , {xl }pl=1 ) to (3.1) is uniquely identifiable with high probability when mp ≥
(np + m) · poly(log(np + m)) and kZ ∗ Z − Ck ≤ 12 .
Proof: [Proof of Theorem 3.3] From (3.1), we let Aw = Aw,0 + δA where Aw,0 is the noiseless
part of Aw . By definition of Aw,0 , we know that αz0 is the solution to the overdetermined system
0
Aw,0 z =
where α = w∗cz0 . Now, (5.15) gives
c
A∗w,0 Aw,0 = A∗0 A0 + ww ∗ = P ∗ Z ∗ Q∗ QZP + ww ∗ .
Define xmax := max1≤l≤p kxl k and xmin := min1≤l≤p kxl k. From Proposition 5.4 and Theorem 1
in [36], we know that the eigenvalues {λj }1≤j≤np+m of Z ∗ Z fulfill λ1 = 0 and λj ≥ 12 for j ≥ 2
since kZ ∗ Z − Ck ≤ 12 ; and the eigenvalues of C are 0, 1 and 2 with multiplicities 1, np + m − 2, 1
respectively.
The key is to obtain a bound for κ(Aw,0 ). From (5.15),
λmax (A∗w,0 Aw,0 ) ≤ kP k2 kQk2 kZ ∗ Zk + kwk2 ≤ 3x2max λmax (P ∗ P ) + kwk2
kwk2
2
λmax (P ∗ P )
≤ 3xmax 1 +
m
(5.17)
where x2max λmax (P ∗ P ) ≥ x2max x2m ≥ m. On the other hand, (5.16) gives
min
λmax (A∗w,0 Aw,0 ) ≥ max {kA∗l Al k} ≥
1≤l≤p
since
1
∗
m Al Al
m
2
− In ≤ 12 . For λmin (A∗w,0 Aw,0 ),
e .
A∗w,0 Aw,0 λmin (Q∗ Q)P ∗ Z ∗ ZP + ww ∗ x2min P ∗ Z ∗ ZP + ww ∗ =: P ∗ CP
#
" 1
√ 1m
m
e = x2 Z ∗ Z + w
ew
e ∗ where w
e = P −1 w. By
Denote u1 := √12 √
such that Zu1 = 0 and C
1
min
v
p
using the same procedure as (5.6),
where u :=
Since
Pnp+m
j=1
|u∗1 (P −1 )∗ w|2
e ≥ x2
u∗ Cu
min
αj uj with
=
P
1
∗
2mp |w z0
np+m
X
j=2
λj |αj |2 + |α1 |2 |u∗1 P −1 w|2
1
2
j |αj | = 1 and λj ≥ 2 for
|2 , the smallest eigenvalue of
j ≥ 2 follows from Proposition 5.4.
e satisfies
C
|w ∗ z0 |2
x2min
min 1,
2
mpx2min
1
x2min
2
2
min 1, kwk | Corr(w, z0 )|
≥
2
m
e ≥
λmin (C)
25
∗
2
∗
2
z0 |
z0 |
≥ |w
≥ kwk2 | Corr(w, z0 )|2 follows from kz0 k2 ≥ px2min .
where |w
kz0 k2
px2min
Therefore, the smallest eigenvalue of A∗w,0 Aw,0 satisfies
e min (P ∗ P )
λmin (A∗w,0 Aw,0 ) ≥ λmin (C)λ
x2
≥ min min m, kwk2 | Corr(w, z0 )|2 λmin (P ∗ P )
2m
(5.18)
Combining (5.17) and (5.18) leads to
κ(A∗w,0 Aw,0 ) ≤
Applying Proposition 5.1 and η =
6x2max (m + kwk2 )
κ(P ∗ P ).
x2min min{m, kwk2 | Corr(w, z0 )|2 }
2kδAk
√
m
≥
kδAk
kAw,0 k ,
we have
2
kẑ − αz0 k
≤ κ(Aw,0 )η 1 +
,
kαz0 k
1 − κ(Aw,0 )η
α=
c
w ∗ z0
if κ(Aw,0 )η < 1 where κ(Aw,0 ) obeys
v
u
max{pd2max , x2m }
u
6x2max (m + kwk2 )
min
t
.
κ(Aw,0 ) ≤
x2min min{m, kwk2 | Corr(w, z0 )|2 } min{pd2min , x2m }
max
In particular, if kwk =
√
m, then κ(Aw,0 ) has the following simpler upper bound:
v
u
√
u max{pd2max , x2m }
2 3xmax
min
t
κ(Aw,0 ) ≤
|Corr(w, z0 )|xmin min{pd2min , x2m }
max
which finishes the proof of Theorem 3.3.
5.2.2
Proof of Proposition 5.4(a)
In this section, we will prove that Proposition 5.4(a) if al,i ∼ √12 N (0, In ) + √i2 N (0, In ) where
al,i ∈ Cn is the i-th column of A∗l . Before moving to the proof, we compute a few quantities which
will be used later. Define zl,i as the ((l − 1)m + i)-th column of Z ∗ ,
"
#
√1 hal,i , vl iei
p
zl,i :=
, 1 ≤ l ≤ p, 1 ≤ i ≤ m
− √1m eel ⊗ al,i
(np+m)×1
m
where {ei }m
el }pl=1 ∈ Rp are standard orthonormal P
basis P
in Rm and Rp respectively;
i=1 ∈ R and {e
p
∗
“⊗” denotes Kronecker product. By definition, we have Z ∗ Z = l=1 m
i=1 zl,i zl,i and all zl,i are
independent from one another.
#
"
1
1
2
∗
− √mp
hvl , al,i i ei (e
e∗l ⊗ a∗l,i )
p | hal,i , vl i | ei ei
∗
zl,i zl,i =
1
1
el ee∗l ⊗ al,i a∗l,i
hal,i , vl i (e
el ⊗ al,i )e∗i
− √mp
me
26
and its expectation is equal to
"
∗
)=
E(zl,i zl,i
Pp
It is easy to verify that C =
l=1
1
∗
p ei ei
1
− √mp
(e
el ⊗ vl )e∗i
Pm
∗
i=1 E(zl,i zl,i ).
#
1
− √mp
ei (e
e∗l ⊗ vl∗ )
.
1
el ee∗l ⊗ In
me
Proof: [Proof of Proposition 5.4(a)] The
is to use apply the matrix Bernstein inPpkeyPhere
m
∗ . Let Z := z z ∗ − E(z z ∗ ) and
∗
equality in Theorem 5.13. Note that Z Z = l=1 i=1 zl,i zl,i
l,i l,i
l,i
l,i l,i
we have
1
1 1
1
2
∗
2
2
kZl,i k ≤ kzl,i k + k E(zl,i zl,i )k ≤ | hal,i , vl i | + kal,i k + 2 max
,
p
m
p m
∗ )k ≤ 2 max{ 1 , 1 } follows from Lemma 5.10. Therefore, the exponential norm of
since k E(zl,i zl,i
p m
kZl,i k is bounded by
1
1
n
1
1 1
2
2
kZl,i kψ1 ≤ 2
(| hal,i , vl i | )ψ1 + (kal,i k )ψ1 + 2 max
,
+
≤C
p
m
p m
p m
n
.
which follows from Lemma 5.14 and as a result R := maxl,i kZl,i kψ1 ≤ C p1 + m
P
P
p
m
∗
Now we proceed by estimating the variance σ02 :=
i=1 Zl,i Zl,i . We express Zl,i as
l=1
follows:
#
"
1
1
2
∗
− √mp
ei (e
e∗l ⊗ (vl∗ (al,i a∗l,i − In )))
p (| hal,i , vl i | − 1)ei ei
.
Zl,i =
1
1
el ee∗l ⊗ (al,i a∗l,i − In )
(e
el ⊗ ((al,i a∗l,i − In )vl ))e∗i
− √mp
me
∗ are
The (1, 1)-th and the (2, 2)-th block of Zl,i Zl,i
∗
)1,1 =
(Zl,i Zl,i
and
2
1
1 ∗
| hal,i , vl i |2 − 1 ei e∗i +
v (al,i a∗l,i − In )2 vl ei e∗i ,
2
p
mp l
∗
)2,2 = eel ee∗l ⊗
(Zl,i Zl,i
1
1
(al,i a∗l,i − In )vl vl∗ (al,i a∗l,i − In ) + 2 (al,i a∗l,i − In )2 .
mp
m
Following from (5.36), (5.37) and (5.38), we have
1
1
n
n
∗
∗
∗
∗
E(Zl,i Zl,i )1,1 =
+
In + 2 In .
ei ei , E(Zl,i Zl,i )2,2 = eel eel ⊗
p2 mp
mp
m
Due to Lemma 5.10, there holds,
σ02 :=
p X
m
X
l=1 i=1
=
2
1
p
∗
Zl,i Zl,i
≤2
+
n
m
0
Im
p X
m
X
E(Zl,i Z ∗ )1,1
l,i
l=1 i=1
0
1
p
+
n
m
27
Inp
0
≤2
0
∗ )
E(Zl,i Zl,i
2,2
n
1
+
p m
.
P P
∗
Note that σ02 ≥ k pl=1 m
i=1 E(Zl,i Zl,i )1,1 k =
matrix. By applying (5.34),
∗
kZ Z − Ck =
p X
m
X
l=1 i=1
1
p
+
Zl,i =
nr 1
n
m
∗ ) is a positive semi-definite
since E(Zl,i Zl,i
p X
m
X
l=1 i=1
∗
)−C
E(zl,i zl,i
np
≤ C0 max
+
t + log(np + m),
p m
o 1
1
n
+
(t + log(np + m)) log(np + m) ≤ .
p m
2
∗
With
t = γlog(mp + n), there holds kZ Z − Ck ≤
n
(γ + 1) log2 (np + m) ≤ 14 .
C0 1p + m
5.2.3
1
2
with probability at least 1 − (np + m)−γ if
Proof of Proposition 5.4(b)
We prove Proposition 5.4 based on assumption (3.2). Denote al,i and hl,i as the i-th column of A∗l
and Hl∗ and obviously we have al,i = Ml hl,i . Denote Λl = diag(Al vl ) and let Zl be the l-th block
of Z ∗ in (5.15), i.e.,
#
"
√1 Λl
p
∈ C(np+m)×m .
Zl =
− √1m eel ⊗ A∗l
With A∗l Al = mIn , we have
"
Zl Zl∗ =
1
∗
p Λl Λl
1
eel ⊗ (Λl Al )∗
− √mp
#
1
− √mp
ee∗l ⊗ (Λl Al )
∈ C(np+m)×(np+m)
(e
el ee∗l ) ⊗ In
where the expectation of i-th row of Λl Al yields E(Λl Al )i = E(vl∗ al,i a∗l,i ) = vl∗ . Hence E(Λl Al ) =
1m vl∗ ∈ Cm×n . Its expectation equals
"
#
1
∗ ⊗ (1 v ∗ )
√1 e
e
I
−
m
m
l
p
mp l
E(Zl Zl∗ ) =
.
1
− √mp
eel ⊗ (vl 1∗m )
eel ee∗l ⊗ In
Proof: [Proof of Proposition 5.4(b)] Note that each block Zl is independent and we want
to apply the matrix Bernstein inequality to achieve the desired result. Let Zl := Zl Zl∗ − E(Zl Zl∗ )
and by definition, we have
#
"
1
1
∗
− √mp
(Λl Al − 1m vl∗ )
p (Λl Λl − Im )
kZl k =
1
− √mp
(Λl Al − 1m vl∗ )∗
0
≤
1
1
kΛl Λ∗l − Im k + √ kΛl Al − 1m vl∗ k.
p
mp
Note that
√
√
kΛl Al − 1m vl∗ k ≤ kΛl kkAl k + k1m vl∗ k ≤ mkΛl k + m.
p
Since (5.7) implies that P max1≤l≤p kΛl k ≥ 2γ log(mp) ≤ 2(mp)−γ+1 , we have
p
2γ log(mp) + 1
2γ log(mp) + 1
kΛl k2 + 1 kΛl k + 1
≤
+
+
R := max kZl k ≤
√
√
1≤l≤p
p
p
p
p
28
with probability at least 1 − 2(mp)−γ+1 . We proceed with estimation of σ02 :=
looking at the (1, 1)-th and (2, 2)-th block of Zl Zl∗ , i.e.,
(Zl Zl∗ )1,1 =
(Zl Zl∗ )2,2 =
Pp
∗
l=1 E(Zl Zl )
by
1
1
(Λl Λ∗l − Im )2 +
(Λl Al − 1m vl∗ )(Λl Al − 1m vl∗ )∗ ,
2
p
mp
1
(e
el ee∗l ) ⊗ ((Λl Al − 1m vl∗ )∗ (Λl Al − 1m vl∗ )).
mp
Note that E(Λl Λ∗l − Im )2 = E((Λl Λ∗l )2 ) − Im . The i-th diagonal entry of (Λl Λ∗l )2 is | hal,i , vl i |4 =
| ml , diag(h̄l,i )vl |4 where al,i = Ml hl,i = diag(hl,i )ml and (5.42) implies E(| hal,i , vl i |4 ) ≤ 3
since diag(h̄l,i )vl is still a unit vector (note that diag(h̄l,i ) is unitary since hl,i is a column of Hl∗ ).
Therefore,
E(Λl Λ∗l − Im )2 = E((Λl Λ∗l )2 ) − Im 3Im − Im 2Im .
(5.19)
By using Lemma 5.18, we have
E(Λl Al − 1m vl∗ )(Λl Al − 1m vl∗ )∗
= E(Λl Al A∗l Λ∗l ) − 1m 1∗m
m(n − 1)Im
(n − 1)(mIm − 1m 1∗m )
.
=
m−1
m−1
(5.20)
With al,i = Ml hl,i and independence between {hl,i }m
i=1 and Ml , we have
E((Λl Al − 1m vl∗ )∗ (Λl Al − 1m vl∗ ))
= E(A∗l Λ∗l Λl Al ) − mvl vl∗
=E
m
X
i=1
=E
m
X
i=1
| hal,i , vl i |2 al,i a∗l,i
!
− mvl vl∗
∗
!
diag(hl,i ) | ml , diag(h̄l,i )vl |2 ml ml diag(h̄l,i )
3mIn − mvl vl∗ 3mIn
− mvl vl∗
(5.21)
where E | ml , diag(h̄l,i )vl |2 ml m∗l 3In follows from (5.41) and diag(hl,i ) diag(h̄l,i ) = In . By
using (5.19), (5.20), (5.21) and Lemma 5.10, σ02 is bounded above by
!
p
p
X
X
E(Zl Zl∗ )1,1
0
∗
2
≤2
Zl Zl
σ0 := E
0
E(Zl Zl∗ )2,2
l=1
l=1
#
"
p
2
n−1
X
I
0
+
m
2
(m−1)p
p
≤ 2
3
el ee∗l ) ⊗ In
0
p (e
l=1
"
#
2
n−1
+ m−1
Im
0
1
n−1
p
≤ 2
≤6
+
.
3
p m−1
Inp
0
p
29
n
o
p
Conditioned on the event max1≤l≤p kΛl k ≥ 2γ log(mp) , applying (5.32) with t = γ log(np+m)
gives
p
nr 1
X
n−1p
(γ + 1) log(np + m),
+
≤ C0 max
Zl
p m−1
l=1
s
o 1
2γ log(mp)
log(mp) log(np + m) ≤
(γ + 1)
p
2
−γ
−γ+1 and it suffices to require the condition
with
at least 1 − (np + m) − 2(mp)
probability
n−1
γ 3 log4 (np + m) ≤ 41 .
C0 1p + m−1
5.3
Blind Calibration from multiple snapshots
Recall that Aw in (3.3) and the only difference from (3.1) is that here all Al are
εl = 0, Aw (excluding the last row) can be factorized into
√
pD √ 0
diag(Av
A
1)
√
− √m · · ·
0
mIn
kx1 kIm · · ·
0
0
p
kx1 k
.
.
.
.
.
.
.
..
..
..
..
..
..
..
A0 :=
..
...
.
diag(Av
)
p
A
0
· · · kxp kIm
√
0
· · · − √m
p
|
{z
}|
0
0
{z
}
Q
{z
|
mp×(np+m)
Z∈C
equal to A. If
···
···
..
.
···
0
0
..
.
√
mIn
kxp k
P
}
(5.22)
xi
n is the normalized x .
where vi = kx
∈
C
i
ik
Before we proceed to the main result in this section we need to introduce some notation. Let
ai be the i-th column of A∗ , which is a complex Gaussian random matrix; define Zi ∈ C(np+m)×p
to be a matrix whose columns consist of the i-th column of each block of Z ∗ , i.e.,
#
"
√1 hai , v1 iei · · ·
√1 hai , vp iei
p
p
Zi =
− √1m ee1 ⊗ ai · · · − √1m eep ⊗ ai
(np+m)×p
m
where “⊗” denotes Kronecker product and both {ei }m
el }pl=1 ∈ Rp are the standard
i=1 ∈ R and {e
orthonormal basis in Rm and Rp , respectively. In this way, the Zi are independently from one
another. By definition,
" 1 Pp
#
2 e e∗ − √1 e v ∗ (I ⊗ a a∗ )
|
ha
,
v
i
|
i
i
i
p
i
l
i
i
l=1
p
mp
Zi Zi∗ =
1
1
∗
(Ip ⊗ ai a∗i )ve∗i
− √mp
m Ip ⊗ (ai ai )
where
by
Pp
el
l=1 e
⊗ vl = v ∈ Cnp×1
v1
√
and v = ... with kvk = p. The expectation of Zi Zi∗ is given
vp
#
∗
√1 ei v ∗
e
e
−
i
i
mp
,
E(Zi Zi∗ ) =
1
1
ve∗i
− √mp
m Inp
"
C :=
m
X
i=1
30
"
#
√1 1m v ∗
I
−
m
mp
E(Zi Zi∗ ) =
.
1
v1∗m
Inp
− √mp
Our analysis depends on the mutual coherence of {vl }pl=1 . One cannot expect to recover all
{vl }pl=1 and D if all {vl }pl=1 are parallel to each other. Let G be the Gram matrix of {vl }pl=1 with
p ≤ n, i.e., G ∈ Cp×p and Gk,l = hvk , vl i , 1 ≤ k ≤ l ≤ p and in particular, Gl,l = 1, 1 ≤ l ≤ p. Its
eigenvalues are denoted by {λl }pl=1 with λp ≥ · · · ≥ λ1 ≥ 0. Basic linear algebra tells that
p
X
U GU ∗ = Λ,
λl = p,
(5.23)
l=1
v1∗
where U ∈ Cp×p is unitary and Λ = diag(λ1 , · · · , λp ). Let V = ... ∈ Cp×n , then there holds
v∗
√
√p
∗
∗
∗
Λ = U GU = U V (U V ) since G = V V . Here 1 ≤ kGk ≤ p and p ≤ kGkF ≤ p. In
√
particular, if hvk , vl i = δkl , then G = Ip ; if | hvk , vl i | = 1 for all 1 ≤ k, l ≤ p, then kGk = p and
kGkF = p.
We are now ready to state and prove the main result in this subsection.
Proposition 5.6. There holds
kZ ∗ Z − Ck =
m
X
i=1
Zi Zi∗ − C ≤
1
2
with probability at least 1 − 2m(np + m)−γ if
kGk kGk2F
n
1
C0 max
,
+
log2 (np + m) ≤
2
p
p
m
16(γ + 1)
(5.24)
and each al is i.i.d. complex Gaussian, i.e., ai ∼ √12 N (0, In ) + √i2 N (0, In ). In particular, if
√
G = Ip and kGkF = p, (5.24) becomes
n
1
1
+
.
(5.25)
C0
log2 (np + m) ≤
p m
16(γ + 1)
Remark 5.7. The proof of Theorem 3.5 follows exactly from that of Theorem 3.3 when Proposition 5.6 holds. Hence we just give a proof of Proposition 5.6.
Proof: [Proof of Proposition 5.6] Let Zi Zi∗ − E(Zi Zi∗ ) =: Zi,1 + Zi,2 , where Zi,1 and Zi,2 are
defined as
1 Pp
(| hai , vl i |2 − 1)ei e∗i 0
l=1
p
Zi,1 :=
,
0
0
"
#
1
ei v ∗ (Ip ⊗ (ai a∗i − In ))
0
− √mp
Zi,2 :=
.
1
1
∗
− √mp
(Ip ⊗ (ai a∗i − In ))ve∗i
m Ip ⊗ (ai ai − In )
Estimation of k
p
X
l=1
Pm
i=1 Zi,1 k
Following from (5.23), we have
p
2
2
2
1/2
| hai , vl i | = kV ai k = kU V ai k = kΛ
31
−1/2
Λ
1X
2
λl ξi,l
U V ai k =
2
2
l=1
where Λ−1/2 U V is a p×n matrix with orthonormal
rows and hence each ξi,l is Rayleigh distributed.
√
(We say ξ is Rayleigh distributed if ξ = X 2 + Y 2 where both X and Y are standard real Gaussian
variables.)
Due to the simple form of Zi,1 , it is easy to see from Bernstein’s inequality for scalar random
variables (See Proposition 5.16 in [42]) that
m
X
i=1
p
Zi,1
1X
≤ max
| hai , vl i |2 − 1 ≤ C0 max
1≤i≤m p
l=1
with probability 1 − me−t . Here
Therefore,
p
X
Var
1
p
2
λl (ξi,l
l=1
2
max (λl |ξi,l
1≤l≤p
Pp
2
l=1 | hai , vl i |
!
− 1)
− 1|)ψ1
≤
2
Var(ξi,1
1
p
− 1)
tkGk
,
p
Pp
p
X
l=1
2
l=1 λl (ξi,l
√
tkGkF
p
− 1) where
(5.26)
Pp
l=1 λl
= p.
λ2l = c0 kGk2F ,
1≤l≤p
Pm
∗
i=1 E(Zi,2 Zi,2 )k
in order to bound k
Denote zi := (Ip ⊗ (ai a∗i − I))v and there holds,
Estimation of kZi,2 kψ1
≤ c1 max λl = c1 kGk.
Now we only need to find out kZi,2 kψ1 and σ02 = k
kZi,2 k ≤
−1 =
Pm
i=1 Zi,2 k.
1
1
1
1
max{kai k2 , 1} + √ kzi e∗i k =
max{kai k2 , 1} + √ kzi k.
m
mp
m
mp
For kai k2 , its ψ1 -norm is bounded by C0 n due to Lemma 5.14 and we only need to know kzi kψ1 :
v
u p
uX
√
∗
∗
| hai , vl i |2 kai k + p.
kzi k = k(Ip ⊗ (ai ai − In ))vk ≤ kIp ⊗ (ai ai )vk + kvk = t
l=1
Let u =
Pp
2
l=1 | hai , vl i |
=
√
( uv)ψ1
Pp
2
l=1 λl ξi,l
and v = kai k2 . The ψ1 -norm of kIp ⊗ (ai a∗i )vk satifies
√
= sup q −1 (E( uv)q )1/q ≤ sup q −1 (E uq )1/2q (E v q )1/2q
q≥1
≤
r
q≥1
√
sup q −1 (E uq )1/q sup q −1 (E v q )1/q ≤ C1 np
q≥1
q≥1
where the second inequality follows from the Cauchy-Schwarz inequality, kukψ1 ≤ C1 p, and kvkψ1 ≤
√
C1 n. Therefore, kzi kψ1 ≤ C2 np and there holds
kZi,2 kψ1 ≤ C0
r
√
np
n
n
n
+√
+
.
≤ C0
m
mp
m
m
32
(5.27)
Pm
∗
Estimation of σ02 Note that σ02 :=
i=1 E(Zi,2 Zi,2 ) . Let Ai,1,1 and Ai,2,2 be the (1, 1)-th and
∗ respectively, i.e.,
(2, 2)-th block of Zi,2 Zi,2
Ai,1,1 =
Ai,2,2 =
=
1
k(Ip ⊗ (ai a∗i − In ))vk2 ei e∗i ,
mp
1
1
(Ip ⊗ (ai a∗i − In ))vv ∗ (Ip ⊗ (ai a∗i − In )) + 2 Ip ⊗ (ai a∗i − In )2
mp
m
1
1
[(ai a∗i − In )vk vl∗ (ai a∗i − In )]1≤k≤l≤p + 2 Ip ⊗ (ai a∗i − In )2 .
mp
m
(5.28)
∗ is a positive semi-definite matrix, Lemma 5.10 implies
Since Zi,2 Zi,2
σ02
m
X
Ai,1,1
0
E
.
≤2
0
Ai,2,2
(5.29)
i=1
So we only need to compute E(Ai,1,1 ) and E(Ai,2,2 ).
E k(Ip ⊗
(ai a∗i
2
− In ))vk =
p
X
l=1
E k(ai a∗i
where E(ai a∗i − In )2 = nIn . Now we have E Ai,1,1 =
2
− In )vl k = n
n
∗
m ei ei .
p
X
l=1
kvl k2 = np
For E(Ai,2,2 ), note that
E((ai a∗i − In )vk vl∗ (ai a∗i − In )) = Tr(vk vl∗ )In = hvl , vk i In = Gk,l In
which follows from (5.35). By (5.28), (5.29) and Lemma 5.10, there holds,
n
m n
X
ei e∗i
0
0
n
kGk
2
m
m
≤2
+
.
≤2
σ0 :≤ 2
n
n
1
Inp
0
0 1p G ⊗ In + m
p
m
mp G ⊗ In + m2 Inp
i=1
P
n
∗ ) is positive
each E(Zi,2 Zi,2
One lower bound of σ02 is σ02 ≥ k m
i=1 E(Ai,1,1 )k = m because
√
√
√
≤ log 2C0 √ n
≤ log(2C0 m) where R :=
semi-definite. Therefore, we have log σmR
0
n/m
pn
n
max1≤i≤m kZi,2 kψ1 ≤ CP
0( m +
m ) and m ≥ n.
Applying (5.34) to m
Z
i=1 i,2 with (5.29) and (5.27), we have
m
n n r n
X
+
(t + log(np + m)) log(np + m),
Zi,2
≤ C0 max
m
m
i=1
s
o
kGk
n
+
(log(np + m) + t)
(5.30)
p
m
with
1 − e−t . By combining (5.30) with (5.26) and letting t = γ log(np + m), we have
Pm probability
∗
k i=1 Zi Zi − Ck ≤ 21 with probability 1 − 2m(np + m)−γ if
p
γ log(np + m)kGkF
kGk
1
n
1
≤ , C0 (γ + 1)
+
log2 (np + m) ≤
C0
p
4
p
m
16
n
o
kGk2F
n
1
or equivalently, C0 max kGk
+m
.
log2 (np + m) ≤ 16(γ+1)
p , p2
33
5.4
Proof of the spectral method
We provide the proof of the spectral method proposed in Section 2.3. The proof follows two steps: i)
we provide an error bound for the recovery under noise by using singular value/vector perturbation.
The error bound involves the second smallest singular value of S0 and the noise strength kδSk; ii)
we give a lower bound for σ2 (S0 ), which is achieved with the help of Proposition 5.2, 5.4 and 5.6
respectively for three different models.
The first result is a variant of perturbation theory for the singular vector corresponding to the
smallest singular value. A more general version can be found in [45, 36].
Lemma 5.8. Suppose S = S0 + δS where S0 is the noiseless part of S and δS is the noisy part.
Then
kδSk
kα0 ẑ − z0 k
(I − ẑ ẑ ∗ )z0
≤
=
.
min
kz
k
kz
k
[σ
(S
α0 ∈C
0
0
2 0 ) − kδSk]+
where σ2 (S0 ) is the second smallest singular value of S0 , z0 satisfies S0 z0 = 0, and ẑ is the right
singular vector with respect to the smallest singular value of S, i.e., the solution to (2.10).
s
Proof: By definition, there holds S0 z0 = 0 where z0 = 0 is the ground truth and also the
x0
right singular vector corresponding to the smallest singular value. Without loss of generality, we
assume kz0 k = kẑk = 1. For S, we denote its singular value decomposition as
S := S − σ1 (S)ûẑ ∗ + σ1 (S)ûẑ ∗
{z
} | {z }
|
B1
B0
where σ1 (S), û and ẑ are the smallest singular value/vectors of S. By definition, the vector ẑ is
also the solution to (2.10).
First note that I − ẑ ẑ ∗ = B1∗ (B1∗ )† where (B1∗ )† is the pseudo-inverse of B1 . Therefore, we
have
(I − ẑ ẑ ∗ )z0
kz0 k
= kz0 z0∗ B1∗ (B1∗ )† k = kz0 z0∗ (S0∗ + (δS)∗ − (B0 )∗ )(B1∗ )† k
= kz0 z0∗ (δS)∗ (B1∗ )† k ≤ kδSkk(B1∗ )† k ≤
kδAk
kδSk
≤
σ2 (S)
[σ2 (S0 ) − kδSk]+
where S0 z0 = 0 and (B0 )∗ (B1∗ )† = 0. And the last inequality follows from |σ2 (S0 ) − σ2 (S)| ≤ kδSk
and σ2 (S) ≥ [σ2 (S0 ) − kδSk]+ .
The second smallest singular value σ2 (S0 ) is estimated by using the proposition 5.2, 5.4 and 5.6
and the following fact:
Lemma 5.9. Suppose P is an invertible matrix, and A is a positive semi-definite matrix with the
dimension of its null space equal to 1. Then the second smallest singular value of P AP ∗ is nonzero
and satisfies
σ2 (P AP ∗ ) ≥ σ2 (A)σ12 (P )
where σ1 (·) and σ2 (·) denotes the smallest and second smallest singular values respectively.
34
Proof: The proof is very straightforward. Note that k(P AP ∗ )† k =
deficient. Also from the property of pseudo inverse, there holds
k(P AP ∗ )† k = k(P −1 )∗ A† P −1 k ≤ kA† kkP −1 k2 =
1
σ2 (P AP ∗ )
since A is rank-1
1
1
.
2
σ2 (A) σ1 (P )
Hence, σ2 (P AP ∗ ) ≥ σ2 (A)σ12 (P ).
Proof: [Proof of Theorem 3.9] Combined with Lemma 5.8, it suffices to estimate the lower
bound of the second smallest singular P
value of S0 for the proposed three models. We start with
∗
(a). From (5.1), we know that S0 S0 = pl=1 P Zl Zl∗ P ∗ where
"
#
∗
Λl
D kxk
0
(m+n)×m
√
∈ C(m+n)×(m+n) .
Zl :=
∈
C
,
P
:=
− √1m A∗l
mIn
0
P
From Proposition 5.2, we know that the second smallest eigenvalue of pl=1 Zl Zl∗ is at least
it is also rank-1 deficient. Applying Lemma 5.9 gives
!
p
X
√
p
Zl Zl∗ σ1 (P P ∗ ) ≥ (min{ m, dmin kxk})2
σ2 (S0∗ S0 ) ≥ σ2
2
p
2
and
l=1
and hence σ2 (S0 ) ≥
q
p
2
√
min{ m, dmin kxk}.
Since (b) and (c) are exactly the same, it suffices to show (b). From (5.15), we know that S0
can be factorized into A0 := QZP and Proposition 5.4 implies
"
#
√1 1m v ∗
I
−
m
1
mp
.
kZ ∗ Z − Ck ≤ , C := E(Z ∗ Z) =
1
− √mp
v1∗m
Inp
2
Therefore, S0∗ S0 = QZ ∗ P P ∗ ZQ∗ and σ2 (Z ∗ Z) ≥ 12 . Applying Lemma 5.9 leads to
σ2 (S0∗ S0 ) ≥ σ2 (Z ∗ P P ∗ Z)σ12 (Q) ≥ σ2 (Z ∗ Z)σ12 (P )σ12 (Q)
√ 2
√
m
1 2
pdmin ,
≥ xmin min
2
xmax
where xmin = min{kxl k} and xmax = max{kxl k}.
Appendix
S11 S12
, there
Lemma 5.10. For any Hermitian positive semi-definite matrix S with S :=
∗
S12
S22
holds,
S11 0
S11 S12
2
.
∗
S22
S12
0 S22
In other words, kSk ≤ 2 max{kS11 k, kS22 k}.
35
Lemma 5.11. Corollary 7.21 in [17]. Let a ∈ CM and ε = (ε1 , · · · , εM ) be a Rademacher sequence,
then for u > 0,
M
X
P
εj aj ≥ kaku ≤ 2 exp(−u2 /2).
j=1
For Gaussian and random Hadamard cases, the concentration inequalities are slightly different.
The following theorem is mostly due to Theorem 6.1 in [39] as well as due to [42].
Theorem 5.12. Consider a finite sequence of Zl of independent centered random matrices
PL with
dimension M1 × M2 . We assume that kZl k ≤ R and introduce the random matrix S := l=1 Zl .
Compute the variance parameter
σ02
L
L
n X
o
X
∗
= max k
E(Zl Zl )k, k
E(Zl∗ Zl )k ,
l=1
then for all t ≥ 0
kSk ≤ C0 max{σ0
p
(5.31)
l=1
t + log(M1 + M2 ), R(t + log(M1 + M2 ))}
(5.32)
with probability at least 1 − e−t where C0 is an absolute constant.
The concentration inequality is slightly different from Theorem 5.12 if kZl k is a sub-exponential
random variable. Here we are using the form in [24]. Before presenting the inequality, we introduce
the sub-exponential norm k · kψ1 of a matrix, defined as
kZkψ1 := inf {u : E[exp(kZk/u)] ≤ 2}.
u≥0
(5.33)
One can find more details about this norm and norm on Orlicz spaces in [42] and [40].
Theorem 5.13. For a finite sequence of independent M1 × M2 random matrices Zl with R :=
max1≤l≤L kZl kψ1 and σ02 as defined in (5.31), we have the tail bound on the operator norm of S,
!
√
p
LR
(t + log(M1 + M2 ))}
(5.34)
kSk ≤ C0 max{σ0 t + log(M1 + M2 ), R log
σ0
with probability at least 1 − e−t where C0 is an absolute constant.
The estimation of the ψ1 -norm of a sub-exponential random variable easily follows from the
following lemma.
Lemma 5.14 (Lemma 2.2.1 in [40]). Let z be a random variable which obeys P{|z| > u} ≤ ae−bu ,
then kzkψ1 ≤ (1 + a)/b.
Remark 5.15. A direct implication of Lemma 5.14 gives the ψ1 -norm of kak2 where a ∼ N (0, 12 In )+
iN (0, 21 In ) is a complex Gaussian random vector, i.e., (kak2 )ψ1 ≤ C0 n for the absolute constant
C0 .
36
Lemma 5.16. For a ∼ N (0, 12 In ) + iN (0, 21 In ), there holds
E(kak2 aa∗ ) =
∗
∗
E((a Xa)aa ) =
(n + 1)In ,
X + Tr(X)In ,
(5.35)
for any fixed x ∈ Cn and X ∈ Cn×n . In particular, we have
E | ha, xi |2 aa∗
(5.36)
4
= kxk2 In + xx∗ ,
= 2kxk ,
(5.37)
2
= nIn .
(5.38)
E | ha, xi |
∗
E(aa − In )
4
Lemma 5.17. Suppose that a ∈ Rn is a Rademacher sequence and for any fixed x ∈ Cn and
X ∈ Cn×n , there holds
E(kak2 aa∗ )
E((a∗ Xa)aa∗ )
= nIn ,
(5.39)
= Tr(X)In + X + X T − 2
n
X
Xkk Ekk
(5.40)
k=1
where Ekk is an n × n matrix with only one nonzero entry equal to 1 and at position (k, k). In
particular, setting X = xx∗ gives
E | ha, xi |2 aa∗
E | ha, xi |4
= kxk2 In + xx∗ + xx∗ − 2 diag(x) diag(x̄) 3kxk2 In ,
= 2kxk4 +
n
X
2
x2i
k=1
−2
n
X
k=1
|xk |4 ≤ 3kxk4 .
(5.41)
(5.42)
Proof: Since a is a Rademacher sequence, i.e, each ai takes ±1 independently with equal probability, this implies a2i =P1 and
kak2 = n. Therefore, E(kak2 aa∗ ) = n E(aa∗ ) = nIn . The (k, l)-th
P
entry of (a∗ Xa)aa∗ is ni=1 nj=1 Xij ai aj ak al .
1. If k = l,
n
n
n X
X
X
2
E(Xii |ai |2 |ak |2 ) = Tr(X)
Xij ai aj |ak |
=
E
i=1
i=1 j=1
where E(ai aj |ak |2 ) = 0 if i 6= j.
2. If k 6= l,
n X
n
X
Xij ai aj ak al = Xkl E(|ak |2 |al |2 ) + Xlk E(|ak |2 |al |2 ) = Xkl + Xlk .
E
i=1 j=1
Hence, we have E((a∗ Xa)aa∗ ) = Tr(X)In + X + X T − 2
Lemma 5.18. There holds
E(ΛA(ΛA)∗ ) =
Pn
k=1 Xkk Ekk .
(n − 1)(mIm − 1m 1∗m )
+ 1m 1∗m
m−1
37
where A = HM , Λ = diag(Av) and v = (v1 , · · · , vn )T ∈ Cn is a deterministic unit vector.
H ∈ Cm×n is a random partial Fourier/Hadamard matrix with H ∗ H = mIn and m ≥ n, i.e., the
columns of H are uniformly sampled without replacement from an m × m DFT/Hadamard matrix;
M is a diagonal matrix with entries random sampled from ±1 with equal probability; moreover, we
assume M and H are independent from each other. In particular, if m = n = 1, E(ΛA(ΛA)∗ ) = 1.
Proof: We only prove the case when A is a random Fourier matrix since the Hadamard case is
essentially the same modulo very minor differences. By definition,
ΛAA∗ Λ∗ = ΛHH ∗ Λ∗ = diag(HM v)HH ∗ diag(HM v).
Let hi be the i-th column of H ∗ and the (i, j)-th entry of diag(HM v)HH ∗ diag(HM v) is
hhi , M vi hhj , M vi hhi , hj i where hu, vi = u∗ v. The randomness of ΛAA∗ Λ∗ comes from both H
and M and we first take the expectation with respect to M .
E(hhi , M vi hhj , M vi |H)
= h∗i E(M vv ∗ M |H)hj = h∗i diag(v) diag(v̄)hj
= (H diag(v) diag(v̄)H ∗ )i,j
where E(M vv ∗ M ) = diag(v) diag(v̄) follows from each entry in M being a Bernoulli random
variable. Hence, E((ΛAA∗ Λ∗ )i,j |H) = (H diag(v) diag(v̄)H ∗ )i,j · (HH ∗ )i,j .
Let uk be the k-th columnPof H and 1 ≤ k ≤ n and “⊙” denotes the Hadamard (pointwise)
product. So we have HH ∗ = nk=1 uk u∗k . There holds,
!
n
X
|vk |2 ūk ū∗k ⊙ HH ∗
E(ΛAA∗ Λ∗ |H) = (H diag(v) diag(v̄)H ∗ ) ⊙ HH ∗ =
k=1
=
n
X
k=1
=
|vk |2 diag(ūk )HH ∗ diag(uk ) =
X
1≤k6=l≤n
2
|vk |
n
n X
X
k=1 l=1
diag(ūk )ul u∗l diag(uk ) +
|vk |2 diag(ūk )ul u∗l diag(uk )
1m 1∗m
(5.43)
where the third equation follows from linearity of the Hadamard product and from
ūk ū∗k ⊙ HH ∗ = diag(ūk )HH ∗ diag(uk ).
The last one uses the fact that diag(ūk )uk = 1m if uk is a vector from the DFT matrix or
Hadamard matrix. By the property of conditional expectation, we know that E(ΛAA∗ Λ∗ ) =
E(E(ΛAA∗ Λ∗ )|H) and due to the linearity of expectation, it suffices to find out for k 6= l,
E(diag(ūk )ul u∗l diag(uk )) where uk and ul , by definition, are the k-th and l-th columns of H
which are sampled uniformly without replacement from an m × m DFT matrix F . Note that
(uk , ul ) is actually an ordered pair of random vectors sampled without replacement from columns
of F . Hence there are in total m(m − 1) different choices of diag(ūk )ul and
P(uk = φi , ul = φj ) =
1
,
m(m − 1)
38
∀1 ≤ i 6= j ≤ m, ∀1 ≤ k 6= l ≤ n
where φi is defined as the i-th column of an m × m DFT matrix F . Now we have, for any k 6= l,
E(diag(ūk )ul u∗l diag(uk )) =
=
=
X
1
diag(φ̄i )φj φ∗j diag(φi )
m(m − 1)
i6=j
X
1
diag(φ̄i )φj φ∗j diag(φi ) − m1m 1∗m
m(m − 1)
1≤i,j≤m
!
m
X
1
∗
diag(φ̄i ) diag(φi ) − 1m 1m
m−1
i=1
=
mIm − 1m 1∗m
.
m−1
(5.44)
P
∗
∗ ∗
where diag(φ̄i )φi = 1m and m
i=1 φi φi = mIm . Now we return to E(ΛAA Λ ). By substituting (5.44) into (5.43), we get the desired formula:
X
E(ΛAA∗ Λ∗ ) = E(E(ΛAA∗ Λ∗ |H)) = E
|vk |2 diag(uk )ul u∗l diag(uk ) + 1m 1∗m
1≤k6=l≤n
=
where
P
2
1≤k6=l≤n |vk |
1m 1∗m
mIm −
m−1
=n−
Pn
X
1≤k6=l≤n
2
k=1 |vk |
|vk |2 + 1m 1∗m =
= n − 1 follows from
(n − 1)(mIm − 1m 1∗m )
+ 1m 1∗m ,
m−1
Pm
2
k=1 |vk |
= 1.
References
[1] A. Ahmed, A. Cosse, and L. Demanet, A convex approach to blind deconvolution with
diverse inputs, in 2015 IEEE 6th International Workshop on Computational Advances in MultiSensor Adaptive Processing (CAMSAP), IEEE, 2015, pp. 5–8.
[2] A. Ahmed and L. Demanet, Leveraging diversity and sparsity in blind deconvolution, arXiv
preprint arXiv:1610.06098, (2016).
[3] A. Ahmed, F. Krahmer, and J. Romberg, Empirical chaos processes and blind deconvolution, arXiv preprint arXiv:1608.08370, (2016).
[4] A. Ahmed, B. Recht, and J. Romberg, Blind deconvolution using convex programming,
IEEE Transactions on Information Theory, 60 (2014), pp. 1711–1732.
[5] S. Bahmani and J. Romberg, Lifting for blind deconvolution in random mask imaging:
Identifiability and convex relaxation, SIAM Journal on Imaging Sciences, 8 (2015), pp. 2203–
2238.
[6] L. Balzano and R. Nowak, Blind calibration of sensor networks, in Proceedings of the 6th
International Conference on Information processing in sensor networks, ACM, 2007, pp. 79–88.
[7] L. Balzano and R. Nowak, Blind calibration of networks of sensors: Theory and algorithms,
in Networked Sensing Information and Control, Springer, 2008, pp. 9–37.
39
[8] C. Bilen, G. Puy, R. Gribonval, and L. Daudet, Convex optimization approaches for
blind sensor calibration using sparsity, IEEE Transactions on Signal Processing, 62 (2014),
pp. 4847–4856.
[9] V. Cambareri and L. Jacques, A greedy blind calibration method for compressed sensing
with unknown sensor gains, arXiv preprint arXiv:1610.02851, (2016).
[10] V. Cambareri and L. Jacques, A non-convex blind calibration method for randomised
sensing strategies, in Compressed Sensing Theory and its Applications to Radar, Sonar and
Remote Sensing (CoSeRa), 2016 4th International Workshop on, IEEE, 2016, pp. 16–20.
[11] V. Cambareri and L. Jacques, Through the haze: A non-convex approach to blind calibration for linear random sensing models, arXiv preprint arXiv:1610.09028, (2016).
[12] P. Campisi and K. Egiazarian, Blind Image Deconvolution: Theory and Applications, CRC
press, 2007.
[13] E. J. Candes, Y. C. Eldar, T. Strohmer, and V. Voroninski, Phase retrieval via
matrix completion, SIAM Review, 57 (2015), pp. 225–251.
[14] S. Curtis, J. Lim, and A. Oppenheim, Signal reconstruction from one bit of fourier transform phase, in 1984 IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), vol. 9, IEEE, 1984, pp. 487–490.
[15] M. A. Davenport and J. Romberg, An overview of low-rank matrix recovery from incomplete observations, IEEE Journal of Selected Topics in Signal Processing, 10 (2016), pp. 608–
622.
[16] J. R. Fienup, Reconstruction of an object from the modulus of its fourier transform, Optics
Letters, 3 (1978), pp. 27–29.
[17] S. Foucart and H. Rauhut, A Mathematical Introduction to Compressive Sensing, Springer,
2013.
[18] B. Friedlander and T. Strohmer, Bilinear compressed sensing for array self-calibration,
in 2014 48th Asilomar Conference on Signals, Systems and Computers, Asilomar, 2014.
[19] L. Gan, T. T. Do, and T. D. Tran, Fast compressive imaging using scrambled block
hadamard ensemble, in 2008 16th European Signal Processing Conference, IEEE, 2008, pp. 1–
5.
[20] R. Gribonval, G. Chardon, and L. Daudet, Blind calibration for compressed sensing by
convex optimization, in 2012 IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), IEEE, 2012, pp. 2713–2716.
[21] G. Harikumar and Y. Bresler, Perfect blind restoration of images blurred by multiple
filters: Theory and efficient algorithms, IEEE Transactions on Image Processing, 8 (1999),
pp. 202–219.
40
[22] A. Ito, A. C. Sankaranarayanan, A. Veeraraghavan, and R. G. Baraniuk, Blurburst: Removing blur due to camera shake using multiple images, ACM Trans. Graph., Submitted, 3 (2014).
[23] M. Kech and F. Krahmer, Optimal injectivity conditions for bilinear inverse problems with
applications to identifiability of deconvolution problems, SIAM Journal on Applied Algebra and
Geometry, 1 (2017), pp. 20–37.
[24] V. Koltchinskii et al., Von Neumann entropy penalization and low-rank matrix estimation,
The Annals of Statistics, 39 (2011), pp. 2936–2973.
[25] K. Lee, Y. Li, M. Junge, and Y. Bresler, Blind recovery of sparse signals from subsampled
convolution, IEEE Transactions on Information Theory, 63 (2017), pp. 802–821.
[26] X. Li, S. Ling, T. Strohmer, and K. Wei, Rapid, robust, and reliable blind deconvolution
via nonconvex optimization, arXiv preprint arXiv:1606.04933, (2016).
[27] Y. Li, K. Lee, and Y. Bresler, Identifiability in blind deconvolution with subspace or
sparsity constraints, IEEE Transactions on Information Theory, 62 (2016), pp. 4266–4275.
[28] Y. Li, K. Lee, and Y. Bresler, Optimal sample complexity for blind gain and phase calibration, IEEE Transactions on Signal Processing, 64 (2016), pp. 5549–5556.
[29] S. Ling and T. Strohmer, Blind deconvolution meets blind demixing: Algorithms and performance bounds, arXiv preprint arXiv:1512.07730, (2015).
[30] S. Ling and T. Strohmer, Self-calibration and biconvex compressive sensing, Inverse Problems, 31 (2015), p. 115002.
[31] J. Lipor and L. Balzano, Robust blind calibration via total least squares, in 2014 IEEE
International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE, 2014,
pp. 4244–4248.
[32] V. I. Morgenshtern, E. Riegler, W. Yang, G. Durisi, S. Lin, B. Sturmfels, and H. Bölcskei, Capacity pre-log of noncoherent SIMO channels via Hironaka’s Theorem, IEEE Transactions on Information Theory, 59 (2013), pp. 4213–4229,
http://www.nari.ee.ethz.ch/commth/pubs/p/mrydlsb12.
[33] M. Pollefeys, R. Koch, and L. Van Gool, Self-calibration and metric reconstruction inspite of varying and unknown intrinsic camera parameters, International Journal of Computer
Vision, 32 (1999), pp. 7–25.
[34] Y. Saad, Iterative Methods for Sparse Linear Systems, SIAM, 2003.
[35] P. J. Shin, P. E. Larson, M. A. Ohliger, M. Elad, J. M. Pauly, D. B. Vigneron,
and M. Lustig, Calibrationless parallel imaging reconstruction based on structured low-rank
matrix completion, Magnetic Resonance in Medicine, 72 (2014), pp. 959–970.
[36] G. W. Stewart, Perturbation theory for the singular value decomposition, Technical Report
CS-TR 2539, University of Maryland, (September 1990).
41
[37] G. Tang and B. Recht, Convex blind deconvolution with random masks, in Computational
Optical Sensing and Imaging, Optical Society of America, 2014, pp. CW4C–1.
[38] L. Tong, G. Xu, B. Hassibi, and T. Kailath, Blind identification and equalization based
on second-order statistics: A frequency domain approach, IEEE Transactions on Information
Theory, 41 (1995), pp. 329–334.
[39] J. A. Tropp, User-friendly tail bounds for sums of random matrices, Foundations of Computational Mathematics, 12 (2012), pp. 389–434.
[40] A. Van der Vaart and J. Wellner, Weak Convergence and Empirical Processes: with
Applications to Statistics, Springer Series in Statistics, Springer-Verlag, New York, 1996.
[41] S. Verdú and S. Shamai, Spectral efficiency of cdma with random spreading, IEEE Transactions on Information theory, 45 (1999), pp. 622–640.
[42] R. Vershynin, Introduction to the non-asymptotic analysis of random matrices, in Compressed Sensing: Theory and Applications, Y. C. Eldar and G. Kutyniok, eds., Cambridge
University Press, 2012, ch. 5.
[43] L. Wang and Y. Chi, Blind deconvolution from multiple sparse inputs, IEEE Signal Processing Letters, 23 (2016), pp. 1384–1388.
[44] L. Wang, A. Singer, and Z. Wen, Orientation determination of cryo-em images using
least unsquared deviations, SIAM journal on imaging sciences, 6 (2013), pp. 2450–2483.
[45] P.-Å. Wedin, Perturbation bounds in connection with singular value decomposition, BIT Numerical Mathematics, 12 (1972), pp. 99–111.
[46] M. Wei, The perturbation of consistent least squares problems, Linear Algebra and its Applications, 112 (1989), pp. 231–245.
[47] A. J. Weiss and B. Friedlander, Eigenstructure methods for direction finding with sensor
gain and phase uncertainties, Circuits, Systems, and Signal Processing, 9 (1990), pp. 271–300.
42
| 7cs.IT
|
arXiv:1712.04544v1 [cs.CR] 12 Dec 2017
HIERARCHICAL BLOOM FILTER TREES FOR
APPROXIMATE MATCHING
Frank Breitinger
Cyber Forensics Research
and Education Group
Tagliatela College of Engineering
University of New Haven
[email protected]
Mark Scanlon
Forensics and Security Research Group
School of Computer Science
University College Dublin
[email protected]
David Lillis
Forensics and Security Research Group
School of Computer Science
University College Dublin
[email protected]
ABSTRACT
Bytewise approximate matching algorithms have in recent years shown significant promise in detecting files that are similar at the byte level. This is very useful for digital forensic investigators,
who are regularly faced with the problem of searching through a seized device for pertinent data.
A common scenario is where an investigator is in possession of a collection of “known-illegal” files
(e.g. a collection of child abuse material) and wishes to find whether copies of these are stored on
the seized device. Approximate matching addresses shortcomings in traditional hashing, which can
only find identical files, by also being able to deal with cases of merged files, embedded files, partial
files, or if a file has been changed in any way.
Most approximate matching algorithms work by comparing pairs of files, which is not a scalable
approach when faced with large corpora. This paper demonstrates the effectiveness of using a “Hierarchical Bloom Filter Tree” (HBFT) data structure to reduce the running time of collection-againstcollection matching, with a specific focus on the MRSH-v2 algorithm. Three experiments are discussed,
which explore the effects of different configurations of HBFTs. The proposed approach dramatically
reduces the number of pairwise comparisons required, and demonstrates substantial speed gains,
while maintaining effectiveness.
Keywords: approximate matching, hierarchical bloom filter trees, mrsh-v2
1.
INTRODUCTION
ment agencies worldwide, this has resulted in
significant evidence backlogs becoming commonplace (Scanlon, 2016), frequently reaching 1824 months (Casey et al., 2009) and exceeding
4 years in extreme cases (Lillis et al., 2016). The
backlogs have grown due to a number of factors
including the volume of cases requiring analysis,
the number of devices per case, the volume of
data on each device, and the limited availability
Current digital forensic process models are
surprisingly arduous, inefficient, and expensive. Coupled with the sheer volume of digital forensic investigations facing law enforceThis paper is an extended version of Lillis et al.
(2017), which was presented at the 9th EAI International Conference on Digital Forensics and Cyber Crime
(ICDF2C), Prague, Czech Republic, 9-11 October, 2017.
1
This paper presents an improvement in the
runtime efficiency of approximate matching techniques, primarily through the implementation of
a Hierarchical Bloom Filter Tree (HBFT). Additionally, it examines some of the tunable parameters of the algorithm to gauge their effect on the
required running time. A number of experiments
were conducted, using two different formulations
of a HBFT, which indicated a substantial reduction in the running time, in addition to which
the final experiment achieved a 100% recall rate
for identical files and also for files that have a
MRSH-v2 similarity above a reasonable threshold
of 40%.
Section 2 outlines the prior work that has been
conducted in the area of approximate matching.
The operation of MRSH-v2 is discussed in Section 3. HBFTs are introduced in Section 4. Section 5 presents the series of experiments designed
to evaluate the effectiveness of the HBFT approach, and finally Section 6 concludes the paper
and outlines directions for further work.
of skilled experts (Quick & Choo, 2014). Automated techniques are in continuous development
to aid investigators, but due to the sensitive nature of this work, the ultimate inferences and
decisions will always be made by skilled human
experts (James & Gladyshev, 2015).
Perhaps the most common (and most timeconsuming) task facing digital investigators involves examination of seized suspect devices
to determine if pertinent evidence is contained
therein. Often, this examination requires significant manual, expert data processing and analysis during the acquisition and analysis phases
of an investigation. A number of techniques
have been created or are in development to expedite/automate parts of the typical digital forensic process. These include triage (Rogers et
al., 2006), distributed processing (Roussev &
Richard III, 2004), Digital Forensics as a Service (DFaaS) (van Baar et al., 2014), workflow
management and automation (de Braekt et al.,
2016; J. N. Gupta et al., 2016). While these
techniques can help to alleviate the backlog, the
premise behind many of them involves evidence
discovery based on exact matching of hash values (e.g., MD5, SHA1). Typically, this requires
a set of hashes of known incriminating/pertinent
content. The hash of each artefact from a suspect device is then compared against this set.
This approach falls short against basic counterforensic techniques (e.g., content editing, content
embedding, data transformation).
Approximate matching (also referred to
as “fuzzy hashing”) is one technique used
to aid the discovery of these obfuscated
files (Breitinger, Guttman, et al., 2014). A number of algorithms have been developed including ssdeep (Kornblum, 2006), sdhash (Roussev,
2010), and MRSH-v2 (Breitinger & Baier, 2012).
This paper focuses specifically on MRSH-v2. This
algorithm operates by generating a “similarity
digest” for each file, represented as Bloom filters
(Bloom, 1970). An all-against-all pairwise comparison is then required to determine if files from
a set of desired content is present in a corpus of
unanalysed digital material. Thus, MRSH-v2 does
not exhibit strong scalability for use with larger
datasets.
2. BACKGROUND:
APPROXIMATE MATCHING
Bytewise approximate matching for digital forensics gained popularity in 2006 when Kornblum
(2006) presented context-triggered piecewise
hashing (CTPH) including an implementation
called ssdeep. It was at that time referred
to as “fuzzy hashing”. Later, this term converted to “similarity hashing” (most likely due
to sdhash which stands for “similarity digest
hash” (Roussev, 2010)). In 2014, the National
Institute of Standards and Technology (NIST)
developed Special Publication 800-168, which
outlines the definition and technology for these
kinds of algorithms (Breitinger, Guttman, et al.,
2014).
In addition to the prominent aforementioned implementations, there are several others.
MinHash (Broder, 1997) and SimHash (Sadowski
& Levin, 2007) are ideas on how to detect/identify small changes (up to several bytes),
but were not designed to compare hard disk
images with each other. Oliver et al. (2013)
presented an algorithm named TLSH, which is
2
in a similarity score. Each similarity digest is a
collection of Bloom filters (Bloom, 1970).
To create the similarity digest, MRSH-v2 splits
an input into chunks (also known as “subhashes”) of approximately 160 bytes. These
chunks are hashed using FNV (a fast noncryptographic hash function), which is used to
set 5 bits of the Bloom filter. To divide the input
into chunks, it uses a window of 7 bytes, which
slides through the input byte-by-byte. The content of the window is processed and whenever it
hits a certain value (based on a modulus operation), the end of a chunk is identified. Thus, the
actual size of each chunk varies. Each Bloom filter has a specific capacity. Once this has been
reached, any further chunks are inserted into a
new Bloom filter that is appended to the digest. Approximate matching occurs by comparing similarity digests against one another. To
compare two file sets, an all-against-all pairwise
comparison is required.
One way to improve upon the all-against-all
comparison is to use the file-against-set strategy
outlined by Breitinger, Baier, and White (2014).
An alternative strategy that has not yet been
fully evaluated is to use a hierarchical Bloom
filter tree (HBFT), as suggested by Breitinger,
Rathgeb, and Baier (2014). This approach is intended to achieve speed benefits over a pairwise
comparison while supporting the identification
of specific matching files. The primary contribution of this paper is to investigate the factors
that affect the runtime performance of this latter
approach, compared to the pairwise comparisons
required by the original algorithm.
premised on locality sensitivity hashing (LSH).
There are significantly more algorithms, but to
explain all of them would be beyond the scope
of this paper; a good summary is provided by
Harichandran et al. (2016).
While these algorithms have great capabilities,
they suffer one significant drawback, which we
call the “database lookup problem”. In comparison to traditional hash values which can be
sorted and have a lookup complexity of O(1)
(hashmap) or O(log(n)) (binary tree; where n
is the number of entries in the database), looking up a similarity digest usually requires an
all-against-all comparison (O(n2 )) to identify all
matches. To overcome this drawback, Breitinger,
Baier, and White (2014) presented a new idea
that overcomes the lookup complexity (it is approximately O(1)) but at the cost of inaccuracy.
More specifically, the method allows item vs. set
queries, resulting in the answer either being “yes,
the queried item is in the set” or “no, it is not”;
one cannot say against which item it matches.
As a means of addressing these drawbacks,
Breitinger, Rathgeb, and Baier (2014) presented
a further article where they offered a theoretical solution to the lookup problem, based on a
tree of Bloom filters. However, an implementation (and thus a validation) has not been conducted to date. We refer to this as a Hierarchical Bloom Filter Tree (HBFT). The focus of the
present work is the empirical evaluation of this
approach, so as to demonstrate its effectiveness
and to investigate some practical factors that affect its performance.
3.
THE MRSH-V2 ALGORITHM
4.
The work in this paper is intended to improve
upon the performance of the MRSH-v2 algorithm.
Therefore, it is important to firstly outline its
operation in informal terms, which will aid the
discussion later. A more detailed, formal description of the algorithm can be found in the
paper by Breitinger and Baier (2012). The primary goal of MRSH-v2 is to compress any byte
sequence and output a similarity digest. Similarity digests are created in a way that they can
be compared with each other, which will result
HIERARCHICAL BLOOM
FILTER TREES (HBFT)
In a Hierarchical Bloom Filter Tree (HBFT), the
root node of the tree is a Bloom filter that represents the entire collection. A key feature of a
Bloom filter is that it can say only whether an
item is probably contained in it, or definitely not
contained in it. Thus it is possible to give false
positive results, but not false negatives. The
rate of false positives depends on the size of the
Bloom filter and the number of items it contains.
3
Figure 1: Hierarchical Bloom Filter Tree (HBFT) structure using (a) variable-width and (b) fixedwidth Bloom filters.
usually relates to multiple files.
Depending on the design of the tree, a leaf
node may represent multiple files. Thus a search
that reaches a leaf node will still require a pairwise comparison with each file in this subset, using MRSH-v2. However, given that most searches
will reach only a subset of the root nodes, the
number of pairwise comparisons required for
each file is greatly reduced.
The process to check if a file matches a Bloom
filter node is similar to the process of inserting
a file into the tree. However, instead of inserting each hash into the node, its subhashes are
instead checked against the Bloom filter to see
if they are contained in it. If a specific number
of consecutive hashes are contained in the node,
this is considered to be a match. The number of
consecutive hashes is configurable as a parameter named min run. The first experiment in this
paper (discussed in Section 5.2.1) explores the
effects of altering this value.
In constructing a HBFT, memory constraints
will have a strong influence on the design of the
tree. In practical situations, a typical workstation is unlikely (at present) to have access to over
16GiB of main memory. Thus trade-offs in the
design of the tree are likely. Larger Bloom filters have lower false positive rates (assuming the
quantity of data is constant), but lead to shallower trees (thus potentially increasing the number of pairwise comparisons required).
Of the two proposed designs for a HBFT, each
has its own theoretical advantages. One aim of
the ensuing experiments is to identify if any of
these has more influence in practice. Some considerations worthy of note include:
When searching for a file, if a match is found at
the root of the tree, its child nodes can then be
searched. Although this structure is inspired by
a classic binary search tree, a match at a particular node in a HBFT does not indicate whether
the search should continue in the left or right
subtree. Instead, both child nodes need to be
searched, with the search path ending when a
leaf node is reached or a node does not match.
Two forms of tree layout are shown in Figure 1: one uses Bloom filters of different sizes (referred to as a “variable-width” HBFT), whereas
the other uses a single fixed size for each Bloom
filter. For the variable-width tree, each level in
the tree is allocated an equal amount of memory.
Thus each Bloom filter occupies half the memory
of its parent, and also represents a file set that
is half the size of its parent. The expected false
positive rates will be approximately equal at all
levels in the tree. In contrast, a fixed-width tree
uses the same size for every Bloom filter. Thus
each level of the tree occupies twice the space of
the level above.
When a collection is being modelled as a
HBFT, each file is inserted into the Bloom filter at some leaf node in the tree, and also into
its ancestor nodes. The mechanism of inserting a
file into a Bloom filter is the same as for the single Bloom filter approach outlined by Breitinger,
Baier, and White (2014), which is also very similar to the approach taken by the classic MRSH-v2
algorithm outlined in Section 3. The key difference is that instead of creating a similarity digest
of potentially multiple small Bloom filters for an
individual file, each subhash is used to set 5 bits
of the larger Bloom filter within a tree node that
4
5.
• Calculating the union of two Bloom filters
of equal size is trivially performed using a
bitwise-OR operation. Thus a fixed-width
HBFT can be constructed in a bottom-up
manner, whereby each file needs only be inserted into a leaf Bloom filter. Once all files
have been processed, these leaves can be recursively merged to create the parent nodes.
In contrast, a variable-width tree design requires each file to be inserted separately into
an appropriate node at every level in the
tree.
EXPERIMENTS
As part of this work, a number of experiments
were conducted to examine the factors that affect the performance of the HBFT structure. In
each case, a HBFT was used to model the contents of a dataset. Files from another dataset
were then searched for in the tree, and the results reported. Because the speed of execution is
of paramount importance, and because the original MRSH-v2 implementation was written in C,
the HBFT implementation used for these experiments was also written in that language. The
source code has been made available under the
Apache 2.0 licence1 .
The workstation used for the experiments contains a quad-core Intel Core i7 2.67GHz processor, 12GiB of RAM and uses a solid state drive
for storage. The operating system is Ubuntu
Linux 16.04 LTS. The primary constraint this
system imposes on the design of experiments is
that of the memory that is available for storing
the HBFTs. For all experiments, the maximum
amount of memory made available for the HBFT
was 10GiB. The size of the individual Bloom filters within the trees then depended on the number of nodes in the tree (which in turn depends
on the number of leaf nodes).
For each experiment, the number of leaf nodes
(n) is specified in advance, from which the total
number of nodes can be computed (since this is
a complete binary tree). Given the upper total
memory limit (u, in bytes), and that the size of
each Bloom filter (in bytes) should be a power of
two (per Breitinger, Baier, and White (2014)),
it is possible to calculate the maximum possible
size of each Bloom filter.
For a variable-width tree, all levels in the
tree are allocated the same amount of memory.
Therefore the size of the root Bloom filter in
bytes (r) is given by:
• Another consequence of the above observation is that if the HBFT is to be distributed
over multiple computational nodes, this has
consequences for the quantity of data that
must be shared between nodes when building the tree. Instead of sending the hashes
of all files throughout the system, Bloom filters can be shared and locally merged where
necessary. This is outside of the scope of
this paper, but is discussed as future work
in Section 6.
• The memory required at each level of
a fixed-width tree increases exponentially.
This means that for an equal amount of
available memory, a fixed-width tree must
necessarily either be built to be shallower
than the variable-width tree, or make use of
smaller Bloom filters towards the top of the
tree. The latter approach results in a higher
false-positive rate in this part of the tree,
which will likely lead to deeper searches for
files that are not matched with anything in
the corpus. If the size of the leaf nodes is
equal, then the overall false positive rate of
the two trees will be equivalent.
r = 2blog2 (u/(log2 (n)+1))c
For both types of tree, larger Bloom filters result in lower false positive rates at the expense of
a shallower tree (since memory is limited). In a
shallower tree, each leaf node represents a larger
subset of the corpus, which may require a greater
number of pairwise comparisons for each search.
(1)
The size of the other nodes in bytes is then 2rd
where d is the depth of the node in the tree (i.e.
the size of a Bloom filter is half the size of its
parent).
1
5
Available at http://github.com/ishnid/mrsh-hbft
• The win7 dataset is a fresh installation of a
Windows 7 operating system, with default
options selected during installation. It consists of 48,384 files (excluding symbolic links
and zero-byte files) and occupies approximately 10GiB.
For a HBFT with fixed-sized Bloom filters, all
nodes have size equal to that of the root node.
Here, r is given by:
r = 2blog2 (u/(2n−1))c
(2)
The ultimate goal of the experiments is to
demonstrate that the HBFT approach can improve the running time of an investigation
over the all-against-all comparison approach of
MRSH-v2 without suffering a degradation in effectiveness. It achieves this by narrowing the
search space so that each file that is searched for
need only be compared against a subset of the
dataset.
Using a HBFT, the final outcome will be a set
of similarity scores. This score is calculated by
using MRSH-v2 to compare the search file with all
files contained in any leaves that are reached during the search. Therefore, the HBFT approach
will not identify a file as being similar if MRSH-v2
does not also do so.
In these experiments, the similarity scores generated by MRSH-v2 are considered to be ground
truth. Evaluating the degree to which this agrees
with the opinion of a human judge, or how it
compares with other algorithms, is outside the
scope of this paper. The primary difference between the outputs is that the HBFT may fail to
identify files that MRSH-v2 considers to be similar (i.e. false negatives) due to an appropriate
leaf node not being reached.
Therefore the primary metric used, aside from
running time, is recall: the proportion of knownsimilar (or known-identical) files for which the
HBFT search reaches the appropriate leaf node.
5.1
The first two experiments use one or both of
these datasets directly. The final experiment includes some modifications, as outlined in Section 5.2.3.
5.2
Experiment Overview
The following sections present three experiments
that were conducted to evaluate the HBFT approach. Section 5.2.1 compares the t5 dataset
with itself. This is intended to find whether the
HBFT approach is effective in finding identical
files, and to investigate the effect of varying certain parameters when designing and searching a
HBFT. It also aims to demonstrate the extent
to which the number of pairwise comparisons required can be reduced by using this technique.
Section 5.2.2 uses disjoint corpora of different sizes (t5 and win7). In a typical investigation, there may be a large difference between the
size of the collection of search files and a seized
hard disk. This experiment aims to investigate
whether it is preferable to use the tree to model
the smaller or the larger corpus. Additionally, it
examines the performance characteristics of fixed
and variable width HBFTs.
Finally, Section 5.2.3 uses overlapping corpora
where a number of files have been planted on the
disk image. These files are identical to, or similar
to, files in the search corpus. This experiment
demonstrates that using a HBFT is substantially
faster than the pairwise approach.
Datasets
Two datasets were used as the basis for the experiments conducted in this paper:
5.2.1
Experiment 1: t5 vs. t5
For the initial experiment, the HBFT was constructed to represent the t5 corpus. All files
from t5 were also used for searching. Thus every file searched for is also located in the tree
and should be found. Conducting an all-againstall pairwise comparison using MRSH-v2 required a
total of 19,864,849 comparisons, which took 319
seconds.
• The t5 dataset (Roussev, 2011) is frequently
used for approximate matching experimentation. It consists of 4,457 files (approximately 1.8 GiB) taken from US government
websites. It includes plain text files, HTML
pages, PDFs, Microsoft Office documents
and image files.
6
are 000462.text, 001774.html, 003225.html.
These files are 6.5 KiB, 6.6 KiB and 4.5 KiB
in size respectively. Although each chunk is approximately 160 bytes, this is variable depending
on the file content. While these are relatively
small files, they are not the smallest in the corpus. This shows that even when the file is large
enough to contain 8 chunks of the average size,
a min run requirement of 8 successive matches
may still not be possible. Similarly, using 6 as the
min run value results in two files being missed.
The type of HBFT used did not alter these results.
To construct the tree, the smallest number of
leaf nodes was 32. Following this, the number of
leaf nodes was doubled each time (maintaining
a balanced tree). The exception was that after
the experiment with 2,048 leaf nodes, 4,457 leaf
nodes were used for the final run, thereby representing a single file from the corpus in each leaf.
The aims of this experiment were:
1. Evaluate the effectiveness of the HBFT approach for exact matching (i.e. finding identical files) using recall.
2. Identify an appropriate value for MRSH-v2’s
min run parameter.
Table 1: Effect of min run on recall: identical
files.
min run
Recall
3. Investigate the relationship between the size
of the tree and the time taken to build and
search the tree.
4
6
8
4. Investigate the relationship between the size
of the tree and the number of pairwise comparisons that are required to calculate a similarity score.
100%
99.96%
99.93%
It should be acknowledged that if the aim is
solely to identify identical files, then existing
hash-based techniques will take less time and
yield more reliable results. Intuitively, however,
a system that is intended to find similar files
should also find identical files. While the chunk
size of 160 bytes will always fail to match very
small files, it is desirable to find matches when
file sizes are larger.
Figure 2 shows the time taken to build the
tree and search for all files. As the number of
When running the experiment, it became apparent that the first two aims are linked. Table 1 shows the recall associated with three values of min run: 4, 6 and 8. Using a min run
value of 4 resulted in full recall. However, increasing min run to 6 or 8 resulted in a small
number of files being omitted. When min run is
set to 8, three files are not found in the tree.
This indicates the dangers inherent in requiring longer matching runs. The files in question
Figure 2: Effect of varying number of leaf nodes on time taken: t5 vs. t5
7
comparisons required by MRSH-v2. Thus reducing this search space is the primary function of
the tree. Larger trees tend to result in a smaller
number of comparisons. For the largest trees
(with 4,457 leaves), the min run value does not
have a material effect on the number of comparisons required, regardless of whether the tree is
variable-width or fixed-width. This implies that
although searches tend to reach deeper into the
tree (hence the longer running time), they do not
reach substantially more leaves.
From this experiment, it can be concluded that
using a min run value of 4 is desirable in order to
find exact matches. This causes the time taken
to search to be slightly longer, while having a
negligible impact on the number of pairwise comparisons required afterwards. Fixed-width and
variable-width HBFTs exhibit similar characteristics for a corpus of this size
leaf nodes in the tree increases, so too does the
time taken to search the tree. Higher values of
min run use slightly less time, due to the fact
that it is more difficult for a search to descend
to a lower level when more matches are required
to do so. However, as the recall for these higher
values is lower, 4 was used as the min run value
for further experiments.
The times shown here relate only to building
the tree and searching for files within it, and do
not include the time for the pairwise comparisons at the leaves. Therefore, although using
32 leaf nodes results in the shortest search time
(due to the shallower tree), it would require a
most comparisons, as each leaf node represents
1
32 of the entire corpus. As an illustration, using a variable-width tree with 32 leaf nodes and
min run value of 4 requires 8,538,193 pairwise
comparisons after searching the tree. A similar
tree with 4,457 leaves requires 617,860 comparisons.
5.2.2
Experiment 2: t5 vs. win7 and
win7 vs. t5
One issue that is important to note is that
the time required to perform a full pairwise comparison is 319 seconds. However, for the largest
trees, the times for building and searching the
tree are 274 and 309 seconds for a variable and
fixed HBFT respectively. Thus, for a relatively
small collection such as this, the use of the tree
is unlikely to provide benefits in terms of overall
running time.
The second experiment was designed to operate
with larger dataset sizes. In this experiment, t5
was used as a proxy for a set of known-illegal
files, and win7 was used to represent a seized
disk.
The aims of this experiment were:
Figure 3 plots the number of leaf nodes against
the total number of comparisons required to
complete the investigation. As the size of corpora increases, so does the number of pairwise
2. Contrast the performance of variable-width
and fixed-width trees.
1. Investigate whether the HBFT should represent the smaller or larger corpus.
3. Measure the effect on overall running time
Figure 3: Effect of varying number of leaf nodes on number of comparisons: t5 vs. t5
8
Figure 4: Time to search for win7 in a t5 tree.
Figure 5: Time to search for t5 in a win7 tree.
of using a HBFT.
for both types of tree. This is unsurprising in
the context of disjoint corpora. Most files will
not match, so many searches will end at the root
node, or at an otherwise shallow depth.
Figure 5 shows results when the tree models
win7. With only 32 leaf nodes, it is notable that
all four experimental runs take approximately
the same total time, regardless of the type of tree
or the dataset that is chosen for the tree to represent. Due to its size, the build times for the win7
trees are substantially longer than for t5. The
search time exhibits a generally upward trend as
the number of leaf nodes increases: a trend that
is far more pronounced for the fixed-width tree.
This is a consequence of the hardware constraints associated with the setup of the experiment. Because memory footprint is constrained,
a tree with 48,384 leaf nodes will contain Bloom
filters that are much smaller than for trees with
In pursuit of the first objective, the experiment was first run by building a tree to represent t5 and then searching for the files contained
in win7. The number of leaf nodes in this tree
was varied in the same way as in Experiment 1.
Then this was repeated by inserting win7 into a
tree and searching for the files from t5. Again
the number of leaf nodes was doubled every time,
with the exception that the largest tree contained
one leaf node for every file in the collection (i.e.
48,384 leaves). This procedure was followed for
both forms of tree.
The time taken to build and search the trees
are shown in Figures 4 and 5. Figure 4 shows
the results when the tree represents t5, with the
time subdivided into the time spent building the
tree and the time spent searching for all the files
from win7. The total time is relatively consistent
9
fewer nodes. For the variable-width tree representing win7, although its leaf nodes are 8KiB
in size, its root node is 512MiB. In the corresponding fixed-width tree, the Bloom filters are
all 64KiB. The false positive rate associated with
Bloom filters is much higher for smaller Bloom
filters. Thus even where two corpora have no files
in common, searches int he fixed-width tree will
descend deeper due to false positives higher in
the tree, hence increasing the search time. This
is likely to be even more pronounced in corpora
that have a substantial number of similar files.
Therefore, the fixed-width tree in its current design is unlikely to successfully scale to very large
corpora.
Overall, the total time taken is less when the
tree represents the smaller dataset. As with the
first experiment, the total number of pairwise
comparisons decreases as the number of leaves
increases. Table 2 shows the total number of
comparisons that are required when using the
largest number of leaf nodes (i.e. 4,457 when
the tree represents t5 and 48,384 when win7 is
stored in the tree). Both types of tree require
a smaller number of comparisons when the tree
models t5. This, combined with the lower build
and search time suggest that the preferred approach should be to use the smaller corpus to
construct the HBFT.
isons, was 1,094 seconds. In comparison, the
time taken to perform a full pairwise comparison using MRSH-v2 is 2,858 seconds.
5.2.3
The final experiment involved overlapping
datasets, constructed as follows:
• A set of simulated “known-illegal” files:
4,000 files from t5.
• A simulated seized hard disk: the win7 image, plus 140 files from t5, as follows:
– 100 files that are contained within the
4,000 “illegal” files.
– 40 files that themselves are not contained within the “illegal” files, but
that have a high similarity with files in
the corpus, according to MRSH-v2. 10
of these files have a similarity of 80%
or higher, 10 have a similarity between
60% and 79% (inclusive), 10 have a
similarity between 40% and 59% (inclusive) and 10 have a similarity between 20% and 39% (inclusive).
The aims of this third experiment were:
1. Evaluate the time taken to perform a full
search, compared with the all-against-all
pairwise approach of MRSH-v2.
Table 2: Number of pairwise comparisons required for largest trees: t5 vs. win7
Tree
Search
Fixed
Variable
t5
win7
win7
t5
98,260
193,924
Experiment 3: Planted evidence
2. Evaluate the success of the approach in finding the 100 “illegal” files that are included
verbatim in the hard disk image, and the
40 files from the image that are similar to
“illegal” files, according to MRSH-v2.
98,260
101,386
Memory is an additional consideration. Using
a HBFT to model the larger dataset requires the
similarity hashes of all its files to be cached at
the leaves. This requires more memory than for
the smaller collection, thus reducing the amount
of memory available to store the HBFT itself.
Following these observations, the experiment
was repeated once more. A variable-width tree
was used, which modelled t5 with 4,457 leaves.
All files from win7 were then searched for. The
total running time, including pairwise compar-
For the first aim, the primary metric is the
time taken for the entire process to run, comprising the time to build the tree, the time to
search the tree and the time required to conduct
the pairwise comparisons at the leaves. In evaluating the latter aim, recall is used. Here, “recall”
refers to the percentage of the 100 identical files
that are successfully identified, and “similar recall” refers to the percentage of the 40 similar
files that are successfully found. A file is considered to have been found if the search for the file
10
Figure 6: Time to search for planted evidence (including pairwise comparisons).
it is similar or identical to reaches the leaf node
that contains it, yielding a pairwise comparison.
The total running time for MRSH-v2 was 2,592
seconds. The running times of the HBFT approach are shown in Figure 6. Following the
insights gained in the previous experiment, the
smaller collection of 4,000 “illegal” files was used
to construct the tree and then searches were
conducted for all of the files in the larger corpus. The “Search Time” includes the time spent
searching the tree and the time to perform the
comparisons at the leaves.
As expected, for both types of tree the maximum number of leaf nodes resulted in the
fastest run time. This configuration also yielded
the maximum reduction in the number of pairwise comparisons required, without substantially
adding to the time required to build and search
the tree. The remainder of this analysis focuses
on this scenario, where the tree has 4,000 leaf
nodes.
Using a variable-width tree took 1,182 seconds
(a 54% reduction in the time required for an
all-against-all pairwise comparison). The fixedwidth tree took 1,207 seconds (a 53% reduction).
This illustrates that in terms of run-time, the
HBFT approach offers substantial speed gains
over pairwise comparison. Due to the lack of
scalability of the pairwise approach, this difference is likely to be even more pronounced for
larger datasets.
In terms of effectiveness, all 100 files that were
common to the two corpora were successfully
Table 3: Similar recall for Planted Evidence experiment.
MRSH-v2
Files
Files
Similar
similarity
planted found
recall
80%-100%
60%-79%
40%-59%
20%-39%
10
10
10
10
10
10
10
8
100%
100%
100%
80%
Overall
40
38
95%
found in both tree types. The similar recall is
shown in Table 3 and is the same for both types
of tree. All files with a MRSH-v2 similarity of 40%
or greater with a file in the “illegal” set were successfully identified. Two files with a lower similarity (25% and 26%) were not found. This yields
an overall similar recall score of 95% for all 40
files.
This is an encouraging result, indicating that
the HBFT approach is extremely effective at
finding files that are similar above a reasonable
threshold of 40% and exhibits full recall for identical files. Thus it can be concluded that the
HBFT data structure is a viable alternative to
all-against-all comparisons in terms of effectiveness, while achieving substantial speed gains.
6.
CONCLUSIONS AND
FUTURE WORK
This paper aimed to investigate the effectiveness of using a Hierarchical Bloom Filter Tree
11
A number of avenues for future work are apparent:
(HBFT) data structure to improve upon the allagainst-all pairwise comparison approach used
by MRSH-v2. A number of experiments were conducted with the aim of improving the speed of
the process. Additionally, it was important that
files that should be found were not omitted (i.e.
that recall is maintained).
The first experiment found that while HBFTs
with more leaf nodes take longer to build and
search, they reduce the number of pairwise comparisons required by the greatest degree. It
also suggested the use of a min run value of 4,
as higher values resulted in imperfect recall for
identical files.
The results of the second experiment indicated
that when using corpora of different sizes, it is
preferable to build the tree to model the smaller
collection and then search for the files that are
contained the larger corpus. For larger trees,
it was additionally noted that the fixed-width
HBFT did not scale as well as its variable-width
counterpart. This is due to the small size of the
Bloom filters used in the tree as the number of
nodes increases.
For the final experiment, a Windows 7 image
was augmented by the addition of a number of
files that were identical to those being searched
for, and a further group that were similar. The
HBFT approach yielded a recall level of 100% for
the identical files and of 95% for the similar files,
when using mrsh-v2 as ground truth. On examining the two files that were not found, it was
noted that these had a relatively low similarity
to the search files (25% and 26% respectively),
with all files with a higher similarity score being identified successfully. The run time for this
experiment was substantially quicker than an allagainst-all comparison: a 54% time reduction for
a variable-width tree and a 53% reduction for a
fixed-width tree.
These experiments lead to the conclusion that
the HBFT approach is a highly promising technique. Due the poor scalability of the traditional
all-against-all approach, it can be inferred that
this performance improvement will be even more
pronounced as datasets become larger.
Given the promising results of the experiments
presented in this paper, further work is planned.
• Cuckoo filters have been identified as a
promising replacement for Bloom filters for
approximate matching purposes (V. Gupta
& Breitinger, 2015). It is possible that these
could be incorporate into a similar hierarchical tree structure to produce further improvements.
• Currently, when building the tree, files are
allocated to leaf nodes in a round-robin fashion. For trees with multiple files represented at each leaf, it may be possible that a
more optimised allocation mechanism could
be used for this (e.g. to allocate similar files
to the same leaf node).
• The current model also uses balanced trees,
with the result that all successful searches
reach the same depth in the tree. Further investigation may reveal circumstances where
an unbalanced tree is preferable so as to
shorten some more common searches.
• Parallelisation and distribution are highly
likely to yield further performance improvements, and this should be investigated.
• Fixed-width HBFTs do not scale to the
same extent as variable trees, due to the
high false positive rates that are associated
with the small Bloom filters that result from
using large trees with many nodes. Although the experiments presented in this
paper indicate that variable-width HBFTs
are preferable, there may be circumstances
where fixed-width trees may be useful, due
to their theoretical advantages noted in Section 4.
• While these experiments have used MRSH-v2
as the algorithm for calculating the similarities at the leaf nodes, other algorithms
should be considered also (e.g. sdhash).
12
REFERENCES
9. doi: 10.1016/j.cor.2015.11.003
Gupta, V., & Breitinger, F. (2015). How
cuckoo filter can improve existing approximate matching techniques. In J. James &
F. Breitinger (Eds.), Digital forensics and
cyber crime (Vol. 157, p. 39-52). Springer
International Publishing. doi: 10.1007/
978-3-319-25512-5 4
Harichandran, V. S., Breitinger, F., & Baggili, I.
(2016). Bytewise Approximate Matching:
The Good, The Bad, and The Unknown.
The Journal of Digital Forensics, Security
and Law: JDFSL, 11 (2), 59.
James, J. I., & Gladyshev, P. (2015). Automated
Inference of Past Action Instances in Digital Investigations. International Journal of
Information Security, 14 (3), 249–261. doi:
10.1007/s10207-014-0249-6
Kornblum, J. (2006). Identifying Identical Files
Using Context Triggered Piecewise Hashing. Digital investigation, 3 , 91–97. doi:
10.1016/j.diin.2006.06.015
Lillis, D., Becker, B., O’Sullivan, T., & Scanlon,
M. (2016). Current Challenges and Future
Research Areas for Digital Forensic Investigation. In 11th ADFSL Conference on Digital Forensics, Security and Law (CDFSL
2016). Daytona Beach, FL, USA: ADFSL.
doi: 10.13140/RG.2.2.34898.76489
Lillis, D., Breitinger, F., & Scanlon, M.
(2017). Expediting MRSH-v2 Approximate Matching with Hierarchical Bloom
Filter Trees. In 9th EAI International
Conference on Digital Forensics and Cyber Crime (ICDF2C 2017). Prague, Czech
Republic.
Oliver, J., Cheng, C., & Chen, Y. (2013). TLSH–
A Locality Sensitive Hash. In Fourth
Cybercrime and Trustworthy Computing
Workshop (CTC), 2013 (pp. 7–13). doi:
10.1109/CTC.2013.9
Quick, D., & Choo, K.-K. R. (2014). Impacts
of Increasing Volume of Digital Forensic
Data: A Survey and Future Research Challenges. Digital Investigation, 11 (4), 273–
294. doi: 10.1016/j.diin.2014.09.002
Rogers, M. K., Goldman, J., Mislan, R., Wedge,
T., & Debrota, S. (2006). Computer Foren-
Bloom, B. H. (1970). Space/Time Trade-offs in
Hash Coding with Allowable Errors. Communications of the ACM , 13 (7), 422–426.
Breitinger, F., & Baier, H. (2012). Similarity Preserving Hashing: Eligible Properties
and a New Algorithm MRSH-v2. In International conference on digital forensics and
cyber crime (pp. 167–182). Springer.
Breitinger, F., Baier, H., & White, D. (2014). On
the Database Lookup Problem of Approximate Matching. Digital Investigation, 11 ,
S1–S9. doi: 10.1016/j.diin.2014.03.001
Breitinger, F., Guttman, B., McCarrin, M.,
Roussev, V., & White, D. (2014). Approximate Matching: Definition and Terminology. NIST Special Publication, 800 ,
168.
Breitinger, F., Rathgeb, C., & Baier, H. (2014,
sep).
An Efficient Similarity Digests
Database Lookup - A Logarithmic Divide
& Conquer Approach. Journal of Digital
Forensics, Security and Law , 9 (2), 155–
166.
Broder, A. Z. (1997). On the Resemblance
and Containment of Documents. In Compression and complexity of sequences 1997.
proceedings (pp. 21–29). doi: 10.1109/
SEQUEN.1997.666900
Casey, E., Ferraro, M., & Nguyen, L. (2009). Investigation Delayed is Justice Denied: Proposals for Expediting Forensic Examinations of Digital Evidence. Journal of forensic sciences, 54 (6), 1353–1364.
de Braekt, R. I., Le-Khac, N. A., Farina, J.,
Scanlon, M., & Kechadi, T. (2016, April).
Increasing Digital Investigator Availability Through Efficient Workflow Management and Automation. In 4th international symposium on digital forensic and
security (isdfs) (p. 68-73). doi: 10.1109/
ISDFS.2016.7473520
Gupta, J. N., Kalaimannan, E., & Yoo, S.-M.
(2016). A Heuristic for Maximizing Investigation Effectiveness of Digital Forensic Cases Involving Multiple Investigators.
Computers & Operations Research, 69 , 1–
13
Workshop (Vol. 94).
Sadowski, C., & Levin, G. (2007). Simhash:
Hash-based Similarity Detection. Citeseer.
Scanlon, M. (2016, 08). Battling the Digital
Forensic Backlog through Data Deduplication. In Proceedings of the 6th IEEE
International Conference on Innovative
Computing Technologies (INTECH 2016).
Dublin, Ireland: IEEE.
van Baar, R., van Beek, H., & van Eijk, E.
(2014). Digital Forensics as a Service: A
Game Changer. Digital Investigation, 11,
Supplement 1 , S54 - S62. doi: 10.1016/
j.diin.2014.03.007
sics Field Triage Process Model. Journal of
Digital Forensics, Security and Law , 1 (2),
19–38.
Roussev, V. (2010). Data Fingerprinting with
Similarity Digests. In IFIP International
Conference on Digital Forensics (pp. 207–
226). Springer. doi: 10.1007/978-3-642
-15506-2 15
Roussev, V. (2011). An Evaluation of Forensic
Similarity Hashes. Digital Investigation, 8 ,
S34–S41.
Roussev, V., & Richard III, G. G. (2004). Breaking the Performance Wall: The Case for
Distributed Digital Forensics. In Proceedings of the 2004 Digital Forensics Research
14
| 8cs.DS
|
Published as a conference paper at ICLR 2018
C GAN S WITH
P ROJECTION D ISCRIMINATOR
Takeru Miyato1 , Masanori Koyama2
[email protected]
[email protected]
1
Preferred Networks, Inc. 2 Ritsumeikan University
arXiv:1802.05637v1 [cs.LG] 15 Feb 2018
A BSTRACT
We propose a novel, projection based way to incorporate the conditional information into the discriminator of GANs that respects the role of the conditional
information in the underlining probabilistic model. This approach is in contrast
with most frameworks of conditional GANs used in application today, which use
the conditional information by concatenating the (embedded) conditional vector
to the feature vectors. With this modification, we were able to significantly improve the quality of the class conditional image generation on ILSVRC2012 (ImageNet) 1000-class image dataset from the current state-of-the-art result, and we
achieved this with a single pair of a discriminator and a generator. We were
also able to extend the application to super-resolution and succeeded in producing
highly discriminative super-resolution images. This new structure also enabled
high quality category transformation based on parametric functional transformation of conditional batch normalization layers in the generator. The code with
Chainer (Tokui et al., 2015), generated images and pretrained models are available
at https://github.com/pfnet-research/sngan_projection.
1
I NTRODUCTION
Generative Adversarial Networks (GANs) (Goodfellow et al., 2014) are a framework to construct a
generative model that can mimic the target distribution, and in recent years it has given birth to arrays
of state-of-the-art algorithms of generative models on image domain (Radford et al., 2016; Salimans
et al., 2016; Ledig et al., 2017; Zhang et al., 2017; Reed et al., 2016). The most distinctive feature
of GANs is the discriminator D(x) that evaluates the divergence between the current generative
distribution pG (x) and the target distribution q(x) (Goodfellow et al., 2014; Nowozin et al., 2016;
Arjovsky et al., 2017). The algorithm of GANs trains the generator model by iteratively training
the discriminator and generator in turn, with the discriminator acting as an increasingly meticulous
critic of the current generator.
Conditional GANs (cGANs) are a type of GANs that use conditional information (Mirza & Osindero, 2014) for the discriminator and generator, and they have been drawing attention as a promising
tool for class conditional image generation (Odena et al., 2017), the generation of the images from
text (Reed et al., 2016; Zhang et al., 2017), and image to image translation (Kim et al., 2017; Zhu
et al., 2017). Unlike in standard GANs, the discriminator of cGANs discriminates between the generator distribution and the target distribution on the set of the pairs of generated samples x and its
intended conditional variable y. To the authors’ knowledge, most frameworks of discriminators in
cGANs at the time of writing feeds the pair the conditional information y into the discriminator
by naively concatenating (embedded) y to the input or to the feature vector at some middle layer
(Mirza & Osindero, 2014; Denton et al., 2015; Reed et al., 2016; Zhang et al., 2017; Perarnau et al.,
2016; Saito et al., 2017; Dumoulin et al., 2017a; Sricharan et al., 2017). We would like to however,
take into account the structure of the assumed conditional probabilistic models underlined by the
structure of the discriminator, which is a function that measures the information theoretic distance
between the generative distribution and the target distribution.
By construction, any assumption about the form of the distribution would act as a regularization on
the choice of the discriminator. In this paper, we propose a specific form of the discriminator, a form
motivated by a probabilistic model in which the distribution of the conditional variable y given x is
1
Published as a conference paper at ICLR 2018
(a) cGANs,
input concat
(b) cGANs,
hidden concat
(Mirza & Osindero, 2014)
(Reed et al., 2016)
Adversarial
loss
Adversarial
loss
(c) AC-GANs
(d) (ours) Projection
(Odena et al., 2017)
Adversarial Classificaition
loss
loss
0
Adversarial
loss
Inner
product
y
y
Concat
y
Concat
x y
x
x
x
Figure 1: Discriminator models for conditional GANs
(a) Images generated with the projection model. (left) Tibetan terrier and (right) mushroom.
(b) (left) Consecutive category morphing with fixed z. geyser → Tibetan terrier → mushroom → robin. (right)
category morphing from Tibetan terrier to mushroom with different value of fixed z
Figure 2: The generator trained with the projection model can generate diverse set of images. For
more results, see the experiment section and the appendix section.
discrete or uni-modal continuous distributions. This model assumption is in fact common in many
real world applications, including class-conditional image generation and super-resolution.
As we will explain in the next section, adhering to this assumption will give rise to a structure of the
discriminator that requires us to take an inner product between the embedded condition vector y and
the feature vector (Figure 1d). With this modification, we were able to significantly improve the
quality of the class conditional image generation on 1000-class ILSVRC2012 dataset (Russakovsky
et al., 2015) with a single pair of a discriminator and generator (see the generated examples in
Figure 2). Also, when we applied our model of cGANs to a super-resolution task, we were able to
produce high quality super-resolution images that are more discriminative in terms of the accuracy
of the label classifier than the cGANs based on concatenation, as well as the bilinear and the bicubic
method.
2
T HE ARCHITECTURE OF THE C GAN DISCRIMINATOR WITH A
PROBABILISTIC MODEL ASSUMPTIONS
Let us denote the input vector by x and the conditional information by y 1 . We also denote the
cGAN discriminator by D(x, y; θ) := A(f (x, y; θ)), where f is a function of x and y, θ is the
parameters of f , and A is an activation function of the users’ choice. Using q and p to designate
the true distributions and the generator model respectively, the standard adversarial loss for the
1
When y is discrete label information, we can assume that it is encoded as a one-hot vector.
2
Published as a conference paper at ICLR 2018
discriminator is given by:
L(D) = −Eq(y) Eq(x|y) [log(D(x, y))] − Ep(y) Ep(x|y) [log(1 − D(x, y))] ,
(1)
with A in D representing the sigmoid function. By construction, the nature of the ‘critic’ D significantly affects the performance of G. A conventional way of feeding y to D until now has been
to concatenate the vector y to the feature vector x, either at the input layer (Mirza & Osindero,
2014; Denton et al., 2015; Saito et al., 2017), or at some hidden layer (Reed et al., 2016; Zhang
et al., 2017; Perarnau et al., 2016; Dumoulin et al., 2017a; Sricharan et al., 2017) (see Figure 1a and
Figure 1b). We would like to propose an alternative to this approach by observing the form of the
optimal solution (Goodfellow et al., 2014) for the loss function, Eq. (1), can be decomposed into the
sum of two log likelihood ratios:
f ∗ (x, y) = log
q(x|y)q(y)
q(y|x)
q(x)
= log
+ log
:= r(y|x) + r(x).
p(x|y)p(y)
p(y|x)
p(x)
(2)
Now, we can model the log likelihood ratio r(y|x) and r(x) by some parametric functions f1 and
f2 respectively. If we make a standing assumption that p(y|x) and q(y|x) are simple distributions
like those that are Gaussian or discrete log linear on the feature space, then, as we will show, the
parametrization of the following form becomes natural:
f (x, y; θ) := f1 (x, y; θ) + f2 (x; θ) = y T V φ(x; θΦ ) + ψ(φ(x; θΦ ); θΨ ),
(3)
where V is the embedding matrix of y, φ(·, θΦ ) is a vector output function of x, and ψ(·, θΨ ) is
a scalar function of the same φ(x; θΦ ) that appears in f1 (see Figure 1d). The learned parameters
θ = {V, θΦ , θΨ } are to be trained to optimize the adversarial loss. From this point on, we will
refer to this model of the discriminator as projection for short. In the next section, we would like to
elaborate on how we can arrive at this form.
3
M OTIVATION BEHIND THE projection DISCRIMINATOR
In this section, we will begin from specific, often recurring models and show that, with certain
regularity assumption, we can write the optimal solution of the discriminator objective function
in the form of (3). Let us first consider the a case of categorical variable. Assume that y is a
categorical variable taking a value in {1, . . . , C}, which is often common for a class conditional
image generation task. The most popular model for p(y|x) is the following log linear model:
log p(y = c|x) := vcpT φ(x) − log Z(φ(x)),
(4)
P
L
C
pT
where Z(φ(x)) :=
is the partition function, and φ : x 7→ Rd is the
j=1 exp vj φ(x)
input to the final layer of the network model. Now, we assume that the target distribution q can also
be parametrized in this form, with the same choice of φ. This way, the log likelihood ratio would
take the following form;
r(y|x) = log
q(y = c|x)
= (vcq − vcp )T φ(x) − (log Z q (φ(x)) − log Z p (φ(x))) .
p(y = c|x)
(5)
If we make the values of (vcq , vcp ) implicit and put vc := (vcq − vcp ), we can write f1 (x, y = c) =
vcT φ(x). Now, if we can put together the normalization constant − (log Z q (φ(x)) − log Z p (φ(x)))
and r(x) into one expression ψ(φ(x)), we can rewrite the equation above as
f (x, y) := y T V φ(x) + ψ(φ(x)).
(6)
by using y to denote a one-hot vector of the label y and using V to denote the matrix consisting
of the row vectors vc . Most notably, this formulation introduces the label information via an inner
product, as opposed to concatenation. The form (6) is indeed the form we proposed in (3).
We can also arrive at the form (3) for unimodal continuous distributions p(y|x) as well. Let
y ∈ Rd be a d-dimensional continuous variable, and let us assume that conditional q(y|x)
and p(y|x) are both given by Gaussian distributions, so that q(y|x) = N (y|µq (x), Λ−1
q ) and
3
Published as a conference paper at ICLR 2018
q
p
p(y|x) = N (y|µp (x), Λ−1
p ) where µq (x) := W φ(x) and µp (x) := W φ(x). Then the log
density ratio r(y|x) = log(q(y|x)/p(y|x)) is given by:
s
!
|Λq | exp(−(1/2)(y − µq (x))T Λq (y − µq (x)))
r(y|x) = log
(7)
|Λp | exp(−(1/2)(y − µp (x))T Λp (y − µp (x)))
1
= − y T (Λq − Λp ) y + y T (Λq W q − Λp W p )φ(x) + ψ(φ(x)),
2
(8)
where ψ(φ(x)) represents the terms independent of y. Now, if we assume that Λq = Λp := Λ, we
can ignore the quadratic term. If we further express Λq W q − Λp W p in the form V , we can arrive
at the form (3) again.
Indeed, however, the way that this regularization affects the training of the generator G is a little unclear in its formulation. As we have repeatedly explained, our discriminator measures the divergence
between the generator distribution p and the target distribution q on the assumption that p(y|x) and
q(y|x) are relatively simple, and it is highly possible that we are gaining stability in the training
process by imposing a regularity condition on the divergence measure. Meanwhile, however, the
actual p(y|x) can only be implicitly derived from p(x, y) in computation, and can possibly take
numerous forms other than the ones we have considered here. We must admit that there is a room
here for an important theoretical work to be done in order to assess the relationship between the
choice of the function space for the discriminator and training process of the generator.
4
C OMPARISON WITH OTHER METHODS
As described above, (3) is a form that is true for frequently occurring situations. In contrast, incorporation of the conditional information by concatenation is rather arbitrary and can possibly include
into the pool of candidate functions some sets of functions for which it is difficult to find a logical
basis. Indeed, if the situation calls for multimodal p(y|x), it might be smart not to use the model
that we suggest here. Otherwise, however, we expect our model to perform better; in general, it is
preferable to use a discriminator that respects the presumed form of the probabilistic model.
Still another way to incorporate the conditional information into the training procedure is to directly
manipulate the loss function. The algorithm of AC-GANs (Odena et al., 2017) use a discriminator
(D1 ) that shares a part of its structure with the classifier(D2 ), and incorporates the label information
into the objective function by augmenting the original discriminator objective with the likelihood
score of the classifier on both the generated and training dataset (see Figure 1c). Plug and Play
Generative models (PPGNs) (Nguyen et al., 2017) is another approach for the generative model that
uses an auxiliary classifier function. It is a method that endeavors to make samples from p(x|y)
using an MCMC sampler based on the Langevin equation with drift terms consisting of the gradient
of an autoencoder prior p(x) and a pretrained auxiliary classifier p(y|x). With these method, one
can generate a high quality image. However, these ways of using auxiliary classifier may unwittingly
encourage the generator to produce images that are particularly easy for the auxiliary classifier to
classify, and deviate the final p(x|y) from the true q(x|y). In fact, Odena et al. (2017) reports
that this problem has a tendency to exacerbate with increasing number of labels. We were able to
reproduce this phenomena in our experiments; when we implemented their algorithm on a dataset
with 1000 class categories, the final trained model was able to generate only one image for most
classes. Nguyen et al.’s PPGNs is also likely to suffer from the same problem because they are using
an order of magnitude greater coefficient for the term corresponding to p(y|x) than for the other
terms in the Langevin equation.
5
E XPERIMENTS
In order to evaluate the effectiveness of our newly proposed architecture for the discriminator,
we conducted two sets of experiments: class conditional image generation and super-resolution
on ILSVRC2012 (ImageNet) dataset (Russakovsky et al., 2015). For both tasks, we used the
ResNet (He et al., 2016b) based discriminator and the generator used in Gulrajani et al. (2017),
and applied spectral normalization (Miyato et al., 2018) to the all of the weights of the discriminator
to regularize the Lipschitz constant. For the objective function, we used the following hinge version
4
Published as a conference paper at ICLR 2018
of the standard adversarial loss (1) (Lim & Ye, 2017; Tran et al., 2017)
h
h
ii
L(Ĝ, D) = Eq(y) Eq(x|y) [max(0, 1 − D(x, y)] + Eq(y) Ep(z) max(0, 1 + D Ĝ(z, y), y)
,
h
h
ii
L(G, D̂) = −Eq(y) Ep(z) D̂(G(z, y), y)) ,
(9)
where the last activation function A of D is identity function. p(z) is standard Gaussian distribution
and G(z, y) is the generator network. For all experiments, we used Adam optimizer (Kingma & Ba,
2015) with hyper-parameters set to α = 0.0002, β1 = 0, β2 = 0.9. We updated the discriminator
five times per each update of the generator. We will use concat to designate the models (Figure 1b)2 ,
and use projection to designate the proposed model (Figure 1d) .
5.1
C LASS -C ONDITIONAL I MAGE G ENERATION
The ImageNet dataset used in the experiment of class conditional image generation consisted of
1,000 image classes of approximately 1,300 pictures each. We compressed each images to 128×128
pixels. Unlike for AC-GANs3 we used a single pair of a ResNet-based generator and a discriminator.
Also, we used conditional batch normalization (Dumoulin et al., 2017b; de Vries et al., 2017) for
the generator. As for the architecture of the generator network used in the experiment, please see
Figure 14 for more detail. Our proposed projection model discriminator is equipped with a ‘projection layer’ that takes inner product between the embedded one-hot vector y and the intermediate
output (Figure 14a). As for the structure of the the concat model discriminator to be compared
against, we used the identical bulk architecture as the projection model discriminator, except that we
removed the projection layer from the structure and concatenated the spatially replicated embedded
conditional vector y to the output of third ResBlock. We also experimented with AC-GANs as the
current state of the art model. For AC-GANs, we placed the softmax layer classifier to the same
structure shared by concat and projection. For each method, we updated the generator 450K times,
and applied linear decay for the learning rate after 400K iterations so that the rate would be 0 at the
end. For the comparative experiments, we trained the model for 450K iterations, which was ample
for the training of concat to stabilize. AC-GANs collapsed prematurely before the completion of
450K iterations, so we reported the result from the peak of its performance ( 80K iterations). For all
experiments throughout, we used the training over 450K iterations for comparing the performances.
On a separate note, our method continued to improve even after 450K. We therefore also reported
the inception score and FID of the extended training (850K iterations) for our method exclusively.
See the table 1 for the exact figures.
We used inception score (Salimans et al., 2016) for the evaluation of the visual appearance of the
generated images. It is in general difficult to evaluate how ‘good’ the generative model is. Indeed,
however, either subjective or objective, some definite measures of ‘goodness’ exists, and essential two of them are ‘diversity’ and the sheer visual quality of the images. One possible candidate
for quantitative measure of diversity and visual appearance is FID (Heusel et al., 2017). We computed FID between the generated images and dataset images within each class, and designated the
values as intra FIDs. More precisely, FID (Heusel et al., 2017) measures the 2-Wasserstein distance between the two distributions qy and py , and is given by F (qy , py ) = kµqy − µpy k22 +
trace Cqy + Cpy − 2(Cqy Cpy )1/2 , where {µqy , Cqy }, {µpy , Cpy } are respectively the mean and
the covariance of the final feature vectors produced by the inception model (Szegedy et al., 2015)
from the true samples and generated samples of class y. When the set of generated examples have
collapsed modes, the trace of Cpy becomes small and the trace term itself becomes large. In order
to compute Cqy we used all samples in the training data belonging to the class of concern, and used
5000 generated samples for the computation of Cpy . We empirically observed in our experiments
that intra FID is, to a certain extent, serving its purpose well in measuring the diversity and the visual
quality.
To highlight the effectiveness of our inner-product based approach (projection) of introducing the
conditional information into the model, we compared our method against the state of the art ACGANs as well as the conventional incorporation of the conditional information via concatenation
2
in the preliminary experiments of the image geneation task on CIFAR-10 (Torralba et al., 2008) and
CIFAR-100 (Torralba et al., 2008), we confirmed that hidden concatenation is better than input concatenation
in terms of the inception scores. For more details, please see Table 3 in the appendix section.
3
For AC-GANs, the authors prepared a pair of discriminator and generator for each set classes of size 10.
5
Published as a conference paper at ICLR 2018
35
500
Projection
Concat
20
Projection
25
300
200
100
0
15
10
0.0
400
Projection
inception score
30
1.0
2.0 3.0
iteration
0 100 200 300 400 500
Concat
600
500
400
300
200
100
0
0 100 200 300 400 500 600
ACGANs
4.04.5
1e5
(a) concat vs projection
(b) AC-GANs vs projection
Figure 3:
Learning curves of Figure 4: Comparison of intra FID scores for projection
cGANs with concat and projection concat, and AC-GANs on ImageNet. Each dot corresponds
to a class.
on ImageNet.
Table 1: Inception score and intra FIDs on ImageNet.
Method
Inception Score
Intra FID
AC-GANs
concat
projection
28.5±.20
21.1±.35
29.7±.61
260.0
141.2
103.1
*projection (850K iteration)
36.8±.44
92.4
at hidden layer (concat). As we can see in the training curves Figure 3, projection outperforms
inception score than concat throughout the training. Table 1 compares the intra class FIDs and the
inception Score of the images generated by each method. The result shown here for the AC-GANs is
that of the model at its prime in terms of the inception score, because the training collapsed at the end.
We see that the images generated by projection have lower intra FID scores than both adversaries,
indicating that the Wasserstein distance between the generative distribution by projection to the
target distribution is smaller. For the record, our model performed better than other models on the
CIFAR10 and CIFAR 100 as well (See Appendix A).
Figure 10a and 10b shows the set of classes for which (a) projection yielded results with better
intra FIDs than the concat and (b) the reverse. From the top, the figures are listed in descending
order of the ratio between the intra FID score between the two methods. Note that when the concat
outperforms projection it only wins by a slight margin, whereas the projection outperforms concat
by large margin in the opposite case. A quick glance on the cases in which the concat outperforms
the projection suggests that the FID is in fact measuring the visual quality, because both sets looks
similar to the human eyes in terms of appearance. Figure 5 shows an arbitrarily selected set of
results yielded by AC-GANs from variety of zs. We can clearly observe the mode-collapse on this
batch. This is indeed a tendency reported by the inventors themselves (Odena et al., 2017). ACGANs can generate easily recognizable (i.e classifiable) images, but at the cost of losing diversity
and hence at the cost of constructing a generative distribution that is significantly different from the
target distribution as a whole. We can also assess the low FID score of projection from different
perspective. By construction, the trace term of intra FID measures the degree of diversity within the
class. Thus, our result on the intra FID scores also indicates that that our projection is doing better
in reproducing the diversity of the original. The GANs with the concat discriminator also suffered
from mode-collapse for some classes (see Figure 6). For the set of images generated by projection,
we were not able to detect any notable mode-collapse.
Figure 7a shows the samples generated with the projection model for the classes on which the
cGAN achieved lowest intra FID scores (that is the classes on which the generative distribution
were particularly close to target conditional distribution), and Figure 7b the reverse. While most
of the images listed in Figure 7a are of relatively high quality, we still observe some degree of
mode-collapse. Note that the images in the classes with high FID are featuring complex objects
like human; that is, one can expect the diversity within the class to be wide. However, we note that
we did not use the most complicated neural network available for the experiments presented on this
paper, because we prioritized the completion of the training within a reasonable time frame. It is
very possible that, by increasing the complexity of the model, we will be able to further improve the
6
Published as a conference paper at ICLR 2018
Figure 5: comparison of the images generated by (a) AC-GANs and (b) projection.
Figure 6: Collapsed images on the concat model.
visual quality of the images and the diversity of the distribution. In Appendix ??, we list images of
numerous classes generated by cGANs trained with our projection model.
Category Morphing With our new architecture, we were also able to successfully perform category morphism. When there are classes y1 and y2 , we can create an interpolated generator by
simply mixing the parameters of conditional batch normalization layers of the conditional generator
corresponding to these two classes. Figure 8 shows the output of the interpolated generator with
the same z. Interestingly, the combination is also yielding meaningful images when y1 and y2 are
significantly different.
Fine-tuning with the pretrained model on the ILSVRC2012 classification task. As we mentioned in Section 4, the authors of Plug and Play Generative model (PPGNs) (Nguyen et al., 2017)
were able to improve the visual appearance of the model by augmenting the cost function with that
of the label classifier. We also followed their footstep and augmented the original generator loss with
an additional auxiliary classifier loss. As warned earlier regarding this type of approach, however,
this type of modification tends to only improve the visual performance of the images that are easy
for the pretrained model to classify. In fact, as we can see in Appendix B, we were able to improve
the visual appearance the images with the augmentation, but at the cost of diversity.
5.2
S UPER - RESOLUTION
We also evaluated the effectiveness of (3) in its application to the super-resolution task. Put formally,
the super-resolution task is to infer the high resolution RGB image of dimension x ∈ RRH ×RH ×3
from the low resolution RGB image of dimension y ∈ RRL ×RL ×3 ; RH > RL . This task is very
much the case that we presumed in our model construction, because p(y|x) is most likely unimodal
even if p(x|y) is multimodal. For the super-resolution task, we used the following formulation for
discriminator function:
X
f (x, y; θ) =
(yijk Fijk (φ(x; θΦ ))) + ψ(φ(x; θΦ ); θΨ ),
(10)
i,j,k
where F (φ(x; θΦ )) = V ∗ φ(x; θΦ ) where V is a convolutional kernel and ∗ stands for convolution
operator. Please see Figure 15 in the appendix section for the actual network architectures we used
7
Published as a conference paper at ICLR 2018
(a) Generated images on the class with ‘low’ FID scores.
(b) generated images on the class with ‘high’ FID scores.
Figure 7: 128×128 pixel images generated by the projection method for the classes with (a) bottom five FID scores and (b) top five FID scores. The string and the value above each panel are
respectively the name of the corresponding class and the FID score. The second row in each panel
corresponds to the original dataset.
for this task. For this set of experiments, we constructed the concat model by removing the module in
the projection model containing the the inner product layer and the accompanying convolution layer
altogether, and simply concatenated y to the output of the ResBlock preceding the inner product
module in the original. As for the resolutions of the image datasets, we chose RH = 128 and RL =
32, and created the low resolution images by applying bilinear downsampling on high resolution
images. We updated the generators 150K times for all methods, and applied linear decay for the
learning rate after 100K iterations so that the final learning rate was 0 at 150K-th iteration.
Figure 9 shows the result of our super-resolution. The bicubic super-resolution is very blurry, and
concat result is suffering from excessively sharp and rough edges. On the other hand, the edges
of the images generated by our projection method are much clearer and smoother, and the image
itself is much more faithful to the original high resolution images. In order to qualitatively compare
the performances of the models, we checked MS-SSIM (Wang et al., 2003) and the classification
accuracy of the inception model on the generated images using the validation set of the ILSVRC2012
dataset. As we can see in Table 2, our projection model was able to achieve high inception accuracy
and high MS-SSIM when compared to bicubic and concat. Note that the performance of superresolution with concat model even falls behind those of the bilinear and bicubic super-resolutions
8
Published as a conference paper at ICLR 2018
Figure 8: Category morphing. More results are in the appendix section.
Figure 9: 32x32 to 128x128 super-resolution by different methods
in terms of the inception accuracy. Also, we used projection model to generate multiple batches of
images with different random values of z to be fed to the generator and computed the average of the
logits of the inception model on these batches (MC samples). We then used the so-computed average
logits to make prediction of the labels. With an ensemble over 10 seeds (10 MC in Table 2), we were
able to improve the inception accuracy even further. This result indicates that our GANs are learning
the super-resolution as an distribution, as opposed to deterministic function. Also, the success with
the ensemble also suggests a room for a new way to improve the accuracy of classification task on
low resolution images.
9
Published as a conference paper at ICLR 2018
Table 2: Inception accuracy and MS-SSIM on different super-resolution methods. We picked up
dataset images from the validation set.
6
Method
biliear
bicubic
concat
projection
projection (10 MC)
Inception Acc.(%)
MS-SSIM
23.1
0.835
31.4
0.859
11.0
0.829
35.2
0.878
36.4
-
C ONCLUSION
Any specification on the form of the discriminator imposes a regularity condition for the choice for
the generator distribution and the target distribution. In this research, we proposed a model for the
discriminator of cGANs that is motivated by a commonly occurring family of probabilistic models.
This simple modification was able to significantly improve the performance of the trained generator
on conditional image generation task and super-resolution task. The result presented in this paper is
strongly suggestive of the importance of the choice of the form of the discriminator and the design
of the distributional metric. We plan to extend this approach to other applications of cGANs, such
as semantic segmentation tasks and image to image translation tasks.
ACKNOWLEDGMENTS
We would like to thank the members of Preferred Networks, Inc., especially Richard Calland, Sosuke
Kobayashi and Crissman Loomis, for helpful comments. We would also like to thank Shoichiro
Yamaguchi, a graduate student of Kyoto University, for helpful comments.
10
Published as a conference paper at ICLR 2018
(a) Images better with projection than concat.
(b) Images better with concat than projection.
Figure 10: Comparison of concat vs. projection. The value attached above each panel represents
the achieved FID score.
11
Published as a conference paper at ICLR 2018
R EFERENCES
Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In ICML,
pp. 214–223, 2017.
Harm de Vries, Florian Strub, Jérémie Mary, Hugo Larochelle, Olivier Pietquin, and Aaron C Courville. Modulating early visual processing by language. In NIPS, pp. 6576–6586, 2017.
Emily Denton, Soumith Chintala, Arthur Szlam, and Rob Fergus. Deep generative image models using a
laplacian pyramid of adversarial networks. In NIPS, pp. 1486–1494, 2015.
Vincent Dumoulin, Ishmael Belghazi, Ben Poole, Alex Lamb, Martin Arjovsky, Olivier Mastropietro, and
Aaron Courville. Adversarially learned inference. In ICLR, 2017a.
Vincent Dumoulin, Jonathon Shlens, and Manjunath Kudlur. A learned representation for artistic style. In
ICLR, 2017b.
Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron
Courville, and Yoshua Bengio. Generative adversarial nets. In NIPS, pp. 2672–2680, 2014.
Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron Courville. Improved training
of wasserstein GANs. arXiv preprint arXiv:1704.00028, 2017.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In
CVPR, 2016a.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In
European Conference on Computer Vision, pp. 630–645. Springer, 2016b.
Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, Günter Klambauer, and Sepp
Hochreiter. Gans trained by a two time-scale update rule converge to a nash equilibrium. arXiv preprint
arXiv:1706.08500, 2017.
Taeksoo Kim, Moonsu Cha, Hyunsoo Kim, Jungkwon Lee, and Jiwon Kim. Learning to discover cross-domain
relations with generative adversarial networks. In ICML, pp. 1857–1865, 2017.
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015.
Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew
Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image superresolution using a generative adversarial network. In CVPR, 2017.
Jae Hyun Lim and Jong Chul Ye. Geometric GAN. arXiv preprint arXiv:1705.02894, 2017.
Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784,
2014.
Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative
adversarial networks. In ICLR, 2018.
Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & play generative
networks: Conditional iterative generation of images in latent space. In CVPR, 2017.
Sebastian Nowozin, Botond Cseke, and Ryota Tomioka. f-GAN: Training generative neural samplers using
variational divergence minimization. In NIPS, pp. 271–279, 2016.
Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxiliary classifier
GANs. In ICML, pp. 2642–2651, 2017.
Guim Perarnau, Joost van de Weijer, Bogdan Raducanu, and Jose M Álvarez. Invertible conditional gans for
image editing. In NIPS Workshop on Adversarial Training, 2016.
Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep convolutional
generative adversarial networks. In ICLR, 2016.
Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative
adversarial text to image synthesis. In ICML, pp. 1060–1069, 2016.
Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej
Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale
Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115(3):211–252, 2015.
doi: 10.1007/s11263-015-0816-y.
Masaki Saito, Eiichi Matsumoto, and Shunta Saito. Temporal generative adversarial nets with singular value
clipping. In ICCV, 2017.
Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved
techniques for training GANs. In NIPS, pp. 2226–2234, 2016.
Kumar Sricharan, Raja Bala, Matthew Shreve, Hui Ding, Kumar Saketh, and Jin Sun. Semi-supervised conditional GANs. arXiv preprint arXiv:1708.05789, 2017.
Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan,
Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, pp. 1–9, 2015.
Seiya Tokui, Kenta Oono, Shohei Hido, and Justin Clayton. Chainer: a next-generation open source framework
for deep learning. In Proceedings of workshop on machine learning systems (LearningSys) in the twentyninth annual conference on neural information processing systems (NIPS), 2015.
Antonio Torralba, Rob Fergus, and William T Freeman. 80 million tiny images: A large data set for nonparametric object and scene recognition. IEEE transactions on pattern analysis and machine intelligence, 30
(11):1958–1970, 2008.
12
Published as a conference paper at ICLR 2018
Dustin Tran, Rajesh Ranganath, and David M Blei. Deep and hierarchical implicit models. arXiv preprint
arXiv:1702.08896, 2017.
Zhou Wang, Eero P Simoncelli, and Alan C Bovik. Multiscale structural similarity for image quality assessment. In Asilomar Conference on Signals, Systems and Computers, pp. 1398–1402, 2003.
Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaolei Huang, Xiaogang Wang, and Dimitris Metaxas.
Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In ICCV,
2017.
Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using
cycle-consistent adversarial networks. In ICCV, 2017.
13
Published as a conference paper at ICLR 2018
A
R ESULTS OF CLASS CONDITIONAL IMAGE GENERATION ON CIFAR-10
AND CIFAR-100
As a preliminary experiment, we compared the performance of conditional image generation on
CIFAR-10 and CIFAR-100 3. For the discriminator and the generator, we reused the same architecture used in Miyato et al. (2018) for the task on CIFAR-10. For the adversarial objective functions,
we used (9), and trained both machine learners with the same optimizer with same hyper parameters
we used in Section 5. For our projection model, we added the projection layer to the discriminator in
the same way we did in the ImageNet experiment (before the last linear layer). Our projection model
achieved better performance than other methods on both CIFAR-10 and CIFAR-100. Concatenation
at hidden layer (hidden concat) was performed on the output of second ResBlock of the discriminator. We tested hidden concat as a comparative method in our main experiments on ImageNet,
because the concatenation at hidden layer performed better than the concatenation at the input layer
(input concat) when the number of classes was large (CIFAR-100).
To explore how the hyper-parameters affect the performance of our proposed architecture, we conducted hyper-parameter search on CIFAR-100 about the Adam hyper-parameters (learning rate α
and 1st order momentum β1 ) for both our proposed architecture and the baselines. Namely, we
varied each one of these parameters while keeping the other constant, and reported the inception
scores for all methods including several versions of concat architectures to compare. We tested with
concat module introduced at (a) input layer, (b) hidden layer, and at (c) output layer. As we can see
in Figure 11, our projection architecture excelled over all other architectures for all choice of the
parameters, and achieved the inception score of 9.53. Meanwhile, concat architectures were able
to achieve all 8.82 at most. The best concat model in term of the inception score on CIFAR-100
was the hidden concat with α = 0.0002 and β1 = 0, which turns out to be the very choice of the
parameters we picked for our ImageNet experiment.
Table 3: The performance of class conditional image generation on CIFAR-10 (C10) and CIFAR100 (C100).
Method
Inception score
C10
C100
C10
FID
C100
(Real data)
11.24
14.79
7.60
8.94
8.22
8.25
8.14
8.62
8.80
7.93
8.82
9.04
19.7
19.2
19.2
17.5
25.4
31.4
24.8
23.2
10
9
8
7
6
5
4
3
2
1
0.0001
Inception score
Inception score
AC-GAN
input concat
hidden concat
(ours) projection
Projection
Input concat
Hidden concat
Output concat
AC-GANs
0.0002
0.0005
(a) Varying α (fixing β1 = 0)
10
9
8
7
6
5
4
3
2
1
Projection
Input concat
Hidden concat
Output concat
AC-GANs
0.0
0.5
1
0.9
(b) Varying β1 (fixing α = 0.0002)
Figure 11: Inception scores on CIFAR-100 with different discriminator models varying hyperparameters (α and β1 ) of Adam optimizer.
14
Published as a conference paper at ICLR 2018
Figure 12: Effect of the finetuning with auxiliary classifier loss. Same coordinate in panel (a) and
(b) corresponds to same value of z
Table 4: Inception score and intra FIDs on ImageNet with a pretrained model on classification tasks
for ILSVRC2012 dataset. ‡Nguyen et al. (2017)
Method
Inception Score
Intra FID
‡
47.4
210
N/A
54.2
PPGNs
projection(finetuned)
B
O BJECTIVE FUNCTION WITH AN AUXILIARY CLASSIFIER COST
In this experiment, we followed the footsteps of Plug and Play Generative model (PPGNs) (Nguyen
et al., 2017) and augmented the original generator loss with an additional auxiliary classifier loss. In
particular, we used the losses given by :
h
h
ii
L G, D̂, p̂pre (y|x) = −Eq(y) Ep(z) D̂(G(z, y), y) − LC (p̂pre (y|G(z, y))) ,
(11)
where p̂pre (y|x) is the fixed model pretrained for ILSVRC2012 classification task. For the actual
experiment, we trained the generator with the original adversarial loss for the first 400K updates,
and used the augmented loss for the last 50K updates. For the learning rate hyper parameter, we
adopted the same values as other experiments we described above. For the pretrained classifier,
we used ResNet50 model used in He et al. (2016a). Figure 12 compares the results generated by
vanilla objective function and the results generated by the augmented objective function. As we
can see in Table 4, we were able to significantly outperform PPGNs in terms of inception score.
However, note that the images generated here are images that are easy to classify. The method
with auxiliary classifier loss seems effective in improving the visual appearance, but not in training
faithful generative model.
15
Published as a conference paper at ICLR 2018
C
M ODEL A RCHITECTURES
BN
Conv 3x3
ReLU
ReLU
Conv 3x3
Conv 3x3
BN
ReLU
ReLU
Conv 3x3
(a) ResBlock architecture for the
discriminator. Spectral normalization (Miyato et al., 2018) was
applied to each conv layer.
(b) ResBlock for the generator.
Figure 13: Architecture of the ResBlocks used in all experiments. For the generator generator’s
Resblock, conditional batch normalization layer (Dumoulin et al., 2017b; de Vries et al., 2017) was
used in place of the standard batch normalization layer. For the ResBlock in the generator for the
super resolution tasks that implements the upsampling, the random vector z was fed to the model by
concatenating the vector to the embedded low resolution image vector y prior to the first convolution
layer within the block. For the procedure of downsampling and upsampling, we followed the implementation by Gulrajani et al. (2017). For the discriminator, we performed downsampling (average
pool) after the second conv of the ResBlock. For the generator, we performed upsampling before
the first conv of the ResBlock. For the ResBlock that is performing the downsampling, we replaced
the identity mapping with 1x1 conv layer followed by downsampling to balance the dimension. We
did the essentially same for the Resblock that is performing the upsampling, except that we applied
the upsampling before the 1x1 conv.
16
Published as a conference paper at ICLR 2018
Linear
z
Inner product
Linear, 1024x4x4
Global sum pooling
Embed
ResBlock, Up, 1024
ReLU
y
ResBlock, Up, 512
ResBlock, 1024
ResBlock, Up, 256
ResBlock, Down, 1024
ResBlock, Up, 128
ResBlock, Down, 512
ResBlock, Up, 64
ResBlock, Down, 256
BN, ReLU, Conv 3x3
ResBlock, Down, 128
Tanh
ResBlock, Down, 64
x
x
(b) Generator
(a) Discriminator
Figure 14: The models we used for the conditional image generation task.
17
y
Published as a conference paper at ICLR 2018
y
Linear
Global sum pooling
ReLU
ResBlock, 256
ResBlock, 1024
ResBlock, 256
ResBlock, Down, 1024
ResBlock, 256
ResBlock, Down, 512
ResBlock, 256
ResBlock, Down, 128
ResBlock, Up, 128
Inner product
ResBlock, Down, 256
ResBlock, 128
Conv 3x3
y
ResBlock, 128
ResBlock, Up, 64
ResBlock,64
ResBlock, Down, 128
BN, ReLU, Conv 3x3
ResBlock, 64
Tanh
ResBlock, Down, 64
x
x
(b) Generator
(a) Discriminator
Figure 15: The models we used for the super resolution task.
18
z
z’
Published as a conference paper at ICLR 2018
D
R ESULTS OF C ATEGORY M ORPHING
(a) to Persian cat
(b) to Electric locomotive
(c) to Alp
(d) to Castle
(e) to Chiffonier
Figure 16: Dog (Lhasa apso) to different categories
19
Published as a conference paper at ICLR 2018
(a) Hip to Yellow lady’s slipper
(b) Pirate ship to Yawl
(c) Yurt to Castle
(d) Chiffonier to Chinese cabinet
(e) Valley to Sandbar
Figure 17: Morphing between different categories
20
Published as a conference paper at ICLR 2018
E
M ORE R ESULTS WITH SUPER - RESOLUTION
Figure 18: 32x32 to 128x128 super-resolution results
21
| 1cs.CV
|
Estimating Mixture Entropy with Pairwise Distances
Artemy Kolchinsky
1,∗
and Brendan D. Tracey
1,2,†
1
arXiv:1706.02419v3 [cs.IT] 8 Nov 2017
2
Santa Fe Institute
Massachusetts Institute of Technology
∗
[email protected]
†
[email protected]
Mixture distributions arise in many parametric and non-parametric settings, for example in
Gaussian mixture models and in non-parametric estimation. It is often necessary to compute
the entropy of a mixture, but in most cases this quantity has no closed-form expression,
making some form of approximation necessary. We propose a family of estimators based on
a pairwise distance function between mixture components, and show that this estimator class
has many attractive properties. For many distributions of interest, the proposed estimators
are efficient to compute, differentiable in the mixture parameters, and become exact when the
mixture components are clustered. We prove this family includes lower and upper bounds
on the mixture entropy. The Chernoff α-divergence gives a lower bound when chosen as
the distance function, with the Bhattacharyaa distance providing the tightest lower bound
for components that are symmetric and members of a location family. The Kullback-Leibler
divergence gives an upper bound when used as the distance function. We provide closed-form
expressions of these bounds for mixtures of Gaussians, and discuss their applications to the
estimation of mutual information. We then demonstrate that our bounds are significantly
tighter than well-known existing bounds using numeric simulations. This pairwise estimator
class is very useful in optimization problems involving maximization/minimization of entropy
and mutual information, such as MaxEnt and rate distortion problems.
I.
INTRODUCTION
A mixture distribution is a probability distribution whose density function is a weighted sum of
individual densities. Mixture distributions are a common choice for modeling probability distributions, in both parametric settings, for example learning a mixture of Gaussians statistical model,
and non-parametric settings, such as kernel density estimation.
It is often necessary to compute the differential entropy [6] of a mixture distribution, which is
a measure of the inherent uncertainty in the outcome of X. Entropy estimation arises in image
retrieval tasks [10], image alignment and error correction [33], speech recognition [3, 14], analysis
of debris spread in rocket launch failures [2], and many other settings. Entropy also arises in
optimization contexts [28–30, 33], where it is minimized or maximized under some constraints (e.g.,
MaxEnt problems). Finally, entropy also plays a central role in minimization or maximization of
mutual information, such as in problems related to rate distortion [20].
Unfortunately, in most cases, the entropy of a mixture distribution has no known closed-form
expression. This is true even when the entropy of each component distribution does have a known
closed-form expression. For instance, the entropy of a Gaussian has a well-known form, while the
entropy of a mixture of Gaussians does not. As a result, the problem of finding a tractable and
accurate estimate for mixture entropy has been described as “a problem of considerable current
interest and practical significance” [35].
One way to approximate the entropy is with Monte Carlo (MC) sampling. MC sampling provides an unbiased estimate of mixture entropy, and this estimate can become arbitrarily accurate
by increasing the number of MC samples. Unfortunately, MC sampling is very computationally
intensive as for each sample, the (log) probability of the sample location must be computed under
every component in the mixture. MC sampling typically requires a large number of samples to
2
estimate entropy, especially in high-dimensions. Sampling is thus typically impractical, especially
for optimization problems where for every parameter change a new entropy estimate is required.
An alternative to Monte Carlo sampling is to use an analytic estimator of mixture entropy.
Analytic estimators have estimation bias, but are much more computationally efficient. There are
several existing analytic estimators of entropy, discussed in-depth below. To summarize, however,
commonly-used estimators have significant drawbacks: they have large bias relative to the true
entropy, and/or they are invariant to the amount of “overlap” between mixture components. For
example, many estimators do not depend on the locations of the means in a Gaussian mixture
model.
In this paper, we introduce a novel family of estimators for the mixture entropy. Each member of
this family is defined via a pairwise-distance function between component densities. The estimators
in this family have several attractive properties. They are computationally efficient, as long as the
pairwise-distance function and the entropy of each component distribution are easy to compute.
The estimation bias of any member of this family is bounded by a constant. The estimator is
continuous and smooth and is therefore useful for optimization problems. In addition, we show that
the Chernoff α-divergence (i.e., a scaled Rényi divergence), used as a pairwise-distance function,
gives a lower-bound on the mixture entropy; in particular, the Bhattacharrya distance (α = 0.5)
provides the tightest lower bound of all Chernoff α-divergences when the mixture components are
symmetric and belong to a location family (such as a mixture of Gaussians with equal covariances).
We also show the Kullback-Leibler [KL] divergence, when used as a pairwise-distance function,
gives an upper-bound on the mixture entropy. Finally, our family of estimators can compute the
exact mixture entropy when the component distributions are grouped into well-separated clusters,
a property not shared by other analytic estimators of entropy. In particular, the bounds mentioned
above converge to the same value for well-separated clusters.
A layout of the paper is as follows. We first review mixture distributions and entropy estimation
in Section II. We then present the pairwise distance estimator in Section III, prove bounds on
the error of any estimator in this class, and show distance functions that bound the entropy as
discussed above. In Section IV, we consider the special case of mixtures of Gaussians, and give
explicit expressions for lower and upper bounds on the mixture entropy. When all the Gaussian
components have the same covariance matrix, we show that these bounds have particularly simple
expressions. Then in Section V, we consider the closely related problem of estimating the mutual
information between two random variables, and show that our estimators and bounds can be directly
used to estimate the mutual information. Specializing to the Gaussian case, we show how this can
be used to bound the mutual information across a type of Additive White Noise Gaussian channel.
Finally, in Section VI, we run numerical experiments and compare the performance of our lower
and upper bounds relative to existing estimators. We consider both mixtures of Gaussians and
mixtures of uniform distributions.
II.
BACKGROUND AND DEFINITIONS
We consider the differential entropy, defined as
Z
H(X) := −
pX (x) ln pX (x) dx ,
3
of the mixture distribution pX ,
pX (x) =
N
X
ci pi (x) ,
i=1
where ci indicates the weight of component i (ci ≥ 0,
of component i.
P
i ci
= 1) and pi is the probability density
We can treat the set of component weights as a discrete random variable C with outcomes 1..N ,
where Pr(C = i) = ci . Consider the mixed joint distribution of the discrete random variable C and
the continuous random variable X,
pX,C (x, i) = pi (x)ci ,
and observe that the common identities for conditional and joint entropy still apply [21],
H (X, C) = H (X|C) + H (C) = H (C|X) + H (X)
where we use H for discrete and differential entropy interchangeably. Here the conditional entropies
are
Z
X
pi (x)ci
dx .
H(X|C) =
ci H(pi )
and
H(C|X) = pi (x)ci log
pX (x)
i
Using elementary results from information theory [6], H(X) can be bounded from below by
H(X) ≥ H(X|C) ,
(1)
since conditioning can only decrease entropy. Similarly, H(X) can be bounded from above by
H(X) ≤ H(X, C) = H(X|C) + H(C) ,
(2)
following from H(X) = H(X, C) − H(C|X) and the non-negativity of the conditional discrete
entropy H(C|X). This upper bound on the mixture entropy was previously proposed by Huber et
al. [16].
It is easy to see that the bound in Eq. (1) is tight when all the components have the same
distribution, since then H(pX ) = H(pi ) for all i. The bound in Eq. (2) becomes tight when
H(C|X) = 0, i.e., when any sample from pX uniquely determines the component identity C. This
occurs when the different mixture components have non-overlapping supports, pi (x) > 0 =⇒
pj (x) = 0 for all x and i 6= j. More generally, the bound of Eq. (2) becomes increasingly tight as
the mixture distributions move farther apart from one another.
Assuming that H(pi ), the entropy of each individual component density, has a simple closed
form expression, the bounds in Eq. (1) and Eq. (2) can be computed quickly. However, neither
bound depends on the “overlap” between components. For instance, in a Gaussian mixture model
these bounds are invariant to changes in the component means. The bounds are thus unsuitable for
many problems; for instance, in optimization one typically tunes parameters to adjust component
means, but the above entropy bounds remain the same for regardless of mean location.
There are two other estimators of the mixture entropy that should be mentioned. The first
estimator is based on kernel density estimation [11, 19]. It estimates the entropy using the average
4
probability of the component means, µi ,
ĤKDE (X) := −
X
ci ln
X
i
cj pj (µi ) .
(3)
j
The second estimator is a lower bound that is derived using Jensen’s inequality, −E [ln f (X)] ≥
− ln [E(f (X))], giving
Z X
X
X
X Z
H(X) := −
pi (x)pj (x) dx =: ĤELK . (4)
ci pi (x) ln
cj pj (x) dx ≥ −
ci ln
cj
i
j
i
j
R
In the literature, the term pi (x)pj (x) dx has been referred to as the “Cross Information Potential”
[26, 34] and the “Expected Likelihood Kernel” [17, 18] [ELK] (we use this second acronym to label
this estimator). When the component distributions are Gaussian, pi (x) := N (x; µi , Σi ), the ELK
has a simple closed-form expression, giving the lower bound
X
X
ci ln
cj N (µi ; µj , Σi + Σj ) .
(5)
H(X) ≥ −
i
j
This lower bound was previously proposed for Gaussian mixtures in [16]. Both of the estimators
in Eq. (3) and Eq. (5) are computationally efficient, continuous and differentiable, and depend on
component overlap, making them suitable for optimization. However, as will be shown via numerical
experiments (Section VI), they exhibit significant underestimation bias. However, we will also show
that for Gaussian mixtures with equal covariance, Eq. (3) plus an additive constant is actually an
estimator in our proposed class.
III.
ESTIMATORS BASED ON PAIRWISE-DISTANCES
A.
Overview
Let D(pi kpj ) be some (generalized) distance function between probability densities pi and pj .
Formally, we assume that D is a premetric, meaning that it is non-negative and D(pi kpj ) = 0 if
pi = pj . We do not assume that D is symmetric, nor that it obeys the triangle inequality, nor that
it is strictly greater than 0 when pi 6= pj .
For any allowable distance function D, we propose the following entropy estimator:
X
X
ĤD (X) := H(X|C) −
ci ln
cj exp(−D(pi kpj )) .
i
(6)
j
This estimator can be efficiently computed if the entropy of each component and D(pi kpj ) for all
i, j have simple closed-form expressions. There are many distribution-distance function pairs that
satisfy these conditions (e.g., Kullback-Leibler divergence, Renyi divergences, Bregman divergences,
f-divergences, etc. for Gaussian, uniform, exponential, etc.) [1, 4, 5, 7, 9].
It is straightforward to show that for any D, ĤD falls between the bounds of Eq. (1) and Eq. (2),
H(X|C) ≤ ĤD (X) ≤ H(X, C) .
(7)
5
To do so, consider the “smallest” and “largest” allowable distance functions,
(
0 if pi = pj
Dmin (pi kpj ) = 0
and
Dmax (Xi kXj ) =
∞ otherwise
(8)
For any D and pi , pj , Dmin (pi kpj ) ≤ D(pi kpj ) ≤ Dmax (pi kpj ), thus
ĤDmin (X) ≤ ĤD (X) ≤ ĤDmax (X)
P
Plugging Dmin into Eq. (6) (noting that j cj = 1) gives ĤDmin (X) = H(X|C), while plugging
Dmax into Eq. (6) gives
X
X
X
ĤDmax = H (X|C) −
ci ln ci +
cj e−Dmax (Xi kXj ) ≤ H (X|C) −
ci ln ci = H(X, C)
i
i
j6=i
These two inequalities yield Eq. (7). The true entropy, as shown in Section II, also obeys H (X|C) ≤
H (X) ≤ H (X, C). The magnitude of the bias of ĤD is thus bounded by
ĤD (X) − H(X) ≤ H (X, C) − H (X|C) = H(C) .
In the next two subsections, we improve upon the bounds suggested in Eq. (1) and Eq. (2), by
examining bounds induced by particular distance functions.
B.
Lower Bound
Consider the “Chernoff α-divergence” [7, 22] for some real-valued α, defined as
Z
Cα (pkq) := − ln pα (x)q 1−α (x) dx .
(9)
Note that Cα (pkq) = (1 − α)Rα (pkq), where Rα is Rényi divergence of order α [32].
We show that for any α ∈ [0, 1], ĤCα (X) is a lower bound on the entropy (for α ∈
/ [0, 1], Cα is
not a valid distance function, see the Appendix). To do so, note that
H(X) = H(X|C) + M I(X; C)
(10)
where M I is mutual information. Using a derivation from [12], we bound M I from below,
Z X
pX (x)
dx
M I(X; C) = −
ci pi (x) ln
pi (x)
i
!
P
Z X
Z X
1−α
c
p
(x)
j
j
pX (x)
j
=−
ci pi (x) ln
dx
ci pi (x) ln
αP
1−α dx −
p
(x)
c
p
(x)
pi (x)1−α
i
j j j
i
i
!
P
Z X
1−α
(a)
j cj pj (x)
≥−
ci pi (x) ln
dx
1−α
p
(x)
i
i
Z
X
X
(b)
≥−
ci ln pi (x)α
cj pj (x)1−α dx
i
j
6
=−
X
i
ci ln
X
cj e−Cα (Xi kXj ) .
(11)
j
The inequalities (a) and (b) follow from Jensen’s inequality. This inequality is used directly in (b),
while in (a) it follows from
Z X
Z X
pX (x)
pX (x)
dx
≥
−
ln
ci pi (x)
−
ci pi (x) ln
P
α
1−α
αP
1−α dx
p
(x)
c
p
(x)
p
(x)
c
p
(x)
j
j
j
j
i
i
j
j
i
i
Z X
pX (x)
= − ln
ci pi (x)1−α P
1−α dx
j cj pj (x)
i
X Z
= − ln
ci pX (x) dx = 0
i
Combining gives
H(X) ≥ ĤCα (X) .
This shows that using Cα as a distance function gives a lower bound on the mixture entropy for
any α ∈ [0, 1]. For a general mixture distribution, one could optimize over the value of α to find
the tightest lower bound. We can show that the tightest bound is achieved α = 0.5 in the special
case when all of the mixture components pi are symmetric and come from a location family,
pi (x) = b(x − µi ) = b(µi − x)
Examples of this situation include mixtures of Gaussians with the same covariance (“homoscedastic”
mixtures), multivariate t-distributions with the same covariance, location-shifted bounded uniform
distributions, most kernels used in kernel density estimation, etc.
To show that α = 0.5 is optimal, first define the Chernoff α-coefficient as
Z
cα (pi kpj ) := pi (x)α pj (x)1−α dx .
(Hence, the Chernoff α-divergence is the negative logarithm of the Chernoff α-coefficient.) We show
that for any pair pi , pj of symmetric distributions from a location family, cα (pj kpj ) is minimized by
α = 0.5. This means that all pairwise distances Cα are maximized by α = 0.5, and therefore the
entropy estimator HCα (Eq. (6)) is maximized by α = 0.5.
Define a change of variables
y := µi + µj − x ,
which gives x − µi = µj − y and x − µj = µi − y. This allows us to write the Chernoff α-coefficient
as
Z
cα (pi kpj ) = pi (x)α pj (x)1−α dx
Z
= b(x − µi )α b(x − µj )1−α dx
Z
(a)
= b(µj − y)α b(µi − y)1−α dy
Z
(b)
= b(y − µj )α b(y − µi )1−α dy
7
=c1−α (pi kpj )
where in (a) we’ve substituted variables, and in (b) we used the assumption that b(x) = b(−x).
Since cα (pi kpj ) = c1−α (pi kpj ), cα is symmetric in α about α = 0.5. In the Appendix, we show that
cα (pkq) is everywhere convex in α. Together, this means that cα (pi kpj ) must achieve a minimum
value at α = 0.5. This value of α is a special case of the Chernoff α-coefficient, known as the
Bhattacharyya coefficient, with the corresponding Bhattacharyya distance [8] defined as,
Z p
BD(pkq) := − ln
p(x)q(x)dx = C0.5 (pkq)
Since any Chernoff α-divergence is a lower bound for the entropy, we write the particular case of
Bhattacharyya-distance lower bound as
X
X
H(X) ≥ ĤBD (X) := H(X|C) −
ci ln
cj e−BD(pi kpj ) .
(12)
i
C.
j
Upper bound
The Kullback-Leibler [KL] divergence [6], defined as
Z
p(x)
KL(pkq) := p(x) ln
dx ,
q(x)
provides an upper bound for the estimate of mixture entropy. We show this as follows:
X
X
H(X) = −
ci Epi ln
cj pj (x)
i
(a)
≤−
X
=−
ci ln
X
i
j
X
ci ln
X
i
=
j
X
cj eEpi [ln pj (x)]
cj e−H(pi kpj )
j
ci H(Xi ) −
X
i
ci ln
X
i
cj e−KL(pi kpj )
j
where we use H(·k·) to indicate the cross-entropy function, and employ the identity H(pi kpj ) =
H(pi ) + KL(pi kpj ). The inequality in step (a) uses a variational lower bound on the expectation of
a log-sum [14, 24],
X
X
E ln
Zj ≥ ln
exp (E [ln Zj ]) .
j
j
Combining gives the upper bound
H(X) ≤ ĤKL := H(X|C) −
X
i
ci ln
X
j
cj e−KL(pi kpj ) .
(13)
8
D.
Exact estimation in the “clustered” case
In the previous sections, we described lower and upper bounds on the mixture entropy in the
general case. In this section, we show that in the special case when the mixture components are
“clustered” into well-separated groups, these lower and upper bounds become equal, and so the
pairwise-distance estimate of the entropy is tight. By “clustered”, we mean that there is a grouping
of component distributions such that distributions in the same group are approximately the same,
while distributions assigned to different groups are very different from one another. Though this
situation may seem like an edge case, clustered distributions do arise in mixture estimation, e.g.,
when there are repeated data points, and we have encountered it as solutions to informationtheoretic optimization problems [20]. Note that the number of groups is arbitrary, and therefore
this situation includes the extreme cases of a single group (all component distributions are nearly
the same) as well as N different groups (all component distributions are very different).
Formally, let the function g(i) indicate the group of component i. We define that the components
are “clustered” with respect to grouping g iff KL(pi kpj ) ≤ κ whenever g(i) = g(j) forP
some small κ,
and BD(pi kpj ) ≥ β whenever g(i) 6= g(j) some large β. We use the notation pG (k) = i δ(g(i), k)ci
to indicate the sum of the weights of the components in group k.
We show that when κ is small and β is large, both ĤCα for α ∈ (0, 1] and ĤKL approach
X
H(X|C) −
pG (k) ln pG (k) .
k
Since one is a lower bound and one is an upper bound on the true entropy, the estimators become
exact as they converge in value.
Recall that BD = C0.5 . For α ∈ (0, 1], α−1 Cα (pkq) is a monotonically decreasing function for
α ∈ (0, 1] [27], meaning that Cα ≥ 2αBD(pkq) for α ∈ (0, 0.5]. In addition, (1 − α)−1 Cα (pkq) is a
monotonically increasing function for α > 0 [27], thus Cα ≥ 2(1 − α)BD(pkq) for α ∈ [0.5, 1]. Using
the assumption that BD(pi kpj ) ≥ β and combining gives the bound
Cα (pkq) ≥ (1 − |1 − α|)β
for α ∈ (0, 1], leading to
ĤCα (X) := H(X|C) −
X
ci ln
X
i
≥ H(X|C) −
X
≥ H(X|C) −
X
cj exp(−Cα (pi kpj ))
j
X
X
ci ln
δ(g(i), g(j))ci +
(1 − δ(g(i), g(j)))cj exp(−Cα (pi kpj ))
i
j
j
pG (k) ln [pG (k) + (1 − pG (k)) exp(−(1 − |1 − α|)β)] .
k
In the second line, for i, j in the same group, we’ve used that Cα (pi kpj ) ≥ 0.
For the upper bound ĤKL , we use that KL(pi kpj ) ≤ κ for i and j in the same group, and
otherwise exp(−KL(pi kpj )) ≥ 0. This gives the bound
X
X
ĤKL (X) := H(X|C) −
ci ln
cj exp(−KL(pi kpj ))
i
j
X
X
X
≤ H(X|C) −
ci ln
δ(g(i), g(j))ci e−κ +
(1 − δ(g(i), g(j)))cj exp(−KL(pi kpj ))
i
j
j
9
≤ H(X|C) −
X
pG (k) ln pG (k)e−κ
k
The difference between the bounds is bounded by
ĤKL (X) − ĤCα (X) ≤ κ +
X
≤κ+
X
k
(1 − pG (k)) exp(−(1 − |1 − α|)β)
pG (k) ln 1 +
pG (k)
pG (k)
k
(1 − pG (k)) exp(−(1 − |1 − α|)β)
pG (k)
= κ + (|G| − 1) exp(−(1 − |1 − α|)β)
where |G| is the number of groups. Thus, the difference decreases at least linearly in κ and exponentially in β. This shows that in the clustered case, when κ ≈ 0 and β is very large, our lower
and upper bounds become exact.
It also shows that any distance measure bounded between BD and KL also gives an exact
estimate of entropy in the clustered case. Furthermore, the idea behind this proof can be extended
to estimators induced by other bounding distance function, beyond BD and KL, so as to show that
a particular estimator converges to an exact entropy estimate in the clustered case. Note, however,
that for some distribution-distance pairs the components will never be considered as “clustered”;
e.g., the α-Chernoff distance for α = 0 between any two Gaussians is 0, and so a Gaussian mixture
distributions will never be considered clustered according to this distance.
IV.
GAUSSIAN MIXTURES
Gaussians are very frequently used as components in a mixture distributions. Our family of
estimators is well-suited to estimating the entropies of Gaussian mixtures, since the entropy of a
d-dimensional Gaussian Xi ∼ N (µi , Σi ) has a simple closed-form expression,
H (Xi ) =
1
[ln |Σi | + d ln 2π + d] ,
2
(14)
and because there are many distance functions between Gaussians with closed-form expressions
(KL divergence, the Chernoffα-divergences, Wasserstein distance, etc.). In this section, we consider
Gaussian mixtures and state explicit expression for the lower and upper bounds derived in the
previous section. We also consider these bounds in the special case where all Gaussian components
have the same covariance matrix (homoscedastic mixtures).
We first consider the lower bound, ĤCα , based on the Chernoff α-divergence distance function.
For two multivariate Gaussians p1 ∼ N (µ1 , Σ1 ) and p2 ∼ N (µ2 , Σ2 ), this distance is defined as
[13, Eq. 6]
(1 − α) α
1
|(1 − α)Σ1 + αΣ2 |
−1
T
Cα (p1 kp2 ) =
(µ1 − µ2 ) ((1 − α)Σ1 + αΣ2 ) (µ1 − µ2 ) + ln
.
2
2
|Σ1 |1−α |Σ2 |α
(15)
(As a warning, note that most sources show erroneous expressions for the Chernoff and/or Rényi
α-divergence between two multivariate Gaussians including [25, p. 33], [15, Eq. 2.5], [22, Eq. 25],
[9], [23, Eq. 36], and even a late draft of this manuscript.)
For the upper bound HKL , the KL divergence between two multivariate Gaussians p1 ∼
10
N (µ1 , Σ1 ) and p2 ∼ N (µ2 , Σ2 ) is
i
1h
−1
KL(p1 kp2 ) =
ln |Σ2 | − ln |Σ1 | + (µ1 − µ2 )T Σ−1
(µ
−
µ
)
+
tr
Σ
Σ
−
d
.
1
1
2
2
j
2
(16)
The appropriate lower and upper bounds can be computed by plugging in Eq. (15) and Section IV
into Eq. (6).
These bounds have simple forms when all of the mixture components have equal covariance
matrices, i.e., Σi = Σ for all i. First, define a transformation in which each Gaussian component pj
is mapped to a different Gaussian p̃j,α , which has the same mean but where the covariance matrix
1
is rescaled by α(1−α)
,
pj := N (µj , Σ)
7→
p̃j,α
:= N µj ,
1
Σ .
α(1 − α)
Then, the lower bound of Eq. (11) can be written as
ĤCα =
X
X
d d
+ ln (α(1 − α)) −
ci ln
cj p̃j,α (µi ) .
2 2
i
j
This is derived by combining the expressions for Cα , Eq. (15), the entropy of a Gaussian, Eq. (14),
and the Gaussian density function. For a homoscedastic mixture, the tightest lower bound among
the Chernoff α-divergences is given by α = 0.5, corresponding to the Bhattacharyya distance,
ĤBD =
X
d d 1 X
+ ln −
ci ln
cj p̃j,0.5 (µi ) .
2 2 4
i
j
(This is derived above in Section III B.)
For the upper bound, when all Gaussians have the same covariance matrix, we again combine
the expressions for KL, Section IV, the entropy of a Gaussian, Eq. (14), and the Gaussian p.d.f. to
give
ĤKL (X) =
X
d X
−
ci ln
cj pj (µi ) .
2
i
j
Note that this is exactly the expression for the kernel density estimator ĤKDE (Eq. (3)), plus a dimensional correction. Thus, surprisingly ĤKDE is a reasonable entropy estimator for homoscedastic
Gaussian mixtures, since it is only an additive constant away from KL-distance based estimator
ĤKL (which has various beneficial properties, as described above). This may explain why ĤKDE
has been used effectively in optimization contexts [28–30, 33], where the additive constant is often
irrelevant, despite lacking a principled justification in terms of being a a bound on entropy.
V.
ESTIMATING MUTUAL INFORMATION
It is often of interest, for example in rate distortion problems and related problems [20], to
calculate the mutual information across a communication channel,
M I(X; U ) = H(X) − H(X|U ),
11
where U is the distribution of signals sent across the channel, and X is the distribution of messages
received on the other end of the channel. As with mixture distributions, it is often easy to compute
H(X|U ), the entropy of the received signal given the sent signal (i.e., the distribution of noise on
the channel). The entropy of the received signals, H(X), on the other hand, is often difficult to
compute.
In some cases, the distribution of U may be well approximated by a mixture model. In this case,
we can estimate the entropy of the received signals, H(X), using our pairwise distance estimators,
as discussed in Section III. In particular, we have the lower bound
X
X
M I(X; U ) ≥ −
ci ln
cj exp(−Cα (pi kpj )),
i
j
and the upper bound,
M I(X; U ) ≤ −
X
i
ci ln
X
cj exp(−KL(pi kpj )),
j
where pi is the density of component i, and noting that the H(X|U ) terms cancel in the expression.
P
P
This also illuminates that the actual pairwise portion of the estimator, − i wi ln j wj exp(−D(pi kpj ))
is a measure of the mutual information between the random variable specifying the component
and the random variable generated from the mixture distribution. If the components are identical,
this mutual information is zero, since knowing the component tells you nothing about the value of
X. On the other hand, when all of the components are very different from one another, knowing
the component that generated X is very informative, giving the maximum amount of information,
H(C).
As a practical example, consider a scenario in which U is a random variable representing outside temperature on any particular day. This temperature is measured with a thermometer with
Gaussian measurement noise (the “Additive White Noise Gaussian channel”). This gives our measurement distribution
X = U + N 0, Σ0 .
If the actual temperature distribution is (approximately or exactly) distributed as a mixture of M
Gaussians, each one having mixture weight ci , mean µi , and covariance matrix Σi , then X will also
be distributed as a mixture of M Gaussians, each with weight ci , mean µi , and covariance matrix
Σ0j := Σi + Σ0 . We can then use our estimators to estimate the mutual information between actual
temperature, U , and thermometer measurements, X.
VI.
NUMERICAL RESULTS
In this section, we run numerical experiments and compare estimators of mixture entropy under
a variety of conditions. We consider two different types of mixtures, mixtures of Gaussians and
mixtures of uniform distributions, for a variety of parameter values. We evaluate the following
estimators:
1. The true entropy, H(X), as estimated by a Monte-Carlo sampling of the mixture model.
2,000 samples were used for each MC estimate for the mixtures of Gaussians, and 5,000
samples were used for the mixtures of uniform distributions.
2. Our proposed upper-bound, based on the KL divergence, ĤKL (Eq. (13))
12
3. Our proposed lower-bound, based on the Bhattacharyya distance, ĤBD (Eq. (12))
4. The kernel density estimate based on the component means, ĤKDE (Eq. (3))
5. The lower bound based on the “Expected Likelihood Kernel”, ĤELK (Eq. (4))
6. The lower bound based on the conditional entropy, H(X|C) (Eq. (1))
7. The upper bound based on the joint entropy, H(X, C) (Eq. (2)).
We show the values of the estimators 1-5 as line plots, while the region between the conditional (6)
and joint entropy (7) is shown in shaded green. The code for these figures can be found at http://
www.github.com/btracey/mixent, and uses the Gonum numeric library http://www.gonum.org.
A
B
Entropy (nats)
Entropy (nats)
18
15
12
9
18
15
12
9
-2
-1
0
1
0
2
Ln(Mean Spread)
D
18
Entropy (nats)
Entropy (nats)
C
4
6
Ln(Covariance Similarity)
15
12
40
30
Monte Carlo
Pair Dist: KL
Pair Dist: Bhat.
KDE
Exp. Lik. Ker.
[H(X|C), H(X,C)]
20
10
9
0
-2
-1
0
Ln(Cluster Spread)
8
16
24
Dimension
Figure 1: Entropy estimates for a mixture of a 100 Gaussians. In each plot, the vertical axis shows the
entropy of the distribution, and the horizontal axis changes a feature of the components: A) the distance
between means is increased; B) the component covariances become more similar (at the right side of the
plot, all Gaussians have covariance ≈ I); C) the components are grouped into 5 “clusters”, and the distance
between the locations of the clusters is increased; D) the dimension is increased.
A.
Mixture of Gaussians
In the first experiment, we evaluate the estimators on a mixture of randomly placed Gaussians,
and look at their behavior as the distance between the means of the Gaussians increases. The
mixture is composed of 100 10-dimensional Gaussians, each Gaussian Xi ∼ N (µi , I), with means
are sampled from µi ∼N (0, σI). Fig. 1A depicts the change in estimated entropy as the means grow
farther apart, in particular a function of ln(σ). We see that our proposed bounds are the closest
to the true entropy over the whole range, and in the extremes, we see that our proposed bounds
13
approach the exact value of the true entropy. This is as expected, since as σ → 0 all of the Gaussian
mixture components become identical, and as σ → ∞ all of the Gaussian components grow very
far apart, approaching the case where each Gaussian is in its own “cluster”. The ELK lower bound
is a strictly worse estimate than ĤBD , in this experiment. As expected, we also observe that the
KDE estimator differs by exactly d/2 from the KL estimator.
In the second experiment, we evaluate the entropy estimators as the covariance matrices change
from less to more similar. We again generate 100 10-dimensional Gaussians. Each Gaussian is
1
distributed as Xi ∼ N (µi , Σi ), where now µi ∼N (0, I) and Σi ∼ W( d+n
I, n), where W is a
Wishart distribution with n degrees of freedom. Fig. 1B compares the the estimators with the true
entropy as a function of ln(n). When n is small, the Wishart distribution is broad and the covariance
matrices differ significantly from one another, while as n → ∞, all the covariance matrices become
close to the identity I. Thus, for small n, we essentially recover a “clustered” case, in which every
component is in its own cluster and our lower and upper bounds giving highly accurate estimates.
For large n, we converge to the σ = 1 case of the first experiment.
In the third experiment, we again generate a mixture of 100 10-dimensional Gaussians. Now,
however, the Gaussians are grouped into 5 “clusters”, with each Gaussian component randomly
assigned to one of the clusters. We use g(i) ∈ {1..5} to indicate the group of each Gaussian’s
component i ∈ {1..100}, and each of the 100 Gaussians is distributed as Xi ∼ N (µ̃g(i) , I). The
cluster means µ̃k for k ∈ {1..5} are drawn from N (0, σI). The results are depicted in Fig. 1C. In the
first experiment, we saw that the joint entropy H(X, C) became an increasingly better estimator as
the Gaussians grew increasingly far apart. Here, however, we see that there is a significant difference
between H(X, C) and the true entropy, even as the groups become increasingly separated. Our
proposed bounds, on the other hand, provide accurate estimates of the entropy across the entire
parameter sweep. As expected, they become exact in the limits of all clusters being at the same
location, and all clusters being very far apart from each other.
Finally, we evaluate the entropy estimators while changing the dimension of the Gaussian components. We again generate 100 Gaussian components, each distributed as Xi ∼ N (µi , I (d) ), where
I (d) indicates a d × d dimensional identity matrix. We vary the dimensionality d from 1 to 60.
The results are shown in Fig. 1D. First, we see that when d = 1, the KDE estimator and the
KL-divergence based estimator give a very similar prediction (differing only by 0.5) , but as the
dimension increases, the two estimates diverge at a rate of d/2. Similarly, ĤELK grows increasingly less accurate as the dimension increases. Our proposed lower and upper bounds provide good
estimates of the mixture entropy across the whole dimensional sweep.
B.
Mixture of Uniforms
In the second set of experiments, we consider a mixture of uniform distributions. Unlike Gaussians, uniform distributions are bounded within a hyper-rectangle and do not have full support
over the domain. These means that there can be significant regions within which pX (x) > 0, but
pi (x) = 0 for some i.
Here we list the formulae for pairwise distance measure between uniform distributions. In the
following, we use Vi to indicate the “volume” of distribution pi . Uniform components have a constant
p(x) over their support, and so Vi = 1/pi (x) (for all x where pi (x) > 0). Similarly, we use Vi∩j as
the “volume
of overlap” between pi and pj , i.e., the volume of the intersection of the support of pi
R
and pj , x 1{pi (x) > 0}1{pj (x) > 0}dx. The distance measures between uniforms are then
H(pi ) = ln Vi
14
(
ln(Vj /Vi ) supp pi ⊆ supp pj
KL(pi kpj ) =
∞
o.w.
BD(pi ||pj ) = 0.5 ln Vi + 0.5 ln Vj − ln Vi∩j
X
X Vi∩j
ĤELK = −
wi ln
wj
Vi Vj
i
A
j
B
Entropy (nats)
Entropy (nats)
11
10
9
12
9
6
8
7
3
-4
-3
-2
-1
0
1
-2
0
Ln(Mean Spread)
D 15
11
Entropy (nats)
Entropy (nats)
C
2
Ln(Width Variability)
10
9
8
10
Monte Carlo
Pair Dist: KL
Pair Dist: Bhat.
KDE
Exp. Lik. Ker.
[H(X|C), H(X,C)]
5
7
0
-6
-4
-2
Ln(Cluster Spread)
0
5
10
15
Dimension
Figure 2: Entropy estimates for a mixture of a 100 uniform. In each plot, the vertical axis shows the
entropy of the distribution, and the horizontal axis changes a feature of the components: A) the mean
distance is increased; B) the uniform widths become more similar (at the right side of the plot, all
uniforms have width ≈ 2); C) the components are grouped into 5 “clusters”, and the distance between
these clusters is increased; D) the dimension is increased.
Like the Gaussian case, we run four different computational experiments and compare the mixture entropy estimates to the true entropy, as determined by Monte-Carlo sampling.
In the first experiment, the mixture consists of 100 10-dimensional uniform components, with
Xi ∼ U(µi − 1, µi + 1) , and µi ∼N (0, σI). Fig. 2A depicts the change in entropy as the value
of σ increases. For very small σ, the distributions are almost overlapping, while for large σ they
tend very far apart. As expected, the entropy increases with σ. Here, we see that the prediction
of ĤKL is identical to H(X, C), which arises because KL(pi kpj ) is infinite whenever pi has broader
support than pj . Uniform components with equal size and non-equal means must have some region
of non-overlap, and so the KL is infinite between all pairs of components, thus KL is effectively
Dmax (Eq. (8)). In contrast, we see that ĤBD estimates the true entropy quite well. This example
demonstrates that getting an accurate estimate of mixture entropy may require selecting a distance
function that works will with the component distributions. Finally, it turns out that for uniform
components of equal size, ĤELK = ĤBD . This can be seen by moving BD(pi kpj ) out of the
15
exponential, and noting that Vi = Vj when the components have equal size.
In the second experiment, we adjust the size of the uniform components. We again have 100
10-dimensional components, Xi ∼ U(µi − γi , µi + γi ) , where µi ∼N (0, I), and γi ∼ Γ(1 + σ, 1 + σ),
where Γ is the Gamma distribution. Fig. 2B shows the change in entropy as ln σ increases. When
σ is small, the widths have significant spread, while as σ grows the distributions become close to
equally sized. We again see that ĤBD is a good estimator of entropy, outperforming all of the
non-pairwise based estimators. Generally, not all supports will be non-overlapping, so ĤKL will not
necessarily be equal to H(X, C), though we find the two to be numerically quite close. However,
we also find that the difference between the estimates given ĤBD vs. ĤKL provides a tight estimate
of the true entropy.
In the third experiment, we again consider a mixture with clustered components, and evaluate
the entropy estimators as these clusters grow apart. Here, there are 100 components with Xi ∼
U(µg(i) − 1, µg(i) + 1), where g(i) ∈ {1..5} is the identity of the cluster for distribution i, which are
randomly assigned. The cluster centers are generated according to N (0, σI). Fig. 2C shows the
change in entropy as the clusters locations move apart. Note that in this case, the upper bound
ĤKL significantly outperforms H(X, C), unlike the first and second experiment. We again see that
ĤBD provides a relatively accurate lower bound for the true entropy.
In the final experiment, the dimension of the uniform components is varied. There are again 100
components, with Xi ∼ U(µi − 1, µi + 1) , and µi ∼N (0, σI). Fig. 2D shows the change in entropy
as the dimension increases. On the extreme left, the uniforms are univariate, and on the extreme
left the components are 16-dimensional. Interestingly, the low-dimensional case, H(X|C) is a very
close estimate for the true entropy, while in the high-dimensional case the entropy becomes very
close to H(X, C). This is because in the higher dimensional problems there is more space to be far
away from each other. As in the first experiment, ĤKL is equal to H(X; C). Despite this changing
regime, we see that ĤBD gives good estimates of mixture entropy, regardless of dimension.
VII.
DISCUSSION
We have presented a new class of estimators for the entropy of a mixture distribution. We have
shown that any estimator in this class has a bounded estimation bias, and that this class includes
useful lower and upper bounds on the entropy of a mixture. Finally, we show that these bounds
become exact when mixture components are grouped into well-separated clusters.
Our derivation of the bounds make use of some existing results [12, 14]. However, to our
knowledge, these results have not been previously used to estimate mixture entropies. Furthermore,
they have not been compared numerically or analytically to better-known bounds.
We evaluated these estimators using numerical simulations of mixtures of Gaussians as well as
mixtures of bounded (hypercube) uniform distributions. Our results demonstrate that our estimators perform much better than existing well-known estimators.
This estimator class can be especially useful for optimization problems that involve minimization
of entropy or mutual information. If the distance function used in the pairwise estimator class is
continuous and smooth in the parameters of the mixture components, then the entropy estimate is
also continuous and smooth. This enables use within gradient-based optimization techniques, for
example gradient descent, as often used for machine learning problems.
In fact, we have used our upper bound to carry out a non-parametric, nonlinear version of the
“Information Bottleneck” [31]. Specifically, we minimized an upper bound on the mutual information
between input and hidden layer in a neural networks [20]. We found that the optimal distributions
were often clustered (Section III D). That work demonstrated practically the value of having an
accurate, differentiable upper bound on mixture entropy that performs well in the clustered regime.
16
ACKNOWLEDGEMENTS
We thank David H. Wolpert for useful discussions. We would also like to thank the Santa Fe
Institute for helping to support this research. This work was made possible through the support
of AFOSR MURI on multi-information sources of multi-physics systems under Award Number
FA9550-15-1-0038.
AUTHOR CONTRIBUTIONS
AK and BDT designed the method and experiments; BDT performed the experiments; AK and
BDT wrote the paper.
CONFLICTS OF INTEREST
The authors declare no conflict of interest.
Appendix A: Chernoff α-divergence is not a distance function for α ∈
/ [0, 1]
For any pair of densities p and q, consider the Chernoff α-divergence
Z
Cα := − ln pα (x)q 1−α (x) dx = − ln cα
where the quantity cα is called the Chernoff α-coefficient [22]. Taking the second derivative of
cα with respect to α gives
d2
cα =
dα2
Z
α
p (x)q
1−α
p(x)
(x) ln
q(x)
2
dx
Observe that this quantity is everywhere positive, meaning that cα is convex everywhere. For
simplicity, consider the case p 6= q, in which case this function is strictly convex. Because cα is
equal to 1 for α = 0 and α = 1, this means that cα > 1 for α ∈
/ [0, 1]. This in turn means that
the Chernoff α-divergence Cα is strictly negative for α ∈
/ [0, 1]. Thus, Cα is not a valid distance
function for α ∈
/ [0, 1], as defined in Section III.
[1] Arindam Banerjee, Srujana Merugu, Inderjit S Dhillon, and Joydeep Ghosh. Clustering with
bregman divergences. Journal of machine learning research, 6(Oct):1705–1749, 2005.
[2] Francisco M Capristan and Juan J Alonso. Range safety assessment tool (rsat): An analysis
environment for safety assessment of launch and reentry vehicles. In 52nd Aerospace Sciences
Meeting, page 0304, 2014.
[3] Jia-Yu Chen, John R Hershey, Peder A Olsen, and Emmanuel Yashchin. Accelerated monte
carlo for kullback-leibler divergence between gaussian mixture models. In Acoustics, Speech
and Signal Processing, 2008. ICASSP 2008. IEEE International Conference on, pages 4553–
4556. IEEE, 2008.
[4] Andrzej Cichocki and Shun-ichi Amari. Families of alpha-beta-and gamma-divergences: Flexible and robust measures of similarities. Entropy, 12(6):1532–1568, 2010.
17
[5] Andrzej Cichocki, Rafal Zdunek, Anh Huy Phan, and Shun-Ichi Amari. Similarity measures
and generalized divergences. Nonnegative Matrix and Tensor Factorizations: Applications to
Exploratory Multi-Way Data Analysis and Blind Source Separation, pages 81–129, 2009.
[6] Thomas M. Cover and Joy A. Thomas. Elements of information theory. John Wiley & Sons,
2012.
[7] Gavin E Crooks. On Measures of Entropy and Information. Technical Report 009 v0.6, 2017.
[8] Keinosuke Fukunaga. Introduction to statistical pattern recognition. Computer science and
scientific computing. Academic Press, Boston, 2nd ed edition, 1990.
[9] Manuel Gil, Fady Alajaji, and Tamas Linder. Rényi divergence measures for commonly used
univariate continuous distributions. Information Sciences, 249:124–131, 2013.
[10] Jacob Goldberger, Shiri Gordon, Hayit Greenspan, et al. An efficient image similarity measure
based on approximations of kl-divergence between two gaussian mixtures. In ICCV, volume 3,
pages 487–493, 2003.
[11] Peter Hall and Sally C Morton. On the estimation of entropy. Annals of the Institute of
Statistical Mathematics, 45(1):69–88, 1993.
[12] D. Haussler and M. Opper. Mutual information, metric entropy and cumulative relative
entropy risk. Ann Stat, 25(6), 1997.
[13] Alfred O. Hero, Bing Ma, Olivier Michel, and John Gorman. Alpha-Divergence for Classification, Indexing and Retrieval. 2001.
[14] John R. Hershey and Peder A. Olsen. Approximating the Kullback Leibler divergence between
Gaussian mixture models. In 2007 IEEE International Conference on Acoustics, Speech and
Signal Processing-ICASSP’07, volume 4, pages IV–317. IEEE, 2007.
[15] Tomáš Hobza, Domingo Morales, and Leandro Pardo. Rényi statistics for testing equality of
autocorrelation coefficients. Statistical Methodology, 6(4):424–436, July 2009.
[16] Marco F Huber, Tim Bailey, Hugh Durrant-Whyte, and Uwe D Hanebeck. On entropy
approximation for gaussian mixture random vectors. In Multisensor Fusion and Integration
for Intelligent Systems, 2008. MFI 2008. IEEE International Conference on, pages 181–188.
IEEE, 2008.
[17] Tony Jebara and Risi Kondor. Bhattacharyya and expected likelihood kernels. In Learning
theory and kernel machines, pages 57–71. Springer, 2003.
[18] Tony Jebara, Risi Kondor, and Andrew Howard. Probability product kernels. Journal of
Machine Learning Research, 5(Jul):819–844, 2004.
[19] Harry Joe. Estimation of entropy and other functionals of a multivariate density. Annals of
the Institute of Statistical Mathematics, 41(4):683–697, 1989.
[20] Artemy Kolchinsky, Brendan D. Tracey, and David H. Wolpert. Nonlinear Information Bottleneck. arXiv:1705.02436 [cs, math, stat], May 2017. arXiv: 1705.02436.
[21] Chandra Nair, Balaji Prabhakar, and Devavrat Shah. On entropy for mixtures of discrete
and continuous variables. arXiv preprint cs/0607075, 2006.
[22] Frank Nielsen. Chernoff information of exponential families. arXiv preprint arXiv:1102.2684,
2011.
[23] Frank Nielsen. Generalized Bhattacharyya and Chernoff upper bounds on Bayes error using
quasi-arithmetic means. Pattern Recognition Letters, 42:25–34, June 2014.
[24] John Paisley. Two useful bounds for variational inference. Technical report, Technical report,
Department of Computer Science, Princeton University, Princeton, NJ, 2010.
[25] Leandro Pardo. Statistical Inference Based on Divergence Measures. CRC Press, October
2005. Google-Books-ID: ziDGGIkhqlMC.
[26] Jose C Principe, Dongxin Xu, and John Fisher. Information theoretic learning. Unsupervised
adaptive filtering, 1:265–319, 2000.
[27] Igal Sason and Sergio Verdú. f -divergence inequalities. IEEE Transactions on Information
Theory, 62(11):5973–6006, 2016.
[28] Nicol N. Schraudolph. Gradient-based manipulation of nonparametric entropy estimates.
Neural Networks, IEEE Transactions on, 15(4):828–837, 2004.
[29] Nicol Norbert Schraudolph. Optimization of entropy with neural networks. PhD thesis,
Citeseer, 1995.
18
[30] Sarit Shwartz, Michael Zibulevsky, and Yoav Y. Schechner. Fast kernel entropy estimation
and optimization. Signal Processing, 85(5):1045–1058, 2005.
[31] N. Tishby, F. Pereira, and W. Bialek. The information bottleneck method. In 37th Allerton
Conf on Communication, 1999.
[32] Tim Van Erven and Peter Harremos. Rényi divergence and kullback-leibler divergence. IEEE
Transactions on Information Theory, 60(7):3797–3820, 2014.
[33] Paul Viola, Nicol N. Schraudolph, and Terrence J. Sejnowski. Empirical Entropy Manipulation for Real-World Problems. Advances in neural information processing systems, pages
851–857, 1996.
[34] Jian-Wu Xu, António RC Paiva, Il Park, and Jose C Principe. A reproducing kernel hilbert
space framework for information-theoretic learning. IEEE Transactions on Signal Processing,
56(12):5891–5902, 2008.
[35] O Zobay et al. Variational bayesian inference with gaussian-mixture approximations. Electronic Journal of Statistics, 8(1):355–389, 2014.
| 7cs.IT
|
GENERIC MACHINE LEARNING INFERENCE ON HETEROGENOUS TREATMENT
EFFECTS IN RANDOMIZED EXPERIMENTS
arXiv:1712.04802v3 [stat.ML] 23 Feb 2018
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
Abstract. We propose strategies to estimate and make inference on key features of heterogeneous effects in randomized experiments. These key features include best linear predictors of the effects using machine learning proxies, average
effects sorted by impact groups, and average characteristics of most and least impacted units. The approach is valid in high
dimensional settings, where the effects are proxied by machine learning methods. We post-process these proxies into
the estimates of the key features. Our approach is generic, it can be used in conjunction with penalized methods, deep
and shallow neural networks, canonical and new random forests, boosted trees, and ensemble methods. Our approach
is agnostic and does not make unrealistic or hard-to-check assumptions; we don’t require conditions for consistency of
the ML methods. Estimation and inference relies on repeated data splitting to avoid overfitting and achieve validity.
For inference, we take medians of p-values and medians of confidence intervals, resulting from many different data
splits, and then adjust their nominal level to guarantee uniform validity. This variational inference method is shown
to be uniformly valid and quantifies the uncertainty coming from both parameter estimation and data splitting. The
inference method could be of substantial independent interest in many machine learning applications. An empirical
application to the impact of micro-credit on economic development illustrates the use of the approach in randomized
experiments. An additional application to the impact of the gender discrimination on wages illustrates the potential use
of the approach in observational studies, where machine learning methods can be used to condition flexibly on very
high-dimensional controls.
Key words: Agnostic Inference, Machine Learning, Confidence Intervals, Causal Effects, Variational P-values and
Confidence Intervals, Uniformly Valid Inference, Quantification of Uncertainty, Sample Splitting, Multiple Splitting,
Assumption-Freeness
1. Introduction
Randomized experiments play an important role in the evaluation of social and economic programs and medical treatments (e.g., [35, 27]). Researchers and policy makers are often interested
in features of the impact of the treatment that go beyond the simple average treatment effects. In
particular, very often, they want to know whether treatment effect depends on basic covariates,
such as gender, age, etc. It is essential to assess if the impact of the program would generalize to
a different population with different characteristics, and for economists, to better understand the
driving mechanism behind a particular program effects.
One issue with reporting treatment effects split by subgroups, however, is that there are often a large number of potential sample splits: choosing subgroups ex-post opens the possibility
Date: December, 2017.
We thank Susan Athey, Moshe Buchinsky, Denis Chetverikov, Siyi Luo, Max Kasy, Susan Murphy, Whitney Newey,
and seminar participants at UCLA and AEA for valuable comments.
1
2
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
of overfitting. To solve this problem, medical journals and the FDA require pre-registering subsample of interest in medical trials in advance. In economics, this approach has gained some traction, with the adoption of pre-analysis plan (which can be filed in the AEA registry for randomized
experiments). Restricting heterogeneity analysis to pre-registered subgroups, however, amounts
to throwing away a large amount of potentially valuable information, especially now that many
researchers collect large baseline data sets. It should be possible to use the data to discover ex post
whether there is any relevant heterogeneity in treatment effect by co-variates.
To do this in a disciplined fashion and avoid the risk of overfitting, scholars have recently proposed using machine learning tools (see e.g. [6] and below for a review). Indeed, ML tools seem
to be ideal to explore heterogeneity of treatment effect, when researchers have access to a potentially large array of baseline variables to form subgroup, and little guiding principles on which of
those are likely to be relevant. Several recent papers, which we review below, develop methods for
detecting heterogeneity in treatment effects. Empirical researchers have taken notice.1
This paper develops a generic approach to use any of the tools of machine learning to predict
and make inference on heterogeneous treatment or policy effects. A core difficulty of applying
ML tools to the estimation of heterogenous causal effects is that, while they are successful in prediction empirically, it is much more difficult to obtain uniformly valid inference. In fact, in high
dimensional settings, absent strong assumptions, generic ML tools may not even produce consistent estimates of the conditional average treatment effect (or CATE) (the difference in the expected
potential outcomes between treated and control groups conditional on the covariates Z).
Previous attempts to solve this problem focus on specific tools and situations where those assumptions are satisfied. Our approach to resolve the fundamental impossibilities in non-parametric
inference is different. Motivated by [29], instead of attempting to get consistent estimate and uniformly valid inference on CATE itself, we focus on providing valid estimation and inference for
features of CATE. We start by building a ML proxy predictor of CATE, and then develop valid inference for features of the CATE based on this proxy predictor. In particular, we develop valid
inference for the following three objects, which are likely to be of interest to applied researchers
and policy makers: First, the Best Linear Predictor (BLP) of the CATE based on the ML proxy predictor, second, the Sorted Group Average Treatment Effects (GATES), average treatment effect by
heterogeneity groups induced by the ML proxy predictor S(Z); third, the Classification Analysis
(CLAN): the average characteristics of the most and least affected units defined in terms of the ML
proxy predictor. Thus, we can find out if there is detectable heterogeneity in the treatment effect
based on observables, and if there is any, what is the treatment effect for different bins. And finally
we can describe which of the covariates is correlated with this heterogeneity.
1
In the last few months alone, several new empirical papers in economics used ML methods to estimate heterogenous
effects. E.g. [44], which shows that villagers outperform the machine learning tools when they predict heterogeneity in
returns to capital. [25] predict who benefits the most from a summer internship projects. The methodological papers
reviewed later also contain a number of empirical applications.
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
3
Thus, in the trade off between more restrictive assumptions and a more ambitious estimation,
we chose a different point than previous papers: focus on coarser objects of the function rather
than the function itself, but make as little assumptions as possible. This seems to be a worthwhile
sacrifice: the objects that for which we have developed inference appear to us at this point to be
the most relevant, but in the future, one could easily develop methods for other objects of interest.
The Model and Key Causal Functions. Let Y (1) and Y (0) be the potential outcomes in the treatment state 1 and the non-treatment state 0; see [46]. Let Z be a vector of of covariates that characterize the observational units. The main causal functions are the baseline conditional average
(BCA):
b0 (Z) := E[Y (0) | Z],
(1.1)
and the conditional average treatment effect (CATE):
s0 (Z) := E[Y (1) | Z] − E[Y (0) | Z].
(1.2)
Suppose the binary treatment variable D is randomly assigned conditional on Z, with probability of assignment depending only on a subvector of stratifying variables Z1 ⊆ Z, namely
D ⊥⊥ (Y (1), Y (0)) | Z,
(1.3)
and the propensity score is known and is given by
p(Z) := P[D = 1 | Z] = P[D = 1 | Z1 ],
(1.4)
which we assume is bounded away from zero or one:
p(Z) ∈ [p0 , p1 ] ⊂ (0, 1).
(1.5)
The observed outcome is Y = DY (1) + (1 − D)Y (0). Under the stated assumption, the causal
functions are identified by the components of the regression function of Y given D, Z:
Y = b0 (Z) + Ds0 (Z) + U, E[U | Z, D] = 0,
that is,
b0 (Z) = E[Y | D = 0, Z],
(1.6)
s0 (Z) = E[Y | D = 1, Z] − E[Y | D = 0, Z].
(1.7)
and
We observe Data = (Yi , Zi , Di )N
i=1 , consisting of i.i.d. copies of the random vector (Y, Z, D)
having probability law P . Expectation with respect to P is denoted by E = EP . The probability
law of the entire data is denoted by P = PP and the corresponding expectation is denoted by
E = EP .
4
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
Properties of Machine Learning Estimators of s0 (Z) Motivating the Agnostic Approach. Machine learning (ML) is a name attached to a variety of new, constantly evolving statistical learning
methods: Random Forest, Boosted Trees, Neural Networks, Penalized Regression, Ensembles, and
Hybrids (see, e.g., [50] for a recent review, and [28] for a prominent textbook treatment). In modern
high-dimensional settings, machine learning methods effectively explore the various forms of nonlinear structured sparsity to yield “good” approximations to s0 (z) whenever such assumptions are
valid, based on equations (1.6) and (1.7). As a result these methods often work much better than
classical methods in high-dimensional settings, and have found widespread uses in industrial and
academic applications.
Motivated by the practical predictive success of the methods, it is really tempting to apply the
ML methods directly to try to learn the CATE function z 7→ s0 (z) (by learning the two regression
functions for treated and untreated and taking the difference). However, it is hard, if not impossible, to obtain uniformly valid inference on z 7→ s0 (z) using generic ML methods, under credible
assumptions and practical tuning parameter choices. There are several fundamental reasons as
well as huge gaps between theory and practice that are responsible for this.
One fundamental reason is that the ML methods might not even produce consistent estimators
of z 7→ s0 (z) in high dimensional settings. For example, if z has dimension d and the target function
z 7→ s0 (z) is assumed to have p continuous and bounded derivatives, then the worst case (minimax)
lower bound on the rate of learning this function cannot be better than N −p/(2p+d) as N → ∞, as
shown by Stone [47]. Hence if p is fixed and d is also small, slowly increasing with N , such as
d > log N , then there exists no consistent estimator of z 7→ s0 (z) generally.
Hence, ML estimators cannot be regarded as consistent, unless further very strong assumptions
are made. Examples of such assumptions include structured forms of linear and non-linear sparsity and super-smoothness. While these (sometime believable and yet untestable) assumptions
make consistent adaptive estimation possible (e.g.,[17]), inference remains a more difficult problem, as adaptive confidence sets do not exist even for low-dimensional nonparametric problems
([41, 29]). Indeed, adaptive estimators (including modern ML methods) have biases of comparable
or dominating order as compared to sampling error. Further assumptions such as ”self-similarity”
are needed to bound the biases and expand the confidence bands by the size of bias (see [30, 22])
to produce partly adaptive confidence bands. For more traditional statistical methods there are
constructions in this vein that make use of either undersmoothing or bias-bounding arguments
([30, 22]). These methods, however, are not yet available for ML methods in high dimensions (see,
however, [31] for a promising approach called ”targeted undersmoothing” in sparse linear models).
Suppose we did decide to be optimistic (or panglossian) and imposed the strong assumptions,
that made the theoretical versions of the ML methods provide us with high-quality consistent estimators of z 7→ s0 (z) and valid confidence bands based on them. This would still not give as a practical construction we’d want for our applications. The reason is that there is a huge gap between
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
5
theoretical versions of the ML methods appearing in various theoretical papers and the practical
versions (with the actual, data-driven tuning parameters) coded up in statistical computing packages used by practitioners2.3 The use of ML, for example, involves many tuning parameters with
practical rules for choosing them, while theoretical works provide little guidance or backing for
such practical rules; see e.g., the influential book [28] for many examples of such rules.4
In this paper we shall take an agnostic view and we shall not rely on any structured assumptions,
which might be difficult to verify or believe in practice, and we don’t impose conditions that make
the ML estimators consistent. We simply treat ML as providing proxy predictors for the objects of
interest.
Our Agnostic Approach. Our approach will be radically different – we propose strategies for estimation and inference on
key features of s0 (Z) rather than s0 (Z) itself.
Because of this difference in focus we can avoid making strong assumptions about the properties
of the ML estimators.
Let (M, A) denote a random partition of the set of indices {1, . . . , N }. The strategies that we
consider rely on random splitting of
Data = (Yi , Di , Zi )N
i=1
into a main sample, denoted by DataM = (Yi , Di , Zi )i∈M , and an auxiliary sample, denoted by
DataA = (Yi , Di , Zi )i∈A . We will sometimes refer to these samples as M and A. We assume that
the main and auxiliary samples are approximately equal in size, though this is not required theoretically.
From the auxiliary sample A, we obtain ML estimators of the baseline and treatment effects,
which we call the proxy predictors,
z 7→ B(z) = B(z; DataA ) and z 7→ S(z) = S(z; DataA ).
These are possibly biased and noisy predictors of b0 (z) and s0 (z), and in principle, we do not even
require that they are consistent for b0 (z) and s0 (z). We simply treat these estimates as proxies,
which we post-process to estimate and make inference on the features of the CATE z 7→ s0 (z). We
condition on the auxiliary sample DataA , so we consider these maps as frozen, when working with
the main sample.
2There are cases where such gap does not exist, e.g., see [13, 15] for the lasso.
3For example, even the wide use of K-fold cross-validation in high-dimensional settings for machine learning remains
theoretically unjustified (there do exist, however, related subsample-based methods [52, 40] that do achieve near-oracle
performance for tuning selection).
4Unfortunately, theoretical work often only provides existence results: there exist theoretical ranges of the tuning
parameters that make the simple versions of the methods work for predictive purposes (under very strong assumptions),
leaving no satisfactory guide to practice.
6
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
Using the main sample and the proxies, we shall target and develop valid inference about key
features of s0 (Z) rather than s0 (Z), which include
(1) Best Linear Predictor (BLP) of the CATE s0 (Z) based on the ML proxy predictor S(Z);
(2) Sorted Group Average Treatment Effects (GATES): average of s0 (Z) (ATE) by heterogeneity groups induced by the ML proxy predictor S(Z);
(3) Classification Analysis (CLAN): average characteristics of the most and least affected units
defined in terms of the ML proxy predictor S(Z).
Our approach is generic with respect to the ML method being used, and is agnostic about its formal
properties.
Our point estimation will make use of many splits of the data into main and auxiliary samples
to produce robust estimates. Our estimation and inference will systematically account for two
sources of uncertainty:
(I) Estimation uncertainty conditional on the auxiliary sample.
(II) Splitting uncertainty induced by random partitioning of the data into the main and auxiliary samples.
Because we account for the second source, we call the resulting collection of methods as variational
estimation and inference methods (VEINs). For point estimates we report the median of the estimated key features over different random splits of the data. For the confidence intervals we take
the medians of many random conditional confidence sets and we adjust their nominal confidence
level to reflect the splitting uncertainty. We construct p-values by taking medians of many random
conditional p-values and adjust the nominal levels to reflect the splitting uncertainty. Note that
considering many different splits and accounting for variability caused by splitting is very important. Indeed, with a single splitting practice, empiricists may unintentionally look for a ”good” data
split, which supports their prior beliefs about the likely results, thereby invalidating inference.5
Relationship to the Literature. We focus the review strictly on the literatures about estimation
and inference on heterogeneous effects and inference using sample splitting.
We first mention work that uses linear and semiparametric regression methods. A semiparametric inference method for characterizing heterogeneity, called the sorted effects method, was
given in [20]. This approach does provide a full set of inference tools, including simultaneous
5This problem is “solved” by fixing the Monte-Carlo seed and the entire data analysis algorithm before the empirical
study. Even if such a huge commitment is really made and followed, there is a considerable risk that the resulting
data-split may be non-typical. Our approach allows one to avoid taking this risk.
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
7
bands for percentiles of the CATE, but is strictly limited to the traditional semiparametric estimators for the regression and causal functions. [31] proposed a sparsity based method called ”targeted undersmoothing” to perform inference on heterogeneous effects. This approach does allow
for high-dimensional settings, but makes strong assumptions on sparsity as well as additional assumptions that enable the targeted undersmoothing. A related approach, which allows for simultaneous inference on many coefficients (for example, inference, on the coefficients corresponding
to the interaction of the treatment with other variables) was first given in [14] using a Z-estimation
framework, where the number of interactions can be very large; see also [26] for a more recent
effort in this direction, focusing on de-biased lasso in mean regression problems. This approach,
however, still relies on a strong form of sparsity assumptions. [53] proposed a post-selection inference framework within the high-dimensional linear sparse models for the heterogeneous effects.
The approach is attractive because it allows for some misspecification of the model.
Next we discuss the use of tree-based and other methods. [34] discussed the use of a heuristic
support-vector-machine method with lasso penalization for classification of heterogeneous treatments into positive and negative ones. They used the Horvitz-Thompson transformation of the
outcome (e.g., as in [33, 1]) such that that the new outcome becomes an unbiased, noisy version of
CATE. [5] made use of the Horvitz-Thompson transformation of the outcome variable to inform
the process of building causal trees, with the main goal of predicting CATE. They also provide a
valid inference result on average treatment effects for groups defined by the tree leaves, conditional
on the data split in two subsamples: one used to build the tree leaves and the one to estimate the
predicted values given the leaves. Unlike our generic method introduced below, this inference approach is limited to trees and does not account for splitting uncertainty.6 Like our methods, this approach is essentially assumption-free. [49] provided a subsampling-based construction of a causal
random forest, providing valid pointwise inference for CATE (see also the review in [49] on prior
uses of random forests in causal settings) for the case when covariates are very low-dimensional
(and essentially uniformly distributed).7 Unfortunately, this condition rules out the typical highdimensional settings that arise in many empirical problems, including the ones considered in this
paper.
Our approach is radically different from these existing approaches, in that we are changing the
target, and instead of hunting for CATE z 7→ s0 (z), we focus on key features of z 7→ s0 (z). Our
approach is generic, it can be used in conjunction with penalized methods, deep and shallow neural
6In principle, our (generic) inference methods described below applies to the [5]’s problem to provide valid inference
that does account for sample splitting uncertainty. Also, going beyond trees is crucial, since the best ML algorithms are
typically not trees.
7The dimension d is fixed in [49]; the analysis relies on the Stone’s model with smoothness index β = 1, in which no
consistent estimator exists once d > log n. It’d be interesting to establish consistency properties and find valid inferential
procedures for the random forest in high-dimensional (d ∝ n or d n) approximately sparse cases, with continuous and
categorical covariates, but we are not aware of any studies that cover such settings, which are of central importance to
us.
8
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
networks, canonical and new random forests, boosted trees, and ensemble methods. Our approach
is agnostic and does not make unrealistic or hard-to-check assumptions, for example, we don’t
even require conditions for consistency of the ML methods. We simply treat the ML methods as
providing a proxy predictor z 7→ S(z), which we post-process to estimate and make inference
on the key features of the CATE z 7→ s0 (z). Some of our strategies rely on Horvitz-Thompson
transformations of outcome and some do not. The inspiration for our approach draws upon an
observation in [29], namely that some fundamental impossibilities in non-parametric inference
could be avoided if we focus inference on coarser features of the non-parametric functions rather
than the functions themselves.
Our inference approach is also of independent interest, and could be applied to many problems,
where sample splitting is used to produce ML predictions, e.g. [2]. Related references include
[51, 42], where the ideas are related but quite different in details, which we shall explain below.
The premise is the same, however, as in [42, 45] – we should not rely on a single random split
of the data and should adjust inference in some way. Our approach takes the medians of many
conditional confidence intervals as the confidence interval and the median of many conditional
p-values as the p-value, and adjusts their nominal levels to account for the splitting uncertainty.
Our construction of p-values builds upon ideas in [16, 42], though what we propose is radically
simpler, and our confidence intervals appear to be brand new. Of course sample splitting ideas
are classical, going back to [32, 38, 12, 23, 43], though having been mostly underdeveloped and
overlooked for inference, as characterized by [45]. The overlooked status perhaps allowed us to
innovate on the inference side.
2. Main Identification Results and Estimation Strategies
2.1. BLP of CATE. We consider two strategies for identifying and estimating the best linear predictor of s0 (Z) using S(Z):
BLP[s0 (Z) | S(Z)] := arg
min
f (Z)∈Span(1,S(Z))
E[s0 (Z) − f (Z)]2 ,
which, if exists, is defined by projecting s0 (Z) on the linear span of 1 and S(Z) in the space L2 (P ).
BLP of CATE: The First Strategy. Here we shall identify the coefficients of the BLP from the
weighted linear projection:
Y = α0 X1 + β1 (D − p(Z)) + β2 (D − p(Z))(S − ES) + , E[w(Z)X] = 0,
where S := S(Z),
w(Z) = {p(Z)(1 − p(Z))}−1 ,
X1 := X1 (Z),
e.g.,
X := (X1 , X2 )
X1 = [1, B(Z)],
X2 := [D − p(Z), (D − p(Z))(S(Z) − (S − ES)].
Note that the above equation uniquely pins down β1 and β2 under weak assumptions.
(2.1)
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
9
The interaction (D − p(Z))(S − ES) is orthogonal to D − p(Z) under the weight w(Z) and to all
other regressors that are functions of Z under any Z-dependent weight.8
A consequence is our first main identification result, namely that
β1 + β2 (S(Z) − ES) = BLP[s0 (Z) | S(Z)],
in particular β1 = Es0 (Z) and β2 = Cov(s0 (Z), S(Z))/ Var(S(Z)).
Theorem 2.1 (BLP 1). Consider z 7→ S(z) and z 7→ B(z) as fixed maps. Assume that Y and X have
finite second moments and that EXX 0 is full rank. Then, (β1 , β2 ) defined in (2.1) also solves the best linear
predictor/approximation problem for the target s0 (Z):
(β1 , β2 )0 = arg min E[s0 (Z) − b1 − b2 S(Z)]2 ,
b1 ,b2
in particular β1 = ES0 (Z) and β2 = Cov(s0 (Z), S(Z))/ Var(S(Z)).
The identification result is constructive. We can base the corresponding estimation strategy on
the empirical analog:
Yi = α
b0 X1i + βb1 (Di − p(Zi )) + βb2 (Di − p(Zi ))(Si − EN,M Si ) + b
i ,
i ∈ M,
EN,M [w(Zi )b
i Xi ] = 0,
where EN,M denotes the empirical expectation with respect to the main sample, i.e.
X
EN,M g(Yi , Di , Zi ) := |M |−1
g(Yi , Di , Zi ).
i∈M
The properties of this estimator, conditional on the auxilliary data, are well known and follow as
a special case of Lemma B.1 in the Appendix.
Comment 2.1 (Main Implications of the result). If S(Z) is a perfect proxy for s0 (Z), then β2 = 1.
In general, β2 6= 1, correcting for noise in S(Z). If S(Z) is complete noise, uncorrelated to s0 (Z),
then β2 = 0. Furthermore, if there is no heterogeneity, that is s0 (Z) = s, then β2 = 0. Rejecting
the hypothesis β2 = 0 therefore means that there is both heterogeneity and S(Z) is its relevant
predictor.
Figure 1 provides two examples. The left panel shows a case without heterogeneity in the CATE
where s0 (Z) = 0, whereas there right panel shows a case with strong heterogeneity in the CATE
where s0 (Z) = Z. In both cases we evenly split 1000 observations between the auxiliary and main
8The orthogonalization ideas embedded in this strategy do have classical roots in econometrics (going back to at
least Frisch and Waugh in the 30s), and similar strategies underlie the orthogonal or double machine learning approach
(DML) in [21]. Our paper has different goals than DML, attacking the problem of inference on heterogeneous effects
without rate and even consistency assumptions. The strategy here is more nuanced in that we are making it work under
misspecification or inconsistent learning, which is likely to be true in very high-dimensional problems.
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
0.5
−0.5
0.0
S(Z)
0
−4
−1.0
−2
S(Z)
2
1.0
4
10
−1.0
−0.5
0.0
0.5
1.0
−1.0
Z
−0.5
0.0
0.5
1.0
Z
Figure 1. Example. In the left panel we have a homogeneous CATE s0 (Z) = 0; in
the right panel we have heterogeneous CATE s0 (Z) = Z. The proxy predictor S(Z)
is produced by the Random Forest, shown by green line, the true BLP of CATE is
shown by black line, and the estimated BLP of CATE is shown by blue line. The
true and estimated BLP of CATE are more attenuated towards zero than the proxy
predictor.
samples, Z is uniformly distributed in (−1, 1), and the proxy predictor S(Z) is estimated by random
forest in the auxiliary sample following the standard implementation, see e.g. [28]. When there
is no heterogeneity, post-processing the ML estimates helps reducing sampling noise bringing the
estimated BLP close to the true BLP; whereas under strong heterogeneity the signal in the ML
estimates dominates the sampling noise and the post-processing has little effect.
Comment 2.2 (Digression: Naive Strategy that is not Quite Right). It is tempting and “more natural” to estimate
Y = α̃1 + α̃2 B + β̃1 D + β̃2 D(S − ES) + , E[X̃] = 0,
where X̃ = (1, B, D, D(S − ES)). This is a good strategy for predicting the conditional expectation
of Y given Z and D. But, β̃2 6= β2 , and β̃1 + β̃2 (S − ES) is not the best linear predictor of s0 (Z).
BLP of CATE: The Second Strategy. The second strategy makes use of the Horvitz-Thompson
transformation:
D − p(Z)
.
H = H(D, Z) =
p(Z)(1 − p(Z))
It is well known that the transformed response Y H provides an unbiased signal about CATE:
E[Y H | Z] = s0 (Z)
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
11
and it follows that
BLP[s0 (Z) | S(Z)] = BLP[Y H | S(Z)].
This simple strategy is completely fine for identification purposes, but can severely underperform
in estimation and inference due to lack of precision. We can repair the deficiencies as follows.
We consider, instead, the linear projection:
Y H = µ0 X1 H + β1 + β2 (S − ES) + ˜,
E˜
X̃ = 0,
(2.2)
where B := B(Z), S := S(Z), and X̃ := (X10 H, X̃20 )0 , X̃2 = (1, (S − ES)0 )0 , where X1 = X1 (Z), e.g.
B(Z) or (B(Z), S(Z), p(Z))0 . The terms X1 are present in order to reduce noise.
We show that, as a complementary main identification result,
β1 + β2 (S − ES) = BLP[s0 (Z) | S(Z)].
Theorem 2.2 (BLP 2). Consider z 7→ S(z) and z 7→ B(z) as fixed maps. Assume that Y has finite second
moments and X̃ = (X1 H, 1, (S − ES)) is such that EX̃ X̃ 0 is finite and full rank. Then, (β1 , β2 ) defined
in (2.2) solves the best linear predictor/approximation problem for the target s0 (Z):
(β1 , β2 )0 = arg min E[s0 (Z) − b1 − b2 S(Z)]2 ,
b1 ,b2
in particular β1 = Es0 (Z) and β2 = Cov(s0 (Z), S(Z))/ Var(S(Z)).
The corresponding estimator is defined through the empirical analog:
Yi Hi = µ
b0 X1i Hi + βb1 + βb2 (Si − EN,M Si ) + b
i ,
EN,M b
i X̃i = 0,
and the properties of this estimator, conditional on the auxiliary data, are well known and given
in Lemma B.1.
2.2. The Sorted Group ATE. The target parameters are
E[s0 (Z) | G],
where G is an indicator of group membership.
Comment 2.3. There are many possibilities for creating groups based upon ML tools applied to
the auxiliary data. For example, one can group or cluster based upon predicted baseline response
(as in the “endogenous stratification” analysis, [2]) or based upon actual predicted treatment effect
S. We focus on the latter approach for defining groups, although our identification and inference
ideas immediately apply to other ways of defining groups, and could be helpful in these contexts.
We build the groups to explain as much variation in s0 (Z) as possible
Gk := {S ∈ Ik },
k = 1, ..., K,
12
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
where Ik = [`k−1 , `k ) are non-overlaping intervals that divide the support of S into regions [`k−1 , `k )
with equal or unequal masses:
−∞ = `0 < `1 < . . . < `K = +∞.
The parameters of interest are the Sorted Group Average Treatment Effects (GATES):
E[s0 (Z) | Gk ],
k = 1, . . . , K.
Given the definition of groups, it is natural for us to impose the monotonicity restriction
E[s0 (Z) | G1 ] 6 ... 6 E[s0 (Z) | GK ],
which holds asymptotically if S(Z) is consistent for s0 (Z) and the latter has an absolutely continuous distribution. Under the monotonicity condition, the estimates could be rearranged to obey the
weak monotonicity condition, improving the precision of the estimator. The joint confidence intervals could also be improved by intersecting them with the set of monotone functions. Furthermore,
as before, we can test for homogeneous effects, s0 (Z) = s, by testing whether,
E[s0 (Z) | G1 ] = ... = E[s0 (Z) | GK ].
GATES: The First Strategy. Here we shall recover GATES parameters from the weighted linear
projection equation:
0
Y = α X1 +
K
X
γk · (D − p(Z)) · 1(Gk ) + ν,
E[w(Z)νW ] = 0,
(2.3)
k=1
for B := B(Z), S := S(Z), W = (X10 , W20 )0 ,
0
W2 = ({(D − p(Z))1(Gk )}K
k=1 ) .
The presence of D − p(Z) in the interaction (D − p(Z))1(Gk ) orthogonalizes this regressor relative
to all other regressors that are functions of Z. The controls X1 , e.g. B, can be included to improve
precision.
The second main identification result is that the projection coefficients γk are the GATES
parameters:
K
γ = (γk )K
k=1 = (E[s0 (Z) | Gk ])k=1 .
Given the identification strategy, we can base the corresponding estimation strategy on the following empirical analogs:
Yi = α
b0 X1i + γ
b0 W2i + νbi ,
i ∈ M, EN,M [w(Zi )b
νi Wi ] = 0.
(2.4)
The properties of this estimator, conditional on the auxilliary data, are well known and stated as a
special case of Lemma B.1.
A formal statement appears below, together with a complementary result.
13
1.0
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
4
●
●
●
●
●
●
●
●
●
●
●
0.0
●
●
E s(Z)|Group
●
0
●
●
●
●
●
●
●
●
●
−0.5
−2
E s(Z)|Group
2
0.5
●
●
−4
●
●
−1.0
●
1
2
3
4
5
1
2
Heterogeneity Group
3
4
5
Heterogeneity Group
Figure 2. In the left panel we have the homogeneous CATE s0 (Z) = 0; in the right
panel we have heterogeneous CATE s0 (Z) = Z. The proxy predictor S(Z) for CATE
is produced by the random forest, whose sorted averages by groups are shown as
red dots, exhibiting large biases. These are the naive estimates. The true sorted
group average treatment effects (GATES) E[s0 (Z) | Gk ] are shown by black dots, and
estimated GATES are shown by blue dots. The true and estimated GATES correct
for the biases relative to the naive strategy shown in red. The estimated GATES
shown by blue dots are always closer to the true GATEs shown by black dots than
the naive estimates shown in red.
Figure 2 provides two examples using the same designs as in fig. 1. Post-processing the ML
estimates again has stronger effect when there is no heterogeneity, but in both cases help bring the
estimated GATES close to the true GATES.
GATES: The Second Strategy. Here we employ linear projections on Horvitz-Thompson transformed variables:
K
X
Y H = µ0 X1 H +
γk · 1(Gk ) + ν, E[ν W̃ ] = 0,
(2.5)
k=1
for B := B(Z), S := S(Z), W̃ = (X10 H, W̃20 ), W̃20 = ({1(Gk )}K
k=1 ).
Again, we show that the projection parameters are GATES:
K
γ = (γk )K
k=1 = (E[s0 (Z) | Gk ])k=1 .
14
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
Given the identification strategy, we can base the corresponding estimation strategy on the following empirical analogs:
Yi Hi = µ
b0 X1i Hi + γ
b0 W̃2i + νbi ,
i ∈ M, EN,M [b
νi W̃i ] = 0.
(2.6)
The properties of this estimator, conditional on the auxiliary data, are well known and given in
Lemma B.1. The resulting estimator has similar performance to the previous estimator, and under
some conditions their first-order properties coincide.
The following is the formal statement of the identification result.
Theorem 2.3 (GATES). Consider z 7→ S(z) and z 7→ B(z) as fixed maps. Assume that Y has finite
second moments and the W ’s and W̃ defined above are such that EW W 0 and EW̃ W̃ 0 are finite and have full
rank. Consider γ = (γk )K
k=1 defined by the weighted regression equation (2.3) or by the regression equation
(2.5). These parameters defined in two different ways are equivalent and are equal to the expectation of s0 (Z)
conditional on the proxy group {S ∈ Ik }:
γk = E[s0 (Z) | Gk ].
2.3. Classification Analysis (CLAN). When the BLP and GATES analyses reveal substantial heterogeneity, it is interesting to know the properties of the subpopulations that are most and least
affected. Here we focus on the “least affected group” G1 and “most affect group” GK . Under the
monotonicity assumption, it is reasonable that the first and the last groups are the most and least
affected, where the labels “most” and “least” can be swapped depending on the context.
Let g(Y, Z) be a vector of characteristics of an observational unit. The parameters of interest are
the average characteristics of the most and least affected groups:
δ1 = E[g(Y, Z) | G1 ]
and
δK = E[g(Y, Z) | GK ].
The parameters δK and δ1 are identified because they are averages of variables that are directly
observed. We can compare δK and δ1 to quantify differences between the most and least affected
groups. We call this type of comparisons as classification analysis or CLAN.
3. ”Variational” Estimation and Inference Methods
3.1. Estimation and Inference: The Generic Targets. Let θ denote a generic target parameter or
functional, for example,
• θ = β2 is the heterogeneity predictor loading parameter;
• θ = β1 + β2 (S(z) − ES) is the “personalized” prediction of s0 (z);
• θ = γk is the expectation of s0 (Z) for the group Gk ;
• θ = γK − γ1 is the difference in the expectation of s0 (Z) between the most and least affected
groups;
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
15
• θ = δK − δ1 is the difference in the expectation of the characteristics of the most and least
impacted groups.
3.2. Quantification of Uncertainty: Two Sources. There are two principal sources of sampling
uncertainty:
(I) Estimation uncertainty regarding the parameter θ, conditional on the data split;
(II) Uncertainty or ”variation” induced by the data splitting.
Conditional on the data split, quantification of estimation uncertainty is standard. To account
for uncertainty with respect to the data splitting, it makes sense to examine the robustness and
variability of the estimates/confidence intervals with respect to different random splits. One of our
goals is to develop methods, which we call ”variational estimation and inference” (VEIN) methods,
for quantifying this uncertainty, which can be of independent interest in many settings where the
sample splitting is used.
Quantifying Source (I): Conditional Inference. We first recognize that the parameters implicitly
depend on
DataA := {(Yi , Di , Xi )}i∈A ,
the auxiliary sample, used to create the ML proxies B = BA and S = SA . Here we make the
dependence explicit: θ = θA .
All of the examples admit an estimator θbA such that under mild assumptions,
2
θbA | DataA ∼a N (θA , σ
bA
),
in the sense that, as |M | → ∞,
−1 b
P(b
σA
(θA − θA ) 6 z | DataA ) →P Φ(z).
Implicitly this requires the auxiliary data DataA to be ”sufficiently regular”, and this should happen with high probability.
As a consequence, the confidence interval (CI)
[LA , UA ] := [θbA ± Φ−1 (1 − α/2)b
σA ]
covers θA with approximate probability 1 − α:
P[θA ∈ [LA , UA ] | DataA ] = 1 − α − oP (1).
This leads to straighforward conditional inference, which does not account for the sample splitting
uncertainty.
16
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
Quantifying Source (II): “Variational” Inference. Different partitions (A, M ) of {1, ..., N } yield
different targets θA . Conditional on the data, we treat θA as a random variable, since (A, M ) are
random sets that form random partitions of {1, . . . , N } into samples of size n and N − n. Different partitions also yield different estimators θbA and approximate distributions for these estimators.
Hence we need a systematic way of treating the randomness in these estimators and their distributions.
Comment 3.1. In cases where the data sets are not large, it may be desirable to restrict attention to
balanced partitions (A, M ), where the proportion of treated units is equal to the designed propensity score.
We want to quantify the uncertainty induced by the random partitioning. Conditional on Data,
the estimated θbA is still a random variable, and the confidence band [LA , UA ] is a random set. For
reporting purposes, we instead would like to report an estimator and confidence set, which are
non-random conditional on the data.
Adjusted Point and Interval Estimators. Our proposal is as follows. As a point estimator,
we shall report the median of θbA as (A, M ) vary (as random partitions):
θb := Med[θbA | Data].
This estimator is more robust than the estimator based on a single split. To account for partition uncertainty, we propose to report the following confidence interval (CI) with the nominal
confidence level 1 − 2α:
[l, u] := [Med[LA | Data], Med[UA | Data]].
Note that the price of splitting uncertainty is reflected in the discounting of the confidence level
from 1 − α to 1 − 2α. Alternatively, we can report the confidence interval based on inversion
of a test based upon p-values, constructed below.
The above estimator and confidence set are non-random conditional on the data. The confidence set reflects the uncertainty created by the random partitioning of the data into the main and
auxilliary data.
Comment 3.2. For a random variable X with law PX we define
Med(X) := inf{x ∈ R : PX (X 6 x) > 1/2},
Med(X) := sup{x ∈ R : PX (X > x) > 1/2},
Med(X) := (Med(X) + Med(X))/2.
Note that the lower median is the usual definition of the median. The upper median is the next
distinct quantile of the random variable (or it is the usual median after reversing the order on R).
For example, when X is uniform on {1, 2, 3, 4}, then Med(X) = 2 and Med(X) = 3; and if X is
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
17
uniform on {1, 2, 3}, then Med(X) = Med(X) = 2. For continuous random variables the upper
and lower medians defined above coincide. For discrete random variables they can differ, but the
differences will be small for variables that are close to being continuous.
Suppose we are testing H0 : θA = θ0 against H1 : θA < θ0 , conditional on the auxiliary data,
then the p-value is given by
pA = Φ(b
σ −1 (θbA − θ0 )).
A
−1 b
The p-value for testing H0 : θA = θ0 against H1 : θA > θ0 , is given by pA = 1 − Φ(b
σA
(θA − θ0 )).
Under the null hypothesis pA is approximately distributed as the uniform variable, pA ∼ U (0, 1),
conditional on DataA . Note that, conditional on Data, pA still has randomness induced by random
partitioning of the data, which we need to address.
Adjusted P-values. We say that testing the null hypothesis, based on the p-values pA , that
are random conditional on data, has significance level α if
P(pA 6 α/2 | Data) > 1/2
or p.5 = Med(pA | Data) 6 α/2.
That is, for at least 50% of the random data splits, the realized p-value pA falls below the level
α/2. Hence we can call p = 2p.5 the sample splitting-adjusted p-value, and consider its small
values as providing evidence against the null hypothesis.
Comment 3.3. Our construction of p-values builds upon the false-discovery-rate type adjustment
ideas in [16, 42], though what we propose is radically simpler, and is minimalistic for our problem,
whereas the idea of our confidence intervals below appears to be completely new.
The main idea behind this construction is simple: the p-values are distributed as marginal uniform variables {Uj }j∈J , and hence obey the following property.
Lemma 3.1 (A Property of Uniform Variables). Consider M , the (usual, lower) median of a sequence
{Uj }j∈J of uniformly distributed variables, Uj ∼ U (0, 1) for each j ∈ J, where variables are not necessarily
independent. Then,
P(M 6 α/2) 6 α.
Proof. Let M denote the median of {Uj }j∈J . Then M 6 α/2 is equivalent to |J|−1
α/2)] − 1/2 > 0. So
X
P[M 6 α/2] = E1{|J|−1
[1(Uj 6 α/2)] > 1/2}.
P
j∈J [1(Uj
6
j∈J
By Markov inequality this is bounded by
X
2E|J|−1
[1(Uj 6 α/2)] 6 2E[1(Uj 6 α/2)] 6 2α/2 = α.
j∈J
where the last inequality holds by the marginal uniformity.
18
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
Main Inference Result: Variational P-values and Confidence Intervals. We present a formal result on adjusted p-values using this condition:
PV. Suppose that A is a set of regular auxiliary data configurations such that for all x ∈ [0, 1],
under the null hypothesis:
sup |PP [pA 6 x | DataA ∈ A] − x| 6 δ = o(1),
P ∈P
and inf P ∈P PP [DataA ∈ A] =: 1 − γ = 1 − o(1). In particular, suppose that this holds for
the p-values
−1 b
−1 b
pA = Φ(b
σA
(θA − θA )) and pA = 1 − Φ(b
σA
(θA − θA )).
Lemma B.1 shows that this condition is plausible for the least squares estimators defined in the
previous section under mild conditions.
Theorem 3.1 (Uniform Validity of Variational P-Value). Under condition PV and the null hypothesis
holding,
PP (p.5 6 α/2) 6 α + 2(δ + γ) = α + o(1),
uniformly in P ∈ P.
In order to establish the properties of the confidence interval [l, u], we first consider the properties of the related confidence interval, which is based on the inversion of the p-value based tests:
CI := {θ ∈ R : pu (θ) > α/2, pl (θ) > α/2},
(3.1)
for α < .25 , where, for σ
bA > 0,
−1 b
σA
(θA − θ)] | Data),
pl (θ) := Med(1 − Φ[b
(3.2)
−1 b
pu (θ) := Med(Φ[b
(θA − θ)] | Data).
σA
(3.3)
The confidence interval CI has the following representation in terms of the medians of t-statistics
implied by the proof Theorem 3.2 stated below:
h b
i
θA
−1 (1 − α/2) | Data < 0
Med θ−
−
Φ
h σbAb
i
CI = θ ∈ R :
.
(3.4)
Med θ−θA + Φ−1 (1 − α/2) | Data > 0
σ
bA
This CI can be (slightly) tighter than [l, u], while the latter is much simpler to construct.
The following theorem establishes that both confidence sets maintain the approximate confidence level 1 − 2α.
Theorem 3.2 (Uniform Validity of Variational Confidence Intervals). CI can be represented as (3.4)
and CI ⊆ [l, u], and under condition PV,
PP (θA ∈ CI) > 1 − 2α − 2(δ + γ) = 1 − 2α − o(1),
uniformly in P ∈ P.
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
19
4. Other Considerations and Extensions
1. Choosing the Best ML Method Targeting CATE in Stage 1. There are several options. The
best ML method can be chosen using the auxiliary sample, based on either (A) the ability to predict
Y H using BH and S or (B) the ability to predict Y using B and (D − p(Z))(S − E(S)) under the
weight w(Z) (as in the first type of strategies we developed earlier). To be specific, we can solve
either of the following problems:
(A) minimize the errors in the prediction of Y H on BH and S:
X
(B, S) = arg min
[Yi Hi − B(Zi )Hi − S(Zi )]2 ,
B∈B,S∈S
i∈A
where B and S are parameter spaces for z 7→ B(z) and z 7→ S(z); or
(B) minimize the errors in the weighted prediction of Y on B and (D − p(Z))(S − E(S)):
X
(B, S) = arg min
w(Zi )[Yi − B(Zi ) − (Di − p(Zi )){S(Zi ) − S̄(Zi )}]2 ,
B∈B,S∈S
where S̄(Zi ) = |A|−1
z 7→ S(z).
i∈A
P
i∈A S(Zi )
and B and S are parameter spaces for z 7→ B(z) and
This idea improves over simple but inefficient strategy of predicting Y H just using S, which have
been suggested before for causal inference. It also improves over the simple strategy that predicts
Y using B and DS (which chooses the best predictor for E[Y | D, Z] in a given class but not necessarily the best predictor for CATE s0 (Z)). Note that this idea is new and is of major independent
interest.
2. Choosing the Best ML Method BLP Targeting CATE in Stage 2. The best ML method can
also be chosen in the main sample by maximizing
Λ := |β2 |2 Var(S(Z)) = Corr2 (s0 (Z), S(Z))Var(s0 (Z)).
(4.1)
Maximizing Λ is equivalent to maximizing the correlation between the ML proxy predictor S(Z)
and the true score s0 (Z), or equivalent to maximizing the R2 in the regression of s0 (Z) on S(Z).
3. Choosing the Best ML Method GATES Targeting CATE in Stage 2. Analogously, for GATES
the best ML method can also be chosen in the main sample by maximizing
Λ̄ = E
K
X
k=1
!2
γk 1(S ∈ Ik )
=
K
X
γk2 P(S ∈ Ik ).
(4.2)
k=1
P
This is the part of variation Es20 (Z) of s0 (z) explained by S̄(Z) = K
k=1 γk 1(S(Z) ∈ Ik ). Hence
choosing the ML proxy S(Z) to maximize Λ̄ is equivalent to maximizing the R2 in the regression
20
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
of s0 (Z) on S̄(Z) (without a constant). If the groups Gk = {S ∈ Ik } have equal size, namely
P(S(Z) ∈ Ik ) = 1/K for each k = 1, ..., K, then
Λ̄ =
K
1 X 2
γk .
K
k=1
4. Stratified Splitting. The idea is to balance the proportions of treated and untreated in both
A and M samples, so that the proportion of treated is equal to the experiment’s propensity scores
across strata. This formally requires us to replace the i.i.d. assumption by the i.n.i.d. assumption
(independent but not identically distributed observations) when accounting for estimation uncertainty, conditional on the auxiliary sample. This makes the notation more complicated, but the
results in Lemma B.1 still go through with notational modifications.
5. When Proxies have Little Variation. The analysis may generate proxy predictors S that have
little variation, so we can think of them as “weak”, which makes the parameter β2 weakly identified.
We can either add small noise to the proxies, which is called jittering, so that inference results go
through, or we may switch to testing rather than estimation. For practical reasons, we prefer the
jittering approach.
5. Further Potential Applications to Prediction and Causal Inference Problems
Our inference approach generalizes to any problem of the following sort.
Generalization. Suppose we can construct an unbiased signal Ỹ such that
E[Ỹ | Z] = s0 (Z),
where s0 (Z) is now a generic target function. Let S(Z) denote an ML proxy for s0 (Z). Then,
using previous arguments, we immediately can generate the following conclusions:
(1)
(2)
(3)
(4)
(5)
The projection of Ỹ on the ML proxy S(Z) identifies the BLP of s0 (Z) using S(Z).
The grouped averages of the target (GAT) E[s0 (Z) | Gk ] are identified by E[Ỹ | Gk ].
Using ML tools we can train proxy predictors S(Z) to predict Ỹ in auxiliary samples.
We post-process S(Z) in the main sample, by estimating the BLP and GATs.
We apply variational inference on functionals of the BLP and GATs.
The noise reduction strategies, like the ones we used on the context of H-transformed outcomes,
can be useful in these cases, but their construction could depend on the context.
Example 1. Forecasting or Predicting Regression Functions using ML proxies. This is the most
common type of the problem arising in forecasting. Here the target is the best predictor of Y using
Z, namely s0 (Z) = E[Y | Z], and Ỹ = Y trivially serves as the unbiased signal. The interesting part
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
21
here is the use of variational inference tools developed in this paper for constructing confidence
intervals for the predicted values produced by the estimated BLP of s0 (Z) using S(Z).
Example 2. Predicting Structural Derivatives using ML proxies. Suppose we are interested in
best predicting the partial derivative s0 (z), where s0 (z) = ∂x g(x, z), where g(x, z) = E[Y | X =
x, Z = z]. In the context of demand analysis, Y is the log of individual demand, X is the log-price
of a product, and Z includes prices of other products and characteristics of individuals. Then the
unbiased signal is given by Ỹ = Y [∂x log p(X | Z)], where p(x | z) is the conditional density
function of x given z. That is, E[Ỹ | Z] = s0 (Z) under mild conditions on the density using the
integration by parts formula.
6. Empirical Applications and Implementation Algorithms
We consider two empirical applications. The first application is an RCT conducted in Morocco,
which investigates the effects of microfinance access on several outcomes. The second application
is an observational study, which analyzes the gender wage gap using the U.S. Current Population
Survey (CPS). While the results of the second study should not be interpreted as causal, we aim to
obtain the best approximation to the causal effects of gender-based discrimination on wages using
observational data, by flexibly conditioning on many controls using ML methods. We conclude
this section by providing the implementation algorithms for both examples.
6.1. Heterogeneity in the Effect of Microcredit Availability. We analyze a randomized experiment designed to evaluate the impact of microcredit availability on borrowing and self-employment
activities, which was previously studied in [24]. The experiment was conducted in 162 villages
in Morocco, divided into 81 pairs of villages with similar observable characteristics (number of
households, accessibility to the center of the community, existing infrastructure, type of activities
carried out by the households, and type of agriculture activities). One of the villages in each pair
was randomly assigned to treatment and the other to control. Between 2006 and 2007 a microfinance institution started operating in the treated villages. Two years after the intervention an
endline household survey was conducted with 5,551 households, which constitute our sample.
There was no other microcredit penetration in these villages, before and for the duration of the
study. Therefore, we interpret the treatment as the availability of microcredit.
Recent randomized evaluations of access to microcredit at the community level have found limited impacts of microcredit9. Despite evidence that access to microfinance leads to an increase in
borrowing ([4], [10], [48]) and business creation or expansion ([4], [7], [10], [48]), most studies have
found that this does not translate into an increase in economic outcomes such as profit, income,
labor supply and consumption ([4], [10], [24]). Moreover, there is also no evidence of substantial
gains in human development outcomes, such as education and health ([10],[48]). Studies which
9See [11] for a summary of the recent literature
22
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
estimate the impact of microfinance by randomizing microcredit at the individual level confirm
these findings ([8], [36], [37]).
One question that remains elusive is whether the lack of evidence on the average effects masks
heterogeneity, in which there are potential winners and losers of the microcredit expansion. Understanding this heterogeneity can have important implications for evaluating the welfare effects
of microcredit, designing policies and targeting the groups that would benefit from microfinance.
Indeed, the idea that there might be heterogeneity in the impact of microcredit has been a common theme among RCTs evaluating microfinance programs. Having found mostly positive but
insignificant coefficients, the papers cited above attempt to explore heterogeneous treatment effects, mostly using quantile treatment effects. For profits, most studies seem to find positive impact at the higher quantiles (and in the data set we study here, [24] actually find negative impacts at
lower level). Using Bayesian hierarchical methods to aggregate the evidence across studies, Meager (2017) cautions that these results on quantiles may not be generalizable: the profit variables
seems to have too much noise to lend itself to quantile estimation.
A number of recent papers also consider heterogeneous treatment effects by studying the effect
of microfinance on a subpopulation. In a follow-up study of [10], [9] investigates whether the heterogeneity is persistent six years after the microfinance was introduced. They find that credit has a
much bigger impact on the business outcomes of those who started a business before microfinance
entered than of those without prior businesses. Using the same dataset as in this application [24]
classify households into three categories in term of their probability to borrow before the intervention and find that microcredit access has a significant impact on investment and profit, but still no
impact on income and consumption among those who are most likely to borrow. It is worth noting
that the original strategy for this study was to construct groups which, ex ante had different probability to borrow, in order to separately estimate direct effect of microcredit on those most likely
to borrow, and indirect on those very unlikely to borrow. The researchers had initially collected a
short survey on the entire village, and had then tried to predict the probability to borrow using a
model fitted in a first group of villages. But the model proved to have low predictive power in the
other villages, and in their paper then end up predicting the probability to borrow ex-post. In the
paper, they worry that this ex-post classification leads them to over-fit.
The identification strategy developed in this paper provides several advantages in studying heterogeneity in the treatment effects of microfinance. First, contrary to the literature, which relies on
ad hoc subgroup analysis across a few baseline characteristics, we are agnostic about the source of
heterogeneity. While the variable “had a prior business” has proven to be a robust and generalizable predictors of differences in treatment effect (Meager, 2017) and could therefore be pre-specified
in future pre-analysis plans, we have little idea about what else predict heterogeneity. Second, our
approach is valid in high dimensional settings, allowing us to include a rich set of characteristics
in an unspecified functional form. Finally, using the CLAN estimation we are able to identify the
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
23
characteristics of the most and least affected subpopulations, which could be an important input
for a welfare analysis or targeting households who are likely to benefit from access to microfinance.
We focus on heterogeneity in treatment effects on four household outcome variables, Y : the
amount of money borrowed, the output from self-employment activities, profit from self-employment
activities, and monthly consumption. The treatment variable, D, is an indicator for the household
residing in a treated village. The covariates, Z, include some baseline household characteristics
such as number of members, number of adults, head age, indicators for households doing animal husbandry, doing other non-agricultural activity, having an outstanding loan over the past 12
months, household spouse responded to the survey, another household member (excluding the
HH head) responded to the survey, and 81 village pair fixed effects (these are the variables that
are available for all households). We also include indicators for missing observation at baseline as
controls. Table 6 shows some descriptive statistics for the variables used in the analysis (all monetary variables are expressed in Moroccan Dirams, or MAD). Treated and control households have
similar characteristics and the unconditional average treatment effect on loans, output, profit and
consumption are respectively 1,128, 5,237, 1,844 and -31.
Table 1. Descriptive Statistics of Households
All
Treated
Control
Outcome Variables
Total Amount of Loans
2,359
2,930
1,802
Total output from self-employment activities (past 12 months)
32,499
35,148
29,911
Total profit from self-employment activities (past 12 months)
10,102
11,035
9,191
Total monthly consumption
3,012
2,996
3,027
Number of Household Members
3.879
3.872
3.886
Number of Members 16 Years Old or Older
2.604
2.601
2.607
Head Age
35.976
35.937
36.014
Declared Animal Husbandry Self-employment Activity
0.415
0.426
0.404
Declared Non-agricultural Self-employment Activity
0.146
0.129
0.164
Borrowed from Any Source
0.210
0.224
0.196
Spouse of Head Responded to Self-employment Section
0.067
0.074
0.061
Member Responded to Self-employment Section
0.044
0.048
0.041
Baseline Covariates
We implement our methods using the algorithm described in Section 6.3. By design the propensity score p(Zi ) = 1/2 for all the households. Table 2 compares the four ML methods for producing
the proxy predictors S(Zi ) considered in Stage 1. We find that the Random Forest and Elastic Net
24
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
Table 2. Comparison of ML Methods: Microfinance Availability
Elastic Net
Boosting
Neural Network
Random Forest
2,808,960
2,086,256
2,464,276
2,706,767
875
421
630
1253
143,620,159
78,127,786
76,945,520
133,825,537
8,697
2,823
4,219
5,229
32,307,828
17,105,855
20,404,000
39,286,050
4,595
2,319
1,847
4,478
44,583
26,387
34,019
36,927
101
101
96
109
Amount of Loans
Best BLP (Λ)
Best GATES (Λ̄)
Output
Best BLP (Λ)
Best GATES (Λ̄)
Profit
Best BLP (Λ)
Best GATES (Λ̄)
Consumption
Best BLP (Λ)
Best GATES (Λ̄)
Notes: Medians over 100 splits in half.
outperform the Boosted Tree and Neural Network across all outcome variables for both metrics.
Accordingly, we focus on these two methods for the rest of the analysis.10
Table 3 presents results of the BLP of CATE using the ML proxies S(Z) for the four outcome
variables. We report estimates of the coefficients β1 and β2 , which correspond to the ATE and
heterogeneity loading (HET) parameters in the BLP, respectively. In parentheses, we report confidence intervals adjusted for variability across the sample splits using the median method; and
in brackets, we report adjusted p-values. The estimated ATEs of microfinance availability are consistent with the findings of [24] and are similar to the unconditional ATE, as expected by virtue
of the randomization. The ATE on the amount of loans and output are positive and statistically
significant at least at the 10% level with both ML methods. Microfinance availability does not have
a significant impact on profit and consumption.
Turning to the heterogeneity results, we reject the hypothesis that HET is zero at the 10% level
for the amount of loans, output and profit with the elastic net method, suggesting the presence
of heterogeneity in the effect of microfinance availability. The results are consistent across both
ML methods except for output, for which HET coefficient on the Random Forest proxy is not significantly different from zero at the 10%. Finally, the BLP analysis does not reveal any significant
heterogeneity in the effect on consumption. Overall, these results suggest that microfinance availability has heterogenous impacts on business-related outcomes that does not seem to translate into
10The results obtained using Boosted Tree and Neural Network are similar to the results reported, but they are slightly
less precise. These results are not reported but are available from the authors upon requests.
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
25
Table 3. BLP of Microfinance Availability
Elastic Net
Amount of Loans
Output
Profit
Consumption
Random Forest
ATE (β1 )
HET (β2 )
ATE (β1 )
HET (β2 )
1,163
0.238
1,185
0.375
(544,1736)
(0.021,0.448)
(561,1771)
(0.028,0.774)
[0.000]
[0.060]
[0.000]
[0.069]
5,095
0.262
5,027
0.192
(232,10033)
(0.085,0.433)
(-89,10194)
(-0.100,0.508)
[0.079]
[0.008]
[0.109]
[0.391]
1,553
0.244
1,603
0.279
(-1344,4389)
(0.079,0.416)
(-1276,4536)
(0.046,0.518)
[0.584]
[0.008]
[0.521]
[0.039]
-59.1
0.157
-58.6
0.196
(-161.5,44.2)
(-0.058,0.385)
(-166.6, 43.3)
(-0.160,0.574)
[0.514]
[0.278]
[0.508]
[0.553]
Notes: Medians over 100 splits. 90% confidence interval in parenthesis.
P-values for the hypothesis that the parameter is equal to zero in brackets.
a detectable contemporaneous effect on the standard of living as represented by consumption, even
for the most positively affected households.
We next estimate the GATES. We divide the households into K = 5 groups based on the quintiles
of the ML proxy predictor S(Z) and estimate the average effect for each group. Figures 3-6 present
the estimated GATES coefficients γ1 − γ5 along with joint confidence bands. We also report the
ATE and its confidence interval that were obtained in the BLP analysis for comparison. The GATES
provide a richer understanding of the heterogeneity. In particular, the figures reveal that there are
groups of winners, the most affected groups, for which the GATES on amount of loans, output
and profit are significantly different from zero. These groups are likely to drive the heterogeneity
in the treatment effect that we find in the BLP analysis. We further investigate the GATES by
comparing the most and least affected groups in Table 4. Here we find that the difference of GATES
of these two groups is significantly different from zero at least at the 10% level on amount of loans,
level on output and profit, whereas we fail to reject the hypothesis that this difference is zero at
conventional levels on consumption. Looking at the least affected group, it is reassuring to see that
we have no evidence of negative impact on profit and income, mitigating the concerns that there are
adversely affected households. However, there is negative and insignificant effect on consumption
for the same group. A possible explanation for this result is that investment is lumpy and some
households cut back consumption to increase investment. All the results for the GATES are fairly
robust to the ML method.
26
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
Table 4. GATES of 20% Most and Least Affected Groups
Elastic Net
Amount of Loans
Output
Profit
Consumption
Random Forest
20% Most
(γ5 )
20% Least
(γ1 )
Difference
(γ5 − γ1 )
20% Most
(γ5
20% Least
(γ1 )
Difference
(γ5 − γ1 )
2,677
-197
2,995
2,870
94.707
2,814
(1298,4076)
(-1835,1307)
(945,5103)
(1149,4587)
(-1663,1723)
(503,5193)
[0.000]
[1.000]
[0.008]
[0.002]
[1.000]
[0.032]
22,367
-3,039
25,088
21,606
626
21,035
(7678,36920)
(-12546,6535)
(7028,42698)
(5862,38022)
(-11871,13529)
(125,43170)
[0.007]
[1.000]
[0.015]
[0.015]
[1.000]
[0.097]
10,644
-1,152.242
11,768
11,540
-2,031
14,037
(2146,19096)
(-7250,4952)
(1077,22422)
(2965,20955.576)
(-8721,4796)
(2459,25833)
[0.028]
[1.000]
[0.061]
[0.014]
[1.000]
[0.037]
66.4
-333
383
62
-300
332
(-166.2,289.8)
(-695.6,23.2)
(-38.0,805.6)
(-271,346)
(-683,66)
(-196,835)
[1.000]
[0.140]
[0.152]
[1.000]
[0.228]
[0.429]
Notes: Medians over 100 splits. 90% confidence interval in parenthesis.
P-values for the hypothesis that the parameter is equal to zero in brackets.
We conclude by looking at the average characteristics of the most and least affected groups to
understand what generates heterogeneity in the treatment effects. We omit the results for consumption as we do not detect heterogeneity for this outcome. We focus on three characteristics in
this analysis: Age of household head, non-agricultural self-employment activity and whether the
household borrowed from any source. Table 5 reports the CLAN for the 10% least and most affected
groups defined by the deciles of the CATE proxy S(Z) as well as the difference between the two.
We find households with young heads, non-agriculture self-employment and that borrowed less
from any source are likely to borrow more from the microfinance institution, suggesting that formal loans are substitutes rather than complements. For output only the average self-employment
sector indicator is found to be significantly different between the most and least affected groups. Finally, the estimates for profit are similar to the estimates of amount of loans. It is important to note
that the significance of some of these differences is sensitive to the ML method used to generate
the proxy predictors, while the sign of the differences is robust.
6.2. Gender Wage Gap in the U.S.. We consider an application to the gender wage gap using
data from the U.S. March Supplement of the Current Population Survey (CPS) in 2015. We select
white, non-hispanic individuals who are aged 25 to 64 years and work more than 35 hours per week
during at least 50 weeks of the year. We exclude self-employed workers; individuals living in group
quarters; individuals in the military, agricultural or private household sectors; individuals with
inconsistent reports on earnings and employment status; individuals with allocated or missing
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
27
Table 5. CLAN of Microfinance Availability
Elastic Net
Random Forest
10% Most
(δ10 )
10% Least
(δ1 )
Difference
(δ10 − δ1 )
10% Most
(δ10 )
10% Least
(δ1 )
Difference
(δ10 − δ1 )
29.3
(26.3,32.4)
0.199
(0.159,0.238)
0.144
(0.099,0.189)
-
35.2
(32.2,38.2)
0.068
(0.030,0.108)
0.169
(0.124,0.212)
-
-6.6
(-10.9,-2.4)
[0.004]
0.123
(0.069,0.178)
[0.000]
-0.038
(-0.101,0.025)
[0.448]
23.0
(19.8,26.2)
0.134
(0.096,0.173)
0.109
(0.064,0.153)
-
33.8
(30.5,36.9)
0.118
(0.076,0.156)
0.217
(0.175,0.262)
-
-10.4
(-14.9,-6.0)
[0.000]
0.022
(-0.033,0.075)
[0.875]
-0.107
(-0.164,-0.050)
[0.001]
36.280
(33.4,39.1)
0.275
(0.233,0.315)
0.193
(0.142,0.241)
-
36.708
(33.6,39.6)
0.050
(0.007,0.093)
0.215
(0.167,0.262)
-
-0.896
(-5.242,3.432)
[1.000]
0.226
(0.169,0.285)
[0.000]
-0.033
(-0.102,0.034)
[0.687]
29.090
(25.8,32.3)
0.215
(0.172,0.257)
0.165
(0.121,0.208)
-
30.831
(27.5,34.1)
0.088
(0.045,0.129)
0.189
(0.146,0.234)
-
-1.925
(-6.648,2.799)
[0.849]
0.130
(0.070,0.190)
[0.000]
-0.024
(-0.086,0.039)
[0.895]
34.1
(31.2,37.0)
0.181
(0.140,0.222)
0.180
(0.130,0.230)
-
40.4
(37.5,43.4)
0.108
(0.068,0.149)
0.257
(0.207,0.307)
-
-6.5
(-10.7,-2.5)
[0.003]
0.082
(0.022,0.138)
[0.014]
-0.091
(-0.160,-0.022)
[0.020]
29.2
(25.7,32.6)
0.153
(0.113,0.192)
0.144
(0.098,0.190)
-
33.7
(30.390,37.108)
0.099
(0.058,0.139)
0.162
(0.122,0.206)
-
-5.8
(-10.566,-1.217)
[0.029]
0.051
(-0.003,0.105)
[0.129]
-0.032
(-0.095,0.029)
[0.578]
Amount of Loans
Head Age
Non-agricultural self-emp.
Borrowed from Any Source
Output
Head Age
Non-agricultural self-emp.
Borrowed from Any Source
Profit
Head Age
Non-agricultural self-emp.
Borrowed from Any Source
Notes: Medians over 100 splits. 90% confidence interval in parenthesis.
P-values for the hypothesis that the parameter is equal to zero in brackets.
information in any of the variables used in the analysis; and individuals with hourly wage rate
below $3. The resulting sample consists of 32, 523 workers including 18, 137 men and 14, 382 of
women.
We will estimate the BLP, GATES and CLAN of the conditional gender wage gap, i.e., the difference in the average wage between female and male workers with the same observable characteristics; see [18] for a recent survey on this topic. The outcome variable Y is the logarithm of
the hourly wage rate constructed as the ratio of the annual earnings to the total number of hours
worked, which is constructed in turn as the product of number of weeks worked and the usual
28
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
number of hours worked per week. The treatment D is an indicator for female worker, and the observable characteristics Z include 5 marital status indicators (widowed, divorced, separated, never
married, and married); 5 educational attainment indicators (less than high school graduate, high
school graduate, some college, college graduate, and advanced degree); 4 region indicators (midwest, south, west, and northeast); potential experience constructed as the maximum of age minus
years of schooling minus 7 and zero, i.e., experience = max(age − education − 7, 0); 5 occupation
indicators (management, professional and related; service; sales and office; natural resources, construction and maintenance; and production, transportation and material moving); and 12 industry
indicators (mining, quarrying, and oil and gas extraction; construction; manufacturing; wholesale
and retail trade; transportation and utilities; information; financial services; professional and business services; education and health services; leisure and hospitality; other services; and public
administration).11
Table 6. Descriptive Statistics of Workers
All
Women
Men
All
Women
Men
Log wage
3.16
3.02
3.26
O.manager
0.49
0.56
0.43
Female
0.44
1.00
0.00
O.service
0.09
0.10
0.09
MS.married
0.70
0.65
0.73
O.sales
0.22
0.30
0.16
MS.widowed
0.01
0.02
0.01
O.construction
0.09
0.01
0.15
MS.separated
0.02
0.02
0.01
O.production
0.11
0.04
0.17
MS.divorced
0.12
0.15
0.09
I.minery
0.03
0.01
0.04
MS.Nevermarried
0.16
0.16
0.16
I.construction
0.06
0.01
0.09
E.lhs
0.02
0.01
0.03
I.manufacture
0.13
0.08
0.18
E.hsg
0.25
0.21
0.28
I.retail
0.13
0.11
0.14
E.sc
0.28
0.30
0.27
I.transport
0.04
0.02
0.06
E.cg
0.28
0.29
0.27
I.information
0.02
0.02
0.03
E.ad
0.17
0.19
0.15
I.finance
0.08
0.10
0.07
R.northeast
0.26
0.26
0.26
I.professional
0.11
0.10
0.12
R.midwest
0.27
0.28
0.27
I.education
0.25
0.41
0.12
R.south
0.33
0.34
0.33
I.leisure
0.04
0.05
0.04
R.west
0.22
0.21
0.23
I.services
0.03
0.03
0.04
Experience
21.21
21.03
21.35
I.public
0.07
0.07
0.08
Source: March Supplement CPS 2015
Table 6 reports sample means of the variables used in the analysis. Working women are more
highly educated than working men, have about the same potential experience, and are less likely
to be married and more likely to be divorced. They work relatively more often in managerial and
sales occupations and in the industries providing education and health services. Working men
are relatively more likely to work in construction and production occupations within non-service
industries. The unconditional gender wage gap is 24%.
11The sample selection and variable construction is the same as in [20].
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
29
In this application we do not observe the propensity scores. We estimate them from a large
sample of 129,983 workers from the U.S. March Supplement of the CPS pooling the years 2012 to
2017, excluding 2015, as described in Section 6.3. We ignore the sampling error in the estimated
propensity scores for the rest of the analysis because the sample size to estimate these scores is
much larger than the sample size to estimate the parameters of interest.
Table 7. Comparison of ML Methods
Elastic Net Boosting Neural Network Random Forest
Best BLP (Λ)
0.084
0.079
0.076
0.082
Best GATES (Λ̄)
0.060
0.058
0.058
0.059
Notes: Medians over 1,000 splits in half.
Table 7 compares the ML methods. In this case the elastic net comes out as the winner both
based on BLP and GATES targeting of CATE, followed closely by the random forest. As in the
previous section, we focus on these two ML methods for the rest of the analysis. Table 8 shows the
results for the BLP coefficients. The average conditional gender wage gap of −22% comes close to
the unconditional gap, but the slope of the BLP uncovers substantial heterogeneity across workers.
The estimates and joint confidence intervals for the GATES in Figure 7 confirm this finding. Here
we divide the sample in K = 5 groups defined by the quintiles of the CATE proxy S(Z) and show
that average conditional gender wage gap ranges from about −33% to −9% across the groups.
Table 9 compares the averages of the gender wage gap in the first (most affected) and fifth (least
affected) groups. We reject that the average gaps for the 20% most and least affected are equal at
any conventional level. The results for the BLP and GATES are robust to the ML method used in
Stage 1.
Table 8. BLP of Gender Wage Gap
Elastic Net
ATE (β1 )
HET (β2 )
Random Forest
ATE (β1 )
-0.229
0.726
-0.228
(-0.251,-0.207) (0.540,0.912) (-0.249,-0.207)
[0.000]
[0.000]
[0.000]
HET (β2 )
0.552
(0.409,0.694)
[0.000]
Notes: Medians over 1,000 splits. 90% confidence interval in parenthesis.
P-values for the hypothesis that the parameter is equal to zero in brackets.
Table 10 reports the results of a classification analysis where we divide the workers in K = 10
groups defined by the deciles of the CATE proxy S(Z) and compare the average characteristics of
30
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
Table 9. GATES of Gender Wage Gap
20% Least Affected 20% Most Affected Difference
Elastic Net
Random Forest
(γ5 )
(γ1 )
(γ5 − γ1 )
-0.088
-0.328
0.240
(-0.136,-0.040)
(-0.377,-0.279)
(0.171,0.308)
[0.001]
[0.000]
[0.000]
-0.095
-0.329
0.234
(-0.142,-0.048)
(-0.377,-0.280)
(0.166,0.301)
[0.000]
[0.000]
[0.000]
Notes: Medians over 1,000 splits. 90% confidence interval in parenthesis.
P-values for the hypothesis that the parameter is equal to zero in brackets.
the most and least affected group. This analysis reveals another dimension of the heterogeneity in
the gender wage gap by showing that the 10% most affected workers are more likely to have either
of the following characteristics relative to the 10% less affected workers on average: high wages,
high experience, not advanced degree, married, not divorced or never married. These differences
are robust to the two ML method. The elastic net and random forest, however, disagree in the
comparison of the averages of some of the education variables such as high school graduate and
some college. Due to this sensitivity, we recommend reporting results from at least the two best
ML methods. Overall, the heterogeneity patterns that we find are consistent with both glass ceiling
theories for the gender wage gap [3], and with heterogenous individual preferences where young,
single, and highly educated women are more career-oriented. [20] found similar patterns applying
semiparametric regression methods to the same data.
6.3. Implementation Algorithm. In this section we describe an algorithm based on the first identification strategy and provide some specific implementation details for the two empirical examples.
Algorithm 1 (Inference Algorithm). The inputs are given by the data on units i ∈ [N ] = {1, ..., N }.
Step 0. Fix the number of splits S and the significance level α, e.g. S = 100 and α = 0.05.
Step 1. Compute the propensity scores p(Zi ) for i ∈ [N ].
Step 2. Consider S splits in half of the indices i ∈ {1, ..., N } into the main sample, M , and the
auxiliary sample, A. Over each split s = 1, .., S, apply the following steps:
a. Tune and train each ML method separately to learn B(·) and S(·) using A. For each i ∈ M ,
compute the predicted baseline effect B(Zi ) and predicted treatment effect S(Zi ). If there is
zero variation in B(Zi ) and S(Zi ) add Gaussian noise with a variance of 0.1 to the proxies.
b. Estimate the BLP parameters by weighted OLS in M , i.e.,
Yi = α
b0 X1i + βb1 (Di − p(Zi )) + βb2 (Di − p(Zi ))(Si − EN,M Si ) + b
i , i ∈ M
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
31
Table 10. CLAN of Gender Wage Gap
Elastic Net
Log Wage
Experience
Random Forest
10% Least
10% Most
Difference
10% Least
10% Most
Difference
(δ10 )
(δ1 )
(δ10 − δ1 )
(δ10 )
( δ1 )
(δ10 − δ1 )
2.949
(2.918,2.981)
-
3.436
(3.404,3.467)
-
-0.484
(-0.528,-0.440)
[0.000]
3.074
(3.043,3.106)
-
3.233
(3.201,3.264)
-
-0.157
(-0.201,-0.113)
[0.000]
12.902
25.664
-12.778
14.176
25.109
-10.893
(12.386,13.419) (25.14,26.187) (-13.513,-12.055) (13.666,14.687) (24.602,25.622) (-11.615,-10.176)
[0.000]
[0.000]
Less than H. School
0.039
(0.030,0.048)
-
0.023
(0.014,0.032)
-
0.017
(0.004,0.028)
[0.016]
0.025
(0.016,0.034)
-
0.039
(0.030,0.047)
-
-0.013
(-0.025,-0.001)
[0.078]
High School
0.246
(0.226,0.266)
-
0.167
(0.147,0.187)
-
0.080
(0.052,0.109)
[0.000]
0.162
(0.141,0.182)
-
0.268
(0.248,0.288)
-
-0.106
(-0.134,-0.078)
[0.000]
Some College
0.184
(0.166,0.201)
-
0.094
(0.076,0.110)
-
0.089
(0.066,0.112)
[0.000]
0.186
(0.166,0.206)
-
0.238
(0.217,0.258)
-
-0.050
(-0.079,-0.021)
[0.001]
College Graduate
0.326
(0.303,0.35)
-
0.570
(0.546,0.594)
-
-0.235
(-0.269,-0.202)
[0.000]
0.361
(0.337,0.384)
-
0.354
(0.33,0.377)
-
0.010
(-0.023,0.043)
[1.000]
Advanced Degree
0.201
(0.182,0.22)
-
0.135
(0.116,0.154)
-
0.066
(0.04,0.091)
[0.000]
0.259
(0.241,0.278)
-
0.096
(0.078,0.115)
-
0.164
(0.137,0.19)
[0.000]
Married
0.101
(0.088,0.114)
-
0.953
(0.94,0.966)
-
-0.851
(-0.869,-0.832)
[0.000]
0.359
(0.338,0.38)
-
0.861
(0.84,0.881)
-
-0.499
(-0.529,-0.47)
[0.000]
Widowed
0.010
(0.005,0.015)
-
0.008
(0.003,0.012)
-
0.003
(-0.004,0.009)
[0.728]
0.008
(0.004,0.013)
-
0.011
(0.006,0.016)
-
-0.002
(-0.009,0.005)
[0.981]
Divorced
0.092
(0.08,0.103)
-
0.019
(0.008,0.03)
-
0.072
(0.056,0.088)
[0.000]
0.095
(0.081,0.109)
-
0.083
(0.069,0.097)
-
0.012
(-0.008,0.031)
[0.478]
Separated
0.028
(0.021,0.035)
-
0.010
(0.003,0.017)
-
0.017
(0.008,0.027)
[0.001]
0.013
(0.006,0.019)
-
0.021
(0.014,0.027)
-
-0.008
(-0.017,0.001)
[0.170]
Nevermarried
0.764
(0.749,0.779)
-
0.003
(-0.011,0.018)
-
0.76
(0.738,0.781)
[0.000]
0.523
(0.505,0.541)
-
0.022
(0.004,0.04)
-
0.500
(0.474,0.526)
[0.000]
Notes: Medians over 1,000 splits. 90% confidence interval in parenthesis.
P-values for the hypothesis that the parameter is equal to zero in brackets.
0 , D − p(Z ), (D − p(Z ))(S − E
0
such that EN,M [w(Zi )b
i Xi ] = 0 for Xi = [X1i
i
i
i
i
i
N,M Si )] , where
w(Zi ) = {p(Zi )(1 − p(Zi ))}−1 and X1i includes a constant, B(Zi ) and S(Zi ).
32
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
c. Estimate the GATES parameters by weighted OLS in M , i.e.,
0
Yi = α
b X1i +
K
X
γ
bk · (Di − p(Zi )) · 1(Si ∈ Ik ) + νbi , i ∈ M,
k=1
0 , {(D − p(Z ))1(S ∈ I )}K ]0 , where w(Z ) =
such that EN,M [w(Zi )b
νi Wi ] = 0 for Wi = [Xi1
i
i
i
i
k k=1
{p(Zi )(1 − p(Zi ))}−1 , X1i includes a constant, B(Zi ) and S(Zi ), Ik = [`k−1 , `k ), and `k is the
(k/K)-quantile of {Si }i∈M .
d. Estimate the CLAN parameters in M by
δb1 = EN,M [g(Yi , Zi ) | Si ∈ I1 ]
and
δbK = EN,M [g(Yi , Zi ) | Si ∈ IK ],
where Ik = [`k−1 , `k ) and `k is the (k/K)-quantile of {Si }i∈M .
e. Compute the two performance measures for the ML methods
b = |βb2 |2 Var(S(Z))
d
Λ
K
X
b̄ = 1
γ
bk2 .
Λ
K
k=1
b̄ over the splits.
b and Λ
Step 3: Choose the best ML methods based on the medians of Λ
Step 4: Compute the estimates, (1 − α)-level conditional confidence intervals and conditional
p-values for all the parameters of interest. Monotonize the confidence intervals if needed. For
example, construct a (1 − α) joint confidence interval for the GATES as
{b
γk ± b
c(1 − α)b
σk , k = 1, . . . , K},
(6.1)
where b
c(1 − α) is a consistent estimator of the (1 − α)-quantile of maxk∈1,...,K |b
γk − γk |/b
σk and σ
bk
is the standard error of γ
bk conditional on the data split. Monotonize the band (6.1) with respect to
k using the rearrangement method of [19].
Step 5: Compute the adjusted (1 − 2α)-confidence intervals and adjusted p-values using the
VEIN methods described in Section 3.
Comment 6.1 (ML Methods). We consider four ML methods to estimate the proxy predictors:
elastic net, boosted trees, neural network with feature extraction, and random forest. The ML
methods are implemented in R using the package caret [39]. The names of the elastic net, boosted
tree, neural network with feature extraction, and random forest methods in caret are glmnet, gbm,
pcaNNet and rf, respectively. For each split of the data, we choose the tuning parameters separately
for B(z) and S(z) based on mean squared error estimates of repeated 2-fold cross-validation, except
for random forest, for which we use the default tuning parameters to reduce the computational
time.12 In tuning and training the ML methods we use only the auxiliary sample. In all the methods
we rescale the outcomes and covariates to be between 0 and 1 before training.
12We have the following tuning parameters for each method: Elastic Net: alpha (Mixing Percentage), lambda (Reg-
ularization Parameter), Boosted trees: n.trees (Number of Boosting Iterations), interaction.depth (Max Tree Depth),
shrinkage (Shrinkage), n.minobsinnode (Min. Terminal Node Size), size (Number of Hidden Units) , decay (Weight
Decay), mtry (Number of Randomly Selected Predictors).
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
33
Comment 6.2 (Microfinance Application). We adopt two strategies to improve precision. First,
the linear projections of the BLP and GATES control for village pair fixed effects along with the
predicted baseline effect, B(z) and predicted treatment effect, S(z). Second, as suggested in Section
4, we use stratified sample splitting where the strata are village pairs. We cluster the standard errors
at the village level to account for potential correlated shocks within each village. All reported
results are medians over S = 100 splits and α = 0.05.
Comment 6.3 (Gender Wage Gap Application). We estimate the propensity scores using a neural
network with feature extraction from a large sample of 129,983 workers from the U.S. March Supplement of the CPS pooling the years 2012 to 2017, excluding 2015. We select the neural network
among elastic net, boosted tree, neural network, and random forest, and the tuning parameters by
repeated 2-fold cross validation on the mean squared error. To avoid tuning ML methods for each
data split separately, we tune ML methods by using 20% random extract from the 2012-2017 pooled
sample and use those tuning parameters for all splits. This allows us to use S = 1, 000 splits in this
empirical application. Finally, to reduce the disproportionate impact of extreme propensity score
weights we trimmed the predicted propensity scores between 0.01 and 0.99 in the main sample.
These dropped observations make up approximately 4% of the sample.
7. Concluding Remarks
We propose to focus inference on key features of heterogeneous effects in randomized experiments, and develop the corresponding methods. These key features include best linear predictors
of the effects and average effects sorted by groups, as well as average characteristics of most and
least affected units. Our new approach is valid in high dimensional settings, where the effects are
estimated by machine learning methods. The main advantage of our approach is its credibility: the
approach is agnostic about the properties of the machine learning estimators, and does not rely on
incredible or hard-to-verify assumptions. Estimation and inference relies on data splitting, where
the latter allows us to avoid overfitting and all kinds of non-regularities. Our inference quantifies uncertainty coming from both parameter estimation and the data splitting, and could be of
independent interest. Empirical applications illustrate the practical uses of the approach.
Appendix A. Proofs
Proof of Theorem 2.1. The subset of the the normal equations, which correspond to β1 and β2 , are
given by E[w(Z)(Y − α0 X1 − β 0 X2 )X2 ] = 0. Substituting Y = b0 (Z) + s0 (Z)D + U , and using the
definition X2 = X2 (Z, D) = [D − p(Z), (D − p(Z)(S − ES)]0 , that X1 = X1 (Z), and the law of
34
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
iterated expectations, we notice that:
E[w(Z)b0 (Z)X2 ] = E[w(Z)b0 (Z) E[X2 | Z]] = 0,
=0
E[w(Z)U X2 ] = E[w(Z) E[U | Z, D] X2 (Z, D)] = 0,
0
E[w(Z)X1 X2 ] = E[w(Z)X1 (Z) E[X2 (Z, D) | Z]] = 0.
=0
Hence the normal equations simplify to: E[w(Z)(s0 (Z)D − β 0 X2 )X2 ] = 0. Since
E[{D − p(Z)}{D − p(Z)} | Z] = p(Z)(1 − p(Z)) = w−1 (Z),
and S = S(Z), the components of X2 are orthogonal by the law of iterated expectations:
Ew(Z)(D − p(Z))(D − p(Z))(S − ES) = E(S − ES) = 0.
Hence the normal equations above further simplify to
E[w(Z){s0 (Z)D − β1 (D − p(Z))}(D − p(Z))] = 0,
E[w(Z){s0 (Z)D − β2 (D − p(Z))(S − ES)}(D − p(Z))(S − ES)] = 0.
Solving these equations and using the law of iterated expectations, we obtain
β1 =
Ew(Z)s0 (Z)w−1 (Z)
Ew(Z){s0 (Z)D(D − p(Z))}
=
= Es0 (Z),
Ew(Z)(D − p(Z))2
Ew(Z)w−1 (Z)
Ew(Z){s0 (Z)D(D − p(Z))(S − ES)}
Ew(Z)(D − p(Z))2 (S − ES)2
Ew(Z)s0 (Z)w−1 (Z)(S − ES)
=
= Cov(s0 (Z), S)/Var(S).
Ew(Z)w−1 (Z)(S − ES)2
The conclusion follows by noting that these coefficients also solve the normal equations
β2 =
E{[s0 (Z) − β1 − β2 (S − ES)][1, (S − ES)]0 } = 0,
which characterize the optimum in the problem of best linear approximation/prediction of s0 (Z)
using S.
Proof of Theorem 2.2. The normal equations defining β = (β10 , β20 ) are given by E[(Y H − µ0 X1 H −
β 0 X̃2 )X̃2 ] = 0. Substituting Y = b0 (Z) + s0 (Z)D + U , and using the definition X̃2 = X̃2 (Z) =
[1, (S(Z) − ES(Z))]0 , that X1 = X1 (Z), and the law of iterated expectations, we notice that:
E[b0 (Z)H X̃2 (Z)] = E[b0 (Z) E[H | Z] X̃2 (Z)] = 0,
=0
E[U H X̃2 (Z)] = E[E[U | Z, D] H(D, Z)X̃2 (Z)] = 0,
0
E[X1 (Z)H X̃2 (Z)] = E[X1 (Z) E[H | Z] X̃2 (Z)] = 0.
=0
Hence the normal equations simplify to:
E[(s0 (Z)DH − β 0 X̃2 )X̃2 ] = 0.
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
35
Since 1 and S − ES are orthogonal, the normal equations above further simplify to
E{s0 (Z)DH − β1 } = 0,
E[{s0 (Z)DH − β2 (S − ES)}(S − ES)] = 0.
Using that
E[DH | Z] = [p(Z)(1 − p(Z))]/[p(Z)(1 − p(Z))] = 1,
S = S(Z), and the law of iterated expectations, the equations simplify to
E{s0 (Z) − β1 } = 0,
E{s0 (Z) − β2 (S − ES)}(S − ES) = 0.
These are normal equations that characterize the optimum in the problem of best linear approximation/prediction of s0 (Z) using S. Solving these equations gives the expressions for β1 and β2
stated in the theorem.
Proof of Theorem 2.3. The proof is similar to the proof of Theorem 2.1- 2.2. Moreover, since the
proofs for the two strategies are similar, we will only demonstrate the proof for the second strategy.
The subset of the normal equations, which correspond to γ = (γk )K
k=1 , are given by E[(Y H −
0
0
µ W̃1 − β W̃2 )W̃2 ] = 0. Substituting Y = b0 (Z) + s0 (Z)D + U , and using the definition W̃2 =
0
W̃2 (Z) = [1(S ∈ Ik )K
k=1 ] , that W̃1 = X1 (Z)H, and the law of iterated expectations, we notice that:
E[b0 (Z)H W̃2 (Z)] = E[b0 (Z) E[H | Z] W̃2 (Z)] = 0,
=0
E[U H W̃2 (Z)] = E[E[U | Z, D] H(D, Z)W̃2 (Z)] = 0,
0
E[W̃1 W̃2 (Z)] = E[X1 (Z) E[H | Z] W̃2 (Z)] = 0.
=0
Hence the normal equations simplify to:
E[{s0 (Z)DH − β 0 W̃2 }W̃2 ] = 0.
0
Since components of W̃2 = W̃2 (Z) = [1(Gk )K
k=1 ] are orthogonal, the normal equations above further simplify to
E[{s0 (Z)DH − γk 1(Gk )}1(Gk )] = 0.
Using that
E[DH | Z] = [p(Z){1 − p(Z)}]/[p(Z){1 − p(Z)}] = 1,
that S = S(Z), and the law of iterated expectations, the equations simplify to
E[{s0 (Z) − γk 1(Gk )}1(Gk )] = 0 ⇐⇒
γk = Es0 (Z)1(Gk )/E[1(Gk )] = E[s0 (Z) | Gk ].
36
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
Proof of Theorem 3.1. We have that p.5 6 α/2 is equivalent to EP [1(pA 6 α/2) | Data] > 1/2. So
PP [p.5 6 α/2] = EP 1{EP [1(pA 6 α/2) | Data] > 1/2}.
By Markov inequality,
EP 1{EP [1(pA 6 α/2) | Data] > 1/2} 6 2PP [pA 6 α/2]
Moreover,
PP (pA 6 α/2) 6 EP [PP [pA 6 α/2 | DataA ∈ A] + γ] 6 α/2 + δ + γ.
Proof of Theorem 3.2. To show the second claim, we note that
PP (θA 6∈ CI) = PP (pl (θA ) 6 α/2) + PP (pu (θA ) 6 α/2)
6 α + δ + γ + α + δ + γ,
where the inequality holds by Theorem 3.1 on the p-values. The last bound is upper bounded by
2α + o(1) by the regularity condition PV for the p-values, uniformly in P ∈ P.
To show the first claim, we need to show the following inequalities:
sup{θ ∈ R : pu (θ) > α/2} 6 u,
inf{θ ∈ R : pl (θ) > α/2} > l.
We demonstrate the first inequality, and the second follows similarly.
We have that
−1 b
(θA − θ)} | Data] > α/2}
σA
{θ ∈ R : pu (θ) > α/2} = {θ ∈ R : Med[Φ{b
−1 b
= {θ ∈ R : Φ{Med[b
(θA − θ) | Data]} > α/2}
σA
−1 b
= {θ ∈ R : Med[b
(θA − θ) | Data] > Φ−1 (α/2)}
σA
−1
(θ − θbA ) | Data] < Φ−1 (1 − α/2)}
= {θ ∈ R : Med[b
σA
(
"
#
)
θ − θbA
−1
=
θ ∈ R : Med
− Φ (1 − α/2) | Data < 0 ,
σ
bA
where we have used the equivariance of Med and Med to monotone transformations, implied from
their definition. We claim that by the definition of
u := Med[θbA + σ
bA Φ−1 (1 − α/2) | Data],
we have
"
#
u − θbA
−1
Med
− Φ (1 − α/2) | Data > 0.
σ
bA
Indeed, by the definition of u we have that
E 1(u − θbA − σ
bA Φ−1 (1 − α/2) > 0) | Data > 1/2.
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
37
Since σ
bA > 0 by assumption we have
1(u − θbA − σ
bA Φ
−1
(1 − α/2) > 0) = 1
!
u − θbA
−1
− Φ (1 − α/2) > 0 ,
σ
bA
and it follows that
P
u − θb
A
σ
bA
− Φ−1 (1 − α/2) > 0 | Data > 1/2.
The claimed inequality sup{θ ∈ R : pu (θ) > α/2} 6 u follows.
Appendix B. A Lemma on Uniform in P Conditional Inference
Lemma B.1. Fix positive constants c and C, and a small constant δ > 0. Let Ỹ and X denote generic
outcome and generic d-vector of regressors, whose use and definition may differ in different places of the
paper. Assume that for each P ∈ P, EP |Ỹ |4+δ < C and let 0 < w 6 w(Z) 6 w < ∞ denote the generic
weights, and that {(Ỹi , Zi , Di )}N
i=1 are i.i.d. copies of (Ỹ , Z, D). Let {DataA ∈ AN } be the event such
that the ML algorithm, operating only on DataA , produces a vector XA = X(Z, D; DataA ) that obeys, for
A = Ỹ − X 0 βA defined by: EP [A w(Z)XA | DataA ] = 0, the following inequalities, uniformly in P ∈ P
0
0
EP [kXA k4+δ | DataA ] 6 C, mineig EP [XA XA
| DataA ] > c, mineig EP [2A XA XA
| DataA ] > c.
Suppose that PP {DataA ∈ AN } > 1 − γ → 1 uniformly in P ∈ P, as N → ∞. Let βbA be defined by:
EN,M [w(Z)XAb
A ] = 0,
b
A = YA − X 0 βbA .
0 −1 be an estimator of
0 (E
0 )−1 E
2A XA XA
Let VbN,A := (EN,M XA XA
N,M XA XA )
N,M b
0
0
0
VN,A = (EP [XA XA
| DataA ])−1 EP [2A XA XA
| DataA ](EP [XA XA
| DataA ])−1 .
Then for any convex set R in Rd , we have that uniformly in P ∈ P:
−1/2
PP [VbN,A (βbA − βA ) ∈ R | DataA ] →P P(N (0, I) ∈ R),
−1/2
PP [VbN,A (βbA − βA ) ∈ R | {DataA ∈ AN }] → P(N (0, I) ∈ R),
and the same results hold with VbN,A replaced by VN,A .
Proof. It suffices to demonstrate the argument for an arbitrary sequence {Pn } in P. Let z 7→
X̃A,N (z) be a deterministic map such that the following inequalities hold, for ẽA defined by
EPn [ẽA w(Z)X̃A,N (Z)] = 0
and X̃A = X̃A (Z):
0
0
EPn [kX̃A,N k4 ] < C, mineig EPn [X̃A,N X̃A,N
] > c, mineig EPn [ẽ2A X̃A,N X̃A,N
] > c.
Let d := dim βA . Then we have that (abusing notation):
BN := sup
sup
X̃A,N h∈BL1
(Rd )
−1/2
|EPn h(ṼN,A (βbA − βA ) | X̃A,N ) − Eh(N (0, I))| → 0,
38
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
by the standard argument for asymptotic normality of the least squares estimator, which utilizes
the Lindeberg-Feller Central Limit Theorem. Here
0 −1
0
0 −1
) ,
(EX̃A X̃A
) E˜
2A X̃A X̃A
ṼN,A := (EX̃A X̃A
and BL1 (Rd ) denotes the set of Lipschitz maps h : Rd → [0, 1] with the Lipschitz coefficient
bounded by 1.
Then, for the stochastic sequence XA,N = XA,N (DataA ), we have
sup
h∈BL1
(Rd )
−1/2
|EPn [h(VN,A (βbA − βA )) | XA,N ] − E[h(N (0, I))]| 6 BN + 2(1 − 1{DataA ∈ AN }) →Pn 0.
1/2 −1/2
Since under the stated bounds on moments, VbN,A VN,A →Pn I by the standard argument for consistency of the Eicker-Huber-White sandwich, we further notice that
sup
h∈BL1 (Rd )
−1/2
−1/2
|EPn [h(VbN,A (βbA − βA )) | XA,N ] − EPn [h(VN,A (βbA − βA )) | XA,N ]|
−1/2 1/2
−1/2
6 EPn [kVbN,A VN,A − Ik ∧ 1 · kVN,A (βbA − βA )k ∧ 1 | XA,N ] →Pn E[0 ∧ 1 · kN (0, I)k ∧ 1] = 0,
in order to conclude that
sup
h∈BL1
(Rd )
−1/2
EPn [h(VbN,A (βbA − βA )) | XA,N ] − E[h(N (0, I))] →P 0.
−1/2
−1/2
Moreover, since EPn [h(VbN,A (βbA − βA )) | XA,N ] = EPn [h(VbN,A (βbA − βA )) | DataA ], the first
−1/2
conclusion follows: PPn [VbN,A (βbA − βA ) ∈ R | DataA ] →Pn P(N (0, I) ∈ R), by the conventional
smoothing argument (where we approximate the indicator of a convex region by a smooth map
with finite Lipschitz coefficient). The second conclusion
−1/2
PPn [VbN,A (βbA − βA ) ∈ R | DataA ∈ AN ] → P(N (0, I) ∈ R)
follows from the first by
−1/2
PPn [VbN,A (βbA − βA ) ∈ R | DataA ∈ AN ] =
−1/2
= EPn [PPn [VbN,A (βbA − βA ) ∈ R | DataA ]1({DataA ∈ AN })/PPn {DataA ∈ AN }]
→ E[P(N (0, I) ∈ R) · 1],
using the definition of the weak convergence, implied by the convergence to the constants in probability.
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
39
References
[1] Alberto Abadie. Semiparametric difference-in-differences estimators. The Review of Economic Studies, 72(1):1–19,
2005.
[2] Alberto Abadie, Matthew M Chingos, and Martin R West. Endogenous stratification in randomized experiments.
Technical report, National Bureau of Economic Research, 2017.
[3] James Albrecht, Anders Bjorklund, and Susan Vroman. Is there a glass ceiling in sweden? Journal of Labor Economics,
21(1):145–177, 2003.
[4] Manuela Angelucci, Dean Karlan, and Jonathan Zinman. Microcredit impacts: Evidence from a randomized microcredit program placement experiment by compartamos banco. American Economic Journal: Applied Economics,
7(1):151–182, 2015.
[5] Susan Athey and Guido Imbens. Recursive partitioning for heterogeneous causal effects. Proceedings of the National
Academy of Sciences, 113(27):7353–7360, 2016.
[6] Susan Athey and Guido W Imbens. The econometrics of randomized experiments. Handbook of Economic Field Experiments, 1:73–140, 2017.
[7] Orazio Attanasio, Britta Augsburg, Ralph De Haas, Emla Fitzsimons, and Heike Harmgart. The impacts of microfinance: Evidence from joint-liability lending in mongolia. American Economic Journal: Applied Economics, 7(1):90–122,
2015.
[8] Britta Augsburg, Ralph De Haas, Heike Harmgart, and Costas Meghir. Microfinance, poverty and education. 2012.
[9] Abhijit Banerjee, Emily Breza, Esther Duflo, and Cynthia Kinnan. Do credit constraints limit entrepreneurship?
heterogeneity in the returns to microfinance. Evanston, USA: Department of Economics Northwestern University, 2015.
[10] Abhijit Banerjee, Esther Duflo, Rachel Glennerster, and Cynthia Kinnan. The miracle of microfinance? evidence
from a randomized evaluation. American Economic Journal: Applied Economics, 7(1):22–53, 2015.
[11] Abhijit Vinayak Banerjee. Microcredit under the microscope: what have we learned in the past two decades, and
what do we need to know? Annu. Rev. Econ., 5(1):487–519, 2013.
[12] G. Barnard. Discussion of “Cross-validatory choice and assessment of statistical predictions” by Stone. Journal of
the Royal Statistical Society. Series B (Methodological), page 133?135, 1974.
[13] A. Belloni, V. Chernozhukov, and C. Hansen. Inference on treatment effects after selection amongst highdimensional controls. Review of Economic Studies, 81:608–650, 2014.
[14] Alexandre Belloni, Victor Chernozhukov, and Kengo Kato. Uniform post selection inference for lad regression
models. arXiv preprint arXiv:1304.0282, 2013.
[15] Alexandre Belloni, Victor Chernozhukov, and Lie Wang. Pivotal estimation of nonparametric functions via squareroot lasso. arXiv preprint arXiv:1105.1475, 2011.
[16] Yoav Benjamini and Yosef Hochberg. Controlling the false discovery rate: a practical and powerful approach to
multiple testing. Journal of the royal statistical society. Series B (Methodological), pages 289–300, 1995.
[17] P. J. Bickel, Y. Ritov, and A. B. Tsybakov. Simultaneous analysis of Lasso and Dantzig selector. Annals of Statistics,
37(4):1705–1732, 2009.
[18] Francine D. Blau and Lawrence M. Kahn. The gender wage gap: Extent, trends, and explanations. Journal of Economic
Literature, 55(3):789–865, September 2017.
[19] V. Chernozhukov, I. Fernández-Val, and A. Galichon. Improving point and interval estimators of monotone functions by rearrangement. Biometrika, 96(3):559–575, 2009.
[20] V. Chernozhukov, I. Fernandez-Val, and Y. Luo. The Sorted Effects Method: Discovering Heterogeneous Effects
Beyond Their Averages. ArXiv e-prints, December 2015.
[21] Victor Chernozhukov, Denis Chetverikov, Mert Demirer, Esther Duflo, Christian Hansen, Whitney Newey, and
James Robins. Double/debiased machine learning for treatment and structural parameters. The Econometrics Journal, 2017.
40
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
[22] Victor Chernozhukov, Denis Chetverikov, and Kengo Kato. Anti-concentration and honest, adaptive confidence
bands. The Annals of Statistics, 42(5):1787–1818, 2014.
[23] DR Cox. A note on data-splitting for the evaluation of significance levels. Biometrika, 62(2):441–444, 1975.
[24] Bruno Crépon, Florencia Devoto, Esther Duflo, and William Parienté. Estimating the impact of microcredit on those
who take it up: Evidence from a randomized experiment in morocco. American Economic Journal: Applied Economics,
7(1):123–150, 2015.
[25] Jonathan Davis and Sara B Heller. Rethinking the benefits of youth employment programs: The heterogeneous
effects of summer jobs. Technical report, National Bureau of Economic Research, 2017.
[26] Ruben Dezeure, Peter Bühlmann, and Cun-Hui Zhang. High-dimensional simultaneous inference with the bootstrap. arXiv preprint arXiv:1606.03940, 2016.
[27] Esther Duflo, Rachel Glennerster, and Michael Kremer. Using randomization in development economics research:
A toolkit. Handbook of development economics, 4:3895–3962, 2007.
[28] Jerome Friedman, Trevor Hastie, and Robert Tibshirani. The elements of statistical learning, volume 1. Springer series
in statistics New York, 2001.
[29] Christopher Genovese and Larry Wasserman. Adaptive confidence bands. The Annals of Statistics, pages 875–905,
2008.
[30] Evarist Giné and Richard Nickl. Confidence bands in density estimation. The Annals of Statistics, 38(2):1122–1170,
2010.
[31] Christian Hansen, Damian Kozbur, and Sanjog Misra. Targeted undersmoothing. arXiv preprint arXiv:1706.07328,
2017.
[32] John A Hartigan. Using subsample values as typical values. Journal of the American Statistical Association,
64(328):1303–1317, 1969.
[33] Keisuke Hirano, Guido W. Imbens, and Geert Ridder. Efficient estimation of average treatment effects using the
estimated propensity score. Econometrica, 71(4):1161–1189, 2003.
[34] Kosuke Imai and Marc Ratkovic. Estimating treatment effect heterogeneity in randomized program evaluation. The
Annals of Applied Statistics, 7(1):443–470, 2013.
[35] Guido W Imbens and Donald B Rubin. Causal inference in statistics, social, and biomedical sciences. Cambridge University Press, 2015.
[36] Dean Karlan and Jonathan Zinman. Expanding credit access: Using randomized supply decisions to estimate the
impacts. The Review of Financial Studies, 23(1):433–464, 2009.
[37] Dean Karlan and Jonathan Zinman. Microcredit in theory and practice: Using randomized credit scoring for impact
evaluation. Science, 332(6035):1278–1284, 2011.
[38] Leslie Kish and Martin Richard Frankel. Inference from complex samples. Journal of the Royal Statistical Society. Series
B (Methodological), pages 1–37, 1974.
[39] Max Kuhn. Caret package. Journal of Statistical Software, 28(5):1–26, 2008.
[40] Guillaume Lecué and Charles Mitchell. Oracle inequalities for cross-validation type procedures. Electronic Journal
of Statistics, 6:1803–1837, 2012.
[41] Mark G Low et al. On nonparametric confidence intervals. The Annals of Statistics, 25(6):2547–2554, 1997.
[42] Nicolai Meinshausen, Lukas Meier, and Peter Bühlmann. P-values for high-dimensional regression. Journal of the
American Statistical Association, 104(488):1671–1681, 2009.
[43] Frederick Mosteller and John Wilder Tukey. Data analysis and regression: a second course in statistics. AddisonWesley Series in Behavioral Science: Quantitative Methods, 1977.
[44] Natalia Rigol, Reshmaan Hussam, and Benjamin Roth. Targeting high ability entrepreneurs using community information: Mechanism design in the field, 2016.
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
41
[45] Alessandro Rinaldo, Larry Wasserman, Max G’Sell, Jing Lei, and Ryan Tibshirani. Bootstrapping and sample splitting for high-dimensional, assumption-free inference. arXiv preprint arXiv:1611.05401, 2016.
[46] Donald B Rubin. Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of
educational Psychology, 66(5):688, 1974.
[47] Charles J. Stone. Optimal global rates of convergence for nonparametric regression. Ann. Statist., 10(4):1040–1053,
1982.
[48] Alessandro Tarozzi, Jaikishan Desai, and Kristin Johnson. The impacts of microcredit: Evidence from ethiopia.
American Economic Journal: Applied Economics, 7(1):54–89, 2015.
[49] Stefan Wager and Susan Athey. Estimation and inference of heterogeneous treatment effects using random forests.
Journal of the American Statistical Association, (just-accepted), 2017.
[50] Larry Wasserman. Machine learning overview. In Becker-Friedman Institute, Conference on ML in Economics, 2016.
[51] Larry Wasserman and Kathryn Roeder. High dimensional variable selection. Annals of statistics, 37(5A):2178, 2009.
[52] Marten Wegkamp et al. Model selection in nonparametric regression. The Annals of Statistics, 31(1):252–273, 2003.
[53] Qingyuan Zhao, Dylan S Small, and Ashkan Ertefaie. Selective inference for effect modification via the lasso. arXiv
preprint arXiv:1705.08020, 2017.
42
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
Elastic Net
Random Forest
ATE
90% CB(ATE)
ATE
90% CB(ATE)
GATES
90% CB(GATES)
GATES
90% CB(GATES)
1
2
4000
Treatment Effect
Treatment Effect
4000
2000
2000
0
0
−2000
−2000
1
2
3
4
5
Group by Het Score
3
4
5
Group by Het Score
Figure 3. GATES of Microfinance Availability: Amount of Loans. Point estimates
and 90% adjusted confidence intervals uniform across groups based on 100 random
splits in half
Elastic Net
40000
ATE
90% CB(ATE)
GATES
90% CB(GATES)
Random Forest
40000
90% CB(ATE)
GATES
90% CB(GATES)
1
2
Treatment Effect
20000
Treatment Effect
20000
ATE
0
0
−20000
−20000
1
2
3
Group by Het Score
4
5
3
4
Group by Het Score
Figure 4. GATES of Microfinance Availability: Output. Point estimates and 90%
adjusted confidence intervals uniform across groups based on 100 random splits in
half
5
GENERIC ML FOR FEATURES OF HETEROGENOUS TREATMENT EFFECTS
Elastic Net
ATE
90% CB(ATE)
GATES
90% CB(GATES)
Random Forest
20000
10000
Treatment Effect
Treatment Effect
20000
43
0
−10000
ATE
90% CB(ATE)
GATES
90% CB(GATES)
1
2
10000
0
−10000
1
2
3
4
5
3
Group by Het Score
4
5
Group by Het Score
Figure 5. GATES of Microfinance Availability: Profit. Point estimates and 90% adjusted confidence intervals uniform across groups based on 100 random splits in
half
Elastic Net
ATE
90% CB(ATE)
GATES
90% CB(GATES)
500
Treatment Effect
Treatment Effect
500
Random Forest
0
−500
ATE
90% CB(ATE)
GATES
90% CB(GATES)
1
2
0
−500
1
2
3
Group by Het Score
4
5
3
4
Group by Het Score
Figure 6. GATES of Microfinance Availability: Consumption. Point estimates and
90% adjusted confidence intervals uniform across groups based on 100 random
splits in half
5
44
V. CHERNOZHUKOV, M. DEMIRER, E. DUFLO, I. FERNANDEZ-VAL
Elastic Net
Random Forest
ATE
90% CB(ATE)
ATE
90% CB(ATE)
GATES
90% CB(GATES)
GATES
90% CB(GATES)
−0.1
Treatment Effect
Treatment Effect
−0.1
−0.2
−0.3
−0.2
−0.3
−0.4
−0.4
1
2
3
Group by Het Score
4
5
1
2
3
4
Group by Het Score
Figure 7. GATES of gender wage gap for K = 5 groups defined by the quintiles of
S(Z). Point estimates and 90% adjusted confidence intervals uniform across groups
based on 1,000 random splits in half
5
| 10math.ST
|
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED
GROUPS
arXiv:1706.07571v1 [math.GR] 23 Jun 2017
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
Abstract. We have developed polynomial-time algorithms to generate terms of the
cogrowth series for groups Z o Z, the lamplighter group, (Z o Z) o Z and the Navas-Brin
group B. We have also given an improved algorithm for the coefficients of Thompson’s
group F, giving 32 terms of the cogrowth series. We develop numerical techniques to
extract the asymptotics of these various cogrowth series. We present improved rigorous
lower bounds on the growth-rate of the cogrowth series for Thompson’s group F using
the method from [18] applied to our extended series. We also generalise their method by
showing that it applies to loops on any locally finite graph. Unfortunately, lower bounds
less than 16 do not help in determining amenability.
Again for Thompson’s group F we prove that, if the group is amenable, there cannot be
a sub-dominant stretched exponential term in the asymptotics1. Yet the numerical data
provides compelling evidence for the presence of such a term. This observation suggests
a potential path to a proof of non-amenability: If the universality class of the cogrowth
sequence can be determined rigorously, it will likely prove non-amenability.
We estimate the asymptotics of the cogrowth coefficients of F to be
cn ∼ c · µn · κn
σ
logδ n
· ng ,
where µ ≈ 15, κ ≈ 1/e, σ ≈ 1/2, δ ≈ 1/2, and g ≈ −1. The growth constant µ must
be 16 for amenability. These two approaches, plus a third based on extrapolating lower
bounds, support the conjecture [7, 18] that the group is not amenable.
1. Introduction
In an attempt to find compelling evidence for the amenability or otherwise of Thompson’s
group F , we have studied, numerically, the co-growth sequence of a number of infinite,
finitely generated amenable groups whose asymptotics are, in most cases, partially or fully
known. We have chosen a number of examples with increasingly complex asymptotics.
Using the experience and insights gained from these examples, we turn to a study of
Thompson’s group F , having first developed an improved algorithm for the generation of
the co-growth sequence, which we evaluate to O(x32 ).
The cogrowth series of a group G with finite, inverse closed, generating set S is
X
CG =
cn xn ,
n≥0
where cn is the number of words w of length 2n over the alphabet S, which satisfy w =G 1
i.e. w is the identity in the group G. There are many equivalent definitions of amenability.
A standard one is that a group G is amenable if it admits a left-invariant finitely additive
1
2
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
probability measure µ. A consequence of the Grigorchuk-Cohen [11, 5] theorem is that G is
amenable if and only if the radius of convergence of CG is 1/|S|2 . In particular, Thompson’s
group F amenable if and only if its cogrowth sequence has exponential growth rate 16.
We have developed new, polynomial-time algorithms to generate coefficients for the
lamplighter group, and for general wreath product groups, Wd = Z od Z. We also give a
polynomial time algorithm for the cogrowth coefficients of the Navas-Brin group, and an
improved algorithm to generate the coefficients of Thompson’s group F, generating the
cogrowth sequences to O(x128 ) and O(x32 ) for B and F respectively.
The amenable group introduced independently by Navas [19] and Brin [4], which we call
the Navas-Brin group B, is a subgroup of Thompson’s group F , and is defined as an infinite
wreath product, with an extra generator which commutes each generator of the infinite
wreath product to the next one. It has 2 generators, so the growth rate of the cogrowth
sequence is 16. It also has a sub-exponential growth term that is very close to exponential,
and so makes the growth rate difficult to estimate accurately with the number of terms at
our disposal.
Using results of Pittet and Sallof-Coste [20, 21], we prove that the cogrowth coefficients
cn of Thompson’s group F satisfy
cn < 16n · λ−n
κ
for any real numbers κ < 1, and λ > 1. That is to say, if Thompson’s group F is amenable,
then its asymptotics cannot contain a stretched-exponential term1. Such a term is present
in the asymptotics of the lamplighter group L and the family of groups Wd . Furthermore,
our numerical study reveals compelling evidence for the presence of such a term in the
asymptotics of the coefficients of F. This is our first strong evidence that Thompson’s
group F is not amenable. Our second piece of evidence is the estimation of the growth
constant. For amenability, the growth constant must be 16. We find that it is very close
to 15.0 (we do not suggest it is exactly 15, but that is certainly a possibility).
Our numerical analysis relies on a number of methods that are well-known in the statistical mechanics and enumerative combinatorics community. Many are reviewed in [13] and
[16]. For studies of the cogrowth asymptotics we primarily rely on the behaviour of the
ratio of successive coefficients, as irrespective of the sub-dominant asymptotics, this ratio
must go to the growth constant in the limit as the order of the coefficients goes to infinity.
One new technique that we make use of in our study of the groups B and F is that of
series extension [15]. In the case of group B, we have 128 exact coefficients, but predict a
further 590 ratios (and terms) with an estimated accuracy of, at worst, 1 part in 5 × 10−7 .
Having these extra (approximate) terms greatly improves the quality of the analysis we
can perform. Similarly, for group F, we use 32 exact terms to predict a further 200 ratios
1We define stretched exponential more broadly than usual. It normally refers to a term of the form e−tβ ,
β
δ
with t > 0 and 0 < β < 1. We allow behaviour such as e−t ·log t , or indeed any appropriate logarithmic
δ
term. We do not have a name for sub-exponential growth of the form e−t/ log t , with δ > 0 (or appropriate
logarithmic function) which is the type of term that must be present in the cogrowth series of the Navas-Brin
group, and indeed in Thompson’s group F if it were amenable.
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
3
(and terms) with an estimated accuracy of 1 part in 4 × 10−5 . This level of accuracy is
more than sufficient for the graphical techniques we use to extract the asymptotics.
Another approach to estimating the growth rate was introduced by Haagerup, Haagerup
and Ramirez-Solano in [18] who proved that the cogrowth sequence of Thompson’s group
F is given by the moments of a probability measure. We extend this to prove that this
observation applies to the cogrowth sequence of any Cayley graph. In this way a sequence
of rigorous lower bounds to the growth constant of the cogrowth series can be constructed.
This approach also gives some stronger, non-rigorous, pseudo-bounds. Further details of
this method, and some results, are given in section 4.
The simplest examples of groups we have chosen have asymptotics of the form
cn ∼ c · µn · ng ,
where c is a constant, µ is the growth constant and g is an exponent.
The first example of such a group is Z2 , which is a particularly simple case as both
2
the coefficients and generating function
are exactly known. In fact cn = 2n
n , and the
√
generating function CZ2 = 2K 4 π x , where K is the complete elliptic integral of the first
kind.
The second example is the Heisenberg group, for which the asymptotic form of the
coefficents is known [10] to be cn ∼ 0.5 · 16n · n−2 , corresponding to a generating function
1
CHeisenberg ∼ (1 − 16x) log(1 − 16x).
2
We have calculated 90 terms of the generating function, and show that this is sufficient to
get a very precise asymptotic representation of the coefficients.
The next level of asymptotic complexity arises when there is an additional stretchedexponential term, so that the coefficients of the generating function behave as
σ
cn ∼ c · µn · κn · ng ,
where 0 < κ < 1, and 0 < σ < 1. There is no known simple expression for the corresponding
generating function in such cases2. The lamplighter group L is the wreath product of the
group of order two with the integers, L = Z2 o Z. The growth rate is known, µ = 9, and
from Theorem 3.5 of [21] it follows that σ = 1/3, and from [22] we know that the exponent
1/3
g = 1/6. So for the lamplighter group, cn ∼ c · 9n · κn · n1/6 . Methods to extract the
asymptotics from the coefficients have been developed, and are described in [14]. We give
a polynomial time algorithm to generate the coefficients, and use it to determine the first
201 coefficients, from which we are able to estimate the correct values of the parameters
µ, σ and g.
We next consider wreath products Wd = Z od Z. In that case the exponent of the
stretched-exponential term also includes a fractional power of a logarithm. Coefficients
of the generating function behave as given by Theorem 3.11 in [21], so that
σ
cn ∼ c · µn · κn
logδ n
· ng ,
2See, for example [14] for a discussion of this point, and further examples of such generating functions.
4
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
where 0 < κ < 1, and 0 < σ, δ < 1.
For d = 1, one has µ = 16, σ = 1/3, δ = 2/3 and g is not known. For d = 2, one has,
again by Theorem 3.11 in [21], µ = 36, σ = 1/2 and δ = 1/2. For general d, µ = (2d)2 ,
σ = d/(d + 2), and δ = 2/(d + 2).
Note that this dimensional dependence of the exponent σ of the stretched-exponential
term appears to be a common feature among a broad class of problems. For example, if one
considers the problem of a self-avoiding walk attached to a surface at its origin (or a Dyck
path or a Motzkin path) and pushed toward the surface at its end-point (or its highest
vertex), then, as shown in [1] there is a stretched-exponential term in the asymptotics of
the coefficients, with exponent σ = 1/(1 + df ), where df is the fractal dimension of the
walk/path. Whether this dimensional dependence is in fact a ubiquitous feature of such
stretched-exponential terms remains an open question.
We have studied two examples, W1 = Z o Z and W2 = (Z o Z) o Z, based on the series we
have generated of 276 and 133 terms respectively. We find that the presence of the confluent
logarithmic term in the exponent makes the analysis significantly more difficult, but we
can nevertheless accurately estimate the growth constant µ and less precisely estimate the
sub-dominant growth rate κ and the exponents σ and δ. Our estimates of the exponent g
are not precise enough to be useful.
We then turn to a contrived example, a constructed series with the asymptotics of Wd =
Z od Z, with d = 98. As d increases, the exponent in the stretched-exponential term gets
closer to 1, and so this term behaves more and more like the dominant exponential growth
term µn . We show that estimating the correct growth constant even approximately requires
careful analysis, and appropriate techniques. This serves as a caution, and underlies that
our conclusions regarding the non-amenability of Thompson’s group F assumes the absence
of some unknown functional pathology.
Finally we study two groups whose behaviour is not fully known. The first is the NavasBrin group B. We give a polynomial-time algorithm to generate the coefficients, and in
this way generate the first 128 terms, then use these to estimate the next 590 ratios. This
group has a sub-exponential growth term that is very close to exponential, and so makes the
growth rate difficult to estimate accurately with the number of terms at our disposal. The
second is Thompson’s group F where we have 32 exactly known terms, and 200 estimated
ratios of terms.
The makeup of the paper is as follows. In Section 2 we describe the algorithms developed
for the cogrowth series of the lamplighter group L, W1 , W2 , B and Thompson’s group F.
In Section 3 we discuss the possible asymptotic form of the cogrowth series for Thompson’s
group F, and prove the absence of a stretched-exponential term. In Section 4 we develop
the idea that the cogrowth coefficients can be represented as the sequence of moments of
a probability measure. With this identification we establish rigorous lower-bounds on the
growth constant for Thompson’s group F. In Section 5 we analyse the series expansions
for the cogrowth series of all the groups we have mentioned above, apart from B and F.
Section 6 is devoted to a description of the method of series extension that we employ,
and in Sections 7 and 8 we use this method and the techniques discussed in the previous
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
5
section to analyse the Navas-Brin group B and Thompson’s group F. Section 9 comprises
a discussion and conclusion.
2. Series generation
In this section we describe the algorithms we have used to compute the terms of the
cogrowth sequence of various groups. We start by describing polynomial time algorithms
which we have found and used for the groups L, W1 , W2 , and B. Finally we describe the
algorithm which we have used for Thompson’s group F . The first 50 coefficients for the
group B are given in Table 1, while the coefficients of the cogrowth series of F are given
in Table 2.
2.1. Wreath Products G o Z. Let G be a group with finite generating set S. We will
describe a polynomial time algorithm for computing the cogrowth series of G o Z, with
respect to the generating set {a} ∪ S, where a generates Z, given the corresponding series
for G. In particular, this give a polynomial time algorithm to compute the cogrowth of the
lamplighter group Z2 o Z as well as groups such as Z o Z and (Z o Z) o Z.
Let ak be the number of loops of length k in G. For example, if G = Z, then a2k = 2k
k
and a2k+1 = 0 for all k ∈ Z≥0 . Then for each positive integer n, define the generating
function Pn (x) by
∞
X
j+n−1
Pn (x) =
aj xj .
n−1
j=0
This is the generating function for n-tuples of words w1 , w2 , . . . , wn ∈ S ∗ such that w1 . . . wn =
1, counted by the length of the word w1 . . . wn .
Given a loop l in G o Z, we define the base loop l0 of l to be the loop in Z made up of
only the terms a and a−1 in l. For each positive integer i, let ci be the number of steps in
the baseloop l0 from ai−1 to ai (which is the same as the number of steps from ai to ai−1 )
and let di be the number of steps from a−i+1 to a−i . Let m and n be maximal such that
cm , dn > 0. Then the length of l0 is equal to
m
n
X
X
2ci +
2dj .
i=1
j=1
l0
Let = a1 a2 . . . a|l0 | and l = w1 a1 w2 . . . a|l0 | w|l0 |+1 , where each wi is a word in (S ∪ S −1 )∗ .
We say that the height of one of the subwords wi is equal to the integer p which satisfies
ap = a1 . . . ai . Then l is a loop if and only if for any height h, concatening all of the words
wi at height h creates a loop in G. Hence the generating function for the sections at height
h is Pr (x) where r is the number of these sections. If h > 0 then r = ch + ch+1 , if h < 0
then r = d−h + d−h+1 and if h = 0 then r = c1 + d1 + 1. Hence, by considering the sections
of l at each height separately, we see that the generating function for loops l with base loop
l0 is equal to
(1)
0
x|l | Pdn (x)Pcm (x)Pc1 +d1 +1 (x)
m−1
Y
i=1
Pci +ci+1 (x)
n−1
Y
j=1
Pdj +dj+1 (x),
6
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
assuming that m, n ≥ 1. Similarly, if m = 0 and n ≥ 1, the generating function is
|l0 |
x Pdn (x)Pd1 +1 (x)
n−1
Y
Pdj +dj+1 (x).
j=1
If n = 0 and m ≥ 1, the generating function is
0
x|l | Pcm (x)Pc1 +1 (x)
m−1
Y
Pci +ci+1 (x).
i=1
Finally, if m = 0 and n = 0, then the generating function is P1 (x). So we now need to sum
this over all possible base loops l0 .
For a given pair of sequences c1 , . . . , cm , d1 , . . . , dn , the number of such base loops is
equal to
(2)
m−1
n−1
c1 + d1 Y ci + ci+1 − 1 Y dj + dj+1 − 1
.
c1
ci − 1
dj − 1
i=1
j=1
This is because from each vertex i > 0 we can choose the order of the outgoing steps, except
that the last one must be a left step, and there are ci − 1 other left steps and ci+1 right
i+1 −1
steps. Hence there are ci +c
possible orders of the steps leaving any vertex i > 0, and
ci −1
dj +dj+1 −1
similarly
possible orders of the steps leaving any vertex −j for j > 0. Finally,
dj −1
c1 +d1
there are c1
possible orders of the steps leaving the vertex 0. It is easy to see that for
any possible choice of these orders there is exactly one corresponding base loop l0 .
Now using (1) and (7) it follows that for any pair of sequences c1 , . . . , cm , d1 , . . . , dn ,
with m, n ≥ 1, the generating function for the corresponding loops l in G o Z is equal to
(3)
x2c1 +2d1
m−1
n−1
c + d
c + c
d + d
Y
Y
1
1
i
i+1 − 1
j
j+1 − 1
Pdn Pcm Pc1 +d1 +1
x2ci+1
x2dj+1
Pci +ci+1
Pdj +dj+1 .
c1
ci − 1
dj − 1
i=1
j=1
If m = 0 and n ≥ 1, the generating function is
(4)
x
2d1
Pdn Pd1 +1
n−1
Y
x
2dj+1
j=1
dj + dj+1 − 1
Pdj +dj+1 .
dj − 1
If m ≥ 1 and n = 0 we get a similar generating function, and if m = n = 0 we get P1 (x).
To calculate these we define some new power series Ωd (x) by
Ωd (x) =
X
Pdn
n−1
Y
j=1
x
2dj+1
dj + dj+1 − 1
Pdj +dj+1 (x),
dj − 1
where the sum is over all sequences n, d1 , d2 , . . . , dn with d1 = d. Then it follows immediately from (3) and (4) that the generating function F for the cogrowth series series of G o Z
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
is given by
(5)
F (x) =
∞
X
c,d=1
7
!
∞
X
2c+2d c + d
2d
Pc+d+1 (x)Ωd (x)Ωc (x) + 2
x
x Pd (x)Ωd (x) + P1 (x).
c
d=1
So now we just need to calculate Ωd (x) for each positive integer d. First, the contribution
to Ωd from the case where n = 1 is Pdn = Pd1 = Pd . The contribution from the case where
n = 1 and d2 = b for some fixed positive integer b is
d+b−1
x2b
Pd+b (x)Ωb (x).
d−1
Hence, we have the equation
(6)
Ωd (x) = Pd (x) +
∞
X
b=1
b+d−1
x
Pb+d (x)Ωb (x).
d−1
2b
Using this equation we can calculate the coefficient of xk in Ωd of x in terms of coefficients
of xj in Ωb (x) where we only need to consider j, b satisfying 2b + j ≤ k (hence j ≤ k − 2).
This takes polynomial time using a simple dynamic program.
2.2. The Navas-Brin group B. In this section we adapt the previous algorithm to calculate the cogrowth series for the Navas-Brin group B. Again this is a polynomial time
algorithm, however the polynomial has higher degree than the one for the previous section.
The group B is defined as the semi-direct product
(. . . o Z o Z o Z o Z o . . .) o Z,
where the copies of Z in the wreath product are generated by . . . , a2 , a1 , a0 , a−1 , a−2 , . . .
and the generator t of the other copy of Z satisfies tai t−1 = ai+1 for each i. Note that
the group B is generated by the two elements t and a = a0 . The group B was described
independently in [19] and on page 638 in [4], where Brin showed that is an amenable
supgroup of Thompson’s group F . In that paper it is the group generated by f and h.
We define the t-height of a word over the generating set {a, t, a−1 , t−1 } to be the sum
of the powers of t. Before counting the total number of loops, we will count the number
of loops where any initial subword has non-negative height. Let G(x, y) be the generating
function for these, where x counts the total length and y counts the number of steps of
the loop which end at height 0. For each positive integer n, let Hn (x, y) be the generating
function for n-tuples w1 , w2 , . . . , wn of words in {a, a−1 , t, t−1 }∗ which each end at height 0
and which have no a or a−1 steps at height 0, such that w1 . . . wn = 1. In this generating
function, x counts the total length of w1 . . . wn and y counts the total number of steps
which end at height 0. Given such a loop l, let the baseloop l0 be the subword consisting
of all a and a−1 steps at t-height 0. Similarly to the previous algorithm, we let ci be the
number of steps in l0 from ai−1 to ai , and di be the number of steps in l0 from a−i+1 to
8
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
a−i . Then the length |l0 | of l0 is equal to
m
X
2ci +
i=1
n
X
2dj .
j=1
As in the previous subsection, for a given pair of sequences c1 , . . . , cm , d1 , . . . , dn , the number of such base loops is equal to
m−1
n−1
c1 + d1 Y ci + ci+1 − 1 Y dj + dj+1 − 1
(7)
.
c1
ci − 1
dj − 1
i=1
j=1
Let l0 = a1 a2 . . . a|l0 | , where each ai ∈ {a, a−1 }, and let l = w1 a1 w2 . . . a|l0 | w|l0 |+1 be the
decomposition where each step ai is at t-height 0. We say that the a-height of one of the
subwords wi is equal to the integer p which satisfies ap = a1 . . . ai . Then l is a loop if and
only if for any height h, concatenating all of the words wi at a-height h creates a loop.
Note that each word wi must have height 0 and have no a or a−1 steps at height 0. As in
the previous section we define another generating function Λd (x, y) by
n−1
Y
X
2dj+1 2dj+1 dj + dj+1 − 1
Λd (x, y) =
Hdn
x
y
Hdj +dj+1 (x),
dj − 1
j=1
where the sum is over all sequences n, d1 , d2 , . . . , dn with d1 = d. In the same way as in
the previous section we get the following equations, which are essentially the same as (5)
and (6).
G(x, y) =
∞
X
c,d=1
∞
X
+2
2c+2d 2c+2d
x
y
c+d
Hc+d+1 (x, y)Λd (x, y)Λc (x, y)
c
x2d y 2d Hd (x, y)Λd (x, y)
d=1
(8)
(9)
+H1 (x, y).
Λd (x) = Hd (x, y) +
∞
X
b=1
2b 2b
x y
b+d−1
Hb+d (x, y)Λb (x).
d−1
So now to calculate G(x, y), we just need to calculate the generating functions Hn (x, y).
For each k ∈ Z≥0 , let Jk (x) be the generating function for loops in B which have exactly
k steps which end at t-height 0, none of which are a or a−1 steps, and which never go
below height 0. For each such word w, the number of ways of breaking it into n words
w1 , w2 , . . . , wn where each ends at height 0, such that w1 . . . wn = w is equal to
k+n−1
.
n−1
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
9
Therefore, we can calculate each generating function Hn (x, y) in terms of the generating
functions Jk (x) as follows:
∞
X
k k+n−1
(10)
Hn (x, y) =
Jk (x).
y
n−1
k=0
Finally, we will calculate the generating functions Jk (x). Trivially we have J0 (x) = 1.
For k > 0, let l be a loop counted by Jk (x). Then l must contain exactly k steps which end
at height 0, which are not a or a−1 steps. Hence they must all be t−1 steps. Therefore, l
decomposes as
l = tu1 t−1 tu2 t−1 . . . tuk t−1 ,
where each word uk ends at height 0 and never goes below height 0. Moreover, since l
is a loop, we must have u1 . . . uk = 1. Hence the word u = u1 . . . uk is counted by the
generating function G(x, y). Moreover, if u contains m steps which end at height 0, then
there are exactly
m+k−1
k−1
ways to decompose u into subwords u1 , . . . , uk which each end at height 0. Hence we get
the equation
∞
X
2k m + k − 1
(11)
Jk (x) =
x
[y m ]G(x, y).
k−1
m=0
Now using equations (8), (9), (10) and (11) as well as the base case J0 (x) = 1, we can
calculate the coefficients of G(x, y) in polynomial time using a dynamic program. Finally
we need to relate these coefficients to the total number of loops in B. We claim that for
each n, the number of loops bn of length n in B over the generating set {a, t, a−1 , t−1 } is
equal to
∞
X
n m n
[y ][x ]G(x, y).
bn =
m
m=0
The reason for this is that the contribution to both sides of the equation from any set
of n loops which are cyclic permutations of each other is the same. That is, if we take
n loops xi . . . xn x1 . . . xi−1 for 1 ≤ i ≤ n, and m of these are counted by G(x, y), then
they will each contribute xn y m to G(x, y), so altogether these will contribute n to both
sides of the equation. If two or more of these loops are identical, then we must have
x1 . . . xn = (x1 . . . xp )q for some p, q satisfying pq = n. In this case, assuming that q is
maximal, the contribution to each side is n/q instead of n, since we overcounted by a
factor of q.
Using the last equation we can quickly calculate the coefficients of the cogrowth generating function CB (x) using those of G(x, y). In Table 1 we give the first 50 coefficients of
this generating function. In fact we have 128 terms.
10
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
1
4
28
232
2092
19864
195352
1970896
20275692
211825600
2240855128
23952786400
258287602744
2806152315048
30686462795856
337490492639512
3730522624066540
41422293291178872
461802091590831904
5167329622166765872
58012358366319158872
653272479274904359312
7376993667962247094112
83518163933592420945440
947797532286760923097848
10779770914124700529470264
122856228305621394118000520
1402877847412263986004347872
16048147989560391552043686160
183892883412730524613883088808
2110556326150834244975990231512
24259510831181186885644198829344
279244563297679787781517160899820
3218641495385722409923501191862264
37146337262307758446419466115479416
429227600058421313330040967935014416
4965493663308539362541734301378311648
57506535582014868288482236767840209688
666700108804771886996957763509359246064
7737176908622194648339548498436658811432
89878279784970230837678375953110478795352
1045033044367535197025078407316665177933928
12161645115366917947524997117208173413019632
141653302005285175865456465524239660635389712
1651274058730064356309776255817393993665780288
19264448513399180870635082273788105896265150480
224919270246185854430934219198103161122414157760
2627954546552385827255336138747466100454012242528
30726935577139566309665785537931570627782996384120
359517978960007312327796870699755173605904761839752
Table 1. The first 50 coefficients of the cogrowth series for the Navas-Brin
group B.
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
11
2.3. A General Algorithm. Before we describe the algorithm which we use for Thompson’s group F , we will describe a general algorithm which can be applied to any group
admitting certain functions which can be computed very quickly. In the next subsection
we will describe how we apply this algorithm to F . This algorithm could also be applied to
any of the other groups which we have discussed, however it would be much less efficient
than the specific algorithms described previously in this section.
Our algorithm can be seen as a significantly more memory efficient version of the algorithm in [6]. First we describe that algorithm. Given a loop γ = a0 a1 . . . a2n , where each
ai ∈ V (Γ) and a2n = a0 = e, we define the midpoint of γ to be the vertex an . Then γ is
made up of a walk of length n from e to its midpoint followed by a walk of length n from
its midpoint to 1. Hence, the number of loops in Γ of length 2n with midpoint m is the
square of the number of walks of length n from 1 to m.
Using a simple dynamic program, the algorithm calculates the number of walks to each
vertex in B(e, n), the ball of radius n in Γ. Then one sums the squares of these numbers
to calculate the number of loops of length 2n. Note also that for each walk from e to m,
there is a corresponding walk from e to m−1 , so it is only necessary to calculate the number
of walks to either m or m−1 . The problem with this algorithm is that it is necessary to
store a large proportion of the ball of radius n in memory at the same time. As a result
it is essentially impossible to get any more than 24 coefficients of the cogrowth series for
Thompson’s group F using this algorithm. Our algorithm is very similar except that we
only store the ball of radius k in memory, where k ≈ n/2. Importantly, we do this without
significantly increasing the running time of the program.
Let G be a group with inverse closed generating set S. Let Γ(G, S) denote the Cayley
graph of G with respect to the generating set S. We will often refer to this as simply Γ.
We will assume that every loop has even length, however this algorithm could easily be
altered to apply when this is not the case.
Let O be an object in the program which represents an element of G. We require the
following functions to be implemented:
• init(). This returns an object O which represents the identity in G.
• val(O). This returns a value which is uniquely determined by the element of G
which the object O represents. In other words, val(O1 ) = val(O2 ) if and only if O1
and O2 represent the same element of G.
• For each generator λ ∈ S, we have an operation O.doλ . If O initially represents the
element g ∈ G, this changes O to an object which represents gλ.
• For each generator λ ∈ S, we have a function lλ (O), defined by lλ (O) = |gλ| − |g|,
where g is the element of G which O represents. That is, lλ (O) = 1 if applying λ
moves g away from the identity.
The speed of our algorithm depends entirely on the efficiency of these functions. For
Thompson’s group our implementations of these all take constant time. Importantly, we
do not require an inverse of val to be implemented.
Given these functions, the algorithm proceeds as follows:
Step 1: Assign an arbitrary order to the generating set S and set k = d n2 e.
12
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
Step 2: Using a simple dynamic program, construct an associative array An−k , implemented as a hash table, with a key value pair (kg , ag ) for each element g ∈ G within the
ball of radius n − k. The key kg is given by val(O) where O is any object which represents
g and the value ag is equal to the number of walks of length n − k in Γ from e to g. We
will write ag = A[kg ]. For a number x which is not a key in An−k , we set A[x] = 0.
Step 3: Construct a tree Tk which contains one vertex vg for each element g of G within
the ball of radius k, such that each vertex vg , apart from ve , is connected to exactly one
vertex vh satisfying |h| = |g| − 1, and g = hλ for some λ ∈ S. If there are multiple possible
choices of h, we choose the element h which minimises λ, according to the order we assigned
in step 1. The edge (h, g) is then labelled with λ. Each vertex vg is also labelled with the
number p(vg ) of paths of length k in Γ from e to g.
Step 4: We now create a function numpaths(O, d) whose input is an object O and a
positive integer d, which, assuming that d = |g|, outputs the number of paths of length n
in Γ from e to g, where g is the group element represented by O. During the calculation
of numpaths the object O may change, but at the end it must represent the same group
element g. Each path of length n from e to g −1 in Γ can be written in a unique way as a
path of length k from e to some vertex h in Γ followed by a path of length n − k from h to
g −1 . For a given h, the number of these paths is equal to p(vh )A[kh−1 g−1 ] = p(vh )A[kgh ].
Hence, the number which we need to return is
X
p(vh )A[kgh ].
h∈G
Note also that the summand is 0 unless |h| ≤ k and |gh| ≤ n − k, so we only need to sum
over values of h which satisfy these two inequalities. To do this we perform a depth first
search of the tree Tk , skipping any sections where we can be sure that there are no vertices
vh such that h satisfies the two inequalities. We start the search at the root vertex ve of Tk
and initialise r = 0 and total = 0. Whenever we move from a vertex vh to vhλ we change
d to d + lλ (O) and then apply the operation O.doλ . That way whenever we are at a vertex
vh , the object O represents gh and d = |gh|. We also increase x by 1 whenever we move
to a child vertex and decrease x by 1 when we backtrack so that we always have x = |h|.
Then we add p(vh )A[kgh ] = p(vh )A[val(O)] to the sum total if and only if d ≤ n − k, since
x = |h| ≤ k for every vertex vh in Tk . Since d decreases by at most 1 when we move to a
child vertex, and x always increases by 1, the value x + d never decreases when we move
to a child vertex. So if x + d > n when we are at a vertex vh , then we do not traverse the
children of vh . At the end of the search we return to the root vertex so that O is back to
its original value and then return the value total.
Step 5: For the last step we just need to add up the value of numpaths for every
vertex g in the ball of radius n such that |g| has the same parity as n. To accomplish this
we perform a depth first search of the tree Tn , which is defined in the same way as Tk .
However, we do not explicitly construct Tn as doing so would use too much memory. In
order to perform the depth first search, we just need a function isedgeλ (O) for each λ ∈ S
which returns 1 if and only if there is an outward edge from vg to vgλ in Tn , where g is
the group element that O represents. This will be the case if and only if |gλ| = |g| + 1
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
13
and |gλµ| = |gλ| + 1 for each µ ∈ S with µ < λ−1 . We test this using the functions lλ ,
doλ and lµ . During the depth first search, we keep track of the distance d = |g|, where g is
the group element represented by O. Now, to calculate the number numloops of loops of
length 2n, we first set numloops = 0, then run the depth first search, and when we visit
each vertex of Tn , add numpaths(O, d)2 to numloops. At the end of this process numloops
is equal to the number of loops of length 2n, so we return numloops and terminate the
algorithm.
The advantage of this algorithm is that it only stores Tk and An−k in memory, rather
than all of Tn . This also allows us to parallelise step 5.
2.4. Thompson’s group F . In this section we describe how the object O, the operation
doλ and the functions val and lλ are implemented for Thompson’s group F . We use the
standard generating set S = {a, b, a−1 , b−1 }, which yields the presentation
F = ha, b|a2 ba−2 = baba−1 b−1 , a3 ba−3 = ba2 ba−2 b−1 i.
For O we use the forest representation given by Belk and Brown in [2]. We simultaneously
store the forest diagram as a graph P as well as a pair of binary strings a, b. A forest
diagram is defined as a pair of sequences of binary trees, with one tree highlighted in each
sequence. A single binary tree with m leaves corresponds to a unique binary string s of
length 2m − 2 with the property that s has an equal number of 1’s and 0’s and the number
of 1’s in any initial substring is at least equal to the number of 0’s in that substring. This
is defined by doing a depth first search of the tree and writing a 1 whenever we move
down an edge from a vertex to its left subtree and writing a 0 whenever we backtrack
along such an edge. Now to convert a sequence of binary trees to a binary string, we first
convert each individual tree to a binary string, insert the string 01 before each such string,
then concatenate the results. We then change the 01 before the string corresponding to
the highlighted tree to 00. This is how the strings a and b are defined. We also store
the numbers pa and pb in O, which define the positions of the 00 before the highlighted
tree in each of a and b. The strings a and b each have length at most 2n, so they can be
represented as 64 bit numbers as long as n ≤ 32. The operation doλ is defined easily for the
effect on the graph P . The effect on the binary strings a and b is a bit more complicated
and requires some bit shifting. The entire length of an element of Thompson’s group F
can be determined by its forest diagram, as shown in [2], so we could use this to determine
lλ by using the graph P and simply subtracting the calculated length |g| from the length
we calculate for |gλ|. In fact we do it more efficiently than this, as the difference |gλ| − |g|
is determined entirely by the highlighted tree and the surrounding trees. Finally, val(O)
simply returns the pair (a, b).
In Table 2 we give the first 32 coefficients of the cogrowth generating function for Thompson’s group F. This is 7 further terms than given in [18].
14
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
Coefficients
1
4
28
232
2092
19884
196096
1988452
20612364
217561120
2331456068
25311956784
277937245744
3082543843552
34493827011868
389093033592912
4420986174041164
50566377945667804
581894842848487960
6733830314028209908
78331435477025276852
915607264080561034564
10750847942401254987096
126768974481834814357308
1500753741925909645997904
17833339046478612301547884
212663448005862463186139032
2544535423071442709522261116
30542557512715560857221200908
367718694478039302564802454628
4439941127401928226610731571976
53756708216952135677787623701460
Table 2. Terms in the cogrowth sequence of Thompson’s group F .
3. Possible cogrowth of Thompson’s Group
In this section we will show that if a0 , a1 , . . . is the cogrowth sequence for Thompson’s
group F , then for any real numbers a < 1 and λ > 1, the inequality
a
an < 16n λ−n
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
15
holds for all sufficiently large integers n. As a result, if Thompson’s group is amenable,
then the sequence cannot grow at the rate
a
16n λ−n ,
For any fixed a < 1. This result follows quite readily from results in [21] and [20], however
we will need some definitions before we can see how they apply. Let G be a group with
finite generating set S. Then we define the function φS : Z>0 → R>0 by setting φS (n)
to be the probability that a random walk in (G, S) of length 2n finishes at the origin. In
other words, |S|2n φS (n) is the number of loops of length 2n in the Cayley graph Γ(G, S).
Now, for two different (non-increasing) functions φ1 and φ2 , we say that φ1 φ2 , if there
is some C ∈ R>0 such that φ1 (n) ≤ Cφ2 (n/C), where each φi is extended to the reals by
linear interpolation. Finally we say that φ1 ≈ φ2 if both φ1 φ2 and φ2 φ1 . We recall
Theorem 3.1 from [20]:
Theorem 3.1. Let G be a group with finite, symmetric generating set S and let H be a
subgroup of G and let T be a finite symmetric generating set of H. Then
φS φT .
The other result we need concerns wreath products with Z. In [21], Pittet and SaloffCoste show (in a remark just below Theorem 8.11) that for a finite generating set T of
Z od Z, we have
φT (n) ≈ exp −nd/(d+2) (log n)2/(d+2) .
Now, since Z od Z is a subgroup of Thompson’s group F , we must have
φS (n) φT (n) ≈ exp −nd/(d+2) (log n)2/(d+2) ,
where S is the standard generating set of F . Hence, for any positive integer d, there is a
positive real number C such that
φS (n) ≤ C exp −(n/C)d/(d+2) (log(n/C))2/(d+2) .
Now we are ready to prove our theorem.
Theorem 3.2. Let an be the number of loops of length 2n in the standard Cayley graph
for Thompson’s group. Then for any real numbers a < 1 and λ > 1, the inequality
a
an < 16n λ−n
holds for all sufficiently large integers n.
d
> a. Then there is some C ∈ R>0 such
Proof. Let d be a positive integer such that d+2
that
φS (n) ≤ C exp −(n/C)d/(d+2) log(n/C)2/(d+2)
16
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
for all n ∈ Z>0 . For n sufficiently large, we have log(n/C) > 0, so
C exp −(n/C)d/(d+2) log(n/C)2/(d+2) <C exp −(n/C)d/(d+2)
= exp log(C) − C −d/(d+2) nd/(d+2) ,
Hence, for all n sufficiently large we have
φS (n) < exp(−nα ).
Therefore,
an = 16n φS (n) < 16n exp(−nα )
Note that the same result holds if we replace Thompson’s group F with the Navas-Brin
group B, since it also contains every wreath product Z od Z as a subgroup.
4. Moments
In [18], Haagerup, Haagerup and Ramirez-Solano prove that the cogrowth sequence
a0 , a1 , . . . for Thomson’s group F is the sequence of moments of some probability measure
µ on [0, ∞), in other words, the sequence is a Stieltjes moment sequence. In fact, their
proof applies to the cogrowth series of any (locally finite) Cayley graph Γ. In this section,
we generalise the result further, to any locally finite graph. First we give some background
on the Stieltjes and Hamburger moment problems.
4.1. Stieltjes and Hamburger moment sequences. In the following, for the sequence
(n)
a = a0 , a1 , . . ., and n ≥ 0, we define the matrix H∞ (a) by
an an+1 an+2 . . .
an+1 an+2 an+3 . . .
(n)
H∞
(a) = an+2 an+3 an+4 . . .
..
..
..
..
.
.
.
.
Theorem 4.1. (Stieltjes [23], GantmakherKrein [9] ) For a sequence a = a0 , a1 , . . ., the
following are equivalent:
• There exists a positive measure µ on [0, ∞) such that
Z
an = xn dµ(x).
(0)
(1)
• The matrices H∞ (a) and H∞ (a) are both positive semidefinite.
• There exists a sequence of real numbers α0 , α1 , . . . ≥ 0 such that the gnerating
function A(t) for the sequence a0 , a1 , . . . satisfies
∞
X
α0
A(t) =
an t n =
α1 t
1−
n=0
α2 t
1−
1 − ...
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
17
A sequence which satisfies the conditions of the theorem above is called a Stieltjes
moment sequence.
Theorem 4.2. For a sequence a = a0 , a1 , . . ., the following are equivalent:
• There exists a positive measure µ on (−∞, ∞) such that
Z
an = xn dµ(x).
(0)
• The matrix H∞ (a) is positive semidefinite.
A sequence which satisfies the conditions of the theorem above is called a Hamburger
moment sequence. From either definition of Hamburger moment sequence, it follows immediately that any Stieltjes moment sequence is a Hamburger moment sequence. Carleman’s
condition states that the measure µ is unique if
∞
X
−
1
a2n2n = +∞,
n=0
This is certainly true when the sequence grows at most exponentially, as is the case for all
of our examples. For Stieltjes moment sequences, the following weaker condition implies
that the measure µ is unique:
∞
X
− 1
an 2n = +∞.
n=0
For a Hamburger moment sequence a, which grows at most exponentially, the radius of
convergence of A(t) = a0 + a1 t + a2 t2 + . . . is equal to
1
.
|µ|
In particular, this means that if a is a Stieltjes moment sequence, the exponential growth
rate of the sequence is equal to the minimum value in the support of µ.
One benefit of proving that a sequence a is a Stieltjes moment sequence is that it allows
us to compute good lower bounds for the exponential growth rate of the sequence using
only finitely many terms. This method was described in [18], but we repeat the description
here, using the continued fraction form of a. We consider the generating function
∞
X
α0
.
α1 t
1−
n=0
α2 t
1−
1 − ...
Using the terms a0 , . . . , an , we calculate the terms α0 , . . . , αn . It is easy to see that A(t)
is nondecreasing in each αj . Hence, the minimum possible value An (t) is achieved by
setting αn+1 , αn+2 , . . . to 0. Therefore, the radius of convergence tc of A(t) is bounded
above by the radius of convergence tn of An (t). Therefore, bn = 1/tn is a lower bound
for the exponential growth rate of a. It is easy to check that the sequence b1 , b2 , . . . is
A(t) =
an tn =
18
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
nondecreasing, and bnn > an /a0 . It follows that this sequence of lower bounds converges to
the exponential growth rate of a.
If we assume further that the sequences α0 , α2 , α4 , . . . and α1 , α3 , . . . are non-decreasing,
as seems to be true for many of the cases we consider, we can get stronger lower bounds
for the growth rate by setting αn+1 , αn+3 , . . . to αn−1 and αn+2 , αn+4 , . . . to αn . For this
√
√
sequence the exponential growth rate of corresponding sequence a is ( αn + αn−1 )2 .
4.2. Applications of moments to the cogrowth series. Here we describe how to
compute lower bounds for the growth rate of the cogrowth sequence of Thompson’s group
F.
Theorem 4.3. Let Γ be a locally finite graph with a fixed base vertex v0 . For each n ∈ Z≥0 ,
let tn be the number of loops of length n in Γ which start and end at v0 . Then there exists
a probability measure µ on R such that for each n ∈ Z≥0 , the nth moment of µ is given by
Z ∞
xn dµ = tn .
−∞
In other words, t0 , t1 , . . . is a Hamburger moment sequence.
Proof. The sequence t = t0 , t1 , . . . is a Hamburger
matrix
t0 t1 t2
t1 t2 t3
(0)
H∞
(t) = t2 t3 t4
.. .. ..
. . .
moment sequence if and only if the
...
. . .
. . .
..
.
is positive semidefinite. To prove this, we will show that this is the matrix representation
of a positive definite bilinear form.
Let M be the inner product space over R defined by the orthonormal basis {bv |v ∈ V (Γ)}.
For each n ∈ Z we let xn ∈ M be the element defined by
X
xn =
pv bv ,
v∈V (Γ)
where pv is the number of paths of length n in Γ from v0 to v. Then it is easy to see that
for any non-negative integers m and n, the value hxn , xm i is equal to the number tm+n of
paths of length m + n in Γ from v0 to itself. Now let X be the subspace of M spanned by
{x0 , x1 , . . .}. Then A is the matrix representation of the inner product h, i, restricted to X,
(0)
with respect to the spanning set {x0 , x1 , . . .}. Therefore, H∞ (t) is positive semidefinite.
(0)
Note that if {x0 , x1 , . . .} are linearly independent, then H∞ (t) is positive definite.
Theorem 4.4. Let C ∈ Z>0 and let Γ be a graph with a fixed base vertex v0 , such that
each vertex in Γ has degree at most C. For each n ∈ Z≥0 , let tn be the number of loops of
length n in Γ which start and end at v0 . There exists a probability measure µ on R>0 such
that for each n ∈ Z≥0 , the nth moment of µ is equal to t2n . In other words, t0 , t2 , t4 , . . . is
a Stieltjes moment sequence.
Moreover, µ is unique and its support is contained in the interval [0, C 2 ].
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
Proof. In order to show that the sequence s = t0 , t2 , t4 , . . . is a
it suffices to prove that the two matrices
t0 t2 . . .
t2
t2 t4 . . .
t4
(0)
(1)
H∞ (s) =
and H∞ (s) =
.. .. . .
..
.
. .
.
are positive semidefinite. From the previous theorem,
t0 t1 t2
t1 t2 t3
(0)
H∞
(t) = t2 t3 t4
.. .. ..
. . .
19
Stieltjes moment sequence,
t4 . . .
t6 . . .
.. . .
.
.
we know that the matrix
...
. . .
. . .
..
.
(0)
is positive semidefinite. Hence any principal submatrix of H∞ (t) (using the same rows
(0)
(1)
and columns) is also positive semidefinite. Since both the matrices H∞ (s) and H∞ (s)
(0)
are such principal submatrices of H∞ (t), each of these matrices is positive semidefinite.
Therefore, the sequence t0 , t2 , t4 , . . . is a Stieltjes moment sequence. Now, since each vertex
of the graph has degree at most C, the number of paths of length n is at most C n . Hence
we have the inequality
Z
∞
xn dµ = t2n ≤ C 2n .
0
Therefore, the support of µ must be contained in the interval [0, C 2 ]. This also implies
that µ is unique.
In particular, if we let G be a finitely generated group, with inverse closed generating
set S, and Γ be the corresponding Cayley graph, then the even terms of the cogrowth
sequence for Γ form a Stieltjes moment sequence. Moreover, each vertex has degree |S| so
the support of the corresponding measure µ is contained in the interval [0, |S|]. As described
in the previous subsection, we can compute lower bounds bn for the exponential growth rate
of any such sequence. Turning our attention to Thompson’s group, using 31 terms of the
cogrowth sequence, we have computed the corresponding terms α0 , α1 , . . . , α31 . Using these
we have computed the rigorous lower bound b31 ≈ 13.269 for the exponential growth rate
of the cogrowth sequence of Thompson’s group. If we assume that the sequences α0 , α2 , . . .
√
√
and α1 , α3 , . . . are increasing, we get the stronger lower bound ( α30 + α31 )2 ≈ 13.706.
In Section 8 below we extrapolate the sequence of bounds {bn } to estimate the growth
constant µ, and find µ ≈ 15.0.
5. Series Analysis
We have series for six groups, which we will consider in order. Firstly, the group Z2 ,
then the Heisenberg group, the lamplighter group L = Z2 o Z, the two groups Z o Z and
(Z o Z) o Z, the Navas-Brin group B [19, 4] and finally Thompson’s group F . We will analyse
each of these in turn.
20
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
In all cases our initial analysis is based on the behaviour of the ratio of successive terms,
with other methods deployed as appropriate. In the simplest situation we consider, which
is when the asymptotic form of the coefficients is cn ∼ c · µn · ng , one has that the ratio of
successive coefficients is asymptotically linear when plotted against 1/n, as
g
1
cn
=µ 1+ +o
.
(12)
rn =
cn−1
n
n
It is therefore natural to plot the ratios rn against 1/n. If the correction term o n1 can be
ignored3, such a plot will be linear, with gradient µ · g, and intercept µ at 1/n = 0. If the
growth constant µ is known, or can be guessed, better estimates of the exponent g can be
made by extrapolating the sequence
gn = (rn /µ − 1) · n = g + o(1).
More complicated asymptotic forms for the coefficients can give rise to different expressions for the ratios, as we show below.
5.1. The group Z2 . For the group Z2 , the coefficients of the cogrowth series are known
2
exactly, cn = 2n
n , and so the ratio of successive terms is
1
cn
1
rn =
= 16 1 − + 2 .
cn−1
n 4n
A plot of the ratios against 1/n is shown in Figure 1, based on the first 50 coefficients.
It is clearly going to the expected limit of 16. The exponent g should be −1, and we plot
estimators gn against 1/n in Figure 2, which is also clearly going to the expected limit −1.
This corresponds to a logarithmic singularity of the generating function,
CZ2 (x) ∼ c · log(1 − 16x).
For this simple example one can do much better by using the package gfun, available
in Maple, and asking for the underlying ordinary differential equation for the generating
function, given the first 20 or so coefficients. In this way one immediately obtains the result
for the generating function
√
X
4 x
n
cn x = 2K
,
CZ2 (x) =
π
where K is the complete elliptic integral of the first kind.
5.2. The Heisenberg group. We have calculated 90 terms of the generating function,
and show that this is sufficient to obtain a very precise asymptotic representation of the
coefficients. The leading order asymptotics of the coefficients is known [10] to be cn ∼
16n /(2n2 ), corresponding to a generating function
1
CHeisenberg ∼ (1 − 16x) log(1 − 16x).
2
We have analysed this series in the same way as described above for the group Z2 .
3In the simplest cases, such as the present one, the correction term will be O
1
n2
.
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
21
Figure 1. Plot of Z2 ratios against 1/n.
Figure 2. Estimators of exponent g for Z2
vs. 1/n.
Figure 3. Plot of Heisenberg group ratios
against 1/n.
Figure 4. Estimators of exponent g for
the Heisenberg group vs. 1/n.
A plot of the ratios against 1/n is shown in Figure 3. It is clearly going to the expected
limit of 16. The exponent g should be −2, and we plot estimators gn against 1/n in Figure
4, which are also clearly going to the expected limit −2.
22
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
In order to obtain higher-order asymptotic terms, we subtract the known leading-order
term from the coefficients, forming the sequence
c(1) (n) = cn − 16n /(2n2 ).
A ratio analysis of this sequence strongly suggests that c(1) (n) ∼ const/n, implying that
cn ∼ 16n /(2n2 )+const./n3 . Such behaviour is consistent with a simple algebraic singularity
of the generating function. Accordingly, we attempted a linear fit to the assumed form
cn /16n = 1/(2n2 ) + k1 /n3 + k2 /n4 + k3 /n5 . We did this by solving the linear system given
by setting n = m − 1, n = m, n = m + 1 in the preceding equation, and solving for
k1 , k2 , k3 , with m ranging from 20 to the maximum possible value 89. We obtain an
m-dependent sequence of estimates of the amplitudes k1 , k2 , k3 , which we extrapolated
against appropriate powers of 1/m.
In this way we estimate k1 = 0.93341, k2 = 1.530, and k3 = 3.30, where we expect errors
in these estimates to be confined to the last quoted digit.
To summarise, we find the asymptotics of the coefficients of the cogrowth series of the
Heisenberg group to be
1
0.93341 1.530 3.30
1
n
.
+
+
+ 5 +O
cn = 16
2
3
4
2n
n
n
n
n6
5.3. The lamplighter group. The lamplighter group L is the wreath product of the
group of order two with the integers, L = Z2 o Z. From [22] we know that for this group,
(13)
cn ∼ c · 9n · κn
1/3
· n1/6 .
1/3
So in this example we see the presence of a stretched-exponential term, κn , which
makes the analysis more difficult. As remarked above, we have generated 201 terms of
the cogrowth series, and show how these terms can be used to estimate the asymptotic
behaviour of the coefficients.
If the coefficients of a series include a stretched-exponential term, so that
σ
an ∼ c · µn · κn · ng ,
with 0 < σ, κ < 1, then the ratio of successive terms behaves as
σ log κ g
an
rn =
∼ µ 1 + 1−σ + + · · · .
an−1
n
n
Experimentally, the presence of such a stretched-exponential term is signalled by the fact
that the ratio plots against 1/n exhibit curvature, and that this curvature can be eliminated,
or at least substantially reduced, by plotting the ratios against 1/n1−σ , where σ is roughly
estimated by choosing its value so as to maximise linearity. This theory is developed in
greater detail, along with several examples, in [14].
Because of the presence of two terms in the ratio plots, one of order O(nσ−1 ) the other
of order O(1/n), there is some competition between these two terms, which can make it
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
23
difficult to estimate the value of σ just from the linearity of the ratio plots. So we first
eliminate the O(1/n) term by calculating the modified ratios
1
σ 2 log κ
(1)
.
(14)
rn = n · rn − (n − 1) · rn−1 = µ 1 + 1−σ + o
n
n
In Figure 5 we show the modified ratios plotted against 1/n2/3 , which is seen to be linear,
and extrapolating to the known
√ growth constant of 9. While not shown, we also plotted
the modified ratios against 1/ n and against 1/n3/4 . These were visibly convex upward
and concave downward, respectively. One would conclude that 1/2 < σ < 3/4, and bearing
in mind that in all known such behaviour, σ is a simple rational fraction (arguably simply
related to dimensionality), one would conjecture that κ = 2/3. However, we can also
estimate the value of σ by other means.
Figure 5. Modified lamplighter group ratios vs. n−2/3 .
(1)
If we assume µ = 9, then from (14) it follows that a plot of ln = log |1 − rn /µ| against
log(n) should be linear with gradient σ − 1. This plot (not shown) is indeed visually linear.
To calculate the gradient, which will vary slightly with n, we calculate the local gradient
(ln − ln−1 )/(log(n) − log(n − 1)), and show this plotted against 1/n4/3 in Figure 6. This
plot is clearly going to a limit very close to −2/3, as expected.
One can also find estimators for the exponent σ without assuming or knowing the value
of the growth constant µ. Taking the ratio of the modified ratios eliminates the growth
constant µ, so that
(1)
rn(2) =
rn
(1)
rn−1
=1−
σ 2 (1 − σ) log κ
+ o(nσ−2 ).
n2−σ
24
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
Figure 6. Estimates of σ − 1 vs. n−4/3 .
Figure 7. Estimates of σ − 2 vs. n−2/3 .
(2)
So a plot of log |rn − 1| against log n should be linear with gradient σ − 2. As above,
we don’t show this uninteresting linear plot, but instead show the local gradient, plotted
against 1/n2/3 , in Figure 7, which appears to be going to a value around −1.67, consistent
with the known exact value −5/3.
Assuming the values µ = 9, and σ = 1/3, we estimate the remaining parameters in
the asymptotic expression by direct fitting to the logarithm of the coefficients. From
1/3
cn ∼ c · 9n · κn · ng we get
log cn − n · log 9 ∼ n1/3 · log κ + g · log n + log c.
As in the preceding analysis of the Heisenberg group coefficients, we fit successive triples
of coefficients to get estimates of the three unknowns, log κ, g and log c. The results are
shown in Figures 8, 9, and 10 respectively.
From these plots, we estimate log κ ≈ −2.78, g ≈ 0.17, and log c ≈ −0.6. If we use the
fact that we know that the exponent g = 1/6, we can get refined estimates of the remaining
parameters, giving log κ ≈ −2.775, and log c ≈ −0.55, so that κ ≈ 0.0623, and c ≈ 0.58.
As far as we are aware, these two constants have not previously been estimated.
5.4. Analysis of group Z o Z. As discussed in the introduction, for the groups Z od Z,
there is an additional logarithmic factor associated with the stretched-exponential term.
For d = 1 the group Z o Z has coefficients that behave as
an ∼ const · µn · κn
σ
logδ n
· ng , with σ = 1/3 and δ = 2/3.
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
25
Figure 9. Estimates of
Figure 8. Estimates of log κ exponent g vs. 1/n. The Figure 10. Estimates
vs. 1/n.
log c vs. n−2/3 .
exact value is 1/6.
of
It follows that the ratio of successive coefficients behaves as
(15)
an
σ · log κ · logδ n δ · log κ · logδ−1 n g
rn =
∼µ 1+
+
+ + ···
an−1
n1−σ
n1−σ
n
!
.
We have generated series to order x276 for this group. A simple ratio plot against 1/n is
strongly concave downwards. Plotting the ratios against 1/n2/3 gives a plot which is√much
closer to linear, but still displays a slight concavity. A simple ratio plot against 1/ n by
contrast, displays slight convexity.
As we noted in our analysis of the lamplighter group, the term g/n in eqn. (15) also
makes a contribution (as does the logarithmic term logδ n), so a clearer picture emerges if
this term is eliminated, which we do by forming the modified ratios (14), which behave in
this case as
log κ 2/3
−1/3
−4/3
−5/3+
(1)
n − 2 log
n + o(n
) .
(16)
rn = µ 1 + 2/3 log n + 4 log
9n
√
Plots of the modified ratios are shown in Figures 11, 12, and 13, against 1/ n, 1/n2/3
and 1/n3/4 respectively. It is clear that the plot against 1/n2/3 is the closest to linear,
corresponding to κ = 1/3. However, there is still some downward concavity, due to the associated logarithmic terms.
To see this even more clearly, weshow in Figure 14 a plot of the
modified ratios against log2/3 n + 4 log−1/3 n − 2 log−4/3 n /n2/3 , which is the expected
asymptotic behaviour, see (16). This is indistinguishable from linearity.
To date we haven’t tried to estimate µ, known to be exactly 16. One way to do this is
from the modified ratio plots shown above. All are seen to be tracking towards a value
very close to 16.
It is also possible to estimate the exponent σ directly from the ratios, even without
knowing the dominant exponential growth constant µ. One first forms the ratio of successive
26
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
Figure
11. Modified ratios for Z o Z vs.
√
1/ n.
Figure 12. Modified ratios for Z o Z vs.
n−2/3 .
Figure 13. Modified ratios for Z o Z vs.
n−3/4 .
Figure
14. Modified ratios for Zo Z vs.
2/3
log n + 4 log−1/3 n − 2 log−4/3 n n−2/3 .
ratios, so that
(17)
rrn(1)
=
rn
rn−1
log κ · logδ n
= 1+
n2−σ
δ(2σ − 1) δ(δ − 1)
g
σ(σ − 1) +
+
− 2 + o(1/n2 ).
2
log n
n
log n
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
27
As we did above with the ratios, we eliminate the O(1/n2 ) term by constructing a
modified ratio-of-ratios,
(1)
(18)
rrn(2) =
(1)
n2 rrn − (n − 1)2 rrn−1
c log δ n
= 1 + 2−σ (1 + O(1/ log n)) ,
2n − 1
n
where the constant c = (σ 2 (σ − 1) log κ)/2.
(2)
Then a plot of log |rrn − 1| against log n should be close to linear, as the logarithmic
term will vary very slowly over the range of n-values at our disposal, with gradient σ − 2.
Such a plot (not shown) is visually linear, but in order to calculate the gradient we find
(2)
(2)
the (local) gradient of the segment joining rrn and rrn−1 , which should approach the
“correct” value as n increases. This is shown, plotted against 1/n in Figure 15. It appears
to be going to a limit around −1.62 to − 1.61, which would imply σ ≈ 0.38 or 0.39, rather
than the known value of 1/3.
Figure 15. Estimators of exponent σ − 2 vs. 1/n.
However, if we assume we know that δ = 2/3, and include the confluent logarithmic
term log2/3 n in the exponent of the stretched-exponential term, plotting instead
!
(2)
rn − 1
log
log2/3 n
against log n, the plot is again visually linear. However the corresponding plot of the local
gradient, shown in Figure 16, is clearly going to a limit around −5/3, consistent with the
known value σ = 1/3.
28
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
Figure 16. Estimators of exponent σ − 2
vs. 1/n, assuming a confluent logarithmic
term.
Figure 17. Estimates of log κ vs. 1/n.
Assuming the values µ = 16, and σ = 1/3 and κ = 2/3, we can estimate the remaining
parameters in the asymptotic expression by direct fitting to the logarithm of the coefficients.
2/3
1/3
From cn ∼ c · 16n · κn log n · ng we get
log cn − n · log 16 ∼ n1/3 · log2/3 n · log κ + g · log n + log c.
As in the preceding analysis of the lamplighter group coefficients, we fit successive triples
of coefficients to get n−dependant estimates of the three unknowns, log κ, g and log c. The
results are shown in Figures 17, 18, and 19 respectively.
From these plots, we estimate log κ ≈ −1.64, but it is difficult to estimate g. It appears
to be quite small, close to zero, and could even be negative. It is even more difficult to
extrapolate the plot for log c, though one might conclude the bound log c ≥ −2. These
estimates correspond to κ ≈ 0.194, g ≈ 0, and c > 0.13. As far as we are aware, these three
constants have not previously been studied.
In anticipation of our analysis of Thompson’s group F, where the growth constant µ
is not known, we attempt to estimate both the exponents σ and δ without knowing the
value of µ. Forming the ratios (15) eliminates the constant c in the asymptotic form of the
coefficients, and the ratio of ratios (17) eliminates µ. If we now form the sequence
(1)
(19)
tn =
rrn − 1
(1)
rrn−1 − 1
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
Figure 18. Estimates of exponent g vs.
n−1/3 .
29
Figure 19. Estimates of log c vs. n−1/3 .
this eliminates the base κ of the stretched-exponential term, and in fact
n(tn − 1) ∼ σ − 2 +
δ
.
n log n
So plotting n(tn − 1) against 1/(n log n) should give an estimate of σ − 2. To estimate δ,
we form the sequence
n log2 n(n(tn − 1) − (n − 1)(tn−1 − 1)) ∼ −δ + O(1/ log n).
We show these plots in Figures 20 and 21 respectively. The estimate of σ − 2 appears to
be going to a limit of around -1.6 or below, c.f. the known exact value of −5/3, while the
estimate of δ is harder to estimate, but the plot is certainly consistent with the known
value 2/3. As can be seen, this exponent is difficult to estimate without many more terms
than we currently have.
5.5. Analysis of group (Z o Z) o Z. For this group we have 132 terms in the cogrowth
series, just less than half the number we have for Z o Z, so the results are not quite as
precise. We analysed this series the same way as for the group Z o Z. For this group it is
known that the coefficients grow exponentially, and that the dominant term is 36n . The
1/2
1/2
sub-dominant term is κn log n , which again follows from Theorem 3.11 in [21]. Again,
there is presumably a sub-sub dominant term ng .
In this case we have for the ratio of successive terms:
!
an
log κ log1/2 n
log κ
g
(20)
rn =
∼µ 1+
+
+ + ··· .
an−1
2n1/2
2n1/2 log1/2 n n
30
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
Figure 20. Estimates of σ − 2 vs.
1/(n log n)
Figure 21. Estimates of exponent −δ vs.
1/ log n.
We eliminate the O(1/n) term by forming the modified ratios (14) which behave as
(21)
rn(1)
log κ p
−1/2
−3/2
−3/2+
log n + 2 log
n − log
n + o(n
) .
=µ 1+ √
4 n
First, we remark that extrapolating the ratios against 1/n gives a plot with considerable
curvature (not shown). We plotted the modified ratios, defined above, against 1/nσ for
several values of σ. We show the results for σ = 1/2 and σ = 1/3 in Figures 22 and 23
respectively. Surprisingly, the latter is closer to linear, however it extrapolates to a value
of µ rather larger than the actual value, µ = 36. However if we include the effect of the
logarithmic
term in the exponent, and plot (see equation (21)) the modified ratios against
q
log n
n ,
the modified ratio plot, shown in Figure 24, is indistinguishable from linearity and
extrapolates to the correct value of µ.
Repeating the analysis of the previous section, we attempted to estimate the exponent σ
(2)
without assuming the value of the growth constant µ. A plot of log |rrn − 1| (18) against
log n should be close to linear, (as the logarithmic term will vary only slowly over the range
of n-values at our disposal), with gradient σ − 2. Such a plot (not shown) is visually linear,
but in order to calculate the gradient we find the (local) gradient of the segment joining
(2)
(2)
rrn and rrn−1 , which should approach the “correct” value as n increases. This is shown,
plotted against 1/n in Figure 25. It appears to be going to a limit below −1.42 which
would imply σ < 0.58, compared to the known value of 1/2.
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
31
√
Figure 22. Modified ratios for (Z o Z) o Z vs. 1/ n.
Figure 23. Modified ratios for (Z o Z) o Z
vs. n−1/3 .
Figure
p 24. Modified ratios for (Z o Z) o Z
vs. log n/n.
However, if we assume we know that δ = 1/2, and include the confluent logarithmic
term log1/2 n in the exponent of the stretched-exponential term, plotting instead
!
(2)
rn − 1
log
log1/2 n
32
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
Figure 25. Estimators of exponent σ − 2
vs. 1/n.
Figure 26. Estimators of exponent σ − 2
vs. 1/n, assuming a confluent logarithmic
term.
against log n, the plot is again visually linear. Moreover the corresponding plot of the local
gradient, shown in Figure 26, is going to a limit around −3/2, consistent with the known
value σ = 1/2.
Assuming the values µ = 16, σ = 1/2 and κ = 1/2, we can estimate the remaining
parameters in the asymptotic expression by direct fitting to the logarithm of the coefficients.
1/2
1/2
From cn ∼ c · 36n · κn log n · ng we get
log cn − n · log 36 ∼ n1/2 · log1/2 n · log κ + g · log n + log c.
As in the preceding analysis of Z o Z, we fit successive triples of coefficients to get estimates
of the three unknowns, log κ, g and log c. The results for the first two are shown in Figures
27 and 28 respectively. From this, and further analysis with an additonal term in the
assumed asymptotic form, we estimate log κ ≈ −2.3 and g ≈ 3.3.
Again repeating the analysis of the previous section, we tried to estimate σ and δ directly
without knowing µ or κ. Plotting n(tn − 1) (19) against 1/(n log n) should give an estimate
of σ − 2, and plotting the sequence n log2 n(n(tn − 1) − (n − 1)(tn−1 − 1)) against 1/ log n
should give estimates of −δ. We show these plots in Figures 29 and 30 respectively. The
estimate of σ − 2 appears to be going to a limit of below -1.39 or so, c.f. the known exact
value of −1.5, while it is not possible to estimate δ from this plot, but it is not inconsistent
with the known value 1/2.
5.6. The group Z od Z. In the previous sections we have considered the analysis of the
groups Z od Z for d = 1 and d = 2. We have shown how the stretched-exponential term
slows the rate of convergence of the ratios, but that appropriate analysis can still reveal
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
33
√
Figure 27. Estimates of log κ vs. 1/ n.
Figure
28. Estimates of exponent g vs.
√
1/ n.
Figure 29. Estimates of σ −2 for (ZoZ)oZ
vs. 1/(n log n)
Figure 30. Estimates of exponent δ for
(Z o Z) o Z vs. 1/ log n.
much asymptotic information. However as d increases, it becomes increasingly difficult to
extract the asymptotics from a hundred or so terms of the cogrowth series. To see this, we
34
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
consider the case d = 98. Then we know the asymptotic form of the coefficients is
σ
cn ∼ c · µn · κn
logδ n
· ng ,
where σ = 49/50 and δ = 1/50 [21].
While we could have generated 100 or so terms of this series from the algorithms described above, it will be more instructive to generate a test series with the given asymptotic
behaviour, as then we can generate thousands of terms essentially immediately.
δ
σ
So we have generated coefficients defined by cn = c · µn · κn log n · ng with c = 1, µ = 4,
κ = 0.7, g = 0.5, σ = 49/50 and δ = 1/50. The ratio of successive terms must go to 4.0,
the value of the growth constant4. Using 128 terms of this test series, we show a plot of
the ratios against 1/n in Figure 31. It is not possible to assert that, as n → ∞ the ratios
will go to 4.0. In Figure 32 we show the same plot with 1280 terms. While this curve
is steeply increasing, it is still not possible to assert that the limiting value is 4.0. Using
10000 terms, and plotting the ratios against 1/n1/50 (not shown), we finally see evidence
that the extrapolated limit is around 3.8 or 3.9.
Figure 31. The first 128 ratios for Z o98 Z
vs. 1/n.
Figure 32. The first 1280 ratios for Z o98 Z
vs. 1/n.
For this series the asymptotic form of the ratios is
49 log κ
0.5
rn = µ 1 +
+
+ o(1/n) ,
n
50 · n1/50
so we might expect more informative results if we eliminate the term O(1/n), which we
can do by forming the modified ratios. These are shown, plotted against 1/n1/50 in Figures
4The growth constant is actually 4(d + 1)2 , but for this exercise the actual value is irrelevant, so we have
chosen a much smaller value.
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
35
33 and 34, based on the first 128 terms and the first 10000 terms. Extrapolating these to
n → ∞ again gives a limit around 3.9.
Figure 33. The first 128 modified ratios
for Z o98 Z vs. n−1/50 .
Figure 34. The first 10000 modified ratios
for Z o98 Z vs. n−1/50 .
It is possible to estimate the exponent σ without knowing µ, as we showed in previous
examples above. In particular, using the method based on equation (17), and described
immediately below that equation, we show in Figure 35 a plot of estimators of σ −2 against
1/n, based on a 10000 term series, and it is persuasively going to the known value −1.02.
Unfortunately, for no interesting problem is it realistic to get 10000 terms, so this example, and the next, must remain as a cautionary tale, to the extent that there can and
do exist groups whose cogrowth series exhibit asymptotic behaviour that is difficult to estimate by numerical methods of the type we have considered. Another example of similar
difficulty is given by the Navas-Brin group B, discussed in the next section.
6. Series extension
In this section we develop one further tool that will be extremely useful in our analysis
of the series for Thompson’s group F, where we have only 32 terms, rather than a hundred
or more as in the examples we have been considering. It will also be very helpful in our
analysis of the Navas-Brin group B, discussed in the next section.
Recall that our analysis of the more complex asymptotic forms that include stretchedexponential terms is based on ratios of successive terms, whereas for simpler groups, with
simpler asymptotics, we used the method of differential approximants (DAs). It is obviously
highly desirable to have further terms (in particular, further ratios), for all series with nonsimple asymptotics, and particularly in those cases where we have comparatively short
36
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
Figure 35. Estimators of σ − 2 for Z o98 Z against 1/n for n ≤ 10000.
series, such as the 32 term series we have for Thompson’s group F. In order to obtain further
ratios (or terms), we use the method of differential approximants to predict subsequent
ratios/terms. The detailed description as to how this is done is given in [15].
We will give two demonstrations of the effectiveness of this method. In the first, we take
the first 32 terms of the series for Z o Z discussed above, (we have more than 200 terms for
this series), and use these to predict the next 89 ratios, from 5th order DAs. As well as
the mean ratio, we calculate the standard deviation. We show, in Table 3, a comparison
between the actual error in the predicted ratios and the standard deviation of the estimated
ratios. It can be seen that the true error lies between 1 and 1.5 standard deviations, which
provides some confidence that the predicted ratios are accurate to within an error of 1.5
standard deviations.
For the series simulating the coefficients of the group Zo98 Z, we showed the importance of
long series to reveal the asymptotic behaviour with some precision. In this second example,
we take the first 100 terms of this series, and use them to predict the next 315 ratios. That
is, we estimate cn /cn−1 for n = 101 · · · 415.
To see how precisely these ratios can be predicted, we plot the difference between the
actual ratios and those calculated by 4th order differential approximants in Figure 36.
It can be seen that the error is less than 2 parts in 1020 for all n < 416. Just to make
this perfectly clear, given 100 coefficients, we have predicted the next 315 ratios with an
accuracy of some 20 significant digits.
In a similar fashion, using 4th order DAs, we were able to get 200 extra ratios for the
32-term series for Thompson group F. The maximum error (as estimated by 1.5 s.d. of the
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
Figure 36. Absolute error in predicted ratios of Z o98 Z for 100 < n < 416.
k
1
5
10
20
30
40
50
60
70
80
89
37
Figure 37. The first 128 modified ratios
of the Navas-Brin group B vs. 1/n.
Actual error 1 standard deviation
2.69 × 10−17
2.02 × 10−17
−13
1.14 × 10
7.85 × 10−14
−11
3.37 × 10
2.08 × 10−11
2.22 × 10−8
1.23 × 10−8
−7
9.63 × 10
5.39 × 10−7
−5
1.22 × 10
6.88 × 10−6
−5
7.59 × 10
4.73 × 10−5
3.13 × 10−4
2.23 × 10−4
−4
9.39 × 10
8.11 × 10−4
−3
2.44 × 10
2.44 × 10−3
−3
4.63 × 10
5.38 × 10−3
Table 3. Actual error in coefficient O(z 31+k ) and 1 standard deviation
from the mean of the estimated coefficient.
DAs) is 1 part in 4 × 10−5 , which is graphically imperceptible. In the Appendix we give
the (predicted) next 200 ratios, and their standard deviations.
7. Analysis of the Navas-Brin group B.
This is an amenable group introduced independently by Navas [19] and Brin [4], so we
call it the Navas-Brin group B, and is defined in subsection 2.2. It has 2 generators, so
38
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
the growth rate of the cogrowth sequence is 16. We gave a polynomial-time algorithm to
generate the coefficients above, and have used this to generate 128 terms of the co-growth
series. We then used the method of series extension, described above, to give a further 590
ratios, the last of which we expect to be accurate to 1 part in 5 × 10−7 , while all earlier
ratios will have a lower associated error. We first show a plot of the modified ratios (14)
against 1/n in Figure 37. Even if we knew nothing about the asymptotics of this group,
the curvature of this plot provides strong evidence for a sub-exponential term, and we have
proved that it cannot be a regular stretched-exponential term.
That is to say, the asymptotics for this series must grow more slowly than
σ
cn ∼ c · µn · κn · ng ,
where µ = 16, 0 < σ < 1, and 0 < κ < 1. Possible behaviour might be
cn ∼ c · µn · κn/ log n · ng ,
corresponding to a numerical value σ = 1, which of course hides the logarithmic component.
In that case the ratios will be
constant g
cn
+ + ··· .
∼µ 1+
rn =
cn−1
log n
n
Note that we do not insist the the first correction term is O(1/ log n), it could be a power
of a logarithm, or some other weakly decreasing function, but it cannot have a power-law
increase. For our purposes it suffices to take this term to be O(1/ log n). We show the
modified ratios (this gets rid of the O(1/n) term in the asymptotics) in Figures 38 and 39
which are the same plot, but the first uses only the 128 exact coefficients, while the second
uses the exact plus predicted ratios. From the first plot, it is clear that it would be an
article of faith that the locus is going to 16 as n → ∞. By contrast, the second plot makes
this conclusion far more plausible.
We next try and estimate the exponent σ, which should be 1, without assuming µ =
16. We use the method described below equation (17). With the 128 known terms, the
estimators of σ−2 are shown in Figure 40 and show no evidence of approaching the expected
value of −1. If however we use twice as many terms, so using the next 128 predicted ratios,
we get the plot shown in Figure 41, which is plausibly approaching −1.
This highlights the value of numerically predicting further terms wherever possible.
8. Analysis of Thompson’s group F
For Thompson’s group F it is known that the series grows exponentially like µn . If
µ = 16, the group is amenable. If it is amenable, there cannot be a sub-dominant term of
σ
the form κn with 0 < σ < 1, because the group contains the wreath products ZoZoZo· · ·oZ
as subgroups. This is a consequence of Theorem 1.3 in [20] and results in [21], and is proved
as Theorem 3.2 in Section 3.
We first study the modified ratios, defined by (14). The modified ratio plot against
1/n is shown in figure 42 and displays considerable curvature. By contrast, the same data
plotted against n−1/5 , and shown in figure 43 shows curvature in the opposite direction.
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
39
Figure 38. The first 128 modified ratios
for the Navas-Brin group B vs. 1/ log n.
Figure 39. The first 718 modified ratios
for the Navas-Brin group B vs. 1/ log n.
Figure 40. Estimates of σ − 2 from 128
terms of the Navas-Brin group B.
Figure 41. Estimates of σ − 2 from 256
terms of the Navas-Brin group B.
This is strong evidence for the presence of a conventional stretched-exponential term of the
sort we have seen in our study of the lamplighter group and the family Wd . As mentioned
above, the presence of such a term is incompatible with amenability. This is our first piece
40
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
of evidence that the group is not amenable. Note too that this is quite different to the
behaviour observed for the coefficients of the Navas-Brin group B.
Figure 42. Modified ratios vs.
Thompson’s group F.
1/n for
Figure 43. Modified ratios vs. n−1/5 for
Thompson’s group F.
In our subsequent analysis, we use both the exact coefficients and the extrapolated
coefficients. While all extrapolated terms can be used in calculating the ratios, once one
calculates first and second differences, errors are amplified, and so fewer terms can be used.
That is why we quote the number of terms used for different calculations, as it is only to the
quoted order that we are confident that the calculated quantities are accurate to graphical
accuracy.
To estimate the exponents in the stretched-exponential term we use the procedure described in Section 5.4, given by eqn. (19) and subsequent equations. This procedure allows
for the presence of a confluent power of a logarithm, so that the stretched-exponential
δ
σ
term is κn log n . In this way, based on a series of length 80, we show plots of estimators of
2 − σ and −δ in Figures 44 and 45, plotted against 1/n. Extrapolating these, we estimate
σ ≈ 1/2, and δ ≈ 1/2. Recall that this is exactly the stretched-exponential behaviour of
(Z o Z) o Z.
√
Reverting to the modified ratios, briefly discussed above, we plot these against 1/ n in
figure 46, using 186 terms. One observes that the plot still p
displays a little curvature, but
in Figure 47 the plot of these same modified ratios against log n/n, is essentially linear.
This is the appropriate power to extrapolate against, given our estimates of the stretchedexponential exponents. Extrapolating this to n → ∞ we estimate the limit, which gives
the growth constant, to be 14.8 − 15.1. This is well away from 16, which would be required
for amenability.
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
41
Figure 44. Estimators of σ−2 for Thompson’s group F vs. 1/n.
Figure 45. Estimators of −δ for Thompson’s group F vs. 1/n.
Figure 46. The first 186 modified
ratios
√
for Thompson’s group F vs. 1/ n.
Figure 47. The first 186 modified
ratios
p
for Thompson’s group F vs. log n/n.
One simple test for amenability uses the fact that the ratio of successive coefficients
asymptotes to the growth constant µ. For the lamplighter group, this ratio behaves as
c
1
(L)
rn = 9 1 + 2/3 + o
.
n
n2/3
42
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
For Z o Z one has
rn(2)
c · log2/3 n
+o
= 16 1 +
n2/3
log2/3 n
n2/3
!!
,
and for the triple wreath product, W2 , the corresponding result is
√
√
c log n
log n
(3)
+o
,
rn = 36 1 +
n1/2
n1/2
while for Thompson’s group F all we know is
rn = µ (1 + lower order terms) ,
where we suspect that the correction term is similar to that of the triple wreath product
of Z.
So, a simple test for amenability is to look at the three quotients
9rn
rn
4rn
, (2) , and
.
(L)
(3)
16rn
rn
9rn
If Thompson’s group F is amenable, these
p quotients should all go to 1. In Figures 48, 49,
50 we show these ratios plotted against log n/n, which is the appropriate power, though
this choice is not critical. The ratios do not appear to be going to 1 in any of the three
cases. For all cases we have used 200 ratios. To do this, we used the extended ratios
for Thompson’s group F and also extended the ratios for W2 from the known 132 ratios.
Indeed, all three cases are consistent with a limit around 0.93 ± 0.02, corresponding to
µ = 14.9 ± 0.3. This is entirely consistent with our previous estimate of µ ≈ 15.0.
Figure
48. Quotient
of Thompson group and
lamplighter group ratios
using 200 terms.
Figure 49. Quotient of
Thompson group and Z o Z
ratios using 200 terms.
Figure 50. Quotient of
Thompson group and (Z o
Z) o Z ratios using 200
terms.
Finally, we take the approach of extrapolating the lower bounds produced in Section 4.
√
Note that the sequence of bounds {bn } are bounds on µ. We have no expectation as to
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
43
how this sequence should approach its limit, so we first plot the bounds against 1/n in
figure 51. Some curvature is seen, which, as we have shown above, is evidence that the
locus behaves as
c2
c1
+ ··· .
bn ∼ b∞ 1 + α +
n
n
(1)
We remove the term O(1/n) in this case by forming the sequence bn = (n · bn − (n − 2) ·
bn−2 )/2 where we have shifted n by 2 to remove the effect of a small odd-even oscillation
√
(1)
if one shifts only by 1. We found that plotting bn against 1/ n gave a visually linear
plot, and this is shown in figure 52. Linearly extrapolating the last two entries gives the
estimate b∞ ≈ 3.875, so that µ ≈ 15.02, in agreement with previous estimates above.
Figure 51. Plot of bounds bn for Thompson’s group F against 1/n.
(1)
Figure 52. Plot of modified bounds
√ bn
for Thompson’s group F against 1/ n.
9. Conclusion
We have have given polynomial-time algorithms to generate terms of the cogrowth series
for several groups. In particular, we have given the first series for the Navas-Brin group B.
We have also given an improved algorithm for the coefficients of Thompson’s group F, giving
32 terms of the cogrowth series, extending previous enumerations by 7 terms. We analysed
these various series to develop numerical techniques to extract the asymptotics, and gave
improved asymptotics for the Heisenberg group. We gave an improved lower bound on the
growth-rate of the cogrowth series for Thompson’s group F, µ ≥ 13.2693 using the method
in [18]. We generalised their method, showing that the cogrowth sequences for all these
groups can be represented as the moments of a distribution. Extrapolation of the sequence
of bounds suggests the limit is around 15.0, which is incompatible with amenability.
44
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
For Thompson’s group F we proved that, if the group is amenable, there cannot be a
sub-dominant stretched exponential term in the asymptotics. The numerical data however
provides compelling evidence for the presence of such a term. This observation suggests a
potential path to a proof of non-amenability.
We have extended the sequence of 32 terms for group F by a further 200 terms (or, as
appropriate, 200 ratios of successive terms), which we demonstrate are sufficiently accurate
for the graphical approaches to analysis that we have taken.
A numerical study of the cogrowth sequence cn gives
σ
cn ∼ c · µn · κn
logδ n
· ng ,
where µ ≈ 15, κ ≈ 1/e, σ ≈ 1/2, δ ≈ 1/2, and g ≈ −1. The growth constant µ must be 16
for amenability. This estimate of the growth constant is the same as that obtained from
the extrapolated bounds. These three approaches to the study of amenabilty lead us to
the strong belief that Thompson’s group F is not amenable.
The difficulties we encountered in analysing Z o98 Z and the Navas-Brin group B does
imply that there do exist groups whose cogrowth series are difficult to analyse. Nevertheless, in both those cases we were able to extract the correct asymptotics. Furthermore, the
cogrowth series for Thompson’s group F did not behave like either of these two “difficult”
groups, and indeed appeared to have a stretched exponential term with exponent values
that were readily estimable. While we cannot rule out the presence of some previously
unsuspected pathology in the asymptotic form, we believe that we have presented strong
evidence for the belief that Thompson’s group F is not amenable
10. Appendix
Here are the (predicted) next 200 ratios for Thompson’s group F . That is, the first
ratio here is the coefficient of z 32 divided by the coefficient of z 31 . One standard deviation is 1.6 × 10−21 for the first ratio, 8.4 × 10−16 for the tenth ratio in this list, then
2.2 × 10−13 , 3.8 × 10−11 , 8.3 × 10−10 , 7.5 × 10−9 , 3.1 × 10−8 , 3.5 × 10−7 , 1.2 × 10−6 , 5.3 ×
10−6 , 3.3 × 10−5 , 9.0 × 10−5 , 1.3 × 10−4 , for the twentieth, thirtieth, fourtieth, fiftieth,
seventieth, ninetieth and hunderd and tenthth, hundred and thirtieth, hundred and fiftieth,
hundred and seventy-fifth and two hundredth ratios, respectively.
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
12.139382519134640546100910550116506
12.227584513675824849745149961326117
12.306364858371755791208652657890394
12.377372393555580912478863352420963
12.441869014094574198504147444796094
12.500840543278316369259845344418039
12.555070835196117213101528584412301
12.605192754131413602669204643768679
12.651724136967118435759923721500215
12.695093725180061251425934047225685
12.735660229810914461533323666877360
12.773726591097905739200796016562265
12.809550810208112073629601806420114
12.843354291696056209817237940576090
12.875328347113943574400714300355914
12.905639332151085077301331779253254
12.934432723922797241056261010730083
12.961836439248190289410424756818502
12.987963464694516602362349323441444
13.012914070454692135046857458367114
13.036777659579423368614400305996310
13.059634066593784780607389266957567
13.081555222866179498657087709047136
13.102605564520384538741633073251136
13.122843467661052218534876988958406
13.142322234451185942341571946995161
13.161089481701929123360329350075889
13.179189107961895176661932566775513
13.196660991623399487925223492990383
13.213542543992015500848473483969209
13.229866114857320555465055797943269
13.245661794080940364573430988057957
13.260959939896671980546840767426691
13.275786005549024172548265823401768
13.290161972217217703156676577608563
13.304113073325880062519253728849004
13.317658188282116335265357329472658
13.330818453996063928243407490059108
13.343606937084326659382677923429016
13.356041865699128533066777923821716
13.368138079655518061737105379768840
13.379909230285796422796584282677189
13.391391013985309598551994527199420
13.402540990852287889254736126437580
13.413411407143425386912606665853825
13.424062672750839568072161252780645
13.434394074827496721484051571897707
13.444482843412927251038029918715121
13.454316759036914691255964338022543
13.463903116875215568913679808238264
13.473213802552343969176414720413487
13.482346421441562120101495301037430
13.491179584354353752611550171676095
13.499895368095523512785232116864023
13.508333602964302346321586079889387
13.516654351365134683610700727371316
13.525000203908312007517604856476096
13.532917871456062697723758553984577
13.540554273647847830847507251037425
13.548063007209007082860967768914526
13.555319528240038802193441194513629
13.562641374331046546303564833879801
13.569777875695742862069575716786180
13.576510610487748479571836167835301
13.584419560999735066923148194498099
13.590831265318841752277570171711095
13.596998211959713327539754998388164
12.169952350800835818835333877031972
12.254800541517346423861272221078461
12.330828666879163731631451465102197
12.399541172868336566805611620482939
12.462097792906309126366784632966479
12.519410149156226556638850314986873
12.572206972475257876321270606856264
12.621079365262119276409607772782102
12.666512975937091192133384667614540
12.708911355958942268539084708497350
12.748613234321123627064424451294341
12.785905507056272163479082083260714
12.821033151370240355220794516217934
12.854206894770548244369354384398327
12.885609214745162856554802222623966
12.915399101706732471897030527017181
12.943715839380256116094239676192280
12.970682107319771206377319240620875
12.996406405676872159827925022024219
13.020985259779764800684267751152161
13.044504679013068113321447729698093
13.067041693422999439936815568787524
13.088665732145507263332155324286416
13.109439117936998695729698444879843
13.129418244175823190427250023546405
13.148654712658935599802876138058446
13.167195007769142327138772890144629
13.185080837993801098011934614407469
13.202352151456208348963284966653968
13.219044110867187297414288617379158
13.235188321357468124261384585429120
13.250815166741899371063964386311321
13.265952758167346821144011231194731
13.280627056720430964223829034834261
13.294858653114477361638435828704210
13.308670446314943572746000301996392
13.322086115777917425146418299128619
13.335121539230967797838900373066820
13.347786320277756923575569080786819
13.360110874363270214471235529540315
13.372097273034426235102779701961991
13.383762909644345312962700551534152
13.395144666969365978147066024081715
13.406196253710405879515140382321201
13.416988156570367066227724319777381
13.427535999385894978339537743097391
13.437779418859588380431049420003005
13.447782268650965022269207427038709
13.457544856354882926610701767183908
13.467042454042638170007261022024409
13.476302647287223508754966768268084
13.485293318045282335887952362450478
13.494084747058831414056697557041621
13.502701722883121030342927584362952
13.511107842566379075799459233305947
13.519431847277609941232036080607544
13.527641315577516333811095751023520
13.535493461803957879581733368101521
13.543084323366200838300039176315332
13.550504412576211288095020257342519
13.557762668158072770433170818766971
13.564958044304722734009028672257896
13.572043884450805383578987603438786
13.578837748814414710933467280925295
13.586612599432547359359203779080153
13.592959911744010496809257658001969
13.599059952154497874738089781219857
12.199326127345853916009149880943422
12.281040527431431456883217428846113
12.354482413853570301131793493786704
12.421027974747801610713466511114207
12.481745360450139667192011174605863
12.537479148348791828180725386201064
12.588907920692053314139061768312908
12.636584558849435543137582196072181
12.680965096098118923436248549296051
12.722429798828036178804146087289492
12.761298847540600696316471357098745
12.797844188914176808112395397996995
12.832298623805456848288841282862554
12.864862864837654839427475941273709
12.895711085169493589386718357886965
12.924995323677800300649200885376963
12.952849018349319276368496909220240
12.979389881806587929393146535001536
13.004722173676474514325001997839461
13.028938801180381908385125175110548
13.052122569524550318903068986491322
13.074347966865955839456185038291272
13.095681797656212049861838152969811
13.116184350936469838871049112592397
13.135910414942072482775413880736439
13.154910018341609905494554223902216
13.173227515058212755159290331997399
13.190904563910270596599158822879997
13.207979050670762502542808795553103
13.224484921862603128812749940479344
13.240453522986667992419868738673504
13.255914147651533345355162750305551
13.270895130906341393601998277628536
13.285419125190171039873138415974114
13.299508767603123209430266638744025
13.313185140342015149024233508130128
13.326471695739198888194127423197993
13.339384163779883682357243658284554
13.351930748190988263794674502771721
13.364142777237566574247813605863778
13.376020835131593723786616298972917
13.387582311376027194727073026599142
13.398865377039202234681959848510147
13.409819665483104102788219559926390
13.420539883769483957581690933051480
13.430979699159296367928654845659964
13.441136013513055381428134566110019
13.451072044972393249466607491105591
13.460737144623883340126835494564910
13.470155360061132928954773338423225
13.479337389227819997111748990692973
13.488249148435842618350149753783486
13.497011084300140602722763616549800
13.505547791585498408392624616902183
13.513936036377083173562688008219201
13.522232264247358404187723455476999
13.530300873952004160088528545275672
13.538046382601181802194312557622836
13.545599058351387123157120673148797
13.552923260697021481475229482726246
13.560302341521164494357743132135167
13.567252292374466826596567631465810
13.574288148599622255870938533612831
13.581263278311749328511560181131020
13.588785775454742632035040598882148
13.594915982519419341022644469998104
45
46
ANDREW ELVEY PRICE AND ANTHONY J GUTTMANN
11. Acknowledgements
We wish to thank Andrew Rechnitzer for many stimulating discussions on this topic, and
Murray Elder for helpful comments on the manuscript. AJG wishes to thank Nathan Clisby
for his vastly superior version of the program to use differential approximants to predict
further terms and ratios. AEP wishes to thank ACEMS for financial support through a
PhD top-up scholarship.
References
[1] N Beaton, A J Guttmann, I Jensen and G Lawler, Compressed self-avoiding walks, bridges and polygons, J. Phys A: Math. Theor. 48 454001 (27pp) (2015).
[2] J Belk and K S Brown, Forest diagrams for elements of Thompson’s group F , Int. J. Alg. Comp. 15
815-850 (2005).
[3] S Blake Fordham, Minimal length elements of Thompsons group F. Geom. Dedicata 99 179-220 (2003).
[4] M G Brin, Elementary amenable subgroups of R. Thompson’s group F, Intl. J. Alg. and Comp. 15
619-642 (2005).
[5] J M Cohen, Cogrowth and amenability of discrete groups, J. Funct. Anal, 48 301-309 (1982).
[6] M Elder, E Fusy and A R Rechnitzer, Counting elements and geodesics in Thompson’s group F , J.
Alg. 324 102-121 (2010).
[7] M Elder, A Rechnitzer and E J Janse van Rensburg, Random Sampling of Trivial Words in Finitely
Presented Groups, Expr Math 24 391-409 (2015).
[8] M Elder, A R Rechnitzer and T Wong, On the cogrowth of Thompson’s group F , Groups - Complexity
- Cryptology, 4 301-320 (2012).
[9] F Gantmakher and M Krein, Sur les matrices completement non-négatives et oscillatoires, Compositio
Mathematica 4 445-476 (1937).
[10] D Gretete, Random walks on a discrete Heisenberg group, Rend. Circ. Mat. Palermo 60 329-335 (2011).
[11] R I Grigorchuk, Symmetric random walks on discrete groups, in Multicomponent random systems (eds.
R L Dobrushin and Ya. G Sinai), Nauka, Moscow (1978).
[12] V S Guba, On the properties of the Cayley graph of Richard Thompson’s group F, Internat. J. Alg.
Comp. 14 (56) 677-702, International Conference on Semigroups and Groups in honor of the 65th
birthday of Prof. John Rhodes (2004).
[13] A J Guttmann, in Phase Transitions and Critical Phenomena, vol 13, eds. C Domb and J Lebowitz,
Academic Press, London and New York, (1989).
[14] A J Guttmann, Analysis of series expansions for non-algebraic singularities, J. Phys A: Math. Theor.
48 045209 (33pp) (2015).
[15] A J Guttmann, Series extension: Predicting approximate series coefficients from a finite number of
exact coefficients, J. Phys A: Math. Theor. 49 415002 (27pp) (2016).
[16] A J Guttmann and I Jensen, Series Analysis. Chapter 8 of Polygons, Polyominoes and Polycubes
Lecture Notes in Physics 775, ed. A J Guttmann, Springer, (Heidelberg), (2009).
[17] A J Guttmann and G S Joyce, A new method of series analysis in lattice statistics, J Phys A, 5 L81–
84, (1972).
[18] S Haagerup, U Haagerup and M Ramirez-Solano, A computational approach to the Thompson group
F , Int. J. Alg. and Comp. 25 381-432 (2015).
[19] A Navas, Quelques groupes moyennables de difféomorphismes de l’intervalle, Bol. Soc. Mat. Mexicana
10 219-244 (2004).
[20] C. Pittet and L Saloff-Coste, On the stability of the behavior of random walks on groups, J. Geom.
Anal. 10 No 4 713-737 (2000).
[21] C. Pittet and L Saloff-Coste, On random walks on wreath products, Ann. of Probab. 30 No.2 948-977
(2002).
NUMERICAL STUDIES OF THOMPSON’S GROUP F AND RELATED GROUPS
47
[22] D Revelle, Heat kernel asymptotics on the lamplighter group, Elect. Comm. in Probab. 8 142-154
(2003).
[23] T-J Stieltjes, Recherches sur les fractions continues, Annales de la Faculté des sciences de Toulouse:
Mathématiques. 8 No. 4. J1-J122 (1894).
| 4math.GR
|
Full-Frame Scene Coordinate Regression for
Image-Based Localization
Xiaotian Li, Juha Ylioinas and Juho Kannala
arXiv:1802.03237v1 [cs.CV] 9 Feb 2018
Aalto University
[email protected]
Abstract—Image-based localization, or camera relocalization,
is a fundamental problem in computer vision and robotics, and it
refers to estimating camera pose from an image. Recent state-ofthe-art approaches use learning based methods, such as Random
Forests (RFs) and Convolutional Neural Networks (CNNs), to
regress for each pixel in the image its corresponding position
in the scene’s world coordinate frame, and solve the final pose
via a RANSAC-based optimization scheme using the predicted
correspondences. In this paper, instead of in a patch-based
manner, we propose to perform the scene coordinate regression
in a full-frame manner to make the computation efficient at
test time and, more importantly, to add more global context
to the regression process to improve the robustness. To do so,
we adopt a fully convolutional encoder-decoder neural network
architecture which accepts a whole image as input and produces
scene coordinate predictions for all pixels in the image. However,
using more global context is prone to overfitting. To alleviate this
issue, we propose to use data augmentation to generate more
data for training. In addition to the data augmentation in 2D
image space, we also augment the data in 3D space. We evaluate
our approach on the publicly available 7-Scenes dataset, and
experiments show that it has better scene coordinate predictions
and achieves state-of-the-art results in localization with improved
robustness on the hardest frames (e.g., frames with repeated
structures).
I. I NTRODUCTION
Image-based localization is an important problem in computer vision and robotics. It addresses the problem of estimating the 6 DoF camera pose in an environment from an image.
It is a key component of many computer vision applications
such as navigating autonomous vehicles and mobile robotics
[7], simultaneous localization and mapping (SLAM) [8], and
augmented reality [4, 18].
Currently, there are plenty of image-based localization
methods proposed in the literature. Many state-of-the-art approaches [25, 26] are based on local features, such as SIFT,
ORB, or SURF [17, 23, 1], and efficient 2D-to-3D matching.
Given a 3D scene model, where each 3D point is associated
with the image features from which it was triangulated,
localizing a new query image against the model is solved
by first finding a large set of matches between 2D image
features and 3D points in the model via descriptor matching,
and then using RANSAC [9] to reject outlier matches and
estimate the camera pose on inliers. Although these local
feature based methods have been proven to be very accurate
and robust in many situations, due to the limitations of the
hand-crafted feature detector and descriptor, extremely large
viewpoint changes, repetitive structures and textureless scenes
often produce a large number of false matchings which will
lead to localization failure.
Recently, some promising neural network based localization
approaches have been proposed. Approaches such as PoseNet
[13] formulate 6 DoF pose estimation as a regression problem,
and thus no traditional feature extraction, feature description,
or feature matching processes are required. It has been shown
that these approaches overcome the limitations of the local
feature based approaches, i.e., they are able to recover the camera pose in very challenging indoor environments where the
traditional methods fail. However, their localization accuracy is
still far below traditional approaches in other situations where
local features perform well. Unlike PoseNet, approaches such
as DSAC [3] obtain the 6 DoF pose by a two-stage pipeline:
first, regressing a 3D point position for each pixel in an image,
and then determining the camera pose using RANSAC based
on these correspondences as in the conventional localization
pipeline. Results have shown that these methods are the current
state-of-the-art on the standard 7-Scenes benchmark [28].
In this paper, we follow the two-stage RGB-only localization pipeline and present a full-frame Coordinate CNN for
the scene coordinate regression in the first stage. This fullframe Coordinate CNN is trained to generate dense scene
coordinate predictions from a given RGB image. Compared
to the patch-based Coordinate CNN adopted in the DSAC
pipeline [3], which is trained to generate a single scene
coordinate prediction from an image patch centered around
a pixel, our network considers more global information and
can be evaluated efficiently at test time. Nevertheless, encoding more global context during regression is more prone to
overfitting than using local patches since local patches are
more invariant to viewpoint changes. Therefore, we propose
to use data augmentation to mitigate the problem. We show
with experiments on the widely used 7-Scenes dataset that
our approach has improved scene coordinate predictions and
achieves state-of-the-art results in localization with strong
robustness.
The contributions of this paper can be summarized as
follows:
• We adopt a fully convolutional encoder-decoder network
for the scene coordinate regression in the two-stage
image-based localization pipeline to encode more global
information.
• We do data augmentation in both 2D and 3D space for
the scene coordinate regression, and we show that data
RANSACBased
Optimization
6 DoF
Pose
RGB Image
Convolution
Scene Coordinate Prediction
Upconvolution
Fig. 1. Overview of our system. In this two-stage pipeline, the encoder-decoder full-frame Coordinate CNN accepts a whole RGB image as input and
produces scene coordinate predictions for all pixels in the image, and then the final pose is solved via a RANSAC-based optimization scheme using the
predicted correspondences.
•
augmentation is necessary for training the full-frame Coordinate CNN to achieve good localization performance.
Our approach achieves state-of-the-art results and shows
improved robustness on the hardest frames.
II. R ELATED W ORK
Solving the image-based localization problem consists of
estimating the 6 DoF camera pose of a query image in an
arbitrary scene. This is also known as camera relocalization.
More specifically, the localization system takes only one image
as input, and then outputs camera pose according to a given
representation of the scene. The representation of the scene
depends on the approach used. It can be a database of images,
a reconstructed 3D model, or a trained deep neural network.
Here we only consider one image as input, although the imagebased localization task can be extended by using a sequence
of images as input.
In the earliest stages, image-based localization was solved
by treating it as a location recognition problem [34]. In these
approaches, image retrieval techniques are often applied to
determine the location of the query image by matching it to
a database of images and finding the most similar images to
it. The locations of the database images are known. Thus, the
location of the query image can be estimated according to the
known locations. Initially, these methods worked on databases
consisting of tens of thousands of images and then they were
improved to deal with more than a million images. However,
the localization performance provided by these approaches is
Convolution
limited by the accuracy of the known locations of the database
Upconvolution
images [24] and the density of the key-frames.
To obtain better localization performance, more recent techniques [11] are based on more detailed and structured information, i.e., they use a reconstructed 3D point cloud model,
usually obtained from Structure-from-Motion, to represent the
scene. This is enabled by some powerful SfM approaches such
as Bundler [29]. It is even possible to construct huge models
with millions of points [30]. Thus, image-based localization
systems are enabled to work on a city-scale level. Instead
of matching the query image to the database images, these
keypoint based methods solve the localization task by finding
correspondences between the query image and the 3D model.
This is achieved by the use of the conventional hand-crafted
features, e.g., SIFT [17]. The features of the 3D points are
computed during the 3D reconstruction, and the features of the
query image are detected and extracted at test time. Therefore,
the correspondence search is formulated as a descriptor matching problem. After establishing the 2D-3D correspondences,
the camera pose is determined by a Perspective-n-Point [14]
solver inside a RANSAC [9] loop.
It is obvious that the pose estimation step can only succeed
if the quality of the established correspondences is good
enough. When the 3D model is too big in size or too complex,
an efficient and effective descriptor matching step is essential.
There are several techniques in the literature to handle this
problem, such as 2D-to-2D-to-3D matching [11], 3D-to-2Dmatching [15], prioritized matching [24] and co-visibility
filtering [25]. However, even if the matching process is enough
effective and efficient, limitations of the hand-crated features
will still cause these keypoint based localization systems to
fail.
Recently, it has been shown that machine learning methods
have great potential to tackle the problem of image-based
localization. PoseNet [13] demonstrates the feasibility of formulating 6 DoF pose estimation as a regression problem.
The pose of a query image is directly regressed by a deep
CNN with GoogLeNet [31] architecture pre-trained on largescale image classification data. However, its performance is
still below the state-of-the-art keypoint based methods and
far from ideal for practical camera relocalization. In order
to improve the accuracy of PoseNet, several variants have
been proposed in recent papers. For example, LSTM-Pose
[33] makes use of LSTM units [10] on the CNN output to
exploit the structured feature correlation. The LSTM units
play the role of a structured dimensionality reduction on the
feature vector and lead to drastic improvements in localization
performance. Another variant, Hourglass-Pose [21], is based
on hourglass architecture which consists of a chain of convolution and upconvolution layers followed by a regression
part. The upconvolution layers are introduced to preserve
the fine-grained information of the input image and this
mechanism has been proven to be able to further improve
the accuracy of image-based localization using CNN based
architectures. Besides, it has been shown that the use of a
novel loss function based on scene reprojection error can also
significantly improve PoseNet’s performance [12].
Unlike PoseNet, the scene coordinate regression forests
(SCoRF) approach [28] adopts a regression forest [6] to
generate 2D-3D matches from an RGB-D input image instead
of directly regressing the camera pose. The final camera
pose is then determined via a RANSAC-based solver. This
whole pipeline is similar to the keypoint based localization
approaches, but no traditional feature extraction, feature description, or feature matching processes are required. The
original SCoRF pipeline is further improved by exploiting
uncertainty in the model [32]. Training the random forest to
predict multimodal distributions of scene coordinates results
in increased pose accuracy.
In order to localize RGB-only images as well, the original
SCoRF is extended by utilizing the increased predictive power
of an auto-context random forest [2]. In the most recent
DSAC paper [3], a neural network based SCoRF pipeline
for RGB-only images is proposed. In contrast to previous
SCoRF pipelines, two CNNs are adopted for predicting scene
coordinates and for scoring hypotheses respectively. Moreover,
the conventional RANSAC is replaced by a new differentiable
RANSAC, which enables the whole pipeline to be trained endto-end. In [19], Massiceti et al. further explore efficient versus
non-efficient and RF- versus non-RF-derived NN architectures
for camera localization and propose a new fully-differentiable
robust averaging technique for regression ensembles.
III. M ETHOD
Our CNN based approach presented in this paper is inspired
by the state-of-the-art RGB-only DSAC pipeline [3], and it
consists of two stages. In the first stage, a Coordinate CNN
is adopted to generate 2D-3D correspondences from a given
RGB-only image. In the second stage, a RANSAC-based
scheme is performed to determine the final pose estimate. The
overall pipeline is illustrated in Figure 1.
The Coordinate CNN used in the original DSAC pipeline
[3] is trained and evaluated in a patch-based manner. Similar
to the patch-based approaches for semantic segmentation,
patches need to be sampled during both training and test
time. The patch-based Coordinate CNN is trained using image
patches of fixed size (42 × 42) centered around pixels and the
corresponding world coordinates of these pixels. At test time,
patches of the same size are used as inputs of the Coordinate
CNN to generate 2D-3D correspondences. However, obtaining
a set of 2D-3D correspondences in such a way is very
inefficient. In the DSAC pipeline [3], 1600 correspondences
are needed for pose estimation. Thus 1600 patches should
be extracted and fed into the deep neural network. This is
obviously time-consuming and also requires much memory.
Moreover, the fixed patch size limits the size of the receptive
field, and thus limits the information that can be processed by
the network. This makes it difficult for the Coordinate CNN
to predict accurate 2D-3D correspondences when the scene
exhibits ambiguities (e.g., repeated structures), since the global
context cannot be used by the network.
To overcome such limitations, we propose to use a encoderdecoder CNN which accepts a whole image as input and
produces scene coordinate predictions for all pixels in the
image. This is inspired by recently presented architectures
for solving per-pixel tasks, such as semantic segmentation,
depth estimation, and disparity estimation. Performing scene
coordinate regression in this way, patches are not required to
be sampled and the 2D-3D correspondences can be generated efficiently at test time. Since the receptive field of the
encoder-decoder architecture covers almost the entire input
image, more global context is added to the regression process and more overall information is considered. However,
the global image appearance may change a lot due to only
small viewpoint changes. That is, it is not as invariant to
viewpoint changes as local appearance. Therefore, regressing
the scene coordinates using more global information is prone
to overfitting. We propose to use data augmentation to mitigate
the problem. We call our network full-frame Coordinate CNN.
A. Network Architecture
Our full-frame Coordinate CNN is based on the architecture
of DispNet [20]. It is fully convolutional [16] such that
dense per-pixel scene coordinate predictions can be generated
from arbitrary-sized input images. We call the outputs scene
coordinate images. The network consists of a contractive part
and an expanding part. The input image is first spatially
compressed via the contractive part and then refined via the
expanding part. Moreover, shortcut connections are added in
between to overcome the data bottleneck. Unlike DispNet [20],
there is only one final output layer at the end of the network
and no multi-scale side predictions are used. Instead of ReLU
[22], we use ELU [5] for the nonlinearity between layers. The
details of our network architecture are described in Table I.
Layers start with upconv are upconvolutional layers and all
other layers are convolutional. Also for each layer, the size of
the kernel, the stride, the number of input channels, the number
of output channels, and the input (+ is a concatenation) are
given.
B. Training Loss
Similar to the patch-based Coordinate CNN used in the
original DSAC pipeline, we train our full-frame Coordinate
CNN by minimizing the Euclidean distance between the scene
coordinate ground truth and the prediction as given in Equation
1. Unlike training in a patch-based manner where patches and
single predictions are used and only patches centered at pixels
with valid depth values are sampled for training, our network
uses pairs of color image and scene coordinate image (dense
scene coordinate values) as training samples where the ground
truth scene coordinates can be missing for some pixels. Here
we simply ignore the pixels without the ground truth scene
coordinates and mask out their contributions to the final loss.
TABLE I
O UR NETWORK ARCHITECTURE .
Name
conv1a
conv1b
conv2a
conv2b
conv3a
conv3b
conv4a
conv4b
conv5a
conv5b
conv6a
conv6b
conv7a
conv7b
upconv6
iconv6
upconv5
iconv5
upconv4
iconv4
upconv3
iconv3
upconv2
iconv2
upconv1
iconv1
upconv0
iconv0
coord pred
Kernel
7×7
7×7
5×5
5×5
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
3×3
Str.
2
1
2
1
2
1
2
1
2
1
2
1
2
1
2
1
2
1
2
1
2
1
2
1
2
1
2
1
1
Ch I/O
3/32
32/32
32/64
64/64
64/128
128/128
128/256
256/256
256/512
512/512
512/512
512/512
512/512
512/512
512/512
1024/512
512/512
1024/512
512/256
512/256
256/128
256/128
128/64
128/64
64/32
64/32
32/16
16/16
16/3
Input
image
conv1a
conv1b
conv2a
conv2b
conv3a
conv3b
conv4a
conv4b
conv5a
conv5b
conv6a
conv6b
conv7a
conv7b
upconv6+conv6b
iconv6
upconv5+conv5b
iconv5
upconv4+conv4b
iconv4
upconv3+conv3b
iconv3
upconv2+conv2b
iconv2
upconv1+conv1b
iconv1
upconv0
iconv0
More formally, loss for a training sample can be written as:
X
loss =
Mij kŶij − Yij k
(1)
i,j
where Y and Ŷ are ground truth scene coordinate image and
scene coordinate image prediction respectively, M is a mask
and (i, j) is a 2D pixel coordinate. The mask M has the same
resolution as the color image and the scene coordinate image.
A pixel of M is set to 1 if the corresponding scene coordinate
ground truth exists and 0 otherwise.
C. Data Augmentation
While our network has larger receptive field and can encode
more global context information for better understanding of
the scene, it is more prone to overfitting compared to a network
trained and tested in a patch-based manner. This is because
local patches are more invariant to viewpoint changes and we
can usually sample sufficiently enough patches for training.
However, in our case, we need more training images taken
from a lot of different viewpoints to regularize the network,
since global image appearance is more sensitive to viewpoint
changes. Unfortunately, a common problem for image-based
localization is that the training poses in the whole pose space
is not complete enough to regularize such a network. Also,
image-based localization is typically performed scene-wise,
which means that only the images belonging to a particular
scene can be used during training. It does not make sense
to use data belonging to other scenes, and unlike for other
computer vision tasks such as semantic segmentation, it is not
Fig. 2. Data augmentation. Left: original RGB and sence coordinate images.
Middle: 2D transformation. Right: 3D transformation.
straightforward to apply transfer learning technique for imagebased localization. Hence, overfitting can be a serious problem
in training our full-frame Coordinate CNN for solving imagebased localization.
To alleviate the overfitting problem, we propose to use data
augmentation to generate more data for training. A simple
and common way of doing data augmentation in the context
of solving per-pixel tasks is to apply 2D affine geometric
transformations to the training images. In our case, we apply
2D transformations to both the input RGB images and the
ground truth scene coordinate images. The transformations we
perform include translation, rotation, and scaling.
However, these 2D affine transformations do not always
generate correct input RGB images which are consistent with
3D scenes (although they are always consistent with their
transformed ground truth scene coordinate images). Since the
camera is in 3D space and 3D points are projected onto
the image plane according to a camera model, if camera
pose is changed, the transformation between the old and new
2D coordinates of 3D points is much more complex than
the combination of 2D transformations in the image plane.
Therefore, we propose to augment data also in 3D space. For
each training sample, since we already have the ground truth
scene coordinate image, we can generate a local 3D point
cloud from that. We then add a small random transformation
to the ground truth pose to synthesize a new camera pose.
Finally, using the new camera pose, the local 3D point cloud
and the known camera model, we project the points to the new
camera plane to generate a new RGB image and its ground
truth scene coordinate image.
D. Camera Pose Optimization
In the second stage of the two-stage pipeline, a RANSAC
algorithm is used in combination with a PnP algorithm to
determine the final pose estimation from the 2D-3D correspondences generated in the first stage. Here the RANSAC
pose optimization algorithm we use is similar to the one used
in the DSAC pipeline [3]. The only differences are that firstly
we simply score the hypotheses by counting inliers instead of
using a Score CNN [3] to regress the scores of hypotheses
from their reprojection error images, and secondly we use
the conventional argmax hypothesis selection instead of the
probabilistic selection. We believe the Score CNN used in [3]
is prone to overfitting.
The detailed steps of the RANSAC-based pose optimization
algorithm used in this paper are described in Algorithm 1.
Here, score(Hi ) is the number of inliers, i.e., the number of
correspondences with reprojection errors less than a threshold
τ , and the final pose Hf is the one with highest inlier count.
Following [3], 40 × 40 correspondences are sampled (N =
1600), 256 hypotheses are sampled per image (K = 256), the
inlier threshold τ is set to 10 pixels, 8 refinement steps are
performed (R = 8), in each refinement step at most 100 inliers
are chosen (P = 100), and the refinement is stopped in case
less than 50 inliers have been found (Q = 50).
Algorithm 1 Pseudocode for RANSAC optimization
01: evaluate the Coordinate CNN to obtain N correspondences
02: i ← 0
03: while i < K do
04:
sample 4 2D-3D correspondences
05:
generate a pose hypothesis h using PnP
06:
if all the 4 correspondences are inliers of h then
07:
Hi ← h
08:
Si ← score(Hi )
09:
i←i+1
10: select the final pose Hf according to the scores
11: j ← 0
12: while j < R do
13:
sample at most P inliers
14:
if less than Q inliers are found then
15:
break
16:
recalculate the pose Hf from the inliers
17: return pose Hf
IV. E XPERIMENTS
In this section, we evaluate the performance of our system,
and compare it with other state-of-the-art methods. We use the
7-Scenes dataset [28] for benchmarking.
A. Evaluation Dataset
For the experiments, we use the 7-Scenes dataset [28]
provided by Microsoft Research. The 7-Scenes dataset is a
widely used RGB-D dataset which consists of seven different
indoor environments. The RGB-D images are captured using a
handheld Kinect camera at 640×480 resolution and associated
with 6 DoF ground truth camera poses obtained via the
KinectFusion system. A dense 3D model is also available for
each scene. Each scene contains multiple sequences of tracked
RGB-D camera frames and these sequences are split into training and testing data. The dataset exhibits several challenges,
namely motion blur, illumination changes, textureless surfaces,
repeated structures (e.g., in the Stairs dataset), reflections (e.g.,
in the Redkitchen dataset), and sensor noise. Following [3],
we use the depth images to generate the ground truth scene
coordinates. At test time, we only use RGB images to estimate
poses.
B. Implementation Details
We train our network from scratch for 800 epochs with a
batch size of 16 (i.e., 200k updates for Chess) using the Adam
optimizer where β1 = 0.9, β2 = 0.999, and = 10−8 . The
loss is computed as described in Section III-B. The initial
learning rate is set to 0.0001 and is halved every 200 epochs
until the end.
As mentioned in Section III-C, we perform data augmentation online during network training. We perform the 2D
affine transformation with a 40% chance, perform the 3D
transformation with a 50% chance, and use the original image
with a 10% chance. For the 2D transformation, we uniformly
sample translation from the range [−20%, 20%] of the image
width and height for x and y respectively, sample rotation from
[−45◦ , 45◦ ] and sample scaling from [0.7, 1.5]. All pixels
without valid RGB values are padded with the same random
value. For the 3D transformation, we uniformly sample the
rotational axis and the rotational angle is uniformly sampled
from [0◦ , 60◦ ]. The direction of the translation vector is again
uniformly sampled, and its magnitude in mm is sampled from
[0, 200]. All pixels without valid RGB values are padded with
different random values. Figure 2 shows an example of the data
augmentation.
At test time, although our full-frame Coordinate CNN can
generate 640 × 480 scene coordinate predictions, we only
sample 40×40 of the predictions for the pose estimation stage
to make it consistent with the DSAC pipeline [3].
C. Results
We report both median error of camera pose estimations
and accuracy measured as the percentage of query images
for which the camera pose error is below 5◦ and 5cm. The
results of our method on the 7-Scenes dataset together with
two current state-of-the-art methods are shown in Table II.
Following [3], Complete denotes the combined set of frames
(17000) of all scenes. We also plot the cumulative histograms
of localization errors for a more detailed comparison and
report the results of [25] and DSAC* (i.e., DSAC [3] patchbased scene coordinate regression + conventional RANSAC).
Note that for the DSAC pipeline [3], only the accuracy is
reported in the original paper, and we obtain the additional
results using the publicly available source code and trained
models provided by the authors. For [25], we use Colmap
[27] to construct models from scratch and register the models
against the ground truth poses of the training images, and again
we use the publicly available implementation of [25].
As we can see, the overall performance of DSAC* is
better than the original DSAC [3]. While achieving almost
the same results on the easy frames, DSAC* has superior
performance on the harder ones, i.e., DSAC* provides better
0.95 quantiles, especially for scenes that have fewer training
images (Fire, Heads). More importantly, DSAC* is able to
produce reasonable median localization error for Stairs, though
TABLE II
R ESULTS OF OUR APPROACH ON THE 7-S CENES DATASET. O UR APPROACH OUTPERFORMS TWO STATE - OF - THE - ART METHODS AND HAS SIGNIFICANTLY
IMPROVED PERFORMANCE ON S TAIRS . T RAINING THE FULL - FRAME C OORDINATE CNN WITHOUT OUR PROPOSED DATA AUGMENTATION
DRAMATICALLY DEGENERATES THE LOCALIZATION PERFORMANCE .
Scene
Chess
Fire
Heads
Office
Pumpkin
Red Kitchen
Stairs
Average
Complete
Brachmann
et al. [2]
94.9%
73.5%
48.1%
53.2%
54.5%
42.2%
20.1%
55.2%
55.2%
5◦ , 5cm
Brachmann
et al. [3]
94.6%
74.3%
71.7%
71.2%
53.6%
51.2%
4.5%
60.1%
62.5%
Ours
noaug
43.2%
10.7%
14.1%
30.6%
18.5%
38.8%
0.2%
22.3%
28.0%
Ours
88.5%
62.3%
75.1%
73.1%
51.4%
60.4%
29.5%
62.9%
64.9%
the accuracy on the hardest frames of Stairs is still terrible.
The results of DSAC* suggest that the use of the Score CNN
[3] could make the entire localization pipeline less robust.
According to the results, our proposed method outperforms
the two state-of-the-art methods [2, 3] and the traditional
keypoint based method [25], and its overall localization performance is more robust. To be specific, our method significantly
improves the performance on the hardest frames. We find that
while DSAC [3] (as well as DSAC*) and [25] often have
extreme localization errors, i.e., rotational errors close to 180◦
and rotational errors larger than 2-5m, the maximum errors
of our methods are always reasonable. As shown in Figure
3, the Ours curves always reach 1 very quickly compared to
others. The results suggest that the global information could
help the system to perform better in challenging scenarios.
Remarkably, our method is able to provide significantly better
localization performance for Stairs. Even the hardest frames
can be localized with the errors less than 0.41m and 13◦ .
It shows that the full-frame Coordinate CNN with enlarged
receptive field can better cope with the repetitive structures,
while the patch-based Coordinate CNN [3] and the local
keypoint based method [25] are limited due to their local
nature. Our method is also on-par with [19]. However, in [19],
only the overall accuracy is reported and this prevents us from
carrying out a more detailed comparison.
To show that our full-frame Coordinate CNN is more
efficient at test time, we present the runtimes of the CNNs
in Table III. We see that our full-frame Coordinate CNN is
one order of magnitude faster than the patch-based one [3].
And this comparison is even not fair since our full-frame
Coordinate CNN produces 640 × 480 predictions while only
40 × 40 are generated by the patch-based one [3]. In addition,
our full-frame Coordinate CNN has a relatively small model
size compared to the regression forest based model [19] as
shown in Table IV.
Following [19], in addition to the direct measure of localization performance, we present the accuracy of the intermediate
scene coordinate predictions on the test images. In Table V,
we report the percentage of scene coordinate inliers and the
mean Euclidean distance between the inliers and their ground
truth scene coordinate labels. A prediction is considered as
Brachmann
et al. [2]
1.5cm, 1.3◦
3.0cm, 1.4◦
5.9cm, 3.4◦
4.7cm, 1.7◦
4.3cm, 2.1◦
5.8cm, 2.2◦
17.4cm, 7.0◦
6.1cm, 2.7◦
-
Median Error
Brachmann
Ours
et al. [3]
noaug
2.1cm, 0.7◦
5.9cm, 1.9◦
2.8cm, 1.0◦
13.2cm, 4.5◦
2.0cm, 1.3◦
12.6cm, 9.4◦
3.4cm, 1.0◦
7.0cm, 2.0◦
4.7cm, 1.3◦
10.5cm, 2.7◦
5.0cm, 1.5◦
6.1cm, 1.7◦
1.9m, 49.4◦
43.4cm, 10.1◦
30cm, 8.0◦
14.1cm, 4.6◦
3.9cm, 1.6◦
8.0cm, 2.3◦
Ours
2.4cm,
3.7cm,
2.4cm,
3.5cm,
4.9cm,
4.2cm,
7.9cm,
4.1cm,
3.8cm,
0.8◦
1.4◦
1.7◦
1.0◦
1.3◦
1.2◦
2.1◦
1.4◦
1.2◦
TABLE III
T HE RUNTIMES OF THE FULL - FRAME AND PATCH - BASED C OORDINATE
CNN S .
GPU
NVIDIA GeForce GT 750M
NVIDIA GeForce GTX 1080
Full-frame
∼0.3s
∼0.02s
Patch-based
∼5s
∼0.3s
TABLE IV
T HE MEMORY CONSUMPTION OF OUR MODIFIED D ISP N ET COMPARED TO
THE REGRESSION FOREST BASED MODEL FROM [19].
Ours
∼120MB
Most compact from [19]
∼250MB
an inlier if its Euclidean distance to its ground truth label
is less than 10mm [19]. The normalized histograms of scene
coordinate errors are illustrated in Figure 4. As we can see,
our full-frame Coordinate CNN is able to produce significantly
better scene coordinate predictions. This shows why it is
more robust than the patch-based one. However, this does not
directly lead to equally better localization accuracy. This is
because the RANSAC-based optimizer is highly robust and
non-deterministic [19].
To evaluate the effectiveness of data augmentation, we also
report the performance of our method without the proposed
data augmentation. According to the results, training the fullframe Coordinate CNN without data augmentation dramatically degenerates the localization performance on all the
scenes. Without data augmentation, the full-frame Coordinate
CNN is unable to generalize well to unseen images. This
proves the importance of data augmentation which is used for
regularizing our full-frame Coordinate CNN during training.
TABLE V
T HE PERCENTAGES OF THE SCENE COORDINATE PREDICTION INLIERS
AND THE MEAN ERRORS OF THE INLIERS . O UR PREDICTIONS ARE
CLEARLY THE MOST ACCURATE .
DSAC [3]
59.4%, 41mm
Best from [19]
45.3%, 47mm
Ours
90.3%, 28mm
Fig. 3. Localization performance presented by cumulative histograms (normalized) of errors. Compared to other methods, our method is more robust, i.e.,
it has better performance on the hardest frames, since the curves always reach 1 very quickly.
V. C ONCLUSION
R EFERENCES
In this paper, we present a image-based localization approach using fully convolutional encoder-decoder network to
predict scene coordinates of RGB image pixels in a full-frame
manner. The proposed full-frame Coordinate CNN can process
more global information to better understand the scene. It
takes a whole image as input and produces scene coordinate
predictions for all pixels in the image and can be efficiently
evaluated at test time. The proposed data augmentation is
crucial for our approach to achieve good localization performance. The results show that our method provides more
accurate scene coordinate predictions, achieves state-of-theart localization performance and has improved robustness in
challenging scenarios.
[1] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf:
Speeded up robust features. In ECCV, 2006.
[2] Eric Brachmann, Frank Michel, Alexander Krull,
Michael Ying Yang, Stefan Gumhold, and Carsten
Rother. Uncertainty-driven 6d pose estimation of objects
and scenes from a single RGB image. In CVPR, 2016.
[3] Eric Brachmann, Alexander Krull, Sebastian Nowozin,
Jamie Shotton, Frank Michel, Stefan Gumhold, and
Carsten Rother. Dsac - differentiable ransac for camera
localization. In CVPR, 2017.
[4] Robert Oliver Castle, Georg Klein, and David W. Murray.
Video-rate localization in multiple maps for wearable
augmented reality. In ISWC, 2008.
[5] Djork-Arné Clevert, Thomas Unterthiner, and Sepp
Hochreiter. Fast and accurate deep network learning by
Fig. 4.
Histograms (normalized) of scene coordinate errors. Our method provides better scene coordinate predictions.
exponential linear units (elus). In ICLR, 2016.
[6] Antonio Criminisi and Jamie Shotton. Decision Forests
for Computer Vision and Medical Image Analysis.
Springer Publishing Company, Incorporated, 2013. ISBN
1447149289, 9781447149286.
[7] Mark Cummins and Paul Newman. Fab-map: Probabilistic localization and mapping in the space of appearance.
The International Journal of Robotics Research, 27(6):
647–665, 2008.
[8] Ethan Eade and Tom Drummond. Scalable monocular
slam. In CVPR, 2006.
[9] Martin A. Fischler and Robert C. Bolles. Random sample
consensus: A paradigm for model fitting with applications to image analysis and automated cartography.
Communications of the ACM, 24(6):381–395, 1981.
[10] Sepp Hochreiter and Juergen Schmidhuber. Long ShortTerm Memory. 9(8):1735–1780, 1997.
[11] Arnold Irschara, Christopher Zach, Jan-Michael Frahm,
and Horst Bischof. From structure-from-motion point
clouds to fast location recognition. In CVPR, 2009.
[12] Alex Kendall and Roberto Cipolla. Geometric loss
functions for camera pose regression with deep learning.
In CVPR, 2017.
[13] Alex Kendall, Matthew Grimes, and Roberto Cipolla.
Posenet: A convolutional network for real-time 6-dof
camera relocalization. In ICCV, 2015.
[14] Laurent Kneip, Davide Scaramuzza, and Roland Siegwart. A novel parametrization of the perspective-threepoint problem for a direct computation of absolute camera position and orientation. In CVPR, 2011.
[15] Yunpeng Li, Noah Snavely, and Daniel P. Huttenlocher.
Location recognition using prioritized feature matching.
In ECCV, 2010.
[16] Jonathan Long, Evan Shelhamer, and Trevor Darrell.
Fully convolutional networks for semantic segmentation.
In CVPR, 2015.
[17] David G. Lowe. Distinctive image features from scaleinvariant keypoints. International Journal of Computer
Vision, 60(2):91–110, 2004.
[18] Simon Lynen, Torsten Sattler, Michael Bosse, Joel A.
Hesch, Marc Pollefeys, and Roland Siegwart. Get out of
my lab: Large-scale, real-time visual-inertial localization.
In RSS, 2015.
[19] Daniela Massiceti, Alexander Krull, Eric Brachmann,
Carsten Rother, and Philip H. S. Torr. Random forests
versus neural networks - what’s best for camera relocalization? In ICRA, 2017.
[20] Nikolaus Mayer, Eddy Ilg, Philip Häusser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas
Brox. A large dataset to train convolutional networks
for disparity, optical flow, and scene flow estimation. In
CVPR, 2016.
[21] Iaroslav Melekhov, Juha Ylioinas, Juho Kannala, and
Esa Rahtu. Image-based localization using hourglass
networks. In ICCVW, 2017.
[22] Vinod Nair and Geoffrey E. Hinton. Rectified linear units
improve restricted boltzmann machines. In ICML, 2010.
[23] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and
Gary R. Bradski. ORB: an efficient alternative to SIFT
or SURF. In ICCV, 2011.
[24] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Fast
image-based localization using direct 2d-to-3d matching.
In ICCV, 2011.
[25] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Improving image-based localization by active correspondence
search. In ECCV, 2012.
[26] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Efficient
effective prioritized matching for large-scale image-based
localization. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 39(9):1744–1756, 2017.
[27] Johannes L. Schönberger and Jan-Michael Frahm.
Structure-from-motion revisited. In CVPR, 2016.
[28] Jamie Shotton, Ben Glocker, Christopher Zach, Shahram
Izadi, Antonio Criminisi, and Andrew W. Fitzgibbon.
Scene coordinate regression forests for camera relocalization in rgb-d images. In ICCV, 2013.
[29] Noah Snavely, Steven M. Seitz, and Richard Szeliski.
Modeling the world from internet photo collections.
International Journal of Computer Vision, 80(2):189–
210, 2008.
[30] Christoph Strecha, Timo Pylvninen, and Pascal Fua.
Dynamic and scalable large scale image reconstruction.
In CVPR, 2010.
[31] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott E. Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going
deeper with convolutions. In CVPR, 2015.
[32] Julien P. C. Valentin, Matthias Nießner, Jamie Shotton,
Andrew W. Fitzgibbon, Shahram Izadi, and Philip H. S.
Torr. Exploiting uncertainty in regression forests for
accurate camera relocalization. In CVPR, 2015.
[33] Florian Walch, Caner Hazirbas, Laura Leal-Taixe,
Torsten Sattler, Sebastian Hilsenbeck, and Daniel Cremers. Image-based localization using lstms for structured
feature correlation. In ICCV, 2017.
[34] Wei Zhang and Jana Kosecka. Image based localization
in urban environments. In 3DPVT, 2006.
| 1cs.CV
|
Improved Adaptive Resolution Molecular Dynamics
Simulation
Iuliana Marin1, Virgil Tudose2, Anton Hadar2, Nicolae Goga1,3, Andrei Doncescu4
Affiliation 1: Faculty of Engineering in Foreign Languages
University POLITEHNICA of Bucharest, Bucharest, Romania, [email protected]
Affiliation 2: Department of Strength of Materials, University POLITEHNICA of Bucharest, Bucharest, Romania
Affiliation 3: Molecular Dynamics Group, University of Groningen, Groningen, Netherlands
Affiliation 4: Laboratoire d'analyse et d'architecture des systemes, Université Paul Sabatier de Toulouse, Toulouse, France
Abstract—Molecular simulations allow the study of properties
and interactions of molecular systems. This article presents an
improved version of the Adaptive Resolution Scheme that links two
systems having atomistic (also called fine-grained) and coarsegrained resolutions using a force interpolation scheme. Interactions
forces are obtained based on the Hamiltonian derivation for a given
molecular system. The new algorithm was implemented in
GROMACS molecular dynamics software package and tested on a
butane system. The MARTINI coarse-grained force field is applied
between the coarse-grained particles of the butane system. The
molecular dynamics package GROMACS and the Message Passing
Interface allow the simulation of such a system in a reasonable
amount of time.
Keywords—adaptive resolution scheme; molecular dynamics;
MPI; stochastic dynamics; coarse-grained; fine-grained.
INTRODUCTION
The traditional computational molecular modeling
indicates the general process of describing complex chemical
systems in terms of a realistic atomic model, with the goal of
understanding and predicting macroscopic properties based on
detailed knowledge at an atomic scale [1]. In literature, the
study of a system at atomistic level can be defined as finegrained (FG) modeling [2].
I.
Coarse-graining is a systematic way of reducing the
number of degrees of freedom for a system of interest.
Typically whole groups of atoms are represented by single
beads and the coarse-grained (CG) force fields describe their
effective interactions. Coarse-grained models are designed to
reproduce certain properties of a reference system. This can be
either a full atomistic model or a set of experimental data. The
coarse-grained potentials are state dependent and should be reparameterized depending on the system of interest and the
simulation conditions [1].
The main disadvantage of a coarse-grained model is that
the precise atomistic details are lost. In many applications it is
important to preserve atomistic details for some region of
special interest [2]. To combine the two systems, fine-grained
and coarse-grained, it was developed a so-called multiscale
simulation technique [3, 4, 5, 6, 7]. This method combines the
two resolutions by coupling them with a mixing parameter .
The multiscale interaction forces are computed as a weighted
sum of the interactions on fine-grained and coarse-grained
levels [2].
The adaptive resolution scheme (AdResS) couples two
systems with different resolutions by a force interpolation
scheme. The two resolutions are called atomistic and coarsegrained [3, 4]. The atomistic representation of system
describes the phenomena produced at the atomic level (noted
fine-grained or FG) and the other one, coarse-grained,
describes the system at molecular level (noted coarse-grained
or CG).
The AdResS scheme works in the following conditions:
•
the forces between molecules are scaled and the
potential energy is not scaled. Consequently, the Hamiltonian
cannot be used to obtain the energies;
•
just non-bonded interactions between molecules are
scaled. In this way, some computational resources are saved;
•
particles are kept together via bonded atomistic
forces that are computed everywhere, including the coarsegrained part (no pure coarse-grained in that part);
•
in the transition regions there is a potential which is
added (that takes into account the difference in chemical
potentials between the two representations). In this paper it is
denoted by .
In the current paper, it is proposed a method for obtaining
the interaction forces by applying the Hamiltonian derivation,
where potentials can be scaled. The advantage of this method
is that, derivation is based on the sound Hamiltonian model and
energies can be reported correctly. The improved algorithm
was implemented in GROMACS, a molecular dynamics
software, that runs in parallel using the Message Passing
Interface (MPI). Simulations have been done on butane. The
MARTINI coarse-grained force field which was implemented
by the Molecular Dynamics Group from the University of
Groningen was also applied.
The paper is organized as follows. In the next section the
improved AdResS multiscaling resolution scheme is presented.
Section 3 describes the implementation of AdResS in
GROMACS using MPI and its testing on a butane molecular
system. Section 4 outlines the obtained results. The last section
presents the final conclusions.
II. RELATION TO EXISTING THEORIES AND WORK
In this section it is presented the relevant theory for the
algorithm implemented in GROMACS – the improved
AdResS multiscaling resolution scheme [3,4].
A convention is made: for a coarse-grained molecule are
used capital letters for coordinates and for a fine-grained
molecule, small letters. For simplifying the presentation, it is
considered that the space factor is function only on the
coordinate of the reference system. On the and coordinate
axis, is constant.
It is denoted by
the
coordinate of a fine-grained
particle with the number
in the coarse-grained molecule
with number . The coarse-grained molecule is centered in
the center of mass ( ) of the ensemble of corresponding finegrained particles .
The relation between the two parameters is:
∑
(1)
∑
Fig. 2. The coordinates
is
for a coarse-grained particle
A. Multiscaling force
Starting from the Hamiltonian that scales the non-bonded
forces, it is added an extra potential in the transition region
that adjusts for the difference in chemical potentials between
the fine-grained and coarse-grained regions.
As explained above, fine-grained bonded interactions are
computed everywhere, including the coarse-grained region.
The potential energy in the multiscaling model
is a
function of coordinates only and it is chosen as: The
multiscaling factor is defined as follows:
"/
-.
4,
(2)
where
is the mass of the fine-grained particle and
the mass of the coarse-grained particle.
and
"∙
.
+,
-.
+,
/
.
4,
/ 01 2
" /
"3 ∙
(4)
The multiscaling factor is defined as follows:
#
!" $
In Figure 1 more coarse-grained molecules with
corresponding fine-grained particles and the parameters
and are represented.
%
∙ (
'
(5)
Using (4) in (5), it is obtained:
25
6
6
89
:7
∙
/
/
89
57
6 957
6
Because is function of
variable. Then
Fig. 1. The coordinates
and
<=
In AdResS, the space factor is function of the coordinate
of a coarse-grained particle ( ):
#
!" $
%
∙ (
'
(3)
where
is the coordinate in the reference system for the
center of fine-grained region, ) is half of the length of the
fine-grained region and * is the length of transition (hybrid
fine-grained - coarse-grained) region. In Figure 2 are
displayed these parameters.
6 89
57
6
6 9:7
6
/ 12
/
6 ;
6
∙
6 89
:7
6
2
6
6
∙
(6)
, it is necessary to change the
<=
<
for a coarse-grained particle
/
∙
<
∙
<
<
(7)
According to (1), it is obtained
<
<
(8)
Using (8) in (7), results
<=
<
∙ ′
where ′ is the derivative of (3) with respect
computed at the end of section (see formula (18)).
(9)
which is
Taking that into account
BC
<?@A
<
25 -.
(10)
25 .
(11)
C
<?@A
<
are the non-bonded, respectively bonded, forces in the atom
and
BC
<?DA
<
,
25 -.
(12)
25 .
(13)
C
<?DA
<
are the non-bonded, respectively bonded, force in the coarsegrained molecule , formula (6) becomes:
6
6
25
2
∙ ′∙
89
:7
∙
∙
′
89
57
2
6 9:7
2 59 /
6
By making the notation
∙
6
6
∙ 589 / 1 2
/
∙
6 ;
6
6 89
:7
6
The implementation of the new algorithm is based on MPI
parallelization (see Figure 3). MPI is used by dividing the
simulation box into several boxes. The number of processors
involved in computation is equal to the number of the smaller
simulation boxes. Data is passed between the neighboring
simulation boxes through the use of messages which is
constant. The MARTINI coarse-grained force field which is
applied on the particles of the system lowers the initial energy
of atoms until the reference temperature is reached and it is
maintained at that state [8].
The parallelization and the improved AdResS algorithm are
depicted in Figure 3.
∙
(14)
25
<?E
<
III. TECHNOLOGY APPROACH
The implementation of the improved AdResS is done in
GROMACS, a package specialized for running molecular
dynamics simulations. GROMACS was firstly developed at the
University of Groningen. Because the simulation time might
last for a long time, even in terms of months, the Message
Passing Interface (MPI) is used for parallelizing the
computations within the molecular system. The thermostat that
was used is stochastic dynamics.
(15)
it is obtained:
25
2
∙
′
∙
89
:7
∙
′
∙
89
57
2 59 2 59 ∙
2
∙ 589 2 1 2
2 5;
∙ 589 ∙
(16)
Then it results:
5
5
89
:7
2
/ 12
I ′∙
89
:7
∙
′
∙ 589 ∙
G2
/ 12
∙
89
57
∙
′
<=
K
'
%
∙ cos O
P
!
P
!
/ 59 ∙
89
57
/
/ 5;
%
# $
∙ Q ∙ sin O
'
∙
′
⇒
∙ 589 / 59 H /
with respect
!
#
∙ 589 / 59 /
∙ 589 / 59 J / 5;
Finally, the derivative of
formula (3):
<
∙
/
∙
(17)
is obtained from
P
!
%
# $
∙ Q∙
'
(18)
Fig. 3. The simulation box and its processes
Parallelization is done through MPI. The computation is
divided between the involved processors. Each processor will
follow the algorithm that starts by firstly computing the bonded
.
.
fine-grained and coarse-grained interactions +,
, 4,
. The next
step is the computation of the non-bounded fine-grained and
-.
-.
, 4,
. The multiscaling forces
coarse-grained interactions +,
are computed according to formula (17). The velocities and the
positions of the particles get after that updated.
IV.
FINDINGS
The simulation was done on an Asus ZenBook Pro
UX501VW computer having an Intel Core i7-6700HQ
processor with 8 threads and a frequency of 2.6 GHz. The
RAM has a capacity of 16 GB. The simulation box comprises
butane with 36900 atoms. The dimension of an atom is equal to
10 nm on each direction (x, y, z). The reference temperature is
set at 323 K. The system of atoms has been simulated on a
different number of processors for 10,000 simulation steps. For
each processor, a number of eight executions were made, each
of the reported values being an average of the eight execution
times.
The execution time expressed in ns/day and in hour/ns are
presented in Table 1.
TABLE I.
EXECUTION TIMES AND TEMPERATURE FOR A DIFFERENT
NUMBER OF PROCESSORS
Performance
Temperature
Test
No.
[ns/day]
[hour/ns]
[K]
1
7.421
3.234
326.487
2
15.326
1.566
326.497
3
20.146
1.191
326.440
4
26.127
0.919
326.481
5
19.328
1.242
326.543
6
28.193
0.851
326.498
7
23.697
1.013
326.525
8
35.591
0.674
326.466
In the table above the standard deviations are no larger than
0.036 for the execution time and 0.110 for the temperature.
Execution time [hour/ns]
A different number of processors was used to analyze how
parallelization evolves at run time for the butane system. The
execution time according to the number of processors is
presented in Fig. 4.
4
1.566 1.191
0
1
The reference temperature for the system was of 323 K.
The obtained value with a difference of around ± 3K is within
the statistical accepted errors for such simulations.
V. CONCLUSIONS
The article presents the improved AdResS through the MPI
parallelization implemented in the GROMACS molecular
dynamics software and its testing on a butane molecular
system, along with the experimental results. Usually the
simulation time is large. For this reason parallelization is used
for obtaining the results in a shorter period of time.
The algorithm depends on the hardware used and on the
number of processors on which it is run. There is a significant
difference in time for the simulation where only one processor
is involved as compared to the case when several processors
interact through the use of MPI - a speedup of about four times
more was obtained when six processors were used as compared
with the case of one processor. The temperature is kept within
the accepted statistical ranges for such simulations.
Further work include the testing of this implementation on
more systems, generalizing the theory for cylinder coordinates,
further improvements for AdResS algorithm.
3
2
3.234
1
Fig. 5. Temperature variation in time
2
3
0.919 1.242 0.851 1.013 0.674
4
5
6
7
REFERENCES
[1]
8
Processor No
Fig. 4. Execution time expressed in hour/ns as a function of the processor
number used
It can be observed that as the number of processors
increased, the execution time linearly decreased, after which a
plateau was obtained for the considered butane system with
36900 atoms.
There is a significant difference in time for the simulation
where only one processor is involved compared to the case
when several processors interact through the use of MPI - a
speedup of about four times more was obtained when six
processors were used as compared with the case of one
processor.
In Figure 5 is represented the variation of temperature in
time. By using the stochastic dynamics thermostat and the
improved AdResS, the temperature of the system is kept
around the value of 326 K as it can be observed in Table 1.
[2]
[3]
[4]
[5]
[6]
[7]
[8]
M. J. Abraham, D. Van Der Spoel, E. Lindahl, B. Hess and the Gromacs
Development Team, “GROMACS User Manual version 4.6.7”,
www.gromacs.org, p. 110, 2014.
N. Goga et al., "Benchmark of Schemes for Multiscale Molecular
Dynamics Simulations", J. Chem. Theory Comp., vol. 11, 2015, pp.
1389-1398.
M. Praprotnik, L. Delle Site, K. Kremer, "Adaptive resolution
molecular-dynamics simulation: Changing the degrees of freedom on the
fly", J. Chem. Phys. 2005.
M. Praprotnik, L. Delle Site, K. Kremer, "Multiscale simulation of soft
matter: From scale bridging to adaptive resolution", Annu. Rev. Phys.
Chem., vol. 59, 2008, pp.545-571.
M. Christen, W. F. van Gunsteren, "Multigraining: An algorithm for
simultaneous fine-grained and coarse-grained simulation of molecular
systems", J. Chem. Phys., vol. 124, 2006.
B. Ensing, S. O. Nielsen, P. B. Moore, M. L. Klein, M. J. Parrinello,
"Energy conservation in adaptive hybrid atomistic/coarse-grain
molecular dynamics", J. Chem. Theory Comp., vol. 3, 2007, pp. 11001105.
L. Delle Site, "What is a Multiscale Problem in Molecular Dynamics?",
Entropy, vol. 15, 2014, pp. 23-40.
S. J. Marrink, H.J. Risselada, S. Yefimov, D. P. Tieleman, A. H. de
Vries, “The MARTINI Force Field: Coarse Grained Model for
Biomolecular Simulations”, J. Phys. Chem., 111, pp. 7812-7824, 2007.
| 5cs.CE
|
Estimation under group actions: recovering orbits from invariants
arXiv:1712.10163v1 [math.ST] 29 Dec 2017
Afonso S. Bandeira∗1,2 , Ben Blum-Smith†1 , Amelia Perry‡3 , Jonathan Weed§3 , and
Alexander S. Wein¶k3
1
Department of Mathematics, Courant Institute of Mathematical Sciences, New York
University, New York NY, USA
2
Center for Data Science, New York University, New York NY, USA
3
Department of Mathematics, Massachusetts Institute of Technology, Cambridge MA, USA
January 1, 2018
Abstract
Motivated by geometric problems in signal processing, computer vision, and structural biology, we
study a class of orbit recovery problems where we observe noisy copies of an unknown signal, each acted
upon by a random element of some group (such as Z/p or SO(3)). The goal is to recover the orbit of the
signal under the group action. This generalizes problems of interest such as multi-reference alignment
(MRA) and the reconstruction problem in cryo-electron microscopy (cryo-EM). We obtain matching
lower and upper bounds on the sample complexity of these problems in high generality, showing that the
statistical difficulty is intricately determined by the invariant theory of the underlying symmetry group.
In particular, we determine that for cryo-EM with noise variance σ 2 and uniform viewing directions,
the number of samples required scales as σ 6 . We match this bound with a novel algorithm for ab initio
reconstruction in cryo-EM, based on invariant features of degree at most 3. We further discuss how to
recover multiple molecular structures from heterogeneous cryo-EM samples.
1
Introduction
Many computational problems throughout the sciences exhibit rich symmetry and geometry, especially in
fields such as signal and image processing, computer vision, and microscopy. This is exemplified in cryoelectron microscopy (cryo-EM) [ADLM84, SS11, Nog16], an imaging technique in structural biology that
was recently awarded the 2017 Nobel Prize in Chemistry. This technique seeks to estimate the structure
of a large biological molecule, such as a protein, from many noisy tomographic projections (2-dimensional
images) of the molecule from random unknown directions in 3-dimensional space.
In cryo-EM, our signal of interest is the density θ of the molecule, considered as an element of the
vector space of functions on R3 . We have access to observations of the following form: our microscopy
sample contains many rotated copies Ri θ of the molecule, where Ri ∈ SO(3) are random, unknown 3D
∗ E-mail address: [email protected]. Supported in part by NSF grant DMS-1712730 and NSF grant DMS-1719545.
Part of this work was done while ASB was visiting the Simons Institute for the Theory of Computing, this visit was partially
supported by the DIMACS/Simons Collaboration on Bridging Continuous and Discrete Optimization through NSF grant
#CCF-1740425.
† E-mail address: [email protected].
‡ E-mail address: [email protected]. Supported in part by NSF CAREER Award CCF-1453261 and a grant from the
MIT NEC Corporation.
§ E-mail address: [email protected]. Supported in part by NSF Graduate Research Fellowship DGE-1122374.
¶ E-mail address: [email protected].
k Authors are ordered alphabetically, please direct correspondence to [email protected].
1
rotations, and we observe the noisy projections Π(Ri θ) + ξi , where Π denotes tomographic projection (in a
fixed direction) and ξi is a large noise contribution, perhaps Gaussian. This specific problem motivates the
following general abstraction.
Fix a compact group G acting (by orthogonal transformations) on a vector space V . Throughout, the
vector space will be taken to be Rp and the group can be thought of as a subgroup of O(p), the orthogonal
group1 . Let θ ∈ V be the signal we want to estimate. We receive noisy measurements of its orbit as follows:
for i = 1, . . . , n we observe a sample of the form
yi = gi · θ + ξi
where gi is drawn randomly (in Haar measure2 ) from G and ξi ∼ N (0, σ 2 I) is noise. The goal is to recover
the orbit of θ under the action of G. We refer to this task as the orbit recovery problem.
This abstraction, already a rich object of study, neglects the tomographic projection in cryo-EM; we
will also study a generalization of the problem which allows such a projection. We will also consider the
additional extension of heterogeneity [Jon16, LS16, LS17, BBLS17], where mixtures of signals are allowed:
we have K signals θ1 , . . . , θK , and each sample yi = gi · θki + ξi comes from a random choice 1 ≤ ki ≤ K
of which signal is observed. This extension is of paramount importance for cryo-EM in practice, since the
laboratory samples often contain one protein in multiple conformations, and understanding the range of
conformations is key to understanding the function of the protein.
1.1
Prior work
Several special cases of the orbit recovery problem have been studied for their theoretical and practical
interest. Besides cryo-EM, another such problem is multi-reference alignment (MRA) [BCSZ14, BRW17,
PWB+ 17], a problem from signal processing [ZvdHGG03, PZAF05] with further relevance to structural
biology [Dia92, TS12]. In this problem, one observes noisy copies of a signal θ ∈ Rp , each with its coordinates
permuted by a random cyclic shift. This is an example of the orbit recovery problem when G is taken to be
the cyclic group Z/p acting by cyclic permutations of the coordinates. Since the cyclic group Z/p is simpler
than SO(3), understanding MRA has been seen as a useful stepping stone towards a full statistical analysis
of cryo-EM.
Many prior methods for orbit recovery problems employ the so-called synchronization approach where the
unknown group elements gi are estimated based on pairwise comparison of the samples yi . If the samples
were noiseless, one would have gi gj−1 yj = yi ; thus noisy samples still give some weak information about
gi gj−1 . Synchronization is the problem of using such pairwise information to recover all the group elements
gi (up to a global right-multiplication by some group element). Once the group elements gi are known, the
underlying signal can often be easily recovered.
The synchronization approach to cryo-EM can be summarized as follows [VG86, VH87, SS11]. By the
Fourier projection–slice theorem, the Fourier transforms of the tomographic projections are 2D slices of the
Fourier transform of the molecule density. Given a hypothesis as to the angles of two slices, we can predict a
1D line of intersection along which those slices should agree. By measuring correlation along that common
line, we obtain some weak information by which to confirm or refute our hypothesized angles. Indeed, this
test only depends on the relative angle of the slices, thus providing weak information about the value of
gi gj−1 . We then use a synchronization algorithm to recover the gi using this information.
Methods to solve the synchronization problem over various groups include spectral methods [Sin11, SS11],
semidefinite programming [Sin11, SS11, BCSZ14, BCS15, Ban15, BZS15], and approximate message passing
(AMP) [PWBM16a]. A general Gaussian model for synchronization problems over any compact group is
studied in [PWBM16b, PWBM16a]; ideas from statistical physics suggest that for this model, AMP is
1 We alert the reader to the fact that we will use O(p) to refer to the group of orthogonal matrices in dimension p and
O(g(n)) as the standard big-O notation: f (n) = O(g(n)) if and only if there exists a constant C such that f (n) ≤ Cg(n) for all
n sufficiently large. It will be clear from context which one is meant.
2 We note that any distribution of g can be reduced to Haar by left multiplying y by a Haar-distributed group element.
i
i
However, as illustrated in [ABL+ 17], it is sometimes possible to exploit deviations from Haar measure.
2
optimal among all efficient algorithms [PWBM16a]. However, the model does not capture the underlying
signal θ and instead assumes that for every pair i, j of samples, independent noisy measurements of the
relative group elements gi gj−1 are observed. This independence does not correctly capture problems like
MRA and cryo-EM, which have independent noise on each sample, rather than on each pairwise comparison
of samples.
The synchronization approach has proven to be effective both in theory and practice when the noise is
sufficiently small. However, once the noise level is large, no consistent estimation of the group elements
gi is possible [ADBS16]. Moreover, it is the high-noise regime that is the practically relevant one for
many applications, including cryo-EM, where the presence of large noise is a primary obstruction to current
techniques [Sig16]. As a result, recent work has focused on approaches to cryo-EM and MRA which provably
succeed even in the large-noise limit. One striking finding of this line of work is that the sample complexity
of the statistical estimation problem increases drastically as the noise level increases. For instance, for the
multi-reference alignment problem with noise variance σ 2 , consistent estimation of typical signals requires
Ω(σ 6 ) samples [BRW17, APS17], with significantly worse rates for atypical signals. By contrast, when σ 2
is smaller than some threshold, only O(σ 2 ) samples are required. Moreover, in contrast with the O(σ 2 )
rate—which would hold even in the absence of a group action—the Ω(σ 6 ) bound obtained in previous works
depends on particular properties of the cyclic group. In this work, we significantly extend this prior work by
determining the sample complexity of the estimation problem in the high-noise regime for general groups.
The leading theoretical framework for the high-noise regime is the invariant features approach [BRW17,
BBM+ 17, PWB+ 17, BBLS17, LBB+ 17]. This approach has a long history in the signal processing literature [Kam80, Sad89, SG92] and is analogous to the well known “method of moments” in statistics [vdV98].
In brief, the invariant features approach bypasses entirely the problem of estimating the group elements and
focuses instead on estimating features of the signal which are preserved by the action of the group. So long as
these invariant features uniquely specify the orbit of the original signal, the invariants are sufficient statistics
for the problem of recovering the orbit of the original signal. This simple approach yields optimal dependence
of the sample complexity on the noise level for the multi-reference alignment problem [BRW17, PWB+ 17].
The application of invariant features to cryo-EM dates back to 1980 with the work of Kam [Kam80], who
partially solved cryo-EM by means of degree-2 invariant features, reducing the unknown molecule structure
to a collection of unknown orthogonal matrices. Subsequent work has explored methods to estimate these
orthogonal matrices [BZS15], including recent work showing how two noiseless tomographic projections
suffice to recover these orthogonal matrices [LBB+ 17]. Our work can be viewed as a degree-3 extension of
Kam’s method that fully solves cryo-EM while circumventing the orthogonal retrieval issue, and without
requiring any noiseless projections. Our approach is ab initio, i.e. it does not require an initial guess of what
the molecule looks like and thus cannot suffer from model bias, which is a documented phenomenon [Coh13]
where the initial guess can have a significant effect on the result. Ab initio estimates are particularly useful
to serve as a model-free starting point for popular iterative refinement algorithms such as RELION [Sch12].
Throughout, we focus on the case where the group elements are Haar-distributed. In the basic orbit
recovery problem (projection), any distribution of gi can be reduced to Haar by left-multiplying each sample
yi by a Haar-distributed group element. However, as illustrated in [ABL+ 17], it is sometimes possible to
exploit deviations from Haar measure. The situation is different when we add projection to the problem
setup, as is the case with Cryo-EM; if the viewing direction is not distributed uniformly then there may exist
parts of the molecule that are systematically imaged less than others, which can cause serious difficulties in
reconstruction.
The present paper connects the orbit recovery problem to the invariant theory of groups, a classical
and well-developed branch of algebra (see for example [Kač94, Dol03, Stu08, DK15]). Invariant theory’s
traditional goal is to describe the ring of all polynomial functions on a vector space that are invariant under
the action of a group – the invariant algebra. Since the 19th century, culminating in the pioneering work of
David Hilbert [Hil90, Hil93], it has been known that the invariant algebra is finitely generated in many cases
of interest, and so a fundamental problem has been to bound the degrees of the generators. In 2002, Derksen
and Kemper [DK15] introduced the notion of a separating algebra – a subring of the invariant algebra that
separates all orbits of the group action which are separated by the full invariant algebra. Our connection
3
to orbit recovery motivates the question of bounding the degree required to generate a separating algebra
(see Section 4.4), a problem which has been recently studied [KK10, Dom16]. Our work also motivates the
question of bounding the degree at which the field of invariant rational functions is generated as a field (see
Section 4.3), which does not appear to have been the focus of research attention before.
1.2
Our contributions
In this paper we extend the results of [BRW17] and show that the method of moments yields optimal sample
complexity for orbit recovery problems over any compact group. Specifically, we show that optimal sample
complexity is achieved by an algorithm that estimates the moments from the samples and then solves a
polynomial system of equations in order to find a signal θ that would produce such moments. As the sample
complexity depends on the number of moments used, this gives rise to the algebraic question of how many
moments suffice to determine the orbit of θ. Using tools from invariant theory and algebraic geometry, we
investigate this question for various success criteria and obtain sharp results in a number of settings. Our
main focus is on the case where the signal is assumed to be generic and the goal is to output a finite list
of signals, one of which is the truth. In this case we give a simple efficient algorithm for determining the
number of moments required for any given orbit recovery problem. The main step of the algorithm is to
compute the rank of a particular Jacobian matrix.
We note that ours is an information-theoretic result rather than a computational one because even with
knowledge of the number of moments required, estimating the original signal still requires solving a particular
polynomial system of equations and we do not attempt to give a computationally-efficient method for this.
There are fast non-convex heuristic methods to solve these systems in practice [BBLS17] but we leave for
future work the question of analyzing such methods rigorously and exploring whether or not they reach the
information-theoretic limits determined in this paper. For the case of finite groups, another efficient method
for solving the polynomial system is via tensor decomposition, which has been analyzed for MRA [PWB+ 17].
Concrete results for problems such as MRA and cryo-EM are in Section 5.
1.3
Motivating examples
In addition to the examples of MRA and cryo-EM, it is helpful to have the following motivating examples
in mind:
1. Learning a “bag of numbers”: let G be the symmetric group Sp , acting on V = Rp by permutation
matrices. Thus we observe random rearrangements of the entries of a vector, plus noise.
2. Learning a rigid body: let G be the rotation group SO(p), acting on the matrix space V = Rp×m by
left-multiplication. We imagine the columns of our matrix as vertices defining a rigid body; thus we
observe random rotations of this rigid body (with vertices labeled) plus noise.
3. S 2 registration: Let S 2 ⊆ R3 be the unit sphere. Let V be the finite-dimensional vector space of
functions on S 2 → R that are band-limited, i.e. linear combinations of spherical harmonics up to some
fixed degree (spherical harmonics are the appropriate “Fourier basis” for functions on the sphere); let
θ ∈ V be such a function S 2 → R. Let G = SO(3), acting on the sphere by 3-dimensional rotation;
this induces an action on V via (g · θ)(x) = θ(g −1 · x). Thus we observe many noisy copies of a fixed
function on the sphere, each rotated randomly.
1.4
Problem statement
Throughout, we consider a compact (topological) group G acting linearly, continuously, and orthogonally
on a finite-dimensional real vector space V = Rp . In other words, G acts on V via a linear representation
ρ : G → O(V ), and ρ itself is a continuous function. Here O(V ) denotes the space of real orthogonal p × p
matrices. Let Haar(G) denote Haar measure (i.e., the “uniform distribution”) on G. We define the orbit
recovery problem as follows.
4
Problem 1.1 (orbit recovery). Let V = Rp and let θ ∈ V be the unknown signal. Let G be a compact group
that acts linearly, continuously, and orthogonally on V . For i ∈ [n] = {1, 2, . . . , n} we observe
yi = gi · θ + ξi
where gi ∼ Haar(G) and ξi ∼ N (0, σ 2 Ip×p ), all independently. The goal is to estimate θ. Note that we can
only hope to recover θ up to action by G; thus we aim to recover the orbit {g · θ : g ∈ G} of θ.
In practical applications, σ is often known in advance and, when it is not, it can generally be estimated
accurately on the basis of the samples. We therefore assume throughout that σ is known and do not pursue
the question of its estimation in this work.
Our primary goal is to study the sample complexity of the problem: how must the number of samples n
scale with the noise level σ (as σ → ∞ with G and V fixed) in order for orbit recovery to be statistically
possible? All of our results will furthermore apply to a generalized orbit recovery problem (Problem 2.3)
allowing for projection and heterogeneity (see Section 1.6).
Our work reveals that it is natural to consider several different settings in which to state the orbit recovery
problem. We consider the following two decisions:
1. Do we assume that θ is a generic signal, or do we allow for a worst-case signal? (Here generic means
that there is a measure-zero set of disallowed signals.)
2. Do we want to output a θ′ such that θ′ (approximately) lies in the orbit of θ (unique recovery), or
simply a finite list θ1 , . . . , θs of candidates such that one of them (approximately) lies in the orbit of θ
(list recovery)?
The terminology “list recovery” is borrowed from the idea of list decoding in the theory of error-correcting
codes [Eli57]. By taking all combinations of the two options above, there are four different recovery criteria.
Strikingly, these different recovery criteria can be very different in terms of sample complexity, as the following
examples show (see Section 5 for more details):
1. Multi-reference alignment (MRA): Recall that this is the case G = Z/p acting on V = Rp by cyclic
shifts. It is known [PWB+ 17] that if θ is generic then unique recovery is possible with O(σ 6 ) samples.
However, for a worst-case θ, many more samples are required (even for list recovery); as shown in
[BRW17], there are some very particular infinite families of signals that cannot be distinguished without
Ω(σ 2p ) samples. This illustrates a large gap in difficulty between the generic and worst-case problems.
2. Learning a rigid body: Let G be the rotation group SO(p) acting on the matrix space Rp×m by left
multiplication. We imagine the columns of our matrix as vertices defining a rigid body; thus we observe
random rotations of this rigid body (with vertices labeled) plus noise. With O(σ 4 ) samples it is possible
to recover the rigid body up to reflection, so that list recovery (with a list of size 2) is possible. However,
unique recovery (even for a generic signal) requires drastically more samples: Ω(σ 2p ).
We will address all four recovery criteria but our main focus will be on the case of generic list recovery,
as it is algebraically the most tractable to analyze. For the following reasons we also argue that it is perhaps
the most practically relevant case. Clearly real-world signals are generic. Also, unique recovery is actually
impossible in some practical applications; for instance, in cryo-EM it is impossible to determine the chirality
of the molecule. Furthermore, one could hope to use application-specific clues to pick the true signal out
from a finite list; for instance, in cryo-EM we might hope that the spurious solutions in our finite list do not
look like “reasonable” molecules and can be thrown out.
1.5
Method of moments
Our techniques rely on estimation of the following moments:
Definition 1.2 (moment tensor). The order-d moment tensor is Td (θ) := Eg [(g · θ)⊗d ] where g ∼ Haar(G).
5
Pn
We can estimate Td (θ) from the samples by computing n1 i=1 yi⊗d plus a correction term to cancel bias
from the noise terms (see Section 7 for details). The moments Td (θ) are related to polynomials that are
invariant under the group action, which brings us to the fundamental object in invariant theory:
Definition 1.3 (invariant ring). Let x = (x1 , . . . , xp ) be a set of coordinate functions on V = Rp , i.e. a basis
for the dual V ∗ , so that R[x] := R[x1 , . . . , xp ] is the ring of polynomial functions V → R. We have an action
of G on R[x] given by (g · f )(·) = f (g −1 (·)). (If we fix a basis for V , we can think of x as indeterminate
variables corresponding to the entries of θ ∈ V .) The invariant ring R[x]G ⊆ R[x] is the ring consisting of
polynomials f that satisfy g · f = f for all g ∈ G. An element of the invariant ring is called an invariant
polynomial (or simply an invariant ). Invariant polynomials can be equivalently characterized as polynomials
of the form Eg [g · f ] where f ∈ R[x] is any polynomial and g ∼ Haar(G).
The two objects above are equivalent in the following sense. The moment tensor Td (θ) contains the same
information as the set of evaluations f (θ) for all f ∈ R[x]G that are homogeneous of degree d. In particular,
for any such polynomial f , f (θ) is a linear combination of the entries of Td (θ).
The following algebraic question will be of central importance: when do the values of invariant polynomials
(of degree ≤ d) of θ determine the orbit of θ (in the appropriate sense)? As we see below, the sample
complexity of the statistical problem is completely characterized by the answer to this question.
1.5.1
Warm up: hypothesis testing
Consider for now the simple problem of distinguishing between two fixed hypotheses θ = τ1 and θ = τ2 ,
where τ1 and τ2 are two fixed vectors in V . One method is to find an invariant polynomial f for which
f (τ1 ) 6= f (τ2 ) and to estimate f (θ) using the samples. The sample complexity of this procedure depends
on the degree of f because if f has degree d, we need O(σ 2d ) samples to accurately estimate f (θ). We will
prove the following (see Section 3).
Theorem 1.4 (distinguishing upper bound). Fix τ1 , τ2 ∈ V . If there exists a degree-d invariant polynomial
f ∈ R[x]G with f (τ1 ) 6= f (τ2 ) then, using O(σ 2d ) samples, it is possible to distinguish between θ = τ1 and
θ = τ2 with type-I and type-II error probabilities each at most 1/3.
Here, O(· · · ) hides factors that depend on G (and its action on V ), τ1 , and τ2 , but not σ; we are most
interested in how the sample complexity scales as σ becomes large (with everything else held fixed). The
error probability 1/3 is arbitrary and can be boosted by taking more samples (see Theorem 3.2).
Furthermore, we have a matching lower bound to show that the method of moments is optimal: the
sample complexity is driven by the minimum degree of an invariant polynomial that separates τ1 and τ2 .
Theorem 1.5 (distinguishing lower bound). Fix τ1 , τ2 ∈ V . Let d∗ be the smallest positive integer d for
∗
which Td (τ1 ) 6= Td (τ2 ). Then Ω(σ 2d ) samples are required to distinguish between θ = τ1 and θ = τ2 with
type-I and type-II error probabilities each at most 1/3.
See Section 3 for more details.
1.5.2
Recovery
We now address the problem of recovering the signal θ from the samples. Our goal is to recover the orbit of
θ, defined as follows.
G
G
Definition 1.6. For θ1 , θ2 ∈ V , define an equivalence relation ∼ by letting θ1 ∼ θ2 if there exists g ∈ G
G
such that g · θ1 = θ2 . The orbit of θ (under the action of G) is the equivalence class of θ under ∼, i.e. the
set {g · θ : g ∈ G}. Denote by V /G the set of orbits of V , that is, the equivalence classes of V modulo the
G
relation ∼.
We need the following definitions to capture the notion of approximately recovering the orbit of θ.
6
Definition 1.7. For θ1 , θ2 ∈ V , let
dG (θ1 , θ2 ) = min kθ1 − g · θ2 k2 .
g∈G
This pseudometric induces a metric on the quotient space V /G in the obvious way, so we can write dG (o1 , o2 )
for o1 , o2 ∈ V /G. By slight abuse of notation, we write dG (θ1 , o2 ) for dG (o1 , o2 ), where o1 is the orbit of θ1 .
Theorem 1.5 already shows that if the orbit of θ is not determined by knowledge of the first d − 1 moment
tensors, then at least Ω(σ 2d ) samples are required to recover the orbit of θ. We are now ready to (informally)
state our main result on recovery (see Section 3 for more details), which provides a matching upper bound.
Theorem 1.8 (recovery upper bound, informal). Fix θ ∈ V . If the moments T1 (θ), · · · , Td (θ) uniquely
b ≤ε
determine the orbit of θ, then using O(σ 2d ) samples, we can produce an estimator θb such that dG (θ, θ)
with high probability.
The analogous result holds for list recovery (see Section 3): if the moments determine a finite number of
possibilities for the orbit of θ then we can output a finite list of estimators, one of which is close to the orbit
of θ.
Thus, we have reduced to the algebraic question of determining how many moments are necessary to
determine the orbit of θ (either uniquely or in the sense of list recovery). In Section 4 we will use tools from
invariant theory and algebraic geometry in order to address these questions.
1.6
Extensions: projection and heterogeneity
We now consider some extensions to the basic orbit recovery problem (Problem 1.1), motivated by the
application of cryo-EM:
1. Projection: In cryo-EM, we do not observe a noisy 3-dimensional model of the rotated molecule;
we only observe a 2-dimensional projection of it. We will model this projection by a linear map
Π : Rp → Rq that maps a 3-dimensional model to its 2-dimensional projection (from a fixed viewing
direction). The samples are then given by yi = Π(gi · θ) + ξi where ξi ∼ N (0, σ 2 I).
2. Heterogeneity: In cryo-EM we observe images of many different copies of the same molecule, each
rotated differently. However, if our sample is not pure, we may have a mixture of different molecules
and want to recover the structure of all of them. We will model this by taking K different unknown
signals θ1 , . . . , θK along with positive mixing weights w1 , . . . , wK which sum to 1. Each sample takes
the form yi = gi · θki + ξi where ki is chosen at random according to the mixing weights.
In Section 2 we will formally define a generalization of the orbit recovery problem that allows for either (or
both) of the above extensions. All of our methods will apply to this general case.
1.7
Organization of the remainder of the paper
In Section 2, we define a generalization of Problem 1.1 which encompasses projection and heterogeneity,
and specify the basic algebraic objects which relate to our generalized problem. In Section 3, we formally
state statistical upper and lower bounds for the generalized orbit recovery problem in terms of invariants. In
Section 4, we establish our basic algebraic results and specify the algebraic criteria that correspond to the
different recovery criteria defined in Section 1.4. We also give an efficient algorithm to decide the algebraic
criterion corresponding to generic list recovery. Finally, in Section 5, we apply our work to several examples
of the orbit recovery problem, including MRA and cryo-EM. We conclude in Section 6 with questions for
future work.
Sections 7–9 contain proofs of results from preceding sections. Appendix A contains an account of the
invariant theory of SO(3), and Appendix B contains a proof (using Galois theory) of generic unique recovery
in a particular special case of the orbit recovery problem (the regular representation of a finite Abelian
group).
7
2
General problem statement
Our results will apply not only to the basic orbit recovery problem (Problem 1.1) but to a generalization
(Problem 2.3 below) that captures the projection and heterogeneity extensions discussed in Section 1.6. We
first define mixing weights for heterogeneous problems.
PK
Definition 2.1 (mixing weights). Let w = (w1 , . . . , wK ) ∈ ∆K := {(z1 , . . . , zK ) : zk ≥ 0 ∀k, k=1 zk = 1}.
w
Let k ∼ [K] indicate that k is sampled from [K] = {1, . . . , K} such that k = ℓ with probability wℓ . We
will sometimes instead parametrize the mixing weights by w k = wk − 1/K so that w lies in the vector space
PK
∆ := {(z1 , . . . , zK ) :
k=1 zk = 0}.
In a heterogeneous problem with K different signals, we can only hope to recover the signals up to permutation. To formalize this, our compound signal will lie in a larger vector space V and we will seek to recover
its orbit under a larger group G.
Definition 2.2 (setup for heterogeneity). Let G̃ be a compact group acting linearly, continuously, and
orthogonally on Ṽ = Rp . Let V = Ṽ ⊕K ⊕ ∆K , so that θ ∈ V encodes K different signals along with mixing
weights: θ = (θ1 , . . . , θK , w). We let an element (g1 , . . . , gK , π) of the Cartesian product set G̃K × SK act on
V as follows: first, each gk acts on the corresponding θk , and then π permutes the θk and the coordinates of
w. Note that this action is linear and orthogonal (where ∆ uses the usual inner product inherited from RK ).
There is a natural group structure G on the set G̃K × SK such that the action just described is actually a
group action by G: the semidirect product G = G̃K ⋊ϕ SK , where ϕ denotes the action of SK on G̃K by
permutations of the factors. This is also called the wreath product of G̃ by SK and written G̃ ≀ SK . The
product topology on G̃K × SK makes G a topological group; it is compact with respect to this topology since
all the factors are compact, and the action described above is continuous.
Of course, by taking K = 1 we recover the basic setup (without heterogeneity) as a special case. We are
now ready to give the general problem statement.
Problem 2.3 (generalized orbit recovery). Let Ṽ = Rp and W = Rq . Let G̃ be a compact group acting
linearly, continuously, and orthogonally on Ṽ . Let Π : Ṽ → W be a linear map. Let θ = (θ1 , . . . , θK , w) ∈
V := Ṽ ⊕K ⊕ ∆K be an unknown collection of K signals with mixing weights w ∈ ∆K . For i ∈ [n] =
{1, 2, . . . , n} we observe
yi = Π(gi · θki ) + ξi
w
where gi ∼ Haar(G̃), ki ∼ [K], ξi ∼ N (0, σ 2 Iq×q ), all independently. The goal is to estimate the orbit of θ
under G := G̃K ⋊ SK .
Note that this serves as a reduction from the heterogeneous setup to the basic setup in the sense that we are
still only concerned with recovering the orbit of a vector θ under the action of some compact group.
As discussed previously, we apply the method of moments. The moments are now defined as follows.
Definition 2.4 (moment tensor). For the generalized orbit recovery problem (Problem 2.3), the order-d
w
moment tensor is Td (θ) := Eg,k [(Π(g · θk ))⊗d ] where g ∼ Haar(G̃) and k ∼ [K]. Equivalently, Td (θ) =
PK
⊗d
].
k=1 wk Eg [(Π(g · θk ))
The invariant ring is defined as in Definition 1.3 but now for the larger group G acting on the larger V :
Definition 2.5 (invariant ring). Note that dim(V ) = Kp + K − 1 and let x = (x1 , . . . , xdim(V ) ) be a basis
for V ∗ ; here the last K − 1 variables correspond to ∆, e.g. they can correspond to w 1 , . . . , w K−1 . We then
let R[x]G ⊆ R[x] be the polynomials in x that are invariant under the action of G (as in Definition 1.3).
Recall that in the basic orbit recovery problem, Td (θ) corresponds precisely to the homogeneous invariant
polynomials of degree d; now Td (θ) corresponds to a subspace of the homogeneous invariant polynomials of
degree d. Specifically, the method of moments gives us access to the following polynomials (evaluated at θ):
8
Definition 2.6. Let UdT be the subspace (over R) of the invariant ring R[x]G consisting of all R-linear
T
combinations of entries of Td (x). Let U≤d
= U1T ⊕ · · · ⊕ UdT ⊆ R[x]G . Here we write Td (x) for the collection
of polynomials (one for each entry of Td (θ)) that map θ to Td (θ).
T
We will be interested in whether the subspace U≤d
contains enough information to uniquely determine
the orbit of θ (or determine a finite list of possible orbits) in the following sense.
Definition 2.7. A subspace U ⊆ R[x]G resolves θ ∈ V if there exists a unique o ∈ V /G such that f (θ) = f (o)
for all f ∈ U . Similarly, U list-resolves θ if there are only finitely many orbits o1 , . . . , os such that f (θ) = f (oi )
for all f ∈ U .
Here we have abused notation by writing f (o) to denote the (constant) value that f takes on every θ ∈ o.
The following question is of central importance.
T
Question 2.8. Fix θ ∈ V . How large must d be in order for U≤d
to uniquely resolve θ? How large must d
T
be in order for U≤d to list-resolve θ?
The answer depends on G and V but also on whether θ is a generic or worst-case signal, and whether
we ask for unique recovery or list recovery. Our statistical results in Section 3 will show that the sample complexity of the generalized orbit recovery problem is Θ(σ 2d ) where d is the minimal d from Question 2.8. More specifically, the recovery procedure that obtains this bound is based on estimating the
moments T1 (θ), . . . , Td (θ) and solving a system of polynomial equations to (approximately) recover θ. Our
algebraic results in Section 4 will give general methods to answer Question 2.8 for any G and V .
3
Statistical results
In this section, we state upper and lower bounds on the performance of optimal estimators for the orbit
recovery problem. Proofs are deferred to Section 7. Our approach will be the method of moments introduced
in Section 1.5. We assume for normalization purposes that there exists a constant c ≥ 1 such that c−1 ≤
kθk ≤ c, so that σ captures entirely the signal-to-noise ratio of the problem. We denote by Θ the subset of
V consisting of vectors satisfying this requirement.
Denote by Pθ the distribution of a sample arising from the generalized orbit recovery problem (Problem 2.3) with parameter θ.
Definition 3.1. Given θ ∈ Θ, the order-d matching set for θ, Mθ,d, is the set consisting of all τ ∈ V such
T
.
that f (τ ) = f (θ) for all f ∈ U≤d
T
T
list-resolves θ when
resolves θ exactly when Mθ,d contains a single orbit, and U≤d
We note that U≤d
Mθ,d is the union of a finite number of orbits.
We are now ready to state a formal theorem justifying Theorems 1.4 and 1.8, above. The following
theorem establishes that we can approximately learn the order-d matching set for θ with probability at least
1 − δ on the basis of O(σ 2d log(1/δ)) samples. Denote by Mεθ,d the ε-fattening of Mθ,d, i.e., the set of all
φ ∈ Θ such that minτ ∈Mθ,d kφ − τ k ≤ ε.
Theorem 3.2. For any positive integer n, noise level σ ≥ maxθ∈Θ kθk, and accuracy parameter δ > 0, there
cn = M
cn (y1 , . . . , yn ) ⊆ V such that, for any positive constant ε, if y1 , . . . , yn ∼ Pθ i.i.d.
exists an estimator M
and n ≥ cθ,ε,d log(1/δ)σ 2d , then with probability at least 1 − δ,
cn ⊆ Mεθ,d .
Mθ,d ⊆ M
The constant cθ,ε,d in Theorem 3.2 can be replaced by cθ,d ε−2 in the unique recovery setting if θ is
suitably generic, but the dependence on ε can be worse in general. What is key is that cθ,ε,d does not depend
on σ, so that Theorem 3.2 captures the behavior of the sample complexity in the large σ limit.
9
Theorem 3.2 essentially follows from the observation that, since the variance of yi is O(σ 2 ), a degree-d
T
polynomial in the entries of yi has variance O(σ 2d ). This implies that for any f ∈ U≤d
, the evaluation f (θ)
2d
can be accurately estimated on the basis of O(σ ) samples. By inverting a suitable polynomial system, we
can thereby identify Mθ,d , at least approximately. A full proof of Theorem 3.2 appears in Section 7.
Theorem 3.2 captures the behavior described in both Theorem 1.4 and Theorem 1.8. Indeed, if τ1 and τ2
differ on some f ∈ UdT , then Mτ1 ,d is separated from Mτ2 ,d , and Theorem 3.2 implies that we can therefore
distinguish between the two distributions when ε is sufficiently small. Moreover, as the following corollary
cn allows us to recover or list-recover the orbit of θ.
shows, Theorem 3.2 implies that the confidence set M
Corollary 3.3. Suppose that Mθ,d is the union of M orbits o1 , . . . , oM , where M is finite. There exists an
εθ such that, for ε < εθ , if n ≥ cθ,ε,d log(1/δ)σ 2d , then on the basis of n i.i.d. samples from Pθ we can produce
M estimators θb1 , . . . θbM such that, with probability at least 1 − δ, there exists a permutation π : [M ] → [M ]
satisfying
dG (θbi , oπ(i) ) ≤ ε
for all i ∈ [N ].
Proof. Since G is a compact group acting continuously on V , the orbits are compact. By assumption Mθ,d ,
is a union of a finite number of orbits, so there exists an εθ such that dG (oi , oj ) ≥ 4εθ for any i 6= j. For
cn as in Theorem 3.2. Theorem 3.2 implies the
any ε < εθ , let N be an ε/2-net of V /G, and construct M
existence of a constant cθ,ε,d such that as long as n ≥ cθ,ε,d log(1/δ)σ 2d , with probability at least 1 − δ,
cn ⊆ Mε/2 .
Mθ,d ⊆ M
θ,d
cn ) ≤ ε/2. With probability 1 − δ, any element
Consider the set C consisting of o ∈ N such that dG (o, M
of C is within ε of oi for some i ∈ [M ], and for each oi there exists an o ∈ C that is at most ε away. By
assumption, distinct orbits in Mθ,d are separated by more than 4ε, so if any two elements of C are separated
by at most 2ε, then they are close to the same element of Mθ,d . As a result, it is possible to partition C
into M sets C1 , . . . , CM such that dG (o, o′ ) ≤ 2ε if o, o′ are in the same set, and dG (o, o′ ) > 2ε if o and o′
are in different sets. For i ∈ [M ], let θbi be any element of V such that the orbit of θbi lies in Ci . The claim
follows.
The constant εθ in the statement of Corollary 3.3 will not be known in general. However, a weaker
cn under the projection V 7→ V /G.
statement still holds when ε ≥ εθ . Indeed, consider the image of the set M
There exists a finite partition of the resulting set such that any two orbits in the same cluster are closer
than any two orbits in different clusters. (Note that the clustering into a single set always satisfies this
requirement.) If we are able to choose this partition such that the diameter of each set is at most ε′ , then
by choosing a representative from each cluster, we obtain a finite set of estimators, at least one of which is
guaranteed to be ε′ -close to θ with high probability. Corollary 3.3 implies that this partition can be taken
to consist of at most M clusters for ε′ arbitrarily small, as long as n ≥ Cε′ σ 2d .
We now prove a lower bound showing that the dependence on σ in Theorem 3.2 is tight. We show that
T
if U≤d−1
fails to resolve (or list-resolve) θ, then Ω(σ 2d ) samples are necessary to recover (or list-recover) the
orbit of θ. Together with Theorem 3.2, this lower bound implies that if d∗ is the smallest positive integer
∗
for which UdT∗ resolves (or list-resolve) θ, then Θ(σ 2d ) samples are required to recover (or list-recover) the
orbit of θ. We make this lower bound precise in Theorem 3.4.
Theorem 3.4. For any positive integer d, there exists a constant cd such that if τ1 and τ2 are elements
in Mθ,d−1 lying in different orbits, then no procedure can distinguish between Pnτ1 and Pnτ2 with probability
greater than 2/3 if n ≤ cd σ 2d .
Note that via Le Cam’s method [LeC73], Theorem 3.4 translates into lower bounds for the problem of
recovering θ. The proof of Theorem 3.4 relies on a tight bound for the Kullbeck-Leibler divergence between
the distributions Po1 and Po2 established in [BRW17]. More details appear in Section 7.
10
4
Algebraic results
In this section, we will consider the four recovery criteria defined in Section 1.4, and give algebraic characterizations of each case. The results of Section 3 imply that it suffices to focus our attention on deciding
when a subspace U resolves (or list-resolves) a parameter θ. We show below how to answer this question
by purely algebraic means. Moreover, for generic list recovery, we show how this question can be answered
algorithmically in polynomial time. For generic and worst-case unique recovery, we also give algorithms to
decide the corresponding algebraic condition; however, these algorithms are not efficient.
Throughout, we assume the setup defined in Section 2 for the generalized orbit recovery problem. In
particular, G is a compact group acting linearly and continuously on a finite-dimensional real vector space
V (although we do not require in this section that the action be orthogonal). We have the invariant ring
T
R[x]G corresponding to the action of G on V , and a subspace U ⊆ R[x]G (e.g. U≤d
) of invariants that we
have access to. We are interested in whether the values f (θ) for f ∈ U determine the orbit of θ ∈ V under
T
G. The specific structure of G and U≤d
(as defined in Section 2) will be largely unimportant and can be
abstracted away.
4.1
Invariant theory basics
We will often need the following basic operator that averages a polynomial over the group G.
Definition 4.1 (Reynolds operator). The Reynolds operator R : R[x] → R[x]G is defined by
R(f ) =
E
[g · f ].
g∼Haar(G)
Note that the Reynolds operator is a linear projection from R[x] to R[x]G that preserves the degree of
homogeneous polynomials (i.e. a homogeneous polynomial of degree d gets mapped either to a homogeneous
polynomial of degree d or to zero).
Observation 4.2. Let R[x]G
d denote the vector space consisting of homogeneous invariants of degree d. We
can obtain a basis for R[x]G
d by applying R to each monomial in R[x] of degree d. (This yields a spanning
set which can be pruned to a basis if desired.)
In our setting, we have the following basic fact from invariant theory.
Theorem 4.3 (e.g. [Kač94] Theorem 4.1-3). The invariant ring R[x]G is finitely generated as an R-algebra.
In other words, there exist generators f1 , . . . , fm ∈ R[x]G such that R[f1 , . . . , fm ] = R[x]G .
Furthermore, there is an algorithm to find a generating set; see Section 8.1. Another basic fact from invariant
theory implies that the entire invariant ring is sufficient to determine the orbit of θ. (This is not always true
for non-compact groups; see Example 2.3.1 in [DK15].)
Theorem 4.4 ([Kač94] Theorem 6-2.2). The full invariant ring R[x]G resolves every θ ∈ V .
Proof. Let o1 , o2 ∈ V /G be distinct (and therefore disjoint) orbits. Since G is compact and acts continuously,
o1 and o2 are compact subsets of V . Thus by Urysohn’s lemma there exists a continuous function f˜ : V → R
such that f˜(τ ) = 0 ∀τ ∈ o1 and f˜(τ ) = 1 ∀τ ∈ o2 . The Stone–Weierstrass theorem states that a continuous
function on a compact domain can be uniformly approximated to arbitrary accuracy by a polynomial. This
means there is a polynomial f ∈ R[x] with f (τ ) ≤ 1/3 ∀τ ∈ o1 and f (τ ) ≥ 2/3 ∀τ ∈ o2 . It follows that
h = R(f ) is an invariant polynomial that separates the two orbits: h(o1 ) ≤ 1/3 and h(o2 ) ≥ 2/3.
Thus, in order to determine the orbit of θ it is sufficient to determine the values of all invariant polynomials.
(This condition is clearly also necessary in the sense that if the orbit is uniquely determined then so are the
values of all invariants.)
11
Remark 4.5. In what follows we will be discussing algorithms that take the problem setup as input (includT
ing G̃ and its action on Ṽ , along with Π, K) and decide whether or not U≤d
(for some given d) is capable of a
particular recovery task (e.g. list recovery of a generic θ ∈ V ). We will always assume that these algorithms
have a procedure to compute a basis for UdT (for any d) in exact symbolic arithmetic. This is non-trivial
in some cases because Td (x) (and thus UdT ) involves integration over the group (and may involve irrational
values), but we will not worry about these details here. For the important case of SO(3), it is possible to
write down a basis for the invariants in closed form (see Appendix A).
Remark 4.6. We will draw from various references for algorithmic aspects of invariant theory. The case of
finite groups is treated by [Stu08]. Although the invariant ring is sometimes taken to be C[x]G instead of
R[x]G , this is unimportant in our setting because the two are essentially the same: since our group action
is real, a basis for R[x]G (over R) is a basis for C[x]G (over C). The case of infinite groups is covered by
[DK15]. Here the group is assumed to be a reductive group over C (or another algebraically-closed field).
This means in particular that the group is a subset of complex-valued matrices that is defined by polynomial
constraints. Although compact groups such as SO(3) do not satisfy this, the key property of a reductive
group is the existence of a Reynolds operator satisfying certain properties; since this exists for compact
groups (Definition 4.1), some (but not all) results still hold in our setting.
4.2
Generic list recovery
We will see that the case of list recovery of a generic signal is governed by the notion of algebraic independence.
Definition 4.7. Polynomials f1 , . . . , fm ∈ R[x] are algebraically dependent if there exists a nonzero polynomial P ∈ R[y1 , . . . , ym ] such that P (f1 , . . . , fm ) = 0 (i.e. P (f1 (x), . . . , fm (x)) is equal to the zero polynomial).
Otherwise, they are algebraically independent.
Definition 4.8. The transcendence degree of a subspace U ⊆ R[x], denoted trdeg(U ) is the maximum value
of m for which there exist algebraically independent f1 , . . . , fm ∈ U . A set of trdeg(U ) such polynomials is
called a transcendence basis of U .
We now present our algebraic characterization of the generic list recovery problem.
Theorem 4.9 (generic list recovery). Let U ⊆ R[x]G be a finite-dimensional subspace. If trdeg(U ) =
trdeg(R[x]G ) then there exists a set S ⊆ V of full measure such that if θ ∈ S then U list-resolves θ.
Conversely, if trdeg(U ) < trdeg(R[x]G ) then there exists a set S ⊆ V of full measure such that if θ ∈ S then
U does not list-resolve θ.
The proof is deferred to Sections 8.2 and 8.3. A set has full measure if its complement has measure zero.
The intuition behind Theorem 4.9 is that trdeg(R[x]G ) is the number of degrees of freedom that need to be
pinned down in order to learn the orbit of θ, and so we need this many algebraically independent constraints
(invariant polynomials). Note that we have not yet given any bound on how large the finite list might be;
we will address this in Section 4.3.
In order for Theorem 4.9 to be useful, we need a way to compute the transcendence degree of both
R[x]G and U . In what follows, we will discuss methods for both of these: in Section 4.2.1 we show how to
compute trdeg(R[x]G ) analytically, and in Section 4.2.2 we give an efficient algorithm to compute trdeg(U )
T
for a subspace U . By taking U = U≤d
this yields an efficient algorithm to determine the smallest degree d
T
at which U≤d list-resolves a generic θ (thereby answering Question 2.8 for the case of generic list recovery).
4.2.1
Computing the transcendence degree of R[x]G .
Intuitively, the transcendence degree of R[x]G is the number of parameters required to describe an orbit of
G. For finite groups, this is simply the dimension of V :
Proposition 4.10 ([Stu08] Proposition 2.1.1). If G is a finite group, trdeg(R[x]G ) = dim(V ).
12
For infinite groups, the situation may be slightly different. For instance, if SO(3) acts on V = R3 in
the standard way (rotations in 3 dimensions), then a generic orbit is a sphere, with dimension two. This
means there is only one parameter to learn, namely the 2-norm, and we expect R[x]G to have transcendence
degree 1 accordingly. On the other hand, if SO(3) acts on a rich class of functions S 2 → R (as in the S 2
registration problem; see Section 5.4) then each orbit resembles a copy of SO(3) which has dimension 3. This
is formalized in the following.
Proposition 4.11 ([Dol03] Corollary 6.2). If G is an algebraic group, then
trdeg(R[x]G ) = dim(V ) − dim(G) + min dim(Gv ),
v∈V
where Gv is the stabilizer at v of the action of G (that is, the subgroup of all g ∈ G fixing v).
An alternate approach to the transcendence degree of R[x]G uses a central object in invariant theory:
the Hilbert series (see e.g. [DK15]).
G
Definition 4.12. Let R[x]G
consisting of homogeneous invariants of
d be the subspace (over R) of R[x]
G
degree d. The Hilbert series of R[x] is the formal power series
H(t) :=
∞
X
d
dim(R[x]G
d )t .
d=0
For a given G acting on V , there is an explicit formula (Molien’s formula) for the Hilbert series:
Proposition 4.13 ([Kač94] Remark 3-1.8). Let ρ : G → GL(V ) be the representation by which G acts on
V . Then for |t| < 1, H(t) converges and we have
H(t) =
E
g∼Haar(G)
det(I − t ρ(g))−1 .
This formula is tractable to compute, even for complicated groups; see Section 5.4 for details in the case of
SO(3). Once we have the Hilbert series, it is easy to extract trdeg(R[x]G ) as follows.
Proposition 4.14. The order of the pole at t = 1 of H(t) is equal to trdeg(R[x]G ).
The proof comes from [DK15]; see Section 8.4 for more details.
For heterogeneous problems (K > 1), the transcendence degree can be computed easily from the transcendence degree of the corresponding homogeneous (K = 1) problem.
Proposition 4.15. Let G̃ be a compact group acting linearly and continuously on Ṽ , and let G = G̃K ⋊ SK
act on V = Ṽ ⊕K ⊕ ∆K as in Definition 2.2. Let R[x]G be the invariant ring corresponding to the action of
G on V , and let R[x̃]G̃ be the invariant ring corresponding to the action of G̃ on Ṽ (i.e. the K = 1 problem).
Then trdeg(R[x]G ) = K · trdeg(R[x̃]G̃ ) + K − 1.
The proof can be found in Section 8.5. Note, however, that the result is intuitively reasonable by counting
parameters. We know trdeg(R[x̃]G̃ ) is the number of parameters required to describe an orbit of G̃ acting
on Ṽ . Thus, in the heterogeneity problem we have trdeg(R[x̃]G̃ ) parameters for each of the K signals, plus
an additional K − 1 parameters for the K mixing weights (since they sum to 1).
4.2.2
Algorithm for transcendence basis of U .
In this section we prove the following.
Theorem 4.16. There is an efficient algorithm to perform the following task. Given a basis {u1 , . . . , us }
for a finite-dimensional subspace U ⊆ R[x], output a transcendence basis for U .
13
Our first ingredient is the following simple classical test for algebraic independence (see, e.g., [ER93,
BMS13] for a proof).
Definition 4.17 (Jacobian). Given polynomials f1 , . . . , fm ∈ R[x] = R[x1 , . . . , xp ], we define the Jacobian
matrix Jx (f1 , . . . , fm ) ∈ (R[x])m×p by (Jx (f1 , . . . , fm ))ij = ∂xj fi where ∂xj denotes formal partial derivative
with respect to xj .
Proposition 4.18 (Jacobian criterion for algebraic independence). Polynomials f = (f1 , . . . , fm ) are algebraically independent if and only if the Jacobian matrix Jx (f ) has full row rank (over the field R(x)).
It suffices to test the rank of the Jacobian at a generic point x.
Corollary 4.19. Fix f = (f1 , . . . , fm ). Let z ∼ N (0, Ip×p ). If f is algebraically dependent then Jx (f )|x=z
does not have full row rank. If f is algebraically independent then Jx (f )|x=z has full row rank with probability
1.
Proof. An m × p matrix has deficient row rank if and only if either m > p or every maximal square
submatrix has determinant zero. Every such determinant of Jx (f ) is a polynomial in x; if this polynomial is
not identically zero then plugging in generic values for x will not cause it to vanish.
Remark 4.20. In practice we may choose to plug in random rational values for x so that the rank computation can be done in exact symbolic arithmetic. The Jacobian test will still succeed with overwhelming
probability (provided we use a fine enough mesh of rational numbers). Also note that if we find any value
of x for which the Jacobian has full row rank, this constitutes a proof of algebraic independence.
Remark 4.21. In some cases (e.g. if the polynomials involve irrational values) it may be slow to compute the
Jacobian rank in exact symbolic arithmetic. We can alternatively compute the singular values numerically
and count how many are reasonably far from zero. This method works reliably in practice (i.e., it is
extremely clear how to separate the zero and nonzero singular values) but does not constitute a rigorous
proof of algebraic independence.
Curiously, although the Jacobian criterion gives an efficient test for algebraic dependence, it is much
harder (#P -hard) to actually find the algebraic dependence (i.e., the polynomial relation) when one exists
[Kay09].
The Jacobian criterion implies the well-known fact that the collection of algebraically independent subsets
of R[x] form a matroid ; this is called an algebraic matroid (see e.g. [Sch03]). In particular, we have the
following exchange property:
Proposition 4.22. Let I, J be finite subsets of R[x], each algebraically independent. If |I| < |J| then there
exists f ∈ J r I such that I ∪ {f } is algebraically independent.
We next note that in the task from Theorem 4.16, a transcendence basis can always be taken from the
basis {u1 , . . . , us } itself.
Lemma 4.23. Let U be a finite-dimensional subspace of R[x] with basis B = {u1 , . . . , us }. If U contains r
algebraically independent elements, then so does B.
Proof. Let B ′ ⊆ B be a maximal set of algebraically independent elements of B. If |B| < r then by the
exchange property
(Proposition 4.22) there exists v ∈ U r B ′ such that B ′ ∪ {v} is algebraically independent.
Ps
Write v = i=1 αi ui . Since B ′ is maximal, we have from the Jacobian criterion (Proposition 4.18) that
for all 1 P
≤ i ≤ s, the row vector Jx (ui ) lies in the R(x)-span of B := {Jx (b)}b∈B ′ . But this means that
Jx (v) = si=1 αi Jx (ui ) lies in the R(x)-span of B. By the Jacobian criterion this contradicts the fact that
B ′ ∪ {v} is algebraically independent.
Proof of Theorem 4.16.
Let {u1 , . . . , us } be a basis (or spanning set) for U . From above we have that the transcendence degree of U
14
is the row rank of the Jacobian Jx (u1 , . . . , us ) evaluated at a generic point x. A transcendence basis for U
is the set of ui corresponding to a maximal linearly independent set of rows
We can use the following simple greedy algorithm to construct a transcendence basis. As input, receive
a list of polynomials {u1 , . . . , us }. Initialize I = ∅. For i = 1, . . . , s, add {ui } to I if I ∪ {ui } is algebraically
independent, and do nothing otherwise. (Note that this condition can be efficiently tested by Corollary 4.19.)
Output the resulting set I.
We now show correctness. Let Ii be the set after item ui has been considered (and possibly added), and
set I0 = ∅. It suffices to show that for each i ∈ {0, . . . , s}, Ii is a maximal independent subset of {u1 , . . . , ui }.
We proceed by induction. The claim is vacuously true when i = 0. Assume it holds for i − 1. If Ii is not
a maximal independent subset of {u1 , . . . , ui }, then there exists an independent set J ⊆ {u1 , . . . , ui } with
|J| > |I|, so by the exchange property (Proposition 4.22) there exists a uj with j ≤ i such that uj ∈
/ Ii and
Ii ∪ {uj } is independent. In particular, the subset Ij−1 ∪ {uj } of Ii ∪ {uj } is independent. But the fact that
uj was not added at the (j − 1)th step implies that Ij−1 ∪ {uj } is not independent, a contradiction. So Ii is
indeed maximal.
We obtain that I = Is is a maximal independent subset of {u1 , . . . , us }, and hence by Lemma 4.23 a
transcendence basis of U .
4.3
Generic unique recovery
For list recovery problems, the following gives an explicit upper bound on the size of the list.
Theorem 4.24. Let U be a subspace of the invariant ring R[x]G . Let FG be the field of fractions of R[x]G .
If [FG : R(U )] = D < ∞ then there exists a set S ⊆ V of full measure such that for any θ ∈ S, U list-resolves
θ with a list of size ≤ D.
The proof is deferred to Section 8.2. Here R(U ) is the smallest subfield of FG containing both R and U , and
[FG : R(U )] denotes the degree of a field extension; see Section 8.2 for more details. Since [FG : R(U )] = 1
is equivalent to R(U ) = FG , we have the following criterion for unique recovery.
Corollary 4.25 (generic unique recovery). If R(U ) = FG then there exists a set S ⊆ V of full measure such
that if θ ∈ S then U resolves θ.
The intuition here is that we want to be able to learn every invariant polynomial by adding, multiplying,
and dividing polynomials from U (and scalars from R). We need θ to be generic so that we never divide by
zero in the process.
Theorem 4.26. For a finite-dimensional subspace U ⊆ R[x]G , there is an algorithm to compute the degree
of the field extension from Theorem 4.24. As input, the algorithm requires a basis for U and the ability to
compute the Reynolds operator (Definition 4.1).
We give the algorithm and the proof in Section 8.6. The algorithm uses Gröbner bases and is unfortunately
inefficient to run in practice.
Additionally, in Appendix B we present a method for proving field generation using Galois theory. There
we use it to show generic unique recovery in the case where G is a finite Abelian group and V is its regular
representation (over R). These ideas may be helpful for proving generic unique recovery in other settings.
4.4
Worst-case unique recovery
We give a sufficient algebraic condition for worst-case unique recovery:
Theorem 4.27 (worst-case unique recovery). Let U ⊆ R[x]G be a finite-dimensional subspace with basis
{f1 , . . . , fm }. If U generates R[x]G as an R-algebra (i.e. R[f1 , . . . , fm ] = R[x]G ) then U resolves every θ ∈ V .
Proof. Every element of R[x]G can be written as a polynomial in the fi (with coefficients in R). This means
the values f1 (θ), . . . , fm (θ) uniquely determine all the values f (θ) for f ∈ R[x]G and so the result follows
because R[x]G resolves every θ ∈ V (Theorem 4.4).
15
Theorem 4.28. There is an algorithm to test whether or not U generates R[x]G as an R-algebra. As input,
the algorithm requires a basis for U and the ability to compute the Reynolds operator (Definition 4.1).
We give the algorithm and the proof in Section 8.6. The algorithm uses Gröbner bases and is unfortunately
inefficient to run in practice.
If G is a finite group, it is known that R[x]G has a generating set for which all elements have degree at
most |G| (this is Noether’s degree bound ; see Theorem 2.1.4 in [Stu08]). It follows that R[x]G
≤|G| resolves
every θ ∈ V . Recall (from Section 1.4) that this is tight for MRA: degree |G| is necessary for worst-case
signals.
A precise characterization of when U resolves every θ ∈ V is (by definition) that U should be a separating
set or (equivalently) should generate a separating algebra (see [DK15] Section 2.4). The notions of generating and separating sets do not always coincide, as illustrated by Example 2.4.2 in [DK15]. Furthermore,
generating sets may require strictly higher maximum degree [Dom16].
4.5
Worst-case list recovery
We give a sufficient algebraic condition for worst-case list recovery:
Theorem 4.29 (worst-case list recovery). Let U ⊆ R[x]G be a subspace with finite basis {f1 , . . . , fm }. If
R[x]G is finitely generated as a R[f1 , . . . , fm ]-module, then U list-resolves every θ ∈ V .
In other words, this condition says that there exists a basis g1 , . . . , gs ∈ R[x]G such that every element of
R[x]G can be written as a linear combination of g1 , . . . , gs with coefficients from R[f1 , . . . , fm ]. It is sufficient
to take U to be a set of primary invariants from a Hironaka decomposition (see Section 8.4).
Proof. Since R[x]G finitely generated as an R-algebra (Theorem 4.3), if R[x]G is finitely generated as a
R[f1 , . . . , fm ]-module then it follows that (see [Sha94] Section 5.3) every h ∈ R[x]G satisfies a monic polynomial
hk + ck−1 hk−1 + · · · + c1 h + c0 = 0
with ci ∈ R[f1 , . . . , fm ]. Letting h1 , . . . , hs be generators for R[x]G (as an R-algebra), we have that the values
f1 (θ), . . . , fm (θ) determine a finite set of possible values for h1 (θ), . . . , hs (θ), each of which determines (at
most) one orbit for θ.
5
Examples
In this section we work out some specific examples, determining the degree at which generic list recovery is
possible using the methods of Section 4.2. (We focus on generic list recovery because our algorithms for the
other recovery criteria are unfortunately too slow even for quite small examples.) We obtain several recovery
theorems for problems such as MRA and cryo-EM within finite ranges of parameters where we have verified
the Jacobian criterion using a computer, and beyond these parameter ranges, we state conjectural patterns.
The following themes emerge in the examples studied in this section. First, we see that many problems
are possible at degree 3, which is promising from a practical standpoint. Second, we do not encounter any
unexpected algebraic dependencies, and so we are able to show that heuristic parameter-counting arguments
are correct. In particular, we see that if there are enough linearly independent invariants, there are also
enough algebraically independent invariants.
5.1
Learning a bag of numbers
Let G be the symmetric group Sp acting on V = Rp by permutation matrices. The invariant ring consists of
the symmetric polynomials, which are generated by the elementary symmetric polynomials e1 , . . . , ep where
ei has degree i. Worst-case unique recovery is possible at degree p since R[x]G
≤p generates the full invariant
ring. Furthermore, degree p is actually required, even for generic list recovery. This is because any invariant
16
of degree ≤ p − 1 can be expressed as a polynomial in e1 , . . . , ep−1 and thus trdeg(R[x]G
≤p−1 ) = p − 1. So
this problem has a steep sample complexity of order σ 2p .
5.2
Learning a rigid body
Let G be the rotation group SO(p) acting on the matrix space Rp×m by left multiplication. We imagine
the columns of our matrix as vertices defining a rigid body; thus we observe random rotations of this rigid
body (with vertices labeled) plus noise. Let U ∈ Rp×m be such a matrix signal. With O(σ 4 ) samples,
we can estimate the degree-2 Gram matrix U ⊤ U ; taking a Cholesky factorization, we recover U up to left
action by an element of the larger group O(p). Thus we recover the rigid body up to a reflection ambiguity,
demonstrating list recovery (with a list of size 2). Surprisingly, assuming m ≥ p, we do not uniquely resolve
a generic signal until degree p, where with O(σ 2p ) samples we can estimate a p × p minor of U , which is a
degree-p invariant that changes sign under reflection.
The impossibility of unique recovery until degree p is a consequence of the “first fundamental theorem”
for the special orthogonal group SO(p), which asserts that the invariant ring is generated by the entries of
the Gram matrix U ⊤ U together with the p × p minors of U (see for instance [Kač94]); thus the invariants of
degree 3, . . . , p − 1 carry no information in addition to the degree-2 invariants.
5.3
Multi-reference alignment (MRA)
Recall that this is the case of G = Z/p acting on V = Rp by cyclic shifts. It is already known that for the
basic MRA problem (without projection or heterogeneity), generic unique recovery is possible at degree 3 for
any p [BRW17]. The methods of Section 4.2 confirm the weaker result that generic list recovery is possible
at degree 3 (at least for the values of p that we tested). Note the stark contrast in difficulty from the case
of the full symmetric group G = Sp above.
Remark 5.1. This result for MRA is actually a special case of a more general phenomenon. Let G be any
finite group and let V be the regular representation i.e. the space of functions f : G → R with the action
(g · f )(h) = f (g −1 h). (Note that for G = Z/p this is precisely the MRA problem.) It is known [Kak09]
that for this setup, the triple correlation (a collection of degree-3 invariants) is sufficient to resolve a generic
signal, and thus generic unique recovery is possible at degree 3. In Appendix B we give a short proof for the
special case where the group is Abelian; our proof illustrates a method based on Galois theory which may
be useful to prove generic unique recovery in other settings.
We can also verify that for MRA with p ≥ 3, generic list recovery is impossible at degree 2. This follows
from Theorem 4.9 because trdeg(R[x]G ) = p (since G is finite) but the number of algebraically independent
invariants of degree ≤ 2 is at most ⌊p/2⌋+1. We can see this as follows. A basis for the invariants of degree ≤ 2
is {R(x1 ), R(x21 ), R(x1 x2 ), R(x1 x3 ), . . . , R(x1 xs )} with s = ⌊p/2⌋+1. Here R denotes the Reynolds operator,
which averages over cyclic shifts of the variables. For instance, R(x1 x2 ) = p1 (x1 x2 +x2 x3 +x3 x4 +· · ·+xp x1 ).
Note that the basis above has size ⌊p/2⌋ + 2 but there is an algebraic dependence within it because R(x1 )2
can be written in terms of the other basis elements. The claim now follows.
Generic list recovery is possible at degree 1 for p = 1 and at degree 2 for p = 2. (This is true even for
worst-case unique recovery; recall from Section 4.4 that degree |G| is always sufficient for this.)
We now move on to variants of the MRA problem.
5.3.1
MRA with projection
We now consider MRA with a projection step. We imagine that the coordinates of the signal are arranged
in a circle so that G acts by rotating the signal around the circle. We then observe a projection of the circle
onto a line so that each observation is the sum of the two entries lying “above” it on the circle. This is
intended to resemble the tomographic projection in cryo-EM. We formally define the setup as follows.
17
Problem 5.2 (MRA with projection). Let p ≥ 3 be odd. Let V = Rp and G = Z/p acting on V by cyclic
shifts. Let q = (p − 1)/2 and W = Rq . Let Π : V → W be defined by
Π(v1 , . . . , vp ) = (v1 + vp , v2 + vp−1 , . . . , v(p−1)/2 + v(p+3)/2 ).
We call the associated generalized orbit recovery problem (Problem 2.3) MRA with projection. (We consider
the homogeneous case K = 1.)
Note that since p is odd, there is one entry v(p+1)/2 which is discarded by Π. The reason we consider the
odd-p case rather than the seemingly more elegant even-p case is because generic list recovery is actually
impossible in the even-p case. This is because the signals θ and θ + (c, −c, c, −c, . . .) cannot be distinguished
from the samples, even if there is no noise.
Restricting now to odd p, note that we cannot hope for generic unique recovery because it is impossible to
tell whether the signal is wrapped clockwise or counterclockwise around the circle. In other words, reversing
the signal via (θ1 , . . . , θp ) 7→ (θp , . . . , θ1 ) does not change the distribution of samples. We can still hope for
generic list recovery, hopefully with a list of size exactly 2. This degeneracy is analogous to the chirality
issue in cryo-EM: it is impossible to determine the chirality of the molecule (i.e. if the molecule is reflected
about some 2-dimensional plane through the origin, this does not change the distribution of samples).
It appears that, as in the basic MRA problem, generic list recovery is possible at degree 3. We proved
this for p up to 21 by checking the Jacobian criterion (see Section 4.2) on a computer, and we conjecture
that this trend continues.
Conjecture 5.3. For MRA with projection, for any odd p ≥ 3, generic list recovery is possible at degree 3.
Note that generic list recovery is impossible at degree 2 because the addition of the projection step to basic
T
MRA can only make it harder for U≤d
to list-resolve θ.
5.3.2
Heterogeneous MRA
We now consider heterogeneous MRA, i.e. the generalized orbit recovery problem (Problem 2.3) with G̃ = Z/p
acting on Ṽ = Rp by cyclic shifts, K ≥ 2 heterogeneous components, and no projection (i.e., Π is the identity).
We will see that generic list recovery is possible at degree 3 provided that p is large enough compared
to K. First note that the number of degrees of freedom to be recovered is trdeg(R[x]G ) = Kp + K − 1 (see
Propositions 4.10 and 4.15). Let us now count the number of distinct entries of Td (x) for d ≤ 3. Note that
Td (x) is symmetric (under permutations of indices) but we also have additional symmetries given by cyclic
shifts, e.g. (T3 (x))i,j,k = (T3 (x))i+c,j+c,k+c where c is an integer and the sums i + c, j + c, k + c are computed
modulo p. One can compute that T1 (x) has 1 distinct entry, T2 (x) has ⌊p/2⌋ + 1 distinct entries, and T3 (x)
has p + ⌈(p − 1)(p − 2)/6⌉ distinct entries. The total number of distinct entries is
U := p + 2 + ⌊p/2⌋ + ⌈(p − 1)(p − 2)/6⌉.
By Theorem 4.9, list recovery is impossible when U < Kp + K − 1. By testing the Jacobian condition, we
observe that the converse also appears to hold. We tested this up to K = 15 and up to the corresponding
critical p value.
Conjecture 5.4. For heterogeneous (K ≥ 2) MRA, generic list recovery is possible at degree 3 precisely if
U ≥ Kp + K − 1. This condition on U can be stated more explicitly as follows:
• K = 2 requires p ≥ 1.
• K = 3 requires p ≥ 12.
• K = 4 requires p ≥ 18.
• Each K ≥ 5 requires p ≥ 6K − 5.
18
Recent work [BBLS17] also studies the heterogeneous MRA problem. Similarly to the present work, they
apply the method of moments and solve a polynomial system of equations in order to recover the signals.
To solve the system they use an efficient heuristic method that has no provable guarantees but appears to
work well in practice. Their experiments suggest that if the signals have i.i.d. Gaussian entries, this method
√
succeeds only when (roughly) K ≤ p instead of the condition (roughly) K ≤ p/6 that we see above (and
that [BBLS17] also identified based on parameter-counting). Exploring this discrepancy is an interesting
direction for future work.
One question of particular interest is whether this example evinces a statistical-computational gap,
√
whereby all polynomial-time methods fail to succeed once K exceeds p. Some evidence for why we might
expect this is as follows. Recent work on tensor decomposition [MSS16] gives a polynomial-time algorithm
PN
p
to decompose a third order tensor of the form i=1 a⊗3
i + E where ai ∈ R are i.i.d. from the unit sphere
1.5
and E is noise, provided N ≤ p (up to factors of log p). The heterogeneous MRA problem can be cast as
such a tensor decomposition problem with N = Kp components ai ; the components are the p cyclic shifts
of each of the K signals. Although these ai are not independent, we expect that if the signals are random
then the ai are “random enough” for the same tensor decomposition result to hold, which exactly yields the
√
condition K ≤ p (up to factors of log p).
5.4
S 2 registration
Recall that this is the case where the signal θ is a real-valued function defined on the unit sphere S 2 in R3 .
The formal setup is as follows.
Let G = SO(3). For each ℓ = 0, 1, 2, . . . there is an irreducible representation Vℓ of SO(3) of dimension
2ℓ + 1. These representations are of real type, i.e. they can be defined over the real numbers so that
Vℓ = R2ℓ+1 . Let F be a finite subset of {0, 1, 2, . . .} and consider the orbit recovery problem in which G acts
on V = ⊕ℓ∈F Vℓ .
As intuition for the above setup, Vℓ is a basis for the degree-ℓ spherical harmonic functions S 2 → R
defined on the surface of the unit sphere S 2 ⊆ R3 . The spherical harmonics are a complete set of orthogonal
functions on the sphere and can be used (like a “Fourier series”) to represent a function S 2 → R. Thus the
signal θ ∈ V can be thought of as a function on the sphere, with SO(3) acting on it by rotating the sphere.
See Appendix A for details on spherical harmonics.
The primary case of interest is F = {1, . . . , F } for some F (the number of “frequencies”). We will see
that generic list recovery is possible at degree 3 so long as F ≥ 10. We will see that it is convenient to not
include 0 ∈ F , but we now justify why this is without loss of generality. V0 is the trivial representation, i.e.
the 1-dimensional representation on which every group element acts as the identity. In the interpretation of
spherical harmonics, the V0 -component is the mean value of the function over the sphere. We claim that the
S 2 registration problem with 0 ∈ F can be easily reduced to the problem with F ′ = F r {0}. This is because
the V0 -component is itself a degree-1 invariant; given the value of this invariant, one can subtract it off and
reduce to the case without a V0 -component (i.e. the case where the function on the sphere is zero-mean).
Thus we have that e.g. generic list recovery is possible (at a given degree) for F if and only if it is possible
for F ′ .
Using Proposition 4.11 we compute that trdeg(R[x]G ) = p − p′ , where
X
p = dim(V ) =
(2ℓ + 1)
ℓ∈F
and
0 ℓmax = 0
2 ℓmax = 1
where ℓmax = max ℓ.
p′ =
ℓ∈F
3 ℓmax ≥ 2
After all, V0 is the trivial representation on the 1-dimensional vector space, with 3-dimensional stabilizer
SO(3), and V1 is the standard 3-dimension representation of SO(3) on R3 by rotations, which yields a onedimensional SO(2) stabilizer at each nonzero point. When ℓmax ≥ 2, the representation V is known to have
zero-dimensional stabilizer at some points (see e.g. [Ete96]).
19
In the following we restrict to the case 0 ∈
/ F for simplicity (but recall that this is without loss of
G
generality). There are therefore no degree-1 invariants, i.e. R[x]G
1 is empty. By Theorem 4.9, if dim(R[x]2 ) +
G
G
dim(R[x]3 ) < trdeg(R[x] ) then generic list recovery is impossible at degree 3; this rules out generic list
recovery for F = {1, 2, . . . , F } when F ≤ 9. (We will see below how to compute dim(R[x]G
d ).) Beyond this
threshold, the situation is more hopeful:
Theorem 5.5. If F = {1, 2, . . . , F } and 10 ≤ F ≤ 16 then the degree-3 method of moments achieves generic
list recovery.
This theorem is based on computer verification of the Jacobian criterion for 10 ≤ F ≤ 16 using exact
arithmetic in a finite extension of Q. This result lends credence to the following conjecture.
Conjecture 5.6. Consider the S 2 registration problem with 0 ∈
/ F . We conjecture the following.
G
G
• Generic list recovery is possible at degree 3 if and only if dim(R[x]G
2 ) + dim(R[x]3 ) ≥ trdeg(R[x] )
G
G
(where trdeg(R[x] ) is computed above and dim(R[x]d ) can be computed from Proposition 5.7 below).
• In particular, if F = {1, 2, . . . , F } then generic list recovery is possible at degree 3 if and only if F ≥ 10.
The reason it is convenient to exclude the trivial representation is because it simplifies the parametercounting: if we use the trivial representation then we have a degree-1 invariant f and so there is an algebraic
relation between the degree-2 invariant f 2 and the degree-3 invariant f 3 .
We now discuss how to compute dim(R[x]G
d ). Using the methods in Section 4.6 of [DK15], we can give a
formula for the Hilbert series of R[x]G ; see Section 9.1. However, if one wants to extract a specific coefficient
dim(R[x]G
d ) of the Hilbert series, we give an alternative (and somewhat simpler) formula:
Proposition 5.7. Consider S 2 registration with frequencies F . Let χd (φ) : R → R be defined recursively by
χ0 (φ) = 1,
χ1 (φ) =
X
ℓ∈F
"
1+2
ℓ
X
m=1
#
cos(m φ) , and
d
χd (φ) =
Then we have
dim(R[x]G
d) =
1X
χ1 (iφ)χd−i (φ).
d i=1
1
π
Z
π
0
(1 − cos φ)χd (φ) dφ.
We give the proof in Section 9.2. Additionally, in Appendix A.6 we give explicit formulas for the invariants
(up to degree 3), which yields a combinatorial analogue of Proposition 5.7 (up to degree 3).
5.5
Cryo-EM
We adapt the following simple model for the cryo-EM reconstruction problem. We will use properties of the
3-dimensional Fourier transform, including the projection-slice theorem; see e.g. [Osg07] for a reference.
The signal is a 3-dimensional molecule, which we can think of as encoded by a density function f : R3 → R.
The 3-dimensional Fourier transform of f is fb : R3 → C given by
Z ∞Z ∞Z ∞
b
f (kx , ky , kz ) =
e−2πi(xkx +yky +zkz ) f (x, y, z) dx dy dz.
(1)
−∞
−∞
−∞
It is sufficient to learn fb because we can then recover f using the inverse Fourier transform. SO(3) acts
on the molecule by rotating it in 3-dimensional space (keeping the origin fixed). When f is rotated in
20
(x, y, z) coordinates, fb is also rotated in (kx , ky , kz )-coordinates by the same rotation. Each observation is
a 2-dimensional image obtained by first rotating f by a random element of SO(3) and then projecting f
parallel to the z axis. Specifically, the projection of f is fproj : R2 → R given by
Z ∞
fproj(x, y) =
f (x, y, z) dz.
−∞
By the projection-slice theorem, the 2-dimensional Fourier transform of fproj is equal to the slice fbslice : R2 →
C given by
fbslice (kx , ky ) = fb(kx , ky , 0).
Thus we think of fb as our unknown signal with SO(3) acting by rotation, and with post-projection which
reveals only the slice of fb lying in the plane kz = 0.
This does not yet conform to our definition of a (generalized) orbit recovery problem because the signal
needs to lie in a finite-dimensional real vector space. Instead of thinking of fb as a function on R3 , we fix a
finite number S of nested spherical shells in R3 , each of different radius and all centered at the origin. We
consider only the restriction of fb to these shells. We fix a finite number F of frequencies and on each shell
we expand fb (restricted to that shell) in the basis of spherical harmonics, truncated to 1 ≤ ℓ ≤ F . (As in S 2
registration, we can discard the trivial representation ℓ = 0 without loss of generality, and it is convenient
to do so.) Being the Fourier transform of a real-valued function, fb satisfies
fb(−kx , −ky , −kz ) = fb(kx , ky , kz )
(2)
(see (1)) and so we can use a particular basis Hℓm of spherical harmonics for which the expansion coefficients
are real; see Appendix A. We have now parametrized our signal by a finite number of real values θsℓm with
1 ≤ s ≤ S, 1 ≤ ℓ ≤ F , and −ℓ ≤ m ≤ ℓ. In particular, the restriction of fb to shell s has expansion
X
X
θsℓm Hℓm .
1≤ℓ≤F −ℓ≤m≤ℓ
SO(3) acts on each shell by 3-dimensional rotation; see Section A for the details of how SO(3) acts on spherical
harmonics. The projection Π reveals only the values on the equator z = 0 (or in spherical coordinates,
θ = π/2) of each shell. Using again the property (2), the output of Π on each shell has an expansion with
real coefficients in a particular finite basis hm ; see Section A.4.
Remark 5.8. There are various other choices one could make for the basis in which to represent the
(Fourier transform of the) molecule. Each of our basis functions is the product of a spherical harmonic and
a radial delta function (i.e. a delta function applied to the radius, resulting in a spherical shell). Another
common basis is the Fourier–Bessel basis (used in e.g. [LBB+ 17]) where each basis function is the product
of a spherical harmonic and a radial Bessel function. More generally we can take the product of spherical
harmonics with any set of radial basis function. It turns out that the choice of radial basis is unimportant
because the resulting problem will be isomorphic to our case (spherical shells) and so the same results hold.
We now present our results on the above cryo-EM model. We focus on identifying the regime of parameters
for which generic list recovery is possible at degree 3. Again using Proposition 4.11, we have for F ≥ 2:
trdeg(R[x]G ) = dim(V ) − 3 = S
F
X
ℓ=1
(2ℓ + 1) − 3 = S(F 2 + 2F ) − 3
(3)
where again we have a zero-dimensional stabilizer.
T
In Appendix A we give an explicit construction of the invariant polynomials in U≤3
. By testing the
Jacobian criterion in exact arithmetic on small examples, we arrive at the following theorem:
Theorem 5.9. Consider the homogeneous (K = 1) cryo-EM problem with S shells and F frequencies.
21
• If S = 1 then for any F ≥ 2, generic list recovery is impossible at degree 3.
• If 2 ≤ S ≤ 4 and 2 ≤ F ≤ 6, the degree-3 method of moments achieves generic list recovery.
The first assertion results from a simple counting argument: there are fewer invariants at degree ≤ 3 than
degrees of freedom. The second part is by confirming that the Jacobian of the invariants has rank equal to
trdeg(R[x]G ), through computer-assisted exact arithmetic over an appropriate finite extension of Q.
In floating-point arithmetic, we have further verified that the Jacobian appears to have appropriate rank
for 2 ≤ S ≤ 10 and 2 ≤ F ≤ 10, leading us to conjecture the following:
Conjecture 5.10. If S ≥ 2 then the degree-3 method of moments achieves generic list recovery (regardless
of F ).
T
Intuitively, when there is a single shell (S = 1) there are simply not enough invariants in U≤3
. However,
when S ≥ 2, the number of invariants increases dramatically due to cross-terms that involve multiple shells.
5.5.1
Heterogeneous cryo-EM
We now consider heterogeneous cryo-EM (K ≥ 2). By combining (3) with Proposition 4.15 we can compute
trdeg(R[x]G ). Based on testing the Jacobian criterion on small examples, we conjecture that the degree-3
method of moments achieves generic list recovery if and only if dim(U2T ) + dim(U3T ) ≥ trdeg(R[x]G ). In
T
other words, we expect no unexpected algebraic dependencies among U≤3
. (Recall that there are no degree-1
invariants since we are not using the trivial representation ℓ = 0).
In Section A.6 we give a conjectured formula for the exact value of dim(U2T )+dim(U3T ) for all S ≥ 1, F ≥ 2.
As a result we can determine for any given S ≥ 1 and F ≥ 2, the exact condition on K for which we believe
generic list recovery is possible. For S and F large, this condition is approximately K ≤ S 2 /4.
6
Open questions
We leave the following as directions for future work.
1. Our methods require testing the rank of the Jacobian on a computer for each problem size. It would
be desirable to have analytic results for e.g. (variants of) MRA in any dimension p.
2. We have given an efficient test for whether generic list recovery is possible, but have not given a
similarly efficient test for generic unique recovery. In cases where unique recovery is impossible, it
would be nice to give a tight bound on the size of the list; for instance, for MRA with projection, we
conjecture that the list has size exactly 2 (due to “chirality”), but we lack a proof for this fact. Our
algorithms for testing generic unique recovery are based on Gröbner bases, the calculation of which is
known to be computationally hard in the worst case [Huỳ86]. Unfortunately, the algorithms we have
proposed are also extremely slow in practice, though a faster implementation may be possible.
3. Our procedure for recovering θ from the samples involves solving a polynomial system of equations.
While solving polynomial systems is NP-hard in general, the fact that the polynomials used in the orbit
recovery problem have special structure leaves open the possibility of finding an efficient (polynomial
time) method with rigorous guarantees. Possible methods include tensor decomposition [PWB+ 17]
and non-convex optimization [BBLS17].
4. We have addressed the statistical limits of orbit recovery problems. However, prior work has indicated
the presence of statistical-to-computational gaps in related synchronization problems [PWBM16a], and
we expect such gaps to appear in orbit recovery problems too. As discussed in Section 5.3.2, the results
of [BBLS17] suggest a possible gap of this kind for heterogeneous MRA.
22
7
Proofs for Section 3: statistical results
We first prove Theorem 3.2. This theorem in fact holds for more general mixture problems, not merely those
arising from the orbit recovery problems defined in Problem 2.3. For convenience, we will state and prove
the theorem in its general form.
Problem 7.1 (mixture recovery). Let V = Rp , and let Θ ⊂ V be compact. For θ ∈ Θ, let µθ be a measure
on Rp whose support is contained in the unit ball, and assume the map θ 7→ µθ is continuous with respect to
the weak topology. Let D be a known distribution on R with finite moments of all orders, and let σ ≥ 1. For
i ∈ [n] = {1, 2, . . . , n}, we observe
yi = xi + σξi ,
where xi ∼ µθ and the entries of ξi are independently drawn from D. The goal is to estimate θ.
Write Pθ for the distribution arising from the parameter θ, and let Eθ be expectation with respect to
this distribution. We denote by Enθ the expectation taken with respect to n i.i.d. samples from Pθ . Where
there is no confusion, we also write Eθ for expectation with respect to the distribution µθ .
We require that µθ have bounded support; the requirement that it be supported in the unit ball is
for normalization purposes only. We assume throughout that σ ≥ 1. The following definiton gives the
generalization of Definition 3.1 to the mixture recovery problem.
Definition 7.2. Given a positive integer d and θ ∈ V , the order-d matching set Mθ,d is the set consisting of
all φ ∈ V such Eθ [x⊗k ] = Eφ [x⊗k ] for k = 1, . . . , d, where Eζ represents expectation with respect to x ∼ µζ .
Problem 7.1 generalizes Problem 2.3 by allowing the random vector xi to arise from more general mixtures
than those arising from group actions. Note that when the mixtures do arise from a generalized orbit recovery
problem, i.e., when xi = Π(gi · θki ), where gi and ki are distributed as in Problem 2.3, then Definition 7.2
reduces to Definition 3.1.
Having made these definitions, our goal in this section is to show that Theorem 3.2 holds word-for-word
in the setting of mixture recovery. To do so, we show that entries of the moment tensors Eθ [x⊗k ] can be
estimated on the basis of O(σ 2k ) samples from Pθ .
7.1
Estimation of moments
Our estimators will be based on a system of orthogonal univariate polynomials under the measure corresponding to D. Let H0 (x) = 1, and for k ≥ 1, define
Hk (x) = xk −
k−1
X
Eξ∼D [ξ k Hj (ξ)]
j=0
Eξ∼D [Hj (ξ)2 ]
Hj (x) .
It is easy to check that these polynomials are orthogonal under the inner product given by hf, gi =
Eξ∼D [f (ξ)g(ξ)] and that the polynomials H0 , . . . , Hk form a basis for the space of polynomials of degree
at most k. We denote them by Hk because they coincide with the classic Hermite polynomials when D is
Gaussian.
Like the Hermite polynomials, they satisfy the identity
Eξ∼D [Hk (x + ξ)] = xk .
Indeed, we can expand Hk (x + ξ) in the basis of the orthogonal polynomials as
Hk (x + ξ) =
k
X
j=0
23
αj (x)Hj (ξ) ,
(4)
where αj (x) is a polynomial in x of degree at most k−j. Since Hj (ξ) has zero mean for j ≥ 1 by construction,
we obtain
E[Hk (x + ξ)] = α0 (x) ,
and since Hk is a monic polynomial of degree k, we must have α0 (x) = xk .
We briefly review multi-index notation.
Definition 7.3. Q
A p-dimensional multi-index is a tuple α = (α1 , . . . , αp ) of nonnegative integers. For
α
p
x ∈ Rp , let xα = j=1 xj j .
Pp
For any multi-index α, we write |α| = j=1 αj . Given independent samples y1 , . . . , yn from Pθ , consider
the estimate for Ex∼µθ [xα ] given by
n
α :=
xf
α is unbiased.
We first show that xf
p
1 X Y αj
σ Hαj (σ −1 yi ) .
n i=1 j=1
α ] = E [xα ].
Lemma 7.4. For all θ ∈ Θ, Enθ [xf
θ
α is a sum of i.i.d. terms, it suffices to prove the claim for a single sample. By (4),
Proof. Since xf
Eθ
p
hY
j=1
h
i
σ αj Hαj (σ −1 yi ) = E
x∼µθ
= E
x∼µθ
= E
x∼µθ
E
ξ1 ,...,ξp ∼D
p
hY
E
ξj ∼D
j=1
p
hY
p
Y
j=1
i
σ αj Hαj (σ −1 (xj + σξj )) x
i
σ αj Hαj (σ −1 xj + ξj ) x
i
σ αj (σ −1 xj )αj = Eθ [xα ] .
j=1
It remains to bound the variance.
α] ≤
Proposition 7.5. For any multi-index α, there exists a constant cα such that for all θ ∈ Θ, Varθ [xf
−1 2|α|
cα n σ .
α is a sum of i.i.d. terms, it suffices to prove the claim for n = 1. Given a multi-index α, let
Proof.QSince xf
p
cα = j=1 supxj ∈[−1,1] Eξj ∼D [Hαj (xj + ξj )2 ], and note that cα is independent of σ. We obtain
α] ≤
Varθ [xf
E
x∼µθ
≤
p
hY
sup
E
ξ ∼D
j=1 j
p
Y
E
x:kxk≤1 j=1 ξj ∼D
≤ cα σ 2|α| ,
i
σ 2αj Hαj (σ −1 xj + ξj )2 x
2αj
σ Hαj (σ −1 xj + ξj )2
as claimed.
Finally, we apply the “median-of-means” trick [NY83] to show that we can combine the estimators defined
above to obtain estimates for the moment tensors Eθ [x⊗k ] for k ≤ d which are close to their expectation
with high probability.
24
Proposition 7.6. Let y1 , . . . , yn be i.i.d. samples from Pθ . For any degree d and accuracy parameter δ,
α = x
α (y , . . . , y ) for all α with |α| ≤ d such that with probability at least 1 − δ,
c
there exist estimators xc
1
n
r
log(p/δ)
α
d
α
c
max |Eθ [x ] − x | ≤ cd σ
,
α:|α|≤d
n
for some constant cd .
Proof. Split the samples into m subsamples of equal size, for some m to be specified, and for each α construct
α, . . . , x
α on the basis of the m subsamples. (We assume for convenience that m divides
f
the m estimators xf
m
1
α, . . . , x
α
α.
c
f
n.) Let x be the median of xf
m
1
Chebyshev’s inequality together with Proposition 7.5 implies that there exists a constant cd such that,
for each j = 1, . . . , m and multi-index α,
r i
h
m
α
d
α
Pr |xf
≤ 1/4 ,
j − Eθ [x ]| > cd σ
n
α, . . . , x
α are independent, a standard concentration argument shows that
f
and since the estimators xf
m
1
r i
h
m
α − E [xα ]| > c σ d
≤ e−m/4 .
Pr |xc
θ
d
n
By a “stars-and-bars” counting argument [Fel68], there are p+d
multi-indices α satisfying |α| ≤ d, so taking
d
a union bound over all choices of α yields
r
m
cα | ≤ cd σ d
max |Eθ [xα ] − x
α:|α|≤d
n
−m/4
with probability at least 1 − p+d
. Choosing m = 4 log( p+d
d e
d /δ) and taking cd sufficiently large in
the statement of the theorem yields the claim.
Note that the constant cd in the statement of Proposition 7.6 can be made explicit, given knowledge of
the distribution D.
7.2
Robust solutions to polynomial systems
We now show that approximate knowledge of the moment tensors Eθ [x⊗k ] for k = 1, . . . , d suffices to
approximately recover θ.
Lemma 7.7. For all θ ∈ Θ and ε > 0, there exists a ε′ > 0 such that, if φ ∈ Θ satisfies maxk≤d kEθ [x⊗k ] −
Eφ [x⊗k ]k∞ < ε′ , then there exists a τ ∈ Mθ,d such that kφ − τ k < ε.
Proof. We employ a simple compactness argument. Consider the set F = {φ ∈ Θ : ∀ τ ∈ Mθ,d kφ − τ k ≥ ε}.
Since Θ is compact, so is F . Set
ε′ = min max kEθ [x⊗k ] − Eφ [x⊗k ]k∞ .
φ∈F k≤d
⊗k
⊗k
Clearly if maxk≤d kEθ [x ] − Eφ [x ]k∞ < ε′ for some φ ∈ Θ, then there exists a τ ∈ Mθ,d such that
kφ − τ k < ε, so it remains to check that ε′ > 0.
Since θ 7→ µθ is continuous with respect to the weak topology and µθ is supported on a compact set
for all θ ∈ Θ, the moment map θ 7→ Eθ [x⊗k ] is also continuous for all k ≤ d. If φ ∈ F , then in particular
φ∈
/ Mθ,d, so there exists a k ≤ d for which Eθ [x⊗k ] 6= Eφ [x⊗k ]. Therefore ε′ > 0, as desired.
Lemma 7.7 is simply stating that the function φ 7→ minτ ∈Mθ,d kφ − τ k is continuous at θ with respect to
the topology induced by the moment maps. Note that, for generic θ when µθ arises from an orbit recovery
problem, the moment map will be continuously differentiable with a nonsingular Jacobian, so the inverse
function theorem implies ε′ can be taken to be Ω(ε). In general, however, the dependence could be worse.
25
7.3
Proof of Theorem 3.2
cα as in Proposition 7.6, and let
Construct the estimators x
(
)
r
log(p/δ)
α | ≤ c σd
cn = φ ∈ Θ : max |Eφ [xα ] − xc
M
d
n
α:|α|≤d
cn and that, for all
Applying Proposiiton 7.6, we have with probability at least 1 − δ that Mθ,d ⊆ M
c
φ ∈ Mn ,
r
log(p/δ)
⊗k
⊗k
α
α
d
.
max kEφ [x ] − Eθ [x ]k = max |Eφ [x ] − Eθ [x ]| ≤ 2cd σ
k≤d
n
α:|α|≤d
q
By Lemma 7.7, there exists an ε′θ,ε such that, as long as 2cd σ d log(p/δ)
< ε′θ,ε , then with probability at
n
cn ⊆ Mε .
least 1 − δ, we have the desired inclusion M
θ,d
Therefore taking n > (2cd /ε′θ,ε )2 log(p/δ)σ 2d = cθ,ε,d log(1/δ)σ 2d suffices.
7.4
Information geometry of gaussian mixtures
In this section, we establish an upper bound on the Kullbeck-Leibler divergence between different gaussian
mixtures, which we denote by D(· k ·).
The proof follows the outline used in [BRW17], based off a technique developed in [LNS99, CL11].
Proposition 7.8. Let θ, φ ∈ Θ, let the distribution D be N (0, 1) for some σ ≥ 1, and let d be any positive
integer.
There exist universal constants C and c such that if Eθ [x⊗k ] = Eφ [x⊗k ] for k ≤ d − 1, then
D(Pθ k Pφ ) ≤ C
(cσ)−2d
.
d!
Proof. We first establish the claim when d = 1. Note that the condition on the moment tensors is vacuous
in this case. By the convexity of the divergence,
D(Pθ k Pφ ) ≤
2
′
2
E D(N (x, σ ), N (x , σ )) =
x∼µ
θ
x′ ∼µφ
E
x∼µ
θ
x′ ∼µφ
kx − x′ k2
≤ 2σ −2 ,
2σ 2
where in the last step we used the fact that x and x′ lie in the unit ball almost surely.
Now, assume d > 1, so in particular Eθ [x] = Eφ [x]. Denote their common mean by v. For ζ ∈ {θ, φ}
denote by µζ the distribution of x − v when x ∼ µζ , and let Pζ denote distribution of y when y = x + ξ
for x ∼ µζ and ξ ∼ N (0, σ 2 I). Since this transformation is a deterministic bijection, the data processing
inequality implies D(Pθ k Pφ ) = D(Pθ k Pφ ).
Note that Eµθ [x] = Eµθ [x] = 0 and Eµθ [x⊗k ] = Eµφ [x⊗k ] for k ≤ d − 1. Hence without loss of generality
we can reduce to the case where µθ and µφ are both centered and are supported in a ball of radius 2.
We bound the χ2 -divergence between Pθ and Pφ . Let f be the density of a standard p-dimensional
Gaussian and for ζ ∈ Θ, let fζ be the density of Pζ , which can be written explicitly as
1
fζ (y) = Ex∼µζ σ −p f (σ −1 (y − x)) = σ −p f (σ −1 y)Ex∼µζ e− 2σ2 (x
2
−2y ⊤ x)
.
Since kxk ≤ 2 almost surely with respect µζ , Jensen’s inequality implies that
1
fζ (y) ≥ σ −p f (σ −1 y)e− 2σ2 (4−2y
⊤
Eζ x)
Recall that the χ2 divergence is defined by
χ2 (Pθ , Pφ ) =
Z
(fθ (y) − fφ (y))2
dy .
fθ (y)
26
2
= σ −p f (σ −1 y)e− σ2 .
(5)
Applying (5) to the denominator, expanding the definitions of fθ and fφ , and applying a change of variables
yields
Z
2
⊤
2
⊤
1
1
2
χ2 (Pθ , Pφ ) ≤ e2/σ
(Eθ e− 2σ2 (x −2y x) − Eφ e− 2σ2 (x −2y x) )2 σ −p f (σ −1 y) dy
Z
⊤
−1
−1
2
⊤
−1
−1
2
1
1
2/σ2
=e
(Eθ ey (σ x)− 2 kσ xk − Eφ ey (σ x)− 2 kσ xk )2 f (y) dy
2
= e2/σ E(Eθ eg
⊤
(σ−1 x)− 12 kσ−1 xk2
− Eφ eg
⊤
(σ−1 x)− 12 kσ−1 xk2 2
g ∼ N (0, I) .
) ,
(6)
Given ζ, ζ ′ ∈ Θ, let x ∼ µζ and x′ ∼ µζ ′ be independent. Then interchanging the order of expectation
and using the expression for the moment generating function of a standard Gaussian random variable, we
obtain
x⊤ x′
⊤
−1
′
−1
2
−1 ′ 2
1
Eζ,ζ ′ Eg eg (σ (x+x ))− 2 (kσ xk +kσ x k ) = Eζ,ζ ′ e σ2 .
Applying this expression to (6) after expanding the square produces
2
χ2 (Pθ , Pφ ) ≤ e2/σ E e
x∼µ
x⊤ x′
σ2
θ
−2 E e
x∼µ
θ
x′ ∼µθ
x′ ∼µφ
x⊤ x′
σ2
+
E e
x⊤ x′
σ2
x∼µφ
x′ ∼µφ
,
where in each expectation the random variables x and x′ are independent. Since µθ and µφ are compactly
supported, Fubini’s theorem implies we can expand each term as a power series and interchange expectation
and summation to produce
∞
−2k
2 X σ
E (x⊤ x′ )k − 2 E (x⊤ x′ )k + E (x⊤ x′ )k
χ2 (Pθ , Pφ ) ≤ e2/σ
x∼µφ
x∼µθ
x∼µθ
k!
′
′
′
k=0
= e2/σ
= e2/σ
= e2/σ
2
2
2
∞
X
k=0
∞
X
k=0
∞
X
k=d
x ∼µφ
x ∼µθ
σ
−2k
k!
x ∼µp hi
hEθ x⊗k , Eθ x⊗k i − 2hEθ x⊗k , Exφ x⊗k i + hEφ x⊗k , Eφ x⊗k i
σ −2k
kEθ [x⊗k ] − Eφ [x⊗k ]k2HS
k!
σ −2k
kEθ [x⊗k ] − Eφ [x⊗k ]k2HS ,
k!
where h·, ·i denotes the Frobenius inner product between tensors and k · kHS denotes the Hilbert-Schmidt
norm. Since under both µθ and µφ , kxk ≤ 2 almost surely, we have for all k ≥ 2,
kEθ [x⊗k ] − Eφ [x⊗k ]k2HS ≤ 2kEθ [x⊗k ]k2HS + 2kEφ [x⊗k ]k2HS ≤ 4k+1 .
Therefore
χ2 (Pθ , Pφ ) ≤ 4e2/σ
2
2
∞
d −2d
X
2 4 σ
4k σ −2k
≤ 4e6/σ
,
k!
d!
k=d
and applying the inequality D(Pθ k Pφ ) ≤ χ (Pθ , Pφ ) [Tsy09] proves the claim.
7.5
Proof of Theorem 3.4
If τ1 and τ2 are both in Mθ,d−1, then by Proposition 7.8, the corresponding distributions Pτ1 and Pτ2 satisfy
D(Pτ1 k Pτ2 ) ≤ (cσ)C2d d! . By the Neyman-Pearson lemma, for any test ψ using n samples,
s
r
1
Cn
n
n
Pr(ψ = 2) + Pr(ψ = 1) ≥ 1 − dTV (Pτ1 , Pτ2 ) ≥ 1 −
D(Pnτ1 k Pnτ2 ) = 1 −
,
τ2
τ1
2
2(cσ)2d d!
27
where we have applied Pinsker’s inequality and the chain rule for divergence. Therefore, to achieve an error
probability of at most 1/3, we must have n ≥ 2(cσ)2d d!/(9C) = cd σ 2d , as claimed.
8
8.1
Proofs for Section 4: algebraic results
Algorithm for generators of R[x]G
We know that R[x]G is finitely generated as an R-algebra (Theorem 4.3). There are various algorithms to
compute a finite set of generators for R[x]G [Stu08, DK15]. However, some require the group to be finite or to
be reductive over an algebraically-closed field. One algorithm that certainly works in our context (compact
groups) is Algorithm 2.2.5 in [Stu08]. As input it requires the Hilbert series of R[x]G (which can be computed
by Proposition 4.13) and a procedure to compute a basis for R[x]G
d (which can be done with the Reynolds
operator by Observation 4.2). The idea is as follows. We keep a set of proposed generators f1 , . . . , fm .
At each step we compare the Hilbert series of R[x]G with the Hilbert series of R[f1 , . . . , fm ] (which can be
computed using Gröbner bases). If these series differ at the td term, this means we are missing an invariant
at degree d. To remedy this, we create a new homogeneous invariant of degree d using the Reynolds operator,
and add it to our set of proposed generators. We repeat until the Hilbert series match.
8.2
Bounding the list size for generic signals
In this section we prove Theorem 4.24 and the first part of Theorem 4.9 (see Section 8.3 for the second part).
Recall the following basic definitions and facts from field theory.
Definition 8.1. If F2 is a subfield of F1 , we write F1 /F2 and call this a field extension. The degree of the
extension, denoted [F1 : F2 ], is the dimension of F1 as a vector space over F2 .
Proposition 8.2. Let R ⊆ F2 ⊆ F1 with F1 finitely generated (as a field) over R. Let r be the transcendence
degree of F1 (over R). The field extension F1 /F2 has finite degree if and only if F1 contains r algebraically
independent elements.
Proof. This is a basic fact of field theory. If F1 contains r algebraically independent elements then the
extension F1 /F2 is algebraic and finitely generated, and therefore has finite degree. Otherwise, the extension
is transcendental and has infinite degree.
In light of the above (and using the fact that R[x]G is finitely generated), Theorem 4.24 implies the first
part of Theorem 4.9 and so it remains to prove Theorem 4.24 (i.e. list size is bounded by D := [FG : R(U )]).
Proof of Theorem 4.24.
Write FU := R(U ). In characteristic zero, every algebraic extension is separable, so by the primitive element
theorem, FG = FU (α) for some α ∈ FG . Since α generates a degree-D extension, α is the root of a degree-D
polynomial
αD + bD−1 αD−1 + · · · + b1 α + b0
(7)
with coefficients bi ∈ FU . Furthermore, every element of FG can be expressed as
c0 + c1 α + · · · + cD−1 αD−1
with ci ∈ FU . In particular, let g1 , . . . , gk be generators for R[x]G (as an R-algebra) and write
(i)
(i)
(i)
gi = c0 + c1 α + · · · + cD−1 αD−1 .
(8)
(i)
Let S ⊆ V be the subset for which α and all the (finitely-many) coefficients bi , cj have nonzero denominators;
S is a non-empty Zariski-open set and thus has full measure. Now fix θ ∈ S. Given the values f (θ) for all
f ∈ U , each bi takes a well-defined value in R and so from (7) there are at most D possible values that α(θ)
can take. From (8), each value of α(θ) uniquely determines all the values gi (θ) and thus uniquely determines
all the values f (θ) for f ∈ R[x]G . Since R[x]G resolves θ (Theorem 4.4), this completes the proof.
28
8.3
Generic list recovery converse
In this section we prove the second part of Theorem 4.9 (the converse).
Let p = dim(V ), trdeg(U ) = q, and trdeg(R[x]G ) = r so that q < r ≤ p. Let f = {f1 , . . . , fm } be a basis
for U , and let g = {g1 , . . . , gr } be a transcendence basis for R[x]G . Let S ⊆ V be the set of points θ for
which the Jacobian Jx (f )|x=θ has row rank q and the Jacobian Jx (g)|x=θ has row rank r; by the Jacobian
criterion (see Corollary 4.19), S is a non-empty Zariski-open set and thus has full measure.
Fix θ ∈ S. For a sufficiently small open neighborhood X ⊆ S containing θ we have the following. The
Jacobian criterion on f implies that {τ ∈ X : f (τ ) = f (θ)} has dimension p − q. The Jacobian criterion on
g implies that every z ∈ g(X) has a preimage g−1 (z) := {τ ∈ X : g(τ ) = z} of dimension p − r. Since
p − q > p − r it follows that there are infinitely many θ1 , θ2 , . . . ∈ X such that f (θi ) = f (θ) but the values
g(θ1 ), g(θ2 ), . . . are all distinct (and thus the θi belong to distinct orbits). Therefore U does not list-resolve
θ.
8.4
Hilbert series and Hironaka decomposition
In this section we prove Proposition 4.14 on extracting the transcendence degree from the Hilbert series
(as the pole order at t = 1). While this is a general property of finitely generated algebras over a field,
there is an easy proof for invariant rings stemming from a key structural property of such rings called the
Cohen-Macaulay property or Hironaka decomposition.
Theorem 8.3 ([DK15] Section 2.6). The invariant ring R[x]G has the following structure. There exist
homogeneous primary invariants f1 , . . . , fr ∈ R[x]G and homogeneous secondary invariants g1 , . . . , gs ∈
R[x]G such that
• {f1 , . . . , fr } are algebraically independent, and
• any element of R[x]G can be written uniquely as a linear combination of g1 , . . . , gs with coefficients
from R[f1 , . . . , fr ].
The proof can be found in Section 2.6 of [DK15]; note that the only property of the group that is used is
the existence of a Reynolds operator (and so the proof is valid for compact groups).
Proof of Proposition 4.14.
The Hironaka decomposition above implies that the Hilbert series takes the form
Ps
tdeg(gj )
Qr j=1 deg(f )
i )
i=1 (1 − t
(this is equation (2.7.3) in [DK15]). It is now clear that the order of the pole at t = 1 is precisely r. But
we can see as follows that f1 , . . . , fr is a transcendence basis for R[x]G and so r = trdeg(R[x]G ). As in the
proof of Theorem 4.29, since R[x]G is a finitely generated R[f1 , . . . , fr ]-module, every h ∈ R[x]G satisfies a
polynomial with coefficients in R[f1 , . . . , fr ], which is an algebraic dependence among {f1 , . . . , fr , h}.
8.5
Transcendence degree for heterogeneity
In this section we prove Proposition 4.15. To recall the setup, we have G̃ acting on Ṽ with associated
variables x̃. We also have G = G̃K ⋊ SK acting on V = Ṽ ⊕K ⊕ ∆K with associated variables x. Let us also
introduce an intermediate group: G′ = G̃K , acting on V (with associated variables x).
(k)
(k)
Partition the variables x as follows. For k = 1, . . . , K, let x(k) = (x1 , . . . , xp ) be the variables
corresponding to signal k. Let z = (z1 , . . . , zK−1 ) be the variables corresponding to the mixing weights
PK−1
w 1 , . . . , w K−1 . Whenever we refer to zK , this is just shorthand for − k=1 zk .
We first prove a simpler version of the result without the action of SK .
29
Lemma 8.4. Let r̃ = trdeg(R[x̃]G̃ ) and let r = K r̃ + K − 1. Then
′
trdeg(R[x]G ) = r.
′
Proof. To show ‘≥’ we need to exhibit r algebraically independent elements of R[x]G . Letting f1 , . . . , fr̃ be
a transcendence basis for R[x̃]G̃ , it suffices to take
I := {fi (x(k) )}1≤i≤r̃,1≤k≤K ∪ {z1 , . . . , zK−1 }.
′
To show ‘≤’ we first recall that we can obtain a spanning set for the subspace R[x]G
d by applying the
Reynolds operator R (for G′ ) to each degree-d monomial (in the variables x). Such a monomial takes the
form
K
Y
mk (x(k) )
m(x) = M (z)
k=1
where M, mk are monomials. Applying the Reynolds operator yields
R(m(x)) =
E
g1 ,...,gk ∼G̃
M (z)
K
Y
k=1
mk (gk · x(k) ) = M (z)
K
Y
(k)
E mk (gk · x ).
k=1 gk ∼G̃
Note that R(m(x)) is the product of pure invariants, i.e. invariants that only involve variables from a single
one of the blocks x(1) , . . . , x(K) , z. It is clear that I (from above) is a maximal set of algebraically independent
pure invariants. It is now easy to show using the Jacobian criterion (Proposition 4.18) that if any R(m(x))
is added to I, it will no longer be algebraically independent. The result now follows using basic properties
of algebraic independence (Proposition 4.22 and Lemma 4.23).
Proof of Proposition 4.15.
′
Since R[x]G ⊆ R[x]G , it is clear (in light of the above) that trdeg(R[x]G ) ≤ r. Thus we need only to show
trdeg(R[x]G ) ≥ r by demonstrating r algebraically independent invariants. Let e1 , . . . , eK be the elementary
symmetric functions in K variables. With fi as above, we take the invariants
{ek (fi (x(1) ), . . . , fi (x(K) ))}1≤i≤r̃,1≤k≤K ∪ {e2 (z1 , . . . , zK ), . . . , eK (z1 , . . . , zK )}.
Note that e1 (z1 , . . . , zK ) is not included because it is equal to 0. The fact that ek (fi (x(1) ), . . . , fi (x(K) )) are
algebraically independent can be seen because {e1 , . . . , eK } is algebraically independent and {fi (x(k) )}i,k
is algebraically independent. We can see that {ek (z1 , . . . , zK )}k≥2 are algebraically independent as follows.
An algebraic dependence would be a polynomial P such that P (e2 (z2 , . . . , zK ), . . . , eK (z1 , . . . , zK )) (now
P
treating zK as a separate variable) has a root zK = − K−1
k=1 zk and thus has e1 (z1 , . . . , zK ) as factor.
But this contradicts the fact that any symmetric polynomial has a unique representation in terms of the
elementary symmetric polynomials.
8.6
Gröbner bases
In this section we show how to use Gröbner bases to test various algebraic conditions. In particular, we
prove Theorems 4.26 and 4.28. The ideas from this section are mostly standard in the theory of Gröbner
bases; see e.g. [CLO07] for a reference.
Definition 8.5. A monomial order on R[x] is a well-ordering on the set M of all (monic) monomials,
satisfying M ≤ N ⇔ M P ≤ N P for all M, N, P ∈ M. We will say that a monomial order favors a variable
xi if the monomial xi is larger (with respect to the monomial order) than any monomial not involving xi .
We write LM(f ) to denote the leading monomial of a polynomial f , i.e. the monomial occurring in f that is
largest (with respect to the monomial order); LM(f ) does not include the coefficient.
30
Definition 8.6. A Gröbner basis of an ideal I ⊆ R[x] is a finite subset B ⊆ I such that for every f ∈ I
there exists b ∈ B such that LM(f ) is a multiple of LM(b). We call B a reduced Gröbner basis if all its
elements are monic and it has the additional property that for every pair of distinct b, b′ ∈ B, no monomial
occurring in b is a multiple of LM(b′ ).
The following basic facts about Gröbner bases are proved in [CLO07]. A Gröbner basis is indeed a basis, in
that it generates the ideal. Every ideal I ⊆ R[x] has a Gröbner basis, and has a unique reduced Gröbner
basis. Buchberger’s algorithm computes the reduced Gröbner basis of an ideal I = hf1 , . . . , fm i, given a list
of generators fi . (It is not a polynomial-time algorithm, however.)
Suppose we are interested in the relations between polynomials f1 , . . . , fm ∈ R[x]. Introduce additional
variables t = (t1 , . . . , tm ) and consider the ideal I := hf1 (x) − t1 , . . . , fm (x) − tm i ⊆ R[x, t]. Given f1 , . . . , fm
there is an algorithm to compute a Gröbner basis for the elimination ideal
J := hf1 (x) − t1 , . . . , fm (x) − tm i ∩ R[t].
In fact, the algorithm is simply to compute a Gröbner basis for I using a particular monomial order and then
keep only the elements that depend only on t (see Chapter 3 of [CLO07]). The elimination ideal consists
precisely of the polynomial relations among f1 , . . . , fm :
Lemma 8.7. For any polynomial P ∈ R[t] we have: P ∈ J if and only if P (f1 (x), . . . , fm (x)) ≡ 0.
Proof. The direction ‘⇒’ is clear because if we let ti = fi (x) for all i then the generators of I vanish and
so every element of I vanishes. To show the converse, it suffices to show that for any polynomial P ∈ R[t],
P (f1 (x), . . . , fm (x)) − P (t1 , . . . , tm ) ∈ I. This can be shown inductively using the following key idea:
x1 x2 − t1 t2 =
1
1
(x1 − t1 )(x2 + t2 ) + (x2 − t2 )(x1 + t1 )
2
2
and so x1 x2 − t1 t2 ∈ hx1 − t1 , x2 − t2 i.
Generation as an R-algebra. Suppose we want to know whether fm ∈ R[f1 , . . . , fm−1 ]. This is equivalent
to asking whether there exists P ∈ J of the form
P (t) = tm − Q(t1 , . . . , tm−1 )
(9)
for some Q ∈ R[t1 , . . . , tm−1 ]. Suppose that J contains an element P of the form (9). Compute a Gröbner
basis B for J with respect to a monomial order that favors tm . The leading monomial of P is tm so by the
definition of a Gröbner basis there must be an element b ∈ B whose leading monomial divides tm . Since
1 ∈
/ J (by Lemma 8.7), the leading monomial of b is exactly tm and so b takes the form (9). Therefore,
fm ∈ R[f1 , . . . , fm−1 ] if and only if B contains an element of the form (9).
We can now prove Theorem 4.28: to test whether R[f1 , . . . , fm ] = R[x]G , compute generators g1 , . . . , gs
for R[x]G (see Section 8.1) and use the above to test whether each gi is in R[f1 , . . . , fm ].
Generation as a field. Suppose we want to know whether fm ∈ R(f1 , . . . , fm−1 ). This is equivalent to
asking whether fm can be expressed as a rational function of f1 , . . . , fm−1 (with coefficients in R), which is
equivalent (by multiplying through by the denominator) to asking whether there exists P ∈ J of the form
P (t) = tm Q1 (t1 , . . . , tm−1 ) − Q2 (t1 , . . . , tm−1 ) with Q1 ∈
/ J.
(10)
Suppose that J contains an element P of the form (10). Compute a reduced Gröbner basis B for J with
respect to a monomial order that favors tm . It is a basic property of Gröbner bases that P can be written as
X
pi (t)bi (t)
P (t) =
i
where pi ∈ R[t] and bi ∈ B with LM(pi ) ≤ LM(P ) and LM(bi ) ≤ LM(P ). If no bi involves the variable
tm then Q1 ∈ J, a contradiction. Therefore some bj must have degree 1 in tm . Since B is a reduced
Gröbner basis it cannot contain any element of the form (10) with Q1 ∈ J. This completes the proof that
fm ∈ R(f1 , . . . , fm−1 ) if and only if B contains an element of the form (10).
31
Degree of field extension. Consider the setup from Theorem 4.24: given a finite set U = {f1 , . . . , fm } ⊆
R[x]G , we want to compute [FG : FU ] where FU = R(U ) and FG is the field of fractions of R[x]G . We can
assume [FG : FU ] is finite (since we can efficiently test whether this is the case using Proposition 8.2 and
the methods of Section 4.2). Let d be such that R[x]G
≤d generates FG as a field (over R). (It is sufficient
G
for R[x]G
to
generate
R[x]
as
an
R-algebra;
such
a
d
can be computed via Section 8.1. If G is finite then
≤d
d = |G| is sufficient; see Section 4.4.) A generic element of R[x]G
≤d will generate the field extension:
Lemma 8.8. For all but a measure-zero set of α ∈ R[x]G
≤d , FG = FU (α).
This fact is related to the primitive element theorem. We include a proof for completeness.
Proof. The field extension FG /FU is finite and separable (since we’re in characteristic zero), so by the
fundamental theorem of Galois theory, there are only finitely many intermediate fields. (Take the normal
closure of FG /FU ; then the intermediate fields are in bijection with a finite group, and only some of them
lie inside FG .) Let L be the collection of intermediate fields of FG /FU that are proper subfields of FG . We
know R[x]G
≤d is a subspace of FG that generates FG and therefore is not contained by any field in L. This
G
means each field L ∈ L intersects R[x]G
≤d at a proper subspace VL of R[x]≤d . The finite union ∪L∈L VL is a
measure-zero subset of R[x]G
≤d , and any α outside of it satisfies FG = FU (α).
Let α be a generic element of R[x]G
≤d . In light of the above, [FG : FU ] is equal to the smallest positive
integer D for which there exists a relation
QD (f1 , . . . , fm )αD + · · · + Q1 (f1 , . . . , fm )α + Q0 (f1 , . . . , fm ) ≡ 0
for polynomials Qi with QD (f1 , . . . , fm ) 6≡ 0. This can be tested similarly to field generation. Compute
a reduced Gröbner basis B for the elimination ideal J ⊆ R[t1 , . . . , tm , τ ] consisting of the relations among
f1 , . . . , fm , α; use a monomial order that favors τ . Then [FG : FU ] is equal to the smallest positive integer
D for which B contains an element of degree D in τ (or ∞ if B contains no element that involves τ ). This
proves Theorem 4.26.
Remark 8.9. An alternative to using Gröbner bases for the above tasks is to solve a (very large) linear
system in order to find the minimal relation among a set of polynomials. There are bounds on the maximum
possible degree of such a relation (if one exists) [Kay09].
9
Proofs for S 2 registration
9.1
Formula for Hilbert series of R[x]G
We can derive the Hilbert series of R[x]G for S 2 registration using the methods in Section 4.6 of [DK15].
Proposition 9.1. Consider S 2 registration with frequencies F . For |t| < 1, the Hilbert series of R[x]G is
given by
X
H(t) =
Res(f, z)
z∈P
where
f (z) =
z
1 − 21 (z + 1/z)
−z N −2(1 − z)2
hQ
i
= Q
Qℓ
Qℓ
ℓ
m
m
m
2 ℓ∈F
ℓ∈F
m=−ℓ (1 − tz )
m=0 (1 − tz )
m=1 (z − t)
Q
P
with N = 12 ℓ∈F ℓ(ℓ + 1). Here Res(f, z) denotes the residue (from complex analysis) of the function f at
the point z, and P is the set of poles of f (z) inside the unit circle (in C). Namely, P contains t1/m e2πik/m
for all m ∈ {1, 2, . . . , maxℓ∈F ℓ} and for all k ∈ {0, 1, . . . , m − 1}. If N ≤ 1, P also contains 0.
32
Proof. Recall Molien’s formula (Proposition 4.13):
H(t) =
E
g∼Haar(G)
det(I − t ρ(g))−1 .
Note that det(I − t ρ(g)) depends only on the conjugacy class of g. In SO(3), two elements are conjugate
if and only if they rotate by the same angle φ. When g ∼ Haar(SO(3)), the angle φ = φ(g) is distributed
with density function π1 (1 − cos φ) on [0, π] (see e.g. [Sal79]). If g has angle φ, the matrix ρℓ (g) by which
it acts on the irreducible representation Vℓ has eigenvalues e−iℓφ , e−i(ℓ−1)φ , . . . , eiℓφ (see e.g. [Vve01]). The
matrix ρ(g) by which g acts on V = ⊕ℓ∈F Vℓ is block diagonal with blocks ρℓ (g). Using the above we write
an expression for the Hilbert series:
H(t) =
1
π
Z
π
0
1
1 − cos φ
dφ =
Qℓ
imφ )
2π
ℓ∈F
m=−ℓ (1 − te
Q
Z
0
2π
1 − 21 (eiφ + e−iφ )
dφ.
Qℓ
imφ )
ℓ∈F
m=−ℓ (1 − te
Q
Now write this as a complex contour integral around the unit circle in C and apply the residue theorem from
complex analysis to arrive at the result.
9.2
Formula for dimension of R[x]G
d
The dimension of R[x]G can be extracted as the coefficient of td in the Hilbert series from the previous
section, but here we give a different formula based on character theory from representation theory. The
character of a representation ρ : G → GL(V ) (where V is a finite-dimensional real vector space) is the
function χV : G → R defined by χV (g) = tr(ρ(g)).
In our case, using the eigenvalues of ρℓ (g) from the previous section, we have
χVℓ (g) = 1 + 2
ℓ
X
cos(m φ(g))
m=1
P
where φ(g) is the angle of rotation of g. For V = ⊕ℓ∈F Vℓ we then have χV (g) = ℓ∈F χVℓ (g).
As a representation of G = SO(3), R[x]d is (isomorphic to) the dth symmetric power of V , denoted
S d (V ). (This is using the fact that a real representation is isomorphic to its dual.) There is a recursive
formula for the character of S d (V ):
d
χS d (V ) (g) =
1X
χV (g i )χS d−i (V ) (g).
d i=1
This comes from the Newton–Girard formula for expressing complete homogeneous symmetric polynomials
in terms of power sum polynomials.
The representation R[x]d = S d (V ) decomposes as the direct sum of irreducible representations Vℓ . The
subspace of R[x]d consisting of all copies of the trivial representation V0 (the 1-dimensional representation
G
on which every group element acts as the identity) is precisely R[x]G
d . Thus, dim(R[x]d ) is the number of
copies of the trivial representation in the decomposition of R[x]d . This can be computed using characters:
dim(R[x]G
d ) = hχS d (V ) , χV0 i = hχS d (V ) , 1i where hf1 , f2 i := Eg∼Haar(G) [f1 (g)f2 (g)]. Since characters are class
functions (i.e. they are constant on conjugacy classes), we can compute this inner product by integrating
over the angle φ (as in the previous section). This yields the formula stated in Proposition 5.7.
Acknowledgements
ASW is grateful for feedback given by several members of the audience when he presented a preliminary
version of these results at a workshop dedicated to multi-reference alignment, organized as part of the Simons
Collaboration on Algorithms & Geometry.
33
References
[ABL+ 17]
Emmanuel Abbe, Tamir Bendory, William Leeb, João Pereira, Nir Sharon, and Amit Singer.
Multireference alignment is easier with an aperiodic translation distribution. arXiv preprint
arXiv:1710.02793, 2017.
[ADBS16]
Cecilia Aguerrebere, Mauricio Delbracio, Alberto Bartesaghi, and Guillermo Sapiro. Fundamental limits in multi-image alignment. IEEE Trans. Signal Process., 64(21):5707–5722, 2016.
[ADLM84]
Marc Adrian, Jacques Dubochet, Jean Lepault, and Alasdair W McDowall. Cryo-electron
microscopy of viruses. Nature, 308(5954):32–36, 1984.
[APS17]
Emmanuel Abbe, Joao Pereira, and Amit Singer. Sample complexity of the boolean multireference alignment problem. arXiv preprint arXiv:1701.07540, 2017.
[Ban15]
Afonso S. Bandeira. Convex Relaxations for Certain Inverse Problems on Graphs. PhD thesis,
Princeton University, June 2015.
[BBLS17]
Nicolas Boumal, Tamir Bendory, Roy R. Lederman, and Amit Singer. Heterogeneous multireference alignment: a single pass approach. arXiv preprint arXiv:1710.02590, 2017.
[BBM+ 17]
Tamir Bendory, Nicolas Boumal, Chao Ma, Zhizhen Zhao, and Amit Singer. Bispectrum
inversion with application to multireference alignment. arXiv preprint arXiv:1705.00641, 2017.
[BCS15]
Afonso S. Bandeira, Yutong Chen, and Amit Singer. Non-unique games over compact groups
and orientation estimation in cryo-EM. arXiv preprint arXiv:1505.03840, 2015.
[BCSZ14]
Afonso S. Bandeira, Moses Charikar, Amit Singer, and Andy Zhu. Multireference alignment
using semidefinite programming. In Proceedings of the 5th conference on Innovations in theoretical computer science, pages 459–470. ACM, 2014.
[BFB97]
Miguel A Blanco, M Flórez, and M Bermejo. Evaluation of the rotation matrices in the basis
of real spherical harmonics. Journal of Molecular Structure: Theochem, 419(1):19–27, 1997.
[BMS13]
Malte Beecken, Johannes Mittmann, and Nitin Saxena. Algebraic independence and blackbox
identity testing. Information and Computation, 222:2–19, 2013.
[Böh13]
Arno Böhm. Quantum mechanics: foundations and applications. Springer Science & Business
Media, 2013.
[BRW17]
Afonso Bandeira, Philippe Rigollet, and Jonathan Weed. Optimal rates of estimation for
multi-reference alignment. arXiv preprint arXiv:1702.08546, 2017.
[BZS15]
Tejal Bhamre, Teng Zhang, and Amit Singer. Orthogonal matrix retrieval in cryo-electron
microscopy. In Biomedical Imaging (ISBI), 2015 IEEE 12th International Symposium on,
pages 1048–1052. IEEE, 2015.
[CL11]
T. Tony Cai and Mark G. Low. Testing composite hypotheses, Hermite polynomials and
optimal estimation of a nonsmooth functional. Ann. Statist., 39(2):1012–1041, 2011.
[CLO07]
David Cox, John Little, and Don O’Shea. Ideals, varieties, and algorithms: an introduction to
computational algebraic geometry and commutative algebra. Springer New York, 2007.
[Coh13]
J. Cohen. Is high-tech view of hiv too good to be true? Science, 6145(341):443–444, 2013.
[Dia92]
R. Diamond. On the multiple simultaneous superposition of molecular structures by rigid body
transformations. Protein Science, 1(10):1279–1287, October 1992.
34
[DK15]
Harm Derksen and Gregor Kemper. Computational invariant theory. Springer, 2015.
[Dol03]
Igor Dolgachev. Lectures on invariant theory, volume 296. Cambridge University Press, 2003.
[Dom16]
M. Domokos. Degree bounds for separating invariants of abelian groups. arXiv preprint
arXiv:1602.06597, 2016.
[Eli57]
Peter Elias. List decoding for noisy channels. Technical Report 335, Research Laboratory of
Electronics, Massachusetts Institute of Technology, 1957.
[ER93]
Richard Ehrenborg and Gian-Carlo Rota. Apolarity and canonical forms for homogeneous
polynomials. European Journal of Combinatorics, 14(3):157–181, 1993.
[Ete96]
Gábor Etesi. Spontaneous symmetry breaking in the SO(3) gauge theory to discrete subgroups.
Journal of Mathematical Physics, 37(4):1596–1601, 1996.
[Fel68]
W. Feller. An Introduction to Probability Theory and Its Applications, volume 1. Wiley, 1968.
[Hil90]
David Hilbert. Über die Theorie der algebraischen Formen. Mathematische Annalen, 36:473–
531, 1890.
[Hil93]
David Hilbert. Über die vollen Invariantensysteme. Mathematische Annalen, 42:313–370, 1893.
[Huỳ86]
Dũng T. Huỳnh. A superexponential lower bound for Gröbner bases and Church-Rosser commutative Thue systems. Inform. and Control, 68(1-3):196–206, 1986.
[Jon16]
Slavica Jonić. Cryo-electron microscopy analysis of structurally heterogeneous macromolecular
complexes. Computational and structural biotechnology journal, 14:385–390, 2016.
[Kač94]
Victor Kač. Invariant theory [lecture notes]. https://people.kth.se/~laksov/notes/invariant.pdf,
1994.
[Kak09]
Ramakrishna Kakarala. Completeness of bispectrum on compact groups. arXiv preprint
arXiv:0902.0196, 1, 2009.
[Kam80]
Zvi Kam. The reconstruction of structure from electron micrographs of randomly oriented
particles. Journal of Theoretical Biology, 82(1):15–39, 1980.
[Kay09]
Neeraj Kayal. The complexity of the annihilating polynomial. In 24th Annual IEEE Conference
on Computational Complexity (CCC’09), pages 184–193. IEEE, 2009.
[KK10]
Martin Kohls and Hanspeter Kraft. Degree bounds for separating invariants. arXiv preprint
arXiv:1001.5216, 2010.
[LBB+ 17]
Eitan Levin, Tamir Bendory, Nicolas Boumal, Joe Kileel, and Amit Singer. 3D ab initio
modeling in cryo-EM by autocorrelation analysis. arXiv preprint arXiv:1710.08076, 2017.
[LeC73]
L. LeCam. Convergence of estimates under dimensionality restrictions. Ann. Statist., 1:38–53,
1973.
[LNS99]
O. Lepski, A. Nemirovski, and V. Spokoiny. On estimation of the Lr norm of a regression
function. Probab. Theory Related Fields, 113(2):221–253, 1999.
[LS16]
Roy R. Lederman and Amit Singer. A representation theory perspective on simultaneous
alignment and classification. arXiv preprint arXiv:1607.03464, 2016.
[LS17]
Roy R Lederman and Amit Singer. Continuously heterogeneous hyper-objects in cryo-EM and
3-D movies of many temporal dimensions. arXiv preprint arXiv:1704.02899, 2017.
35
[MSS16]
Tengyu Ma, Jonathan Shi, and David Steurer. Polynomial-time tensor decompositions with
sum-of-squares. In Foundations of Computer Science (FOCS), 2016 IEEE 57th Annual Symposium on, pages 438–446. IEEE, 2016.
[Nog16]
Eva Nogales. The development of cryo-EM into a mainstream structural biology technique.
Nature methods, 13(1):24–27, 2016.
[NY83]
A. S. Nemirovsky and D. B. and Yudin. Problem complexity and method efficiency in optimization. A Wiley-Interscience Publication. John Wiley & Sons, Inc., New York, 1983. Translated
from the Russian and with a preface by E. R. Dawson, Wiley-Interscience Series in Discrete
Mathematics.
[Osg07]
Brad Osgood.
Chapter 8:
n-dimensional Fourier Transform.
Lecture Notes for EE 261:
The Fourier Transform and its Applications.
https://see.stanford.edu/materials/lsoftaee261/book-fall-07.pdf, 2007.
[PWB+ 17]
Amelia Perry, Jonathan Weed, Afonso Bandeira, Philippe Rigollet, and Amit Singer. The
sample complexity of multi-reference alignment. arXiv preprint arXiv:1707.00943, 2017.
[PWBM16a] Amelia Perry, Alexander S Wein, Afonso S. Bandeira, and Ankur Moitra. Message-passing algorithms for synchronization problems over compact groups. arXiv preprint arXiv:1610.04583,
2016.
[PWBM16b] Amelia Perry, Alexander S. Wein, Afonso S. Bandeira, and Ankur Moitra. Optimality and
sub-optimality of PCA for spiked random matrices and synchronization. arXiv:1609.05573,
2016.
[PZAF05]
R. G. Pita, M. R. Zurera, P. J. Amores, and F. L. Ferreras. Using multilayer perceptrons to
align high range resolution radar signals. In Artificial Neural Networks: Formal Models and
Their Applications - ICANN 2005, pages 911–916. Springer Berlin Heidelberg, 2005.
[Ros57]
Morris E Rose. Elementary theory of angular momentum. Physics Today, 10:30, 1957.
[Sad89]
B. M. Sadler. Shift and rotation invariant object reconstruction using the bispectrum. In
Workshop on Higher-Order Spectral Analysis, pages 106–111, Jun 1989.
[Sal79]
Eugene Salamin. Application of quaternions to computation with rotations. Technical report,
Working Paper, 1979.
[Sch03]
Alexander Schrijver. Combinatorial optimization: polyhedra and efficiency. Springer Science
& Business Media, 2003.
[Sch12]
S. H. W. Scheres. RELION: implementation of a bayesian approach to cryo-EM structure
determination. J. Struct. Biol., 180(3):519–530, 2012.
[SG92]
Brian M Sadler and Georgios B Giannakis. Shift-and rotation-invariant object reconstruction
using the bispectrum. JOSA A, 9(1):57–69, 1992.
[Sha94]
Igor R. Shafarevich. Basic algebraic geometry. Springer, 1994.
[Sig16]
Fred J. Sigworth. Principles of cryo-EM single-particle image processing. Microscopy, 65(1):57–
67, 2016.
[Sin11]
Amit Singer. Angular synchronization by eigenvectors and semidefinite programming. Applied
and computational harmonic analysis, 30(1):20–36, 2011.
36
[SS11]
Amit Singer and Yoel Shkolnisky. Three-dimensional structure determination from common
lines in cryo-EM by eigenvectors and semidefinite programming. SIAM Journal on Imaging
Sciences, 4(2):543–572, 2011.
[Stu08]
Bernd Sturmfels. Algorithms in invariant theory. Springer Science & Business Media, 2008.
[TS12]
D. L. Theobald and P. A. Steindel. Optimal simultaneous superpositioning of multiple structures with missing data. Bioinformatics, 28(15):1972–1979, 2012.
[Tsy09]
Alexandre B. Tsybakov. Introduction to nonparametric estimation. Springer Series in Statistics.
Springer, New York, 2009. Revised and extended from the 2004 French original, Translated
by Vladimir Zaiats.
[vdV98]
A. W. van der Vaart. Asymptotic statistics, volume 3 of Cambridge Series in Statistical and
Probabilistic Mathematics. Cambridge University Press, Cambridge, 1998.
[VG86]
BK Vainshtein and AB Goncharov. Determination of the spatial orientation of arbitrarily
arranged identical particles of unknown structure from their projections. In Soviet Physics
Doklady, volume 31, page 278, 1986.
[VH87]
Marin Van Heel. Angular reconstitution: a posteriori assignment of projection directions for
3D reconstruction. Ultramicroscopy, 21(2):111–123, 1987.
[Vve01]
Dimitri
Vvedensky.
Chapter
8:
Irreducible
Representations
of
SO(2)
and
SO(3).
Group
theory
course
[lecture
notes].
http://www.cmth.ph.ic.ac.uk/people/d.vvedensky/groups/Chapter8.pdf, 2001.
[ZvdHGG03] J. P. Zwart, R. van der Heiden, S. Gelsema, and F. Groen. Fast translation invariant classification of HRR range profiles in a zero phase representation. Radar, Sonar and Navigation,
IEE Proceedings, 150(6):411–418, 2003.
A
A.1
Spherical harmonics and SO(3) invariants
Spherical harmonics
We follow the conventions of [BFB97]. Parametrize the unit sphere by angular spherical coordinates (θ, φ)
with θ ∈ [0, π] and φ ∈ [0, 2π). (Here θ = 0 is the north pole and θ = π is the south pole.) For integers ℓ ≥ 0
and −ℓ ≤ m ≤ ℓ, define the complex spherical harmonic
Yℓm (θ, φ) = (−1)m Nℓm Pℓm (cos θ)eimφ
with normalization factor
Nℓm =
s
(2ℓ + 1)(ℓ − m)!
4π(ℓ + m)!
where Pℓm (x) are the associated Legendre polynomials
Pℓm (x) =
1
2ℓ ℓ!
(1 − x2 )m/2
dℓ+m 2
(x − 1)ℓ .
dxℓ+m
In the S 2 registration problem we are interested in representing a real-valued function on the sphere, in
which case we use an expansion (with real coefficients) in terms of the real spherical harmonics:
(−1)m
√
m
m > 0,
√2 (Yℓm (θ, φ) + Yℓm (θ, φ)) = 2Nℓm Pℓ (cos θ) cos(mφ)
Yℓ0 (θ, φ) = Nℓ0 Pℓ0 (cos θ)
m = 0,
Sℓm (θ, φ) =
√
m
|m|
(−1)
√ (Yℓ|m| (θ, φ) − Yℓ|m| (θ, φ)) =
2Nℓ|m| Pℓ (cos θ) sin(|m|φ) m < 0.
i 2
37
Here Yℓm is the complex conjugate of Yℓm , which satisfies the identity
Yℓm (θ, φ) = (−1)m Yℓ(−m) (θ, φ).
(11)
−m
m
= (−1)m Nℓm Pℓm .
Above we have also used the identity Pℓ−m = (−1)m (ℓ−m)!
(ℓ+m)! Pℓ , which implies Nℓ(−m) Pℓ
In the cryo-EM problem we are instead interested in representing the Fourier transform of a real-valued
function. Such a function f has the property that if ~r and −~r are antipodal points on the sphere, f (−~r) =
f (~r). For this type of function we use an expansion (with real coefficients) in terms of a new basis of spherical
harmonics:
1
ℓ+m
Yℓ(−m) (θ, φ))
m > 0,
√2 (Yℓm (θ, φ) + (−1)
ℓ
i Yℓ0 (θ, φ)
m = 0,
Hℓm (θ, φ) =
√i (Yℓ|m| (θ, φ) − (−1)ℓ+m Yℓ(−|m|) (θ, φ)) m < 0.
2
One can check that Hℓm (−~r) = Hℓm (~r) using (11) along with the fact Yℓm (−~r) = (−1)ℓ Yℓm (~r) which comes
from Pℓm (−x) = (−1)ℓ+m Pℓm (x).
A.2
Wigner D-matrices
We will mostly work in the basis of complex spherical harmonics Yℓm since the formulas are simpler. The
analogous results for the other bases can be worked out by applying the appropriate change of basis.
Let Vℓ ≃ C2ℓ+1 be the vector space consisting of degree-ℓ complex spherical
harmonics represented
P
in the basis {Yℓm }−ℓ≤m≤ℓ , i.e. v ∈ C2ℓ+1 encodes the spherical harmonic ℓm=−ℓ vm Yℓm . These Vℓ (for
ℓ = 0, 1, 2, . . .) are the irreducible representations of SO(3). Each can also be defined over the real numbers
by changing basis to the real spherical harmonics Sℓm .
A group element g ∈ SO(3) acts on a (spherical harmonic) function f : S 2 → R via (g · f )(x) = f (g −1 x).
The action of g on Vℓ is given by the Wigner D-matrix Dℓ (g) ∈ C(2ℓ+1)×(2ℓ+1) defined as in [BFB97].
We will need the following orthogonality properties of the Wigner D-matrices. First, the standard Schur
orthogonality relations from representation theory yield
′
ℓ (g)D ℓ
Dmk
m′ k′ (g) =
E
g∼Haar(SO(3))
1
1ℓ=ℓ′ 1m=m′ 1k=k′ .
2ℓ + 1
We also have [Ros57]
ℓ
ℓ′
Dmk
(g)Dm
′ k′ (g)
=
′
ℓ+ℓ
X
L
hℓ m ℓ′ m′ |L (m + m′ )ihℓ k ℓ′ k ′ |L (k + k ′ )iD(m+m
′ )(k+k′ ) (g)
L=|ℓ−ℓ′ |
where hℓ1 m1 ℓ2 m2 |ℓ mi is a Clebsch-Gordan coefficient. There is a closed-form expression for these coefficients [Böh13]:
s
(2ℓ + 1)(ℓ + ℓ1 − ℓ2 )!(ℓ − ℓ1 + ℓ2 )!(ℓ1 + ℓ2 − ℓ)!
hℓ1 m1 ℓ2 m2 |ℓ mi =1m=m1 +m2
×
(ℓ1 + ℓ2 + ℓ + 1)!
p
(ℓ + m)!(ℓ − m)!(ℓ1 − m1 )!(ℓ1 + m1 )!(ℓ2 − m2 )!(ℓ2 + m2 )! ×
X
(−1)k
k
k!(ℓ1 + ℓ2 − ℓ − k)!(ℓ1 − m1 − k)!(ℓ2 + m2 − k)!(ℓ − ℓ2 + m1 + k)!(ℓ − ℓ1 − m2 + k)!
where the sum is over all k for which the argument of every factorial is nonnegative.
38
A.3
Moment tensor
Let F be a multi-set of frequencies from {1, 2, . . .} and consider the action of G = SO(3) on V = ⊕ℓ∈F Vℓ .
Recall that we want an explicit formula for Td (x) = Eg [(Π(g · x))⊗d ] with g ∼ Haar(G) (where Π can be the
identity in the case of no projection). We have
Eg [(Π(g · x))⊗d ] = Π⊗d Eg [ρ(g)⊗d ]x⊗d
(where x⊗d is a column vector of length dim(V )d ) and so we need an explicit formula for the matrix
Eg [ρ(g)⊗d ]. Here ρ(g) is block diagonal with blocks Dℓ (g) for ℓ ∈ F . There are no degree-1 invariants
since we have excluded the trivial representation ℓ = 0. For the degree-2 invariants Eg [ρ(g)⊗2 ], consider a
particular block Eg [Dℓ1 (g) ⊗ Dℓ2 (g)] for some pair (ℓ1 , ℓ2 ). The entries in this block can be computed using
0
the above orthogonality relations (and using D00
(g) = 1):
Eg [(Dℓ1 (g))m1 k1 (Dℓ2 (g))m2 k2 ] = 1ℓ1 =ℓ2 1m1 =−m2 1k1 =−k2 hℓ1 m1 ℓ2 m2 |0 0ihℓ1 k1 ℓ2 k2 |0 0i
= 1ℓ1 =ℓ2 1m1 =−m2 1k1 =−k2
(−1)m1 +k1
2ℓ1 + 1
ℓ1 +m1
√
using the special case hℓ1 m1 ℓ2 m2 |0 0i = 1ℓ1 =ℓ2 1m1 =−m2 (−1)
.
2ℓ1 +1
Similarly, for degree 3 we have
Eg [(Dℓ1 (g))m1 k1 (Dℓ2 (g))m2 k2 (Dℓ3 (g))m3 k3 ] =
1|ℓ2 −ℓ3 |≤ℓ1 ≤ℓ2 +ℓ3 1m1 +m2 +m3 =0 1k1 +k2 +k3 =0
A.4
(−1)m1 +k1
hℓ2 m2 ℓ3 m3 |ℓ1 (m2 + m3 )ihℓ2 k2 ℓ3 k3 |ℓ1 (k2 + k3 )i.
2ℓ1 + 1
Projection
Let V = ⊕ℓ∈F Vℓ with F a subset of {1, 2, . . .}. Let Π : V → W be the projection that takes a complex
spherical harmonic function and reveals only its values on the equator θ = π/2. In cryo-EM this projection
is applied separately to each shell (see Section 5.5). Letting L = maxℓ∈F ℓ, the functions b−L , b−L+1 , . . . , bL
(from the circle S 1 to R) form a basis for W , where bm (φ) = eimφ . The projection Π takes the form
Π(Yℓm ) = (−1)m Nℓm Pℓm (0)bm
extended by linearity. By taking a binomial expansion of (x2 − 1)ℓ it can be shown that
(
0
(ℓ + m) odd,
m
Pℓ (0) =
(−1)(ℓ−m)/2
ℓ
(ℓ+m)/2 (ℓ + m)! (ℓ + m) even.
2ℓ ℓ!
(12)
For cryo-EM, if we use the basis Hℓm so that the expansion coefficients are real, the output of the
projection can be expressed (with real coefficients) in the basis
1
imφ
+ (−1)m e−imφ )
m > 0,
√2 (e
1
m = 0,
hm (φ) =
√i (ei|m|φ − (−1)m e−i|m|φ ) m < 0,
2
where the projection Π takes the form
|m|
Π(Hℓm ) = (−1)m Nℓ|m| Pℓ
extended by linearity.
39
(0)hm
A.5
Explicit construction of invariants
Consider the cryo-EM setup (see Section 5.5) with S shells and F frequencies. We will cover S 2 registration as
the special case S = 1 (without projection). Use the basis of complex spherical harmonics, with corresponding
variables xsℓm with 1 ≤ s ≤ S, 1 ≤ ℓ ≤ F , and −ℓ ≤ m ≤ ℓ. One can change variables to Sℓm or Hℓm but
for our purposes of testing the rank of the Jacobian it is sufficient to just work with Yℓm (since the change
of variables has no effect on the rank of the Jacobian).
Recall that in Section A.3 we computed expressions for the matrices Eg [Dℓ1 (g)⊗ Dℓ2 (g)] and Eg [Dℓ1 (g)⊗
ℓ2
D (g) ⊗ Dℓ3 (g)], and in particular they are rank-1. Using this we can explicitly compute the entries of Td (x)
and thus extract a basis for U2T and U3T . We present the results below.
A.5.1
Degree-2 invariants
Without projection, the degree-2 invariants are
I2 (s1 , s2 , ℓ) =
X
1
(−1)k xs1 ℓk xs2 ℓ(−k)
2ℓ + 1
|k|≤ℓ
for s1 , s2 ∈ {1, . . . , S} and ℓ ∈ {1, . . . , F }. Swapping s1 with s2 results in the same invariant, so take s1 ≤ s2
to remove redundancies.
With projection, the degree-2 invariants are
X
Nℓm Nℓ(−m) Pℓm (0)Pℓ−m (0)I2 (s1 , s2 , ℓ)
P2 (s1 , s2 , m) = (−1)m
ℓ≥|m|
with s1 , s2 ∈ {1, . . . , S} and m ∈ {−F, . . . , F }. Negating m or swapping s1 with s2 results in the same
invariant (up to sign) so take s1 ≤ s2 and m ≥ 0 to remove redundancies. Recall the expression (12) for
Pℓm (0).
A.5.2
Degree-3 invariants
Let ∆(ℓ1 , ℓ2 , ℓ3 ) denote the predicate |ℓ2 − ℓ3 | ≤ ℓ1 ≤ ℓ2 + ℓ3 (which captures whether ℓ1 , ℓ2 , ℓ3 can be the
side-lengths of a triangle). Without projection, the degree-3 invariants are
I3 (s1 , ℓ1 , s2 , ℓ2 , s3 , ℓ3 ) =
1
2ℓ1 + 1
X
(−1)k1 hℓ2 k2 ℓ3 k3 |ℓ1 (−k1 )ixs1 ℓ1 k1 xs2 ℓ2 k2 xs3 ℓ3 k3
k1 +k2 +k3 =0
|ki |≤ℓi
for si ∈ {1, . . . , S} and ℓi ∈ {1, . . . , F } satisfying ∆(ℓ1 , ℓ2 , ℓ3 ). There are some redundancies here. First,
permuting the three (si , ℓi ) pairs (while keeping each pair in tact) results in the same invariant (up to scalar
multiple). Also, some of the above invariants are actually zero; specifically, this occurs when (s1 , ℓ1 ) =
(s2 , ℓ2 ) = (s3 , ℓ3 ) with ℓ1 odd, or when (s1 , ℓ1 ) = (s2 , ℓ2 ) 6= (s3 , ℓ3 ) with ℓ3 odd (or some permutation of this
case).
With projection, the degree-3 invariants are
P3 (s1 , m1 , s2 , m2 , s3 , m3 ) =
X
1
2
3
(0)Pℓm
(0)Pℓm
(0)hℓ2 m2 ℓ3 m3 |ℓ1 (−m1 )iI3 (s1 , ℓ1 , s2 , ℓ2 , s3 , ℓ3 )
Nℓ1 m1 Nℓ2 m2 Nℓ3 m3 Pℓm
(−1)m1
1
2
3
ℓ1 ,ℓ2 ,ℓ3 : ∆(ℓ1 ,ℓ2 ,ℓ3 )
for si ∈ {1, . . . , S} and mi ∈ {−F, . . . , F } such that m1 + m2 + m3 = 0. There are again redundancies under
permutation: permuting the three (si , mi ) pairs results in the same invariant. Negating all three m’s also
results in the same invariant. There are additional non-trivial linear relations (see Section A.6 below).
40
A.6
A.6.1
Counting the number of invariants
S 2 registration
For the case of S 2 registration (S = 1) the above degree-2 and degree-3 invariants without projection (with
G
redundancies removed as discussed above) form a basis for R[x]G
2 ⊕ R[x]3 (although we have not made this
rigorous). Thus, counting these invariants gives a combinatorial analogue of Proposition 5.7.
A.6.2
Cryo-EM
T
In this section we give a formula for trdeg(U≤3
) for (heterogeneous) cryo-EM (with projection), valid for all
K ≥ 1, S ≥ 1 and F ≥ 2. The formula is conjectural but has been tested (via the Jacobian criterion) for
various small values of K, S, F .
The number of algebraically independent degree-2 invariants turns out to be the number of distinct I2
invariants (i.e. without projection), since the projected invariants P2 are linear combinations of these. The
number of such invariants is 21 S(S + 1)F .
For degree 3, things are more complicated because the projected invariants P3 have smaller dimension
than the I3 . We start by counting the number of distinct (up to scalar multiple) P3 invariants. To this
end, let X (S, F ) be the set of equivalence classes of tuples (s1 , m1 , s2 , m2 , s3 , m3 ) with si ∈ {1, . . . , S} and
mi ∈ {−F, . . . , F }, modulo the relations
(s1 , m1 , s2 , m2 , s3 , m3 ) ∼ (s2 , m2 , s1 , m1 , s3 , m3 ) ∼ (s1 , m1 , s3 , m3 , s2 , m2 )
(s1 , m1 , s2 , m2 , s3 , m3 ) ∼ (s1 , −m1 , s2 , −m2 , s3 , −m3 )
(permutation)
(negation).
There are some non-trivial linear relations among the distinct P3 invariants, which we must now account
for. The number of such relations is
E(S) := 2S + 4S(S − 1) + S(S − 1)(S − 2).
This can be broken down as follows. For each k ∈ {1, 2, 3} there are 2k relations for each size-3 multisubset {s1 , s2 , s3 } of {1, . . . , S} with exactly k distinct elements. (We do not currently have a thorough
understanding of what exactly the linear relations are, but we have observed that the above pattern holds.)
We can now put it all together and state our conjecture. We will also use the formula (3) for trdeg(R[x]G ),
extended to the heterogeneous case via Proposition 4.15.
Conjecture A.1. Consider cryo-EM with F ≥ 2 frequencies.
• trdeg(R[x]G ) = K[S(F 2 + 2F ) − 3] + K − 1,
• dim(U2T ) = 12 S(S + 1)F ,
• dim(U3T ) = |X (S, F )| − E(S),
• generic list recovery is possible at degree 3 if and only if dim(U2T ) + dim(U3T ) ≥ trdeg(R[x]G ).
When S and F are large, the dominant term in dim(U2T ) + dim(U3T ) is |X (S, F )| ≈ S 3 F 2 /4 and so generic
list recovery is possible when (roughly) K ≤ S 2 /4.
B
Unique recovery for the regular representation
Let G be a finite abelian group. Let V be the regular representation of G over R, i.e. V = R|G| with
basis indexed by the elements of G, and the action of G permutes the basis elements vg according to group
multiplication: h · vg = vhg . Note that for the cyclic group G = Z/p, this is precisely the MRA problem. We
will show that with this setup, generic unique recovery is possible at degree 3:
41
Theorem B.1. Let G be a finite abelian group and let V be the regular representation of G over R. Then
T
R(U≤3
) is equal to the field of fractions of R[x]G .
Recall that the equality of the two fields above implies generic unique recovery at degree 3 by Corollary 4.25.
Generic unique recovery at degree 3 is actually true in the more general setting where V is the regular
representation over R of any finite group [Kak09].
T
Proof. It is sufficient to show C(U≤3
) = C(R[x]G ). Let p = |G| = dim(V ). The polynomial algebra on V
gets an action of G in the usual way and the degree-1 component is isomorphic to the regular representation.
Over C, the regular representation decomposes as a direct sum of all the characters χ(1) , . . . , χ(p) of G
(each of which is a group homomorphism G → C× = C r {0}). (Here we have used the fact that G is finite
and abelian, and so its irreducible representations are 1-dimensional.)
b = {χ(1) , . . . , χ(p) } be the set of characters. For each χ ∈ G
b let yχ ∈ C[V ] be the associated
Let G
eigenvector (in the degree-1 component of the polynomial algebra) so that g · yχ = χ(g)yχ for all g ∈ G.
b forms an abelian group (the character group or Pontryagin dual of G) under pointwise multiplication:
G
(χ1 χ2 )(g) = χ1 (g)χ2 (g).
b yχ yχ−1 is a degree-2 invariant, and for any χ1 , χ2 ∈ G,
b yχ1 yχ2 y(χ χ )−1 is a
Observe that for any χ ∈ G,
1 2
b we have
degree-3 invariant. For each χ ∈ G
Y
Y
yχ′ yχ′−1 =
yχp
yχ yχ′ y(χχ′ )−1
b
χ′ ∈G
b
χ′ ∈G
and so yχp lies in the field generated by the invariants of degree ≤ 3.
The field extension C(yχ(1) , . . . , yχ(p) )/C(yχp (1) , . . . , yχp (p) ) is Galois with Galois group H = (Z/p)p ; the
generator of the ith copy of Z/p acts by multiplying yχ(i) by a fixed primitive pth root of unity. (Recall that
the Galois group is the automorphisms of the larger field that fix the smaller field pointwise.)
By the fundamental theorem of Galois theory, intermediate extensions of the above Galois extension are
in bijection with subgroups of H. The correspondence maps a field to the group that fixes it (pointwise) and
maps a group to the field that it fixes. The field C(R[x]G ) is an intermediate extension with corresponding
T
group G. The field C(U≤3
) is also an intermediate extension, and we will show that its corresponding group
T
is also G, thus proving C(U≤3
) = C(R[x]G ) as desired.
T
T
Clearly G fixes C(U≤3 ) (since U≤3
are invariants), so it remains to show that there are no other automorT
phisms of C(yχ(1) , . . . , yχ(p) ) that fix U≤3
. Let φ ∈ H. As an element of (Z/p)p , we can write φ = (φ1 , . . . , φp ).
If χ = χ(i) we will also write φχ for φi . If φ fixes the degree-2 invariant yχ yχ−1 , we must have φχ−1 = φ−1
χ . If
.
In
particular,
=
φ
φ also fixes the degree-3 invariant yχ1 yχ2 y(χ1 χ2 )−1 , we must have φχ1 φχ2 = φ−1
χ
χ
−1
1 2
(χ1 χ2 )
b → C× given by χ 7→ φχ is a group homomorphism. Thus, any φ fixing U T can be identified
the map G
≤3
b
b , which is isomorphic to G; in particular there are only p
with an element of the Pontryagin double dual G
such elements φ, which completes the proof.
42
| 10math.ST
|
Joint Uplink-Downlink Cell Associations for
Interference Networks with Local Connectivity
Manik Singhal and Aly El Gamal
arXiv:1701.07522v3 [cs.IT] 22 Feb 2018
ECE Department, Purdue University
Email: {msingha,elgamala}@purdue.edu
Abstract—We study information theoretic models of interference networks that consist of K Base Station (BS) Mobile Terminal (MT) pairs. Each BS is connected to the MT
carrying the same index as well as L following MTs, where the
connectivity parameter L ≥ 1. We fix the value of L and study
large networks as K goes to infinity. We assume that each
MT can be associated with Nc BSs, and these associations are
determined by a cloud-based controller that has a global view
of the network. An MT has to be associated with a BS, in
order for the BS to transmit its message in the downlink, or
decode its message in the uplink. In previous work, the cell
associations that maximize the average uplink-downlink per
user degrees of freedom (puDoF) were identified for the case
when L = 1. Further, when only the downlink is considered, the
problem was settled for all values of L when we are restricted to
use only zero-forcing interference cancellation schemes. In this
work, we first propose puDoF inner bounds for arbitrary values
of L when only the uplink is considered, and characterize
the uplink puDoF value when only zero-forcing schemes are
allowed. We then introduce new achievable average uplinkdownlink puDoF values. We show that the new scheme is
optimal for the range when Nc ≤ L2 when we restrict our
attention to zero forcing schemes. Additionally we conjecture
that the having unity puDoF during uplink is optimal when
Nc ≥ L.
I. I NTRODUCTION
The fifth generation of cellular networks is expected
to bring new paradigms to wireless communications, that
exploit recent technological advancements like cloud computing and cooperative communication (also known as Coordinated Multi-Point or CoMP). In particular, the rising
interest in Cloud Radio Access Networks (C-RAN) (see
e.g., [1]-[6]) holds a promise for such new paradigms. These
paradigms require new information theoretic frameworks to
identify fundamental limits and suggest insights that are
backed by rigorous analysis. The focus of this work is to
identify associations between cell edge mobile terminals
and base stations, that maximize the average rate across
both uplink and downlink sessions, while allowing for
associating one mobile terminal with more than one base
station and using cooperative transmission and reception
schemes between base stations in the downlink and uplink
sessions, respectively. With a cloud-based controller, optimal
decisions for these associations can take into account the
whole network topology, with the goal of maximizing a sum
rate function.
Cloud-based CoMP communication is a promising new
technology that could significantly enhance the rates of cell
edge users (see [7] for an overview of CoMP). In [8], an
information theoretic model was studied where cooperation
was allowed between transmitters, as well as between receivers (CoMP transmission and reception). CoMP transmission and reception in cellular networks are applicable in the
downlink and uplink, respectively. The model in [8] assumed
that each message can be available at Mt transmitters and
can be decoded through Mr received signals. It was shown
that full Degrees of Freedom (DoF) can be achieved if
Mt + Mr ≥ K + 1, where K is the number of transmitterreceiver paris (users) in the network.
Recently in [9], alternative frameworks for cooperation
in both downlink and uplink were introduced. The new
frameworks are based on the concept of message passing
between base stations. In the downlink, quantized versions
of the analog transmit signals are being shared between
base station transmitters. The supporting key idea is that
information about multiple messages can be shared from
one transmitter to another with the cost of sharing only
one whole message (of the order of log P , where P is
the transmit power), if we only share information needed
to cancel the interference caused by the messages at unintended receivers, through dirty paper coding (see [10]).
In the uplink, decoded messages are shared from one base
station receiver to another, where they are used to cancel
interference. It was shown in [9] that there is a duality
in this framework between schemes that are used in the
downlink and those that are used for the uplink, with the
clear advantage that the same backhaul infrastructure can
be used to support both scenarios.
In this work, we first characterize the puDoF of message passing decoding in the uplink of locally connected
interference networks when Nc < L2 . We then consider the
problem of jointly optimizing the assignment of messages
over the backhaul to maximize the average puDoF across
both downlink and uplink sessions. We assume that each
base station can be associated with Nc mobile terminals, and
that an association is needed whenever a mobile terminal’s
message is used by a base station in either the downlink
or the uplink. This usage of a message could be either for
delivering the message in downlink, decoding the message
in uplink, or for interference cancellation. This problem
was first considered in [13], where the average puDoF was
characterized for the case when L = 1. Here, we consider
general values of L, and first show how our new result
for the uplink settles the average puDoF problem when
Nc ≤ L2 . We then tackle this problem when Nc > L,
by fixing the uplink scheme to the optimal uplink-only
scheme, that associates each mobile terminal with the L + 1
base stations connected to it, and characterize the optimal
downlink scheme under this constraint. The intuition behind
this step is that full DoF is achieved in the uplink when
Nc > L through associating each mobile terminal with all
L + 1 base stations connected to it: any change in that cell
association is expected to decrease the uplink puDoF with
a factor greater than the gain achieved for the downlink
puDoF.
When considering this work, it is important to note that
the assumptions in a theoretical framework need not reflect
directly a practical setting, but are rather used to define a
tractable problem whose solution can lead to constructive
insights. For example, it was shown in [11] that imposing
a downlink backhaul constraint where each message can be
available at a specified maximum number of transmitters
(maximum transmit set size constraint), can lead to solutions
that are also useful to solve the more difficult and more
relevant to practice problem, where an average transmit
set size constraint is used instead of the maximum. Also,
in [12], it was shown that solutions obtained for the locally
connected network models, that are considered in this work,
can be used to obtain solutions for the more practical cellular
network models, by viewing the cellular network as a set
of interfering locally connected subnetworks and designing
a fractional reuse scheme that avoids interference across
subnetworks.
A. Prior Work
In [13], the considered problem was studied for Wyner’s
linear interference networks (channel model was introduced
in [14]). The optimal message assignment and puDoF value
were characterized. Linear networks form the special case
of our problem when L = 1. Here, all our results are
for general values of the connectivity parameter L. Also,
in [11], the downlink part of our problem was considered,
and the optimal message assignment (cell association) and
puDoF value were characterized for general values of the
connectivity parameter L, when we restrict our attention to
zero-forcing (or interference avoidance) scheme.
B. Document Organization
In Section II, we present the problem setup. In Section III,
we discuss previous work on zero-forcing CoMP transmission schemes for the downlink. We then present bounds
for the puDoF of the uplink in Section IV, and prove the
converse in Sections V and VI. In Section VII, we present
new achievable puDoF values when the average of the uplink
and downlink is considred. We finally present concluding
remarks in Section IX.
II. S YSTEM M ODEL AND N OTATION
For each of the downlink and uplink sessions, we use the
standard model for the K−user interference channel with
single-antenna transmitters and receivers,
Yi (t) =
K
X
Hi,j (t)Xj (t) + Zi (t),
(1)
j=1
where t is the time index, Xj (t) is the transmitted signal of
transmitter j, Yi (t) is the received signal at receiver i, Zi (t)
is the zero mean unit variance Gaussian noise at receiver i,
and Hi,j (t) is the channel coefficient from transmitter j to
receiver i over time slot t. We remove the time index in the
rest of the paper for brevity unless it is needed. The signals
Yi and Xi correspond to the receive and transmit signals
at the ith base station and mobile terminal in the uplink,
respectively, and the ith mobile terminal and base station in
the downlink, respectively. For consistency of notation, we
will always refer to Hi,j as the channel coefficient between
mobile terminal i and base station j.
A. Channel Model
We consider the following locally connected interference
network. The mobile terminal with index i is connected to
base stations {i, i − 1, · · · , i − L}, except the first L mobile
terminals, which are connected only to all the base stations
with a similar or lower index. More precisely,
Hi,j = 0 iff i ∈
/ {j, j + 1, · · · , j + L}, ∀i, j ∈ [K],
(2)
and all non-zero channel coefficients are drawn from a
continuous joint distribution. Finally, we assume that global
channel state information is available at all mobile terminals
and base stations.
B. Cell Association
For each i ∈ [K], let Ci ⊆ [K] be the set of base
stations, with which mobile terminal i is associated, i.e.,
those base stations that carry the terminal’s message in the
downlink and will have its decoded message for the uplink.
The transmitters in Ci cooperatively transmit the message
(word) Wi to mobile terminal i in the downlink. In the
uplink, one of the base station receivers in Ci will decode Wi
and pass it to the remaining receivers in the set. We consider
a cell association constraint that bounds the cardinality of
the set Ci by a number Nc ; this constraint is one way to
capture a limited backhaul capacity constraint where not all
messages can be exchanged over the backhaul.
|Ci | ≤ Nc , ∀i ∈ [K].
(3)
We would like to stress on the fact that we only allow full
messages to be shared over the backhaul. More specifically,
splitting messages into parts and sharing them as in [15], or
sharing of quantized signals as in [9] is not allowed.
C. Degrees of Freedom
Let P be the average transmit power constraint at each
transmitter, and let Wi denote the alphabet for message
Wi . Then the rates Ri (P ) = log n|Wi | are achievable if
the decoding error probabilities of all messages can be
simultaneously made arbitrarily small for a large enough
coding block length n, and this holds for almost all channel
realizations. The degrees of freedom di , i ∈ [K], are defined
i (P )
as di = limP →∞ Rlog
P . The DoF region D is the closure
of the set of all achievable DoF tuples. The total number of
degrees of freedom (η) is the maximum value ofPthe sum of
the achievable degrees of freedom, η = maxD i∈[K] di .
For a K-user locally connected with connectivity parameter L, we define η(K, L, Nc ) as the best achievable η on
average taken over both downlink and uplink sessions over
all choices of transmit sets satisfying the backhaul load
constraint in (3). In order to simplify our analysis, we define
the asymptotic per user DoF (puDoF) τ (L, Nc ) to measure
how η(K, L, Nc ) scales with K while all other parameters
are fixed,
η(K, L, Nc )
.
(4)
K→∞
K
We further define τD (L, Nc ) and τU (L, Nc ) as the puDoF
when we optimize only for the downlink and uplink session,
respectively.
τ (L, Nc ) = lim
IV. U PLINK -O NLY S CHEME
D. Interference Avoidance Schemes
We consider in this work the class of interference avoidance schemes, where every receiver is either active or
inactive. An active receiver can observe its desired signal
with no interference. In the downlink, we are considering
cooperative zero-forcing where a message’s interference is
cancelled over the air through cooperating transmitters. In
the uplink, we are considering message passing decoding
where a decoded message is passed through a cooperating
receiver to other receivers wishing to remove the message’s
interference.
We add the superscript zf to the puDoF symbol when
we impose the constraint that the coding scheme that can
be used has to be a zero-forcing scheme. For example,
τUzf (L, Nc ) denotes the puDoF value when considering
only the uplink and impose the restriction to zero-forcing
schemes.
III. P RIOR W ORK : D OWNLINK -O NLY S CHEME
In [11], the considered setting was studied for only downlink transmission. When restricting our choice of coding
scheme to zero-forcing schemes, the puDoF value was
characterized as,
zf
τD
(L, Nc ) =
2Nc
,
2Nc + L
transmitters in each subnetwork are inactive to avoid intersubnetwork interference. The zero-forcing scheme aims to
deliver 2Nc messages free of interference in each subnetwork, so that the acheived puDoF value is as in (5). In order
to do that with a cooperation constraint that limits each message to be available at Nc transmitters, we create two Multiple Input Single Output (MISO) Broadcast Channels (BC)
within each subnetwork; each with Nc transmitter-receiver
pairs, and ensure that interference across these channels is
eliminated. We now discuss the cell association in the first
subnetwork, noting that the remaining subnetworks follow
an analogous pattern. The first MISO BC consists of the first
Nc transmitter-receiver pairs. For each i ∈ {1, 2, · · · , Nc },
message Wi is associated with base stations with indices in
the following set, Ci = {i, i+1, · · · , Nc }. The second MISO
BC consists of the Nc transmitters with indices in the set
{Nc +1, Nc +2, · · · , 2Nc } and the Nc receivers with indices
in the set {Nc + L + 1, Nc + L + 2, · · · , 2Nc + L}. Note that
the middle L receivers in each subnetwork are deactivated
to eliminate interference between the two MISO BCs. For
each i ∈ {Nc + L + 1, Nc + L + 2, · · · , 2Nc + L}, message
Wi is associated with transmitters that have indices in the
set Ci = {i−L, i−L−1, · · · , Nc +1}. It was shown in [11]
that the puDoF value of (5) achieved by this scheme is that
best achievable value in the downlink using the imposed
cooperation constraint and zero-forcing schemes.
(5)
and the achieving cell association was found to be the
following. The network is split into subnetworks; each with
consecutive 2Nc + L transmitter-receiver pairs. The last L
We discuss in this section backhaul designs that optimize
only the uplink rate, and consider only zero-forcing coding schemes. More precisely, we show that the following
theorem holds.
Theorem 1: The asymptotic puDoF for the uplink when
considering message passing schemes is characterised by the
following equation:
L + 1 ≤ Nc ,
1
zf
Nc +1
L
(6)
τU (L, Nc ) =
L+2
2 ≤ Nc ≤ L,
2Nc
L
1 ≤ Nc ≤ 2 − 1.
2Nc +L
The cell association that is used to achieve the above
is as follows. When Nc ≥ L + 1, each mobile terminal
is associated with the L + 1 base stations connected to
it. The last base station, with index K, in the network
decodes the last message and then passes it on to the L
other base stations connected to the K th mobile terminal,
eliminating all interference caused by that mobile terminal.
Each preceding base station then decodes its message and
passes it on to the other base stations, eliminating the
interference caused by the message. Thus, one degree of
freedom is achieved for each user.
In the second range L2 ≤ Nc ≤ L, the cell association
c +1
that is used to achieve a puDoF value of NL+2
is as follows.
The network is split into subnetworks, each with consecutive
L + 2 transmitter-receiver pairs. In each subnetwork, the
last Nc + 1 words are decoded. For each i ∈ {L + 2, L +
1, · · · , L + 2 − Nc + 1}, message Wi is associated with base
stations {i, i−1, · · · , L+2−Nc +1} ⊆ Ci . Thus the last Nc
words are decoded. The base stations with indices in the set
{2, 3, · · · , L + 2 − Nc } are inactive as there is interference
from the last transmitter in the subnetwork which cannot
be eliminated. The first base station decodes WL+2−Nc . To
eliminate the interference caused by the transmitters in the
set S = {L + 2 − Nc + 1, L + 2 − Nc + 2, · · · , L + 1} at the
first base station of the subnetwork, we add the first base
station to each Ci , ∀i ∈ S. Now for messages with indices
in the set S, we have used βi = 2 + i − (L + 2 − Nc + 1)
associations up to this point; the factor of two comes from
the base station resolving Wi and the first base station of
the subnetwork. But each transmitter with indices in the
set S\{L + 1} also interferes with the subnetwork directly
preceding this subnetwork. ∀i ∈ S\{L + 1}, the message
Wi interferes with the bottom L + 1 − i base stations of
the preceding subnetwork, which is precisely the number
of associations left for the respective message i.e. Nc −
βi = L + 1 − i, thus inter-subnetwork interference can be
eliminated at those base stations.
In the third range 1 ≤ Nc ≤ L2 −1, the cell association that
c
is used to achieve the lower bound of 2N2N
is similar to the
c +L
one described in Section III for the downlink. The network is
split into disjoint subnetworks; each with consecutive 2Nc +
L transmitter-receiver pairs. For the uplink, we consider
two sets of indices for transmitters AT = {1, 2, · · · , Nc }
and BT = {Nc + L + 1, Nc + L + 2 · · · , 2Nc + L}, and
corresponding sets of receivers AR = {1, 2, · · · , Nc } and
BR = {Nc + 1, Nc + L + 2 · · · , 2Nc }. For each i ∈ AT ,
the message Wi is associated with the receivers receiving
it in AR . Receiver i decodes Wi and the other associations
in Ci exist for eliminating interference. Similarly For each
j ∈ BT , the message Wj is associated with the receivers
receiving it in BR , but now receiver j − L decodes Wj and
the other associations in Cj are for eliminating interference.
We observe that if we were not restricted to the zeroforcing coding scheme then for the third range, we could
achieve 12 puDoF using asymptotic interference alignment [16], which is higher than the value achieved by zeroforcing. The next sections complete the proof of Theorem
1
V. C ONVERSE P ROOF WHEN
L
2
≤ Nc ≤ L
In this section, we provide a converse proof for the second
range of (6). More precisely, we show that the following
holds.
τUzf (L, Nc ) =
Nc + 1
,
L+2
L
≤ Nc ≤ L.
2
(7)
We start by proving the case when Nc = L, The optimal
zero-forcing puDoF for the uplink can be characterised as:
τUzf (L, L) =
L+1
.
L+2
(8)
We begin by dividing the network into subnetworks of
L + 2 consecutive transmitters-receiver pairs. We observe
that in any subnetwork, if we have Nc + 1 = L + 1 consecutive active receivers (base stations), then the transmitter
connected to all these receivers must be inactive, because a
message’s interference cannot be canceled at Nc or more receivers. Let ΓBS be the set of subnetworks where all Nc +2
receivers are active, and ΦBS be the set of subnetworks with
at most Nc active receivers. Similarly, let ΓM T and ΦM T
be the subnetworks with Nc + 2 active transmitters and at
most Nc active transmitters, with respect to order. To be
able to achieve a higher puDoF than (8), it must be true that
both conditions hold: |ΓBS | > |ΦBS | and |ΓM T | > |ΦM T |.
Now note that for any subnetwork that belongs to ΓBS , at
most Nc transmitters will be active, because the interference
caused by any message cannot be canceled at Nc or more
receivers. Hence ΓBS ⊆ ΦM T . Further, the same logic
applies to conclude that for any subnetwork with Nc + 1
active receivers, the number of active transmitters is at
most Nc + 1, and hence ΓM T ⊆ ΦBS . It follows that
if |ΓBS | > |ΦBS |, then |ΓM T | < |ΦM T |, and hence the
statement in (8) is proved.
To aid in the next step we define MT-BS pairs (mi , bj ) as
decoding pairs if Wi is decoded at base station j. To prove
c +1
when L2 ≤ Nc < L, we use the
that τUzf (L, Nc ) = NL+2
following lemmas :
Lemma 1: For any zero-forcing scheme, one of the
following is true for any two decoding pairs (mi1 , bj1 )
and (mi2 , bj2 ): j2 ∈
/ {i1 , i1 − 1, · · · , i1 − L} or j1 ∈
/
{i2 , i2 − 1, · · · , i2 − L}.
Proof: If the claim were not true, i.e. j2 ∈ {i1 , i1 −
1, · · · , i1 − L} and j1 ∈ {i2 , i2 − 1, · · · , i2 − L} then
Wi1 and Wi2 would interfere with one another and could
not be decoded using the zero-forcing scheme. This is a
consequence of the work done in [17].
Lemma 2: For any set L ⊆ [K] of L + 1 consecutive
indices, a maximum of Nc mobile terminals with indices in
L can be decoded at base stations with indices in L for any
zero-forcing scheme.
Proof: We prove this claim by contradiction. If Nc + 1
or more mobile terminals with indices in L are decoded
at base stations with indices in L, then at least one of the
mobile terminals would be associated with more than Nc
base stations. This violates the constraint in (3).
From Lemma 1, we have the following corollary:
Corollary 1: For any two decoding pairs (mi1 , bj1 ) and
(mi2 , bj2 ) in a zero-forcing scheme, if i1 > i2 then j1 > j2
and vice versa.
Immediately from Lemma 2 we have that subnetwork only
decoding, i.e. transmissions from a subnetwork are decoded
in the same subnetwork, can only decode at most Nc + 1
words in each subnetwork of L+2 consecutive BS-MT pairs.
Our proof will be based on the concept that to break the
inner bound described in (6), at least one subnetwork of L+2
consecutive MT-BS pairs must have more than Nc +1 active
mobile terminals. And all such subnetworks must borrow
base stations from the subnetwork above it to decode words
corresponding to its own mobile terminals. This happens
because for a consecutive set of L + 2 mobile terminals, the
only base stations that can help decode their transmissions
are the corresponding base stations or the other L base
stations with preceding indices that are connected to the
set.
To aid in the writing we define : αk = (L + 2) × (k − 1),
here αk denotes the first index of each subnetwork Lk .
In this sense Lk is topologically below Lk−1 , i.e. mobile
terminals from Lk are connected to some base stations in
Lk−1 . Additionally M Ti denotes mobile terminal i, and
BSj denotes base station j
We use the above lemmas and definitions to define a best
case scenario for inter-subnetwork interference. A best case
scenario is where the interference from one subnetwork’s
(e.g., Lk ) mobile terminals to another subnetwork’s (e.g.,
Lk−1 ) base stations is focused on the bottom most base
stations. This is defined as the best case scenario because
from Lemma 1, we know that for Lk−1 ’s own mobile
terminals to be decoded in Lk−1 , we need base stations
that are indexed outside the range of the interference from
the mobile terminals of Lk . We also define that if there
exists decoding pairs (mi , bj ) such that i ≥ αk and j < αk ,
i.e. the mobile terminal is in Lk and the base station is in
Lk−1 , then Lk borrows a base station from Lk−1 . Similarly,
if there exists certain consecutive base stations in Lk−1
indexed by (αk − µ, αk − µ + 1, · · · αk − 1) such that no
words can be decoded here in the zero-forcing scheme due
to the cooperation constraint being met in Lk , we say that
Lk blocks µ base stations in Lk−1 .
We introduce two new variables x and δ. Here x defines
the number of extra mobile terminals (beyond Nc +1) active
in a subnetwork of L + 2 consecutive mobile terminals and
base stations, and δ defines the number of base stations
that Lk borrows from Lk−1 to help decode words from
Lk . When L2 ≤ Nc ≤ L it follows from the network
topology and the defined cooperation constraint that we have
1 ≤ x, δ ≤ Nc .
c +1
when L >
We want to show that τUzf (L, Nc ) ≤ NL+2
L
Nc ≥ 2 . It follows from the pigeonhole principle that to
break this bound, there must be at least one subnetwork (say
Lk ) where we have Nc +1+x mobile terminals active. Now
by Lemma 2, we have that Lk must borrow at least x base
stations from Lk−1 , thus x ≤ δ ≤ Nc . We now consider
possible cases for the value of δ.
When δ = 1, thus x = 1, so Lk has Nc + 2 active mobile
terminals. As Lk is borrowing one base station, say base
station j, Nc + 1 words must have been decoded in Lk . By
Lemma 2, there exists at least one decoding pair (mi , bn )
where i, n ≥ αk , such that bn is not connected to the highest
indexed active mobile terminal in Lk . Due to the size of the
subnetwork, this forces n = αk . Hence, mobile terminal i’s
transmission is decoded at the first base station of Lk . By
Lemma 1, this implies that j ∈
/ {i, i − 1, ...i − L}. It follows
that the best case scenario occurs when i = αk + (L + 2 −
(Nc + 1)), making j ≤ αk − Nc = αk−1 + L + 2 − Nc .
Let the number of available base stations left in Lk−1 be
θ. As Nc ≥ L2 , it follows that θ ≤ L + 2 − Nc ≤ Nc + 2.
Additionally, due to the borrowed base station, the number
of associations allowed for M Tαk−1 +L+1 (the last mobile
terminal in Lk−1 ) has effectively reduced by one. From
Lemma 2, we have that a maximum of Nc mobile terminals
can be decoded in Lk−1 . It follows that either the average
number of active mobile terminals over the two subnetworks
is still Nc + 1 per subnetwork, or Lk−1 will have to borrow
at least one base station from Lk−2 . We do not consider
the former case, as we just have to restart our argument
from Lk−2 because all subnetworks with higher indexes
will have an average of Nc + 1 active mobile terminals per
subnetwork. Hence, we only consider the latter case where
Lk−1 borrows at least one base station from Lk−2 . As base
station αk−1 is being used in Lk−1 , the lowest possible
indexed base station that Lk−1 borrows from Lk−2 is base
station αk−2 + (L + 2 − Nc ). Therefore the argument for
Lk−2 borrowing base stations from Lk−3 is exactly the same
as the argument shown for Lk−1 borrowing from Lk−2 . It
follows that this borrowing will continue till either we stop
borrowing at some subnetwork Li , where i < k, or L1
needs to borrow at least one more base station, which is
not possible. If Li does not borrow from Li−1 , we have
that Li and Lk have at most Nc and Nc + 1 + 1 active
mobile terminals, respectively, and all other subnetworks
between them have at most Nc + 1 active mobile terminals,
resulting in an average of Nc + 1 active mobile terminals
per subnetwork over these k − i subnetworks. Thus we can
discard them as they do not break the inner bound and
start the same argument over from Li−1 . If we continue
borrowing till L1 , we have that L1 and Lk have at most
Nc and Nc + 1 + 1 active mobile terminals respectively and
all other subnetworks have at most Nc + 1 active mobile
terminals, resulting that the average number of active mobile
terminals over the whole network is Nc + 1 per subnetwork
c +1
. This presents the
which implies that τUzf (L, Nc ) ≤ NL+2
simplest case for our iterative argument.
When δ > 1, we have a similar argument as described
in the previous paragraph. By Lemma 1, we have that
the borrowed base stations in Lk−1 will have to send the
associations downwards, i.e. the lowest indexed borrowed
base station in Lk−1 will have to be exclusively connected
to the lowest indexed active mobile terminal of Lk . As
the index of the lowest active mobile terminal in Lk is
at most αk + (L + 2 − (Nc + 1 + x)) − 1, we have
that the index of the lowest borrowed base station in
Lk−1 is αk−1 + (L + 3 − Nc − x). Therefore the number
of available base stations in Lk−1 can be expressed as
L + 3 − Nc − x. These available base stations must at least
decode Nc + 1 + (1 − x) mobile terminals’ transmissions
to have an average greater than Nc + 1 active mobile
terminals per subnetwork over Lk and Lk−1 without Lk−1
borrowing base stations from Lk−2 . This cannot happen
when L + 3 − Nc − x < Nc + 1 + 1 − x, which is only
L+1
possible when Nc > L+1
2 . Hence, the condition Nc > 2
implies that Lk−1 has to borrow at least one base station
from Lk−2 , which presents an iterative argument as the one
shown when δ = 1.
Now we consider the case when L2 ≤ Nc ≤ L+1
2 .
By Lemma 2, we also have that the maximum number of
mobile terminals from Lk−1 decoded in Lk−1 ’s available
base stations is Nc . As we only need Nc + 2 − x active
mobile terminals decoded to break the inner bound defined,
Lk−1 will not have to borrow from Lk−2 when x ≥ 2.
At least Nc + 2 − x mobile terminals’ transmissions must
be decoded in Lk−1 , but M Tαk−1 +L+1 has its associations
reduced by δ ≥ x. Using M Tαk−1 +L+1 , a maximum of
Nc − δ + 1 ≤ Nc + 1 − x transmitted words can be decoded
within Lk−1 , which will lead us to have Lk−1 borrowing
at least one base station from Lk−2 . This presents another
iterative argument, akin to the one shown above.
In order to achieve a case where Lk−1 does not have
to borrow base stations from Lk−2 , our best case scenario guides us to find the first mobile terminal in Lk−1 ,
which is connected to at most x − 2 base stations that
are being borrowed by Lk , but still connected to at least
Nc + 2 − x available base stations in Lk−1 . Assume that
the index of that mobile terminal is αk−1 + ν. Clearly,
ν ≤ (L+2−Nc −x)+(x−2) = L−Nc . So in Lk−1 we have
Nc + 2 − x active mobile terminals without borrowing from
Lk−2 , but mobile terminal αk−1 + ν has already used up all
its associations and it is connected to some base stations in
Lk−2 , specifically at least Nc base stations. Hence, Lk−2 has
a maximum L+2−Nc ≤ Nc +2 base stations available to decode more transmissions, and we need at least Nc +1 words
to be decoded here, which can be done, but this would imply
that at least two mobile terminals are associated with Nc
base stations. These two mobile terminals are indexed higher
than κ, where κ = αk−2 + L + 1 − (Nc + 1). Hence, Lk−2
blocks at least Nc of the bottom L base stations in Lk−3 , and
one can see that each further subnetwork blocks at least one
base station from the preceding subnetwork for the average
number of active mobile terminals per subnetwork to remain
above Nc + 1. If say Li does not block any base stations in
Li−1 , then Li can have at most Nc active mobile terminals
decoded in Li . It follows that either Li borrows from Li−1
or only has Nc active mobile terminals. If Li borrows from
Li−1 we have a similar iterative argument as shown above.
Otherwise, Li has only Nc active mobile terminals, making
the average number of active mobile terminals through
the considered k − i subnetworks Nc + 1 per subnetwork.
Hence, each subnetwork continues blocking base stations
in the preceding subnetwork and the extra active mobile
terminals in the whole network does not scale and is fixed
by the constant x, which shows that the average number of
active mobile terminals asymptotically approaches Nc + 1
for every subnetwork of size L + 2.
We have shown that if any subnetwork has more than
Nc + 1 active mobile terminals when L ≥ Nc ≥ L2 , either
the number of extra active mobile terminals do not scale
with size of the network, or the average over the whole
network remains bounded by Nc +1 active mobile terminals
per subnetwork. This forces that the average number of
decoded words per subnetwork is at most Nc + 1, implying
that the asymptotic puDoF during the uplink using zero
c +1
. We have shown in Section IV
forcing, τUzf (L, Nc ) ≤ NL+2
Nc +1
zf
c +1
that τU (L, Nc ) ≥ L+2 , implying that τUzf (L, Nc ) = NL+2
L
whenever 2 ≤ Nc ≤ L. The proof of (7) is thus complete.
VI. C ONVERSE P ROOF FOR UPLINK WHEN Nc <
L
2
In this section, we provide a converse proof for the third
range of (6). More precisely, we show that the following
holds.
L
2Nc
,
Nc < .
(9)
τUzf (L, Nc ) =
2Nc + L
2
Similar to Section V Our proof will be based on the
concept that to break the inner bound described in (6),
at least one subnetwork of 2Nc + L consecutive MT-BS
pairs must have more than 2Nc active mobile terminals.
And all such subnetworks must either borrow or block base
stations from the subnetwork above it to decode words
corresponding to its own mobile terminals.
This happens because for a consecutive set of 2Nc + L
mobile terminals, the only base stations that can help decode
their transmissions are the corresponding base stations or
the other L base stations with preceding indices that are
connected to the set.
Using the Lemmas and definitions from Section V we
start our proof. We present cases on x, and δ.
Firstly we notice as each subnetwork has 2Nc +L mobileterminals, base-stations pairs so there is a possibility of
decoding more then 2Nc transmissions using subnetworkonly-decoding (SO-decoding) in Lk . We first show that only
a maximum of 2Nc + 1 transmissions can be decoded using
SO-decoding.By Lemma 2, to decode 2Nc +2 transmissions
using SO-decoding you need a subnetwork of 2 ∗ L + 1
mobile-terminal and base stations pairs which is larger than
2Nc + L. Thus Lk can decode at most 2Nc + 1 using SOdecoding
Thus, our first case is x = 1 and δ = 0. In Lk we have
that the highest indexed base stations that can decode the
2Nc + 1 transmissions are the L + 1 + Nc + 1 = L + Nc + 2
highest indexed base stations. But to transmit and decode
2Nc + 1 transmissions in Lk there will be at least 3 mobile
terminals that achieve the maximum number of associations.
By topology, only the lowest indexed such terminal is the
one that blocks base stations in Lk−1 . Specifically it blocks
L − 2Nc + 3 base stations, thus only 4Nc − 3 base stations
are left to decode at least 2Nc + 1 − 1 base stations.
From these 4Nc − 3, Lk−1 needs at least L + 2 − (L −
2Nc + 3) base stations to decode Nc + 1 transmissions.
But at least 2 of the mobile terminals transmitting to these
base stations would meet its max constraint, and if you
utilize exactly L + 2 − (L − 2Nc + 3) base stations then
by Corollary 1 only 1 of the highest L − 2Nc + 3 can
be transmitting. Similar to above we only consider the
lower indexed mobile terminal which reaches its maximum
association constraint. This mobile terminal is indexed at
most αk−1 + 4Nc − 5, which results that including the
base stations which are decoding transmissions and those
which cant due association constraints at least L + 2 of
the 4Nc − 3 base stations are used up. We are left with
4Nc − 3 − (L + 2) = 4Nc − L − 5 ≤ 2Nc − 6 base
stations as Nc ≤ L−1
2 . Now these 2Nc − 6 base stations
must decode at least NC − 1 transmissions, thus at least
Nc − 2 of these transmissions must come from lowest
indexed 2Nc −7 mobile terminals, and the other transmission
is transmitted from at most the αk−1 + (2Nc + L − (L +
2)) − 1 = αk−1 + (2Nc − 3). Which forces that in Lk−2
the highest indexed L + 1 − (2Nc − 2) = L − 2Nc + 3
can decode at most 1 transmission and the highest indexed
L + 1 − (2Nc − 7) = L − 2Nc + 8 can decode at most 2
transmissions.
So in Lk−2 the remaining 2Nc + L − (L − 2Nc + 8) =
4Nc − 8 base stations must decode at least 2Nc − 2. Using
a similar argument as above L + 2 of these available base
stations are used to decode Nc + 1 transmissions. So in
Lk−2 , 4Nc − 8 − (L + 2) = 4Nc − L − 10 ≤ 2Nc − 11
of the lowest indexed base stations must decode at least
Nc − 3 transmissions. So compared to Lk−1 where the
lowest indexed 2Nc − 6 base stations had to decode Nc − 1
transmissions, in Lk−2 the lowest indexed 2Nc − 11 have
to decode at least Nc − 3 transmissions, so even though the
number of needed transmissions decreased by 2, the number
of available base stations decreased by 5. This propagation
of interference would continue through all preceding subnetworks and the number of available base stations would keep
decreasing faster than the number of transmissions to be
decoded. So either we reach L1 or we stop this propagation
of interference at some Li , where i < k. If the latter
happens, then in Li one could decode at most Nc +1, which
would bring the average number of decoded words between
c
the k − i subnetworks to less than 2N2N
, so we just restart
c +L
our argument from Li−1 . If the former happens then the
number of extra decoded transmissions did not scale with
the network size, and thus the asymptotic puDOF is still
2Nc
2Nc +L .
If x = 1, and δ = 1, one observes that the interference
from Lk to Lk−1 is actually worse than as described above.
This is due to the fact that the extra active mobile terminal’s
transmission will be either decoded at one of the highest
indexed L−2Nc +3 base stations of Lk−1 or a lower indexed
base station. This either causes the same interference as
described above or by Lemma 1 the effective max constraint
of some of the higher indexed mobile terminals is reduced,
which is worse than before. Thus a similar argument follows.
If x = 1, and δ > 1, then by Lemma 1, we have that
the borrowed base stations in Lk−1 would be lower indexed
than the highest indexed L − 2Nc + 3 base stations, thus the
interference is worse than the first argument, which leads
to the same conclusion that the asymptotic puDoF is still
2Nc
2Nc +L .
Now if x > 1, by Lemma 1, and the first argument we
have that either the highest indexed L−2Nc +3 base stations
are blocked in Lk−1 and some lower indexed base stations
are borrowed, or all borrowed base stations have a lower
index than the highest indexed L − 2Nc + 3 base stations,
which from above arguments leads us to the same conclusion
c
.
that the asymptotic puDoF is still 2N2N
c +L
So we have shown that We have shown that if any
subnetwork has more than 2Nc active mobile terminals when
Nc < L2 , either the number of extra active mobile terminals
do not scale with size of the network, or the average over
the whole network remains bounded by 2Nc active mobile
terminals per subnetwork. This forces that the average
number of decoded words per subnetwork is at most 2Nc ,
implying that the asymptotic puDoF during the uplink using
c
zero forcing, τUzf (L, Nc ) ≤ 2N2N
. We have shown that
c +L
2Nc
zf
c
τU (L, Nc ) ≥ 2Nc +L , implying that τUzf (L, Nc ) = 2N2N
c +L
L
whenever Nc < 2 .
VII. AVERAGE U PLINK -D OWNLINK D EGREES OF
F REEDOM
In [13], the puDoF value τ (L = 1, Nc ) was characterized.
Here, we present zero-forcing schemes, with the goal of
optimizing the average rate across both uplink and downlink
for arbitrary values of L ≥ 2. We propose the following
theorem
Theorem 2: The average uplink-downlink puDoF that can
be achieved utilizing the interference avoidance schemes
described in Section II is characterized by
(
1
(1 + γD (Nc , L)) L + 1 ≤ Nc ,
zf
τ (L, Nc ) ≥ 2 2Nc
(10)
1 ≤ Nc ≤ L,
2Nc +L
where γD (Nc , L) is the downlink component of the puDoF
when Nc ≥ L + 1, and is given by
γD (L, Nc ) =
2
L+1
L+2
+ Nc − (L + 1)
2L+1
.
+ Nc − (L + 1)
2
(11)
The coding scheme that achieves the inner bound for
the second range of (10) is essentially the union of the
scheme described in Section III and the scheme that achieves
the third range of (6). The network is split into disjoint
subnetworks; each with consecutive 2Nc + L transmitterreceiver pairs. We consider two sets of base stations ABS =
{1, 2, · · · , Nc } and BBS = {Nc + 1, Nc + 2 · · · , 2Nc }, and
two sets of mobile terminals AM T = {1, 2, · · · , Nc } and
BM T = {Nc +L+1, Nc +L+2 · · · , 2Nc +L}. Now for each
i ∈ AM T , Ci = ABS . Similarly for each j ∈ BM T , Cj =
BBS . Thus, for the downlink and uplink, we can get the optimal puDoF described in Sections III and IV when Nc < L2 .
For the case where Nc ≥ L + 1, the coding scheme that
achieves the inner bound in (10) is as follows. First, we
associate each mobile terminal with the L + 1 base stations
connected to it. This achieves the puDoF value of unity
during the uplink in the same way as the scheme that
achieves it in Section IV. Hence, we know so far that
Ci ⊇ {i, i − 1, i − 2, · · · , i − L} ∩ [K], ∀i ∈ [K]. When
sending messages from base stations to mobile terminals,
the cooperative zero-forcing scheme works in a ”downward”
fashion as shown in [13]. Due to the network topology,
the uplink message passing scheme that achieves the unity
puDoF works in an ”upward” manner as shown in Section
IV. So to maximize the downlink puDoF, we need to find a
coding scheme that optimizes these opposing trends.
We define CiD as the set of extra associations that the
downlink scheme requires for MT i. Thus, ∀i ∈ [K] we
have that Ci = CiD ∪ {i, i − 1, · · · , i − L}.
For the downlink, we divide the network
into disjoint sub-
+ Nc − (L + 1)
networks; each consists of L + 2 L+1
2
consecutive transmitter-receiver pairs. We define = d L+1
2 e,
and κ = +Nc −(L+1). The cell association has a repeated
pattern every 2κ + L BS-MT pairs, and hence, it suffices to
describe it for the first 2κ + L BS-MT pairs. We consider
two cases based on the parity of the connectivity parameter
L. If L is odd, we partition the indices of mobile terminals
in the subnetwork into three sets:
Fig. 1: Scheme for average uplink (green shade) and downlink (blue shade) communication when Nc ≤ L
(a) L odd
S1
= {, + 1, · · · , + κ − 1},
S2
= {2 + κ, 2 + κ + 1 · · · , 2 + 2κ − 1},
S3
= {1, 2, · · · , L + 2κ} \ (S1 ∪ S2 ).
(b) L even
Fig. 2: Scheme for downlink, with all the associations
needed for optimal uplink, that achieves the lower bound
defined in equation (10) when Nc ≥ L + 1
The mobile terminals indexed in S3 are kept inactive. The
cell associations for downlink are given by the following words during the downlink, and hence our average puDoF
description.
during the downlink is
(
{1,
2,
·
·
·
,
κ
−
1},
∀i
∈
S
,
1
2( L+1
2κ
CiD =
2 + (Nc − (L + 1)))
=
.
{ + κ, + κ + 1, · · · , + 2κ − 1},
∀i ∈ S2 .
L + 2κ
L + 2( L+1
2 + (Nc − (L + 1)))
If L is even, we partition the indices of mobile terminals
A similar argument follows for the case when L is even. We
in the subnetwork into three sets:
have that (10) is valid thus Theorem 2 holds.
Figures 1 and 2 serve as examples for the average uplinkS 0 1 = {, + 1, · · · , + κ − 1},
downlink inner bounds defined in this section .
0
S 2 = {2 + κ − 1, 2 + κ + 1 · · · , 2 + 2κ − 2},
In the case of L = 1, the optimal puDoF is characterized
S 0 3 = {1, 2, · · · , L + 2κ} \ (S1 ∪ S2 ).
in [13]. The findings there coincide with our findings, as
L = 1 we find that for Nc ≥ L + 1, it directly implies
The mobile terminals indexed in S30 are kept inactive. The that = 1, andκ = (N − L) = N − 1 which results in
c
c
2(Nc −1)
cell associations
are given by the following description.
(
.
γD (Nc , L) = 2(N
c −1)+1
{1, 2, · · · , κ − 1},
∀i ∈ S 0 1 ,
CiD =
{ + κ, + κ + 1, · · · , + 2κ − 1},
∀i ∈ S 0 2 . VIII. AVERAGE U PLINK -D OWNLINK D O F WITH F ULL
U PLINK D O F
So If L is odd we have a subnetwork of L+2κ transmitterWe show that the downlink puDoF as described in (11)
receiver pairs and we decode
is optimal when we have unity DoF for uplink, i.e., each
( + κ − 1 − + 1) + (2 + 2κ − 2 − (2 + κ) + 1) = 2κ mobile terminal is associated with all the base stations con-
nected to it. In other words, we are restricted in this section
to cell associations that satisfy the following definition.
Definition 1: We say that an association scheme is called
a Full coverage association if each mobile terminal is
associated with all the base stations connected to it. More
precisely, ∀i ∈ K, {i, i − 1, · · · i − L} ∈ Ci .
We then have the following theorem:
Theorem 3: Optimal downlink puDoF when we have a
full coverage association and Nc > L is characterized as
γD (L, Nc ) =
.
2(d L+1
2 e + (Nc − L + 1))
L + 2(d L+1
2 e + (Nc − L + 1))
To help in proving Theorem 3, we define the following:
= d L+1
2 e and κ = + Nc − (L + 1). In order to prove
Theorem 3, we first break up the network into subnetworks
of 2κ + L consecutive base station (transmitter) and mobile
terminal (receiver) pairs. Label each subnetwork Lk such
that the lowest index of a base station and mobile terminal
in Lk is k×(2κ+L). Our proof will be based on the concept
that to break the puDoF described in Theorem 3, at least one
subnetwork of 2κ + L consecutive MT-BS pairs must have
more than 2κ + 1 active mobile terminals.
First, we show that without borrowing base stations
from preceding subnetworks, surpassing the puDoF value in
Theorem 3 is impossible. To do so, we first define a special
class of downlink schemes and then formulate a lemma.
Definition 2: We say that a downlink scheme relies
on Subnetwork-only downlink decoding if transmissions
from a subnetwork can only be decoded in the same subnetwork.
Lemma 3: Utilizing subnetwork-only downlink when we
already have a full coverage association scheme, a subnetwork can decode a maximum of 2κ words. Furthermore, for
a subnetwork to decode 2κ words it must be that the last
receiver in the subnetwork is active.
Proof: We prove Lemma 3 by contradiction. Say 2κ+1
words are decoded, then at least 2κ + 1 transmitters are
active in the subnetwork. So there exists at least one active
transmitter, say BS i, indexed between the indices of two
sets of κ active transmitters. Let ν1 and ν2 be the sets of
indices of the active transmitters above and below i, with
respect to order. Also, let x = max ν1 , and y = min ν2 .
We first observe that the smallest possible value for x
is κ and similarly the largest possible value of y is κ + ,
so we have i ∈ ν3 = {κ + 1, κ + 2, · · · , κ + − 1}. For
there to be 2κ + 1 words decoded words in the subnetwork,
we also need 2κ + 1 active receivers, thus we can make
corresponding disjoint index sets ν10 , ν20 , and ν30 . Where
receivers with indices in ν10 receive the first κ words,
receivers with indices in ν20 receive the last κ words, and ν30
is the set of indices of receivers that can decode the extra
word, say Wj . Now, x0 = max ν10 and y 0 = min ν20 , so we
observe that the smallest possible value for x0 is κ + − 1
and the largest possible value of y 0 is κ+L. Hence, we have
j ∈ ν30 = {κ + , · · · κ + L − 1} ∩ {i, i + 1, · · · i + L}.
We then observe that at least +1 active receivers indexed
in the set ν 0 = ν10 ∪ ν20 observe the extra transmitted signal.
To aid in writing, we define µ0j as the index set of receivers
listening to transmitters which are transmitting Wj . Hence,
for every receiver indexed µ0j there must be a unique active
transmitter indexed in ν = ν1 ∪ ν3 such that the transmitter
is associated with Wj to deliver the message at MT j and
cancel interference at all other receivers indexed in ν 0 ; call
this set of indices µj . With just the full association scheme
there are at most − 1 such active transmitters, thus we
must activate more transmitters indexed in ν3 , or add more
active transmitters indexed in ν1 ∪ ν2 to Cj . Either of those
actions increases the size of µ0j thus we must then increase
the size of µj to cancel interference. So to cancel the
interference of the extra transmitter we will always introduce
more interference in the channel. So we cannot get more
than 2κ words decoded through Subnetwork-only decoding.
Additionally if the last receiver was not active we would
have that the largest possible value of y 0 is κ + L − 1 which
would imply that it would be observing transmissions from
BS x, and to erase that interference an extra transmitter
indexed in {κ + 1, · · · y 0 } must be activated during the
downlink and associated with everything that BS x is
associated with during the downlink. But this would mean
at least one of the messages associated with BS x during
the downlink would be associated with more than Nc base
stations overall (uplink and downlink), which is not possible,
thus the last active receiver must be active. This method of
reasoning validates Lemma 3
If Theorem 3 were not true then we would have that at
least one subnetwork in the entire network must decode at
least 2κ + x words. We present cases on x.
If x = 1, let the first subnetwork that decodes 2κ + 1
words be Lk . Lemma 3 would then imply that at least one
base station is active in Lk−1 such that this base station
is either canceling the interference induced from the extra
base stations active in Lk or is the extra base station that
is carrying the extra word for Lk , but this implies that in
Lk−1 at least the last mobile terminal cannot be active,
which using Lemma 3 implies that at most 2κ − 1 words
can be decoded in Lk−1 . Thus the average over the two
subnetworks is still 2κ words decoded per subnetwork.
When x > 1, one observes that the maximum number
of base stations that can help from Lk−1 when trying to
decode words in Lk is L−1. These extra L−1 base stations
are being used to cancel interference induced from extra
base stations in Lk . That would imply that the preceding
subnetwork would only help with decoding a maximum of
L + Nc − (L + 1) words, so the smallest index of an active
receiver in Lk that is being helped by transmitters in Lk−1 is
at most L + Nc − (L + 1), which implies that the next active
receiver must have an index that is at least 2L+Nc −L + 1,
which leaves at most κ + − L receivers to decode at least
2κ + 1 − (L + Nc − (L + 1)) = κ + − L + 1 words.
Thus, in order to decode 2κ + x words where x > 1 in
subnetwork Lk , we require that κ + − L receivers decode
at least κ + − L + 1 words. Clearly this is impossible.
Thus, only an average of 2κ words per subnetwork can be
2κ
words decoded per receiver,
decoded, which results in 2κ+L
which is exactly what Theorem 3 states.
IX. C ONCLUDING R EMARKS
In this work, we presented an effort to understand optimal
cell association decisions in locally connected interference
networks, focusing on optimizing for the average uplinkdownlink puDoF problem. We consider a backhaul constraint that allows for associating each mobile terminal with
Nc base stations (cells), and an interference network where
each base station is connected to a corresponding mobile
terminal as well as L mobile terminals with succeeding
indices. We characterized the optimal association and puDoF
for the uplink problem when zero-forcing schemes are
considered. We also found that the characterization of the
optimal association for the average uplink-downlink puDoF
problem when Nc < L2 follows from our uplink characterization and previous work for the downlink problem. We
also presented the optimal zero-forcing downlink scheme if
we fix the uplink scheme to the uplink-only-optimal scheme
when Nc ≥ L + 1. We conjecture that it is in fact optimal to
have full DoF in the uplink when Nc ≥ L + 1, and hence it
would follow that the presented cell association and average
puDoF are optimal in this case.
R EFERENCES
[1] S. Veetil, K. Kuchi and R. K. Ganti. (2015, Dec.). Performance of cloud radio access networks. [Online]. Available:
http://arxiv.org/pdf/1512.05904v1.pdf
[2] A. Checko, H. L. Christiansen, Y. Yan, L. Scolari, G. Kardaras,
M. S. Berger and L. Dittmann, “Cloud RAN for mobile networks
- a technology overview,” IEEE Communication Surveys Tutorials,
vol. 17, no. 1, pp. 405-426, First Quart. 2015.
[3] China Mobile, “Next generation fronthaul interface,” White Paper,
Oct. 2015.
[4] The 5G Infrastructure Public Private Partnership. (2015, Jan.). 5GXhaul Project. [Online]. Available: https://5g-ppp.eu/5g-xhaul/
[5] O. Simeone, A. Maeder, M. Peng, O. Sahin and W. Yu.
(2015, Dec.). Cloud radio access network: Virtualizing wireless access for dense heterogeneous systems. [Online]. Available:
http://arxiv.org/abs/1512.07743.
[6] S. -H. Park, O. Simeone and S. Shamai. (2016, Jan.). Joint optimization of cloud and edge processing for fog radio access networks.
[Online]. Available: http://arxiv.org/abs/1601.02460.
[7] P. Marsch and G. P. Fettweis, Coordinated Multi-Point in Mobile
Communications: From Theory to Practice, 1st ed. New York, NY:
Cambridge, 2011.
[8] V. S. Annapureddy, A. El Gamal, V. V. Veeravalli, “Degrees of
Freedom of Interference Channels with CoMP Transmission and
Reception,” IEEE Trans. Inf. Theory, vol. 58, no. 9, pp. 5740-5760,
Sep. 2012.
[9] V. Ntranos, M. Maddah-Ali, G. Caire. (2014, Jul.). On
uplink-downlink duality for cellular IA. [Online]. Available:
https://arxiv.org/abs/1407.3538.
[10] M. H. M. Costa, “Writing on dirty paper (corresp.),” IEEE Trans. Inf.
Theory, vol. 29, pp. 439-441, May 1983.
[11] A. El Gamal, V. S. Annapureddy, and V. V. Veeravalli, “Interference
channels with coordinated multi-point transmission: Degrees of freedom, message assignment, and fractional reuse”, IEEE Trans. Inf.
Theory, vol. 60, no. 6, pp. 3483-3498, Mar. 2014.
[12] M. Bande, A. El Gamal, V. V. Veeravalli. (2016, Oct.). Degrees
of Freedom in Wireless Interference Networks with Cooperative
Transmission and Backhaul Load Constraints. [Online]. Available:
https://arxiv.org/abs/1610.09453.
[13] A. El Gamal, “Cell associations that maximize the average uplinkdownlink degrees of freedom,” in Proc. IEEE International Symposium on Information Theory (ISIT), Barcelona, Spain, Jul. 2016.
[14] A. Wyner, “Shannon-theoretic approach to a Gaussian cellular
multiple-access channel,” IEEE Trans. Inf. Theory, vol. 40, no. 5,
pp. 1713-1727, Nov. 1994.
[15] M. Wigger, R. Timo and S. Shamai (2016, Mar.). Conferencing in
Wyner’s Asymmetric Interference Network: Effect of Number of
Rounds. [Online]. Available: http://arxiv.org/abs/1603.05540
[16] V. Cadambe and S. A. Jafar, “Interference alignment and degrees of
freedom of the K-user interference channel,” IEEE Trans. Inf. Theory,
vol. 54, no. 8, pp. 3425-3441, Aug. 2008.
[17] V. Ntranos, M. Maddah-Ali, G. Caire. (2014, Feb.). Cellular Interference Alignment [Online]. Available: https://arxiv.org/abs/1402.3119.
| 7cs.IT
|
High order exactly divergence-free Hybrid Discontinuous Galerkin Methods
for unsteady incompressible flows
Christoph Lehrenfelda , Joachim Schöberlb
a Institute
for Computational and Applied Mathematics, WWU Münster, Münster, Germany
for Analysis and Scientific Computing, TU Wien, Wien, Austria
arXiv:1508.04245v2 [math.NA] 26 Apr 2016
b Institute
Abstract
In this paper we present an efficient discretization method for the solution of the unsteady incompressible
Navier-Stokes equations based on a high order (Hybrid) Discontinuous Galerkin formulation. The crucial
component for the efficiency of the discretization method is the disctinction between stiff linear parts and
less stiff non-linear parts with respect to their temporal and spatial treatment.
Exploiting the flexibility of operator-splitting time integration schemes we combine two spatial discretizations which are tailored for two simpler sub-problems: a corresponding hyperbolic transport problem and
an unsteady Stokes problem.
For the hyperbolic transport problem a spatial discretization with an Upwind Discontinuous Galerkin
method and an explicit treatment in the time integration scheme is rather natural and allows for an efficient
implementation. The treatment of the Stokes part involves the solution of linear systems. In this case a discretization with Hybrid Discontinuous Galerkin methods is better suited. We consider such a discretization
for the Stokes part with two important features: H(div)-conforming finite elements to garantuee exactly
divergence-free velocity solutions and a projection operator which reduces the number of globally coupled
unknowns. We present the method, discuss implementational aspects and demonstrate the performance on
two and three dimensional benchmark problems.
Keywords: Navier-Stokes equations, Hybrid Discontinuous Galerkin Methods, H(div)-conforming Finite
Elements, exactly divergence-free, Operator-Splitting, reduced stabilization
1. Introduction
1.1. Problem statement
We consider the numerical solution of the unsteady incompressible Navier Stokes equations in a velocitypressure formulation:
∂
in Ω
∂t u + div(−ν∇u + u⊗u + pI) = f
(1)
div u = 0 in Ω
with boundary conditions u = uD on ΓD ⊂ ∂Ω and (ν∇u − pI) · n = 0 on Γout = ∂Ω \ ΓD . Here, ν = const
is the kinematic viscosity, u the velocity, p the pressure, and f is an external body force. We consider
a discretization to (1) with high order (Hybrid) Discontinuous Galerkin (DG) finite element methods for
complex geometries with underlying meshes consisting of possibly curved tetrahedra, hexahedra, prisms and
pyramids.
Email addresses: [email protected] (Christoph Lehrenfeld), [email protected]
(Joachim Schöberl)
Preprint submitted to Comp. Meth. Appl. Mech. Eng.
April 27, 2016
1.2. Literature
Since its introduction in the paper of Reed and Hill [1], DG methods have been developed and used
for hyperbolic problems [2, 3, 4, 5, 6] and later extended to second-order elliptic (and parabolic) problems
[7, 8, 9, 10]. DG methods are specifically popular for flow problems [11, 12]. For incompressible flows
different finite element methods have been discussed in the literature [13, 14, 15]. Among these methods
several DG formulations have been considered, e.g. [16, 17, 18, 19, 20].
DG methods provide a flexibility which can be utilized for different purposes. In our case, the motivation
to consider this type of discretizations is twofold. First, relevant flow problems that can be modeled with
the incompressible Navier-Stokes equations are often convection dominated. Using the Upwind mechanism
DG methods offer a natural way to devise stable discretizations of (dominating) convection. Secondly,
abandoning H 1 -conforming finite element spaces, DG methods allow to consider (only) H(div)-conforming
finite elements (see [21]) for (Navier-)Stokes problems. This is attractive as it facilitates the design of a
discretization with exactly solenoidal solution which in turn implies energy-stability of the discretization for
the Navier-Stokes problem, cf. [19, 20, 22].
Compared to Continuous Galerkin (CG) methods the number of degrees of freedom of Discontinuous
Galerkin methods increases significantly. This drawdack is often outwayed by the advantages of the method.
However, when it comes to solving linear systems Discontinuous Galerkin methods suffer most from drastically increased globally coupled degrees of freedom. An approach to compensate for this is the concept of
Hybridization where additional unknowns on element interfaces are introduced. This increases the number
of degrees of freedom, but introduces two advantages. First, the global couplings are reduced and secondly,
the structure of the couplings allows to apply static condensation for the element unknowns. The concept has originally been introduced in the context of mixed finite element methods, cf. [21]. In the last
decade Hybridization has also been applied to Discontinuous Galerkin method for a variety of problems,
cf. [23, 24, 25, 26, 27]. We also mention, that recently similar concepts are also known in the literature
under the name “Hybridized Weak Galerkin” methods [28] where a slightly different framework is used to
derive the methods. With the emphasis to treat general polyhedral meshes “Hybrid High-Order” (HHO)
methods [29, 30, 31] have recently been introduced where a combination of Hybrid (or hybridized) methods
and higher order spaces is considered.
Concerning HDG discretizations for incompressible flows, we also mention the papers [32, 26, 33]. Further, different approaches to implement exact incompressible finite element solutions to the Stokes problem
using Hybridization have been investigated in [34, 35, 36]. In [37] the Stokes-Brinkman problem, the Stokes
problem with an additional zero order term, has been discretized using a hybridized H(div)-conforming
finite element formulation.
An interesting and often praised aspect of Hybrid mixed finite element methods is the fact that a postprocessing step can be used to reconstruct interior unknowns of an increased higher order accuracy for
elliptic problems [21]. The same is also possible for HDG methods in mixed formulation which is another
advantage over conventional DG methods, see e.g. [25]. For elliptic problems an accuracy of order k + 2
in the volume can be achieved if polynomials of order k are used on the element interfaces. One main
contribution in this paper is the introduction of a projection operator which achieves the same effect. In
[22] we already discussed this operator which has recently also been addressed under the name “reduced
stabilization” in [38, 39, 40, 41]. In these papers the construction of the corresponding operator is only
possible in two dimensions using special integration rules. We explain how the operator can be implemented
in a more general setting.
The efficiency and practicability of HDG methods is rarely addressed in the literature. Interesting
exception are [42, 43] where the computational cost of the method is compared to CG and DG methods.
The nature of diffusive (viscous) and convective terms appearing in convection-diffusion-type problems
have a substantially different character. Discretizations of problems involving only one of the two mechanisms
would typically lead to different spatial and temporal discretizations. It is therefore often desirable to consider
operator-splitting time discretization schemes which allow for a separate treatment of the different operators
as in [44, 45, 46, 47, 48, 49]. In these papers the spatial discretization for the stiff (diffusion/Stokes) and the
non-stiff (convection) operators is typically the same. We consider a different treatment of the operators
with respect to their temporal and spatial discretization.
2
1.3. The concept: Efficiency through operator splitting
A crucial ingredient in the considered discretization is the fact that the convection and the Stokes
problem are separated by means of an operator-splitting method. The operator-splitting method is chosen
such that only operator evaluations of the convection operator are required so that the time integration
scheme is explicit in terms of the convection operator. This is often affordable as the time step restrictions
following from the convection operator are the least restrictive ones in the Navier-Stokes problem. Moreover,
the convection operator is non-linear, s.t. the set up of linear systems of equations and corresponding
preconditioners or solvers would have to take place every time step which renders implicit approaches for
the convection very expensive. The Stokes operator is dealt with implicitly, i.e. it appears in linear systems
in every time step. Due to the differential-algebraic structure of the Navier-Stokes equations this is necessary
w.r.t. the pressure and the incompressibility constraint. As completely explicit handling of the viscosity
terms would introduce severe time step restrictions it is also advisable to treat viscous forces implicit. Note
that the Stokes operator is time-independent such that the setup of linear systems and preconditioners or
solvers can be done once and re-used in every time step.
The spatial discretizations for the different operators are designed differently as the treatment of the
convection term can be optimized for operator evaluations while the discretization of the Stokes operator
has to provide an efficient handling of implicit solution steps, i.e. the solution of corresponding linear
systems. This different treatment reflects in the use of two different finite element spaces which are used:
An H(div)-conforming Hybrid DG space for the velocity-pressure-pair for solving Stokes problems and a
DG space for handling convection.
1.4. Main contributions and structure of this paper
In this paper we introduce a new discretization method for the incompressible Navier-Stokes equations.
The discretization is based on a decomposition of the problem into the (unsteady) Stokes problem and
a hyperbolic transport problem. This decomposition reflects in the use of appropriate operator splitting
methods in the time discretization and the use of different finite element spaces for the different spatial
operators.
While we use a rather standard DG formulation for the spatial discretization of the hyperbolic transport
problem, we consider a new HDG formulation for the spatial discretization of the Stokes-type problem. We
use H(div)-conforming functions of element-wise polynomials of degree k for the velocity and discontinuous
pressure functions of degree k − 1. Due to the introduction of additional unknowns of degree k − 1 on
the facets for the approximation of the tangential trace of the velocity static condensation allows to reduce
the unknowns and their couplings significantly. The resulting globally coupled unknowns correspond to the
approximation of the normal trace of the velocity on the facets by order k functions, the tangential trace of
the velocity by order k − 1 functions and the approximation of the pressure field by piecewise constants.
The main contributions of this paper are:
1. Introduction of a new H(div)-conforming high order accurate HDG method with a projected jumps
formulation (also known as reduced stabilization or reduced-order HDG) for the solution of Stokes-type
problems.
2. Presentation of a combined spatial discretization for the Navier-Stokes equations based on a standard
Upwind DG formulation for the hyperbolic transport problem with the new HDG method for StokesBrinkman problems.
3. Discussion of operator-splitting time integration schemes which restrict solutions of linear systems to
Stokes-Brinkman problems.
In section 2 we discuss the discretization of spatial operators. The discussion is divided into two parts, the
discretization of the Stokes-Brinkman problem, the problem which involves all relevant spatial operators
except for the convection and the discretization of the convection operator. For the former part we consider
an HDG discretization, for the latter part a standard DG discretization. In section 3 operator-splitting
time integration schemes are discussed which are tailored for such a situation. Finally, in section 4 we give
numerical examples which demonstrate the accuracy and the performance of the method.
3
1.5. Preliminaries
Before describing the methods, we introduce some basic notation and assumptions: Ω is an open bounded
domain in Rd with a Lipschitz boundary Γ. It is decomposed into a shape regular partition Th of Ω consisting
of elements T which are (curved) simplices, quadrilaterals, hexahedrals, prisms or pyramids. For ease of
presentation we assume that all elements are not curved and consider only homogeneous Dirichlet boundary
conditions. The element interfaces and element boundaries coinciding
boundary are called
S with the domain
S
facets. The set of those facets F is denoted by Fh and there holds T ∈Th ∂T = F ∈Fh F .
In the sequel we distinguish functions with support only on facets indicated by a subscript F and those
with support also on the volume elements which is indicated by a subscript T . Compositions of both types
are used for the HDG discretization of the velocity which is denoted by u = (uT , uF ).
For vector-valued functions the superscripts t denotes the application of the tangential projection: v t =
v − (v·n) · n ∈ Rd . The index k which describes the polynomial degree of the finite element approximation
at many places through out the paper is an arbitrary but fixed positive integer number.
We identify finite
functions u ∈ Xh with their representation in terms of coefficient vectors
PNelement
X
ui ϕi for a corresponding basis {ϕi } of Xh and NX = dim(Xh ). A (generic) bilinear
u ∈ RNX , s.t. u = i=1
Y
form Gh : Xh × Yh → R is identified with the matrix G ∈ RNY ×NX , s.t. Gi,j = Gh (ϕX
j , ϕi ).
2. DG/HDG spatial discretization
In this section we introduce the spatial discretization for the Stokes operator, the convection operator
and transfer operations between both. First, we introduce the H(div)-conforming Hybrid DG discretization
of the Stokes-Brinkman problem, i.e. the stationary Navier-Stokes problem without convection and an
additional zero order term in section in section 2.1. This discretization is improved significantly. Further
on, a modification of the discretization using the idea of projected jumps (also known as reduced stabilization
or reduced-order HDG) is presented in section 2.2 including an explanation of how the projection operator
can be realized. For the convection part of the Navier-Stokes problem we consider a DG discretization using
standard approaches. This is discussed in section 2.3. As the discretization spaces for the Stokes part and
the convection part are different, we present transfer operations between the spaces in section 2.4 which
allow us to finally formulate the semi-discrete problem in section 2.5.
2.1. H(div)-conforming HDG formulation of the Stokes-Brinkman problem
In this part we consider the discretization of the Stokes part of the Navier-Stokes problem. For simplicity
we restrict the discussion to homogeneous Dirichlet boundary conditions. We present the method and
elaborate on important properties. For an error analysis of the method we refer to [22]. The reaction
term τ −1 corresponding to an inertia term stemming from an implicit time integration scheme is further
incorporated. The resulting problem is known as the (stationary) Stokes-Brinkman problem:
−1
τ u + div(−ν∇u + pI) = f
in Ω
(2)
div u = 0
in Ω
We first introduce the finite element spaces, followed by the definition of the bilinear forms corresponding
to the involved operators.
2.1.1. H(div)-conforming Finite Elements for Stokes.
Following [19] a DG formulation for the incompressible Navier Stokes equations which is locally conservative and energy-stable at the same time has to provide discrete solutions which are exactly divergence-free.
This can be achieved with a suitable pair of finite element spaces. We consider the use of H(div)-conforming
Finite Element spaces for the velocity u. H(div)-conformity requires that every discrete function uh is in
H(div, Ω) = {u ∈ [L2 (Ω)]d : div u ∈ L2 (Ω)}
4
We use piecewise polynomials, so that on each element the functions are in H(div, T ). For global conformity,
continuity of the normal component is necessary, resulting in
Y
Wh := {uT ∈
[P k (T )]d , [[uT ·n]]F = 0 ∀ F ∈ Fh } ⊂ H(div, Ω), NW := dim(Wh ),
(3)
T ∈Th
with [[·]]F the usual jump operator and P k the space of polynomials up to degree k. We refer to [21] for
details on the construction of H(div)-conforming finite elements such as the Brezzi-Douglas-Marini finite
element.
The appropriate Finite Element space for the pressure is the space of piecewise polynomials which are
discontinuous and of one degree less:
Y
Qh :=
P k−1 (T ), NQ := dim(Qh ).
(4)
T ∈Th
This velocity-pressure pair Wh /Qh fulfills
div(Wh ) = Qh .
The crucial point of this choice of the velocity-pressure pair is the property:
If a velocity uT ∈ Wh is weakly incompressible, it is also strongly incompressible:
Z
div(uT )q dx = 0 ∀qh ∈ Qh ⇔ div(uT ) = 0 in Ω.
(5)
(6)
Ω
The benefit of (6) is twofold: First, it allows to show energy-stability for the incompressible Navier-Stokes
equations, cf. section 2.5. Secondly, error estimates for the velocity error can be derived which are independent of the pressure field, we comment on this in remark 5 below.
As solutions of the incompressible Navier Stokes equations are [H 1 (Ω)]d × L2 (Ω)-regular and tangential
continuity is not imposed as an essential condition on the finite element space the discrete formulation has
to incorporate the tangential continuity weakly. We do this with a corresponding (Hybrid) DG formulation
for the tangential components across element interfaces
Remark 1 (Reduced spaces). Due to (6) solutions of the discretized (Navier)-Stokes problem will be
exactly divergence-free velocity fields. This a priori knowledge can be exploited. The basis for the space Wh
can be constructed in such a way such that we can discard certain higher order basis functions (with non-zero
divergence) that will have no contribution. A corresponding basis is introduced in [50] in the context of a
set of higher order basis functions which fulfill an exact sequence property. The resulting space Whred has
div(Whred ) = Qred
h := {p|T = const : T ∈ Th } such that also most of the degrees of freedom of the pressure
space can be discarded. The reduction of basis functions for Wh and Qh is explained in more detail in [22,
Chapter 2.2].
2.1.2. The HDG space for the velocity.
To (weakly) enforce continuity we apply a discontinuous Galerkin (DG) formulation such as the Interior
Penalty method [51]. However, to avoid the full coupling of degrees of freedom of neighboring elements,
we introduce additional unknowns on the skeleton, the facet unknowns, which represent an approximation
of the tangential trace of the solution. The DG formulation is then replaced with a corresponding HDG
formulation, s.t. degrees of freedom of neighboring elements couple only through the facet unknowns. We
note that the resulting space is only “hybrid” in the sense of the tangential unknowns. Further, we do not
consider a hybridization with respect to the pressure, cf. also remark 7.
As normal continuity is already implemented in Wh we only need a DG enforcement of continuity in the
tangential direction and hence only introduce the facet unknown for the tangential direction of the trace:
Y
Fh := {uF ∈
[P k (F )]d , uF · n = 0 }, NF := dim(Fh ).
(7)
F ∈Fh
5
H 1 -conforming (CG)
DG
H(div)-conforming DG
HDG
H(div)-conforming HDG
Figure 1: Tangential and normal continuity for different finite element spaces for the velocity.
Functions in Fh have normal component zero. For the discretization of the velocity field we use the composite
space
Uh := Wh × Fh , NU := dim(Uh ).
(8)
Remark 2 (The role of the facet space). We note that the space Fh is only introduced to allow for a
more efficient handling of linear systems. In section 4.3 we consider a numerical example where a vectorvalued Poisson problem is considered for the HDG space Uh in comparison to other (H(div)-conforming) DG
spaces to illustrate the impact of Hybridization. In case that only explicit applications of discrete operators
are used, the introduction of the space Fh entails no advantages. In this paper, facet variables appear only in
the discretization of the viscous forces. Although the facet variable has no contribution in the discretization
of inertia and pressure forces or the incompressibilty constraint we define the corresponding operations for
u ∈ Uh instead of uT ∈ Wh to simplify the presentation.
2.1.3. Viscous forces.
For ease of presentation we consider a hybridized version of the Interior Penalty method for the discretization of the viscous term. In remark 3 we comment on alternatives. In the hybridized version of the
Interior Penalty method, the usual jump across element interfaces of the Interior Penalty method is replaced
with jumps between element interior and facet unknown (in tangential direction) [[ut ]] = utT − uF from both
sides of a facet. The bilinear form corresponding to the HDG discretization of viscous forces is
Z
X Z
∂uT t
Ah (u, v) :=
ν∇uT : ∇vT dx −
ν
[[v ]] ds
∂n
T
∂T
T ∈T
(9)
Zh
Z
∂vT t
α
−
ν
[[u ]] ds +
ν [[ut ]][[v t ]] ds, u = (uT ,uF ), v = (vT ,vF ) ∈ Uh
∂n
∂T
∂T h
In this bilinear form the four terms have different functions. While the first two terms ensure consistency
(in the sense of Galerkin orthogonality) the third and fourth term are tailored to ensure symmetry (adjoint
consistency) and stability, respectively. Due to continuity of the solution ([[·t ]] = 0) the latter terms also
preserve consistency. For a more detailed introduction we refer to [22, section 2.3.1]. With respect to the
discrete norm
)
(
2
X
∂u
1
T
|||u|||21 :=
k∇uT k2T + k[[ut ]]k2∂T + h
h
∂n ∂T
T ∈Th
which is an appropriately modified version of the discrete norm typically used in the analysis of Standard
Interior Penalty methods, the bilinearform Ah is (for a sufficiently large α) consistent, bounded and coercive.
The coupling through facet unknowns enforces the same kind of continuity as in the Interior Penalty
DG formulation while preserving the following structure in the sparsity pattern: element unknowns are only
coupled with unknowns associated with the same element or unknowns associated with aligned facets.
6
k
quadrilateral
triangle
4
8
16
32
0.305
0.167
0.313
0.190
0.315
0.201
0.315
0.205
Table 1: LBB constant in dependence of k for a single element.
Remark 3 (Interior Penalty and alternatives). A drawback of the (Hybrid) Interior Penalty method
is the fact that the stabilization parameter α depends on the shape regularity of the mesh. Often α is chosen
on the safe side, but as was pointed out in [52] the condition number of arising linear systems increases
with α. In the same paper a hybridized variant of the Bassi-Rebay stabilization method (cf. [4, 5, 53]) for
a scalar Poisson equation has been proposed, see also remark 12. Such a variant is used in the numerical
examples, but as it does not have any further consequences for the remainder of this work, we stick to the
well-known (Hybrid) Interior Penalty method for ease of presentation.
2.1.4. Mass bilinear form.
The HDG mass matrix is defined as
Z
uT vT dx,
MU
(u,
v)
:=
h
u = (uT , uF ), v = (vT , vF ) ∈ Uh .
Ω
(10)
Note that we defined the mass matrix for u ∈ Uh although uF has no contribution, cf. remark 2.
2.1.5. Pressure force and incompressibility constraint.
For the pressure force and the incompressibility constraint we define the bilinearform
X Z
Dh (u, p) :=
−
p div uT dx for u = (uT , uF ) ∈ Uh , p ∈ Qh
T ∈Th
for which the LBB-condition
sup
u∈Uh
(11)
T
Dh (u, p)
≥ cLBB kpkL2 ,
|||u|||1
∀ p ∈ Qh
(12)
holds true for a cLBB independent of the mesh size h, cf. [22, Proposition 2.3.5].
Remark 4 (Robustness in polynomial degree k). In numerical experiments we observed that the LBB
constant cLBB in (12) is also robust in k. To this
R end we computed the LBB constant of one element with
Dirichlet boundary conditions and the condition Ω p dx = 0 on the pressure. The results are shown in Table
1 for varying the polynomial degree k. From the boundedness of the LBB constant on one element and the
h-robustness in (12), the robustness in h and k on the global spaces follows. This will is in the forthcoming
master’s thesis of Philip Lederer.
2.1.6. Discretization of the Brinkman-Stokes problem
With the introduced discretizations of the spatial operators the discrete Brinkman-Stokes problem can
be written as:
Find u ∈ Uh and p ∈ Qh , s.t.
−1
τ Mh (u, v) + Ah (u, v) + Dh (v, p) = hf, vi ∀ v ∈ Uh ,
(13)
Dh (u, q) =
0 ∀ q ∈ Qh .
Due to coercivity of Ah (respectively τ −1 Mh +Ah ), the LBB-condition of Dh and consistency and continuity
of all bilinear forms Brezzi’s famous theorem (see [21]) can be applied to obtain optimal order a priori error
estimates. For a discussion of the coupling structure of this discretization we refer to Remark 7 below.
7
Remark 5 (Pressure-independence of velocity error). Classical error estimates for mixed problems
result in error estimates which are formulated in the compound norm of the velocity and pressure space.
This has the disadvantage that the discretization error in the velocity depends on the approximation error
in the pressure. As was pointed out in [54], ideally this should not be the case. Due to the fact that discrete
solutions to (13) are exactly divergence-free an error estimate for the velocity field which does not involve
the pressure can be derived, cf. [22, Lemma 2.3.13].
2.2. Projected jumps: An enhancement of the HDG Stokes discretization
The proposed HDG formulation for viscous forces in section 2.1.3 can also be derived as a hybrid mixed
method with a modified flux. This is done in detail in [25] (see also [22, section 1.2.2] or [33]). In that
setting the unknowns for the primal variable (uT ) are approximated with the same polynomial degree k as
the facet unknowns (uF ). Afterwards, in a postprocessing step approximations for the primal unknown, of
one degree higher, k + 1, are reconstructed in an element-by-element fashion. This approach ends up with
a higher order approximation than the previously introduced HDG method considering the use of the same
polynomial degree k for the facet unknowns. This sub-optimality can be overcome by means of a projection
operator which leads to the projected jumps formulation. We first introduce the method and explain how
the method can be implemented afterwards.
2.2.1. Projected jumps: The method.
The idea of the projected jumps formulation is to reduce the polynomial degree of the facet unknowns in
Fh to k − 1,
Y
[P k−1 (F )]d , uF · n = 0, ∀ F ∈ Fh },
(14)
Fh → F h := {uF ∈
F ∈Fh
while keeping the polynomial degree k in Wh . In order to do this in a consistent fashion we have to modify
the bilinearform Ah . First, we introduce the L2 projection Π for a fixed facet F ∈ Fh :
Z
Z
Π : [P k (F )]d → [P k−1 (F )]d ,
(Π f ) v dx =
f v dx ∀ v ∈ [P k−1 (F )]d
(15)
F
F
k−1
T
Due to the fact that ∂u
(F ) on affine linearly mapped elements, the test functions vT and vF in
∂n ∈ P
the first integral on the element boundary in (9) can be replaced by their L2 (F ) projections ΠvT and ΠvF .
If we additionally use Π[[ut ]] instead of [[ut ]] to symmetrize and stabilize the formulation in (9) we end up
with a modified version of the Hybrid Interior Penalty method:
Z
X Z
∂uT
Arh (u, v) :=
ν∇uT : ∇vT dx −
ν
Π[[v t ]] ds
∂n
T ∈T
(16)
Zh T
Z ∂T
∂vT
α
t
−
ν
Π[[u ]] ds +
ν Π[[ut ]]Π[[v t ]] ds, u, v ∈ Wh × F h
∂n
∂T
∂T h
Note that the first two boundary integrals are just reformulated while only the last integral is really changed.
This modification preserves the important properties of Ah : It is still consistent and bounded. Further, a
deeper look into the coercivity proof (see [22]) reveals that coercivity can easily be shown in the modified
(weaker) norm
(
)
2
X
1
∂uT
2
2
t 2
|||u|||1,∗ :=
k∇uT kT + k[[Π u ]]k∂T + h
.
(17)
h
∂n ∂T
T ∈Th
Notice that, as we do not modify Wh , the normal component of uT ∈ Wh is still a polynomial of degree k.
The idea of applying such a reduced stabilization is also discussed and analyzed in [39, 40]. However, only
for the two-dimensional case the realization of the projection Π is discussed (by means of Gauss quadrature,
cf. [39, section 3.4]). In the next section a simple way to implement this projection operator is presented.
8
Remark 6 (Interplay with operator-splitting). If a Hybrid DG formulation for a problem involving
diffusion (viscosity) and convection is applied, such a reduction of the facet unknowns is only appropriate if
diffusion is dominating. One premise of the operator-splitting in this paper is that convection is not involved
in implicit solution steps. Hence, the facet variable will never be involved in the discretization of convection,
s.t. even if the physical problem is convection dominated, the projected jumps modification can be applied
without loss of accuracy.
Remark 7 (Comparison to a fully hybridized formulation). We note that the formulation in (13)
(and the corresponding reduced formulation with Arh (·, ·)) is not a “hybridized” formulation in the usual
sense, see e.g. [25]. Only for the tangential component of the velocity a new unknown is introduced, such
that we only have a “tangentially hybridized” velocity discretization. The pressure is not hybridized. We
comment on the resulting coupling of this HDG method and compare it to a HDG formulation where only
Hybrid unknowns on the facet exist, for the velocity and the pressure. The velocity unknowns in Uh = Wh ×Fh
(or Wh × F h in the reduced case) can be reduced to unknowns on the skeleton by static condensation. The
remaining facet unknowns are the usual unknowns for the normal component of the H(div)-conforming
finite element space and the (tangential) unknowns in Fh (or F h ). Due to the pressure unknowns, static
condensation of the formulation in (13) does not yield a global system which only involves facet unknowns,
but also element unknowns for the pressure. However, except for the constant pressure all pressure unknowns
can be eliminated by static condensation. We summarize: The globally reduced system only involves velocity
unknowns on the facets and a constant pressure on each element. Alternatively one could formulate a
related HDG formulation which only involves facet unknowns for the velocity and the pressure. In such a
formulation the pressure plays the role of the lagrange multiplier for the normal continuity and only for
the (weak) tangential continuity velocity unknowns have to be added on the facet. To obtain an accurate
(and exactly divergencefree) order k approximation of the velocity and an order k − 1 approximation of the
pressure, the facet unknowns would be an order k approximation of the pressure and an order k (or order k−1
if projected jumps or corresponding postprocessing techniques are applied) approximation of the tangential
velocities. Overall the number of unknowns would be smaller compared to the formulation proposed before
by one (constant) pressure unknown per element. However, the price for this is the fact, that the Lagrange
multiplier space (the pressure unknowns on the facets) is much larger and hence the structure of the arising
linear system is different. It very much depends on the linear solver strategies which of the two Hybrid DG
formulations gives the better perfomance.
2.2.2. Projected jumps: Realization.
The following way to implement projected jumps needed for Arh in (16) relies on an L2 -orthogonal basis
for the facet functions in Fh . To obtain an L2 -orthogonal basis, we take a local coordinate system on
each facet (spanned by d − 1 aligned edges) and an L2 -orthogonal Dubiner basis (see [55]) for each vector
component. We consider the three dimensional case, and denote the vectors of the local coordinate system
by e1 and e2 . Translation to the two dimensional case is then obvious.
On each facet F ∈ Fh , the space of (vector-valued) polynomials up to degree k, span(e1 , e2 ) · P k (F ), can
be split into the orthogonal subspaces
Vk−1 := span{e1 , e2 } · P k−1 (F )
and
⊥
Vk−1
:= span{e1 , e2 } · [P k (F ) ∩ P k−1 (F )
⊥
].
(18)
⊥
On each facet F the facet unknown uF can thus be written as uF = ūF +λ with unique ūF ∈ Vk−1 , λ ∈ Vk−1
.
0
We now replace the highest order function λ with two copies λT and λT , each of these functions is associated
with one of the two neighbouring elements T and T 0 , the functions λT are only defined element-local. λT
can thus be eliminated after the computation of the element matrix and finally, only uT ∈ Wh and ūF ∈ F h
appear in the global system.
λT is only responsible for implicitly realising the projection operator. To see this we now consider one
facet F ∈ Fh and one of the neighboring elements T ∈ Th . We use the decomposition corresponding to (18)
for trial and test functions
uF = ūF + λT and vF = v̄F + µT ,
(19)
9
⊥
with ūF , v̄F ∈ Vk−1 and λT , µT ∈ Vk−1
where λT , µT are only supported on T . In (13) we choose the test
⊥
function v such that vT = v̄F = 0, µT ∈ Vk−1
on T and µT 0 = 0 on every other element T 0 . This yields
Z
Z
Z
∂uT
α
α
⊥
ν
µT ds +
.
(20)
ν (utT − ūF − λT )µT ds =
ν (utT − λT )µT ds = 0, ∀ µT ∈ Vk−1
∂n
F
F h
F h
Hence, there holds λT = (I − Π)utT and we have
[[ut ]] = utT − ūF − λT = Π(utT ) − ūF = Π(utT − ūF ) = Π[[ut ]]
(21)
We conclude that it is sufficient to consider the HDG bilinear form Ah as before, where the local element
matrices are computed according to Fh . For each element matrix the degrees of freedom corresponding to
⊥
Vk−1
are then eliminated forming a corresponding Schur complement. This yiels a final stiffness matrix
which is only set up with respect to space unknowns of Wh × F h .
2.3. DG formulation for the convection
For the discretization of the convection part we consider a standard DG finite element space:
Vh := {u : u ∈ [P k (T )]d ∀ T ∈ Th },
NV := dim(Vh ).
(22)
We define the mass bilinear form
MVh
Z
(w, z) :=
w z dx,
Ω
w, z ∈ Vh .
(23)
Using an L2 -orthogonal basis on each element renders the associated mass matrix MV : RNV → RNV
diagonal. A stable spatial discretization of (1) with respect to the convection part is achieved using a
Standard Upwind DG trilinearform. We assume that the convection velocity uT is exactly divergence-free.
Z
X Z
Ch (uT ; w, z) :=
− w ⊗ uT : ∇z dx +
uT ·n ŵ z ds, uT ∈ Wh , w, z ∈ Vh .
(24)
T
T
∂T
where ŵ denotes the upwind value ŵ = limε&0 w(x − εuT (x)). Standard Upwind DG formulations are stable
in the sense that with ∂Ωin := {x ∈ ∂Ω, uT (x) · n < 0} there holds
Z
1
|uT · n| w2 ds + Ch (uT ; w, w) ≥ 0,
∀ w ∈ Vh , u ∈ Uh , div(uT ) = 0.
(25)
2 ∂Ωin
2.4. Transfer operations and embeddings
The discrete convection and Stokes operators have been defined on different spaces. In order to combine
both we introduce transfer operations between the (finite element) spaces to make both discretizations
compatible. We restrict ourselves to two types of transfer operations which are based on embeddings:
I: With Wh ⊂ Vh we have a canonical embedding of Uh in Vh with the embedding operator I : (uT , uF ) ∈
Uh → uT ∈ Vh . Note that the corresponding operation in terms of coefficient vectors, denoted by
I : RNU → RNV , is not an identity.
IT : The embedding operator I implies the canonical embedding I 0 : Vh0 → Uh0 , which maps functionals
I 0 : f ∈ Vh0 → [(uT , uF ) ∈ Uh → f (uT )] ∈ Uh0 . The corresponding matrix representation is IT : RNV →
RNU .
To realize the operator I we consider the equivalent L2 problem for u ∈ Uh .
Z
Z
(Iu) v dx =
uT v dx ∀ v ∈ Vh .
Ω
Ω
10
(26)
In terms of coefficient vectors this reads as
MV (Iu) = MU,V u, ∀ u ∈ RNU
=⇒
with the mixed mass matrix
Z
U,V
V
Mi,j =
ϕW
j ϕi dx, i = 1, . . . , NV , j = 1, . . . , NW
Ω
I = (MV )−1 MU,V
(27)
and MU,V
i,j = 0, j > NW .
Note that MV is diagonal (for affine linear transformations) such that (MV )−1 can be evaluated very
efficiently. The overall cost of the transfer operator I is essentially that of one sparse matrix multiplication.
Let CV : RNU × RNV → RNV denote the discrete convection operator corresponding to the trilinearform
Ch in (25).
With the operator I we can formulate applications of the convection and the mass operations CV (u), MV :
NV
R
→ RNV with respect to functions in the HDG space Uh and denote the corresponding operators by
U
C (u), MU : RNU → RNU ,
CU (u) := IT CV (u)I,
MU := IT MV I.
(28)
Remark 8 (Restriction on time integration scheme). Note that the restriction to these two transfer
operations implies that we do not allow to apply any part of the Stokes operator to a function in Vh and that
no functional on Uh can be used in solution steps involving the convection operator. This is a restriction on
the time integration scheme.
Remark 9 (Curved elements). If curved elements are considered we no longer have Wh ⊂ Vh due to the
Piola transform usually applied to construct H(div)-conforming finite elements. In this case I : Uh → Vh
is not an embedding, but the L2 projection. Nevertheless, MV is block diagonal with blocks which are not
diagonal matrices only on curved elements. Hence, applications of (MV )−1 are still cheap.
2.5. The semidiscrete formulation
With the definitions of the bilinear forms in the previous sections we arrive at
discrete DAE problem: Find u(t) ∈ Uh and p(t) ∈ Qh , such that
∂
hf, vi
∀ v ∈ Uh ,
MU
h ( ∂t u, v) + Ah (u, v) + Ch (uT ; u, v)+ Dh (v, p) =
Dh (u, q) =
0
∀ q ∈ Qh ,
Mh (u, v) = Mh (u0 , v) ∀ v ∈ Uh ,
the following spatially
t ∈ [0, T ],
t ∈ [0, T ],
t = 0.
(29)
Here, we implicitly used the embedding I to define Ch (uT ; ·, ·) on Uh . Due to (25) we have with
(
∂
1 d
d
u, u)L2 =
kuk2L2 = kukL2 kukL2 = hf, ui − Ah (u, u) − Ch (uT ; u, u) − Dh (u, p) ≤ kf kL2 kukL2 (30)
| {z } |
{z
} | {z }
∂t
2 dt
dt
≥0
≥0
=0
the stability of the kinetic energy:
d
kukL2 ≤ kf (t)kL2
dt
In the next section we discuss operator splitting time integration methods to solve (29) efficiently.
(31)
3. Operator-splitting time integration
In this section we are faced with the problem of solving the semi-discrete Navier-Stokes problem with a
proper time integration scheme. For ease of presentation we neglect external forces (f = 0) in the following
11
and consider the problem in terms of discrete operators corresponding to the bilinear (trilinear) forms
introduced in the previous section: Find u(t) ∈ RNU and p(t) ∈ RNQ , such that
U
in [0, T ],
MU ∂u
∂t + Au + C (u) u +∆t D p = f
(32)
DT u = 0 in [0, T ],
u(t = 0) = u0 .
In the time integration scheme we want to explicitly exploit the properties of the spatial discretization,
i.e. the convection operator C should only be involved explicitly in terms of operator evaluation. Due
to the DAE-structure we require time integration schemes which are stiffly accurate. Hence, the solution
of a Stokes-Brinkman problem should conclude every time step so that the incompressibility constraint is
ensured. For operator splittings of this kind different approaches exist. We briefly discuss three approaches.
In section 3.1 we discuss additive decomposition methods, like the famous class of IMplicit EXplicit (IMEX)
schemes. Product decomposition methods, sometimes also called exponential factor splittings, are discussed
in section 3.2 in the framework of operator-integration-factor splittings introduced in [48]. We discuss the
advantages and disadvantages of both approaches and motivate the consideration of a different approach.
An operator-splitting modification of the famous fractional step method, cf. [56], which eliminates the
most important disadvantages of the additive and multiplicative decomposition methods is then introduced
and discussed in section 3.3. We want to stress that the considered operator splitting approaches are of
convection-diffusion type and should not be confused with projection methods like the Chorin splitting [57].
3.1. Additive decomposition using IMEX schemes
Additive decomposition methods distinguish spatial operators that are treated implicitly and those that
are treated explicitly. The decomposition is additive in the sense that every solution (sub-)step involves
both operators. This is used by IMplicit EXplicit (IMEX) schemes (see [44, 45, 46]). The simplest of these
schemes is the semi-implicit Euler method for which one time step of size ∆t reads as:
(MU + ∆tA)un+1 + ∆tDpn+1 = MU un − ∆tCU (un )un
(33)
DT un+1 = 0
Convection only appears on the r.h.s. so that only operator evaluations occur for the convection operator,
while the remainder appears fully implicit. This scheme is obviously only first order accurate and only conditionally stable. For higher order methods of this decomposition type (e.g. using Multistep or partitioned
Runge-Kutta schemes) we refer to [44, 45, 46].
The major drawback of this type of operator splitting methods is the fact that the time step size of the
explicit and the implicit part of the decomposition have to coincide. Thereby the stability restriction caused
by the explicit treatment of the convection part dictates not only the number of explicit evaluations but also
- which is typically more expensive - the number of solution steps for the implicit part. The decomposition
methods considered in the sections 3.2 and 3.3 overcome this issue.
3.2. Product decomposition with operator-integration-factor splitting
The disadvantage of IMEX methods can be avoided with product decomposition methods, where sequences of separated problems are solved successively. The separated problems then only involve the Stokes
or the convection operator at the same time. The major benefit of this is the fact, that the numerical solution of the sub-problems (e.g. the size of the time steps) can be chosen completely different. The derivation
of the so called operator-integration-factor splitting approach has been derived in [48, Section 2.1]. We use
this idea to obtain an operator decoupling for the DAE which allows to formally rewrite (32) as
∂ t→t∗
∗
u) + Qt→t M−1 (Au + Dp) = 0 in [0, T ],
∂t (Q
(34)
DT u = 0 in [0, T ],
∗
for some arbitrary t∗ ∈ R with the integration factor Qt→t , the propagation operator, specified later in
section 3.2.2.
12
Be aware that M−1 is to be understood only formally for now. We are able to apply any suitable (i.e.
implicit and stiffly accurate) time integration method for (34). We do this for a first order method here to
explain the procedure and refer to the literature [48] for higher order variants.
3.2.1. Implicit Euler.
We employ the implicit Euler method on (34) and arrive at
1
n
∗
n+1
∗
tn+1 →t∗ n+1
u
− Qt →t un ) + Qt →t M−1 (Aun+1 + Dpn+1 ) = 0,
∆t (Q
DT un+1 = 0.
After setting t∗ = tn+1 and multiplication with M this simplifies to
n
n+1
(M + ∆tA) un+1 + ∆tDpn+1 ) = MV,U Qt →t Iun ,
T n+1
D u
= 0,
(35)
(36)
which is the solution of a Stokes-Brinkman problem as in (33), but with a different right hand side.
3.2.2. The propagation operator.
1
2
1
2
The propagation operator Qt →t is defined as Qt →t w = v(t2 ) where v(s) solves
∂v
+ (MV )−1 CV (s)v(s) = 0,
∂s
∀ s ∈ (t1 , t2 ],
v(t1 ) = w.
(37)
Here CV (s) := CV (ú(s)) with ú(s) an extrapolation of divergence-free solutions of previous time steps.
Note, that the extrapolation of convection velocities renders the problem (37) linear hyperbolic and ensures
stability in the sense of (25). After replacing Q by a numerical time integrator Q∆t for (37), the time
integration method is completely specified. The order of accuracy of the extrapolation ú and the time
integrator Q∆t should coincide with the order of the time integration method applied on (34).
3.2.3. Properties of product decompositions.
At this point the advantage of product decomposition methods over IMEX schemes is evident: As the
time integration of (37) is independent of the one for (34) the (not necessarily) explicit time integrator can
deal with stability restrictions (typically using multiple time steps) without influencing the time step for
(34). This separation of the two problems also introduces the biggest disadvantage of product decomposition
methods: an additional consistency error. Even if the DAE (32) has a stable stationary solution, product
decomposition methods may not reach it. This is not the case for monolithic or additive decomposition
approaches. Nevertheless, this splitting error is controlled by the time discretization error.
Remark 10 ((Marchuk-)Yanenko splitting). If the implicit Euler discretization as in (36) is combined
with an Euler method for (37), the famous Yanenko splitting method is recovered.
3.3. A modified fractional-step-θ-scheme
In this section we introduce an approach to circumvent severe CFL-restrictions and splitting errors at
the same time. We no longer ask for sub-problems that only involve the Stokes or the convection operator
(as in section 3.2), but ask for sub-problems which only involve one of both implicitly. The method is based
on [49] where the well-known fractional-step-θ-scheme, cf. [56, Chapter II, section 10], is modified. The
resulting method is an additive decomposition method without the time step restrictions of IMEX schemes.
We start with formally writing down an operator-splitting version of the fractional-step-θ-scheme. The
scheme is divided into three steps, the first and the last step treat the Stokes part implicitly and the
13
convection part explicitly as in (33) while the second step treats convection implicitly and viscosity forces
explicitly.
(M + θ∆tA)un1 + θ∆t D pn1 = Mun0 − θ∆tCun0
Step 1(tn0 → tn1 ) :
(38a)
DT un1 = 0
= Mun1 − θ∗ ∆t(Aun1 + Dpn1 )
Step 2 (tn1 → tn2 ) : { (M + θ∗ ∆tC)un2
(M + θ∆tA)un3 + θ∆t D pn3 = Mun2 − θ∆tCun2
Step 3 (tn2 → tn3 ) :
DT un3 = 0
(38b)
(38c)
The time stages are labeled by the superscripts n0 , n1 , n2 , n3 , respectively, where n0 denotes initial data
∗
and n3 the final time stage. The time steps size
√ for the first and the last step is θ∆t and θ ∆t in the
∗
middle step with θ = 1 − 2θ. Here θ = 1 − 1/ 2. We note that specifications for M and C with respect
to the considered spaces and the convection velocity for C are still missing at this points. The scheme is
second order accurate. In contrast to the unsplit fractional-step-θ-scheme the stability analysis of this time
integration scheme is an open problem. In our experience, however, the stability restrictions are much less
restrictive than those of a comparable IMEX schemes.
3.3.1. Sub-steps in different spaces.
Initially, we stated that we want to avoid solving linear systems involving convection, for efficiency
reasons. This seems to be contradictory to what is formulated in (38b). Nevertheless, as only convection is
involved implicitly an efficient numerical solution is still possible. We apply a simple iterative scheme which
only involves explicit operator evaluations of the convection to do so. We explain this in section 3.3.2. To
do this efficiently, we want Step 2 to be formulated in the space Vh while Step 1 and 3 are to be formulated
in the space (Uh , Qh ). This poses problems the solution of which we discuss in this section.
In Step 1 and Step 3 the adjustments are obvious: Step 1 only depends on initial data in Uh . The
initial data for Step 3 is in Vh but appears only in terms of functionals (Mv and Cv) such that the transfer
operations are clear. Step 2 is more involved, cf. remark 8. Functionals in Uh0 (such as Au) are in general
not functionals in Vh0 . We use the first equation in (38a) to formally define a different representation of the
functionals required in Step 2:
θ∆t gn1 = −θ∆t(Aun1 + Dpn1 ) = M(un1 − un0 ) + θ∆tCun0
(39)
Now we replace M and C with operations suitable for a setting in Vh :
RNV 3 gn1 :=
1
MU,V (un1 − un0 ) + θ∆tCV (un0 )Iun0
θ∆t
We arrive at the modified fractional-step-θ-scheme:
(MU+ θ∆tA)un1 + θ∆tD pn1
Step 1 :
DT un1
Step 2 : (MV + θ∗ ∆tCV (tn2 )vn2
(MU+ θ∆tA)un3 + θ∆tD pn3
Step 3 :
DT un3
(40)
= MU un0 − θ∆t IT CV (un0 ) I un0
=0
(41a)
= MU,Vun1 − θ∗ ∆tgn1
(41b)
V,U
=M
=0
n2
T
V
n2
v − θ∆t I C (t )v
n2
(41c)
where we replaced the generic operators M, C with suitable ones. We recall the definition of the extrapolated
convection operator CV (tn2 ) := CV (ú(tn2 )) the convection velocity of which is (linearly) extrapolated from
exactly divergence-free velocities.
3.3.2. Iterative solution of the implicit convection problem.
In (41b) we need to solve a problem of the form
v + τ ∗ (MV )−1 CV v = g∗
14
(42)
for given τ ∗ , g∗ and constant convection CV . We do this by means of a pseudo time-stepping method, i.e.
we formulate (42) as the stationary solution to
∂
v(s) + v(s) + τ ∗ CV v(s) = g∗ ,
∂s
s ∈ [0, ∞).
(43)
Note that a stationary solution exists as Id + τ ∗ CV only has eigenvalues λ, with Re(λ) > 1. This stationary
solution is approximated with a few explicit Euler time step with an artificial time step size ∆s.
vi+1 = vi + ∆s(g∗ − vi + τ ∗ (MV )−1 CV vi ).
(44)
This procedure can be interpreted as a Richardson iteration applied to (42). With a time step size which is
tailored for stability (as in the numerical solution to (37)) we iterate (44) until the initial residual is reduced
by a prescribed factor. To solve the convection step, Step 2, we only require operator evaluations for the
convection. The time step size ∆s in this iteration is decoupled from the time step size ∆t used of the
overall scheme.
4. Numerical examples
In this section we consider different test problems which essentially purpose three different goals:
1. The validation of convergence properties of the HDG method with and without the projected jumps.
2. The investigation and quantification of the dependency of the sparsity pattern of arising linear systems
on the choice of the discrete velocity spaces.
3. The evaluation of the performance of the space and time discretization on benchmark problem.
The examples in section 4.1 and section 4.2 aim at goal 1 and are two-dimensional test cases with known
exact solutions for a stationary Stokes and Navier-Stokes problem, respectively. These cases allow for a
thorough investigation of the convergence history of the proposed spatial discretizations in the usual norms.
The test case in section 4.3 aims at goal 2 and is concerned with the impact of the choice of discretization
spaces on the complexity of linear systems. For this purpose, a three-dimensional vector-valued Poisson
problem is considered using different H(div)-conforming DG and HDG spaces. The impact of hybridization
and the improvement due to the projected jumps modification is compared to other DG methods. Finally,
we approach goal 3 by considering two challenging transient benchmark problems in sections 4.4 and 4.5.
The benchmark problems have been defined within the DFG Priority Research Programm ’Flow Simulation
on High Performance Computers’ and are formulated in [58]. In section 4.4 a two-dimensional benchmark
problem is considered to demonstrate the accuracy of our method for a demanding test case and to compare
the discussed time integration methods. Finally, in section 4.5 we discuss a three-dimensional, and hence
computationally demanding, benchmark problem from [58]. We compare accuracy and run-time performance
to the data of the studies in [58, 59, 60].
The methods discussed in this paper have been implemented in the add-on package ngsflow [61] for the
high order finite element library NGSolve [62]. The computations in this section have also been carried out
with this software. Throughout this section we only consider direct solvers and comment on linear solvers
below, in remark 12.
For the computations of the benchmark problem in sections 4.4 and 4.5 we used the reduced H(div)conforming space Whred , cf. remark 1 and the projected jumps modification, cf. section 2.2.
4.1. Stokes flow around obstacle
We consider the Stokes problem in the domain Ω = [−2, 2]2 \ Ω− with the circular obstacle Ω− :=
{kxk ≤ 1} and viscosity ν = 1. On the whole boundary except for the outflow boundary {x = 2} we
prescribe Dirichlet boundary conditions, on {x = 2} we prescribe Neumann-kind boundary conditions such
1
that the solution to the problem is given by u = (∂y Ψ, −∂x Ψ) with the potential Ψ = y(1 − x2 +y
2 ). Note
−
that Ψ|∂Ω− is constant so that u · n = ∇Ψ × n = 0. On ∂Ω we have u · n = 0, but u 6= 0, i.e. we have a
slip on the obstacle. Further we have that p = 0 in Ω.
15
k∇(u − uh )kL2 (Ω)
ku − uh kL2 (Ω)
kp − ph kL2 (Ω)
100
10−1
10−2
10−2
10−4
10−4
10−6
10
10−3
k=1
k=2
k=3
k=4
k=5
−8
10−10
1
2
10−6
O(h6 )
3
refinement level
10−8
4
5
10−5
k=1
k=2
k=3
k=4
k=5
1
2
k=1
k=2
k=3
k=4
k=5
10−7
O(h5 )
10
3
refinement level
4
5
−9
1
2
O(h5 )
3
4
5
refinement level
Figure 2: Convergence of divergence-conforming HDG method with (solid) and without (dotted) projected jumps in different
norms for the example in section 4.1. The dotted lines in gray correspond to the HDG method without projected jumps, the
solid lines in color correspond to the HDG method with projected jumps. Except for the error in the pressure the results are
hardly distinguishable.
On an initially coarse unstructured grid with only 72 triangles we obtain the results displayed in Figure
2 after succesive uniform mesh refinements. We compare the exact solution u, p and the computed solution
uh , ph and investigate the convergence of the error in the norms ku − uh kL2 (Ω) , k∇(u − uh )kL2 (Ω) and
kp − ph kL2 (Ω) . We consider the discretization with and without the projected jumps modification. In all
norms we observe optimal order convergence, i.e. ku − uh kL2 (Ω) = O(hk+1 ), k∇(u − uh )kL2 (Ω) = O(hk ) and
kp − ph kL2 (Ω) = O(hk ) where k is the order of the velocity field inside the elements. Note that after static
condensation the remaining degrees of freedoms are the unknowns corresponding to order k polynomials
for the tangential and order k polynomials for the normal component on the facets and one constant per
element for the pressure. For the projected jumps formulation the tangential unknowns are reduced by one
order. The difference between the error of both formulations — with and without projected jumps — is only
marginal in the velocity. In the pressure field we observe a difference between the methods, however both
converge with optimal order and the difference decreases for increasing polynomial degree k. For the impact
of the projected jumps modification concerning the computational effort, we refer to section 4.3 where this
aspect is discussed in more detail.
4.2. Kovasznay flow
As a test case for the stationary Navier-Stokes equations, we consider the famous example by [63]. On
the boundary of the domain Ω = [− 12 , 32 ] × [0, 2] we again prescribe inhomogeneous Dirichlet data, so that
the exact solution to the Navier-Stokes equations (with ν = 1) is
1
1 − eλx cos(2πy)
u(x, y) =
, p(x, y) = − e2λx + p̄ with p̄ ∈ R
λ λx
sin(2πy)
2
2π e
R
2
Here, λ = ν −1 +√−8π
and p̄ so that Ω p dx = 0.
ν −2 +64π 2
To approximate the stationary solution we initially solve the Stokes problem corresponding to the boundary data and use the simple IMEX scheme of first order, cf. (33), to progress to t = 1000 with 50 time steps
of size ∆t = 20. This choice of time discretization parameters gives stable solutions for all considered spatial
discretizations but at the same time provides time discretization errors which are neglegible compared to
the spatial errors. On an initally coarse unstructured grid with only 18 triangles we obtain the results
displayed in Figure 3 after succesive uniform mesh refinements. We observe that the errors converge optimal
in the considered norms for the pressure and the velocity. Further, we observe that the difference between
the results obtained with and without the projected jumps formulation, e.g. with and without a reduction
16
k∇(u − uh )kL2 (Ω)
ku − uh kL2 (Ω)
kp − ph kL2 (Ω)
100
101
101
10−2
10−1
10−1
10−4
10
10−3
k=1
k=2
k=3
k=4
k=5
k=6
−6
10−8
10−10
1
2
k=1
k=2
k=3
k=4
k=5
k=6
10−5
O(h7 )
10−7
3
refinement level
4
5
1
2
k=1
k=2
k=3
k=4
k=5
k=6
10−3
10−5
O(h6 )
10−7
3
4
5
refinement level
1
2
O(h6 )
3
4
5
refinement level
Figure 3: Convergence of divergence-conforming HDG method with (solid) and without (dotted) projected jumps in different
norms for the example in section 4.2. The results are almost identical.
of the degrees of freedoms at the facets, is again only marginal. In fact one only observes a (very small)
difference in the case k = 1.
4.3. Linear systems - A comparison between DG and HDG methods
We consider the comparably simple vector-valued Poisson problem
− ∆u = f in Ω,
u = 0 on ∂Ω,
(45)
discretized with four different methods. The three-dimensional domain Ω is the same as the one in section 4.5
with an unstructured tetrahedral mesh consisting of 3487 elements. The problem (45) leads to discretizations
with symmetric positive definite matrices A. Only the lower triangular part of the sparse matrix has to be
stored and a sparse Cholesky factorization algorithm can be used. In this comparison, we used the sparse
direct solver PARDISO, cf. [64, 65]. We are only concerned with the sparsity pattern of different methods
before and after static condensation and do not compare accuracy or conditioning of the discretizations.
Moreover, for the sake of simplicity we do not apply the reduction of the space Wh as discussed in remark 1.
The following quantities of interest for the linear systems arising from discretizations of (45) are displayed in
#dof[K] : number of unknowns (in thousands)
#cdof[K] : number of unknowns after static condensation (in thousands)
Table 2:
#nzeA[K] : number of non-zero entries in the system matrix A (lower triangle) (in thousands)
#nzeL[K] : number of non-zero entries in the Cholesky factor L (in thousands)
Four different methods are considered on Wh or Uh with varying polynomial degree between k = 1 and
k = 6:
1. HDG: The HDG method proposed in section 2 without projected jumps.
2. PHDG: The HDG method with projected jumps, cf. section 2.2.
3. Std.DG.: A standard DG method using the space Wh where the basis is constructed such that all
degrees of freedom from one element couple with all degrees of freedom from adjacent elements.
4. N.DG: A nodal DG method using the space Wh where basis functions are assumed to be constructed
such that degrees of freedom associated to one element couple only with degrees of freedom from
adjacent elements which have support on the shared facet. At the same time we assume that basis
functions are constructed such that the number of basis functions with support on a facets is minimized,
cf. [11]. In terms of the sparsity pattern this nodal DG method represents the best case for a DG
method without hybridization.
17
Std.DG
N.DG
HDG PHDG
Std.DG
k=1
#dof[K]
#cdof[K]
#nzeA[K]
#nzeL[K]
23
23
732
10 113
23
23
676
10 208
69
69
2 037
17 768
146
146
21 686
261 977
146
146
16 051
260 416
#dof[K]
#cdof[K]
#nzeA[K]
#nzeL[K]
453
453
184 035
2 099 690
453
411
101 473
1 741 072
298
229
22 443
194 524
38
38
637
5 569
67
67
5 073
64 463
PHDG
158
137
8 103
70 138
112
91
3 628
31 415
500
343
50 412
435 129
423
267
30 588
264 183
67
67
4 177
69 831
k=4
237
168
12 124
104 496
271
271
69 525
814 168
271
261
46 912
731 253
681
389
64 816
557 798
702
702
424 764
4 752 072
702
597
200 813
3 519 061
k=5
773
480
98 696
847 913
HDG
k=2
k=3
#dof[K]
#cdof[K]
#nzeA[K]
#nzeL[K]
N.DG
k=6
1 128
640
175 321
1 502 558
1 022
533
121 944
1 045 444
Table 2: Comparison of different DG methods for the vector-valued reaction Poisson problem (45).
In Table 2 we observe that for small polynomial degree k the amount of additional unknowns required for
the HDG formulation is quite large as are the nonzero entries in the system matrix. Nevertheless except
for k = 1, the number of nonzero entries in the Cholesky factor are comparable (k = 2) to the Standard
DG methods or less (k > 2). For high order, i.e. k ≥ 4 the HDG method performs significantly better as it
has less nonzero entries in A and L. The HDG space with the projected jumps modification improves the
situation dramatically. It essentially compensates the overhead of the HDG method for small k. But even
for k = 6 the effect is still significant. We note, that the difference between the first three methods increases
for increasing polynomial degree k whereas the difference between the last two method, PHDG and HDG,
decreases. In all cases the HDG method with projected jumps outperforms all alternatives.
4.4. A two-dimensional benchmark problem
In this section we consider the benchmark problem denoted as “2D-2Z” in [58] where a laminar flow
around a circle-shaped obstacle is considered. The Reynolds number is moderately high (Re = 100) and
results in a periodic vortex street behind the obstacle. We briefly introduce the problem (for more details
we refer to [58]) and the numerical setup to investigate spatial and temporal discretization errors. Finally,
we discuss the obtained results.
4.4.1. Geometrical setup and boundary conditions.
The domain is a rectangular channel without an almost vertically centered circular obstacle, cf. Figure
4,
Ω := [0, 2.2]×[0, 0.41] \ {kx − (0.2, 0.2)k2 ≤ 0.05}.
(46)
The boundary is decomposed into Γin := {x = 0}, the inflow boundary, Γout := {x = 2.2}, the outflow
boundary and ΓW := ∂Ω\(Γin ∪Γout ), the wall boundary. On Γout we prescribe natural boundary conditions
(−ν∇u + pI) · n = 0, on ΓW homogeneous Dirichlet boundary conditions for the velocity and on Γin the
inflow Dirichlet boundary conditions
u(0, y, t) = uD = (3/2 · ū) · 4 · y(dy − y)/d2y · (1, 0, 0).
Here, ū = 1 and the viscosity is fixed to ν = 10−3 which results in a Reynolds number Re = 100.
4.4.2. Drag and Lift.
The quantities of interest in this example are the (maximal and minimal) drag and lift forces cD , cL that
act on the disc. These are defined as
Z
Z
1
∂u
1
∂u
cD := 2
ν
− pn · ex ds,
cL := 2
ν
− pn · ey ds.
ū r Γ◦
∂n
ū r Γ◦
∂n
18
Here ex , ey denote the unit vectors in x and y direction, r = 0.05 is the radius of the obstacle, ū is the
average inflow velocity (ū = 1) and Γ◦ denotes the surface of the obstacle.
4.4.3. Numerical setup.
We use an unstructured triangular grid with an additional layer of quadrilaterals around the disk which
is anisotropically refined towards the disk once. In Figure 4 the geometry, the mesh and a typical solution
is depicted.
2
0
Figure 4: Sketch of the mesh and the solution (color coding corresponding to velocity magnitude kuk2 ) to the problem considered
in section 4.4 at a fixed time t (left) and zoom-in on the boundary layer mesh (right).
In order to be able to neglect time discretization errors, when investigating the spatial accuracy, we
consider the use of a well-known stiffly accurate second order Runge-Kutta-IMEX scheme, taken from [45,
section 2.6], with an extremely small time step size 3.125 · 10−5 which means that roughly 10000 time steps
are used to resolve one full period. The duration of one full cycle is roughly 1/3s. We also fix the mesh and
consider a pure p-refinement, i.e. variations of the polynomial degree k.
For the investigations of the temporal discretization error, we consider a fixed polynomial degree k = 6,
with the same mesh. For the considered example the stability restriction of the second order IMEX scheme is
severe. A time step of below 10−3 has to be considered. The intention of the discussion of operator-splitting
methods in section 3 has been to present strategies to circumvent or relax these severe conditions. We
compare the performance of a product decomposition method, a second order operator-integration-factor
splitting version of the BDF2 method, and the modified fractional-step-θ-scheme discussed in section 3.3.
Note that these methods allows to consider much larger time steps than the IMEX scheme. As references
we give the values from the literature [66] and from the IMEX scheme with ∆t = 10−3 (stability limit) and
the reference solution with ∆t = 3.125 · 10−5 .
4.4.4. Numerical results: spatial discretization.
In Table 3 the quantities of interest are shown for varying polynomial degree k. As a reference we also
show the result obtained by FEATFLOW [67] with a discretization using quadrilateral meshes and continuous
second order finite elements for the velocity with a discontinuous piecewise linear pressure (Q2 /P1disc ). These
results have been made accessible on [66]. We observe a rapid convergence for the p-refinement, i.e. the
#dof
k
k
k
k
k
k
=1
=2
=3
=4
=5
=6
ref. [66]
max cD
min cD
max cL
min cL
211
148
558
441
797
626
2.52594
3.22841
3.23184
3.22714
3.22759
3.22757
2.47871
3.16260
3.16842
3.16401
3.16432
3.16430
0.65728
1.00571
0.98822
0.98431
0.98578
0.98580
-0.81672
-1.03894
-1.02427
-1.01906
-1.02053
-1.02053
167 232
667 264
3.22662
3.22711
3.16351
3.16426
0.98620
0.98658
-1.02093
-1.02129
2
4
6
9
12
16
Table 3: Accuracy of the spatial discretization: results for different polynomial degrees.
increase of the polynomial degree. Compared to the results from the literature [66] the same order of
accuracy is achieved with a lot less degrees of freedoms.
19
4.4.5. Numerical results: temporal discretization.
The results for the temporal discretization are shown in Table 4. First of all, we observe that the second
order IMEX scheme is already very accurate at its stability limit. But the method does not allow to choose
larger time steps. This is in contrast to the alternative methods discussed here. For these methods we
can consider much large time steps and observe a second order convergence. In this example the product
decomposition method is more accurate than the modified fractional step method by one (time) level. We
note that one major concern with this method is however, that splitting errors appear also if a stationary to
the flow problem exists. This is not the case for additive decomposition methods or the proposed modified
fractional step method.
2nd order IMEX
operator-integrationfactor splitting
(BDF2)
modified fractionalstep-θ-scheme
1/∆t
max cD
min cD
max cL
min cL
<1 000
1 000
32 000
3.22754
3.22757
unstable
3.16437 0.98486
3.16431 0.98580
-1.01957
-1.02053
125
250
500
1 000
3.38456
3.25564
3.23239
3.22819
3.28279
3.18643
3.16836
3.16394
1.16052
1.01725
0.98864
0.98320
-1.19939
-1.05276
-1.02378
-1.02013
125
250
500
1 000
3.38272
3.28713
3.25036
3.23656
3.29673
3.21483
3.18127
3.17094
1.07667
1.00937
0.99172
0.98721
-1.12227
-1.04840
-1.02807
-1.02227
Table 4: Accuracy of operator-splitting time integration methods
4.5. A three-dimensional benchmark problem
Finally, we consider a three dimensional benchmark problem, the problem “3D-3Z” in [58]. In contrast
to the problem in section 4.4 the inflow velocity is varied over time and the observed time interval is fixed.
The maximal Reynolds number in this configuration is also Re = 100 as in the previous section. The focus
in this section is on the study of performance in the sense of computational effort over accuracy.
4.5.1. Geometrical setup and boundary conditions.
The geometrical setup is a generalization of the problem in the previous section. A cylindrical-shaped
obstacle is places in a cuboid-shaped channel slightly above the vertical center:
Ω := [0, 2.5]×[0, 0.41] ×[0, 0.41] \ {k(x1 , x2 ) − (0.5, 0.2)k2 ≤ 0.05}.
(47)
Boundary conditions are chosen as in the previous example except for a change to unsteady inflow boundary
conditions:
u(0, y, z, t) = uD (t) = (9/4 · ū(t)) · 16 · y(dy −y)/d2y · z(dz −z)/d2z · (1, 0, 0),
with ū(t) the average inflow velocity which is time-dependent, ū(t) = sin(πt/8). The considered time interval
is [0, 8s], the viscosity is set to ν = 10−3 s.t. the Reynolds number varies between Re = 0 and Re = 100.
4.5.2. Drag and Lift.
Again, the quantities of interest in this example are the (maximal and minimal) drag and lift forces cD ,
cL that act on the disc. These are defined as
Z
Z
∂u
1
∂u
1
ν
− pn · ex ds
cL := 2
ν
− pn · ey ds
cD := 2
ūmax rh Γ◦
∂n
ūmax rh Γ◦
∂n
with ūmax = maxt∈[0,8] ū(t) = 1, r = 0.05, h = 0.41 and Γ◦ the surface of the obstacle.
20
Figure 5: Used mesh (left) and solution at t = 0.4 (right) for the benchmark problem in section 4.5.
4.5.3. Numerical setup.
We use an unstructured tetrahedral mesh consisting of 5922 elements, cf. Figure 5. For the time
discretization we use the modified fractional-step-θ-scheme discussed in section 3.3. To compensate for the
increase in the spatial accuracy for increasing polynomial degree k, we adapt the number of time steps
accordingly. The computations were carried out on a shared-memory computer with 24 cores. We comment
on details of the computing times in remark 11.
k
k
k
k
=2
=3
=4
=5
#ndof [K]
max cD
max cL
min cL
169
343
595
939
3.43046
3.29331
3.29853
3.29798
0.00262
0.00277
0.00278
0.00278
-0.016289
-0.011099
-0.010762
-0.011054
0.0080
0.0080
0.0040
0.0040
0.0028
0.0028
-0.010992
-0.010999
0.01
0.005
ref. [59]
11 432
89 760
3.2963
3.2978
ref. [60]
7 036
3.2968
ref. [58]
[3.2,3.3]
∆t [s]
comp. time [s]
492
964
3 087
6 670
×
×
×
×
24
24
24
24
35 550 × 24
214 473 × 48
-0.011
[0.002,0.004]
Table 5: Numerical results for the benchmark problem “3D3Z” in [58]. Results obtained with different polynomial degrees and
reference values.
4.5.4. Numerical results.
In Table 5 the results obtained are compared with the literature in terms of accuracy and computing
time. We observe that we can achieve the same level of accuracy as the results in [59] (and [60]) with a
computing time which is dramatically smaller. We note that the computing time per degree of freedom is
actually worse than in [59]. Nevertheless, the same accuracy is achieved with much less degrees of freedoms
using the high order method (k > 2), s.t. our computations exceed the performance results in the literature.
In the study [59] one of the conclusions is that their third order method (Q2 /P1disc ) is much more efficient
compared to lower order methods. We extend this conclusion in the sense that the use of even higher order
discretizations, i.e. k > 2, increases efficiency even further. Moreover, high order discretizations can be
implemented efficiently. One important component for the efficient handling of the Navier-Stokes equations
with our high order discretization is the time integration using operator-splitting.
Remark 11 (Computation times). We remark on the computation for the case k = 5. In that computation approximately 65% of the computing time has been spend on the solution of linear systems for Stokes-type
problems, 30% on convection operator evaluations and 5% on the setup and remaining operations. To solve
the implicit convection problem (Step 2 in (41)) an average of 20 iterations has been applied.
Remark 12 (Linear systems). In the test cases in this section we only applied direct solvers which is
possible due to the (comparably) small size of the arising linear systems. For problems with increasing
21
complexity efficient linear solvers are mandatory. The development of suitable preconditioners of the Stokes
problem is based on efficient preconditioning of the bilinear form Ah . For the scalar problem we could
show poly-logarithmic bounds in k for the condition number of standard p-version domain decomposition
preconditioners, cf. [52]. We plan to investigate suitable generalizations of this preconditioner for the Stokes
problem in the future.
5. Conclusion
We presented and discussed a combined DG/HDG discretization tailored for efficiency. We summarize
the core components. We split the Navier-Stokes problem into linear Stokes-type problems and hyperbolic
transport problems by means of operator-splitting time integration. For the Stokes-type problem we use an
H(div)-conforming Hybrid DG formulation with a new modification: the projected jumps formulation. The
Hybrid DG formulation facilitates the efficient solution of linear systems compared to other DG methods.
The projected jumps formulation improves its efficiency even more. For the hyperbolic transport problem we
apply a standard DG formulation. In numerical test cases we demonstrated the performance of the method.
Acknowledgements
The authors greatly appreciate the contribution of Philip Lederer related to the calculation of LBBconstants, cf. remark 4.
References
[1] Reed WH, Hill T. Triangular mesh methods for the neutron transport equation. Los Alamos Report LA-UR-73-479 1973;
.
[2] Lesaint P, Raviart PA. On a finite element method for solving the neutron transport equation. Mathematical aspects of
finite elements in partial differential equations 1974; 33:89–123.
[3] Johnson C, Pitkäranta J. An analysis of the discontinuous Galerkin method for a scalar hyperbolic equation. Mathematics
of computation 1986; 46(173):1–26.
[4] Bassi F, Rebay S. High-order accurate discontinuous finite element solution of the 2D Euler equations. Journal of computational physics 1997; 138(2):251–285.
[5] Bassi F, Rebay S, Mariotti G, Pedinotti S, Savini M. A high-order accurate discontinuous finite element method for
inviscid and viscous turbomachinery flows. Proceedings of 2nd European Conference on Turbomachinery, Fluid Dynamics
and Thermodynamics, Technologisch Instituut, Antwerpen, Belgium, 1997; 99–108.
[6] Cockburn B, Shu CW. The Runge–Kutta discontinuous Galerkin method for conservation laws V: multidimensional
systems. Journal of Computational Physics 1998; 141(2):199–224.
[7] Arnold DN, Brezzi F, Cockburn B, Marini LD. Unified analysis of discontinuous Galerkin methods for elliptic problems.
SIAM journal on numerical analysis 2002; 39(5):1749–1779.
[8] Houston P, Schwab C, Süli E. Discontinuous hp-finite element methods for advection-diffusion-reaction problems. SIAM
Journal on Numerical Analysis 2002; 39(6):2133–2163.
[9] Rivière B. Discontinuous Galerkin methods for solving elliptic and parabolic equations: theory and implementation. Society
for Industrial and Applied Mathematics, 2008.
[10] Di Pietro DA, Ern A. Mathematical Aspects of Discontinuous Galerkin Methods, vol. 69. Springer Science & Business
Media, 2012.
[11] Hesthaven JS, Warburton T. Nodal discontinuous Galerkin methods: algorithms, analysis, and applications. Springer
Science & Business Media, 2007.
[12] Karniadakis G, Sherwin S. Spectral/hp element methods for computational fluid dynamics. Oxford University Press, 2013.
[13] Girault V, Raviart PA. Finite element methods for Navier-Stokes equations: theory and algorithms, vol. 5. Springer
Science & Business Media, 2012.
[14] Elman HC, Silvester DJ, Wathen AJ. Finite elements and fast iterative solvers: with applications in incompressible fluid
dynamics. Oxford University Press, 2014.
[15] Donea J, Huerta A. Finite element methods for flow problems. John Wiley & Sons, 2003.
[16] Toselli A. hp discontinuous Galerkin approximations for the Stokes problem. Mathematical Models and Methods in Applied
Sciences 2002; 12(11):1565–1597.
[17] Schötzau D, Schwab C, Toselli A. Mixed hp-DGFEM for incompressible flows. SIAM Journal on Numerical Analysis 2002;
40(6):2171–2194.
[18] Girault V, Rivière B, Wheeler M. A discontinuous Galerkin method with nonoverlapping domain decomposition for the
Stokes and Navier-Stokes problems. Mathematics of Computation 2005; 74(249):53–84.
22
[19] Cockburn B, Kanschat G, Schötzau D. A locally conservative LDG method for the incompressible Navier-Stokes equations.
Mathematics of Computation 2005; 74(251):1067–1095.
[20] Cockburn B, Kanschat G, Schötzau D. A note on discontinuous Galerkin divergence-free solutions of the Navier–Stokes
equations. Journal of Scientific Computing 2007; 31(1-2):61–73.
[21] Brezzi F, Fortin M. Mixed and hybrid finite element methods, vol. 15. Springer Science & Business Media, 2012.
[22] Lehrenfeld C. Hybrid discontinuous Galerkin methods for solving incompressible flow problems. Rheinisch-Westfalischen
Technischen Hochschule Aachen 2010; .
[23] Egger H, Schöberl J. A hybrid mixed discontinuous Galerkin method for convection–diffusion problems. J. Numer. Anal
2008; .
[24] Nguyen NC, Peraire J, Cockburn B. An implicit high-order hybridizable discontinuous Galerkin method for linear
convection–diffusion equations. Journal of Computational Physics 2009; 228(9):3232–3254.
[25] Cockburn B, Gopalakrishnan J, Lazarov R. Unified hybridization of discontinuous Galerkin, mixed, and continuous
Galerkin methods for second order elliptic problems. SIAM Journal on Numerical Analysis 2009; 47(2):1319–1365.
[26] Cockburn B, Gopalakrishnan J, Nguyen N, Peraire J, Sayas FJ. Analysis of HDG methods for Stokes flow. Mathematics
of Computation 2011; 80(274):723–760.
[27] Cesmelioglu A, Cockburn B, Nguyen NC, Peraire J. Analysis of HDG methods for Oseen equations. Journal of Scientific
Computing 2013; 55(2):392–431.
[28] Zhai Q, Zhang R, Wang X. A hybridized weak Galerkin finite element scheme for the Stokes equations. Science China
Mathematics 2015; :1–18.
[29] Pietro DAD, Ern A, Lemaire S. An arbitrary-order and compact-stencil discretization of diffusion on general meshes based
on local reconstruction operators. Comp. Meth. Appl. Math. 2014; 14(4):461–472.
[30] Di Pietro D, Ern A, Lemaire S. A review of hybrid high-order methods: formulations, computational aspects, comparison
with other methods. Technical Report HAL-01163569 September 2015.
[31] Cockburn B, Pietro DD, Ern A. Bridging the hybrid high-order and hybridizable discontinuous galerkin methods. Technical
Report HAL-01115318 July 2015.
[32] Egger H, Waluga C. hp analysis of a hybrid DG method for Stokes flow. IMA Journal of Numerical Analysis 2012; :drs018.
[33] Nguyen N, Peraire J, Cockburn B. A hybridizable discontinuous Galerkin method for Stokes flow. Computer Methods in
Applied Mechanics and Engineering 2010; 199(9):582–597.
[34] Carrero J, Cockburn B, Schötzau D. Hybridized globally divergence-free LDG methods. part i: The Stokes problem.
Mathematics of computation 2006; 75(254):533–563.
[35] Cockburn B, Gopalakrishnan J. Incompressible finite elements via hybridization. part i: The Stokes system in two space
dimensions. SIAM Journal on Numerical Analysis 2005; 43(4):1627–1650.
[36] Cockburn B, Gopalakrishnan J. Incompressible finite elements via hybridization. part ii: The Stokes system in three space
dimensions. SIAM Journal on Numerical Analysis 2005; 43(4):1651–1672.
[37] Könnö J, Stenberg R. Numerical computations with H(div)-finite elements for the Brinkman problem. Computational
Geosciences 2012; 16(1):139–158.
[38] Qiu W, Shi K. An HDG method for linear elasticity with strong symmetric stresses. arXiv preprint arXiv:1312.1407 2013;
.
[39] Oikawa I. A hybridized discontinuous Galerkin method with reduced stabilization. Journal of Scientific Computing 2014;
:1–14.
[40] Oikawa I. A reduced HDG method for the Stokes equations. arXiv preprint arXiv:1502.01833 2015; .
[41] Qiu W, Shi K. A superconvergent HDG method for the incompressible Navier-Stokes equations on general polyhedral
meshes. CoRR 2015; abs/1506.07543.
[42] Kirby RM, Sherwin SJ, Cockburn B. To CG or to HDG: a comparative study. Journal of Scientific Computing 2012;
51(1):183–212.
[43] Huerta A, Angeloski A, Roca X, Peraire J. Efficiency of high-order elements for continuous and discontinuous Galerkin
methods. International Journal for Numerical Methods in Engineering 2013; 96(9):529–560.
[44] Ascher UM, Ruuth SJ, Wetton BT. Implicit-explicit methods for time-dependent partial differential equations. SIAM
Journal on Numerical Analysis 1995; 32(3):797–823.
[45] Ascher UM, Ruuth SJ, Spiteri RJ. Implicit-explicit Runge-Kutta methods for time-dependent partial differential equations.
Applied Numerical Mathematics 1997; 25(2):151–167.
[46] Kanevsky A, Carpenter MH, Gottlieb D, Hesthaven JS. Application of implicit–explicit high order Runge–Kutta methods
to discontinuous-Galerkin schemes. Journal of Computational Physics 2007; 225(2):1753–1781.
[47] Strang G. On the construction and comparison of difference schemes. SIAM Journal on Numerical Analysis 1968; 5(3):506–
517.
[48] Maday Y, Patera AT, Rønquist EM. An operator-integration-factor splitting method for time-dependent problems: application to incompressible fluid flow. Journal of Scientific Computing 1990; 5(4):263–292.
[49] Chrispell J, Ervin V, Jenkins E. A fractional step θ-method for convection–diffusion problems. Journal of Mathematical
Analysis and Applications 2007; 333(1):204–218.
[50] Zaglmayr S. High order finite element methods for electromagnetic field computation. PhD Thesis, JKU Linz 2006.
[51] Arnold DN. An interior penalty finite element method with discontinuous elements. SIAM Journal on Numerical Analysis
1982; 19(4):742–760.
[52] Schöberl J, Lehrenfeld C. Domain decomposition preconditioning for high order hybrid discontinuous Galerkin methods
on tetrahedral meshes. Advanced finite element methods and applications. Springer, 2013; 27–56.
[53] Brezzi F, Manzini G, Marini D, Pietra P, Russo A. Discontinuous finite elements for diffusion problems. Atti Convegno in
23
onore di F. Brioschi (Milano 1997), Istituto Lombardo, Accademia di Scienze e Lettere 1999; :197–217.
[54] Linke A. On the role of the Helmholtz decomposition in mixed methods for incompressible flows and a new variational
crime. Computer Methods in Applied Mechanics and Engineering 2014; 268:782–800.
[55] Dubiner M. Spectral methods on triangles and other domains. Journal of Scientific Computing 1991; 6(4):345–390.
[56] Glowinski R. Finite element methods for incompressible viscous flow. Handbook of numerical analysis 2003; 9:3–1176.
[57] Chorin AJ. The numerical solution of the Navier-Stokes equations for an incompressible fluid. Bulletin of the American
Mathematical Society 1967; 73(6):928–931.
[58] Schäfer M, Turek S, Durst F, Krause E, Rannacher R. Benchmark computations of laminar flow around a cylinder. Flow
simulation with high-performance computers II 1996; :547–566.
[59] Bayraktar E, Mierka O, Turek S. Benchmark computations of 3D laminar flow around a cylinder with CFX, OpenFOAM
and FeatFlow. International Journal of Computational Science and Engineering 2012; 7(3):253–266.
[60] John V. On the efficiency of linearization schemes and coupled multigrid methods in the simulation of a 3d flow around a
cylinder. International Journal for Numerical Methods in Fluids 2006; 50(7):845–862.
[61] ngsflow: Flow solver package for NGSolve. www.sourceforge.net/projects/ngsflow/.
[62] Schöberl J. C++11 implementation of finite elements in NGSolve. Technical Report ASC-2014-30, Institute for Analysis and Scientific Computing September 2014. URL http://www.asc.tuwien.ac.at/~schoeberl/wiki/publications/
ngs-cpp11.pdf.
[63] Kovasznay L. Laminar flow behind a two-dimensional grid. Mathematical Proceedings of the Cambridge Philosophical
Society, vol. 44, Cambridge Univ Press, 1948; 58–62.
[64] PARDISO Solver project. www.pardiso-project.org.
[65] Schenk O, Gärtner K. Solving unsymmetric sparse systems of linear equations with PARDISO. J. of Future Generation
Computer Systems 2004; 20:475–487.
[66] FEATFLOW Finite element software for the incompressible Navier-Stokes equations. www.featflow.de.
[67] Turek S, Becker C. FEATFLOW: Finite element software for the incompressible Navier-Stokes equations, User Manual
Release 1.1. Preprint, Heidelberg 1998; 4.
24
| 5cs.CE
|
The Case for a Mixed-Initiative Collaborative Neuroevolution Approach
Sebastian Risi, Jinhong Zhang, Rasmus Taarnby, Peter Greve, Jan Piskur, Antonios Liapis, Julian Togelius
arXiv:1408.0998v1 [cs.NE] 5 Aug 2014
IT University Copenhagen, 2300 Copenhagen, Denmark
{sebr, jinh, reta, pgre, japi, anli, juto}@itu.dk
Abstract
It is clear that the current attempts at using algorithms to create artificial neural networks have had mixed success at best
when it comes to creating large networks and/or complex behavior. This should not be unexpected, as creating an artificial
brain is essentially a design problem. Human design ingenuity still surpasses computational design for most tasks in most
domains, including architecture, game design, and authoring
literary fiction. This leads us to ask which the best way is
to combine human and machine design capacities when it
comes to designing artificial brains. Both of them have their
strengths and weaknesses; for example, humans are much too
slow to manually specify thousands of neurons, let alone the
billions of neurons that go into a human brain, but on the other
hand they can rely on a vast repository of common-sense understanding and design heuristics that can help them perform
a much better guided search in design space than an algorithm. Therefore, in this paper we argue for a mixed-initiative
approach for collaborative online brain building and present
first results towards this goal.
With around 200 billion neurons and 125 trillion
synapses, the human brain is the most complex system
known to exist. The brain’s structural complexity, with intricate synaptic motifs that repeat throughout it, gives rise to
our unique mental abilities. Additionally, the brains plasticity allows us to learn new abilities throughout our life (e.g.
learning a new language) and to change our behavior based
on past experience.
Therefore, creating similar artificial structures by recapitulating the process that created intelligence on earth, is an
intriguing possibility. In this context, neuroevolution (i.e.
evolving artificial neural networks (ANNs) via evolutionary
algorithms) has shown promising results in a variety of different domains (Floreano et al., 2008; Stanley and Miikkulainen, 2004; Yao, 1999). However, these results still pale in
comparison to the capabilities of natural brains. The reasons
for this are manifold. Especially the problem of deceptive
fitness landscapes (i.e. mutations increase fitness but actually lead further away from the final objective) has limited
the scope of problems amenable to evolutionary algorithms.
Research in circumventing this problem has mainly focused on two different ideas. First, novelty search (Lehman
and Stanley, 2011), a method that rewards novel behaviors instead of rewarding objective performance, has shown
promise and significantly outperforms objective-based approaches in a variety of different domains. Other approaches
are based on interactive evolutionary computation (IEC)
methods, wherein the human user guides evolution by repeatedly choosing from a set of candidates. Woolley and
Stanley (2011) showed that interactive evolution can help to
discover artifacts which are very hard to evolve with traditional evolutionary approaches. Recently Woolley and Stanley (2014) combined IEC with novelty search, demonstrating that the approaches complement each other and together
address some of the challenges that each method struggles
with by itself (e.g. novelty search can get lost in large search
spaces, interactive evolution is limited by user fatigue). In
a related approach, Bongard and Hornby (2013) recently
demonstrated that human input and objective-based search
can also be combined synergistically to solve challenging
robotic domains.
While these human in the loop approaches have shown
promising results, we argue in this paper that they do not exploit the whole range of human intuition, ability to identify
promising stepping stones, or collaborative problem-solving
skills that could ultimately allow us to create more brain-like
artificial structures. For example, collaborative games like
Foldit hint at the power of crowdsourcing the human brain’s
natural abilities for specific tasks (e.g. pattern matching, spatial reasoning), which are hard to solve by computational approaches. However, in traditional IEC applications the role
of the user is often reduced to solely judging the created artifacts and only “nudging” evolution by deciding between a
discrete choice of candidates. In other words, only the computer creates content (e.g. images, ANNs, etc.) and the role
of the human is to guide evolution to content they prefer.
An approach that does require significantly more input
from the user is a mixed-initiative process (Liapis et al.,
2014), wherein both the computer and the human take turns
in creating content. Yet while mixed-initiative based approaches have shown promise in the context of procedurally
generated content for games (Liapis et al., 2013; Yannakakis
Figure 1: BrainCrafter Web Interface. Beta version online
at: braincrafter.dk
et al., 2014), they have not yet been applied to the evolution of large-scale ANNs for complex tasks. However, the
promise of such an approach is a system that can benefit
from the different skillsets of a human and a computational
creator. For example, while a human user might develop
an intuition about promising domain-dependent network topographies, a computational method is likely more effective
at fine-tuning specific synaptic weights. Therefore, we argue
for a collaborative mixed-initiative approach in which both
the computer and the collaborating human users can take initiative and propose changes to an evolving neural networks.
That is, at any point the human can revert to just doing selection and let the evolutionary algorithm serve up new content
to judge, or decide to jump in and have a more active role in
the design of the ANN.
Our recent work has taken steps towards this goal by focusing on two parallel lines of research. First, before introducing a mixed-initiative approach it is useful to determine
how good we as humans are at building complex ANNs for
certain tasks (without evolution) and if collaborating with
other users proves useful. Insights from this experiment
should in turn provide useful clues about the strengths of
human ANN design and most importantly, non-intuitive aspects of the design process we tend to struggle with (i.e.
aspects which would benefit most from an assisting computational creator). In this context we recently introduced
BrainCrafter (braincrafter.dk), which is an online application that allows the user to build ANNs for specific control programs (e.g. a robot that must traverse a maze) by
adding neurons and connections in a drag&drop like fashion (Figure 1). While building ANNs the users can observe
the resulting simulated robot behaviors in real-time, proving
insights into the effects of different network modifications.
BrainCrafter also allows users to collaborate by building on
high-scoring solutions created by other people.
While the ongoing BrainCrafter experiment should pro-
Figure 2: Picture CPPN-Compiler User Interface. Beta version online at: rasmustaarnby.dk/thesis
vide insights into our abilities to collaboratively construct
complex networks, more challenging domains will likely require the orchestrated effort of both human and computational creators. Therefore extending BrainCrafter to support the users’ collaborative engineering efforts through a
mixed-initiative approach is an important next goal. To allow seamlessly switching between interactive evolution and
manual ANN design will require an ANN representation that
(1) can be edited easily by the user on a local (e.g. individual
neurons) and global level (e.g. overall network topography
and topology), (2) is evolvable and compact, and (3) allows
the system to produce meaningful suggestions based on the
user’s input.
In this context we are building on the a generative encoding called compositional pattern producing networks
(CPPNs; Stanley (2007)) and the recently introduced concept of a CPPN-Compiler (Risi, 2013). CPPNs, which
are based on principles of how natural organisms develop,
allow the compact encoding of complex patterns, from
two-dimensional images (Secretan et al., 2011) and threedimensional forms (Clune and Lipson, 2011) to large-scale
ANNs (Stanley et al., 2009; Clune et al., 2011; Risi and
Stanley, 2012). The idea behind the CPPN-Compiler is to
allow the user to directly compile a high-level description
of the desired starting structure into the CPPN itself. We
are currently exploring the benefits of this approach through
the collaborative interactive evolution of images (Figure 2),
in which the user can draw a vector image and annotate
it with important regularities like symmetry. The CPPNCompiler then compiles this high-level description into the
CPPN itself. Since the compiled CPPN now directly embodies the annotated domain regularities (e.g. bilateral symmetric arms), the produced offspring show meaningful variations that nevertheless share common features. Considering the insight that CPPNs can also produce large-scale
ANNs (Stanley et al., 2009) and can be modified to create
Network
Editing
CPPN-Compiler
Synapse
Phenotype-toGenotype
Symmetry
x1 y1 x2
y2
CPPNRepresentation
ES-HyperNEAT
Neuron
GeneratehVariations
Floreano, D., Dürr, P., and Mattiussi, C. (2008). Neuroevolution:
from architectures to learning. Evol. Intelligence, 1(1):47–62.
Lehman, J. and Stanley, K. O. (2011). Abandoning objectives:
Evolution through the search for novelty alone. Evolutionary Computation, 19(2):189–223.
…
Liapis, A., Shaker, N., and Smith, G. (2014). Mixed-initiative approaches. In Shaker, N., Togelius, J., and Nelson, M. J., editors, Procedural Content Generation in Games: A Textbook
and an Overview of Current Research. Springer.
Task
Liapis, A., Yannakakis, G. N., and Togelius, J. (2013). Sentient sketchbook: Computer-aided game level authoring. In
Proceedings of ACM Conference on Foundations of Digital
Games (FDG).
InteractivehEvolution
Risi, S. (2013). A compiler for cppns: Transforming phenotypic
descriptions into genotypic representations. In 2013 AAAI
Fall Symposium Series.
Figure 3: Mixed-initiative Neuroevolution Framework
Risi, S. and Stanley, K. O. (2012). An enhanced hypercube-based
encoding for evolving the placement, density, and connectivity of neurons. Artificial Life, 18(4):331–363.
certain neural topologies (ES-HyperNEAT; Risi and Stanley (2012)), opens up the intriguing possibility of a neural
CPPN-Compiler. Our current efforts focus on such a compiler that will form the backbone for our mixed-initiative
BrainCrafter application.
The envisioned collaborative mixed-initiative system is
depicted in Figure 3. The users can collaboratively construct neural networks and annotate them with regularities
(e.g. symmetry, repetition, etc.), which are used by the computational creator to construct the internal CPPN model and
in turn propose meaningful variations to the user. At any
point in the process the human can revert to just doing selection or decide to directly edit the ANN, wherein phenotypic edits are directly compiled back into the genotype. The
promise of the proposed system is that it could allow a variety of tasks to be solved by many people online within a
mixed-initiative environment, which have heretofore proven
too difficult. We expect this project will profoundly impact
the fields of ANN research and potentially also deepen our
understanding of the way biological neural networks solve
certain problems.
Secretan, J., Beato, N., D’Ambrosio, D., Rodriguez, A., Campbell,
A., Folsom-Kovarik, J., and Stanley, K. (2011). Picbreeder: A
case study in collaborative evolutionary exploration of design
space. Evolutionary Computation, 19(3):373–403.
References
Bongard, J. C. and Hornby, G. S. (2013). Combining fitness-based
search and user modeling in evolutionary robotics. In Proceeding of the fifteenth annual conference on Genetic and
evolutionary computation conference, pages 159–166. ACM.
Clune, J. and Lipson, H. (2011). Evolving three-dimensional objects with a generative encoding inspired by developmental
biology. In Proceedings of the European Conference on Artificial Life, See http://EndlessForms. com.
Clune, J., Stanley, K. O., Pennock, R. T., and Ofria, C. (2011). On
the performance of indirect encoding across the continuum
of regularity. Evolutionary Computation, IEEE Transactions
on, 15(3):346–367.
Stanley, K. O. (2007). Compositional pattern producing networks:
A novel abstraction of development. Genetic Programming
and Evolvable Machines Special Issue on Developmental
Systems, 8(2):131–162.
Stanley, K. O., D’Ambrosio, D. B., and Gauci, J. (2009). A
hypercube-based indirect encoding for evolving large-scale
neural networks. Artificial Life, 15(2):185–212.
Stanley, K. O. and Miikkulainen, R. (2004). Competitive coevolution through evolutionary complexification. JAIR, 21:63–100.
Woolley, B. G. and Stanley, K. O. (2011). On the deleterious effects of a priori objectives on evolution and representation.
In Proceedings of the 13th annual conference on Genetic and
evolutionary computation, pages 957–964. ACM.
Woolley, B. G. and Stanley, K. O. (2014). A novel humancomputer collaboration: Combining novelty search with interactive evolution. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO-2014), New York,
NY, USA. ACM.
Yannakakis, G. N., Liapis, A., and Alexopoulos, C. (2014). Mixedinitiative co-creativity. In Proceedings of the 9th Conference
on the Foundations of Digital Games.
Yao, X. (1999). Evolving artificial neural networks. Proceedings
of the IEEE, 87(9).
| 9cs.NE
|
arXiv:1704.03767v1 [cs.DC] 12 Apr 2017
Parallelized Kendall’s Tau Coefficient Computation via
SIMD Vectorized Sorting On Many-Integrated-Core
Processors
Yongchao Liua,∗, Tony Pana , Oded Greena , Srinivas Alurua,∗
a School
of Computational Science & Engineering, Georgia Institute of Technology, Atlanta,
GA 30332, USA
Abstract
Pairwise association measure is an important operation in data analytics. Kendall’s
tau coefficient is one widely used correlation coefficient identifying non-linear relationships between ordinal variables. In this paper, we investigated a parallel
algorithm accelerating all-pairs Kendall’s tau coefficient computation via single
instruction multiple data (SIMD) vectorized sorting on Intel Xeon Phis by taking advantage of many processing cores and 512-bit SIMD vector instructions.
To facilitate workload balancing and overcome on-chip memory limitation, we
proposed a generic framework for symmetric all-pairs computation by building
provable bijective functions between job identifier and coordinate space. Performance evaluation demonstrated that our algorithm on one 5110P Phi achieves
two orders-of-magnitude speedups over 16-threaded MATLAB and three ordersof-magnitude speedups over sequential R, both running on high-end CPUs. Besides, our algorithm exhibited rather good distributed computing scalability
with respect to number of Phis. Source code and datasets are publicly available
at http://lightpcc.sourceforge.net.
Keywords: Pairwise correlation; Kendall’s tau coefficient; all-pairs
computation; many integrated core; Xeon Phi
1. Introduction
Identifying interesting pairwise association between variables is an important
operation in data analytics. In bioinformatics and computational biology, one
typical application is to mine gene co-expression relationship via gene expression
data, which can be realized by query-based gene expression database search [1]
∗ Corresponding
author
Email addresses: [email protected] (Yongchao Liu), [email protected] (Tony Pan),
[email protected] (Oded Green), [email protected] (Srinivas Aluru)
1 Preliminary work was presented in the 28th International Symposium on Computer Architecture and High Performance Computing, Los Angeles, USA, 2016
Preprint submitted to Journal of Parallel and Distributed Computing
April 13, 2017
or gene co-expression network analysis [2]. For gene expression database search,
it targets to select the subject genes in the database that are co-expressed
with the query gene. One approach is to first define some pairwise correlation/dependence measure over gene expression profiles across multiple samples
(gene expression profiles for short) and then rank query-subject gene pairs by
their scores. For gene co-expression networks, nodes usually correspond to genes
and edges represent significant gene interactions inferred from the association
of gene expression profiles. To construct a gene co-expression network, allpairs computation over gene expression profiles is frequently conducted based
on linear (e.g. [3] [4] [5]) or non-linear (e.g. [6] [7] [8]) co-expression measures.
A variety of correlation/dependence measures have been proposed in the literature and among them, Pearson’s product-moment correlation coefficient [9]
(or Pearson’s r correlation) is the most widely used correlation measure [10].
However, this correlation coefficient is only applicable to linear correlations. In
contrast, Spearman’s rank correlation coefficient [11] (or Spearman’s ρ coefficient) and Kendall’s rank correlation coefficient [12] (or Kendall’s τ coefficient)
are two commonly used measures for non-linear correlations [13]. Spearman’s
ρ coefficient is based on Pearson’s r coefficient but applies to ranked variables,
while Kendall’s τ coefficient tests the association between ordinal variables.
These two rank-based coefficients were shown to play complementary roles in
the cases when Pearson’s r is not effective [14]. Among other non-linear measures, mutual information [15] [16] [17], Euclidean distance correlation [18] [19],
Hilbert-Schmidt information criterion [20], and maximal information criterion
[21] are frequently used as well. In addition, some unified frameworks for pairwise dependence assessments were proposed in the literature (e.g. [22]).
Kendall’s τ coefficient (τ coefficient for short) measures the ordinal correlation between two vectors of ordinal variables. Given two ordinal vectors
u = {u1 , u2 , ..., un } and v = {v1 , v2 , ..., vn }, where variables ui and vi are both
ordinal (0 ≤ i < n), the τ coefficient computes the correlation by counting
the number of concordant pairs nc and the number of discordant pairs nd by
treating ui and vi as a joint ordinal variable (ui , vi ). For the τ coefficient, a
pair of observations (ui , vi ) and (uj , vj ), where i 6= j, is deemed as concordant
if ui > uj and vi > vj or ui < uj and vi < vj , and discordant if ui > uj and
vi < vj or ui < uj and vi > vj . Note that if ui = uj or vi = vj , this pair is
considered neither concordant nor discordant.
In our study, we will consider two categories of τ coefficient, namely Tau-a
(denoted as τA ) and Tau-b (denoted as τB ). τA does not take into account tied
elements in each vector and is defined as
τA =
nc − nd
n0
(1)
where n0 = n(n − 1)/2. If all elements in each vector are distinct, we have
nc + nd = n0 and can therefore re-write Equation (1) as
τA =
2nd
n0 − 2nd
= 1−
n0
n0
2
(2)
As opposed to τA , τB makes adjustments for ties and is computed as
nc − nd
τB = p
(3)
(n0 − n1 )(n0 − n2 )
P ′ ′
P ′ ′
′
′
where n1 =
i ui (ui − 1)/2 and n2 =
i vi (vi − 1)/2. ui (vi ) denotes the
cardinality of the i-th group of ties for vector u (v). Within a vector, each
distinct value defines one tie group and this value acts as the identifier of the
corresponding group. The cardinality of a tie group is equal to the number of
elements in the vector whose values are identical to the identifier of the tie group.
If all elements in either vector are distinct, τB will equal τA as n1 = n2 = 0.
This indicates τA is a special case of τB and an implementation of τB will cover
τA inherently. In addition, from the definitions of τA and τB , we can see that
the computation of the τ coefficient between u and v is commutable.
Besides the values of correlation coefficients, some applications need to calculate P -value statistics to infer statistical significance between variables. For this
purpose, one approach is permutation test [23]. However, a permutation test
may need a substantial number of pairwise τ coefficient computation [24] even
for moderately large n, thus resulting in prohibitively long times for sequential
execution. In the literature, parallelizing pairwise τ coefficient computation has
not yet been intensively explored. One recent work is from Wang et al. [25],
which accelerated the sequential τ coefficient computation in R [26] based on
Hadoop MapReduce [27] parallel programming model. In [25], the sequential
all-pairs τ coefficient implementation in R was shown extremely slow on largescale datasets. In our study, we further confirmed this observation through our
performance assessment (refer to section 5.2).
In this paper, we parallelized all-pairs τ coefficient computation on Intel
Xeon Phis based on Many-Integrated-Core (MIC) architecture, the first work
accelerating all-pairs τ coefficient computation on MIC processors to the best
of our knowledge. This work is a continuation from our previous parallelization of all-pairs Pearson’s r coefficient on Phi clusters [28] and further enriches
our LightPCC library (http://lightpcc.sourceforge.net) targeting parallel
pairwise association measures between variables in big data analytics. In this
work, we have investigated three variants, namely the naı̈ve variant, the generic
sorting-enabled (GSE) variant and the vectorized sorting-enabled (VSE) variant, built upon three pairwise τ coefficient kernels, i.e. the naı̈ve kernel, the GSE
kernel and the VSE kernel, respectively. Given two ordinal vectors u and v of n
elements each, the naı̈ve kernel enumerates all possible pairs of joint variables
(ui , vi ) (0 ≤ i < n) to obtain nc and nd , resulting in O(n2 ) time complexity. In
contrast, both the GSE and VSE kernels take sorting as the core and manage
to reduce the time complexity to O(n log n).
Given m vectors of n elements each, the overall time complexity would be
O(m2 n2 ) for the naı̈ve variant and O(m2 n log n) for the GSE and VSE variants. The VSE variant enhances the GSE one by exploiting 512-bit wide single
instruction multiple data (SIMD) vector instructions in MIC processors to implement fast SIMD vectorized pairwise merge of sorted subarrays. Furthermore,
3
to facilitate workload balancing and overcome on-chip memory limitation, we
investigated a generic framework for symmetric all-pairs computation by pioneering to build a provable, reversible and bijective relationship between job
identifier and coordinate space in a job matrix.
The performance of our algorithm was assessed using a collection of real
whole human genome gene expression datasets. Our experimental results demonstrates that the VSE variant performs best on both the multi-threaded CPU
and Phi systems, compared to the other two variants. We further compared
our algorithm with the all-pairs τ coefficient implementations in the widely
used MATLAB [29] and R [26], revealing that our algorithm on a single 5110P
Phi achieves up to 812 speedups over 16-threaded MATLAB and up to 1,166
speedups over sequential R, both of which were benchmarked on high-end CPUs.
In addition, our algorithm exhibited rather good distributed computing scalability with respect to number of Phis.
2. Intel Many-Integrated-Core (MIC) Architecture
Intel MIC architecture targets to combine many Intel processor cores onto
a single chip and has already led to the release of two generations of MIC processors. The first generation is code named as Knights Corner (KNC) and
the second generation code named as Knights Landing (KNL) [30]. KNC is
a PCI Express (PCIe) connected coprocessor that must be paired with Intel
Xeon CPUs. KNC is actually a shared-memory computer [31] with full cache
coherency over the entire chip and running a specialized Linux operating system over many cores. Each core adopts an in-order micro-architecture and has
four hardware threads offering four-way simultaneous multithreading. Besides
scalar processing, each core is capable of vectorized processing from a newly
designed vector processing unit (VPU) featuring 512-bit wide SIMD instruction
set architecture (ISA). For KNC, each core has only one VPU and this VPU is
shared by all active hardware threads running on the same core. Each 512-bit
vector register can be split to either 8 lanes with 64 bits each or 16 lanes with 32
bits each. Note that the VPU does not support legacy SIMD ISAs such as the
Streaming SIMD extensions (SSE) series. As for caches, each core has separate
L1 instruction and data caches of size 32 KB each, and a 512 KB L2 cache
interconnected via a bidirectional ring bus with the L2 caches of all other cores
to form a unified shared L2 cache over the chip. The cache line size is 64 bytes.
In addition, two usage models can be used to invoke KNC: offload model and
native model, where the former relies on compiler pragmas/directives to offload
highly-parallel parts of an application to KNC, while the latter treats KNC as
symmetric multiprocessing computers. While primarily focusing on KNC Phis
in this work, we note that our KNC-based implementations can be easily ported
onto KNL processors, as KNL implements a superset of KNC instruction sets.
We expect that our implementations will be portable to future Phis as well.
4
Table 1: A list of notions used
Notion
m
n
u
ui
nc
nd
n0
n1
n2
n3
τA
τB
SX
|SX |
vX
Description
number of vectors
number of elements per vector
vector u of n elements, likewise for v
i-th element of vector u, likewise for vi
number of concordant variable pairs
number of discordant variable pairs
n(n − 1)/2
P
′
′
′
Pi u′i (u′i − 1)/2, u′i is size of the i-th tie group in u
Pi vi (vi − 1)/2, vi is size of the i-th tie group in v
i wi (wi − 1)/2, wi is size of the i-th joint tie group for u and v
(nc − nd )/n
p0
(nc − nd )/ (n0 − n1 )(n0 − n2 )
a sorted subarray
length of the sorted subarray SX
a 512-bit SIMD vector with 16 32-bit integer lanes
3. Pairwise Correlation Coefficient Kernels
For the τ coefficient, we have investigated three pairwise τ coefficient kernels:
the naı̈ve kernel, the GSE kernel and the VSE kernel. From its definition,
it can be seen that the Kendall’s τ coefficient only depends on the order of
variable pairs. Hence, given two ordinal vectors, we can first order all elements
in each vector, then replace the original value of every element with its rank
in each vector, and finally conduct the τ coefficient computation on the rank
transformed new vectors. This rank transformation does not affect the resulting
coefficient value, but could streamline the computation, especially for ordinal
variables in complex forms of representation. Moreover, this transformation
needs to be done only once beforehand for each vector. Hence, we will assume
that all ordinal vectors have already been rank transformed in the following
discussions. For the convenience of discussion, Table 1 shows a list of notions
used across our study.
3.1. Naı̈ve Kernel
The naı̈ve kernel enumerates all possible combinations of joint variables
(ui , vi ) (0 ≤ i < n) and counts the number of concordant pairs nc as well
as the number of discordant pairs nd . As mentioned above, given two joint
variable pairs (ui , vi ) and (uj , vj ) (i 6= j), they are considered concordant if
ui > uj and vi > vj or ui < uj and vi < vj , discordant if ui > uj and vi < vj
or ui < uj and vi > vj , and neither concordant nor discordant if ui = uj or
vi = vj . Herein, we can observe that the two joint variables are concordant
if and only if the value of (ui − uj ) × (vi − vj ) is positive; discordant if and
only if the value of (ui − uj ) × (vi − vj ) is negative; and neither concordant nor
5
discordant if and only if the value of (ui − uj ) × (vi − vj ) is equal to zero. In this
case, in order to avoid branching in execution paths (particularly important for
processors without hardware branch prediction units), we compute the value of
nc − nd by examining the sign bit of the product of ui − uj and vi − vj (refer to
lines 2 and 11 in Algorithm 1).
Algorithm 1 Pseudocode of our naı̈ve kernel
1: function calc sign(v)
2:
return (v > 0) − (v < 0);
3: end function
⊲ return 1 if v > 0, -1 if v < 0 and 0, otherwise
4: function kendall tau a naı̈ve(u, v, n)
5:
norminator = 0;
6:
for i = 1; i < n; ++i do
7:
a = ui ; b = vi ;
8:
#pragma vector aligned
9:
#pragma simd reduction(+:nominator)
10:
for j = 0; j < i; ++j do
11:
nominator += calc sign((a − uj )) × (b − vj ));
12:
end for
13:
end for
;
14:
return nominator
n(n−1)/2
15: end function
⊲ norminator represents nc − nd
⊲ compute nc − nd
Algorithm 1 shows the pseudocode of the naı̈ve kernel. From the code, the
naı̈ve kernel has a quadratic time complexity in a function of n, but its runtime
is independent of the actual content of u and v, due to the use of function
calc sign. Meanwhile, the space complexity is O(1). Note that this naı̈ve
kernel is only used to compute τA .
3.2. Generic Sorting-enabled Kernel
Considering the close relationship between calculating τ and ordering a list of
variables, Knight [32] proposed a merge-sort-like divide-and-conquer approach
with O(n log n) time complexity, based on the assumption that no element tie
exists within any vector. As this assumption is not always the case, Knight did
mention this drawback and suggested an approximation method by averaging
counts, rather than propose an exact solution. In this subsection, we investigate
an exact sorting-enabled solution to address both cases: with or without element
ties within any vector, together in a unified manner.
Given two ordinal vectors u and v, this GSE kernel generally works in the
following five steps.
• Step1 sorts the list of joint variables (ui , vi ) (0 ≤ i < n) in ascending order,
where the joint variables are sorted first by the first element ui and secondarily by the second element vi . In this step, we used quicksort via the
standard qsort library routine, resulting in O(n log n) time complexity.
• Step2 performs a linear-time scan over the sorted list to compute n1 (refer
to Equation (3)) by counting the number of groups consisting of tied
values as well as the number of tied values in each group. Meanwhile, we
compute a new value n3 for joint ties, with respect to the pair (ui , vi ), as
6
P
i wi (wi − 1)/2, where wi represents the number of jointly tied values in
the i-th group of joint ties for u and v.
• Step3 counts the number of discordant pairs nd by re-sorting the sorted
list obtained in Step1 in ascending order of all elements in v via a merge
sort procedure that can additionally accumulate the number of discordant
joint variable pairs each time two adjacent sorted subarrays are merged.
The rationale is as follows. Firstly, when merging two adjacent sorted
subarrays, we count the number of discordant pairs by only performing
pairwise comparison between joint variables from distinct subarrays. In
this way, we can ensure that every pair of joint variables will be enumerated once and only once during the whole Step3 execution. Secondly,
given two adjacent sorted subarrays to merge, it is guaranteed that the
first value (corresponding to u) of every joint variable in the left subarray
(with the smaller indices) is absolutely less than or equal to the first value
of every joint variable in the right subarray (with the larger indices), due
to the sort conducted in Step1. In particular, when the first value is identical for the two joint variables from distinct subarrays, the second value
(corresponding to v) of the joint variable from the left subarray is also absolutely less than or equal to the second value of the joint variable from the
left subarray. This means that discordance occurs only if the second value
of a joint variable from the right subarray is less than the second value of
a joint variable from the left subarray. Therefore, the value of nd can be
gained by accumulating the number of occurrences of the aforementioned
discordance in every pairwise merge of subarrays. Algorithm 2 shows the
pseudocode of counting discordant pairs with out-of-place pairwise merge
of adjacent sorted subarrays, where the time complexity is O(n log n) and
the space complexity is O(n).
• Step4 performs a linear-time scan over the sorted list obtained in Step3
to compute n2 (see Equation (3)) in a similar way to Step2. This works
because the sorted list actually corresponds to a sorted list of all elements
in v.
• Step 5 computes the numerator nc − nd in Equations (1) and (3) as n0 −
n1 − n2 + n3 − 2nd . Note that if there is no tie in each vector, n1 , n2 and
n3 will all be zero. In this case, nc − nd will be equal to n0 − 2nd as shown
in Equation (2).
From the above workflow, we can see that the GSE kernel takes into account
tied elements within each vector. Unlike the naı̈ve kernel that computes τA ,
the GSE kernel targets the computation of τB . As mentioned above, τA is
actually a special case of τB and an algorithm for τB will inherently cover τA .
Therefore, our GSE kernel is able to calculate both τA and τB in a unified
manner. In addition, the GSE kernel has O(n log n) time complexity, since the
time complexity is O(n log n) for both Step1 and Step3, O(n) for both Step2
and Step4, and O(1) for Step 5.
7
Algorithm 2 Pseudocode of Step3 of our GSE kernel
1: function gse merge(in, out, lef t, mid, right)
2:
l = p = lef t; r = mid; nd = 0;
3:
while l < mid && r < right do
4:
if in[r].v < in[l].v then
5:
nd + = mid − l;
6:
out[p++] = in[r++];
7:
else
8:
out[p++] = in[l++];
9:
end if
10:
end while
11:
return nd ;
12: end function
13: function kendall tau b step3 gse(pairs, buf f er, n)
14:
nd = 0; in = pairs; out = buf f er;
15:
for s = 1; s < n; s *= 2 do
16:
for l = 0; l < n; l += 2 * s do
17:
m = min(l + s, n);
18:
r = min(l + 2 ∗ s, n);
19:
nd += gse merge(in, out, l, m, r);
20:
end for
21:
swap(in, out);
22:
end for
23:
if pairs != in then
24:
memcpy(pairs, in, n * sizeof(*pairs));
25:
end if
26:
return nd ;
27: end function
⊲ merge two sorted subarrays
⊲ count discordant pairs
⊲ Perform out-of-place merge
⊲ swap in and out
3.3. Vectorized Sorting-enabled Kernel
The VSE kernel enhances the GSE kernel by employing 512-bit SIMD vector
instructions on MIC processors to implement vectorized pairwise merge of sorted
subarrays. In contrast with the GSE kernel, the VSE kernel has made the
following algorithmic changes. The first is packing a rank variable pair (ui , vi )
into a signed 32-bit integer. In this packed format, each variable is represented
by 15 bits in the 32-bit integer, with ui taking the most significant 16 bits and
vi the least significant 16 bits. In this case, the VSE kernel limits the maximum
allowable vector size n to 215 − 1 = 32, 767. The second is that due to packing,
we replace the generic variable pair quicksort in Step1 with an integer sorting
method, and re-implement the discordant pair counting algorithm in Step3 (see
Algorithm 2) based on integer representation. In Step1, sorting packed integers
is equivalent to sorting generic variable pairs (ui , vi ), since within any packed
integer ui sits in higher bits and vi in lower bits. In Step3, an additional
preprocessing procedure is needed to reset to zero the most significant 16 bits
of each packed integer (corresponding to u). In our implementation, we split a
512-bit SIMD vector into a 16-lane 32-bit-integer vector and then investigated
a vectorized merge sort for Step1 and a vectorized discordant pair counting
algorithm for Step3, both of which use the same pairwise merge method and
also follow a very similar procedure. Algorithm 3 shows the pseudocode of Step3
of our VSE kernel.
In the literature, some research has been done to accelerate sorting algorithms by means of SIMD vectorized pairwise merge of sorted subarrays. Hiroshi et al. [33] employed a SSE vectorized odd-even merge network [34] to
merge sorted subarrays in an out-of-core way. Chhugan et al. [35] adopted
8
Algorithm 3 Pseudocode of Step3 our VSE kernel
1: function vse merge(in, out, lef t, mid, right)
2:
l = lef t; r = mid; p = lef t; nd = 0;
3:
if mid − lef t < 16||right − mid < 16 then
4:
while l < mid && r < right do
5:
if input[r] < input[l] then
6:
nd + = mid − l; out[p++] = in[r++];
7:
else
8:
out[p + +] = in[l + +];
9:
end if
10:
end while
11:
else
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
39:
40:
41:
42:
43:
44:
45:
46:
47:
48:
⊲ correspond to variable v
⊲ count discordant pairs
while l < mid && r < right do
if in[r] < in[l] then
nd + = mid − 1; r + +;
else
l + +;
end if
end while
⊲ correspond to variable v
⊲ merge two sorted subarrays
l = lef t; r = mid;
vMin = mm512 load epi32(in + l); vMax = mm512 load epi32(in + r);
l+ = 16; r+ = 16;
while true do
if mm512 reduce min epi32(vMin) ≥ mm512 reduce max epi32(vMax) then
mm512 store epi32(out + p, vMax); p+ = 16; vMax = vMin;
else if mm512 reduce min epi32(vMax) ≥ mm512 reduce max epi32(vMin) then
mm512 store epi32(out + p, vMin); p+ = 16;
else
⊲ invoke Algorithm 4
bitonic merge 16way(vMin, vMax);
mm512 store epi32(out + p, vMin); p+ = 16;
end if
if l + 16 ≥ mid||r + 16 ≥ right then
break;
end if
A = in[l]; B = in[r]; C = mm512 reduce max epi32(vMax);
if C ≤ A && C ≤ B then
mm512 store epi32(out + p, vMax); p+ = 16;
vMin = mm512 load epi32(in + l); l+ = 16;
vMax = mm512 load epi32(in + r); r+ = 16;
else if B < A then
vMin = mm512 load epi32(in + r); r+ = 16;
else
vMin = mm512 load epi32(in + l); l+ = 16;
end if
end while
end if
⊲ invoke Algorithm 5
bitonic merge 16way leftover(in, out, l, r, p, mid, right, vMax);
return nd ;
end function
49: function kendall tau b step3 vse(pairs, buf f er, n)
50:
nd = 0; in = pairs; out = buf f er;
51:
for s = 1; s < n; s *= 2 do
52:
for l = 0; l < n; l += 2 * s do
53:
m = min(l + s, n);
54:
r = min(l + 2 ∗ s, n);
55:
nd += vse merge(in, out, l, m, r);
56:
end for
57:
swap(in, out);
58:
end for
59:
if pairs != in then
60:
memcpy(pairs, in, n * sizeof(*pairs));
61:
end if
62:
return nd ;
63: end function
9
⊲ Perform out-of-place merge
⊲ swap in and out
u15
u14
u13
u12
u11
u10
u9
u8
u7
u6
u5
u4
u3
u2
u1
u0
v0
v1
v2
v3
v4
v5
v6
v7
v8
v9
v10
v11
v12
v13
v14
v15
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
a0
a1
a2
a3
a4
a5
a6
a7
a8
a9
a10
a11
a12
a13
a14
a15
b0
b1
b2
b3
b4
b5
b6
b7
b8
b9
b10
b11
b12
b13
b14
b15
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
a0
a1
a2
a3
a4
a5
a6
a7
a8
a9
a10
a11
a12
a13
a14
a15
b0
b1
b2
b3
b4
b5
b6
b7
b8
b9
b10
b11
b12
b13
b14
b15
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
a0
a1
a2
a3
a4
a5
a6
a7
a8
a9
a10
a11
a12
a13
a14
a15
b0
b1
b2
b3
b4
b5
b6
b7
b8
b9
b10
b11
b12
b13
b14
b15
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
a0
a1
a2
a3
a4
a5
a6
a7
a8
a9
a10
a11
a12
a13
a14
a15
b0
b1
b2
b3
b4
b5
b6
b7
b8
b9
b10
b11
b12
b13
b14
b15
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
L
H
a0
a1
a2
a3
a4
a5
a6
a7
a8
a9
a10
a11
a12
a13
a14
a15
b0
b1
b2
b3
b4
b5
b6
b7
b8
b9
b10
b11
b12
b13
b14
b15
Figure 1: 16-way bitonic merge network: (1) input data are stored in two 16-lane vectors and
(2) pipelining from left-to-right
similar ideas to [33], but combined a SSE vectorized in-register odd-even merge
sort [34] with an in-memory bitonic merge network. Chen et al. [36] absorbed
the merge-path idea of [37, 38] and extended the SSE vectorized work of [35]
to take advantage of 512-bit SIMD VPUs on KNC Phis and used a vectorized
in-register bitonic merge sort, instead of the in-register odd-even merge sort.
In our VSE kernel, we engineered a 512-bit SIMD vectorized in-register bitonic
merge network as the core of our pairwise merge procedure for sorted subarrays,
which is similar to [36], and further proposed a predict-and-skip mechanism to
reduce the number of comparisons during pairwise merge.
3.3.1. In-register bitonic merge network
The in-register bitonic merge network is the core of our vectorized pairwise
merge of sorted subarrays. In our algorithm, this network has 16 ways and
merges two sorted vectors vM in and vM ax (Figure 1 shows the computation
layout of the network), where all elements in each vector are placed in ascending
order from lane 0 to lane 15. In this case, to generate an input bitonic sequence
from the two vectors, we need to reverse the order of all elements in one and only
one vector (reverse vM ax in our case) and this order reversal is realized by one
permutation instruction, i.e. mm512 permutevar epi32(·). Having completed
the order reversal, we can complete the sorting of vectors vM in and vM ax in
log2 (32) = 5 steps. Algorithm 4 shows the pseudocode for our 16-way in-register
bitonic merge network. From the code, the function bitonic merge 16way is
10
composed of a fixed number of vector instructions and thus has a constant time
complexity.
Algorithm 4 Pseudocode of our 16-way in-register bitonic merge network
1: procedure bitonic merge 16way(vMin, vMax)
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
⊲ constant vector variables and reused
//vReverse = mm512 set epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
//vP ermIndex16 = mm512 set epi32(7, 6, 5, 4, 3, 2, 1, 0, 15, 14, 13, 12, 11, 10, 9, 8)
//vP ermIndex8 = mm512 set epi32(11, 10, 9, 8, 15, 14, 13, 12, 3, 2, 1, 0, 7, 6, 5, 4);
//vP ermIndex4 = mm512 set epi32(13, 12, 15, 14, 9, 8, 11, 10, 5, 4, 7, 6, 1, 0, 3, 2);
//vP ermIndex2 = mm512 set epi32(14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1);
⊲ Reserve vector vMax
vMax = mm512 permutevar epi32(vReverse, vMax);
⊲ Level 1
vL1 = mm512 min epi32(vMin, vMax);
vH1 = mm512 max epi32(vMin, vMax);
⊲ Level 2
vT mp = mm512 permutevar epi32(vP ermIndex16, vL1);
vT mp2 = mm512 permutevar epi32(vP ermIndex16, vH1);
vL2 = mm512 mask min epi32(vL2, 0x00ff, vT mp, vL1);
vH2 = mm512 mask min epi32(vH2, 0x00ff, vT mp2, vH1);
vL2 = mm512 mask max epi32(vL2, 0xff00, vT mp, vL1);
vH2 = mm512 mask max epi32(vH2, 0xff00, vT mp2, vH1);
⊲ Level 3
vT mp = mm512 permutevar epi32(vP ermIndex8, vL2);
vT mp2 = mm512 permutevar epi32(vP ermIndex8, vH2);
vL3 = mm512 mask min epi32(vL3, 0x0f0f, vT mp, vL2);
vH3 = mm512 mask min epi32(vH3, 0x0f0f, vT mp2, vH2);
vL3 = mm512 mask max epi32(vL3, 0xf0f0, vT mp, vL2);
vH3 = mm512 mask max epi32(vH3, 0xf0f0, vT mp2, vH2);
⊲ Level 4
vT mp = mm512 permutevar epi32(vP ermIndex4, vL3);
vT mp2 = mm512 permutevar epi32(vP ermIndex4, vH3);
vL4 = mm512 mask min epi32(vL4, 0x3333, vT mp, vL3);
vH4 = mm512 mask min epi32(vH4, 0x3333, vT mp2, vH3);
vL4 = mm512 mask max epi32(vL4, 0xcccc, vT mp, vL3);
vH4 = mm512 mask max epi32(vH4, 0xcccc, vT mp2, vH3);
⊲ Level 5: vMin and vMax store the sorted sequence
vT mp = mm512 permutevar epi32(vP ermIndex2, vL4);
vT mp2 = mm512 permutevar epi32(vP ermIndex2, vH4);
vMin = mm512 mask min epi32(vMin, 0x5555, vT mp, vL4);
vMax = mm512 mask min epi32(vMax, 0x5555, vT mp2, vH4);
vMin = mm512 mask max epi32(vMin, 0xaaaa, vT mp, vL4);
vMax = mm512 mask max epi32(vMax, 0xaaaa, vT mp2, vH4);
end procedure
3.3.2. Vectorized pairwise merge of sorted subarrays
Our vectorized pairwise merge of sorted subarrays relies on the aforementioned 16-way in-register bitonic merge network and adopted a very similar
procedure to [33]. Given two sorted subarrays SA and SB , we assume that they
are aligned to 64 bytes and their lengths |SA | and |SB | are multiples of 16, for
the convenience of discussion. In this way, the vectorized pairwise merge works
as follows.
1. loads the smallest 16 elements of SA and SB to vM in and vM ax, respectively, and advances the pointer of each subarray.
2. invokes bitonic merge 16way(·) to sort vectors vM in and vM ax, and
then stores the content of vM in, the smallest 16 elements, to the resulting
output array.
11
Algorithm 5 Pseudocode of processing the leftovers in both subarrays
1: procedure bitonic merge 16way leftover(in, out, l, r, p, mid, right, vMax)
2:
vIndexInc = mm512 set epi32(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
⊲ maximum signed integer value
3:
vMaxInt = mm512 set1 epi32(0x7fffffff)
4:
vMid = mm512 set1 epi32(mid);
5:
vRight = mm512 set1 epi32(right);
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
32:
33:
34:
35:
36:
37:
38:
39:
40:
41:
42:
43:
44:
45:
46:
47:
48:
49:
50:
51:
52:
⊲ initialize the vector mask for vMax
maskMax = 0xffff;
while l < mid||r < right do
if l < mid && r < right then
if in[r] < in[l] then
vT mp = mm512 add epi32( mm512 set1 epi32(r), vIndexInc)
maskMin = mm512 cmplt epi32 mask(vT mp, vRight);
vMin = mm512 mask load epi32(vMaxInt, maskMin, in + r); r+ = 16;
else
vT mp = mm512 add epi32( mm512 set1 epi32(l), vIndexInc)
maskMin = mm512 cmplt epi32 mask(vT mp, vMid);
vMin = mm512 mask load epi32(vMaxInt, maskMin, in + l); l+ = 16;
end if
else if l < mid then
vT mp = mm512 add epi32( mm512 set1 epi32(l), vIndexInc);
maskMin = mm512 cmplt epi32 mask(vT mp, vMid);
vMin = mm512 mask load epi32(vMaxInt, maskMin, in + l);
tmp = mm512 mask reduce max epi32(maskMax, vMax);
if tmp ≤ mm512 reduce min epi32(vMin) then
break;
end if
l+ = 16;
else if r < right then
vT mp = mm512 add epi32( mm512 set1 epi32(r), vIndexInc);
maskMin = mm512 cmplt epi32 mask(vT mp, vRight);
vMin = mm512 mask load epi32(vMaxInt, maskMin, in + r);
tmp = mm512 mask reduce max epi32(maskMax, vMax);
if tmp ≤ mm512 reduce min epi32(vMin) then
break;
end if
r+ = 16;
end if
⊲ invoke Algorithm 4
bitonic merge 16way(vMin, vMax);
⊲ calculate new vector masks
nb = mm countbits 32(maskMin) + mm countbits 32(maskMax);
maskMin = nb > 15? 0x0ffff : (1 << nb) − 1;
maskMax = nb < 17? 0 : (1 << (nb − 16)) − 1;
mm512 mask store epi32(out + p, maskMin, vMin);
p += mm countbits 32(maskMin);
end while
⊲ write out vMax
mm512 mask packstorelo epi32(out + p, maskMax, vMax);
mm512 mask packstorehi epi32(out + p + 16, maskMax, vMax);
p += mm countbits 32(maskMax);
⊲ copy out the rest
if l < mid then
memcpy(out + p, in + l, (mid − l)∗ sizeof(int));
else if r < right then
memcpy(out + p, in + r, (right − r)∗ sizeof(int));
end if
end procedure
12
3. compares the next element of SA and SB and loads 16 elements into vM in
from the subarray whose next element is the smaller, followed by pointer
advancing for the subarray.
4. jumps back to the second step and repeats the procedure until all elements
in both subarrays have been processed.
5. completes the merge of SA and SB by writing the content of vM ax to the
output array.
Note that if |SA | or |SB | is not a multiple of 16, we will have to implement
some special treatment to deal with the leftovers in SA and SB . This special
treatment is implemented as Algorithm 5 and invoked by Algorithm 3 (see line
46). From the above procedure, we can see that the time complexity of merging
SA and SB is linear and thus the VSE kernel has a time complexity of O(n log n).
As assumed that both SA and SB are aligned to 64 bytes, we can therefore
use only one mm512 load epi32(·) instruction to load 16 integer elements from
memory to a vector and only one mm512 store epi32(·) instruction to store 16
elements in a vector to memory. This is because memory address is guaranteed
to keep aligned to 64 bytes during pairwise merge. Defining Trd to denote
the latency of memory load (in clock cycles) and Twt to denote the latency of
memory store, the number of clock cycles per element load from memory can be
estimated as (Trd + 16)/16 = Trd/16 + 1 clock cycles and the number of clocks
cycles per element store to memory as (Twt + 16)/16 = Twt /16 + 1 clock cycles.
Moreover, bitonic merge 16way(·) is composed of 27 vector instructions (see
lines 7∼33 in Algorithm 4) and each of its invocations leads to the output of 16
elements. In these regards, the average number of instructions per element can
be estimated as 27/16 ≈ 1.69. Considering that we only used simple integer
and single-source permutation instructions, each of which has a two clock cycle
latency according to the Intel Xeon Phi processor software optimization manual
[39], the average computational cost per element can be estimated as ⌈2 ×
27/16⌉ = 4 clock cycles in the worst cases where consecutive instructions always
have data dependency. In order to reduce inter-instruction latency caused by
data dependency, we have manually tuned the order of instructions by means
of interleaving. Since each instruction has a latency of two clock cycles, this
interleaving could enable to execute one instruction in only one clock cycle,
thus further reducing the cost per element to ⌈1 × 27/16⌉ = 2 clock cycles
optimistically. In sum, the overall cost per element can be favorably estimated
to be (Trd + Twr )/16 + 4 clock cycles.
We have proposed a predict-and-skip scheme to determine (i) whether we
need to invoke the 16-way bitonic merge network and (ii) whether we should
load two vectors of new elements from SA and SB , respectively. For cases (i),
the rationale is if the minimum value in vM in (vM ax) is ≥ the maximum value
in vM ax (vM in), every value in vM in (vM ax) will be ≥ to all values in vM in
(vM ax). In this case, there is no need of invoking the bitonic merge network
(lines 23∼26 in Algorithm 3), since the order of all elements in both vectors
is already deterministic. For cases (ii), we compare the maximum value C in
vM ax with both the smallest element A in the rest of SA and the one B in
13
the rest of SB (lines 34∼38 in Algorithm 3). If C ≤ A and C ≤ B, none
of the elements in vM ax will be greater than any element in the rest of SA
and SB , indicating that the absolute order of all vM ax elements has already
been determined. In this case, we write vM ax to the resulting output array
and load two vectors of new elements from SA and SB to vM in and vM ax,
respectively. The updates in vM in and vM ax will be used for the next iteration
of comparison.
It is worth mentioning that we also developed a 32-way bitonic merge network to implement the VSE kernel. Unfortunately, this new kernel based on
32-way merge network yielded slightly inferior performance to the kernel with
the 16-way network through our tests. In this regard, we adopted the 16-way
merge network in our VSE kernel.
4. Parallel All-Pairs Computation
4.1. All-Pairs Computation Framework
We consider the m × m job matrix to be a 2-dimensional coordinate space
on the Cartesian plane, and define the left-top corner to be the origin, the
horizontal x-axis (corresponding to columns) in left-to-right direction and the
vertical y-axis (corresponding to rows) in top-to-bottom direction. For nonsymmetric all-pairs computation (non-commutative pairwise computation), the
workload distribution over processing elements (PEs), e.g. threads, processes,
cores and etc., would be relatively straightforward. This is because coordinates
in the 2-dimensional matrix corresponds to distinct jobs. Unlike non-symmetric
all-pairs computation, it suffices by only computing the upper-triangle (or lowertriangle) of the job matrix for symmetric all-pairs computation (commutative
pairwise computation). In this case, balanced workload distribution could be
more complex than non-symmetric all-pairs computation.
In this paper, we propose a generic framework for workload balancing in
symmetric all-pairs computation, since the computation of the τ coefficient between any u and v is commutable as mentioned above. This framework works
by assigning each job a unique identifier and then building a bijective relationship between a job identifier Jm (y, x) and its corresponding coordinate (y, x).
We refer to this as direct bijective mapping. While facilitating balanced
workload distribution, this mapping merely relies on bijective functions, which
is a prominent feature distinguished from existing methods. To the best of our
knowledge, in the literature bijective functions have not ever been proposed for
workload balancing in symmetric all-pairs computation. In [40], the authors
used a very similar job numbering approach to ours (shown in this study), but
did not derive a bijective function for symmetric all-pairs computation. Our
framework can be applied to cases with identical (e.g. using static workload
distribution) or varied workload per job (e.g. using a shared integer counter
to realize dynamic workload distribution via remote memory access operations
in MPI [41] and Unified Parallel C (UPC) programming models [42] [43]) and
is also particularly useful for parallel computing architectures with hardware
14
*"& (
"
0
1
2
3
8
9
10 11 12 13 14
4
5
6
7
! " # 0.5
15 16 17 18 19 20
21 22 23 24 25
26 27 28 29
30 31 32
!
)
!
! " # 0.5 #
"$ % " % 0.25 # 2&' % 1(
)
! " # 0.5 %
"$ % " % 0.25 # 2&' % 1(
33 34
35
(a)
(b)
(c)
Figure 2: (a) an example direct bijective mapping between job identifier and coordinate space,
(b) solution region for Equation (7), and (c) solution region for Equation (8)
schedulers such as GPUs and FPGAs. In the following, without loss of generality, we will interpret our framework relative to the upper triangle of the job
matrix by counting in the major diagonal. Nonetheless, this framework can be
easily adapted to the cases excluding the major diagonal.
Direct bijective mapping. Given a job (y, x) in the upper triangle, we compute
its integer job identifier Jm (y, x) as
Jm (y, x) = Fm (y) + x − y, 0 ≤ y ≤ x < m
(4)
for dimension size m. In this equation, Fm (y) is the total number of cells
preceding row y in the upper triangle and is computed as
Fm (y) =
y(2m − y + 1)
2
(5)
where y varies in [0, m]. In this way, we have defined Equation (4) based on our
job numbering policy, i.e. all job identifiers vary in [0, m(m + 1)/2) and jobs
are sequentially numbered left-to-right and top-to-bottom in the upper triangle
(see Fig. 2a for an example).
Reversely, given a job identifier J = Jm (y, x) (0 ≤ J < m(m+1)/2), we need
to compute the coordinate (y, x) in order to locate the corresponding variable
pair. As per our definition, we have
J ≥ Fm (y) ⇔ y 2 − (2m + 1)y + 2J ≥ 0
(6)
J ≤ Fm (y + 1) − 1 ⇔ y 2 − (2m − 1)y + 2(J + 1) − 2m ≤ 0
By solving J ≥ Fm (y) (see Figure 2b), we get
p
y ≤ m + 0.5 − m2 + m + 0.25 − 2J
(7)
while getting
y ≥ m − 0.5 −
p
m2 + m + 0.25 − 2(J + 1)
15
(8)
by
Figure 2c). In this case, by defining ∆ =
p solving J ≤ Fm (y + 1) − 1 (see
√
2 + m + 0.25 − 2(J + 1), ∆′ =
m
m2 + m + 0.25 − 2J and z = m − 0.5 −
p
m2 + m + 0.25 − 2(J + 1), we can reformulate Equations (7) and (8) to be
z ≤ y ≤ z + 1 + ∆ − ∆′ . Because ∆ < ∆′ , we know that [z, z + 1 + ∆ − ∆′ ]
is a sub-range of [z, z + 1) and thereby have z ≤ y < z + 1. Based on our job
numbering policy stated before, as a function of integer y, Equation (6) definitely
has y solutions, meaning that at least one integer exists in [z, z + 1 + ∆ − ∆′ ],
which satisfies Equation (6). Meanwhile, it is known that there always exists
one and only one integer in [z, z + 1) (can be easily proved) and this integer
equals ⌈z⌉, regardless of the value of z. Since [z, z + 1 + ∆ − ∆′ ] is a sub-range
of [z, z + 1), we can conclude that Equation (6) has a unique solution y that is
computed as
l
m
p
y = ⌈z⌉ = m − 0.5 − m2 + m + 0.25 − 2(J + 1)
(9)
Having got y, we can compute the coordinate x as
x = J + y − Fm (y)
(10)
based on Equation (4).
4.2. Tiled Computing
Based on the direct bijective mapping, we adopted tiled computing with
the intention to benefit from L1/L2 caches. The rationale behind the tiled
computing is loading into cache a small subset of the bigger dataset and reusing
this block of data in cache for multiple passes. This technique partitions a
matrix into a non-overlapping set of equal-sized q × q tiles. In our case, we
partition the job matrix and produce a tile matrix of size w × w tiles, where
w = ⌈m/q⌉. In this way, all jobs in the upper triangle of the job matrix are still
fully covered by the upper triangle of the tile matrix. By treating a tile as a
unit, we can assign a unique identifier to each tile in the upper triangle of the
tile matrix and then build bijective functions between tile identifiers and tile
coordinates in the tile matrix, similarly as we do for the job matrix.
Because the tile matrix has an identical structure to the original job matrix,
we can directly apply our bijective mapping to the tile matrix. In this case,
given a coordinate (yq , xq ) (0 ≤ yq ≤ xq < w) in the upper triangle of the tile
matrix, we can compute a unique tile identifier Jw (yq , xq ) as
Jw (yq , xq ) = Fw (yq ) + xq − yq , 0 ≤ yq ≤ xq < w
(11)
where Fw (yq ) is defined similar to Equation (5) as
Fw (yq ) =
yq (2w − yq + 1)
2
(12)
Likewise, given a tile identifier Jw , such that 0 ≤ Jw < w(w + 1)/2, we can
16
reversely compute its unique vertical coordinate yq as
l
m
p
yq = w − 0.5 − w2 + w + 0.25 − 2(Jw + 1)
(13)
and subsequently its unique horizontal coordinate xq as
xq = Jw + yq − Fw (yq )
(14)
As title size is subject to cache sizes and input data, tuning tile size is
believed to be important for gaining high performance. This tuning process,
however, is tedious and has to be conducted case-by-case. For convenience, we
empirically set q to 8 for CPUs and to 4 for Phis. Nevertheless, users can feel
free to tune tile sizes to meet their needs.
4.3. Multithreading
4.3.1. Asynchronous kernel execution
When m is large, we may not have sufficient memory to reside the resulting
m × m correlation matrix Mτ entirely in memory. To overcome memory limitation, we adopted a multi-pass kernel execution model which partitions the tile
identifier range [0, w(w + 1)/2) into a set of non-overlapping sub-ranges (equalsized in our case) and finishes the computation one sub-range after another. In
this way, we do not need to allocate the whole memory for matrix Mτ . Instead,
we only need to allocate a small amount of memory to store the computing
results of one sub-range, thus considerably reducing memory footprint.
When using the KNC Phi, we need to transfer the newly computed results
from the Phi to the host after having completed each pass of kernel execution.
If kernel execution and data transfer is conducted in sequential, the Phi will
be kept idle during the interim of device-to-host data transfer. In this regard,
a more preferable solution would be to employ asynchronous kernel execution
by enabling concurrent execution of host-side tasks and accelerator-side kernel
execution. Fortunately, KNC Phis enable such a kind of asynchronous data
transfer and kernel execution associated with the offload model. This asynchronous execution can be realized by coupling the signal and wait clauses,
both of which are associated with each other via a unique identifier. That is,
a signal clause initiates an asynchronous operation such as data transfer and
kernel execution, while a wait clause blocks the current execution until the associated asynchronous operation is completed. Refer to our previous work [28]
for more details about the host-side asynchronous execution workflow proposed.
4.3.2. Workload balancing
For the variant using the naı̈ve kernel, all jobs have the same amount of
computation (refer to Algorithm 1). Thus, given a fixed number of threads, we
evenly distribute jobs over the set of threads by assigning a thread to process
one tile at a time. In contrast, for the two variants using the sorting-enabled
kernels, different jobs may have different amount of computation because of
17
the two rounds of sort used in Step1 and Step3. Typically, an OpenMP dynamic scheduling policy is supposed to be in favor to address workload irregularity, albeit having relatively heavier workload distribution overhead than static
scheduling. However, it is observed that dynamic scheduling produces slightly
inferior performance to static scheduling through our tests. In this regard, we
adopted static scheduling in our two sorting-enabled variants and assigned one
thread to process one tile at a time, same as the naı̈ve variant.
4.4. Distributed Computing
On KNC Phi clusters, we used MPI offload model which launches MPI
processes just as an ordinary CPU cluster does. In this model, one or more
Phis will be associated to a parental MPI process, which utilizes offload pragmas/directives to interact with the affiliated Phis, and inter-Phi communications
need to be explicitly managed by parental processes. In this sense, Phis may not
perceive the existence of remote inter-process communications. Our distributed
implementations require one-to-one correspondence between MPI processes and
Phis, and adopted a static workload distribution scheme based on tiled computation. This static distribution is inspired by our practice in multithreading
on single Phis. In our implementation, given p processes, we evenly distribute
tiles onto the p processes with the i-th (0 ≤ i < p) process assigned to compute
the tiles whose identifiers are in [i × ⌈ w(w+1)
⌉, (i + 1) × ⌈ w(w+1)
⌉). Within each
2p
2p
process, we adopted asynchronous execution workflow as well.
5. Performance Evaluation
We assessed the performance of our algorithm from four aspects: (i) comparison between our three variants, (ii) comparison with widely used counterparts:
MATLAB (version R2015b) and R (version 3.2.0), (iii) multithreading scalability on a single Phi, and (iv) distributed computing scalability on Phi clusters.
In these tests, we used four real whole human genome gene expression datasets
(refer to Table 2) produced by Affymetrix Human Genome U133 Plus 2.0 Array.
These datasets are publicly available in the GPL570 data collection of SEEK
[1], a query-based computational gene co-expression search engine over large
transcriptomic databases. In this study, unless otherwise specified, we compute
τ coefficients between genes, meaning that m is equal to the number of genes
and n equal to the number of samples for each dataset.
All tests are conducted on 8 compute nodes in CyEnce HPC Cluster (Iowa
State University), where each node has two high-end Intel E5-2650 8-core 2.0
GHz CPUs, two 5110P Phis (each has 60 cores and 8 GB memory) and 128 GB
memory. Our algorithm is compiled by Intel C++ compiler v15.0.1 with option
-fast enabled. In addition, for distributed computing scalability assessment,
when two processes are launched into the same node, we used the environment
variable I MPI PIN PROCESSOR LIST to guide Intel MPI runtime system to pin
the two processes within the node to distinct CPUs (recall that one node has
two CPUs).
18
Table 2: Information of whole human genome gene expression datasets used
Name
DS21050
DS19784
DS13070
DS3526
No. of Genes
17,941
17,941
17,941
17,941
No. of Samples
310
320
324
353
Platform
Affymetrix
Affymetrix
Affymetrix
Affymetrix
5.1. Assessment of Our Three Variants
All of the three variants work on CPUs, Phis and their clusters. For the
naı̈ve and GSE variants, they both used the same C++ core code for CPUand MIC-oriented instances. For the VSE variant, its 512-bit SIMD vectorization is only applicable to Phis. In this regard, to support CPUs, we further
developed a non-vectorized merge sort in Step1 and a non-vectorized discordant pair counting algorithm in Step3 instead, based on the aforementioned
packed integer representation. Considering that this non-vectorized version also
works on Phis, we have examined how much our 512-bit SIMD vectorization
contributes to speed improvement by comparing the vectorized version with the
non-vectorized one on the same 5110P Phi. Interestingly, by measuring the
correlation between genes using the datasets in Table 2, we observed that the
non-vectorized version is on a par with the vectorized one. This may be due to
the relatively small value of n, which is only 353 at maximum. Based on this
consideration, we further evaluated both versions by measuring the correlation
between samples, where m will be equal to the number of samples and n equal to
the number of genes. In this case, n becomes 17,941 for each dataset, more than
50× larger than before. In this context, our performance assessment exposed
that the vectorized version is consistently superior to the non-vectorized one,
yielding very stable speedups averaged to be 1.87 for all datasets. This result
was exciting, proving that our 512-bit SIMD vectorization did boost speed even
for moderately large n values. Hence, we have used the vectorized version of
the VSE variant for performance measurement all throughout our study.
5.1.1. On CPU
We first compared the performance of our three variants on multiple CPU
cores. Table 3 shows the performance of each variant on 16 CPU cores and
Figure 3 shows the speedup of the 16-threaded instance of each variant over its
single-threaded one.
For the naı̈ve variant, its 16-threaded instance achieves a roughly constant
speedup of 13.70 over its single-threaded one. This observation is consistent
with our expectation, as the runtime of the naı̈ve kernel is subject to the vector
size n but independent of actual vector content (refer to the implementation
shown in Algorithm 1). In contrast, the speed of the two sorting-enabled kernels
are sensitive to vector content to some degree. This can be explained by the
following three factors: variable pair sorting in Step1, discordant pair counting
based on pairwise merge of sorted subarrays in Step3, and linear-time scans
19
14.0
Naïve
GSE
VSE
13.9
Speedup
13.8
13.7
13.6
13.5
13.4
DS21050
DS19784
DS13070
DS3526
Figure 3: Speedups of 16-threaded instances over single-thread ones on multiple CPU cores
for determining tie groups in Step2 and Step4. Nevertheless, for both sortingenabled variants, their 16-threaded instances yield relatively consistent speedups
over their single-threaded ones, respectively. For the GSE (VSE) variant, its 16threaded instance runs 13.74× (13.79×) faster than its single-threaded one on
average, with a minimum speedup 13.63 (13.69) and a maximum speedup 13.90
(13.97). For each case, the two sorting-enabled variants both yield superior
performance to the naı̈ve variant, where the average speedup is 1.60 for the
GSE kernel and 2.72 for the VSE kernel.
Table 3: Performance comparison of three variants on multiple CPU cores
Instance
Single-threaded
16-threaded
Dataset
DS21050
DS19784
DS13070
DS3526
DS21050
DS19784
DS13070
DS3526
Naı̈ve
10,052
10,690
10,975
12,993
734
781
801
948
20
Time (s)
GSE
6,609
6,776
6,920
7,619
475
495
504
559
VSE
3,758
3,938
4,041
4,742
275
286
294
340
Speedup
GSE VSE
1.52 2.67
1.58 2.71
1.59 2.72
1.71 2.74
1.54 2.67
1.58 2.73
1.59 2.72
1.70 2.79
Table 4: Performance comparison of the three variants on Phi
Dataset
DS21050
DS19784
DS13070
DS3526
Naı̈ve
400
424
433
506
Time (s)
GSE
419
434
459
488
VSE
261
266
276
296
Speedup
GSE VSE
0.95 1.53
0.98 1.59
0.94 1.57
1.04 1.71
5.1.2. On Xeon Phi
Subsequently, we evaluated the three variants on a single 5110P Phi (see
Table 4). From Table 4, the VSE variant outperforms the naı̈ve variant for each
dataset by yielding an average speedup of 1.60 with the minimum speedup 1.53
and the maximum speedup 1.71. The GSE variant is superior to the naı̈ve one
on the DS3526 dataset, but inferior to the latter on the remaining three datasets.
Interestingly, by revisiting the CPU results shown in Table 3, we recalled that
the former actually performs consistently better than the latter for each dataset
on CPU. Since both the naı̈ve and the GSE variants use the same pairwise τ
coefficient kernel code for their corresponding CPU- and MIC-oriented implementations, this discordant performance ranking between the two variants could
owe to the architectural differences of the two types of processing unit (PU).
For instance, by enabling auto-vectorization, the naı̈ve kernel (see Algorithm
1) can concurrently process 16 integer elements by one 512-bit SIMD vector
instruction, in contrast with 4 integer elements by one 128-bit SSE instruction
on CPU. This fourfold increase in parallelism would enable the naı̈ve kernel to
further boost performance.
Figure 4 shows the speedups of each variant on the Phi over on multiple
CPU cores for each dataset. From Figure 4, it is observed that the Phi instance
of each variant outperforms its corresponding single-threaded and 16-threaded
CPU instances for each dataset. Specifically, the Phi instance of the naı̈ve
variant runs 25.34× faster on average than its single-threaded CPU instance,
with the maximum speedup 25.67, and 1.85× faster on average than the 16threaded CPU instance, with the maximum speedup 1.87. Compared to their
single-threaded CPU instances, the Phi instances of the GSE and VSE variants
produce the average speedups of 15.52 and 14.96 and the maximum speedups
of 15.77 and 16.00, respectively. Meanwhile, in comparison to their 16-threaded
CPU instances, their Phi instances performs 1.13× and 1.08× better on average
and 1.15× and 1.15× better at maximum, respectively.
5.2. Comparison With MATLAB and R
Secondly, we compared our algorithm with all-pairs τ coefficient implementations in the widely used MATLAB (version R2015b) and R (version 3.2.0).
MATLAB and R are executed in one 16-core compute node mentioned above,
where MATLAB runs 16 threads as it supports multithreading and R runs in
sequential. For fair comparison, both MATLAB and R merely compute all-pairs
21
Naïve
30
GSE
VSE
Naïve
2.0
GSE
VSE
1.8
25
1.6
1.4
Speedup
Speedup
20
15
10
1.2
1.0
0.8
0.6
0.4
5
0.2
0
0.0
DS21050
DS19784
DS13070
DS21050
DS3526
(a)
DS19784
DS13070
DS3526
(b)
Figure 4: Speedups of each variant on Phi over on CPU: (a) speedups over one CPU core and
(b) speedups over 16 CPU cores.
Table 5: Performance comparison with MATLAB and R
Dataset
DS21050
DS19784
DS13070
DS3526
Time (s)
MLAB R
182,343 266,001
205,181 284,215
205,063 287,635
240,778 345,668
Speedup over MLAB
Naı̈ve GSE VSE
456
435
699
484
473
772
473
446
743
476
494
812
Speedup over R
Naı̈ve GSE VSE
665
635
1,019
671
655
1,069
664
626
1,042
683
709
1,166
τ coefficient values, without conducting statistical tests, same as our algorithm
does. Table 5 shows the runtimes of the 16-threaded MATLAB and sequential
R instances as well as their comparison with our three variants that execute on
the 5110P Phi. From the table, each variant demonstrates excellent speedups,
i.e. two orders-of-magnitude over 16-threaded MATLAB and three orders-ofmagnitude over sequential R. Compared to 16-threaded MATLAB, the average
and maximum speedups are 472 and 484 for the naı̈ve variant, 462 and 494 for
the GSE variant and 756 and 812 for the VSE variant, respectively. In contrast,
the average and maximum speedups over sequential R are 671 and 683 for the
naı̈ve variant, 656 and 709 for the GSE variant and 1,074 and 1,166 for the VSE
variant, respectively.
5.3. Parallel Scalability Assessment
5.3.1. Multithreading
Thirdly, we evaluated the multithreading scalability of our variants on the
Phi with respect to number of threads. Figure 5 shows the parallel scalability
of the three variants. In this test, we used balanced thread affinity to ensure
that thread allocation is balanced over the cores and the threads allocated to
the same core have consecutive identifiers (i.e. neighbors of each other). This
22
59T
118T
177T
236T
1400
59T
118T
177T
1400
236T
1200
1200
1000
1000
1000
800
600
Time (s)
1200
Time (s)
Time (s)
1400
800
600
177T
236T
600
400
400
200
200
200
0
0
(a)
118T
800
400
DS21050 DS19784 DS13070 DS3526
59T
0
DS21050 DS19784 DS13070 DS3526
(b)
DS21050 DS19784 DS13070 DS3526
(c)
Figure 5: Multithreading scalability of different variants: (a) the naı̈ve variant, (b) the GSE
variant and (c) the VSE variant.
thread affinity configuration is attained by setting the environment variable
KMP AFFINITY to balanced.
From the figures, it is observed that each variant gets performance improved
as the number of active threads per core grows from 1, via 2 and 3, to 4 (corresponding to 59, 118, 177 and 236 threads, respectively). As each core is dual
issue and employs in-order execution, at least two threads per core are required
in order to saturate the compute capacity of each core. Meanwhile, when moving
from one thread per core to two threads per core, we expect that the speedup
could be close to 2. In this regard, we investigated the speedup of the instance
with two threads per core over the one with one thread per core for each variant, and found that the average speedups are 1.58, 1.61 and 1.56 for the naı̈ve,
GSE and VSE variants, respectively. This finding is close to our expectation
largely. Furthermore, since all threads per core share the dual in-order execution pipeline, deploying three or four threads per core may further improve
performance but normally with decreased parallel efficiency (with respect to
threads rather than cores). Note that for the 5110P Phi, only 59 out of 60 cores
are used for computing as one core is reserved for the operating system running
inside. Since each variant reaches peak performance at 236 threads, we used
this number of threads for performance evaluation in all tests, unless otherwise
stated.
5.3.2. Distributed computing
Finally, we measured the parallel scalability of our algorithm by varying the
number of Phis used in a distributed environment (see Figure 6). This test is
conducted in a cluster that consists of 16 Phis and is constituted by the aforementioned 8 compute nodes. Each variant used 236 threads, since this setting
leads to the best performance as mentioned above. From Figure 6, the naı̈ve
variant demonstrates nearly constant speedups for all datasets on a specific
number of Phis, while the two sorting-enabled variants exposed slight fluctuations under the same hardware configuration. This can be explained by the fact
that the runtime of our naı̈ve variant is independent of vector content, whereas
that of each sorting-enabled variant is sensitive to actual vector content to some
degree. Moreover, for each dataset, it is observed that every variant has demon23
18
18
DS21050
DS19784
DS13070
DS3526
14
Speedup
12
DS21050
DS19784
DS13070
DS3526
16
14
12
Speedup
16
10
8
10
8
6
6
4
4
2
2
0
0
2
4
8
Number of Phis
16
2
4
8
Number of Phis
(a)
(b)
DS21050
DS19784
DS13070
DS3526
12
GSE
VSE
500
400
Time (s)
14
Speedup
Naïve
600
18
16
16
10
8
300
200
6
4
100
2
0
0
2
4
8
Number of Phis
1
16
(c)
2
4
8
Number of Phis
16
(d)
Figure 6: Distributed computing scalability: (a) the naı̈ve variant, (b) the GSE variant and
(c) the VSE variant and (d) runtimes of each variant on the DS3526 dataset.
strated rather good scalability with respect to number of Phis. Concretely, the
naı̈ve variant achieves an average speedup of 2.00, 3.99, 7.96 and 15.81, the GSE
one yields an average speedup of 1.93, 3.82, 7.63 and 14.97, and the VSE one
produces an average speedup of 1.99, 3.94, 7.86 and 15.23, on 2, 4, 8 and 16
Phis, respectively.
6. Conclusion
Pairwise association measure is an important operation in searching for
meaningful insights within a dataset by examining potentially interesting relationships between data variables of the dataset. However, all-pairs association
computation are computationally demanding for a large volume of data and may
take a prohibitively long sequential runtime. This computational challenge for
big data motivated us to use parallel/high-performance computing architectures
to accelerate its execution.
In this paper, we have investigated the parallelization of all-pairs Kendall’s τ
coefficient computation using MIC processors as the accelerator. To the best of
our knowledge, this is for the first time that all-pairs τ coefficient computation
24
has been parallelized on MIC processors. For the τ coefficient, we have developed
three variants, namely the naı̈ve variant, the GSE variant and the VSE variant,
based on three pairwise τ coefficient kernels, i.e. the naı̈ve kernel, the GSE
kernel and the VSE kernel. The three variants all can execute on CPUs, Phis and
their clusters. The naı̈ve kernel has a time complexity of O(n2 ), while the other
two sorting-enabled kernels have an improved time complexity of O(n log n).
Furthermore, we have proposed a generic framework for workload balancing
in symmetric all-pairs computation and overcoming memory limitation on the
host or accelerator side. This framework assigns unique identifiers to jobs in the
upper triangle of the job matrix and builds provable bijective functions between
job identifier and coordinate space.
The performance of the three variants was evaluated using a collection of
real gene expression datasets produced from whole human genomes. Our experimental results demonstrated that on one 5110P Phi, the naı̈ve variant runs up
to 25.67× faster than its execution on a single CPU core, and up to 1.87× faster
than on 16 CPU cores. On the same Phi, the GSE and VSE variants run up to
15.77× and 16.00× faster than their executions on a single CPU core, and up
to 1.15× and 1.15× faster than on 16 CPU cores, respectively. Meanwhile, on
the same CPU/Phi hardware configuration, the VSE variant achieves superior
performance to the naı̈ve and GSE variants. Interestingly, the naı̈ve variant was
observed to be inferior to the GSE variant on CPUs for both single-threaded
and 16-threaded settings, but became superior to the latter on the 5110P Phi
for most benchmarking datasets used. As both variants use the same core
C++ code for their corresponding CPU- and MIC-oriented implementations as
mentioned above, this observation suggests that the architectural differences
between different types of PUs may make substantial impact on the resulting
performance. Subsequently, we compared our algorithm with the third-party
counterparts implemented in the popular MATLAB and R, observing that our
algorithm on a single 5110P Phi runs up to 812× faster than 16-threaded MATLAB and up to 1, 166× faster than sequential R, both of which are executed
on high-end Intel E5-2650 CPUs. We further assessed the parallel scalability
of our algorithm with respect to number of Phis in a distributed environment,
exposing that our algorithm demonstrated rather good distributed computing
scalability.
Finally, it is worth mentioning that the current version of our VSE kernel
constrains the largest n value to 215 − 1, due to 32-bit integer packing. One
solution to overcome this constraint is packing a variable pair (ui , vi ) into a
64-bit integer with each variable occupying 32 bits. In this case, we can split
a 512-bit vector into 8 lanes with each lane representing a 64-bit integer, and
then implement vectorized pairwise merge of sorted subarrays based on 64-bit
integers. Nonetheless, it should be noted that this 64-bit solution has twice less
parallelism than 32-bit. In the future, we plan to combine this non-linear measure with other linear or non-linear correlation/dependence measures to generate
fused co-expression networks for genome-wide gene expression data analysis at
population level. In addition, since R programming language is frequently used
in data analytics and the R implementation of all-pairs τ coefficient computation
25
is extremely slow for large-scale datasets, we also plan to release a R package
for public use based on our research in this study.
Acknowledgment
This research is supported in part by US National Science Foundation under
IIS-1416259 and an Intel Parallel Computing Center award. Conflict of interest :
none declared.
References
References
[1] Q. Zhu, A. K. Wong, A. Krishnan, M. R. Aure, A. Tadych, R. Zhang, D. C.
Corney, C. S. Greene, L. A. Bongo, V. N. Kristensen, et al., Targeted exploration and analysis of large cross-platform human transcriptomic compendia, Nature Methods 12 (3) (2015) 211–214.
[2] R. Steuer, J. Kurths, C. O. Daub, J. Weise, J. Selbig, The mutual information: detecting and evaluating dependencies between variables, Bioinformatics 18 (suppl 2) (2002) S231–S240.
[3] A. J. Butte, I. S. Kohane, Unsupervised knowledge discovery in medical
databases using relevance networks., in: Proceedings of the AMIA Symposium, American Medical Informatics Association, 1999, p. 711.
[4] M. Mutwil, S. Klie, T. Tohge, F. M. Giorgi, O. Wilkins, M. M. Campbell, A. R. Fernie, B. Usadel, Z. Nikoloski, S. Persson, Planet: combined
sequence and expression comparisons across plant networks derived from
seven species, The Plant Cell 23 (3) (2011) 895–910.
[5] A. Gobbi, G. Jurman, A null model for pearson coexpression networks,
PloS One 10 (6) (2015) e0128115.
[6] A. A. Margolin, I. Nemenman, K. Basso, C. Wiggins, G. Stolovitzky, R. D.
Favera, A. Califano, ARACNE: an algorithm for the reconstruction of gene
regulatory networks in a mammalian cellular context, BMC Bioinformatics
7 (Suppl 1) (2006) S7.
[7] M. Aluru, J. Zola, D. Nettleton, S. Aluru, Reverse engineering and analysis
of large genome-scale gene networks, Nucleic Acids Research 41 (1) (2013)
e24.
[8] A. Lachmann, F. M. Giorgi, G. Lopez, A. Califano, Aracne-ap: gene network reverse engineering through adaptive partitioning inference of mutual
information, Bioinformatics 32 (14) (2016) 2233–2235.
[9] J. Lee Rodgers, W. A. Nicewander, Thirteen ways to look at the correlation
coefficient, The American Statistician 42 (1) (1988) 59–66.
26
[10] L. Song, P. Langfelder, S. Horvath, Comparison of co-expression measures:
mutual information, correlation, and model based indices, BMC Bioinformatics 13 (1) (2012) 328.
[11] C. Spearman, Spearmans rank correlation coefficient, Amer J Psychol 15
(1904) 72–101.
[12] M. G. Kendall, Rank correlation methods, Griffin, 1948.
[13] Y. Wang, Y. Li, H. Cao, M. Xiong, Y. Y. Shugart, L. Jin, Efficient test
for nonlinear dependence of two continuous variables, BMC Bioinformatics
16 (1) (2015) 1.
[14] W. Xu, Y. Hou, Y. Hung, Y. Zou, A comparative analysis of spearman’s
rho and kendall’s tau in normal and contaminated normal models, Signal
Processing 93 (1) (2013) 261–276.
[15] G. A. Darbellay, I. Vajda, et al., Estimation of the information by an
adaptive partitioning of the observation space, IEEE Transactions on Information Theory 45 (4) (1999) 1315–1321.
[16] C. O. Daub, R. Steuer, J. Selbig, S. Kloska, Estimating mutual information
using b-spline functions–an improved similarity measure for analysing gene
expression data, BMC Bioinformatics 5 (1) (2004) 1.
[17] K.-C. Liang, X. Wang, Gene regulatory network reconstruction using conditional mutual information, EURASIP Journal on Bioinformatics and Systems Biology 2008 (1) (2008) 1–14.
[18] G. J. Székely, M. L. Rizzo, N. K. Bakirov, et al., Measuring and testing
dependence by correlation of distances, The Annals of Statistics 35 (6)
(2007) 2769–2794.
[19] M. R. Kosorok, On brownian distance covariance and high dimensional
data, The Annals of Applied Statistics 3 (4) (2009) 1266.
[20] A. Gretton, O. Bousquet, A. Smola, B. Schölkopf, Measuring statistical
dependence with hilbert-schmidt norms, in: International Conference on
Algorithmic Learning Theory, Springer, 2005, pp. 63–77.
[21] D. N. Reshef, Y. A. Reshef, H. K. Finucane, S. R. Grossman, G. McVean,
P. J. Turnbaugh, E. S. Lander, M. Mitzenmacher, P. C. Sabeti, Detecting
novel associations in large data sets, Science 334 (6062) (2011) 1518–1524.
[22] L. Song, A. Smola, A. Gretton, J. Bedo, K. Borgwardt, Feature selection via
dependence maximization, Journal of Machine Learning Research 13 (May)
(2012) 1393–1434.
[23] B. H. W. Chang, W. Tian, Gsa-lightning: ultra-fast permutation-based
gene set analysis, Bioinformatics 32 (19) (2016) 3029–3031.
27
[24] H. Abdi, The kendall rank correlation coefficient, Encyclopedia of Measurement and Statistics. Sage, Thousand Oaks, CA (2007) 508–510.
[25] S. Wang, I. Pandis, D. Johnson, I. Emam, F. Guitton, A. Oehmichen,
Y. Guo, Optimising parallel R correlation matrix calculations on gene expression data using mapreduce, BMC Bioinformatics 15 (1) (2014) 1.
[26] R. C. Team, et al., R: A language and environment for statistical computing, R Foundation for Statistical Computing, 2013.
[27] J. Dean, S. Ghemawat, Mapreduce: simplified data processing on large
clusters, Communications of the ACM 51 (1) (2008) 107–113.
[28] Y. Liu, T. Pan, S. Aluru, Parallel pairwise correlation computation on
intel xeon phi clusters, in: 28th International Symposium on Computer
Architecture and High Performance Computing, IEEE, 2016, pp. 141–149.
[29] Mathworks, https://www.mathworks.com/products/matlab.html (2015).
[30] A. Sodani, R. Gramunt, J. Corbal, H.-S. Kim, K. Vinod, S. Chinthamani,
S. Hutsell, R. Agarwal, Y.-C. Liu, Knights landing: Second-generation intel
xeon phi product, IEEE Micro 36 (2) (2016) 34–46.
[31] J. Jeffers, J. Reinders, Intel Xeon Phi coprocessor high-performance programming, Morgan Kaufmann, 2013.
[32] W. R. Knight, A computer method for calculating kendall’s tau with ungrouped data, Journal of the American Statistical Association 61 (314)
(1966) 436–439.
[33] H. Inoue, T. Moriyama, H. Komatsu, T. Nakatani, Aa-sort: A new parallel
sorting algorithm for multi-core simd processors, in: Proceedings of the
16th International Conference on Parallel Architecture and Compilation
Techniques, IEEE Computer Society, 2007, pp. 189–198.
[34] K. E. Batcher, Sorting networks and their applications, in: Proceedings of
the April 30–May 2, 1968, spring joint computer conference, ACM, 1968,
pp. 307–314.
[35] J. Chhugani, A. D. Nguyen, V. W. Lee, W. Macy, M. Hagog, Y.-K. Chen,
A. Baransi, S. Kumar, P. Dubey, Efficient implementation of sorting on
multi-core simd cpu architecture, Proceedings of the VLDB Endowment
1 (2) (2008) 1313–1324.
[36] T. Xiaochen, K. Rocki, R. Suda, Register level sort algorithm on multicore simd processors, in: Proceedings of the 3rd Workshop on Irregular
Applications: Architectures and Algorithms, ACM, 2013, p. 9.
[37] S. Odeh, O. Green, Z. Mwassi, O. Shmueli, Y. Birk, Merge path-cacheefficient parallel merge and sort, Tech. rep., Technical report, CCIT Report
(2012).
28
[38] O. Green, R. McColl, D. A. Bader, Gpu merge path: a gpu merging algorithm, in: Proceedings of the 26th ACM international conference on
Supercomputing, ACM, 2012, pp. 331–340.
[39] Intel, Intel xeon phi processor software optimization guide,
https://software.intel.com/en-us/articles/intel-xeon-phi-processorsoftware-optimization-guide (2016).
[40] T. Kiefer, P. B. Volk, W. Lehner, Pairwise element computation with
MapReduce, in: Proceedings of the 19th ACM International Symposium
on High Performance Distributed Computing, ACM, 2010, pp. 826–833.
[41] W. Gropp, T. Hoefler, R. Thakur, E. Lusk, Using advanced MPI: Modern
features of the message-passing interface, MIT Press, 2014.
[42] T. El-Ghazawi, L. Smith, Upc: unified parallel c, in: Proceedings of the
2006 ACM/IEEE conference on Supercomputing, ACM, 2006, p. 27.
[43] J. González-Domı́nguez, Y. Liu, B. Schmidt, Parallel and scalable shortread alignment on multi-core clusters using upc++, PLoS ONE 11 (1)
(2016) e0145490.
29
| 2cs.AI
|
A Scalable Stream-Oriented Framework for Cluster
Applications
Tassos S. Argyros
David R. Cheriton
Distributed Systems Group
Stanford University
arXiv:cs/0504051v1 [cs.DC] 13 Apr 2005
{Argyros, Cheriton}@DSG.Stanford.EDU
ABSTRACT
This paper presents a stream-oriented architecture for structuring cluster applications. Clusters that run applications
based on this architecture can scale to tenths of thousands
of nodes with significantly less performance loss or reliability problems. Our architecture exploits the stream nature
of the data flow and reduces congestion through load balancing, hides latency behind data pushes and transparently
handles node failures. In our ongoing work, we are developing an implementation for this architecture and we are able
to run simple data mining applications on a cluster simulator.
1.
INTRODUCTION
One of the main characteristics of computing these days is
the data explosion; in fact, data nowadays grows at an exponential rate. What is really surprising, however, is that the
growth of the acquired data surpasses the increase rate of
the processing speed of today’s processors. As an example,
the amount of data in GenBank (a genomics database) doubles every 9 months, a much higher rate than the 18-month
doubling period of processors (as dictated by Moore’s law).
On the other hand, the processing of data is of imminent
value, since we continuously discover new ways to mine and
process large amounts of data to extract useful information.
Fields such as biology, physics, earth sciences and even marketing depend more and more on mining huge amounts of
data to break new ground and advance forward.
Where all this leads to is that we need to find a way to process ever-larger amounts of data with systems of reasonable
cost. A possible solution to this problem could be clusters of
commodity machines; after all, the cost of PC and network
hardware keeps falling at impressive rates. Thus, one would
argue, all we have to do to reach our goal is to buy more
PCs, larger switches, and just connect everything together.
follow this approach. First, there are severe scaling problems. Indeed, utilization of under 10% of peak performance
in large Linux clusters has been reported []. The problem
has only begin to show up in current systems, since the majority of today’s clusters has relatively small sizes between 10
and 100 nodes, with only a handful of systems being in the
range over 1000 nodes. However, in order to cope with the
rapidly increasing amounts of data, we should expect that
clusters with many thousands of nodes will become the rule
rather than the exception in a few years; moreover, clusters
with 10,000 nodes or more should start appearing soon. One
can only imagine how severe scaling problems will appear in
systems of such size.
Apart from performance problems, today’s cluster face a
number of other challenges. Node failures is one of them.
Basic probability theory dictates that the more nodes we
have in a cluster, the greater the chance that one of them will
fail within a small period of time. Identifying and correcting node failures is a hard task that today usually requires
human intervention. Moreover, writing an application to
run on a cluster, even a simple one, is much harder than it
seems. With today’s technology one has to use some sort
of message passing interface or start making remote procedure calls to transfer data. Such programming is difficult,
time consuming and hard to debug, especially in a parallel
environment.
The question that naturally follows is what can we do about
these problems. More specifically, how should the data be
send between the nodes so that congestion and latency problems do not arise? How do you deal with node and network
failures that are inevitable in a system of that size? What’s
the most effective way to interconnect such a big network?
And last but not least, how can one program and debug such
large scale applications? On a first thought, these questions
seem orthogonal to each other and inherent in every large
scale system.
Unfortunately, there are multiple problems when we try to
However, we believe that this is not true. We claim that the
problem lies in the conventional way that today’s systems
are engineered, and that an integrated solution to all of the
above questions do exist. Specifically, today’s systems are
build under the convention that the applications determine
data flow in an arbitrary way. What this means is that
a part of an application that runs on a cluster node can
potentially request data from any other node at any time;
and under this assumption, there is no systemic action that
can be taken to prevent scaling problems. Viewed from a
point outside of the application, data requests are random
accesses, and random access does not scale.
The solution that we propose is to build the whole system
based on streams. More specifically, we propose to structure
the applications in a way that they read and write streams
and build an underlying framework that handles all the data
flow issues. There are numerous advantages that result from
this approach. By infusing the stream model into the applications, we can build a framework that has enough knowledge of how the data flows within the cluster to circumvent
the scaling problems. Instead of using data requests to send
data from one node to the other (a latency-prone approach),
data can now be pushed to its destination beforehand.
Moreover, this model significantly simplifies applications programming. Issues that before needed to be handled by the
application, such as data flow and node failures, are now
handled transparently to the application by our framework.
Since all data exchange takes place using structures similar
to files, there are no complicated message-passing code that
is hard to write, prone to errors and sensitive to changes
(e.g. in the size of the cluster).
Finally, node failures become manageable since the framework has enough information to figure out what part of the
data is lost and proceed to corrective action. This way, every application that is written using streams is robust to
failures, at no expense to the programmer.
Figure 1:
In this example we have three
stages. Stage 1 outputs streams “streamA.str” and
“streamB.str”. Stage 2 inputs “streamA.str” and
stage 3 inputs “streamB.str”. Note that all details
of how stream data is delivered to nodes are handled
by the framework.
provided by previous stages),
In the rest of this paper, we present our on-going work which
can be separated to two main tasks. The first is to define the
computational model of streams: their programming representation and their semantics. The second task is the implementation of the programming framework that includes
functions such as stream operations and node failures handling. We begin by presenting the streams model.
2.
2. the declaration of the operations that should be applied on the input streams, and
3. the declaration of the output streams.
Additionally, a stage definition includes the actual function
that operates on the data units that a stream carries.
THE STREAMS MODEL
The goal of our system is to perform efficient processing
on large amounts of data. We refer to an independent and
complete processing application as a task.
An interesting observation is that in most cases a task can
be broken down to a number of independent operations that
apply on data as it flows through them. More specifically, we
define a computational stage to be a logical processing unit:
each stage receives as input a number of streams, and it has
the option to apply some specific operations to the incoming
streams. As an example, a stage could specify that it needs
to receive a particular stream sorted in some sense (we will
define exactly what this means in the next sections). The
stage then can perform any computation on the data units
of a stream, and it can output the results of the computation
as another set of streams, for the next stages to process. For
an illustration of this concept see figure 1
An important point here is to distinguish between the declaration and the definition of a stage. More specifically, a
stage declaration is comprised by:
1. the declaration of the input streams (which must be
The reason why this distinction is important has to do with
the fact that the flow of data depends on the declaration
of a stage solely. In other words, we can determine with
which other stages a stage will exchange data. Thus, the
layout of the data flow is now known before the execution
of the application, just by examining the declarations of the
stages.
A question that naturally arises is how are stages related to
nodes? This issue will be examined in the implementation
section, but as a simplification one can imagine that each
stage is assigned to a set of nodes with each node in the
set running the same piece of code (specifically the stage
function that operates on the stream data units).
2.1 Definition of Streams
“Streams” of data have been around since the first days of
computing. Usually, when one refers to a stream, she implies
a data flow with two main properties: sequentiality and uniformity. That is, a stream can be abstracted as a sequence
of data units with the restriction that these data units are
of the same kind (i.e. they share a common low-level data
representation). As an example, one could naturally define
“a stream of integers” to mean a data sequence with data
units being 32-bit integers.
that end in “.str”, but essentially any unique string is adequate. Having all streams under a unique namespace is an
important simplification, since if a stage outputs a stream
named “sorted.str”, then another stage that is interested
in processing this stream just needs to “ask” for the stream
named “sorted.str” (more details on the programming interface are provided in section 3).
2.2 Stream Windows
Figure 2: Three nodes send data to three other
nodes using a stream of data. Note that this is
a high-level representation. Implementation-wise,
nodes use TCP pair-connections to transfer data.
In our model, we have expanded this classic definition of
streams. We raise the restriction of sequentiality, in the
sense that in some cases the exact order of data is of little
importance. For example, we may have three nodes sending
data to another three nodes using a single stream (see figure
2); because the three sending nodes are not synchronized,
order is difficult to be determined here. The other extension
to the streams definition is that our model requires each data
unit to be associated with a key. We can describe the key
to be just a constant-length additional field to each stream
unit. Thus, each unit of a stream can be abstractly described
as a C struct:
Streams have infinite length in theory, and usually unknown
length in practice. Since most applications simply need results before the end of the stream, we need to find a way
to start producing results while a data stream still flows.
As a solution to this, we propose using stream windows as
computational units, upon which we can perform operations
and produce results. There are many ways to define such a
window; what we have used, and seems to make sense for a
large number of applications, is to define a window based on
a timestamp. This timestamp is defined as a non-negative
integer and it is set by the programmer in the stage where
the data is produced. It has the additional requirement that
it should be increasing; that is, stream data units that are
outputted first (at a stage’s output stream) are expected to
have smaller timestamps than the data units that follow.
More formally, a window of width T is defined to be all the
data units of a stream that have timestamps more or equal
to nT and less than (n + 1)T − 1, for n ∈ ℵ. As an example, the first window will be all the stream data units with
timestamps 0 to T − 1 and the second will have data units
with timestamp from T to 2T − 1.
Having defined the notion of window, we now proceed in
defining the stream operations.
2.3 Operations on Streams
As mentioned before, each stage can apply some operations
on the incoming streams. One could see these operations as
a kind of “queries” on streaming data. The reason why our
framework provides these operations (instead of having the
applications programmer providing them) has to do with
the need of the framework being in control of how the data
flows. In a conventional approach, the application programmer would need to program both the data flow operations
(e.g. using a message passing interface) and the operations
that apply on the data units. Our approach, in contrast,
struct StreamDataUnit {
is to distinguish these two programming tasks, provide the
KeyType key;
data flow operations by the framework and leave up to the
DataType data;
programmer the simpler but more important task of pro};
gramming the data operations. In this way, we achieve both
to give the framework full knowledge and control of the data
flow and to relieve the application programmer from the
Although the data field could change depending on the DataType hard task of programming how the data should move from
that a stream is associated to (i.e. the type representation
one node to the other.
of a stream’s data units), the type of the key, KeyType is
the same for all streams. Intuitively, key is a numeric field
In deciding which operations should be provided by the
that in some way acts as a representative for the respecframework, we wanted to select a relatively small operations
tive data in the various stream operations. In our current
set that would cover the majority of our target applications.
implementation, KeyType is a fixed-size 64-bit integer.
These operations act on windows of data, as defined in the
previous section. More specifically, these operations and
Additionally, one soon realizes the need to uniquely identheir semantics when applied to a stream are:
tify every stream. Towards this purpose, we define a global
stream namespace where each stream is associated with a
unique name. We conventionally refer to streams with names
Sort For each window, the data units of the operated stream
will arrive at each node at a sorted order (either descending or ascending) based on the key value.
Group For each window, the data of the operated stream
will arrive at each node grouped by the key value, i.e.
all stream data units of the same window that have the
same key value will arrive together at each receiving
node.
Figure 3 shows an example of the Group operation. There
we have packets of a stream flowing from the left three nodes
to the right three nodes. The stream has the Group operation applied to it, and thus packets with the same key end
up at the same node.
a boolean formula). In order to implement this for
a stream, we just add some code at the stage that
outputs the stream that only allows a data unit to be
outputted when it satisfies the mentioned criterion.
• Aggregations. Assume that we have a stream and
we want to perform some aggregation operation on it,
e.g. if the stream carries integers we may want to sum
all integers in each window. One way to implement
this is to define a stage that takes the stream to be
aggregated as an input, defines another output stream
that carries the aggregations and have that stream end
up in the same node (by giving each aggregation the
same key and applying the Group operator) where all
the aggregations would sum up to produce the final
number for each window.
One could argue whether these functions should be classified as stream operations or simple applications. The point,
however, is that the Group and Sort operations with data
units computations is a much more powerful combination
that what it seems.
2.4 Load Balancing using Failure Management
There are two reasons why a node may get overwhelmed
with data:
• The node may have lower performance compared to
the other nodes due to a hardware/software problem.
This is the classic definition of “node failure”.
Figure 3: An example of the Group operator in action. Here we have two stages, with stage 1 sending
data to stage 2 using a Grouped stream. Note how
all packets with key 10 end up in the same node.
• The actual amount of data send to the node may be
much greater than the amount send to other nodes;
thus although the node functions properly, it can not
process all the incoming data as fast as the other nodes.
We argue that these two simple operators suffice for a surprising large number of applications. Indeed, in many data
processing applications data is exchanged between the nodes
of a cluster in order to be grouped, to be sorted or to perform a seemingly different operation that on a hindsight it
is again based on sorting or grouping.
In our approach we do not distinguish between these two
cases, although they may seem initially as two totally different problems. We argue that by using the failure recovery
mechanism for both cases is the more appropriate tactic for
the following reasons.
There is a large number of operations that can be implemented using our two basic operations combined with some
in-stage computation. Here are some examples:
• Join. We can perform window-oriented joins between
streams of data. By join we mean that we output a
concatenation of data units from the joined streams
if all these data units share the same key value. The
way that this can be implemented is that we create
a stage with input the streams that we would like to
be joined with the Group operation applied. Then all
the data units with the same key value will end up at
the same node. All we have to do then is to see which
of these key values span across all streams and output
the respective data units.
• Select. A select operation can also be implemented.
By select we mean that we want a stream to contain
only key values that satisfy a specific criterion (perhaps
• In many situations it is hard to distinguish between the
two cases; thus by treating data overloads and failures
differently, there is the risk of making a wrong decision
and make the problem worse instead of fixing it.
• On an afterthought, the two problems are very similar in nature; indeed, the result of both is that a node
is not able to finish a task that it has been assigned;
thus the appropriate corrective action in both situations should be also very similar.
• Some load balancing mechanisms are anyway implemented into the failure management subsystem since
after a failure occurs, the failure must be handled in a
way that will not result in making other (non-failed)
nodes overloaded.
• The design of the system is overall simpler and more
effective when we can solve two important problems
under a single mechanism.
• Under our approach, load balancing takes into consideration all possible causes of load unbalance, such
as failures, bad data partitioning, non-uniformities in
hardware or network and corrects them in run-time.
We thus believe that the above are significant advantages
of our approach compared to other conventional load balancing mechanisms. Note that the definition of failure that
we use, i.e. a node that does not make adequate progress,
means that our load balancing mechanism can handle not
only problems of data partitioning and overloading, but also
hardware and network problems (since e.g. a bad network
link would also cause slow node progress). More details on
the implementation of the above mechanisms are given in
section 4.7.
3.
PROGRAMMING FRAMEWORK
In this section we present the programming framework that
provides the applications with the streams functionality. The
approach that we have taken is to implement the framework
as a C++ library, that is compiled along with the application
code.
processingStage->newOutputStream( dataStream );
or
processingStage->newInputStream(dataStream,SORT);
where in the second case, the second argument is the operation to be applied to the input stream.
3.2 Stage function definition
In our approach streams share many characteristics with
files. In order to read or write data from a stream, one
has to call the getStreamHandle() method that takes as an
argument the string name of the stream and whether it is
an input or an output stream. Here are some examples:
StreamHandle distrStr =
getStreamHandle("distr.str",INPUT);
StreamHandle mergeStr =
getStreamHandle("merge.str",OUTPUT);
Using the StreamHandle objects one can get or send packets
from/to the respective streams. For instance, the expression
More specifically, the programmer needs to perform two
main tasks in order to build a cluster application:
data = getPacket( distrStr );
1. The high level stage and stream declarations must be
given. That is, the stages and the streams must be created and then each stage must specify which are the
input and the output streams that it uses and what operations should be applied into the incoming streams.
2. For each stage, the programmer must provide the actual function that performs the computation on the
data units, that is reads the data from the incoming
streams, modifies it and outputs the results to the outgoing streams.
3.1 Stage and streams declarations
Our aim is to declare stages and streams in a single point
within the code. In our current implementation, the programmer needs to create a function where stages and streams
are declared as in the following example:
Ptr<Stage> processingStage =
stageManager->newStage( 10 );
would get the next packet from the queue of the stream
“distr.str”. Similarly, in order to send a packet to the
stream “merge.str” one simply has to issue the command
sendPacket(mergeStr, data);
In both these examples, data is of type DataUnit, which is
the class representation of a stream packet.
Note that in our implementation there is also a notification
mechanism for signaling the stage function when new packets arrive.
4. IMPLEMENTATION
We are currently implementing a framework that will support cluster applications that use the stream model. To
test and evaluate this framework we are also developing a
simulator of a cluster. In the next sections we present the
details of the framework implementation and the simulation
environment.
4.1 Simulation Environment
where the number 10 declares the nodes to be assigned to
this stage. Also, a stream is declared similarly as
Ptr<Stream> dataStream =
streamManager->newStream( "data.str" );
where the name of the stream is declared. After both a stage
and a stream is declared, the fact that the stage is associated
with a stream should be declared as well:
As mentioned before, we have developed a simulation environment to test and develop our framework.
The lower layer of the simulator is the network, where nodes
and their interconnect are simulated and a basic mechanism
to exchange packets exists. Packet traffic is implemented
using queues for each node and event notifications.
We are currently progressing towards implementing the full
functionality of the framework as described in this paper.
As a future work, we plan to deploy it in a small-size real
cluster for further development and testing.
4.2 Data flow
As mentioned before, an important advantage of the stream
model is that data can be pushed to its destination, instead
of having to be requested in advance. In order to do this,
we are using TCP connections; actually, TCP fits our needs
pretty well since it is stream-oriented in nature.
There is the issue of how we create and manage the connections from the application point of view. The case in most
of today’s systems is that the programmer needs to explicitly identify which connections must be created and between
which nodes specifically (e.g. one may need to specify the
IP’s of the nodes). However, in our implementation after
the declaration of stages and streams is done, a connection
is created automatically by the framework for every pair of
nodes that exchange data using a specific stream. The ID’s
of the connecting nodes and all the other network details
are hidden behind the simple specifications of stages and
streams.
To analyze this in more detail, when a stage is declared,
the number of nodes it is assigned to is also declared. Each
of the assigned nodes receive an initialization packet that
includes the streams that it inputs and outputs along with
the IP’s of the nodes that it should be connected through
these streams. Then the node opens a TCP connection with
each node that is connected through a input stream and
listens for a connection from each node that is connected
through an output stream.
When a node has made all the connections for its input and
output streams, it can start its computation. As soon as
there is some results from a stage, the nodes of that stage
push the data to its destination; for flow management we use
the standard TCP flow control mechanisms. Consecutively,
as soon as the next stages receive the first data packets, they
start processing data and producing results.
4.3 The Control Process
As soon as one tries to implement some control features,
such as failure management and load balance mechanisms,
it becomes evident that we need to synchronize the cluster
nodes on some decisions. As an example, in order to classify
a specific node of a stage as failed, all nodes of the previous
stage must agree on that fact; otherwise we may end up
in the unfortunate situation where some part of the data
still goes to the failed node (by the nodes that do not see
it as failed) while other nodes (that consider it failed) send
related data elsewhere.
In order to implement such features, a node is automatically
chosen to host a control process and it is called the control
node. This process communicates with the nodes of each
stage, receives data and feedback from them and makes decisions such as whether a node should be considered failed
or not (failure management) or how the data should be partitioned to be send to the receiving nodes (load balancing).
Note that we can have one control node per stage or one
control node for all stages or something in between. In other
words, a physical node can run as many control processes
as it can manage; we expect that the control tasks will not
be demanding in terms of processing power and bandwidth
Figure 4: Here we have three stages (1,2, and 3) and
two streams (A and B). The thin lines represent network connections, while thick lines show the logical
flow of the data. Note that we only have one control node that runs two control processes for the two
streams of the cluster.
requirements and thus a single node can act as a control
process for many stages. For an example of this fact, see
figure
4.4 Windows-based Computation and Implicit
State
In section 2.2 we argued on the need to have windows as
a computation unit. Implementation-wise, what this means
is that our total computation task is actually a sequence of
small window computations. Under this model, state is directly related to windows: the framework assumes that the
data of a window are its implicit state. Therefore, during a
window computation the only data (state) that needs to be
preserved is the data of the current window. This model enables us to perform load balancing and failure management
transparently, since the application does not have to declare
explicitly what state should be recovered in case of a failure.
4.5 GROUP Operation
In the Group operation three entities are involved: the sending stage, the receiving stage and the Group’ed stream. Remember that the semantics of the Group operation is that
we need all stream data units with the same key to end up
at the same node.
The main challenge in implementing the Group operation is
how to partition the keys in a way that no node gets either
overloaded (by receiving too much data) or stays idle (by receiving too little data). Related work in this field proposes
to make a pass over the data before deciding on how to distribute it. However, in our case this is too expensive and
often impossible since we do not have the luxury to store
the (possible endless) incoming stream data in order to determine the distribution; let alone the fact that determining
the actual distribution would require sorting the data, which
brings us back to our initial problem of how to split the data.
To tackle this problem we implement the following strategy.
As soon as some data is produced in the sending nodes, instead of streaming it to the next stage we buffer it and measure its distribution. The sending nodes use this information
to determine how the data should be split among the nodes.
After we begin sending data, we consider any overloaded
nodes as failed and use our failure handling mechanism to
redistribute the data of these nodes.
Let’s see the above procedure with some more detail. First,
we partition the data units using a hash function. Note that
we partition based on the keys of the data units, and thus
we ensure that data units with identical keys end up in the
same node. A problem that arises is that we do not know
the size of each partition in advance; nor we can assume
that all partitions have equal sizes. The way we solve this
issue is by making the number of partitions a multiple of the
number of nodes of the receiving stage; in other words, we
hash the data to a number of buckets that is many times
the number of the nodes that these buckets will end up to.
By carefully assigning many buckets per node we are able
to circumvent the problem of buckets having different sizes.
However, we can not make a proper assignment of buckets to nodes if we do not know the size of each bucket. In
order to get an estimate of this, we do not start sending
data as soon as it is available; instead, we partition the data
locally until a predetermined amount of data has been partitioned1 Explain how this is determined.. Then, the sending nodes communicate the partition sizes to the respective
stage control process (see section 4.3). The control process
executes an algorithm that figures out the optimal split of
the data (see Appendix A) and returns this information to
the nodes that immediately start sending the data to the
next stage.
What we achieve with this process is a very good starting
point in the computation of the first window. Moreover,
before we begin computing the next window we repeat a
similar process; but now we can use the actual distribution
of the data send in the previous window to best determine
how we should split the data in the next window.
However, what if in the middle of a window computation
the distribution changes unexpectedly? For example, if we
are dealing with a stock trading processing system, a sudden
increase in the trading of a specific set of stocks that end
up in the same node may overload that node and hold back
the whole computation. As mentioned, our solution in this
case is to handle overloads as failures. We examine this
mechanism in detail in section 4.7.
4.6 SORT operation
The Sort operation share many details with the Group operation. We will describe here in which ways they are similar
and we will focus on the points that they differ.
What is traditionally used for sorting data using a cluster
of machines is to partition the data into key ranges and
assign each range of keys to a specific machine. We argue
that under our stream model we can sort in a much more
efficient way. More specifically, there are two ways to sort
data using a distributed system:
1. We must ensure that each node receives keys that belong to a particular range, e.g. [min value, max value).
After each node receives all data with keys in that
range, it sorts them. Then, whenever some other process needs sorted data that fall into that range, it
queries that node. This is the conventional way of
sorting data in a cluster.
2. An alternative way is to have the sending nodes sort
whatever piece of data they have and then stream out
in sorted order the data to the next stage. Each node
of the next stage will simply need to merge the incoming data; in other words, if the data is send in descending order, the nodes of the next stage simply need to
read each time the data with the highest key from all
the incoming connections. This concept is illustrated
in figure 5.
We argue that the alternative approach has significant advantages, especially in a streaming environment like ours.
As an example, imagine an application that is only interested in the top 10% of the data. With the conventional
approach, we need to send all data through the network
since we do not know in advance which part of the data will
belong in the top 10% (and should be send out) and which
not. In contrast, with our approach the nodes will sort the
data and they will only stream it out until the next stage is
no more interested in reading more data. In other words, no
more data is going to be send over the network than what
is essential for the computation.
4.7 Failure Detection and Handling
The first step in handling a failure is to detect it. However,
since we are building a distributed system, we must make
sure that a node must be considered either as failed or not
failed by every other node that directly communicate with
it. Note that since we have partitioned our computation task
into stages, the only subset of nodes that are concerned with
a failure are these of a previous stage that send data through
a stream to the failed node. Therefore our model achieves
to restrict the impact of a failure to only a small subset of
the total nodes of the cluster.
Our approach to detecting a failure is that each stage independently decides on whether a node of a next stage is
failed or not. The nodes of a stage send periodic progress
reports to the stage control process for all nodes where they
send data. If there is a node failure then the progress report of all nodes will report that the failed node does not
seem to make enough progress; the control process will then
characterize this node as failed and it will trigger the failure
recovery mechanism.
The recovery mechanism is streams-oriented. This means
that the failure will be handled separately for each stream
that sends data to this node. The exact actions depend on
whether this stream has a Group or Sort operation assigned
to it. We begin by looking how to recover a stream with a
Group operation.
Assume that we have the stream “data.str” that is distributed Grouped from stage A to stage B. Node n that receives data in stage B fails, and stage A detects that. There
Figure 5: Sorting by merging sorted streams. In order to receive a data unit from the incoming stream, the
node in stage 2 chooses the packet with the highest key from all the incoming connections of the specific
stream.
will be a set of key buckets that stage A has associated with
node n. In implemented a failure recovery mechanism, the
following issues need to be solved:
stage after B can check these sequence numbers to simply
ignore results that were send to them before by node n.
5. RELATED WORK
1. The key partition that was assigned to the failed node
must be itself partitioned and redistributed to the other
nodes without causing any other overloads.
2. The data of the current window that was send to node
n must have been stored somewhere (apart from node
n).
3. There must be a mechanism to detect which results
n outputted before it failed, so that either stage B
doesn’t output duplicates or the stage that receives
data from stage B can recognize and ignore duplicates.
As soon as the control process decides to declare a node as
failed, it also makes a decision of how to repartition the key
space. In doing this, it assigns bigger part of the data of
the failed node to the those nodes that make the greatest
progress. In this way, a node failure actually results in a
more even load across the nodes of stage B. The algorithm
to do that is similar with the one for the initial partition
(see Appendix A) and it is not presented in detail in this
paper.
The data is redundantly stored in the sending nodes. That
is, as the nodes stream out data they also buffer it as a
back up in case of failures. After the control process has
declared a node failed, it determines the new data partition
and informs the nodes of stage A that immediately start
re-sending the buffered data following the new partition.
We moreover need to detect which results were outputted
by node n before it failed, and which not. Since we can not
count on the failed node n to give us this data, we should not
assign this task in the nodes of stage B to determine this.
Instead, with each data unit that we output we include a
sequence number that depends on the input data units that
were used to produce the specific result. The nodes of the
An other class of related work are database systems for
streams. Recently, several stream database systems have
been build [?, ?, ?]. However, most of them focus on singlenode environments and do not give any insight for implementing such systems in a distributed environment. Moreover, they are focused more on executing pre-determined
queries on data rather than acting as a framework for generic
cluster applications.
The only database projets that is deals with distributed systems issues in streams databases, to the extend of our knowledge, are the Aurora* and Medusa projects [?, ?]. Note
however that these systems are distributed in the sense that
two queries can execute in two different nodes – there is no
possibility of executing a single query using multiple nodes,
something that we attempt to do in our work. Concerning the above systems, in [?], fault recovery is examined
in distributed stream database systems. The paper follows
a replication-based approach, in contrast to our work. In
[?] several approaches on failure recovery are presented and
they are evaluated in a simulated environment.
MapReduce [?] is a programming model for writing cluster applications using two functions: the Map function produces some key/value pairs and the Reduce function merges
the pairs that share the same key. There are several points
where this work differs than ours. First, the streams programming model that we introduce is much more general
than simply using a Map/Reduce function pair, since it also
includes the notion of windows and operations. Moreover,
our system has load balancing mechanisms that are absent
from the MapReduce implementation. Finally, MapReduce
writes intermediate results to secondary storage (hard disks
on nodes) in order to be robust to failures. However, this approach has significant impact on performance. Our aim, in
contrast, is to handle failures on-the-fly and use only memory to store intermediate data.
River [?] is a system that uses streams for load balancing.
However, load balancing with River can only take place at
specific points of the computation; namely, at points where
each data unit from one stage can be send to an arbitrarily chosen node of the next stage under no restrictions (this
excludes the Group and Sort operations for instance). Our
system implements a similar mechanism for load balancing
when there are no restrictions in the flow of data; but we also
achieve to have load balancing at all stages of our computation, a much more general result. Also, like the River paper,
our load balancing mechanism can balance non-uniformities
either in data distribution, in hardware or in the network.
As part of the NOW Berkeley project [?], there has been
some work on how can one use a cluster to sort efficiently
[?]. The approach used is to split the data in ranges, that
(as we argued in section 4.6) we believe that it is not the
best direction to perform a sort operation, at least in our
application model. Also, the authors make the explicit assumption that the data follows a uniform distribution; in
our work, we do not make this assumption but rather we
use data pre-processing to approximate the actual distribution and load balance mechanisms to cope with changes or
bad approximations of the actual distribution. Also, our
algorithms are position to cope with the case that not all
source nodes carry the same amount of data, while this is
an essential assumption in the NOW-Sort paper.
There has also been some work on using streams in building faster microprocessors. Specifically, the Imagine stream
processor [?] uses a stream model to bypass the memory
bandwidth bottleneck. Based on the Imagine stream processor, there is an effort by the Merrimac project to build
a full supercomputer that is composed of stream processors
[?]. We believe that this work differs in goals, assumptions
and potential applications from our work. More specifically, Merrimac aim is to achieve an order of magnitude
more TFLOPS than conventional supercomputers. However, this greater computing capability requires pure scientific applications that perform a large number of numerical
computations and access relatively small amounts of data;
our framework, in contrast, enables generic applications that
process vast amounts of data to achieve theoretical peak performance in commodity hardware.
APPENDIX
A. SKETCH OF BUCKET DISTRIBUTION
ALGORITHM
Consider the following problem: we have a set of buckets, {b1 , b2 , . . . , bk }, each containing a number of items. Let
these numbers of items to be n1 , n2 , . . . , nk respectively. We
also have l nodes. We want to continuously group these
buckets into the l nodes, meaning that each node will get
a group of buckets in the form {bi , bi+1 , . . . , bj−1 , bj }. Our
goal is to find the bucket grouping that minimizes the variance of the number of items that each nodes receives (in
other words, it minimizes the square of the distance of the
number of items of each bucket from the average).
In order to try all possible combinations, we would need exponential time, since by applying some basic combinatorics
we find that the number of all possible instances of the prob-
lem is k+1
. However, there exists a dynamic algorithm
l−1
solution that solves the problem in polynomial time.
The basic idea of the algorithm is that if we have an optimal
distribution of k buckets into l nodes, and the i-th node
contains up to the j-th bucket, then the allocation of the
first j buckets into the first i nodes is also optimal. This is
true since if it was not optimal, the overall bucket allocation
would not be optimal as well.
Based on this observation, we construct a matrix T of dimensions l · k. A matrix cell T [i, j] contains the variance of
the optimal allocation for the first j buckets into the first i
nodes. Also, we have calculated the average number of items
per node, µ. From the mentioned optimality property, we
can calculate the T [i, j] element using values only from the
previous column using the following formula:
(
)
j
X
2
T [i, j] = min
T [i − 1, k] + (
nt − µ)
(1)
k=1...j−1
t=k+1
The variation of the optimal solution is T [l, k]. In order to
find the actual optimal bucket distribution, we construct a
second matrix D of the same dimensions as T and in each
position we write the minimum decision that we make from
expression 1. Then we can reconstruct the optimal distribution by going “backwards” from point D[l, k] to D[1, 1].
The asymptotic running time of the algorithm is O(lk) to for
the outer matrix loop and O(k) for calculating expression
12 Instead of caclculating the sum in 1 each time, we can
initially build an array and cache the sums there. This way,
the amortized time to calculate the sum is O(1).. Thus
the total running time is O(lk2 ). We have implemented the
algorithm and we have find out that it executes in reasonable
time for values of l, k near 1000 (e.g. 1000 nodes and 1000
buckets).
| 6cs.PL
|
Efficient Algorithms for Checking Fast
arXiv:1708.09253v1 [cs.LO] 29 Aug 2017
Termination in VASS
Tomáš Brázdil
Krishnendu Chatterjee
Faculty of Informatics, Masaryk University
IST Austria
[email protected]
[email protected]
Antonín Kučera
Petr Novotný
Faculty of Informatics, Masaryk University
IST Austria
[email protected]
[email protected]
Dominik Velan
Faculty of Informatics, Masaryk University
[email protected]
Abstract
Vector Addition Systems with States (VASS) consists of a finite state space equipped with
d counters (d is called the dimension), where in each transition every counter is incremented,
decremented, or left unchanged. VASS provide a fundamental model for analysis of concurrent
processes, parametrized systems, and they are also used as abstract models for programs for
bounds analysis. While termination is the basic liveness property that asks the qualitative
question of whether a given model always terminates or not, the more general quantitative
question asks for bounds on the number of steps to termination. In the realm of quantitative
bounds a fundamental problem is to obtain asymptotic bounds on termination time. Large
asymptotic bounds such as exponential or higher already suggest that either there is some error
in modeling, or the model is not useful in practice. Hence we focus on polynomial asymptotic
bounds for VASS. While some well-known approaches (e.g., lexicographic ranking functions)
are neither sound nor complete with respect to polynomial bounds, other approaches only
present sound methods for upper bounds. The existing approaches neither provide complete
methods nor provide analysis of precise complexity bounds. In this work our main contributions
are as follows: First, for linear asymptotic bounds we present a sound and complete method
for VASS, and moreover, our algorithm runs in polynomial time. Second, we classify VASS
according the normals of the vectors of the cycles. We show that singularities in the normal
are the key reason for asymptotic bounds such as exponential (even in three dimensions) and
non-elementary (even in four dimensions) for VASS. In absence of singularities, we show
that the asymptotic complexity bound is always polynomial and of the form Θ(nk ), for some
integer k ≤ d. We present an algorithm, with time complexity polynomial in the size of the
VASS and exponential in dimension d, to compute the optimal k. In other words, in absence of
1
singularities, we present an efficient sound and complete method to obtain precise (not only
upper, but matching upper and lower) asymptotic complexity bounds for VASS.
1
Introduction
Static analysis for quantitative bounds. Static analysis of programs reasons about programs without
running them. The most basic and important problem about liveness properties studied in program
analysis is the termination problem that given a program asks whether it always terminates. The
above problem seeks a qualitative or Boolean answer. However, given the recent interest in analysis
of resource-constrained systems, such as embedded systems, as well as for performance analysis, it
is vital to obtain quantitative performance characteristics. In contrast to the qualitative termination,
the quantitative termination problem asks to obtain bounds on the number of steps to termination.
The quantitative problem, which is more challenging than the qualitative one, is of great interest
in program analysis in various domains, e.g., (a) in applications domains such as hard real-time
systems, worst-case guarantees are required; and (b) the bounds are useful in early detection of
egregious performance problems in large code bases [34].
Approaches for quantitative bounds. Given the importance of the quantitative termination problem significant research effort has been devoted, including important projects such as SPEED,
COSTA [34, 35, 1]. Some prominent approaches are the following:
• The worst-case execution time (WCET) analysis is an active field of research on its own
(with primary focus on sequential loop-free code and hardware aspects) [69].
• Advanced program-analysis techniques have also been developed for asymptotic bounds,
such as resource analysis using abstract interpretation and type systems [35, 1, 45, 36, 37],
e.g., linear invariant generation to obtain disjunctive and non-linear upper bounds [19], or
potential-based methods [36, 37].
• Ranking functions based approach provides sound and complete approach for the qualitative
termination problem, and for the quantitative problem it provides a sound approach to obtain
asymptotic upper bounds [7, 9, 20, 59, 67, 21, 70, 63].
In summary, the WCET approach does not consider asymptotic bounds, while the other approaches
consider asymptotic bounds, and present sound but not complete methods for upper bounds.
VASS and their modeling power. Vector Addition Systems (VASs) [50] or equivalently Petri Nets
are fundamental models for analysis of parallel processes [25]. Enriching VASs with an underlying
finite-state transition structure gives rise to Vector Addition Systems with States (VASS). Intuitively,
a VASS consists of a finite set of control states and transitions between the control states, and and
a set of d counters that hold non-negative integer values, where at every transition between the
control states each counter is either incremented or decremented. VASS are a fundamental model
2
for concurrent processes [25], and thus are often used for performing analysis of such processes [22,
30, 48, 49]. Besides that, VASS have been used as models of parametrized systems [6], as abstract
models for programs for bounds and amortized analysis [66], as well as models of interactions
between components of an API in component-based synthesis [27]. Thus VASS provide a rich
modeling framework for a wide class of problems in program analysis.
Previous results for VASS. For a VASS, a configuration is a control state along with the values of
counters. The termination problem for VASS can be defined as follows: (a) counter termination
where the VASS terminates when one of the counters reaches value 0; (b) control-state termination
where given a set of terminating control states the VASS terminates when one of the terminating
states is reached. The termination question for VASS, given an initial configuration, asks whether
all paths from the configuration terminate. The counter-termination problem is known to be
EXPSPACE-complete: the EXPSPACE-hardness is shown in [56, 23] and the upper bound
follows from [71, 5, 26].
Asymptotic bounds analysis for VASS. While the qualitative termination problem has been studied
extensively for VASS, the problem of quantitative bounds for the termination problem has received
much less attention. In general, even for VASS whose termination can be guaranteed, the number
of steps required to terminate can be non-elementary (tower of exponentials) in the magnitude of
the initial configuration (i.e. in the maximal counter value appearing in the configuration). For
practical purposes, bounds such as non-elementary or even exponential are too high as asymptotic
complexity bounds, and the relevant complexity bounds are the polynomial ones. In this work we
study the problem of computing asymptotic bounds for VASS, focusing on polynomial asymptotic
bounds. Given a VASS and a configuration c, let nc denote the maximum value of the counters in
c. If for all configurations c all paths starting from c terminate, then let Tc denote the worst-case
termination time from configuration c (i.e., the maximum number of steps till termination among
all paths starting from c). The quantitative termination problem with polynomial asymptotic bound
given a VASS and an integer k asks whether the asymptotic worst-case termination time is at
most a polynomial of degree k, i.e., whether there exists a constant α such that for all c we have
Tc ≤ α · nkc . Note that with k = 1 (resp., k = 2, 3) the problem asks for asymptotic linear (resp.,
quadratic, cubic) bounds on the worst-case termination time. The asymptotic bound problem is
rather different from the qualitative termination problem for VASS, and even the decidability of
this problem is not obvious.
Limitations of the previous approaches for polynomial bounds for VASS. In the analysis of asymptotic bounds there are three key aspects, namely, (a) soundness, (b) completeness, and (c) precise (or
tight complexity) bounds. For asymptotic bounds, previous approaches (such as ranking functions,
potential-based methods etc) are sound (but not complete) for upper bounds. In other words, if
the approaches obtain linear, or quadratic, or cubic bounds, then such bounds are guaranteed as
asymptotic upper bounds (i.e., soundness is guaranteed), however, even if the asymptotic bound
is linear or quadratic, the approaches may fail to obtain any asymptotic upper bound (i.e., com-
3
pleteness is not guaranteed). Another approach that has been considered for complexity analysis of
programs are lexicographic ranking functions [4]. We show that with respect to polynomial bounds
lexicographic ranking functions are not sound, i.e., there exists VASS for which lexicographic
ranking function exists but the asymptotic complexity is exponential (see Example 4.11). Finally,
none of the existing approaches are applicable for tight complexity bounds, i.e., the approaches
consider O(·) bounds and are not applicable for Θ(·) bounds. In summary, previous approaches do
not provide sound and complete method for polynomial asymptotic complexity of VASS; and no
approach provide techniques for precise complexity analysis.
Our contributions. Our main contributions are related to the complexity of the quantitative termination with polynomial asymptotic bounds for VASS and our results are applicable to counter
termination.
1. We start with the important special case of linear asymptotic bounds. We present the
first sound and complete algorithm that can decide linear asymptotic bounds for all VASS.
Moreover, our algorithm is an efficient one that has polynomial time complexity. This contrast
sharply with EXPSPACE-hardness of the qualitative termination problem and shows that
deciding fast (linear) termination, which seems even more relevant for practical purposes, is
computationally easier than deciding qualitative termination.
2. Next, we turn our attention to polynomial asymptotic bounds. For simplicity, we restrict
ourselves to VASS where the underlying finite-state transition structure is strongly connected
(see Section 7 for more comments). Given such a VASS A, for every short1 cycle C of
the A, the effect of executing the short cycle once can be represented as a d-dimensional
vector, an analogue of loop summary (ignoring any nested sub-loops) for classical programs.
Let Inc denote the set of all increments, i.e., short cycle effects in A. We investigate the
geometric properties of Inc to derive complexity bounds on A. The property playing a key
role is whether all cycle effects in Inc lie on one side of some hyperplane in Rd . Formally,
each hyperplane is uniquely determined by its normal vector n (a vector perpendicular to the
hyperplane), and a hyperplane defined by n covers a vector effects v if v · n ≤ 0, where “·”
is the dot product of vectors. Geometrically, the hyperplane defined by n splits the whole
d-dimensional space into two halves such that the normal n points into one of the halves,
and its negative −n points into the other half. The hyperplane then “covers” vector v if v
points into the same half as the vector −n. We denote by Normals(A) the set of all normals
such that each n ∈ Normals(A) covers all cycle effects in A. Depending on the properties
of Normals(A), we can distinguish the following cases:
(A) No normal: if Normals(A) = ∅ (Fig. 1a);
(B) Negative normal: if all n ∈ Normals(A) have a negative component (Fig. 1b);
(C) Positive normal: if there exists n ∈ Normals(A) whose all components are positive
(Fig. 1c);
1
A cycle C is short if its length is bounded by the number of control states of a given VASS.
4
y
y
x
x
(b) All normals are negative (normal (−1.5, −1)
pictured).
(a) No normal.
y
y
x
x
(d) Singular normal (0, 1).
(c) Positive normal (1.5, 1).
Figure 1: Classification of VASS into 4 sub-classes according to the geometric properties of vectors
of cycle effects, pictured on 2D examples. Each figure pictures (as red arrows) vectors of simple
cycle effects in some VASS (it is easy, for each figure, to construct a VASS whose simple cycle
effects are exactly those pictured). The green dashed line, if present, represents the hyperplane (in
2D it is a line) covering the set of cycle effects. The thick blue arrow represents the normal defining
the covering hyperplane. The pink shaded area represents the cone generated by cycle effects (see
Section 2.3). Intuitively, we seek hyperplanes that do not intersect the interior of the cone (but can
touch its boundary).
(D) Singular normal: if (C) does not hold, but there exists n ∈ Normals(A) such that all
components of n are non-negative (in which some component of n is zero, Fig. 1d);
First, we observe that given a VASS, we can decide to which of the above category it belongs,
in time which is polynomial in the number of control states of a given VASS for every fixed
dimension (i.e., the algorithm is exponential only in the dimension d; see Section 2.2 for more
comments). Second, we also show that if a VASS belongs to one of the first two categories,
then there exist configurations with non-terminating runs from them (see Theorem 4.2).
Hence asymptotic bounds are not applicable for the first two categories and we focus on the
last two categories for polynomial asymptotic bounds.
3. For the positive normal category (C) we show that either there exist non-terminating runs
or else the worst-case termination time is of the form Θ(nk ), where k is an integer and
5
k ≤ d. We show that given a VASS in this category, we can first decide whether all runs
are terminating, and if yes, then we can compute the optimal asymptotic polynomial degree
k such that the worst-case termination time is Θ(nk ) (see Theorem 4.8). Again, this is
achievable in time polynomial in the number of control states of a given VASS for every
fixed dimension. In other words, for this class of VASS we present an efficient approach
that is sound, complete, and obtains precise polynomial complexity bounds. To the best of
our knowledge, no previous work presents a complete approach for asymptotic complexity
bounds for VASS, and the existing techniques only consider O(·) bounds, and not precise
Θ(·) bounds.
4. We show that singularities in the normal are the key reason for complex asymptotic bounds in
VASS. More precisely, for VASS falling into the singular normal category (D), in general the
asymptotic bounds are not polynomial, and we show that (a) by slightly adapting the results of
[58], it follows that termination complexity of a VASS A in category (D) cannot be bounded
by any primitive recursive function in the size of A; (b) even with three dimensions, the
asymptotic bound is exponential in general (see Example 4.9), (c) even with four dimensions,
the asymptotic bound is non-elementary in general (see Example 4.10).
The main technical contribution of this paper is a novel geometric approach, based on hyperplane
separation techniques, for asymptotic time complexity analysis of VASS. Our methods are sound
for arbitrary VASS and complete for a non-trivial subclass.
2
2.1
Preliminaries
Basic Notation
We use N, Q, and R to denote the sets of non-negative integers, rational numbers, and real numbers.
The subsets of all positive elements of N, Q, and R are denoted by N+ , Q+ , and R+ . Further, we
use N∞ to denote the set N ∪ {∞} where ∞ is treated according to the standard conventions. The
cardinality of a given set M is denoted by |M |. When no confusion arises, we also use |c| to denote
the absolute value of a given c ∈ R.
Given a function f : N → N, we use O(f (n)) and Ω(f (n)) to denote the sets of all g : N → N
such that g(n) ≤ a · f (n) and g(n) ≥ b · f (n) for all sufficiently large n ∈ N, where a, b ∈ R+
are some constants. If h(n) ∈ O(f (n)) and h(n) ∈ Ω(f (n)), we write h(n) ∈ Θ(f (n)).
Let d ≥ 1. The elements of Rd are denoted by bold letters such as u, v, z, . . .. The i-th component
of v is denoted by v(i), i.e., v = (v(1), . . . , v(d)). For every n ∈ N, we use ~n to denote the
constant vector where all components are equal to n. The scalar product of v, u ∈ Rd is denoted
P
by v · u, i.e., v · u = di=1 v(i) · u(i). The other standard operations and relations on R such as +,
≤, or < are extended to Rd in the component-wise way. In particular, v is positive if v > ~0, i.e., all
p
components of v are positive. The norm of v is defined by norm(v) = v(1)2 + · · · + v(d)2 .
6
(0,0)
q1
(-1,1)
(0,0)
q2
(1,-1)
(-1,0)
q1
q2
(-1,0)
(-1,1)
(a) Quadratic complexity.
(b) Non-terminating VASS.
(1,-1)
(-1,0)
q1
(-1,1)
(0,0)
q2
(1,-1)
(c) Linear complexity.
Figure 2: An example of 2-dimensional VASS of varying complexity.
Half-spaces and Cones. An open half-space of Rd determined by a normal vector n ∈ Rd ,
where n 6= ~0, is the set Hn of all x ∈ Rd such that x · n < 0. A closed half-space Ĥn is defined in
the same way but the above inequality is non-strict. Given a finite set of vectors U ⊆ Rd , we use
P
cone(U ) to denote the set of all vectors of the form u∈U cu u, where cu is a non-negative real
constant for every u ∈ U .
Example 2.1. In Fig. 1, the cone, or more precisely its part that intersects the displayed area
of R2 , generated by the cycle effects (i.e., by the “red” vectors) is the pink-shaded area. As
for the half spaces, e.g., in Fig. 1d, the closed half-space defined by the normal vector (0, 1) is
the set {(x, y) | y ≤ 0}, while the open half-space determined by the same normal is the set
{(x, y) | y < 0}. Intuitively, each normal vector n determines a hyperplane (pictured by dashed
lines in Fig. 1) that cuts Rd in two halves, and Hn is the half which does not contain n: depending
on whether we are interested in closed or open half-space, we include the separating hyperplane
into Hn or not, respectively.
2.2
Syntax and semantics of VASS
In this subsection we present a syntax of VASS, represented as finite state graphs with transitions
labelled by vectors of counter changes.
Definition 2.2. Let d ∈ N+ . A d-dimensional vector addition system with states (VASS) is a
pair A = (Q, T ), where Q 6= ∅ is a finite set of states and T ⊆ Q × {−1, 0, 1}d × Q is a set of
transitions.
Example 2.3. Fig. 2 shows examples of three small 2-dimensional VASS. The VASS in Fig. 2a
has two states q1 , q2 and four transitions (q1 , (−1, 1), q2 ), (q1 , (0, 0), q2 ), (q2 , (−1, 0), q1 ),
(q2 , (1, −1), q2 ).
In some cases, we design algorithms where the time complexity is not polynomial in ||A|| (i.e., the
size of A), but polynomial in |Q| and exponential just in d. Then, we say that the running time is
polynomial in |Q| for a fixed d.
7
We use simple operational semantics for VASS based on the view of VASS as finite-state machines
augmented with non-negative integer-valued counters.
A configuration of A is a pair pv, where p ∈ Q and v ∈ Nd . The set of all configurations of A is
denoted by C (A). The size of pv ∈ C (A) is defined as ||pv|| = max{v(i) | 1 ≤ i ≤ d}.
A finite path in A of length n is a finite sequence π of the form p0 , u1 , p1 , u2 , p2 , . . . , un , pn where
n ≥ 1 and (pi , ui+1 , pi+1 ) ∈ T for all 0 ≤ i < n. If p0 = pn , then π is a cycle. A cycle is short
if n ≤ |Q|. The effect of π, denoted by eff (π), is the sum u1 + · · · + un . Given two finite paths
α = p0 , u1 , . . . , pn and β = q0 , v1 , . . . , qm such that pn = q0 , we use α
β to denote the finite
path p0 , u1 , . . . , pn , v1 , . . . , qm .
Let π be a finite path in A. A decomposition of π into short2 cycles, denoted by Decomp(π), is a
finite list of short cycles (repetitions allowed) defined recursively as follows:
• If π does not contain any short cycle, then Decomp(π) = [], where [] is the empty list.
• If π = α
γ
β where γ is the first short cycle occurring in π, then Decomp(π) =
Concat([γ], Decomp(α
β)), where Concat is the list concatenation operator.
Observe that if Decomp(π) = [], then the length of π is at most |Q| − 1. Since the length of every
short cycle is bounded by |Q|, the length of π is asymptotically the same as the number of elements
in Decomp(π), assuming a fixed VASS A.
Given a path π = p0 , u1 , p1 , u2 , p2 , . . . , un , pn and an initial configuration p0 v0 , the execution of
π in p0 v0 is a finite sequence p0 v0 , . . . , pn vn of configurations where vi = v0 + u1 + · · · + ui
for all 0 ≤ i ≤ n. If vi ≥ ~0 for all 0 ≤ i ≤ n, we say that π is executable in p0 v0 .
2.3
Termination Complexity of VASS
A zero-avoiding computation of length n initiated in a configuration pv is a finite sequence of
configurations α = q0 z0 , . . . , qn zn initiated in pv such that zi > ~0 for all 0 ≤ i ≤ n, and for
each 0 ≤ i < n there is a transition (qi , u, qi+1 ) ∈ T where zi+1 = zi + u. Every zero-avoiding
computation α initiated in q0 z0 determines a unique finite path πα in A such that α is the execution
of πα in q0 z0 .
Definition 2.4. Let A = (Q, T ) be a d-dimensional VASS. For every configuration pv of A, let
L(pv) be the least ` ∈ N∞ such that the length of every zero-avoiding finite computation initiated
in pv is bounded by `. The termination complexity of A is a function L : N → N defined by
L(n) = max {L(pv) | pv ∈ C (A) where ||pv|| = n} .
If L(n) = ∞ for some n ∈ N, we say that A is non-terminating, otherwise it is terminating.
2
A standard technique for analysing paths in VASS are decompositions into simple cycles, where all states except for
p0 and pn are pairwise different. The reason why we use short cycles instead of simple ones is clarified in Lemma 2.5.
8
Observe that if A is non-terminating, then L(n) = ∞ for all sufficiently large n ∈ N. Further, if A
is terminating, then L(n) ∈ Ω(n). In particular, if L(n) ∈ O(n), we also have L(n) ∈ Θ(n).
Given a path π = p0 , u1 , p1 , u2 , p2 , . . . , un , pn and an initial configuration p0 v0 , the execution of
π in p0 v0 is a finite sequence p0 v0 , . . . , pn vn where vi = v0 + u1 + · · · + ui for all 0 ≤ i ≤ n.
If vi ≥ ~0 for all 0 ≤ i ≤ n, we say that π is executable in p0 v0 .
Let Inc = {eff (π) | π is a short cycle of A} . The elements of Inc are called increments. Note
that if u ∈ Inc, then −|Q| ≤ u(i) ≤ |Q| for all 1 ≤ i ≤ d. Hence, |Inc| is polynomial in |Q|,
assuming d is a fixed constant. Although the total number of all short cycles can be exponential
in |Q|, the set Inc is computable efficiently.3
Lemma 2.5. Let A = (Q, T ) be a d-dimensional VASS, and let p ∈ Q. The set Inc is computable
in time O(||A||d ), i.e., polynomial in |Q| assuming d is a fixed constant.
Proof. The set Inc is computable by the following standard algorithm: For all q, q 0 ∈ Q and
k be the set of all effects of paths from q to q 0 of length exactly k. Observe that
1 ≤ k ≤ n, let Eq,q
0
1 = {u | (q, u, q 0 ) ∈ T } for all q, q 0 ∈ Q;
• Eq,q
0
S
k−1
k =
00
0
• for every 1 < k ≤ |Q|, we have that Eq,q
0
q 00 ∈Q {v+u | v ∈ Eq,q 00 and (q , u, q ) ∈ T }.
S
S
k , and the sets E k
Obviously, Inc = q∈Q nk=1 Eq,q
q,q 0 for k ≤ |Q| are computable in time
polynomial in |Q|, assuming d is a fixed constant.
A strongly connected component (SCC) of A is maximal R ⊆ Q such that for all p, q ∈ R where
p 6= q there is a finite path from p to q. Given a SCC R of Q, we define the VASS AR by restricting
the set of control states to R and the set of transitions to T ∩ (R × {−1, 0, 1}d × R). We say that
A is strongly connected if Q is a SCC of A.
3
Linear Termination Time
In this section, we give a complete and effective characterization of all VASS with linear termination
complexity.
More precisely, we first provide a precise mathematical characterization of VASS with linear
complexity: we show that if A is a d-dimensional VASS, then L(n) ∈ O(n) iff there is an open
half-space Hn of Rd such that n > ~0 and Inc ⊆ H.
Next we show that the mathematical characterization of VASS of linear complexity is equivalent to
the existence of a ranking function of a special form for this VASS. We also show that existence
of such a function for a given VASS A can be decided (and the function, if it exists, synthesized)
3
Note that Lemma 2.5 would not hold if we used simple cycles instead of short cycles, because the problem
whether a given vector v is an effect of a simple cycle is NP-complete, even if d = 1 (NP-hardness follows, e.g., by a
straightforward reduction of the Hamiltonian path problem).
9
in time polynomial in the size of A. Hence, we obtain a sound and complete polynomial-time
procedure for deciding whether a given VASS has linear termination complexity.
We start with the mathematical characterization. Due to the next lemma, we can safely restrict
ourselves to strongly connected VASS. A proof is trivial.
Lemma 3.1. Let d ∈ N, and let A = (Q, T ) be a d-dimensional VASS. Then L(n) ∈ O(n) iff
LR (n) ∈ O(n) for every SCC R of Q, where LR (n) is the termination complexity of AR .
Now we show that if there is no open half-space Hn such that n > ~0 and Inc ⊆ Hn , then there
P
exist short cycles γ1 , . . . , γk and coefficients b1 , . . . , bk ∈ N+ such that the sum ki=1 bi · eff (γi )
is non-negative. Note that this does not yet mean that A is non-terminating—it may happen that
the cycles π1 , . . . , πk pass through disjoint subsets of control states and cannot be concatenated
without including auxiliary finite paths decreasing the counters.
Lemma 3.2. Let A = (Q, T ) be a d-dimensional VASS. If there is no open half-space Hn of Rd
such that n > ~0 and Inc ⊆ Hn , then there exist v1 , . . . , vk ∈ Inc and b1 , . . . , bk ∈ N+ such that
P
k ≥ 1 and ki=1 bi vi ≥ ~0.
Proof. We distinguish two possibilities.
(a) There exists a closed half-space Ĥn of Rd such that n > ~0 and Inc ⊆ Ĥn .
(b) There is no closed half-space Ĥn of Rd such that n > ~0 and Inc ⊆ Ĥn .
Case (a). We show that there exists u ∈ Inc such that u 6= ~0 and −u ∈ cone(Inc). Note that this
immediately implies the claim of our lemma—since −u ∈ cone(Inc), there are v1 , . . . , vk ∈ Inc
P
and c1 , . . . , ck ∈ R+ such that −u = ki=1 ci vi . Since all elements of Inc are vectors of nonnegative integers, we can safely assume ci ∈ Q+ for all 1 ≤ i ≤ k. Let b be the least common
multiple of c1 , . . . , ck . Then bu + (b · c1 )v1 + · · · + (b · ck )vk = ~0 and we are done.
It remains to prove the existence of u. Let us fix a normal vector n > ~0 such that Inc ⊆ Ĥn and
the set Inc n = {v ∈ Inc | v · n < 0} is maximal (i.e., there is no n0 > ~0 satisfying Inc ⊆ Ĥn0
and Inc n ⊂ Inc n0 ). Further, we fix u ∈ Inc such that u · n = 0. Note that such u ∈ Inc
must exist, because otherwise Inc n = Inc which contradicts the assumption of our lemma. We
show −u ∈ cone(Inc). Suppose the converse. Then by Farkas’ lemma there exists a separating
hyperplane for cone(Inc) and −u with normal vector n0 , i.e., v · n0 ≤ 0 for all v ∈ Inc and
−u · n0 > 0. Since n > ~0, we can fix a sufficiently small ε > 0 such that the following conditions
are satisfied:
• n + εn0 > ~0,
• for all v ∈ Inc such that v · n < 0 we have that v · (n + εn0 ) < ~0.
Let w = n + εn0 . Then w > 0, v · w < 0 for all v ∈ Inc n , and u · w = u · n + ε(u · n0 ) =
ε(u · n0 ) < 0. This contradicts the maximality of Inc n .
10
Case (b). Let B = {v ∈ Rd | v ≥ ~0 and 1 ≤
Pd
i=1 v(i)
≤ 2}. We prove cone(Inc) ∩ B 6= ∅,
which implies the claim of our lemma (there are v1 , . . . , vk ∈ Inc and c1 , . . . , ck ∈ Q+ such that
Pk
i=1 ci vi ∈ B). Suppose the converse, i.e., cone(Inc) ∩ B = ∅. Since both cone(Inc) and B are
closed and convex and B is also compact, we can apply the “strict” variant of hyperplane separation
theorem. Thus, we obtain a vector n ∈ Rd and a constant c ∈ R such that x · n < c and y · n > c
for all x ∈ cone(Inc) and y ∈ B. Since ~0 ∈ cone(Inc), we have that c > 0. Further, n ≥ ~0 (to
see this, realize that if n(i) < 0 for some 1 ≤ i ≤ d, then y · n < 0 where y(i) = 1 and y(j) = 0
for all j 6= i; since y ∈ B and c > 0, we have a contradiction). Now we show x · n ≤ 0 for
all x ∈ cone(Inc), which contradicts the assumption of Case (b). Suppose x · n > 0 for some
x ∈ cone(Inc). Then (m · x) · n > c for a sufficiently large m ∈ N. Since m · x ∈ cone(Inc), we
have a contradiction.
Now we give the promised characterization of all VASS with linear termination complexity. Our
theorem also reveals that the VASS termination complexity is either linear or at least quadratic (for
example, it cannot be that L(n) ∈ Θ(n log n)).
Theorem 3.3. Let A = (Q, T ) be a d-dimensional VASS. We have the following:
(a) If there is an open half-space Hn of Rd such that n > ~0 and Inc ⊆ Hn , then L(n) ∈ O(n).
(b) If there is no open half-space Hn of Rd such that n > ~0 and Inc ⊆ Hn , then L(n) ∈ Ω(n2 ).
Proof. We start with (a). Let Hn be an open half-space of Rd such that n > ~0 and Inc ⊆ Hn , and
let qu be a configuration of A. Note that dn · ue ∈ O(||qu)||) because n does not depend on qu.
Let δ = minv∈Inc |v · n|. Each short cycle decreases the scalar product of the normal n and vector
of counters by at least δ. Therefore, for every zero-avoiding computation α initiated in qu we have
that Decomp(α) contains at most O(||qu||) elements, so the length of α is O(||qu||).
Now suppose there is no open half-space Hn of Rd such that n > ~0 and Inc ⊆ Hn . We show
that L(n) ∈ Ω(n2 ), i.e., there exist p ∈ Q and a constant a ∈ R+ such that for all configurations
p~n, where n ∈ N is sufficiently large, there is a zero-avoiding computation initiated in p~n whose
length is at least a · n2 . Due to Lemma 3.1, we can safely assume that A is strongly connected. By
Lemma 3.3, there are v1 , . . . , vk ∈ Inc and b1 , . . . , bk ∈ N+ such that k ≥ 1 and
k
X
bi vi ≥ ~0.
(1)
i=1
As the individual short cycles with effects v1 , . . . , vk may proceed through disjoint sets of states,
they cannot be trivially concatenated into one large cycle with non-negative effect. Instead, we fix
a control state p ∈ Q and a cycle π initiated in p visiting all states of Q (here we need that A is
strongly connected). Further, for every 1 ≤ i ≤ k we fix a short cycle γi such that eff (γi ) = vi .
For every t ∈ N, let πt be a cycle obtained from π by inserting precisely t · bi copies of every γi ,
where 1 ≤ i ≤ k. Observe that the inequality (1) implies
eff (πt ) = eff (π) + t ·
k
X
bi vi ≥ eff (π)
i=1
11
for every t ∈ N.
(2)
For every configuration pu, let t(u) be the largest t ∈ N such that πt is executable in pu and results
in a zero-avoiding computation. If such a t(u) does not exist, i.e. πt is executable in pu for all
t ∈ N, then A is non-terminating (since, e.g. v1 must be non-negative in such a case), and the
proof is finished. Hence, we can assume that t(u) is well-defined for each u. Since the cycles
π and γ1 , . . . , γk have fixed effects, there is b ∈ R+ such that for all configurations pu where
all components of u (and thus also ||pu||) are above some sufficiently large threshold ξ we have
that t(u) ≥ b · ||pu||, i.e. t(u) grows asymptotically at least linearly with the minimal component
of u. Now, for every n ∈ N, consider a zero-avoiding computation α(n) initiated in p~n defined
inductively as follows: Initially, α(n) consists just of pu0 = p~n; if the prefix of α(n) constructed
so far ends in a configuration pui such that t(ui ) ≥ 1 and ui ≥ ξ~ (an event we call a successful hit),
then the prefix is prolonged by executing the cycle πt(ui ) (otherwise, the construction of α(n) stops).
Thus, α(n) is obtained from p~n by applying the inductive rule I(n) times, where I(n) ∈ N∞ is the
number of successful hits before the construction of α(n) stops. Denote by pui the configuration
visited by α(n) at i-th successful hit. Now the inequality (2) implies that ui ≥ ~n + i · eff (π), so
there exists a constant e such that ||pui || ≥ n − i · e. In particular the decrease of all components
of ui is at most linear in i. This means that I(n) ≥ c · n for all sufficiently large n ∈ N, where
~ so
c ∈ R+ is a suitable constant. But at the same time, upon each successful hit we have ui ≥ ξ,
length of the segment beginning with i-th successful hit and ending with the (i + 1)-th hit or with
the last configuration of α(n) is at least b · ||pui || ≥ b · (n − i · e). Hence, the length of α(n) is at
P
least c·n
i=1 b · (n − i · e), i.e. quadratic.
Example 3.4. Consider the VASS in Figure 2c. It consists of two strongly connected components,
{q1 } and {q2 }. In A{q1 } we have Inc = {(−1, 1)}. For n = (1, 21 ) the open half-space Ĥn
contains Inc. Similarly, in A{q2 } we have Inc = {(−1, 0), (1, −1)}. For n = (1, 2) we again have
that Inc is contained in open half-space Ĥn . Hence, the VASS has linear termination complexity.
It is strongly connected and Inc
Now consider the VASS in Figure 2a.
=
{(−1, 1), (−2, 2), (1, −1), (2, −2), (−1, 0)}. But there cannot be an open 2-dimensional halfspace (i.e. an open half-plane) containing two opposite vectors, e.g. (−1, 1) and (1, −1), because
for any line going through the origin such that (−1, 1) does not lie on the line it holds that (1, −1)
lies on the “other side” of the line than (−1, 1). Hence, the VASS in Figure 2a has at least quadratic
termination complexity. The same argument applies to VASS in Figure 2b.
A straightforward way of checking the condition of Theorem 3.3 is to construct the corresponding
linear constraints and check their feasibility by linear programming. This would yield an algorithm
polynomial in |Inc|, i.e., polynomial in |Q| for every fixed dimension d. Now we show that the
condition can actually be checked in time polynomial in the size of A. We do this by showing that
the mathematical condition stated in Theorem 3.3 is equivalent to the existence of a ranking function
of a special type for a given VASS. Formally, a weighted linear map for a VASS A = (Q, T )
is defined by a vector of coefficients c and by a set of weights {hq | q ∈ Q}, one constant for
each state of A. The weighted linear map µ = (c, {hq | q ∈ Q}) defines a function (which we,
slightly abusing the notation, also denote by µ) assigning numbers to configurations as follows:
12
µ(pv) = c · v + hp . A weighted linear map µ is a weighted linear ranking (WLR) function for A
if c ≥ ~0 and there exists > 0 such that for each configuration pv and each transition (p, u, q) it
holds µ(pv) ≥ µ(q(v + u)) + , which is equivalent, due to linearity, to
hp − hq ≥ c · u +
(3)
We show that weighted linear ranking functions provide a sound and complete method for proving
linear termination complexity of VASS.
Theorem 3.5. Let d ∈ N. The problem whether the termination complexity of a given d-dimensional
VASS is linear is solvable in time polynomial in the size of A. More precisely, the termination
complexity of a VASS A is linear if and only if there exists a weighted linear ranking function for
A. Moreover, the existence of a weighted linear ranking function for A can be decided in time
polynomial in ||A||.
Proof Sketch. In the course of the proof we describe a polynomial time-algorithm for deciding
whether given VASS has linear termination complexity. Once the algorithm is described, we will
show that what it really does is checking the existence of a weighted linear ranking function for A.
Let us start by sketching the underlying intuition. Our goal is to decide, in polynomial time, whether
there is an open half-space Hn of Rd such that n > ~0 and Inc ⊆ Hn . Note that this is equivalent to
deciding whether there is an open half-space Hn of Rd such that n ≥ ~0 and Inc ⊆ Hn (since we
demand Hn to be open and the scalar product is continuous, n ≥ ~0 can be slightly tilted by adding
a small ~δ > 0 to obtain a positive vector with the desired property).
Given a vector n ∈ Rd and a configuration qv, we say that v · n is the n-value of qv. Observe
that if there is an open half-space Hn such that n ≥ ~0 and Inc ⊆ Hn , then there is ε > 0 such
that the effect of every short cycle decreases the n-value of a configuration by at least ε. As every
path can be decomposed into short cycles, every path steadily decreases the n-value of visited
configurations. It follows that the mean change (per transition) of the n-value along an infinite path
is bounded from above by −ε/|Q|. On the other hand, if the maximum mean change in n-values
(over all infinite paths) is bounded from above by some negative constant, then every short cycle
must decrease the n-value by at least this constant. So, it suffices to decide whether there is n ≥ ~0
such that for all infinite paths the mean change of the n-value is negative. Thus, we reduce our
problem to the classical problem of maximizing the mean payoff over a decision process with
rewards. Using standard results (see, e.g., [60]), the latter problem polynomially reduces to the
problem of solving a linear program that is (essentially) equivalent to the inequality (3). Finally,
the linear program can be solved in polynomial time using e.g. [51].
Remark 3.6. The weighted linear ranking functions can be seen as a special case of well-known
linear ranking functions for linear-arithmetic programs [20, 59], in particular state-based linear
ranking functions, where a linear function of program variables is assigned to each state of the
control flow graph. WLR ranking functions are indeed a special case, since the linear functions
assigned to various state are almost identical, and they differ only in the constant coefficient hq .
13
Also, as the proof of the previous theorem shows, WLR functions in VASS can be computed directly
by linear programming, without the need for any “supporting invariants,” since effect of a transition
in VASS is independent of the current values of the counters. Also, well-foundedness (i.e. the fact
that the function is bounded from below) is guaranteed by the fact that n ≥ 0 and counter values in
VASS are always non-negative. It is a common knowledge that the existence of a state-based linear
ranking function for a linear arithmetic program implies that the running time of the program is
linear in the initial valuation of program variables. Hence, our main result can be interpreted
as proving that for VASS, state-based linear ranking functions are both sound and complete for
proving linear termination complexity.
4
Polynomial termination time
In this section we concentrate on VASS with polynomial termination complexity. For simplicity,
we restrict ourselves to strongly connected VASS. As we already indicated in Section 1, our analysis
proceeds by considering properties of normal vectors perpendicular to hyperplanes covering the
vectors of Inc.
Definition 4.1. Let A = (Q, T ) be a d-dimensional VASS. The set Normals(A) consists of all
n ∈ Rd such that n 6= ~0 and Inc ⊆ Ĥn (i.e., v · n ≤ 0 for all v ∈ Inc).
Let A be a strongly connected VASS. We distinguish four possibilities.
(A) Normals(A) = ∅.
(B) Normals(A) 6= ∅ and all n ∈ Normals(A) have a negative component.
(C) There exists n ∈ Normals(A) such that n > ~0.
(D) There exists n ∈ Normals(A) such that n ≥ ~0 and (C) does not hold.
Note that one can easily decide which of the four conditions holds by linear programming. Due to
Lemma 2.5, the decision algorithm is polynomial in the number of control states of A (assuming d
is a fixed constant).
We start by showing that a VASS satisfying (A) or (B) is non-terminating. A proof is given in
Section 5.2.
Theorem 4.2. Let A = (Q, T ) be a d-dimensional strongly connected VASS such that (A) or (B)
holds. Then A is non-terminating.
4.1
VASS satisfying condition (C)
Assume A is a d-dimensional VASS satisfying (C). We prove that if A is terminating, then
L(n) ∈ Θ(n` ) for some ` ∈ {1, . . . , d}. Further, there is a polynomial-time algorithm deciding
whether A is terminating and computing the constant ` if it exists (assuming d is a fixed constant).
14
A crucial tool for our analysis is a good normal, introduced in the next definition.
Definition 4.3. Let A be a VASS. We say that a normal n ∈ Normals(A) is good if n > ~0 and for
every v ∈ cone(Inc) we have that −v ∈ cone(Inc) iff v · n = 0.
Example 4.4. Consider the VASS of Fig. 2a. Here, a good normal is, e.g., the vector n = (1, 1).
Observe that the effects of both self-loops (on q1 and q2 ) belong to the hyperplane defined by (1, 1).
Note that these loops compensate each other’s effects so long as we stay in the hyperplane (this is
the defining property of the good normal). This allows us to zig-zag in the plane without "paying"
with decrements in the n-value except when we need to switch between the loops (recall that the
n-value of a configuration qv is the product v · n). This produces a path of quadratic length, which
is asymptotically the worst case.
The next lemma says that a good normal always exists and it is computable efficiently. A proof can
be found in Section 5.3.
Lemma 4.5. Let A be a d-dimensional VASS satisfying (C). Then there exists a good normal
computable in time polynomial in |Q|, assuming d is a fixed constant.
The next theorem is the key result of this section. It allows to reduce the analysis of termination
complexity of a given VASS to the analysis of several smaller instances of the problem, which can
be then solved recursively.
Theorem 4.6. Let A be a VASS satisfying (C), and let n ∈ Normals(A) be a good normal.
Consider a VASS An = (Q, Tn ) where
Tn = {t ∈ T | there is a short cycle γ of A containing t such that eff (γ) · n = 0} .
Further, let C1 , . . . , Ck be all SCC of An with at least one transition. We have the following:
(1) If k = 0 (i.e., if there is no SCC of An with at least one transition), then LA (n) ∈ Θ(n).
(2) If k > 0, all AnC1 , . . . , AnCk are terminating, and the termination complexity of every AnCi is Θ(fi (n)), then A is terminating and LA (n) ∈ Θ(n · max[f1 , . . . , fk ](n)),
where max[f1 , . . . , fk ] : N → N is a function defined by max[f1 , . . . , fk ](n) =
max{f1 (n), . . . , fk (n)}.
To get some intuiting behind the proof of Theorem 4.6, consider the following example.
Example 4.7. Consider the VASS of Fig. 2a. As mentioned in Example 4.4, there is a good normal
n = (1, 1), which gives Tn = {(−1, 1), (1, −1)}. Then Case (2) of Theorem 4.6 gives us two
simpler VASS AnC1 , AnC2 where AnC1 has a single state q1 and a single transition (q1 , (−1, 1), q1 ),
and AnC2 has a single state q2 and a single transition (q2 , (1, −1), q2 ). Observe that both AnC1
and AnC2 can now be considered individually, and both of them have linear complexity. Also, as
mentioned in Example 4.4, the good normal makes sure that the effect of the worst case behavior
in AnC1 can be compensated by a path in AnC2 , and vice versa. Moreover, following the worst
case path in AnC1 and its compensation in AnC2 decreases the final n-value of configurations only
by a constant (caused by the switch between AnC1 and AnC2 ). So, we can follow such “almost
compensating” loop Ω(n) times, and obtain a path of quadratic length.
15
Note that in the general case the situation is more complicated since the compensating path may
need to be composed using paths in several VASS of AnC1 , . . . , AnCk . So, we need to be careful about
the number of switches and about geometry of the compensating path.
Proof sketch for Theorem 4.6. Claim (1) follows easily. It suffices to realize that if there is no SCC
of An with at least one transition, then there is no v ∈ Inc satisfying v · n = 0. Hence, v · n < 0
for all v ∈ Inc, and we can apply Theorem 3.3.
Now we prove Claim (2). Let α be a zero-avoiding computation of A initiated in a configuration
qu. Since the last configuration pv of α satisfies v ≥ ~0, we have that v · n ≥ 0. Hence,
v·n
=
(u + eff (πα )) · n
=
u · n + eff (πα ) · n
≥
0.
Let Decomp(πα ) by a decomposition of πα into short cycles. For every short cycle γ of A we have
that eff (γ) · n ≤ 0. Since πα can contain at most |Q| transitions which are not contained in any
cycle, we have that u · n ≤ v · n + c, where c ∈ N is some fixed constant. This means that ||pv|| is
O(||qu||). Consequently, the same holds also for all intermediate configurations visited by α.
A short cycle γ of A such that eff (γ) · n < 0 is called n-decreasing, otherwise it is n-neutral.
Clearly, the total number of n-decreasing short cycles in Decomp(πα ) is O(||qu||), because each of
them decreases the scalar product with n by a fixed constant bounded away from zero, and u · n is
O(||qu||). This means that the total number of transitions in πα which are not in Tn is O(||qu||) (as
we already noted, πα can also contain transitions which are not contained in any short cycle, but
their total number is bounded by |Q|). Let % be a subpath of πα with maximal length containing only
transitions of Tn . Note that % is a concatenation of at most |Q| subpaths which contain transitions
of the same SCC Ci of An . Each of these subpaths is initiated in a configuration of size O(||qu||),
and hence its length is O(fi (||qu||)). Hence, the length of % is O(||qu|| · max[f1 , . . . , fk ](||qu||)).
It remains to prove that LA (n) ∈ Ω(n · max[f1 , . . . , fk ](n)). Let us fix some i ≤ k. We prove that
there exists a constant a ∈ R+ such that for all sufficiently large n there exists a zero-avoiding
computation αn of length at least a · n · fi (n) initiated in a configuration of size n. The construction
of αn is technically non-trivial, so we first explain the underlying idea informally. A formal proof
is given in Section 5.4.
To achieve the length Ω(n · fi (n)), the computation αn needs to execute Ω(n) paths of length
Θ(fi (n)) “borrowed” from AnCi . The problem is that even after executing just one path π of length
Θ(fi (n)), some counters can have very small values, which prevents the executing of another
path of length Θ(fi (n)). Therefore, we need to “compensate” the effect of π and increase the
counters. This is where we use the properties of a good normal. We can choose π so that it forms
a cycle in AnCi (not necessarily a short one), and we prove that all cycles in AnCi are n-neutral.
From this we get − eff (π) ∈ cone(Inc), and hence the effect of π can be compensated by an
appropriate combination of short cycles of An . So, after executing π, we execute the corresponding
“compensating” path, and this is repeated Ω(n) times. Note that we need to ensure that the
compensating paths do not decrease the counters too much in intermediate configurations, and the
16
compensation ends in a configuration which is sufficiently close to the original configuration where
we started executing π.
Now we can formulate and prove the main result of this section.
Theorem 4.8. Let A be a d-dimensional VASS satisfying (C). The problem whether A is terminating
is decidable in time polynomial in |Q|, assuming d is a fixed constant. Further, if A is terminating,
then L(n) ∈ Θ(nk ), where k ∈ {1, . . . , d} is a constant computable in time polynomial in |Q|,
assuming d is a fixed constant.
Proof. For a given a A, the algorithm starts by computing a good normal n (see Lemma 4.5) and
constructing the VASS An = (Q, Tn ) of Theorem 4.6. Here, the set Tn is computed as follows.
Note that A can be seen as a directed multigraph where the nodes are the states and the edges
correspond to transitions. To every transition (q, u, q 0 ) we assign its weight −u · n. Note that the
multigraph does not contain any negative cycles (a negative cycle in the multigraph would induce a
cycle in A increasing the n-value; however, such a cycle cannot exist with a good normal n). To
decide whether a given transition (q, u, q 0 ) belongs to Tn , it suffices to find a path with the least
accumulated weight from q 0 to q (which can be done using, e.g., Bellman-Ford algorithm [64])
and check whether the accumulated weight is equal to u · n. Hence, Tn is computable in time
polynomial in the size of A (for a given good normal n).
Then, the algorithm proceeds by constructing the SCC C1 , . . . , Ck of An . If k = 0, then L(n) ∈
Θ(n) (see Theorem 4.6 (1)). If k = 1 and AnC1 = A, then A is non-terminating (this is a
consequence of Theorem 4.6 (2); if AnC1 was terminating with termination complexity Θ(f1 (n)),
then by Theorem 4.6 (2), the termination complexity of A = AnC1 is Θ(n · f1 (n)), which is
impossible). Otherwise, the algorithm proceeds by analyzing AnC1 , . . . , AnCk recursively. If some
of them is non-terminating, then A is also non-terminating. Otherwise, the termination complexity
of A is derived from the termination complexity of AnC1 , . . . , AnCk as in Theorem 4.6 (2). Clearly,
we obtain L(n) ∈ Θ(nk ) for some k ∈ {1, . . . , d}. It is easy to verify that the total the number of
recursive calls is polynomial in the size of A.
4.2
VASS satisfying condition (D)
Condition (D) is not sufficiently strong to guarantee polynomial termination time for terminating
VASS. In fact, as d increases, the termination complexity can grow very fast. Even for d = 3, one
can easily construct a terminating VASS satisfying (C) such that L(n) ∈ Ω(2n ).
Example 4.9. Consider the strongly connected 3-dimensional VASS A in Fig. 3. Let n ∈ N
be arbitrary. We construct a zero-avoiding computation α(n) starting in q1~n whose length is
exponential in n. For better readability, denote by x, y, and z the variables representing the first,
second, and third counter, respectively.
17
(1,0,0)
q1
q2
(1,-1,0)
(0,0,-1)
(0,0,0)
(-1,1,0)
q3
q4
(0,1,0)
Figure 3: A 3-dimensional VASS satisfying condition (D) which has an exponential termination
complexity.
The construction consist of iterating several phases.
In Phase (a) we iterate the short
cycle q1 , (1, 0, 0), q2 , (1, −1, 0), q1 as long as y
2.
≥
Then we perform the path
q1 , (1, 0, 0), q2 , (0, 0, 0), q4 to q4 . From there we continue with Phase (b), where we iterate
the short cycle q4 , (0, 1, 0), q3 , (−1, 1, 0), q4 as long as x ≥ 2. After this we perform the path
q4 , (0, 1, 0), q3 , (0, 0, −1), q1 to q1 . There we again switch to Phase (a), repeating the process until
one of the counters hits zero.
One can straightforwardly check that the total effect of performing Phase (a) once is setting y to 1
while setting x to xa +2ya , where xa , ya are the values of x, y before the start of the phase. Similarly,
The total effect of performing Phase (b) once is setting x to 1 while setting y to yb + 2xb , where
xb , yb are the values of x, y before the start of the phase. Hence, the total effect of consecutively
performing Phases (a) and (b) once can be bounded from below as follows: setting x to 1 and
multiplying y by 4. Hence, the total effect of performing N consecutive iterations of Phases (a) and
(b) is setting x to 1, multiplying y by 4N and decreasing z by N . Since z decreases exactly during
the witch from Phase (b) to Phase (a), we can perform exactly n consecutive iterations of (a) and
(b). But increasing y from n to 4n requires at least 4n − n steps in VASS, hence the termination
complexity of A is at least exponential. The matching asymptotic upper bound is easy to get.
The key idea of the previous example can be used as building block for showing that higherdimensional terminating VASS satisfying (D) can have even larger termination complexity than
exponential. Already in dimension 4, the complexity can be non-elementary.
Example 4.10. Consider the 4-dimensional strongly connected VASS in Fig. 4. As before, we
denote by x, y, z, w the individual counters.
For n ∈ N we construct a zero-avoiding computation α(n) started in q1~n whose length in nonelementary. The construction again proceeds by switching between various phases and the phases
we consider are the following: in Phase (a) we iterate cycle q1 , (−1, 1, 0, 0), q2 , (0, 1, 0, 0), q1 from
q1 as long as x ≥ 2. The effect of a single execution of (a) is setting x to 1 and y to ya + 2xa
(as before vp denotes the value of counter v at the start of phase (p)). In Phase (b) we iterate
the self-loop on q3 as long as y ≥ 2, the effect of the phase is setting y to 1 and x to xb + yb .
18
(0,0,-1,0)
(1,-1,0,0)
(-1,1,0,0)
q3
q1
q2
(0,0,0,0)
(0,1,0,0)
(0,0,0,-1)
(0,0,0,0)
q4
(1,-1,1,0)
Figure 4: A 4-dimensional VASS satisfying condition (D) which has a non-elementary termination
complexity.
Phase (c) consists of iterating the self-loop on q4 as long as y ≥ 2 and the effect is setting y to 1
and x and z to xc + yc and zc + yc , respectively. Switching from (b) to (a) or (c) decreases z by 1,
while switching from (c) to (a) or (b) decreases w by 1. Now the construction of α(n) proceeds as
follows: we switch between Phases (a) and (b) as long as z ≥ 2, after which we perform Phase
(a) once more. We call this a Phase (d) and the total effect of (d) is setting x and z to 1, and y
to a number at least 4zd · xd . After Phase (d) we go to q4 and execute Phase (c), after which we
go to q1 and start (d) again, repeating the process until a configuration with a zero counter is hit.
The total effect of a single consecutive execution of (d) and (c) is setting y to 1 and x and z to
a number at least 2zd · xd . Since w is only decremented when switching from (d) to (c), we can
repeat this consecutive execution at least n times. An easy induction shows that after i repeats of
the consecutive executions of (d) and (c) the value of x is at least
n
22
··· 2
ξn := n · 2 | {z } .
n times
Hence, the length of α(n) is at least ξn , i.e. non-elementary.
Figure 3 also provides an example showing that lexicographic ranking functions are not sound
for polynomial bounds on termination complexity. We first define the notion of lexicographic
ranking function for VASS: we specialize the standard definition of a lexicographic ranking
functions for affine automata [4] (a generalization of VASS which models general linear arithmetic
programs). Formally, an m-dimensional lexicographic map for a VASS A = (Q, T ) is a collection
{fqj | q ∈ Q, 1 ≤ j ≤ m} of linear functions of counter values, one function per state and
1 ≤ j ≤ m (we allow m to be different from the dimension d of A). A lexicographic map
{fqj | q ∈ Q, 1 ≤ j ≤ m} is a lexicographic -ranking function for A if each fqj is bounded
from below on N and for each transition (q, u, q 0 ) of A there exists 1 ≤ j ≤ m such that
0
0
f j0 (u) ≤ fqj (~0) − and for all 1 ≤ j 0 < j it holds f j0 (u) ≤ fqj (~0). A standard argument shows
q
q
that if A has a lexicographic -ranking function for, then it is terminating. However, lexicographic
ranking functions are not sound for polynomial complexity bounds.
Example 4.11. Consider the VASS A in Figure 3. Then, denoting the first, second, and third counter
as x, y, z, respectively, there is the following 3-dimensional lexicographic 12 -ranking function for
19
A (we denote fq = (fq1 , . . . , fqm )): fq1 = (z, y, x), fq2 = (z, y − 12 , x), fq4 = (z − 12 , x, y),
fq3 = (z − 12 , x − 12 , y). But as shown in Example 4.9, the VASS has exponential termination
complexity.
5
5.1
Technical Proofs
Proof of Theorem 3.5
We describe a polynomial time-algorithm for deciding whether a given VASS has linear termination
complexity. Recall from the proof sketch that it suffices to solve an equivalent problem whether
there is an open half-space Hn of Rd such that n ≥ ~0 and Inc ⊆ Hn .
Let us formalize our intuition presented in the proof sketch. We need to introduce some additional
notation: An infinite path π is an infinite sequence of the form p0 , u1 , p1 , u2 , p2 , . . . where for
each n ≥ 1 the finite subsequence p0 , u1 , p1 , u2 , p2 , . . . , un , pn is a finite path. We denote by π↓n
the finite prefix p0 , u1 , p1 , u2 , p2 , . . . , un , pn of π. Given an infinite path π, we define the mean
change of n-value as
eff (π↓n ) · n
.
n→∞
n
Consider the following linear program L obtained from [60], Section 8.8, by substituting the reward
MC n (π) = lim inf
r(s, a) with u · n where u is an effect of a transition:
Minimize g with respect to the following constraints:
For all (q, u, q 0 ) ∈ T
g + h(q) − h(q 0 ) ≥ u · n
and
n ≥ ~0.
Here, the variables are g, all h(q), q ∈ Q, and all components of n. By applying the results of [60],
for every optimal solution g, h, n we have that
g = sup MC n (π).
π
Moreover, there is at least one feasible solution.
We prove that there is n ≥ ~0 such that the open half-space Hn contains Inc iff an optimal solution
g, h, n of the above program satisfies g < 0.
Consider an optimal solution g, h, n of the above program. Assume that g < 0. We show that
Inc ⊆ Hn . For the sake of contradiction, assume that there is a short cycle π such that eff (π)·n ≥ 0.
Following the cycle π ad infinitum determines an infinite path π with MC n (π) ≥ 0. However, this
contradicts the fact that 0 > g = supπ MC n (π).
20
Now assume there is n ≥ ~0 such that Inc ⊆ Hn . Let π be an infinite path. Let us fix n ≥ 1
and consider Decomp(π↓n ), the decomposition of π↓n into short cycles. Let Rest(π↓n ) be the
remaining path obtained after removing all short cycles of Decomp(π↓n ) from π↓n . Note that the
~ · n.
length of Rest(π↓n ) is at most |Q|, and hence eff (Rest(π↓n )) · n ≤ |Q|
Now let m be the length (i.e., the number of elements) of the list Decomp(π↓n ). Note that
m ≥ n/|Q| − 1. Consider ε > 0 such that for all short cycles α we have that eff (α) · n ≤ −ε.
Then
eff (π↓n )
≤
~
m·(−ε)+|Q|·n
≤
~
(n/|Q|−1)·(−ε)+|Q|·n
=
~
(n·(−ε)/|Q|)+(|Q|·n+ε)
and thus
eff (π↓n )
n
≤
~ · n + ε)
(n · (−ε)/|Q|) + (|Q|
n
=
~ · n + ε)
−ε (|Q|
+
.
|Q|
n
~ · n + ε)/n = 0, we obtain that MC n (π) ≤ (−ε)/|Q|. As π was chosen
Since limn→∞ (|Q|
arbitrarily, we have that
−ε
< 0.
|Q|
π
Hence, there is a solution g, h, n of the above linear program with g < 0.
sup MC n (π)
≤
In order to decide whether there is an open half-space Hn of Rd such that n ≥ ~0 and Inc ⊆ Hn , it
suffices to compute an optimal solution g, h, n of the above linear program, which can be done in
polynomial time (see, e.g., [51]), and check whether g < 0.
Now we get back to weighted linear ranking functions. Note that each solution g, h, n of the linear
program L in which g < 0 yields a weighted linear ranking function (c, {hq | q ∈ Q}) by putting
c := n and hq := h(q) for each q. Conversely, each weighted linear ranking function yields a
solution of L where g < 0, (we need to put g := −, where is from the definition of a weighted
lin. ranking function). Hence, a VASS A has linear termination complexity if and only if it has a
weighted linear ranking function and this can be decided in polynomial time in size of A.
5.2
Proof of Theorem 4.2
If condition (A) or (B) holds, there is no n ≥ ~0 such that Inc ⊆ Ĥn . We show that then there exists
u ∈ cone(Inc) such that u > ~0. Suppose there is no such u. Let B be the set of all v > ~0. Since
cone(Inc) and B are convex and disjoint, there is a separating hyperplane with normal n ≥ ~0 for
cone(Inc) and B. Since cone(Inc) ⊆ Ĥn , we have a contradiction.
P
So, let u > ~0 such that u = ki=1 ai · vi , where ai ∈ Q+ and vi ∈ Inc for all 1 ≤ i ≤ k. Hence,
P
there also exist b1 , . . . , bk ∈ N+ such that w = k bi · vi > ~0. Let us fix a cycle π in A visiting
i=1
all control states (here we need that A is strongly connected). Clearly, there exists c ∈ N such that
eff (π) + c · w > 0. Let % be a cycle obtained from π by inserting c · bi copies of a short cycle γi ,
where eff (γi ) = vi . Then, eff (%) > ~0, and hence there exists an infinite computation initiated in
p~n for a sufficiently large n ∈ N (the control state p can be chosen arbitrarily).
21
5.3
Proof of Lemma 4.5
Due to condition (C), there exists at least one positive normal. Hence, we can fix a positive
n ∈ Normals(A) such that the set {u ∈ Inc | u · n = 0} is minimal. We show that for every
v ∈ cone(Inc) we have that −v ∈ cone(Inc) iff v · n = 0, i.e., n is a good normal. The “⇒”
direction immediate—if v, −v ∈ cone(Inc), then v · n ≤ 0 and −v · n ≤ 0, which implies
v · n = 0. For the other direction, suppose there exists v ∈ cone(Inc) such that v · n = 0 and
−v 6∈ cone(Inc). Then there also exists u ∈ Inc such that u · n = 0 and −u 6∈ cone(Inc). For
the rest of this proof, we fix such u. By Farkas’ lemma, there exists a separating hyperplane for
cone(Inc) and −u with normal vector n0 , i.e., −u · n0 > 0 and v · n0 ≤ 0 for every v ∈ cone(Inc).
Let us fix a sufficiently small ε > 0 such that n + εn0 > ~0 and v · (n + εn0 ) < 0 for all v ∈ Inc
where v · n < 0. Clearly, n + εn0 is a positive normal. Further, for all v ∈ cone(Inc) such that
v · n < 0 we have that v · (n + εn0 ) < 0. Since u · (n + εn0 ) < 0, we obtain a contradiction with
the minimality of n.
To compute a good normal, first observe that the condition of Definition 4.3 can be safely relaxed
just to the vectors of Inc, i.e., if n ∈ Normals(A) such that n > ~0 and −v ∈ cone(Inc) iff
v · n = 0 for every v ∈ Inc, then n is a good normal. To see this, fix some n with this property,
P
and let u = ki=1 ai · vi , where ai ∈ R+ and vi ∈ Inc for all 1 ≤ i ≤ k. We need to show that
P 0
−u ∈ cone(Inc) iff u · n = 0. If −u ∈ cone(Inc), then −u = ki=1 a0j · vi0 where a0i ∈ R+ and
vi0 ∈ Inc for all 1 ≤ i ≤ k 0 . Hence,
~0 = u + (−u) =
k
X
0
ai · vi +
i=1
k
X
a0i · vi0 .
i=1
Hence,
0 = (u + (−u)) · n =
k
X
0
a i · vi · n +
i=1
k
X
a0i · vi0 · n.
i=1
· n ≤ 0 for all 1 ≤ i ≤ k and all 1 ≤ i0 ≤ k 0 , we obtain vi · n = 0 for all
P
1 ≤ i ≤ k, hence u · n = 0. On the other hand, if u · n = 0, then ki=1 ai · vi · n = 0. Since ai > 0
Since vi · n ≤ 0 and
vi0
and vi · n ≤ 0 for all 1 ≤ i ≤ k, we have that vi · n = 0 for all 1 ≤ i ≤ k. Hence −vi ∈ cone(Inc)
P
for every 1 ≤ i ≤ k (by our assumption), and −u = ki=1 ai · (−vi ) ∈ cone(Inc).
Using the above observation, we can compute a good normal using linear programming as follows:
First, compute the set I = {v ∈ Inc | −v ∈ cone(Inc)}. Note that I can be computed easily by
checking feasibility of the following linear constraints:
−v =
X
au · u
and
au ≥ 0.
u∈Inc
Here, the variables are au . A good normal can be computed using the following linear program:
22
Maximize ε with respect to the following constraints:
u · n = 0 for all u ∈ I
v · n ≤ −ε for all v ∈ Inc r I
n ≥ ~ε.
Here, the variables are ε and all components of n.
Note that there is a good normal iff there is an optimal solution with ε > 0. Moreover, every
optimal solution ε, n with ε > 0 gives a good normal n.
5.4
Proof of Theorem 4.6
Now start by formulating an auxiliary technical lemma which is needed in the proof of Theorem 4.6.
Lemma 5.1. Let A be a VASS satisfying (C), and let n be a good normal. Then there is a constant
κ ∈ R+ such that for every w ∈ cone(Inc), where w · n = 0 and norm(w) = 1, there exist
P
k ∈ N, a1 , . . . , ak ∈ R+ , and v1 , . . . , vk ∈ Inc such that w = kj=1 aj · vj , vj · n = 0 for all
P 0
1 ≤ j ≤ k, and for all k 0 ≤ k, the absolute values of all components of the vector kj=1 aj · vj are
bounded by κ.
Proof. Let w =
Clearly, w · n =
Pk
Pj=1
k
aj · vj where aj ∈ R+ , vj ∈ Inc for all 1 ≤ j ≤ k, and k is minimal.
j=1 aj
· (vj · n) = 0, which implies vj · n = 0 for all 1 ≤ j ≤ k (recall
that vj · n ≤ 0 because n ∈ Normals(A)). First, we show that for every j ≤ k, the vector
−vj does not belong to cone({v1 , . . . , vj−1 , vj+1 , . . . , vk }). Assume the converse, i.e., −v1 ∈
P
cone({v2 , . . . , vk }). Then −v1 = kj=2 bj · vj , where bj ∈ R+ for all 2 ≤ j ≤ k. Further,
w
=
(a1 − c) · v1 + (a2 − cb2 ) · v2 + · · · + (ak − cbk )vk
for every c > 0. Clearly, there exists c > 0 such that at least one of the coefficients (a1 − c),
(a2 − cb2 ), . . . , (ak − cbk ) is zero and the other remain positive, which contradicts the minimality of k. Since {v1 , . . . , vk } ⊆ Ĥn , there must exist n0 > ~0 such that {v1 , . . . , vk } ⊆ Hn0
(otherwise, we can use the same argument as in the proof of Case (a) of Lemma 3.2 to show
that −vj ∈ cone({v1 , . . . , vj−1 , vj+1 , . . . , vk }) for some 1 ≤ j ≤ k). Since vj · n0 < 0 for
all 1 ≤ j ≤ k, each vj moves in the direction of −n by some fixed positive distance. Since
norm(w) = 1, there is a bound δv1 ,...,vk ∈ R+ such that aj ≤ δv1 ,...,vk for all 1 ≤ j ≤ k, because
no aj · vj can go in the direction of −n by more than a unit distance.
The above claim applies to every w ∈ cone(Inc) where w · n = 0. Since Inc is finite, there
are only finitely many candidates for the set of vectors {v1 , . . . , vk } used to express w, and
hence there exists a fixed upper bound δ ∈ R+ for all δv1 ,...,vk . This means that, for every
w ∈ cone(Inc) where w · n = 0, there exist k ∈ N, a1 , . . . , ak ∈ R+ , and v1 , . . . , vk ∈ Inc such
P
that w = kj=1 aj · vj , vj · n = 0, and aj ≤ δ for all 1 ≤ j ≤ k. This immediately implies the
existence of κ.
23
Now can formalize the proof of Theorem 4.6.
All cycles of AnCi are n-neutral. First, realize that for every cycle η of A (not necessarily short)
P
we have that eff (η) · n = γ∈Decomp(η) eff (γ) · n ≤ 0. Now let β = p0 , u1 , p1 , u2 , p2 , . . . , un , pk
be a cycle of AnCi (not necessarily short). Then each transition (pj , uj+1 , pj+1 ) of β is contained in
some n-neutral short cycle γi of A. Let %i be the (unique) path from pj+1 to pj determined by γj ,
P
and let % = %k−1 · · · %0 . Then eff (β)+eff (%) = k−1
j=0 eff (γj ). Hence, eff (β)·n+eff (%)·n =
Pk−1
j=0 eff (γj ) · n = 0. Thus, we obtain eff (β) · n = − eff (%) · n. Since both β and % are cycles
of A, we have that eff (β) · n ≤ 0 and eff (%) · n ≤ 0, which implies eff (β) · n = 0.
Constructing the paths of length Θ(fi (n)). Since the termination complexity of AnCi is Θ(fi (n)),
there is b ∈ R+ such that for all sufficiently large n ∈ N there exist a configuration pn~n and a
zero-avoiding computation βn of length at least b · fi (n) initiated in pn~n. Since πβn inevitably
contains a cycle whose length is at least b0 · fi (n) (for some fixed b0 ∈ R+ independent of βn ), we
can safely assume that πβn is actually a cycle, which implies eff (πβn ) ∈ cone(Inc).
Constructing the compensating path.
Since πβn is n-neutral and eff (πβn ) ∈ cone(Inc), we
have that − eff (πβn ) ∈ cone(Inc). This is where we use the defining property of a good normal.
P
+
Since − eff (πβn ) = m
j=1 aj · vj , where m ∈ N, aj ∈ Q , and vj ∈ Inc for all 1 ≤ j ≤ m, a
straightforward idea is to define the compensating path by “concatenating” baj c copies of γj , where
eff (γj ) = vj , for all 1 ≤ j ≤ m. This would produce the desired effect on the counters, but there
is no bound on the counter decrease in intermediate configurations visited when executing this path.
To overcome this problem, we construct the compensating path for πβn more carefully. Let w be
the normalized eff (πβn ), i.e., w has the same direction as eff (πβn ) but its norm is equal to 1. By
P
+
Lemma 5.1, −w is expressible as −w = m
j=1 aj · vj , where m ∈ N, aj ∈ Q , and vj ∈ Inc, so
that vj · n = 0 for all 1 ≤ j ≤ m, and for all m0 ≤ m, the absolute values of all components of the
P 0
vector m
j=1 aj · vj are bounded by κ, where κ is a constant independent of w. Let us fix some
cycle η of AnCi visiting all of its states (recall that AnCi is strongly connected). The compensating
path for πβn is obtained from η by inserting bnorm(eff (πβn )) · aj c copies of a short cycle with
effect vj , for every 1 ≤ j ≤ m. Observe that the difference between the effect of this compensating
path and − eff (πβn ) is bounded by a constant vector independent of n. Further, when executing
the compensating path, the counters are never decreased by more that κ · norm(eff (πβn )).
Constructing a zero-avoiding computation αn of length Ω(n · fi (n)).
Now we are ready to put
the above ingredients together, which still requires some effort. Let us fix a sufficiently large n ∈ N
and a configuration pv where ||pv|| = n and p is a control state of AnCi . Let q be the first state of
πβn . If we started αn in pv by executing a finite path which changes the control state from p to the
first control state of πβn (which takes at most |Q| transitions) and continued by executing πβn , the
counters could potentially reach values arbitrarily close to zero (it might even happen that πβn is
not executable). Instead, we fix a suitable n0 ≤ n satisfying n − n0 ≥ κ · norm(eff (πβn0 )) + |Q|.
√
√
Since norm(eff (πβn0 )) ≤ d · n0 , we can safely put n0 = (n − |Q|)/(1 + κ d). Now, we can
initiate αn by a short finite path which changes the control state from p to the first control state
24
√
of πβn0 , and continue by executing πβn0 . Note that (1 + κ d) is a constant, so decreasing n to
n0 has no influence in the asymptotic length of the constructed computation. Then, we can safely
execute the compensating path for πβn0 , and thus reach a configuration qu where we continue in
the same way as in pv, i.e., execute another finite path of length Θ(fi (n)) and its corresponding
compensating path. Since the v − u is bounded by a constant vector, this can be repeated Ω(n)
times before reaching a configuration where some counter value is not sufficiently large to perform
another “round”. Hence, the length of the resulting αn is Ω(n · fi (n)).
6
Related Work
In this section we discuss the related work.
Resource analysis. Our work is most closely related to automatic amortized analysis [38, 39, 40, 41,
42, 46, 45, 36, 31], as well as the SPEED project [34, 35, 33]. All these works focus on worst-case
asymptotic bounds for programs, and present sound methods but not complete methods for upper
bounds, i.e., even though the asymptotic bound is linear or quadratic, the approaches may still fail
to provide any upper bound. However, all these works consider general programs rather than the
model of VASS. In contrast, we consider VASS and present sound and complete method to derive
tight (upper and matching lower) polynomial complexity bounds.
Recurrence relations. Other approaches for bounds analysis involve recurrence relations, such
as [32, 29, 1, 2, 3]. Even for relatively simple programs the recurrence are quite complex, and
cannot be obtained automatically. In contrast, we present a polynomial-time approach for optimal
asymptotic bounds for VASS.
Ranking functions and extensions. Ranking functions for intraprocedural analysis have been widely
studied [7, 9, 20, 59, 67, 21, 70, 63]. Most works have focussed on linear or polynomial ranking
functions [20, 59, 67, 21, 70, 63], as well as non-polynomial bounds [14]. Again, these approaches
are sound, but not complete even to derive upper bounds for VASS. The notion of ranking functions
have been also extended to ranking supermartingales [10, 28, 15, 13, 16] for expected termination
time of probabilistic programs, but such approaches do not present polynomial asymptotic bounds.
Results on VASS. The model of VASS [50] or equivalently Petri nets are a fundamental model
for parallel programs [25, 50] as well as parameterized systems [6]. The termination problems
(counter-termination, control-state termination) as well as the related problems of boundedness
and coverability have been a rich source of theoretical problems that have been widely studied [56,
61, 23, 24, 8]. The complexity of the termination problem with fixed initial configuration is
EXPSPACE-complete [56, 71, 5]. Recent work such as [66, 6] shows how VASS and subclass of
VASS (such as lossy VASS) provide a natural model for abstraction and analysis of programs as well
as parametrized systems. The work of [66] also considers lexicographic ranking functions to obtain
sound asymptotic upper bounds for lossy VASS. However, this approach is not complete, and also
25
do not consider tight complexity bounds (but only upper bounds). Besides the termination problem,
the more general reachability problem where given a VASS, an initial and a final configuration,
whether there exists a path between them has also been studied [57, 53, 55]. The reachability
problem is decidable [57, 53, 55], and EXPSPACE-hard [56], and the current best-known upper
bound is cubic Ackermannian [54], a complexity class belonging to the third level of a fast-growing
complexity hierarchy introduced in [62].
Other related approaches are sized types [17, 43, 44], and polynomial resource bounds [65]. Again
none of these approaches are complete for VASS nor they can yield tight asymptotic complexity
bounds.
Hyperplane-separation technique and existence of infinite computation. The problem of existence
of infinite computations in VASS has been studied in the literature. Polynomial-time algorithms
have been presented in [11, 68] using results of [52]. In the more general context of games
played on VASS, even deciding the existence of infinite computation is coNP-complete [11,
68], and various algorithmic approaches based on hyperplane-separation technique have been
studied [12, 47, 18]. In this work we also consider normals of effects of cycles in VASS, which is
related to hyperplane-separation technique. However all previous works consider hyperplane-based
techniques to determine the existence of infinite computations on games played on VASS, and do
not consider asymptotic time of termination. In contrast, we present the first approach to show
that hyperplane-based techniques can be used to derive tight asymptotic complexity bounds on
termination time for VASS.
7
Conclusion
In this paper, we studied the problem of obtaining precise polynomial asymptotic bounds for VASS.
We obtained a full end efficient characterization of all VASS with linear termination complexity.
Then we considered polynomial termination for strongly connected VASS, dividing them into four
disjoint classes (A)–(D). For the first two classes, we proved that the VASS are non-terminating.
For VASS in (C), we obtained a full and effective characterization of termination complexity. For
the last class (D), we have shown that the termination complexity can be exponential even for
dimension three. The results are applicable also to general (i.e., non-strongly connected VASS), by
analyzing the individual SCCs. Some extra effort is needed in (C), because here a possible increase
in the size of configurations accumulated in a given SCC before moving into another SCC must be
taken into account. To keep our proofs reasonably simple, we considered just strongly connected
VASS.
Our result gives rise to a number of interesting directions for future work. First, whether our precise
complexity analysis or the complete method can be extended to other models in program analysis
(such as affine programs with loops) is an interesting theoretical direction to pursue. Second, in the
practical direction, using our result for developing a scalable tool for sound and complete analysis
26
of asymptotic bounds for VASS and their applications in program analysis is also an interesting
subject for future work.
References
[1] Elvira Albert, Puri Arenas, Samir Genaim, Miguel Gómez-Zamalloa, German Puebla, Diana V.
Ramírez-Deantes, Guillermo Román-Díez, and Damiano Zanardini. Termination and cost
analysis with COSTA and its user interfaces. ENTCS, 258(1):109–121, 2009.
[2] Elvira Albert, Puri Arenas, Samir Genaim, and Germán Puebla. Automatic inference of
upper bounds for recurrence relations in cost analysis. In SAS, volume 5079 of LNCS, pages
221–237. Springer, 2008.
[3] Elvira Albert, Puri Arenas, Samir Genaim, Germán Puebla, and Damiano Zanardini. Cost
analysis of java bytecode. In ESOP, volume 4421 of LNCS, pages 157–172. Springer, 2007.
[4] Christophe Alias, Alain Darte, Paul Feautrier, and Laure Gonnord. Multi-dimensional
rankings, program termination, and complexity bounds of flowchart programs. In SAS, pages
117–133. Springer-Verlag, 2010.
[5] Mohamed Faouzi Atig and Peter Habermehl. On yen’s path logic for petri nets. IJFCS,
22(04):783–799, 2011.
[6] Roderick Bloem, Swen Jacobs, Ayrat Khalimov, Igor Konnov, Sasha Rubin, Helmut Veith,
and Josef Widder. Decidability in parameterized verification. SIGACT News, 47(2):53–64,
2016.
[7] Olivier Bournez and Florent Garnier. Proving positive almost-sure termination. In RTA, pages
323–337, 2005.
[8] Laura Bozzelli and Pierre Ganty. Complexity Analysis of the Backward Coverability Algorithm
for VASS. In RP, pages 96–109. Springer, 2011.
[9] Aaron R. Bradley, Zohar Manna, and Henny B. Sipma. Linear ranking with reachability. In
CAV, volume 3576 of LNCS, pages 491–504. Springer, 2005.
[10] Aleksandar Chakarov and Sriram Sankaranarayanan. Probabilistic program analysis with
martingales. In CAV, volume 8044 of LNCS, pages 511–526. Springer, 2013.
[11] Krishnendu Chatterjee, Laurent Doyen, Thomas A. Henzinger, and Jean-Francois Raskin.
Generalized mean-payoff and energy games. In FSTTCS, pages 505–516, 2010.
[12] Krishnendu Chatterjee and Yaron Velner. Hyperplane separation technique for multidimensional mean-payoff games. In CONCUR, pages 500–515. LNCS 8052, Springer, 2013.
27
[13] Krishnendu Chatterjee, Hongfei Fu, and Amir Kafshdar Goharshady. Termination analysis of
probabilistic programs through positivstellensatz’s. In CAV (I), pages 3–22, 2016.
[14] Krishnendu Chatterjee, Hongfei Fu, and Amir Kafshdar Goharshady. Non-polynomial worstcase analysis of recursive programs. In CAV, pages 41–63, 2017.
[15] Krishnendu Chatterjee, Hongfei Fu, Petr Novotný, and Rouzbeh Hasheminezhad. Algorithmic
analysis of qualitative and quantitative termination problems for affine probabilistic programs.
In POPL, pages 327–342. ACM, 2016.
[16] Krishnendu Chatterjee, Petr Novotný, and Dorde Zikelic. Stochastic invariants for probabilistic
termination. In POPL, pages 145–160, 2017.
[17] Wei-Ngan Chin and Siau-Cheng Khoo. Calculating sized types. In Higher-Order and Symbolic
Computation, 14(2-3):261–300, 2001.
[18] Thomas Colcombet, Marcin Jurdzinski, Ranko Lazic, and Sylvain Schmitz. Perfect half space
games. In CoRR, 2017.
[19] Michael Colón, Sriram Sankaranarayanan, and Henny Sipma. Linear invariant generation
using non-linear constraint solving. In CAV, volume 2725 of LNCS, pages 420–432. Springer,
2003.
[20] Michael Colón and Henny Sipma. Synthesis of linear ranking functions. In TACAS, volume
2031 of LNCS, pages 67–81. Springer, 2001.
[21] Patrick Cousot. Proving program invariance and termination by parametric abstraction,
Lagrangian relaxation and semidefinite programming. In VMCAI, volume 3385 of LNCS,
pages 1–24. Springer, 2005.
[22] Emanuele D’Osualdo, Jonathan Kochems, and C. H. Luke Ong. Automatic Verification of
Erlang-Style Concurrency. In SAS, pages 454–476. Springer 2013.
[23] Javier Esparza. Decidability and complexity of petri net problems—an introduction. Petri
nets, pages 374–428, 1996.
[24] Javier Esparza, Ruslán Ledesma-Garza, Rupak Majumdar, Philipp Meyer, and Filip Niksic.
An SMT-Based Approach to Coverability Analysis. In CAV, pages 603–619, Springer, 2014.
[25] Javier Esparza and Mogens Nielsen. Decidability issues for petri nets - a survey. In Bulletin
of the EATCS, 52:245–262, 1994.
[26] Uli Fahrenberg, Line Juhl, Kim G. Larsen, Jiří Srba. Energy Games in Multiweighted
Automata. In Theoretical Aspects of Computing – ICTAC 2011, volume 6916 of LNCS, pages
95–115. Springer, 2011.
[27] Yu Feng, Ruben Martins, Yuepeng Wang, Isil Dillig, and Thomas W. Reps. Component-based
synthesis for complex apis. In POPL, pages 599–612, ACM, 2017.
28
[28] Luis María Ferrer Fioriti and Holger Hermanns. Probabilistic termination: Soundness,
completeness, and compositionality. In POPL, pages 489–501. ACM, 2015.
[29] Philippe Flajolet, Bruno Salvy, and Paul Zimmermann. Automatic average-case analysis of
algorithm. TCS, 79(1):37–109, 1991.
[30] Pierre Ganty and Rupak Majumdar. Algorithmic verification of asynchronous programs.
TOPLAS, 34(1):6:1–6:48, May 2012.
[31] Stéphane Gimenez and Georg Moser. The complexity of interaction. In POPL, pages 243–255.
ACM, 2016.
[32] Bernd Grobauer. Cost recurrences for DML programs. In ICFP, pages 253–264. ACM, 2001.
[33] Bhargav S. Gulavani and Sumit Gulwani. A numerical abstract domain based on expression
abstraction and max operator with application in timing analysis. In CAV, volume 5123 of
LNCS, pages 370–384. Springer, 2008.
[34] Sumit Gulwani. SPEED: symbolic complexity bound analysis. In CAV, volume 5643 of
LNCS, pages 51–62. Springer, 2009.
[35] Sumit Gulwani, Krishna K. Mehra, and Trishul M. Chilimbi. SPEED: precise and efficient
static estimation of program computational complexity. In POPL, pages 127–139. ACM,
2009.
[36] Jan Hoffmann, Klaus Aehlig, and Martin Hofmann. Multivariate amortized resource analysis.
TOPLAS, 34(3):14, 2012.
[37] Jan Hoffmann, Klaus Aehlig, and Martin Hofmann. Resource aware ML. In CAV, volume
7358 of LNCS, pages 781–786. Springer, 2012.
[38] Jan Hoffmann and Martin Hofmann. Amortized resource analysis with polymorphic recursion
and partial big-step operational semantics. In APLAS, volume 6461 of LNCS, pages 172–187.
Springer, 2010.
[39] Jan Hoffmann and Martin Hofmann. Amortized resource analysis with polynomial potential.
In ESOP, volume 6012 of LNCS, pages 287–306. Springer, 2010.
[40] Martin Hofmann and Steffen Jost. Static prediction of heap space usage for first-order
functional programs. In POPL, pages 185–197. ACM, 2003.
[41] Martin Hofmann and Steffen Jost. Type-based amortised heap-space analysis. In ESOP,
volume 3924 of LNCS, pages 22–37. Springer, 2006.
[42] Martin Hofmann and Dulma Rodriguez. Efficient type-checking for amortised heap-space
analysis. In CSL, volume 5771 of LNCS, pages 317–331. Springer, 2009.
[43] John Hughes and Lars Pareto. Recursion and dynamic data-structures in bounded space:
Towards embedded ML programming. In ICFP, pages 70–81. ACM, 1999.
29
[44] John Hughes, Lars Pareto, and Amr Sabry. Proving the correctness of reactive systems using
sized types. In POPL, pages 410–423. ACM Press, 1996.
[45] Steffen Jost, Kevin Hammond, Hans-Wolfgang Loidl, and Martin Hofmann. Static determination of quantitative resource usage for higher-order programs. In POPL, pages 223–236.
ACM, 2010.
[46] Steffen Jost, Hans-Wolfgang Loidl, Kevin Hammond, Norman Scaife, and Martin Hofmann.
"carbon credits" for resource-bounded computations using amortised analysis. In FM, volume
5850 of LNCS, pages 354–369. Springer, 2009.
[47] Marcin Jurdzinski, Ranko Lazic, and Sylvain Schmitz. Fixed-dimensional energy games are
in pseudo-polynomial time. In ICALP pages 260–272, 2015.
[48] Alexander Kaiser, Daniel Kroening, and Thomas Wahl. Dynamic Cutoff Detection in Parameterized Concurrent Programs. In CAV, volume 6174 of LNCS, pages 645–659. Springer,
2010.
[49] Alexander Kaiser, Daniel Kroening, and Thomas Wahl. Efficient Coverability Analysis by
Proof Minimization. IN CONCUR, pages 500–515. Springer, 2012.
[50] Richard M. Karp and Raymond E. Miller. Parallel program schemata. JCSS, 3(2):147–195,
1969.
[51] Leonid Genrikhovich Khachiyan. A polynomial algorithm in linear programming. Doklady
Akademii Nauk SSSR, 244:1093–1096, 1979.
[52] S. Rao Kosaraju and Gregory F. Sullivan. Detecting cycles in dynamic graphs in polynomial
time. In STOC, pages 398–406, 1988.
[53] S. Rao Kosaraju. Decidability of reachability in vector addition systems (preliminary version).
In STOC, pages 267–281. ACM, 1982.
[54] Jérôme Leroux and Sylvain Schmitz. Demystifying reachability in vector addition systems.
In LICS, pages 56–67. ACM/IEEE 2015.
[55] Jérôme Leroux. Vector addition system reachability problem: A short self-contained proof.
In POPL, pages 307–316. ACM, 2011.
[56] Richard J. Lipton. The reachability problem requires exponential space. Technical report 62,
1976.
[57] Ernst W. Mayr. An Algorithm for the General Petri Net Reachability Problem. In SIAM,
13:441–460, 1984.
[58] Ernst W. Mayr and Albert R. Meyer. The complexity of the finite conainment problem for
Petri nets. Journal of the ACM, pages 561–576, 1981.
30
[59] Andreas Podelski and Andrey Rybalchenko. A complete method for the synthesis of linear
ranking functions. In VMCAI, volume 2937 of LNCS, pages 239–251. Springer, 2004.
[60] Martin L. Puterman. Markov Decision Processes. Wiley, 1994.
[61] Charles Rackoff. The Covering and Boundedness Problems for Vector Addition Systems.
TCS, 6:223–231, 1978.
[62] Sylvain Schmitz. Complexity hierarchies beyond elementary. TOTC, 8(1):3:1–3:36, February
2016.
[63] Liyong Shen, Min Wu, Zhengfeng Yang, and Zhenbing Zeng. Generating exact nonlinear
ranking functions by symbolic-numeric hybrid method. J. Systems Science & Complexity,
26(2):291–301, 2013.
[64] Alfonso Shimbel. Applications of matrix algebra to communication nets. The bull. math.
biophysics, 13(3):165–178, Sep 1951.
[65] Olha Shkaravska, Ron van Kesteren, and Marko C. J. D. van Eekelen. Polynomial size
analysis of first-order functions. In TLCA, volume 4583 of LNCS, pages 351–365. Springer,
2007.
[66] Moritz Sinn, Florian Zuleger, and Helmut Veith. A simple and scalable static analysis for
bound analysis and amortized complexity analysis. In CAV, pages 745–761, 2014.
[67] Kirack Sohn and Allen Van Gelder. Termination detection in logic programs using argument
sizes. In PODS, pages 216–226. ACM Press, 1991.
[68] Yaron Velner, Krishnendu Chatterjee, Laurent Doyen, Thomas A. Henzinger, Alexander
Rabinovich, and Jean-Francois Raskin. The complexity of multi-mean-payoff and multienergy games. Inf. Comput., 241:177–196, 2015.
[69] Reinhard Wilhelm. The worst-case execution-time problem - overview of methods and survey
of tools. TECS, 7(3), 2008.
[70] Lu Yang, Chaochen Zhou, Naijun Zhan, and Bican Xia. Recent advances in program
verification through computer algebra. FCS, 4(1):1–16, 2010.
[71] Hsu-Chun Yen. A unified approach for deciding the existence of certain petri net paths.
Information and Computation, 96(1):119–137, 1992.
31
| 8cs.DS
|
The IQ of Neural Networks
Dokhyam Hoshen and Michael Werman
arXiv:1710.01692v1 [cs.LG] 29 Sep 2017
School of Computer Science and Engineering
The Hebrew University of Jerusalem
Jerusalem, Israel
Abstract
IQ tests are an accepted method for assessing human intelligence. The tests consist of several parts that must be solved
under a time constraint. Of all the tested abilities, pattern
recognition has been found to have the highest correlation
with general intelligence. This is primarily because pattern
recognition is the ability to find order in a noisy environment, a necessary skill for intelligent agents. In this paper,
we propose a convolutional neural network (CNN) model for
solving geometric pattern recognition problems. The CNN receives as input multiple ordered input images and outputs the
next image according to the pattern. Our CNN is able to solve
problems involving rotation, reflection, color, size and shape
patterns and score within the top 5% of human performance.
Evaluating machine learning methods against human performance has served as a useful benchmark for artificial intelligence. Deep learning methods have recently achieved
super-human level performance on several important tasks
such as face recognition (Taigman et al. ) and Large Vocabulary Continuous Speech Recognition (LVCSR) (Xiong et
al. ). Although the above tasks might have some correlation
with human intelligence, they are not testing intelligence directly. In this paper, we take a more direct route and evaluate
neural networks directly on tasks designed to test human intelligence.
Intelligence Quotient (IQ) tests are one of the most common methods of defining and testing human computation
and comprehension abilities (Stern 1914). IQ tests are only
estimates of intelligence (as opposed to actual measurements) and their effectiveness is in active debate (Flynn
1987; Siegel 1989). Despite the controversy, IQ tests are
used in practice to make critical decisions, such as school
and workplace placement, and assessment of mental disabilities. IQ scores have also been shown to correlate with
morbidity and mortality, socioeconomic background and
parental IQ.
IQ tests measure different skills: verbal intelligence,
mathematical abilities, spatial reasoning, classification
skills, logical reasoning and pattern recognition. In this paper, we focus on the pattern recognition section. In these
problems, the test-taker is presented with a series of shapes
with geometric transformations between them. As in most
Copyright c 2018, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
IQ tests, the problem is presented with multiple choices for
answer. One of these answers is the (most) correct one. A
limited number of possible transformations are used to produce the questions, thus it is possible to gain experience and
learn to solve them easily and quickly. Nevertheless, pattern
recognition problems have added benefits in comparison to
the other types of questions, as they rely less on a standard
educational background (as in the verbal or mathematical)
and therefore (arguably) are less biased towards upbringing
and formal education. Pattern recognition problems are also
related to real life problems and are therefore assumed to
have a high correlation with the actual intelligence.
In this paper, we solve IQ-test style pattern recognition
questions with a deep network. We train our network on
computer generated pattern recognition problems. The input
to each example consists of two images, where the second
image is a transformed version of the first image. The models infer the transformation and output the prediction of the
next image in two different forms (evaluated separately):
1. Multiple choice: The model is presented with a number of
possible images and choses the most probable one.
2. Regression: The model draws the most likely next image.
Simply by showing the model sequences of images we were
able to grasp geometric concepts such as shape and rotation.
We trained the network on series of images exhibiting the
following transformations (see Fig. 2 for examples):
• Rotation: Each shape is rotated by a constant angle with
respect to the previous shape.
• Size: Each shape is expanded or shrunk by a fixed scaling
factor with respect to the previous shape.
• Reflection: Each shape is transformed by a combination
of a rotation (as before) and a reflection.
• Number: Each image contains one more polygon than in
the previous one i.e if the image contains two triangles
the second and third images will contain three and four
triangles, respectively.
• Color: The model is presented with a series of objects in
different colors and must decide the color of the next object.
• Addition: The model is presented with a series of shapes
that have another shape added to them. The model must
Figure 1: Example of Ravens test style questions. Given
eight shapes the subject must identify the missing piece.
decide what a combination of these previous images looks
like.
We performed extensive experiments investigating the
performance of CNNs at IQ tests. We trained the networks
on a dataset containing all the transformation options. In this
experiment, the operations were also used simultaneously,
i.e there can be two shapes in the image: one rotating and
one expanding. In a different set of experiments, for each
operation, we trained a specialized network on a dataset containing only the specific operation, for analysis purposes.
This research brings us a step closer to comparing deep
learning intelligence to human intelligence. It is also important for evaluating the difficulty of IQ tasks. We can also use
this model in adapting tests to the level of the individual and
so improve the current methods of evaluation. Additionally,
probabilities give an indication to how confusing different
choices were, thus allowing a more accurate evaluation.
Related Work
Psychologists have tried to estimate the effectiveness of different questions in intelligence tests (Crawford et al. 2001).
A commonly used and well validated test for pattern recognition assessment is Ravens Progressive Matrices, the test
taker is typically presented with 60 non-verbal multiple
choice questions (as shown in Figure 1). Such tests are popular mainly due to their independence from language and
writing skills.
One of the advantages of visual pattern recognition problems in IQ tests is that they requires a minimal amount of
prior knowledge. Therefore, such tests can also be used to
quantify the intelligence of very young children who have
not yet learned basic mathematical and verbal concepts. This
measure is thus less polluted by environmental factors. Work
was done on the relationship between the scores that children obtained in these problems in infancy and their results
in later years, in IQ tests in general as well as specifically the
pattern recognition section (DiLalla et al. 1990). The score
was found to be a good predictor for later years, but adding
birth order and parental education resulted in improved predictions.
The subject of solving Raven’s Progressive Matrices
problems computationally was studied by (Kunda, McGreggor, and Goel 2009). The focus of this work was finding a
feature representation for solving Raven’s Progressive Ma-
trices with simple classifiers. While their research was the
first to address these problems as a computational problem,
they did not publish enough experimental results to validate
their method. Our approach is different and more general,
we use deep neural networks to automatically learn the representation jointly with the classifier, rather than handcrafting the feature representation.
Work was done by (Wang et al. 2015) on automatic solution of the verbal reasoning part of IQ tests. Machine learning methods using hand-crafted features were able to automatically solve verbal reasoning questions featuring synonyms and antonyms as well as word analogies. This line of
work is related to our research as it solves ”analogy” problems, where the subject needs to grasp the transformation
rules between words and generalize them. It deals linguistic
transformations rather than visual ones.
(Hoshen and Peleg 2016) analyzed the capability of
DNNs learning arithmetic operations. In this work, the network learned the concept of addition of numbers based on
end-to-end visual learning. This shows the possibility of
learning arithmetic transformations without prior knowledge
of basic concepts such as ”number” or ”addition”.
Much research has been done on Frame Prediction, predicting the next frame given one or more frames. Mathieu et
al. (Mathieu, Couprie, and LeCun 2015) presented a video
next frame prediction network. Due to uncertainty is the output of real video, a GAN loss function was used and shown
to improve over the standard euclidean loss. A CNN system was presented by Oh et al. for frame prediction in Atari
games (Oh et al. 2015). This is simpler than natural video
due to both having simpler images and more deterministic
transformations. Some work been done on object and action
prediction from video frames by (Vondrick, Pirsiavash, and
Torralba 2016), in which the agent learned to predict the FC7
features and used it to classify future actions. In this paper,
we deal with data that uses pre-specified transformations and
simple shapes and is, therefore, more amenable to detailed
analysis than previous research.
Our architecture uses Convolutional Neural Networks
(CNNs) (LeCun et al. 1990). For multiple choice type
problems, we use a classification architecture containing a
discriminator (see e.g. (Krizhevsky, Sutskever, and Hinton
2012)). For open questions, in order to generate an image we
use an autoencoder type architecture (see e.g. (Hinton and
Salakhutdinov 2006)). The hyper-parameters of our CNN architecture, are inspired by the architecture used in DCGAN
(Radford, Metz, and Chintala 2015). We optimized the network parameters using backpropogation (Rumelhart et al. )
(LeCun et al. 2012), with the ADAM (Kingma and Ba 2014)
gradient descent algorithm.
Creating the Dataset
In this paper, we evaluate two types of IQ test scenarios:
solving multiple choice pattern recognition problems and
producing the next image.
Multiple choice questions: This scenario is used in most
IQ tests. The model receives six input images: 2 input images and 4 options for the next image. At train time, it receives an index corresponding to the correct answer. The
Figure 2: Multiple choice question problem types. In this figure each example corresponds to a different problem type. For each
problem, there are 4 possible answers and only 1 is correct (shown in full only for (a), for space considerations). (a) Polygon
rotation: A polygon is rotated by a constant angle θ, between consecutive frames. The correct answer is: Option 1. (b) Squiggle
Rotation: Six random points are sampled and rotated by a constant angle, θ. (c) Size: A polygon is dilated by a constant scale
factor, µ. (d) Reflection: A shape is rotated by a constant angle θ and then reflected around the y-axis. (e) Counting: Each frame
shows one more polygon than in the previous one. (f) Color: The third frame shows one more polygon than the second frame in
the color of the second image (green).(g) Image Addition: The second image adds a line to the first image and the fourth image
adds that same line to the third image.
model outputs the predicted probability for each of the presented options. The most probable option is picked as the
model’s answer.
We form the questions in the following way: for each sample question, we randomly choose a transformation T from
the following: rotation, size, reflection, number, color and
addition. We then randomly produce a 64x64 image containing a shape S. The shape is randomly selected out of
triangles, squares, circles or random squiggles. We apply the
chosen transformation on the first image resulting in the second frame T S. We produce the correct answer, T T S, and
the 3 wrong answers (the other options) using other transformations.
wrong answers are created using different operations. See
Fig. 2(d).
• Number: Between 1 and 4 shapes are selected. For each
successive image we add one further exemplar of the
shape (this shape has the same number of vertices as the
previous shape but will most likely not be a replica of the
previous shape) i.e if the first frame contains a triangle, the
next image will contain two triangles. The wrong images
will show a different number of shapes than the correct
one or it will show the correct number of shapes but of
the wrong shape. See Fig. 2(e).
• Size: We randomly select a scale parameter µ ∈ [0.5..2].
The wrong answers are created by sampling a different
resizing ratio or a different operation. See Fig. 2(c).
• Color: Two colors and a shape are selected at random.
The first image is the shape with the first color and the
second with two shapes with the second color. The solution should be the three copies of the shape with the
second color. In the example in Fig. 2(f) the correct image has three green triangles. The wrong answers will include different colors, shapes and transformations. This is
in essence a combination of the ”number” questions with
color elements.
• Reflection: we randomly select an angle for rotation and
rotate the shape after reflecting it at each stage. The
• Addition: Three shapes are randomly selected.The first
image contains one of these shapes.The second contains
• Rotation: we randomly select an angle θ ∈ [0..2π] and
rotate the shape by θ. The wrong answers are created by
sampling different angles for rotation or a different operation. See Fig. 2(a)(b).
Figure 3: Model architectures: In this figure, we show the architectures of both models, the open questions and the multiple
choice questions. In blue we have the CNN that is used by both models, the encoder network. The inputs of the models are
different: the multiple answer questions require four extra images from which to choose the answer. Thus, the input for the
multiple-choice problems consists of 6 images while the input for the open questions consists of 2 images. After the code
generation, the models split: the autoencoder runs the decoder (green) and outputs an image. The CNN classifier (used for the
multiple-choice problems) has one more affine layer (orange) and outputs four probabilities, one for each option. Note that the
two models are trained separately and do not share features.
both the first and second shapes. The third contains just
the third image. The correct answer is the union of images two and three but not including the part of image
2 that was inherited from image 1. Formally, im4 =
im3∪im2\im1. This problem is a combination of image
addition and subtraction. See Fig. 2(g).
Open questions: In this scenario, the model is required to
predict the next frame. At training time we do not provide
the model with four options to choose from, as the model
generates the next frame on its own. Since it is difficult to
evaluate the correctness of the predicted frame for the ”number” transformation, we did not include it in the open questions dataset. Furthermore, the ”color” transformation in the
this scenario is defined differently. For each sample two colors and shapes are selected at random. The first image shows
the first shape in the first color, the second shows the second shape in the second color. The third image (the solution)
should show the first shape in the second color.
We automatically generated 100K images for training and
1000 images for testing, according to the rules described
above. The data was normalized to have a normal mean and
standard deviation at each pixel.
We will publicly release the dataset generation Python
code.
Model Architecture
For multiple choice problems we used a standard CNN and
for open questions, we used an autoencoder. Both models
were implemented in PyTorch.
Multiple choice questions: We used a CNN architecture
with 4 convolutional layers, each followed by batch normalization and ReLU activation layers. We found this (shallow)
depth to work the best. The input layer receives six images,
the first two frames of the sequence and the four options
that the model must choose the most likely answer from.
The output layer is a linear layer with Softmax activation,
it outputs the probability of each of the 4 optional images
to be the correct answer. The model architecture is sketched
in Fig. 3. We can additionally use the probabilities to infer
which questions were challenging for the model (e.g. by the
probability vector entropy).
Open questions: In this scenario, we generated the answer
image rather than simply selecting from a set of option images. This required a network architecture containing both
an encoder and a decoder (an autoencoder). The choice of
hyper-parameters used in our architecture borrows heavily
from the one used in DC-GAN (Radford, Metz, and Chintala
2015). The encoder consists of 5 convolutional layers, each
with 4x4 convolution kernels with a stride of 2 (no pooling
layers, this is implicitly done by the convolution operation’s
stride). Each convolution is followed by leakyReLU activation layers with a leakage coefficient of 0.2. The input of
the decoder is the first two RGB input frames (total size:
64x64x6). The output of the encoder is a code vector z of
dimension 32. The decoder receives the code z as input and
outputs a full image (size: 64x64x3). The decoder is comprised of 5 deconvolution layers (Transposed Convolution
layers) with 4x4 convolution kernels and stride 2, this effectively upsamples by a factor of 2 with each deconvolution
layer. Each deconvolution layer is followed by an ELU activation layer. See Figure 3 for a diagram of our architecture.
For both scenarios, the same optimization settings were
used. We optimized the model using Adam (Kingma and Ba
2014). The learning rate was annealed by a factor of 0.1 every 100 epochs. We trained the model using mini-batches of
size 64. The loss is calculated by the cross-entropy loss.
The number of training epochs required for convergence
was different for each problem as was their convergence
rate. The number of epochs to convergence ranged between
30 and 100 depending on the problem.
Transformation
Rotation Polygon
Rotation Squiggle
Size
Reflection
Color
Number
Addition
All-Transformation Network
Success rate
98.4%
94%
100%
100%
95.6%
91.4%
100%
92.8%
Table 1: Accuracies for multiple-choice questions: We can
see that the all-transformation network accuracy was lower
than most of the individual scores. This is most likely due
to additional errors in classification of the transformation, a
problem that does not exist in individual trainings. Out of all
the transformations, the the number transformation had the
highest error rate.
Results
In this paper, we set to investigate if neural networks are
able to solve geometric pattern recognition problem as used
in IQ tests. In this section, we evaluate the performance of
our CNN models on learning the transformations to the geometric shapes (rotation, size, reflection etc). For both models, the results convincingly show that the networks have
grasped the basic concepts.
We present quantitative evaluations of model performance.
Multiple-answer questions: As this is a classification
problem, a quantitative evaluation is simple for this scenario.
The accuracy rates of our model on the different tasks are
presented in Tab. 1. We can see that the overall score was
significantly smaller than most of the individual scores. This
is most likely due to error of classification of the transformation, a problem that does not exist in individual trainings.
Out of all the transformations, the one with the highest error
rate was the number transformation. E.g. one failure case
had two similar options, with squares (wrong answer) having two vertices that were illegibly close to one another that
looking very similar to triangles (correct answer).
In Fig. 4 we show an example of the questions where the
network classified the wrong option as the correct answer.
We added the probabilities that the model assigned to these
options as well since this is indicative of the network’s error
rate.
It visually apparent that in this failure case, the options
were very close to one another, and that the network’s output
probabilities were also very close for these options (however
the wrong option had a slightly higher probability). This in-
Figure 4: Wrong Prediction on a rotation problem. True answer: 3 Prediction: 2 Probabilities: Option 1 ≈ 0. Option 2
≈ 0.9. Option 3 ≈ 0.09. Option 4 ≈ 0. We can see that the
network was confused by the thin shape and had difficulty
deciding which direction the shape was rotating.
dicates that the failure does not signify the network’s failure to learn these transformations and the actual error rate is
smaller.
Open questions: We trained and tested the autoencoder
network on the following problems: rotation, size, color, and
reflection. We also tested the capabilities of the agent on
tasks where the transformation was a combination of two
of the basic transformations. Some questions included two
shapes, each simultaneously changing according to a different transformation. In Fig. 5 we show examples from the
different types of questions that were used for evaluation.
It is apparent that while the network has managed to grasp
the concepts of the transformations in all cases, there is some
blurring as the finer details in the image are lost in the encoding (downsampling) stage of the autoencoder. We measure
prediction accuracy by mean squared error (MSE) between
the ground truth and predicted image.
In Tab. 2 we present the average MSEs for each problem
type. We can see from the results that reflection and rotation tasks were the hardest to learn (as shown by the rela-
Figure 5: Open questions Scenario. We present output examples for the following tasks: (a) Rotation (b) Size (c) ”Squiggle”
Rotation (d) Color (e) Reflection (f) Combination of Rotation + Size
Transformation
Rotation Polygon
Rotation Squiggle
Size
Reflection
Color
Overall
Average MSE
4.1E-4
5.7E-4
2.4E-4
4.1E-4
2.1E-4
3.5E-4
Table 2: The results show that the reflection and rotation
transformations are more difficult to learn. Empirical tests
also show that these transformations are the most difficult
for humans to solve.
tively high error). A possible theoretical explanation for the
relative difficulty of learning the rotation transformation is
that more parameters need to be inferred to define the rotation matrix: the center point coordinates and the angle.
As our reflection problems also include a rotation, the same
number of parameters much be inferred. In comparison, for
color problems only a single parameter must be inferred, as
it leaves one image unchanged but multiplies its values by
an identical constant that multiplied the other image. Only
one parameter is also needed for the scale factor in the size
questions. Inferring more parameters can result in a more
challenging task. Furthermore, empirical research has found
that rotation questions are significantly more difficult for humans to solve in IQ tests (Jaarsveld et al. 2010).
Performance on Noisy Images
As pattern recognition is often required in noisy environments, we tested the model’s performance on noisy images.
The setup is identical to the noiseless dataset, but with IID
Gaussian noise of σ = 99 added to all input and output images.
The model performed well in the noisy setting. In the
multiple-choice scenario, the average model error rate was
87% compared to 91% in the noiseless case. In the open
question scenario, the result images were generated correctly without the added noise, as noise cannot be predicted.
Performance on Real Questions
One of the main objectives of this paper is to compare neural network performance on benchmarks used to evaluate
human intelligence. When constructing the dataset, we analyzed IQ tests used in practice in the aim of building a realistic dataset. The rules used for dataset construction are
similar as those used for creating human IQ tests.
We tested our model on a test set of 40 questions taken
from IQ tests used by the National Institute for Testing and
Evaluation in Israel. The questions mostly consisted of reflection and addition problems, which are prominent in human IQ tests. Our model scored 38 correct answers out of 40
(5% error rate). This lies within the top 5% of human performance for all ages in the range between 20 and 70 according to the Institute’s measurements and similar performance
thresholds are given by (Raven 2000). This performance percentile corresponds to an IQ of above 127 (Hunt 2010). The
Figure 6: Sample Question from the Test Set. Answer: 3
true IQ is potentially higher, as the test does not discriminate
between the top IQ percentages.
One difference between our dataset and IQ tests used in
practice is the complexity of the shapes (more edges, more
types of shapes etc.) and the number of options to choose
from. While the number of options can easily be dealt with
by small modification to the architectures, the complexity of
the shapes pose an obstacle for our model. It was not able to
successfully generate crisp pictures in the Open Questions
format. We believe that by making our dataset more picturerealistic we can improve transfer learning in this scenario.
Conclusion and Future Work
We presented an effective neural network model for solving
geometric pattern recognition problems as used in IQ tests.
We also presented a model that is capable of generating the
images of the answers, giving more insight into the patterns
that our model is able to learn.
More research can be done on the types of questions that
CNNs can solve, or more importantly, the types of problems
that they fail at. In the future, we would like to explore more
challenging patterns that capture the limit of CNN learning capacity. Subjects with high scores are typically tested
on questions of ”Advanced Progressive Matrices”, but these
questions were not compatible with the current architecture
of our network. We would like to adjust our model to all
types of questions that appear in Raven’s Progressive Matrices test.
References
[Crawford et al. 2001] Crawford, J. R.; Deary, I. J.; Starr, J.;
and Whalley, L. J. 2001. The nart as an index of prior intellectual functioning: a retrospective validity study covering a
66-year interval. Psychological medicine 31(3):451–458.
[DiLalla et al. 1990] DiLalla, L. F.; Thompson, L.; Plomin,
R.; Phillips, K.; Fagan III, J. F.; Haith, M. M.; Cyphers,
L. H.; and Fulker, D. W. 1990. Infant predictors of preschool
and adult iq: A study of infant twins and their parents. Developmental Psychology 26(5):759.
[Flynn 1987] Flynn, J. R. 1987. Massive iq gains in 14 nations: What iq tests really measure. Psychological bulletin
101(2):171.
[Hinton and Salakhutdinov 2006] Hinton, G. E., and
Salakhutdinov, R. R. 2006. Reducing the dimensionality of
data with neural networks. science 313(5786):504–507.
[Hoshen and Peleg 2016] Hoshen, Y., and Peleg, S. 2016.
Visual learning of arithmetic operation. In AAAI, 3733–
3739.
[Hunt 2010] Hunt, E. 2010. Human intelligence. Cambridge
University Press.
[Jaarsveld et al. 2010] Jaarsveld, S.; Lachmann, T.; Hamel,
R.; and Leeuwen, C. v. 2010. Solving and creating raven
progressive matrices: Reasoning in well-and ill-defined
problem spaces. Creativity Research Journal 22(3):304–
319.
[Kingma and Ba 2014] Kingma, D., and Ba, J. 2014. Adam:
A method for stochastic optimization. arXiv preprint
arXiv:1412.6980.
[Krizhevsky, Sutskever, and Hinton 2012] Krizhevsky, A.;
Sutskever, I.; and Hinton, G. E. 2012. Imagenet classification with deep convolutional neural networks. In Advances
in neural information processing systems, 1097–1105.
[Kunda, McGreggor, and Goel 2009] Kunda, M.; McGreggor, K.; and Goel, A. K. 2009. Addressing the raven’s
progressive matrices test of” general” intelligence. In AAAI
Fall Symposium: Multi-Representational Architectures for
Human-Level Intelligence.
[LeCun et al. 1990] LeCun, Y.; Boser, B. E.; Denker, J. S.;
Henderson, D.; Howard, R. E.; Hubbard, W. E.; and Jackel,
L. D. 1990. Handwritten digit recognition with a backpropagation network. In Advances in neural information
processing systems, 396–404.
[LeCun et al. 2012] LeCun, Y. A.; Bottou, L.; Orr, G. B.; and
Müller, K.-R. 2012. Efficient backprop. In Neural networks:
Tricks of the trade. Springer. 9–48.
[Mathieu, Couprie, and LeCun 2015] Mathieu, M.; Couprie,
C.; and LeCun, Y.
2015.
Deep multi-scale video
prediction beyond mean square error.
arXiv preprint
arXiv:1511.05440.
[Oh et al. 2015] Oh, J.; Guo, X.; Lee, H.; Lewis, R. L.; and
Singh, S. 2015. Action-conditional video prediction using
deep networks in atari games. In Advances in Neural Information Processing Systems, 2863–2871.
[Radford, Metz, and Chintala 2015] Radford, A.; Metz, L.;
and Chintala, S. 2015. Unsupervised representation learning with deep convolutional generative adversarial networks.
arXiv preprint arXiv:1511.06434.
[Raven 2000] Raven, J. 2000. The raven’s progressive matrices: change and stability over culture and time. Cognitive
psychology 41(1):1–48.
[Rumelhart et al. ] Rumelhart, D. E.; Hinton, G. E.;
Williams, R. J.; et al.
Learning representations by
back-propagating errors. Cognitive modeling 5(3):1.
[Siegel 1989] Siegel, L. S. 1989. Iq is irrelevant to the definition of learning disabilities. Journal of learning disabilities
22(8):469–478.
[Stern 1914] Stern, W. 1914. The psychological methods of
testing intelligence. Number 13. Warwick & York.
[Taigman et al. ] Taigman, Y.; Yang, M.; Ranzato, M.; and
Wolf, L. Deepface: Closing the gap to human-level performance in face verification. In CVPR’14.
[Vondrick, Pirsiavash, and Torralba 2016] Vondrick, C.; Pirsiavash, H.; and Torralba, A. 2016. Anticipating visual representations from unlabeled video. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, 98–106.
[Wang et al. 2015] Wang, H.; Tian, F.; Gao, B.; Bian, J.; and
Liu, T.-Y. 2015. Solving verbal comprehension questions
in iq test by knowledge-powered word embedding. arXiv
preprint arXiv:1505.07909.
[Xiong et al. ] Xiong, W.; Droppo, J.; Huang, X.; Seide, F.;
Seltzer, M.; Stolcke, A.; Yu, D.; and Zweig, G. The microsoft 2016 conversational speech recognition system. In
ICASSP’17.
| 1cs.CV
|
arXiv:1501.01346v4 [math.NT] 16 Sep 2016
CONSTRUCTION OF UNIPOTENT GALOIS EXTENSIONS AND MASSEY
PRODUCTS
JÁN MINÁČ AND NGUYỄN DUY TÂN
Dedicated to Alexander Merkurjev
A BSTRACT. For all primes p and for all fields, we find a sufficient and necessary condition of the existence of a unipotent Galois extension of degree p6. The main goal of
this paper is to describe an explicit construction of such a Galois extension over fields
admitting such a Galois extension. This construction is surprising in its simplicity and
generality. The problem of finding such a construction has been left open since 2003.
Recently a possible solution of this problem gained urgency because of an effort to extend new advances in Galois theory and its relations with Massey products in Galois
cohomology.
1. I NTRODUCTION
From the very beginning of the invention of Galois theory, one problem has emerged.
For a given finite group G, find a Galois extension K/Q such that Gal(K/Q ) ≃ G. This
is still an open problem in spite of the great efforts of a number of mathematicians and
substantial progress having been made with specific groups G. (See [Se3].) A more general problem is to ask the same question over other base fields F. This is a challenging
and difficult problem even for groups G of prime power order.
In this paper we make progress on this classical problem in Galois theory. Moreover this progress fits together well with a new development relating Massey products in Galois cohomology to basic problems in Galois theory. For all primes p and all
fields in the key test case of n = 4, we construct Galois extensions with the unipotent
Galois group U n (F p ) assuming only the existence of some Galois extensions of order
p3 . This fits into a program outlined in [MT1] and [MT2], for the systematic construction of Galois p-closed extensions of general fields, assuming only knowledge of Galois
extensions of degree less than or equal to p3 and the structure of pth power roots of
unity in the base field. Thus both the methods and the results in this paper pave the
way to a program for obtaining the structure of maximal pro-p-quotients of absolute
Galois groups for all fields. We shall now describe some previous work of a number of
mathematicians which has influenced our work, as well as its significance for further
developments and applications.
JM is partially supported by the Natural Sciences and Engineering Research Council of Canada
(NSERC) grant R0370A01. NDT is partially supported by the National Foundation for Science and Technology Development (NAFOSTED) grant 101.04-2014.34.
1
2
JÁN MINÁČ AND NGUYỄN DUY TÂN
Recently there has been substantial progress in Galois cohomology which has changed
our perspective on Galois p-extensions over general fields. In some remarkable work,
M. Rost and V. Voevodsky proved the Bloch-Kato conjecture on the structure of Galois
cohomology of general fields. (See [Voe1, Voe2].) From this work it follows that there
must be enough Galois extensions to make higher degree Galois cohomology decomposable. However the explicit construction of such Galois extensions is completely
mysterious. In [MT1], [MT2] and [MT5], two new conjectures, the Vanishing n-Massey
Conjecture and the Kernel n-Unipotent Conjecture were proposed. These conjectures
in [MT1] and [MT2], and the results in this paper, lead to a program of constructing these previously mysterious Galois extensions in a systematic way. In these papers it is shown that the truth of these conjectures has some significant implications
on the structure of absolute Galois groups. These conjectures are based on a number of previous considerations. One motivation comes from topological considerations. (See [DGMS] and [HW].) Another motivation is a program to describe various
n-central series of absolute Galois groups as kernels of simple Galois representations.
(See [CEM, Ef, EM1, EM2, MSp, NQD, Vi].) If the Vanishing n-Massey Conjecture is
true, then by a result in [Dwy], we obtain a program of building up n-unipotent Galois
representations of absolute Galois groups by induction on n. This is an attractive program because we obtain a procedure of constructing larger Galois p-extensions from
smaller ones, efficiently using the fact that certain a priori natural cohomological obstructions to this procedure always vanish.
Recall that for each natural number n, U n (F p ) is the group of upper triangular n × nmatrices with entries in F p and diagonal entries 1. Then U3 (F2 ) is isomorphic to the
dihedral group of order 8, and if p is odd, then U3 (F p ) is isomorphic to the Heisenberg group H p3 of order p3 . For all n ≥ 4 and all primes p, we can think of U n (F p ) as
"higher Heisenberg groups" of order pn(n−1)/2 . It is now recognized that these groups
play a very special role in current Galois theory. Because U n (F p ) is a Sylow p-subgroup
of GLn (F p ), and every finite p-group has a faithful linear n-dimensional representation
over F p , for some n, we see that every finite p-group can be embedded into U n (F p )
for some n. Besides, the Vanishing n-Massey Conjecture and the Kernel n-Unipotent
Conjecture also indicate some deeper reasons why U n (F p ) is of special interest. The
constructions of Galois extensions with the Galois group U3 (F p ) over fields which admit them, are well-known in the case when the base field is of characteristic not p. They
are an important basic tool in the Galois theory of p-extensions. (See for example [JLY,
Sections 6.5 and 6.6]. Some early papers related to these topics like [MNg] and [M] now
belong to classical work on Galois theory.)
In [GLMS, Section 4], a construction of Galois extensions K/F, char( F) 6= 2, with
Gal(K/F) ≃ U4 (F2 ), was discovered. Already at that time, one reason for searching for this construction was the motivation to find ideas to extend deep results on
the characterization of the fixed field of the third 2-Zassenhaus filtration of an absolute Galois group GF as the compositum of Galois extensions of degree at most 8 (see
CONSTRUCTION OF UNIPOTENT GALOIS EXTENSIONS AND MASSEY PRODUCTS
3
[Ef, EM2, MSp, Vi]), to a similar characterization of the fixed field of the fourth 2Zassenhaus filtration of GF . In retrospect, looking at this construction, one recognizes
some elements of the basic theory of Massey products. However at that time the authors of [GLMS] were not familiar with Massey products. It was realized that such a
construction would also be desirable for U4 (F p ) for all p rather than U4 (F2 ), but none
has been found until now.
In [GLMS], in the construction of a Galois field extension K/F with Gal(K/F) ≃
U4 (F2 ), a simple criteria was used for an element in F to be a norm from a bicyclic
extension of degree 4 modulo non-zero squares in the base field F. However in [Me],
A. Merkurjev showed that a straightforward generalization of this criteria for p odd
instead of p = 2, is not true in general. Therefore it was not clear whether such an
analogous construction of Galois extensions K/F with Gal(K/F) ≃ U4 (F p ) was possible for p odd.
On the other hand, a new consideration in [HW], [MT1] and [MT2] led us to formulate the Vanishing n-Massey Conjecture, and the most natural way to prove this conjecture for n = 3 in the key non-degenerate case would be through constructing explicit
Galois U4 (F p )-extensions. In fact we pursued both cohomological variants of proving the Vanishing 3-Massey Conjecture and the Galois theoretic construction of Galois
U4 (F p )-extensions.
The story of proving this conjecture and finally constructing Galois U4 (F p )-extensions
over all fields which admit them, is interesting. First M. J. Hopkins and K. G. Wickelgren
in [HW] proved a result which implies that the Vanishing 3-Massey Conjecture with respect to prime 2, is true for all global fields of characteristic not 2. In [MT1] we proved
that the result of [HW] is valid for any field F. At the same time, in [MT1] the Vanishing n-Massey Conjecture was formulated, and applications on the structure of the
quotients of absolute Galois groups were deduced. In [MT3] we proved that the Vanishing 3-Massey Conjecture with respect to any prime p is true for any global field F
containing a primitive p-th root of unity. In [EMa1], I. Efrat and E. Matzri provided
alternative proofs for the above-mentioned results in [MT1] and [MT3]. In [Ma], E.
Matzri proved that for any prime p and for any field F containing a primitive p-th root
of unity, every defined triple Massey product contains 0. This established the Vanishing
3-Massey Conjecture in the form formulated in [MT1]. Shortly after [Ma] appeared on
the arXiv, two new preprints, [EMa2] and [MT5], appeared nearly simultaneously and
independently on the arXiv as well. In [EMa2], I. Efrat and E. Matzri replace [Ma] and
provide a cohomological approach to the proof of the main result in [Ma]. In [MT5] we
also provide a cohomological method of proving the same result. We also extend the
vanishing of triple Massey products to all fields, and thus remove the restriction that
the base field contains a primitive p-th root of unity. We further provide applications on
the structure of some canonical quotients of absolute Galois groups, and also show that
some special higher n-fold Massey products vanish. Finally in this paper we are able to
provide a construction of the Galois U4 (F p )-extension M/F for any field F which admits such an extension. We use this construction to provide a natural new proof, which
4
JÁN MINÁČ AND NGUYỄN DUY TÂN
we were seeking from the beginning of our search for a Galois theoretic proof, of the
vanishing of triple Massey products over all fields.
Some interesting cases of "automatic" realizations of Galois groups are known. These
are cases when the existence of one Galois group over a given field forces the existence
of some other Galois groups over this field. (See for example [Je, MS2, MSS, MZ, Wh].)
However, nontrivial cases of automatic realizations coming from an actual construction
of embedding smaller Galois extensions to larger ones, are relatively rare, and they are
difficult to produce. In our construction we are able, from knowledge of the existence
of two Heisenberg Galois extensions of degree p3 over a given base field F as above, to
find possibly another pair of Heisenberg Galois extensions whose compositum can be
automatically embedded in a Galois U4 (F p )-extension. (See also Remark 3.3.) Observe
that in all proofs of the Vanishing 3-Massey Conjecture we currently have, constructing
Heisenberg Galois extensions of degree p3 has played an important role. For the sake of
a possible inductive proof of the Vanishing n-Massey Conjecture, it seems important to
be able to inductively construct Galois U n (F p )-extensions. This has now been achieved
for the induction step from n = 3 to n = 4, and it opens up a way to approach the
Vanishing 4-Massey Conjecture.
Another motivation for this work which combines well with the motivation described
above, comes from anabelian birational considerations. Very roughly in various generality and precision, it was observed that small canonical quotients of absolute Galois
groups determine surprisingly precise information about base fields, in some cases entire base fields up to isomorphisms. (See [BT1, BT2, CEM, EM1, EM2, MSp, Pop].) But
these results suggest that some small canonical quotients of an absolute Galois group
together with knowledge of the roots of unity in the base field should determine larger
canonical quotients of this absolute Galois group. The Vanishing n-Massey Conjecture
and the Kernel n-Unipotent Conjecture, together with the program of explicit constructions of Galois U n (F p )-extensions, make this project more precise. Thus our main results, Theorems 3.7, 3.9, 4.3 and 5.5, contribute to this project.
A further potentially important application for this work is the theory of Galois pextensions of global fields with restricted ramification and questions surrounding the
Fontaine-Mazur conjecture. (See [Ko], [La], [McL], [Ga],[Se2].) For example in [McL,
Section 3], there is a criterion for infinite Hilbert p-class field towers over quadratic
imaginary number fields relying on the vanishing of certain triple Massey products.
The explicit constructions in this paper should be useful for approaching these classical
number theoretic problems.
Only relatively recently, the investigations of the Galois realizability of some larger
p-groups among families of small p-groups, appeared. (See the very interesting papers
[Mi1], [Mi2], [GS].) In these papers the main concern is understanding cohomological
and Brauer group obstructions for the realizability of Galois field extensions with prescribed Galois groups. In our paper the main concern is the explicit constructions and
their connections with Massey products. In other recent papers [CMS] and [Sch], the
CONSTRUCTION OF UNIPOTENT GALOIS EXTENSIONS AND MASSEY PRODUCTS
5
authors succeeded to treat the cases of characteristic equal to p or not equal to p, nearly
uniformly. This is also the case with our paper.
Our paper is organized as follows. In Section 2 we recall basic notions about norm
residue symbols and Heisenberg extensions of degree p3 . (For convenience we think
of the dihedral group of order 8 as the Heisenberg group of order 8.) In Section 3 we
provide a detailed construction of Galois U4 (F p )-extensions beginning with two "compatible" Heisenberg extensions of degree p3 . Section 3 is divided into two subsections.
In Subsection 3.1 we provide a construction of the required Galois extension M/F over
any field F which contains a primitive p-th root of unity. In Subsection 3.2 we provide
such a construction for all fields of characteristic not p, building on the results and methods in Subsection 3.1. In Example 3.8 we illustrate our method on a surprisingly simple
construction of Galois U4 (F2 )-extensions over any field F with char( F) 6= 2. In Section
4 we provide a required construction for all fields of characteristic p. After the original
and classical papers of E. Artin and O. Schreier [ASch] and E. Witt [Wi], these constructions seem to add new results on the construction of basic Galois extensions M/F with
Galois groups U n (F p ), n = 3 and n = 4. These are aesthetically pleasing constructions
with remarkable simplicity. They follow constructions in characteristic not p, but they
are simpler. See also [JLY, Section 5.6 and Appendix A1] for another procedure to obtain
these Galois extensions. In Section 5 we provide a new natural Galois theoretic proof of
the vanishing of triple Massey products over all fields in the key non-degenerate case.
We also complete the new proof of the vanishing of triple Massey products in the case
when a primitive p-th root of unity is contained in the base field. Finally we formulate a
necessary and sufficient condition for the existence of a Galois U4 (F p )-extension M/F
which contains an elementary p-extension of any field F (described by three linearly
independent characters), and we summarize the main results in Theorem 5.5.
Acknowledgements: We would like to thank M. Ataei, L. Bary-Soroker, S. K. Chebolu,
I. Efrat, H. Ésnault, E. Frenkel, S. Gille, J. Gärtner, P. Guillot, D. Harbater, M. J. Hopkins, Ch. Kapulkin, I. Kříž, J. Labute, T.-Y. Lam, Ch. Maire, E. Matzri, C. McLeman,
D. Neftin, J. Nekovář, R. Parimala, C. Quadrelli, M. Rogelstad, A. Schultz, R. Sujatha,
Ng. Q. Thắng, A. Topaz, K. G. Wickelgren and O. Wittenberg for having been able to
share our enthusiasm for this relatively new subject of Massey products in Galois cohomology, and for their encouragement, support, and inspiring discussions. We are
very grateful to the anonymous referee for his/her careful reading of our paper, and
for providing us with insightful comments and valuable suggestions which we used to
improve our exposition.
Notation: If G is a group and x, y ∈ G, then [ x, y] denotes the commutator xyx −1 y−1 . For
any element σ of finite order n in G, we denote Nσ to be the element 1 + σ + · · · + σn−1
in the integral group ring Z [ G ] of G.
For a field F, we denote Fs (respectively GF ) to be its separable closure (respectively its
absolute Galois group Gal( Fs /F)). We denote F× to be the set of non-zero elements of
6
JÁN MINÁČ AND NGUYỄN DUY TÂN
F. For a given profinite group G, we call a Galois extension E/F, a (Galois) G-extension
if the Galois group Gal(E/F) is isomorphic to G.
For a unital commutative ring R and an integer n ≥ 2, we denote U n ( R) as the group
of all upper-triangular unipotent n × n-matrices with entries in R. For any (continuous)
representation ρ : G → U n ( R) from a (profinite) group G to U n ( R) (equipped with
discrete topology ), and 1 ≤ i < j ≤ n, let ρij : G → R be the composition of ρ with the
projection from U n ( R) to its (i, j)-coordinate.
2. H EISENBERG
EXTENSIONS
The materials in this section have been taken from [MT5, Section 3].
2.1. Norm residue symbols. Let F be a field containing a primitive p-th root of unity ξ.
For any element a in F× , we shall write χ a for the character corresponding to a via the
Kummer map F× → H 1 (GF , Z/pZ )√= Hom(GF , Z/pZ). From now on we assume that
p
a is not in ( F× ) p . The extension F( √
a)/F is√a Galois extension with the Galois group
hσa i ≃ Z/pZ, where σa satisfies σa ( p a) = ξ p a.
The character χ a defines a homomorphism χ a ∈ Hom(GF , 1p Z/Z ) ⊆ Hom(GF , Q/Z )
by the formula
χa =
1
χa .
p
Let b be any element in F× . Then the norm residue symbol may be defined as
(a, b) := (χa , b) := b ∪ δχa .
Here δ is the coboundary homomorphism δ : H 1 (G, Q/Z ) → H 2 (G, Z ) associated to
the short exact sequence of trivial G-modules
0 → Z → Q → Q/Z → 0.
The cup product χ a ∪ χb ∈ H 2 (GF , Z/pZ ) can be interpreted as the norm residue
symbol (a, b). More precisely, we consider the exact sequence
x 7→ x p
0 −→ Z/pZ −→ Fs× −→ Fs× −→ 1,
where Z/pZ has been identified with the group of p-th roots of unity µ p via the choice
of ξ. As H 1 (GF , Fs× ) = 0, we obtain
i
×p
0−→ H 2 (GF , Z/pZ ) −→ H 2 (GF , Fs× ) −→ H 2 (GF , Fs× ).
Then one has i (χ a ∪ χb ) = (a, b) ∈ H 2 (GF , Fs× ). (See [Se1, Chapter XIV, Proposition 5].)
CONSTRUCTION OF UNIPOTENT GALOIS EXTENSIONS AND MASSEY PRODUCTS
7
2.2. Heisenberg extensions. In this subsection we recall some basic facts about Heisenberg extensions. (See [Sha, Chapter 2, Section 2.4] and [JLY, Sections 6.5 and 6.6 ].)
Assume that a,
b are elements in F× , which are linearly independent modulo ( F× ) p .
√
√
p
Let K = F( p a, b). Then
K/F
is a Galois extension whose Galois √
group is√
generated
√
√
√
√
√
√
p
p
p
p
p
p
p
p
by σa and σb . Here σa ( b) = b, σa ( a) = ξ a; σb ( a) = a, σb ( b) = ξ b.
1 x z
2
We consider a map U3 (Z/pZ ) → (Z/pZ ) which sends 0 1 y to ( x, y). Then
0 0 1
we have the following embedding problem
GF
ρ̄
0
/
Z/pZ
/
U3 (Z/pZ )
/
(Z/pZ)2
/
1,
where ρ̄ is the map (χ a , χb ) : GF → Gal(K/F) ≃ (Z/pZ )2 . (The last isomorphism
Gal(K/F) ≃ (Z/pZ )2 is the one which sends σa to (1, 0) and σb to (0, 1).)
Assume that
√ χa ∪ χb = 0. Then the norm residue symbol (a, b) is trivial. Hence there
exists α in F( p a) such that NF ( √
p a ) /F (α ) = b (see [Se1, Chapter XIV, Proposition 4 (iii)]).
We set
p−2
√
p−2
p−1
p−2
A0 = α σa (α
) · · · σa (α) = ∏ σai (α p−i −1 ) ∈ F( p a).
i =0
Lemma 2.1. Let f a be an element in F× . Let A = f a A0 . Then we have
NF ( √
p a ) /F (α )
b
σa ( A)
=
= p.
p
A
α
α
σa ( A)
σa ( A0 )
Proof. Observe that
=
. The lemma then follows from the identity
A
A0
( s − 1)
p−2
∑ ( p − i − 1) s
i =0
i
=
p−1
∑ si − ps0 .
i =0
Proposition 2.2. Assume that χ a ∪ χb = 0. Let f a be an element in F× . Let A = f a A0 be
defined as above. Then the homomorphism ρ̄ := (χ a , χb ) : GF → Z/pZ × Z/pZ lifts to a
Heisenberg extension ρ : GF → U3 (Z/pZ ).
√
p
Sketch of Proof. Let L := K ( A)/F. Then L/F is Galois extension. Let σ̃a ∈ Gal( L/F)
(resp.
σ̃b ∈ Gal
( L/F)) be an extension of σa (resp.
σb ).
Since σb ( A) = A, we have
√
√
√
p √
p
p
p
p
j
σ̃b ( A ) = ξ A, for some j ∈ Z. Hence σ̃b ( A ) = A. This implies that σ̃b is of
order p.
√
p
√
√
√
b
b
p
p
p
p
i
On the other hand, we have σ̃a ( A ) = σa ( A) = A p . Hence σ̃a ( A ) = ξ A
,
α
α
√
p √
p
p
for some i ∈ Z. Then σ̃a ( A ) = A. Thus σ̃a is of order p.
JÁN MINÁČ AND NGUYỄN DUY TÂN
8
√
√
p
p
If we set σA := [σ̃a , σ̃b ], then σA ( A) = ξ −1 A. This implies that σA is of order p.
Also one can check that
[σ̃a , σA ] = [σ̃b , σA ] = 1.
We can define an isomorphism ϕ : Gal( L/F)
1 1 0
1 0
0
1
0
σa 7→
, σb 7→ 0 1
0 0 1
0 0
ϕ
→ U3 (Z/pZ) by letting
0
1 0 1
1 , σA 7→ 0 1 0 .
1
0 0 1
Then the composition ρ : GF → Gal( L/F) → U3 (Z/pZ ) is the desired lifting of ρ̄.
Note that [ L : F] = p3 . Hence there are exactly p extensions of σa ∈ Gal(E/F) to the
automorphisms in Gal( L/F) since [ L : E] = p3 /p2 = p. Therefore for later use, we can
choose an extension,√still denoted by σa ∈ Gal( L/F), of σa ∈ Gal(K/F) in such a way
p
√
√
b
p
p
that σa ( A) = A
.
α
3. T HE
CONSTRUCTION OF
U4 (F p )- EXTENSIONS :
THE CASE OF CHARACTERISTIC
6= p
3.1. Fields containing primitive p-th roots of unity. In this subsection we assume that
F is a field containing a primitive p-th root ξ of unity. The following result can be
deduced from Theorem 5.5, but for the convenience of the reader we include a proof
here.
Proposition 3.1. Assume that there exists a Galois extension M/F such that Gal( M/F) ≃
U4 (F p ). Then there exist a, b, c ∈ F× such that a, b, c are linearly independent modulo ( F× ) p
√
√ √
p
and (a, b) = (b, c) = 0. Moreover M contains F( p a, b, p c).
Proof. Let ρ be the composite ρ : GF ։ Gal( M/F) ≃ U4 (F p ). Then ρ12 , ρ23 and ρ34 are
elements in Hom(GF , F p ). Hence there are a, b and c in F× such that χ a = ρ12 , χb = ρ23
and χc = ρ34 . Since ρ is a group homomorphism, by looking at the coboundaries of ρ13
and ρ24 , we see that
χ a ∪ χ b = χ b ∪ χ c = 0 ∈ H 2 ( G F , F p ).
This implies that (a, b) = (b, c) = 0 by [Se1, Chapter XIV, Proposition 5].
Let ϕ := (χ a , χb , χc ) : GF → (F p )3 . Then ϕ is surjective. By Galois correspondence,
we have
√ √
√
p
Gal( Fs /F( p a, b, p c)) = ker χ a ∩ ker χb ∩ ker χc = ker ϕ.
√
√ √
p
This implies that Gal( F( p a, b, p c)/F) ≃ (F p )3 . Hence by Kummer theory, we see
√ √
√
p
that a, b and c are linearly independent modulo ( F× ) p . Clearly, M contains F( p a, b, p c).
Conversely we shall see in this section that given these necessary conditions for the
existence of U4 (F p )-Galois extensions over F, as in Proposition 3.1, we can construct a
Galois extension M/F with the Galois group isomorphic to U4 (F p ).
CONSTRUCTION OF UNIPOTENT GALOIS EXTENSIONS AND MASSEY PRODUCTS
9
From now on we assume that we are given elements a, b and c in F× such that a, b and
c are linearly independent modulo ( F× ) p and that (a, b) = (b, √
c) = 0. We shall construct
√
√
p
p
a Galois U4 (F p )-extension M/F such that M contains F( a, b, p c).
√ √
√
√ √
√
p
p
First we note that F( p a, b, p c)/F is a Galois extension with Gal( F( p a, b, p c)/F)
generated by σa , σb , σc . Here
√
√
√
√
√
√
p
p
σa ( p a) = ξ p a, σa ( b) = b, σa ( p c) = p c;
√
√
√
√
√
√
p
p
σb ( p a) = p a, σb ( b) = ξ b, σb ( p c) = p c;
√
√
√
√
√
√
p
p
σc ( p a) = p a, σc ( b) = b, σc ( p c) = ξ p c.
√
√
√ √
Let E = F( p a, p c). Since (a, b) = (b, c) = 0, there are α in F( p a) and γ in F( p c)
(see [Se1, Chapter XIV, Proposition 4 (iii)]) such that
√
NF ( √
p a ) /F (α ) = b = NF ( p c) /F (γ ).
Let G be the Galois group Gal(E/F). Then G = h√
σa , σc i, where σa ∈ G (respec√
√
p
p
tively σc ∈√G) is the restriction of σa ∈ Gal( F( a, b, p c )/F) (respectively σc ∈
√ p √
Gal( F( p a, b, p c)/F)).
√
p
Our next goal is to find an element δ in E× such that the Galois closure of E( δ) is
our desired U4 (F p )-extension of F. We define
C0 =
p−2
√
∏ σci (γ p−i−1 ) ∈ F( p a),
i =0
and define B := γ/α. Then we have the following result, which follows from Lemma 2.1
(see [Ma, Proposition 3.2] and/or [MT5, Lemma 4.2]).
Lemma 3.2. We have
σa ( A0 )
(1)
= Nσc ( B).
A0
σc (C0 )
(2)
= Nσa ( B)−1 .
C0
Remark 3.3. We would like to informally explain the meaning of the next lemma. From
our hypothesis (a, b) = 0 = (b, c) and from
Subsection 2.2, we√see that we can obtain
√
√
√ √
√
p
p
p
p
F( b, p c, p C0 ) of F. Here
two Heisenberg extensions L1 = F( a, b,√ A0 ) and L2 = √
we have chosen specific elements A0 ∈ F( p a) and C0 ∈ F( p c). However we may not
be able to embed the compositum of L1 and L2 into our desired Galois extension M/F
with Gal( M/F) ≃ U4 (F p ). We know that we can modify the element A0 by any element
f a ∈ F× and the element C0 by any element f c ∈ F× obtaining elements A = f a A0 and
C = f c C0 instead of A0 and C0 . This new choice of elements may change the √
fields L1
√
p
p
and
L2 but the new fields will still be Heisenberg extensions containing F( a, b) and
√
√
p
F( b, p c) respectively. The next lemma will provide us with a suitable modification
of A0 and C0 . From the proof of Theorem 3.7 we shall see that the compositum of
JÁN MINÁČ AND NGUYỄN DUY TÂN
10
these modified Heisenberg extensions can indeed be embedded into a Galois extension
M/F with Gal( M/F) ≃ U4 (F p ). This explains our comment in the introduction in the
paragraph related to the "automatic realization of Galois groups".
Lemma 3.4. Assume that there exist C1 , C2 ∈ E× such that
B=
σa (C1 ) C2
.
C1 σc (C2 )
Then Nσc (C1 )/A0 and √
Nσa (C2 )/C0 are in F× . Moreover, if we let A = Nσc (C1 ) ∈ F(
and C = Nσa (C2 ) ∈ F( p c)× , then there exists δ ∈ E× such that
σc (δ)
−p
= AC1 ,
δ
σa (δ)
−p
= CC2 .
δ
Proof. By Lemma 3.2, we have
σa ( A0 )
= Nσc ( B) = Nσc
A0
σa (C1 )
C1
Nσc
Nσc (C1 )
= σa
A0
Nσc (C1 )
A0
C2
σc (C2 )
=
σa ( Nσc (C1 ))
.
Nσc (C1 )
This implies that
Hence
.
√
√
Nσc (C1 )
∈ F ( p c )× ∩ F ( p a )× = F × .
A0
By Lemma 3.2, we have
σc (C0 )
= Nσa ( B−1 ) = Nσa
C0
C1
σa (C1 )
Nσa
σc (C2 )
C2
This implies that
Nσa (C2 )
= σc
C0
Hence
Nσa (C2 )
C0
.
√
√
Nσa (C2 )
∈ F ( p a )× ∩ F ( p c )× = F × .
C0
Clearly, one has
−p
Nσa (CC2 ) = 1,
−p
Nσc ( AC1 ) = 1.
=
σc ( Nσa (C2 ))
.
Nσa (C2 )
√
p
a )×
CONSTRUCTION OF UNIPOTENT GALOIS EXTENSIONS AND MASSEY PRODUCTS
11
We also have
−p
σa ( AC1 )
−p
AC1
−p
− p
C2
σa ( A) σa (C1 ) − p C
−p =
A
C1
σc (C) σc (C2 )
σc (CC2 )
p
b γ
= p B− p
α b
= 1.
CC2
Hence, we have
−p
σa ( AC1 )
−p
AC1
=
−p
σc (CC2 )
−p
CC2
.
From [Co, page 756] we see that there exists δ ∈ E× such that
σc (δ)
−p
= AC1 ,
δ
σa (δ)
−p
= CC2 ,
δ
as desired.
Remark 3.5. The result of I. G. Connell which we use in the above proof, is a variant of
Hilbert’s Theorem 90. This result was independently discovered by S. Amitsur and D.
Saltman in [AS, Lemma 2.4]. (See also [DMSS, Theorem 2] for the case p = 2.)
Lemma 3.6. There exists e ∈ E× such that B =
the following statements are true.
σa σc (e)
. Furthermore, for such an element e
e
(1) If we set C1 := σc (e) ∈ E× , C2 := e−1 ∈ E× , then B =
p−2
(2) If we set C1 := e ∈ E× , C2 := (eB)σc (eB) · · · σc
σa (C1 ) C2
.
C1 σc (C2 )
(eB) ∈ E× , then B =
Proof. We have
Nσa σc ( B) =
Nσa (α)
b
Nσa σc (α)
=
= = 1.
Nσa σc (γ)
Nσc (γ)
b
Hence by Hilbert’s Theorem 90, there exists e ∈ E× such that B =
σa σc (e)
.
e
(1) Clearly, we have
σa (C1 ) C2
σa σc (e)
σa (σc (e)) e−1
=
=
= B.
−
1
C1 σc (C2 )
σc (e) σc (e )
e
σa (C1 ) C2
.
C1 σc (C2 )
12
JÁN MINÁČ AND NGUYỄN DUY TÂN
σa σc (e)
p−1
, we see that eB = σa σc (e). Hence σc (eB) = σa (e). Therefore
e
eB
σa (e)
σa (C1 ) C2
.
B=
=
p
−
1
e σc (eB)
C1 σc (C2 )
√
√
√
√
p
p
p
p
Theorem 3.7. Let the notation and assumption be as in
Lemma
3.4.
Let
M
:
=
E
(
δ,
A,
C,
b ).
√
√
√
p
p
p
Then M/F is a Galois extension, M contains F( a, b, c), and Gal( M/F) ≃ U4 (F p ).
(2) From B =
Proof. Let W ∗ be the F p -vector space in E× /(E× ) p generated by [b] E , [ A] E , [C] E and [δ] E .
Here for any 0 6= x in a field L, we denote [ x ] L the image of x in L× /( L× ) p . Since
(1)
σc (δ) = δAC1
−p
(by Lemma 3.4),
(2)
σa (δ) = δCC2
b
σa ( A) = A p
α
b
σc (C) = C p
γ
−p
(by Lemma 3.4),
(3)
(4)
(by Lemma 2.1),
(by Lemma 2.1),
we see that W ∗ is in fact an F p [ G ]-module. Hence M/F is a Galois extension by Kummer theory.
Claim: dimF p (W ∗ ) = 4. Hence [ L : F] = [ L : E][ E : F] = p4 p2 = p6 .
Proof of Claim: From our hypothesis that dimF p h[ a] F , [b] F , [c] F i = 3, we see that h[b] E i ≃
F p.
Clearly, h[b] E i ⊆ (W ∗ )G . From (3) one gets the relation
[σa ( A)] E = [ A] E [b] E .
This implies that [ A] E is not in (W ∗ )G . Hence dimF p h[b] E , [ A] E i = 2.
From (4) one gets the relation
[σc (C)] E = [C] E [b] E .
This implies that [C] E is not in (W ∗ )σc . But we have h[b] E , [ A] E i ⊆ (W ∗ )σc . Hence
dimF p h[b] E , [ A] E , [C] E i = 3.
Observe that the element (σa − 1)(σc − 1) annihilates the F p [ G ]-module h[b] E , [ A] E , [C] E i,
while by (1) and (3) one has
(σa − 1)(σc − 1)[δ] E =
σa ([ A] E )
= [b] E .
[ A] E
Therefore one has
dimF p W ∗ = dimF p h[b] E , [ A] E , [C] E , [δ] E i = 4.
CONSTRUCTION OF UNIPOTENT GALOIS EXTENSIONS AND MASSEY PRODUCTS
13
√
√
√ √ √
√
p
p
p
p
a, A, b) and H b,c = F( p c, p C, b). Let
√
√
√
√
√
√ √ √
p
p
p
p
p
p
N := H a,b H b,c = F( p a, p c, b, A, C ) = E( b, A, C).
Let H a,b = F(
Then N/F is a Galois extension of degree p5 . This is because Gal( N/E) is dual to the
F p [ G ]-submodule h[b] E , [ A] E , [C] E i via Kummer theory, and the proof of the claim above
shows that dimF p h[b] E , [ A] E , [C] E i = 3. We have the following commutative diagram
Gal( N/F)
//
Gal( H a,b /F)
Gal( H b,c /F)
//
Gal( F(
√
p
b)/F).
So we have a homomorphism η from Gal( N/F) to the pull-back Gal( H b,c /F) ×Gal ( F ( √
p
b)/F )
Gal( H a,b /F):
η : Gal( N/F) −→ Gal( H b,c /F) ×Gal( F ( √
p
Gal( H a,b /F),
b)/F )
which make the obvious diagram commute. We claim that η is injective. Indeed, let σ
be an element in ker η. Then σ | H a,b = 1 in Gal( H a,b /F), and σ | H b,c = 1 in Gal( H b,c /F).
Since N is the compositum of H a,b and H b,c , this implies that σ = 1, as desired.
Gal( H a,b /F)| = p5 = |Gal( N/F)|, we see that η
Since |Gal( H b,c /F) ×Gal( F ( √
p
b )/F )
is actually an isomorphism. As in the proof of
Proposition 2.2, we can choose an ex√
√
p
a,b
p
√
tension σa ∈ Gal( H /F) of σa ∈ Gal( F( a, b)/F) (more precisely, of σa | F ( √
p a, p b) ∈
√
√
p
Gal( F( p a, b)/F)) in such a way that
√
p
√
√
b
p
p
σa ( A ) = A
.
α
Since the square commutative diagram above is a pull-back, we can choose an extension
σa ∈ Gal( N/F) of σa ∈ Gal( H a,b /F) in such a way that
σa | H b,c = 1.
Now we can choose any extension σa ∈ Gal( M/F) of σa ∈ Gal( N/F). Then we have
√
p
√
√
b
p
p
σa ( A ) = A
and σa | H b,c = 1.
α
√
√
p
Similarly, we can choose an extension σc ∈ Gal( M/F) of σc ∈ Gal( F( b, p c)/F) in
such a way that
√
p
√
√
b
p
p
σc ( C) = C
, and σc | H a,b = 1.
γ
14
JÁN MINÁČ AND NGUYỄN DUY TÂN
We define σb ∈ Gal( M/E) to be the element which is dual to [b] E via Kummer theory.
In other words, we require that
√
√
p
p
σb ( b) = ξ b,
√ √
√
p
and σb acts trivially on p A, p C and δ. We consider√σb as an element in Gal( M/F),
√ p √
then it is clear that σb is an extension of σb ∈ Gal( F( p a, b, p c )/F). Let W = Gal( M/E),
and let H = Gal( M/F), then we have the following exact sequence
1 → W → H → G → 1.
By Kummer theory, it follows that W is dual to W ∗ , and hence W ≃ (Z/pZ )4 . In
particular, we have | H | = p6 .
Recall that from [BD, Theorem 1], we know that the group U4 (F p ) has a presentation
with generators s1 , s2 , s3 subject to the following relations
p
p
p
s1 = s2 = s3 = 1,
(R)
[s1 , s3 ] = 1,
[s1 , [s1 , s2 ]] = [s2 , [s1 , s2 ]] = 1,
[s2 , [s2 , s3 ]] = [s3 , [s2 , s3 ]] = 1,
[[s1 , s2 ], [s2 , s3 ]] = 1.
Note that |Gal( M/F)| = p6 . So in order to show that Gal( M/F) ≃ U4 (F p ), we shall
show that σa , σb and σc generate Gal( M/F) and that they satisfy these above relations.
Claim: The elements σa , σb and σc generate Gal( M/F).
Proof of Claim: Let K be the maximal p-elementary subextension of M. Note that Gal( M/F)
is a p-group. So in order to show that σa , σb and σc generate Gal( M/F), we only need to
show that (the restrictions of) these elements generate Gal(K/F) by the Burnside basis
theorem. (See e.g. [Ha, Theorem 12.2.1] or
[Ko, Theorem 4.10].) We shall now determine
√
p
p
the field K. By Kummer theory, K = F( ∆), where ∆ = ( F× ∩ M× )/( F× ) p . Let [ f ] F
be any element in ∆, where f ∈ F× ∩ ( M× ) p ⊆ E× ∩ ( M× ) p . By Kummer theory, one
has W ∗ = (E× ∩ ( M× ) p )/(E× ) p . Hence we can write
[ f ] E = [δ]ǫEδ [ A]ǫEA [C]ǫEC [b]ǫEb ,
ǫ
where ǫδ , ǫ A , ǫC , ǫb ∈ Z. By applying (σa − 1)(σc − 1) on [ f ] E we get [1] E = [b] Eδ . (See
the proof of the first claim of this proof.) Thus ǫδ is divisible by p, and one has
[ f ] E = [ A]ǫEA [C]ǫEC [b]ǫEb .
ǫ
By applying σa − 1 on both sides of this equation, we get [1] E = [b] EA . Thus ǫ A is divisible
by p. Similarly, ǫC is also divisible by p. Hence f = bǫb e p for some e ∈ E. Since b and f
are in F, e p is in F× ∩ (E× ) p and [e p ] F is in h[ a] F , [c] F i. Therefore [ f ] F is in h[ a] F , [b] F , [c] F i
and
∆ = h[ a] F , [b] F , [c] F i.
CONSTRUCTION OF UNIPOTENT GALOIS EXTENSIONS AND MASSEY PRODUCTS
15
√
√ √
p
So K = F( p a, b, p c). Then it is clear that σa , σb and σc generate Gal(K/F) and the
claim follows.
Claim: The order of σa is p.
√
p √
p
p
Proof of Claim: As in the proof of Proposition 2.2, we
see
that
σ
(
A
)
=
A.
a
√
√
√
−p
p
p
p
−1
i
Since σa (δ) = δCC2 (equation (2)), one has σa ( δ ) = ξ δ CC2 for some i ∈ Z.
This implies that
√
√
√
p
p
p
σa2 ( δ) = ξ i σa ( δ)σa ( C)σa (C2 )−1
√
√
p
p
= ξ 2i δ( C)2 C2−1 σa (C2 )−1 .
Inductively, we obtain
√
√
p
p √
p
p
σa ( δ) = ξ pi δ( C ) p Nσa (C2 )−1
√
p
= δ(C) Nσa (C2 )−1
√
p
= δ.
p
Therefore, we can conclude that σa = 1, and σa is of order p.
Claim: The order of σb is p.
√
√
√
√
p
p
p
Proof of Claim: This is clear because σb acts trivially on A, p C, δ, E and σb ( b) = ξ b.
Claim: The order of σc is p.
√
p √
p
p
Proof of Claim: As in the proof of Proposition 2.2, we
see
that
σ
(
C
)
=
C.
c
√
√
√
−p
p
p
p
−1
j
Since σc (δ) = δAC1 (equation (1)), one has σc ( δ) = ξ δ AC1 for some j ∈ Z.
This implies that
√
√
√
p
p
p
σc2 ( δ) = ξ j σc ( δ)σc ( A )σc (C1 )−1
√
√
p
p
= ξ 2j δ( A)2 C1−1 σc (C1 )−1 .
Inductively, we obtain
√
√
p
p
p √
p
σc ( δ) = ξ pj δ( A) p Nσc (C1 )−1
√
p
= δ( A) Nσc (C1 )−1
√
p
= δ.
p
Therefore, we can conclude that σc = 1, and σc is of order p.
Claim: [σa , σc ] = 1.
√
√
Proof of Claim: It is enough to check that σa σc ( p δ) = σc σa ( p δ).
JÁN MINÁČ AND NGUYỄN DUY TÂN
16
We have
√
√
√
p
p
p
σa σc ( δ) = σa (ξ j δ AC1−1 )
√
√
p
p
= ξ j σa ( δ)σa ( A)σa (C1 )−1
√
√ pb
√ √
p
j i p
−1 p
= ξ ξ δ CC2
A
σa (C1 )−1
α
√
p
√
√ √
b
p
p
i+ j p
δ C A
=ξ
(σa (C1 )C2 )−1
α
√
p
√
√ √
b (C1 σc (C2 ))−1
p
p
i+ j p
δ C A
=ξ
α
B
√
p
√
√
√
b
p
p
p
= ξ i+ j δ C A
(C1 σc (C2 ))−1 .
γ
On the other hand, we have
√
√
√
p
p
p
σc σa ( δ) = σc (ξ i δ CC2−1 )
√
√
p
p
= ξ i σc ( δ)σc ( C)σc (C2 )−1
√
p
√
√
√
b
p
p
p
i j
−1
σc (C2 )−1
= ξ ξ δ AC1
C
γ
√
p
√ √
√
b
p
p
i+ j p
=ξ
(C1 σc (C2 ))−1 .
δ A C
γ
√
√
Therefore, σa σc ( p δ ) = σc σa ( p δ), as desired.
Claim: [σa , [σa , σb ]] = [σb , [σa , σb ]] = 1.
Proof of Claim: Since G is abelian, it follows that [σa , σb ] is in W. Now both σb and [σa , σb ]
are in W. Hence [σb , [σa , σb ]] = 1 because W is abelian.
Now we show that [σa , [σa , σb ]] = 1. Since the Heisenberg group U3 (F p ) is a nilpotent
group of nilpotent length √
2, we see
that [σa , [σa , σb ]] = 1 on H a,b and H b,c . So it is enough
√
p
p
to check that [σa , [σa , σb ]]( δ) = δ.
From the choice of σb , we see that
√
√
√
p
p
p
σb σa ( δ) = σa ( δ) = σa σb ( δ ).
√
√
√
p
p
Hence, [σa , σb ]( δ) = δ. Since σa and σb act trivially on p C, and σb acts trivially on E,
we see that
√
√
p
p
[σa , σb ]( C ) = C, and [σa , σb ](C2−1 ) = C2−1 .
CONSTRUCTION OF UNIPOTENT GALOIS EXTENSIONS AND MASSEY PRODUCTS
17
We have
√
√
√
p
p
p
[σa , σb ]σa ( δ) = [σa , σb ](ξ i δ CC2−1 )
√
√
p
p
= [σa , σb ](ξ i )[σa , σb ]( δ)[σa , σb ]( C )[σa , σb ](C2−1 )
√
√
p
p
= ξ i δ CC2−1
√
p
= σa ( δ)
√
p
= σa [σa , σb ]( δ).
√
√
p
p
Thus [σa , [σa , σb ]]( δ) = δ, as desired.
Claim: [σb , [σb , σc ]] = [σc , [σb , σc ]] = 1.
Proof of Claim: Since G is abelian, it follows that [σb , σc ] is in W. Now both σb and [σb , σc ]
are in W. Hence [σb , [σb , σc ]] = 1 because W is abelian.
Now we show that [σc , [σb , σc ]] = 1. Since the Heisenberg group U3 (F p ) is a nilpotent
group of nilpotent length √
2, we see
that [σc , [σb , σc ]] = 1 on H a,b and H b,c . So it is enough
√
p
p
to check that [σc , [σb , σc ]]( δ) = δ.
From the choice of σb , we see that
√
√
√
p
p
p
σb σc ( δ) = σc ( δ) = σc σb ( δ).
√
√
√
p
p
p
Hence, [σb , σc ]( δ) = δ. Since σb and σc act trivially on A, and σb acts trivially on E,
we see that
√
√
p
p
[σb , σc ]( A) = A, and [σb , σc ](C1−1 ) = C1−1 .
We have
√
√
√
p
p
p
[σb , σc ]σc ( δ) = [σb , σc ](ξ j δ AC1−1 )
√
√
p
p
= [σb , σc ](ξ j )[σb , σc ]( δ)[σb , σc ]( A)[σb , σc ](C1−1 )
√
√
p
p
= ξ j δ AC1−1
√
p
= σc ( δ )
√
p
= σc [σa , σb ]( δ).
√
√
p
p
Thus [σc , [σb , σc ]]( δ) = δ, as desired.
Claim: [[σa , σb ], [σb , σc ]] = 1.
Proof of Claim: Since G is abelian, [σa , σb ] and [σb , σc ] are in W. Hence [[σa , σb ], [σb , σc ]] = 1
because W is abelian.
An explicit isomorphism ϕ : Gal( M/F) → U4 (F p ) may be defined as
1 0 0 0
1 0 0 0
1 1 0 0
0 1 0 0
, σb 7→ 0 1 1 0 , σc 7→ 0 1 0 0 .
σa 7→
0 0 1 1
0 0 1 0
0 0 1 0
0 0 0 1
0 0 0 1
0 0 0 1
JÁN MINÁČ AND NGUYỄN DUY TÂN
18
Example 3.8. Let the notation and assumption be as in Lemma 3.4. Let us consider the
α
. (Observe that α + γ 6= 0.) In fact,
case p = 2. In Lemma 3.6, we can choose e =
α+γ
one can easily check that
α
γ α
σa σc (
)=
.
α+γ
αα+γ
(1) If we choose C1 = σc (e) and C2 = e−1 as in Lemma 3.6 part (1), then we have
α2 γ
,
(α + γ)(αγ + b)
(α + γ)(αγ + b)
.
C = Nσa (C2 ) = Nσa (e−1 ) =
bα
α+γ
. In fact, we have
In Lemma 3.4, we can choose δ = e−1 =
α
σc (δ)
= σc (e)−1 e = σc (e)−2 eσc (e) = C1−2 Nσc (e) = AC1−2 ,
δ
σa (δ)
= σa (e−1 )e = e−1 σa (e−1 )e2 = Nσa (e−1 )C2−2 = CC2−2 .
δ
Therefore
s
r
r
√ √ √ √
√
α2 γ
(α + γ)(αγ + b)
α+γ
M = F( b, A, C, δ) = F( b,
,
,
)
(α + γ)(αγ + b)
bα
α
r
√
√
α+γ p
= F( b,
, αγ + b, αγ).
α
A = Nσc (C1 ) = Nσc (e) =
(2) If we choose C1 = e =
then we have
α
γ
and C2 = eB =
as in Lemma 3.6 part (2),
α+γ
α+γ
A = Nσc (C1 ) = Nσc (e) =
α2 γ
,
(α + γ)(αγ + b)
γ2 α
C = Nσa (C2 ) = Nσa (eB) =
.
(α + γ)(αγ + b)
In Lemma 3.4, we can choose δ = (α + γ)−1 . In fact, we have
γ(α + γ)
σc (δ)
=
= AC1−2 ,
δ
αγ + b
σa (δ)
α (α + γ)
=
= CC2−2 .
δ
αγ + b
CONSTRUCTION OF UNIPOTENT GALOIS EXTENSIONS AND MASSEY PRODUCTS
19
Therefore
s
s
α2 γ
αγ2 √
,
, α + γ ).
αγ + b
αγ + b
√
√ √ √
Observe also that M is the Galois closure of E( δ) = F( a, c, α + γ ).
√
√ √ √ √
M = F( b, A, C, δ) = F( b,
3.2. Fields of characteristic not p. Let F0 be an arbitrary field of characteristic 6= p. We
fix a primitive p-th root of unity ξ, and let F = F0 (ξ ). Then F/F0 is a cyclic extension
of degree d = [ F : F0 ]. Observe that d divides p − 1. We choose an integer ℓ such
that dℓ ≡ 1 mod p. Let σ0 be a generator of H := Gal( F/F0 ). Then σ0 (ξ ) = ξ e for an
e ∈ Z \ pZ.
Let χ1 , χ2 , χ3 be elements in Hom(GF0 , F p ) = H 1 (GF0 , F p ). We assume that χ1 , χ2 , χ3
are F p -linearly independent and χ1 ∪ χ2 = χ2 ∪ χ3 = 0. By [MT4, Lemma 2.6], the
homomorphism (χ1 , χ2 , χ3 ) : GF0 → (F p )3 is surjective. Let L0 be the fixed field of ( F0 )s
under the kernel of the surjection (χ1 , χ2 , χ3 ) : GF0 → (F p )3 . Then L0 /F0 is Galois
with Gal( L0 /F0 ) ≃ (F p )3 . We shall construct a Galois extension M0 /F0 such that
Gal( M0 /F0 ) ≃ U4 (F p ) and M0 contains L0 .
The restrictions res GF (χ1 ), res GF (χ2 ), res GF (χ3 ) are elements in Hom(GF , F p ). They
are F p -linearly independent and res GF (χ1 ) ∪ res GF (χ2 ) = res GF (χ2 ) ∪ res GF (χ3 ) = 0.
By Kummer theory there exist a, b, c in F× such that res GF (χ1 ) = χ a , res GF (χ2 ) = χb ,
res GF (χ3 ) = χc . Then we have (a, b) = (b, c) = 0 in H 2 (GF , F p ).
√
√ √
p
Let L = L0 (ξ ). Then L = F( p a, b, p c), and L/F is Galois with Gal( L/F) ≃
Gal( L0 /F0 ) ≃ (F p )3 .
Claim 1: L/F0 is Galois with Gal( L/F0 ) ≃ Gal( F/F0 ) × Gal( L/F).
Proof of Claim: Since L0 /F0 and F/F0 are Galois extensions of relatively prime degrees,
the claim follows.
d−1
We define Φ := ℓ[ ∑ ei σ0−i ] ∈ Z [ H ]. The group ring Z [ H ] acts on F in the obvious
i =0
way, and if we let H act trivially on L0 we get an action on L also. Then Φ determines a
map
Φ : L → L, x 7→ Φ( x ).
For convenience, we shall denote x̃ := Φ( x ).
The claim above implies that Φσ = σΦ for every σ ∈ Gal( L/F).
× )p.
Claim 2: We have ã = a modulo ( F× ) p ; b̃ = b modulo ( F× ) p , c̃ = c modulo ( F√
p
Proof of Claim: A similar
argument as√in the proof of Claim 1 shows that F( a)/F0 √
is
√
p
p
p
Galois with Gal( F( a)/F0 ) = Gal( F( a)/F) × Gal ( F/F0 ). Since both groups Gal
√( F( a )/F)
and Gal( F/F0 ) are cyclic and of coprime orders, we see that the extension F( p a)/F0
is cyclic. By Albert’s result (see [Alb, pages 209-211] and [Wat, Section 5]), we have
JÁN MINÁČ AND NGUYỄN DUY TÂN
20
i
i
σ0 a = ae modulo ( F× ) p . Hence for all integers i, σ0i (a) = ae mod ( F× ) p . Thus σ0−i (ae ) =
a mod ( F× ) p . Therefore, we have
#ℓ
#ℓ "
"
ã = Φ(a) =
d−1
i
∏ σ0−i (ae )
i =0
=
d−1
∏a
= adℓ = a mod ( F× ) p .
i =0
Similarly, we have b̃ = b modulo ( F× ) p , c̃ = c modulo ( F× ) p .
σ0 x̃
d
= σ0 ( x ℓ(1−e )/p ) p ∈ L p .
e
x̃
Proof of Claim: This follows from the following identity in the group ring Z [ H ],
Claim 3: For every x ∈ L, we have
d−1
(σ0 − e)( ∑ ei σ0−i ) = σ0 (1 − ed ) ≡ 0 mod p.
i =0
By our construction of Galois U4 (F p )-extensions over fields containing a primitive pth root of unity (see Subsection 3.1), we have α, γ, B, ..., A, C, δ such that if
we √
let M
:=
√
√
√
√
√
p
p
p
p
p
p
L( A, C, δ), then M/F is a Galois U4 (F p )-extension. We set M̃ := L( Ã, C̃, δ̃).
Claim 4: M̃/F is Galois with Gal( M̃/F) ≃ U4 (F p ).
Proof of Claim: Since Φ commutes with every σ ∈ Gal( L/F), this implies that M̃/F
is Galois. This, together with Claim 2, also implies that Gal( M̃/F) ≃ U4 (F p ) because the construction of M̃ over F is obtained in the same way as in the construction
of M, except that we replace the data { a, b, c, α, γ, B, ...} by their "tilde" counterparts
{ã, b̃, c̃, α̃, γ̃, B̃, ...}.
Claim 5: M̃/F0 is Galois with Gal( M̃/F0 ) ≃ Gal( M̃/F) × Gal( F/F0 ).
Proof of Claim: By Claim 3, we see that σ0 x̃ = x̃ e modulo ( L× ) p for every x̃ in the F p vector subspace W̃ ∗ of L× /( L× ) p generated by Ã, C̃, and δ̃. Hence W̃ ∗ is an F p [Gal( L/F0 )]module. Therefore M̃/F0 is Galois by Kummer theory.
We also have the following exact sequence of groups
1 → Gal( M̃/F) → Gal( M̃/F0 ) → Gal( F/F0 ) → 1.
Since |Gal( M̃/F)| and |Gal( F/F0 )| are coprime, the above sequence is split by SchurZassenhaus’s theorem. (See [Za, IV.7,Theorem 25].) The Galois group Gal( M̃/F0 ) is
the semidirect product of Gal( M̃/F) and H = Gal( F/F0 ), with H acting on Gal( M̃/F)
by conjugation. We need to show that this product is in fact direct, i.e., that the action of H on Gal( M̃/F) is trivial. Note that H has an order coprime to p, and H acts
trivially on Gal( L/F) (see Claim 1) which is the quotient of Gal( M̃/F) by its Frattini
subgroup. Then a result of P. Hall (see [Ha, Theorem 12.2.2]) implies that H acts trivially on Gal( M̃/F).
From the discussion above we obtain the following result.
CONSTRUCTION OF UNIPOTENT GALOIS EXTENSIONS AND MASSEY PRODUCTS
21
Theorem 3.9. Let the notation be as above. Let M0 be the fixed field of M̃ under the subgroup
of Gal( M̃/F0 ) which is isomorphic to Gal( F/F0 ). Then M0 /F0 is Galois with Gal( M0 /F0 ) ≃
Gal( M̃/F) ≃ U4 (F p ), and M0 contains L0 .
Proof. Claim 5 above implies that M0 /F0 is Galois with Gal( M0 /F0 ) ≃ Gal( M̃/F) ≃
U4 (F p ). Since H ≃ Gal( M̃/M0 ) acts trivially on L0 , we see that M0 contains L0 .
Let σ1 := σa | M0 , σ2 := σb | M0 and σ3 := σc | M0 . Then σ1 , σ2 and σ3 generate Gal( M0 /F0 ) ≃
U4 ( Fp ). We also have
χ1 (σ1 ) = 1, χ1 (σ2 ) = 0, χ1 (σ3 ) = 0;
χ2 (σ1 ) = 0, χ2 (σ2 ) = 1, χ2 (σ3 ) = 0;
χ3 (σ1 ) = 0, χ3 (σ2 ) = 0, χ3 (σ3 ) = 1.
(Note that for each i = 1, 2, 3, χi is trivial on Gal( M/M0 ), hence χi (σj ) makes sense for
every j = 1, 2, 3.) An explicit isomorphism ϕ : Gal( M0 /F0 ) → U4 (F p ) may be defined
as
1 0 0 0
1 0 0 0
1 1 0 0
0 1 0 0
, σ2 7→ 0 1 1 0 , σ3 7→ 0 1 0 0 .
σ1 7→
0 0 1 1
0 0 1 0
0 0 1 0
0 0 0 1
0 0 0 1
0 0 0 1
4. T HE
CONSTRUCTION OF
U4 (F p )- EXTENSIONS : T HE
CASE OF CHARACTERISTIC
p
In this section we assume that F is of characteristic p > 0. Although by a theorem of
Witt (see [Wi] and [Ko, Chapter 9, Section 9.1]), we know that the Galois group of the
maximal p-extension of F is a free pro-p- group, finding specific constructions of Galois
p-extensions over F can still be challenging. The following construction of an explicit
Galois extension M/F with Galois group U4 (F p ) is an analogue of the construction in
Subsection 3.1 when we assumed that a p-th root of unity is in F. However we find
the details interesting, and therefore for the convenience of the reader, we are including
them here. Observe that even the case of the explicit construction of Heisenberg extensions of degree p3 in characteristic p is of interest. In the case when F has characteristic
not p, the constructions of Heisenberg extensions of degree p3 are now classical, important tools in Galois theory. We did not find any such constructions in the literature in
the case of characteristic p. Nevertheless the construction in Subsection 4.2 seems to be
simple, useful and aesthetically pleasing. What is even more surprising is that the field
construction of Galois U4 (F p )-extensions over a field F of characteristic p in Subsection
4.3 is almost equally simple. We have to check more details to confirm the validity of this
construction, but the construction of the required Galois extension M itself, is remarkably simple. The possibility of choosing generators in such a straightforward manner
(as described in Theorem 4.3) is striking. It is interesting that the main construction in
Section 3 carries over with necessary modifications in the case of characteristic p.
22
JÁN MINÁČ AND NGUYỄN DUY TÂN
4.1. A brief review of Artin-Schreier theory. (For more details and the origin of this
beautiful theory, see [ASch].) Let F be a field of characteristic p > 0. Let ℘(X ) = X p − X
be the Artin-Schreier polynomial. For each a in F of characteristic p, we let θa be a root
of ℘(X ) = a. We also denote [ a] F to be the image of a in F/℘( F). For each subgroup
U of F/℘( F), let FU := F(θu : [u] F ∈ U ). Then the map U 7→ FU is a bijection between
subgroups of F/℘( F) and abelian extensions of F of exponent dividing p. There is a
pairing
Gal( FU /F) × U → F p ,
defined by hσ, ai = σ(θa ) − θa , which is independent of the choice of root θa . ArtinSchreier theory says that this pairing is non-degenerate.
Now assume that F/k is a finite Galois extension. The Galois group Gal( F/k) acts naturally on F/℘( F). As an easy exercise, one can show that such an extension FU , where
U is a subgroup of F/℘( F), is Galois over k if and only if U is actually an F p [Gal( F/k)]module.
4.2. Heisenberg extensions in characteristic p > 0. For each a ∈ F, let χ a ∈ Hom(GF , F p )
be the corresponding element associated with a via Artin-Schreier theory. Explicitly, χ a
is defined by
χ a (σ ) = σ (θ a ) − θ a .
Assume that a, b are elements in F, which are linearly independent modulo ℘( F). Let
K = F(θa , θb ). Then K/F is a Galois extension whose Galois group is generated by σa
and σb . Here σa (θb ) = θb , σa (θa ) = θa + 1; σb (θa ) = θa , σb (θb ) = θb + 1.
We set A = bθa . Then
σa ( A) = A + b, and σb ( A) = A.
Proposition 4.1. Let the notation be as above. Let L = K (θ A ). Then L/F is Galois whose
Galois group is isomorphic to U3 (F p ).
Proof. From σa ( A) − A = b ∈ ℘(K ), and σb ( A) = A, we see that σ( A) − A ∈ ℘(K )
for every σ ∈ Gal(K/F). This implies that the extension L := K (θ A )/F is Galois. Let
σ̃a ∈ Gal( L/F) (resp. σ̃b ∈ Gal( L/F)) be an extension of σa (resp. σb ). Since σb ( A) = A,
p
we have σ̃b (θ A ) = θ A + j, for some j ∈ F p . Hence σ̃b (θ A ) = θ A . This implies that σ̃b is
of order p.
On the other hand, we have
℘(σ̃a (θ A )) = σa ( A) = A + b.
Hence σ̃a (θ A ) = θ A + θb + i, for some i ∈ F p . Then
p
σ̃a (θ A ) = θ A + pθb + pi = θ A .
This implies that σ̃a is also of order p. We have
σ̃a σ̃b (θ A ) = σ̃a ( j + θ A ) = i + j + θ A + θb ,
σ̃b σ̃a (θ A ) = σ̃b (i + θ A + θb ) = i + j + θ A + 1 + θb .
CONSTRUCTION OF UNIPOTENT GALOIS EXTENSIONS AND MASSEY PRODUCTS
23
We set σ̃A := σ̃a σ̃b σ̃a−1 σ̃b−1 . Then
σ̃A (θ A ) = θ A − 1.
This implies that σ̃A is of order p and that Gal( L/F) is generated by σ̃a and σ̃b . We also
have
σ̃a σ̃A = σ̃A σ̃a , and σ̃b σ̃A = σ̃A σ̃b .
We can define an isomorphism ϕ : Gal( L/F)
1 1 0
1
σ̃a 7→ 0 1 0 , σ̃b 7→ 0
0 0 1
0
→ U3 (Z/pZ) by letting
0 0
1 0 1
1 1 , σ̃A 7→ 0 1 0 .
0 1
0 0 1
Note that [ L : F] = p3 . Hence there are exactly p extensions of σa ∈ Gal(K/F) to the
automorphisms in Gal( L/F) since [ L : K ] = p3 /p2 = p. Therefore for later use, we can
choose an extension of σa ∈ Gal(K/F), which we shall denote σa ∈ Gal( L/F) with a
slight abuse of notation, in such a way that σa (θ A ) = θ A + θb .
Remark 4.2. It is interesting to compare our choices of generators A of Heisenberg extensions over given bicyclic extensions in the case of characteristic p and the case when
the base field in fact contains a primitive p-th root of unity. See the proofs of Proposition 2.2 and the proposition above. Although the form of A in Proposition 2.2 is more
complicated than the strikingly simple choice above, the basic principle for the search
of a suitable A is the same in both cases. We
need to guarantee that σa ( A)/A ∈ (K × ) p
√
and σa ( A) − A ∈ ℘(K ) in order that K ( p A) and K (θ A ) are Galois over F. Further,
in order to guarantee that σa and σb will not commute, we want σa ( A)/A = bk p , for
some k ∈ K × , where σb acts trivially on k. Similarly in the characteristic p case we want
σa ( A) − A = b + ℘(k), where σb (k) = k. If char( F) = p, then this is always possible in a
simple way above with k even being 0. (See also [JLY, Appendix A.1, Example], where
the authors discuss a construction of quaternion Q8 -extensions over fields of characteristic 2.) In the case when F contains a primitive p-th root of unity, the search for a
suitable A leads to the norm condition NF ( √
p a ) /F (α ) = b. For some related considerations see [MS1, Section 2].
4.3. Construction of Galois U4 (F p )-extensions. We assume that we are given elements
a, b and c in F such that a, b and c are linearly independent modulo ℘( F). We shall construct a Galois U4 (F p )-extension M/F such that M contains F(θa , θb , θc ).
First we note that F(θa , θb , θc )/F is a Galois extension with Gal( F(θa , θb , θc )/F) generated by σa , σb , σc . Here
σa (θa ) = 1 + θa , σa (θb ) = θb , σa (θc ) = θc ;
σb (θa ) = θa , σb (θb ) = 1 + θb , σb (θc ) = θc ;
σc (θa ) = θa , σc (θb ) = θb , σc (θc ) = 1 + θc .
JÁN MINÁČ AND NGUYỄN DUY TÂN
24
Recall that A = bθa . We set C := bθc . We set δ := ( AC)/b = bθa θc ∈ E := F(θa , θc ).
Then we have
σa (δ) − δ = bσa (θa )σa (θc ) − bθa θc = b[σa (θa ) − θa ]θc = bθc = C,
σc (δ) − δ = bσc (θa )σc (θc ) − bθa θc = bθa [σc (θc ) − θc ] = bθa = A.
Finally set G := Gal(E/F).
Theorem 4.3. Let M := E(θδ , θ A , θC , θb ). Then M/F is a Galois extension, M contains
F(θa , θb , θc ), and Gal( M/F) ≃ U4 (F p ).
Proof. Let W ∗ be the F p -vector space in E/℘(E) generated by [b] E , [ A] E , [C] E and [δ] E .
Since
σc (δ) = δ + A,
σa (δ) = δ + C,
σa ( A) = A + b,
σc (C) = C + b,
we see that W ∗ is in fact an F p [ G ]-module. Hence M/F is a Galois extension by ArtinSchreier theory.
Claim: dimF p (W ∗ ) = 4. Hence [ L : F] = [ L : E][ E : F] = p4 p2 = p6 .
Proof of Claim: From our hypothesis that dimF p h[ a] F , [b] F , [c] F i = 3, we see that h[b] E i ≃
F p.
Clearly, h[b] E i ⊆ (W ∗ )G . From the relation
[σa ( A)] E = [ A] E + [b] E ,
we see that [ A] E is not in (W ∗ )G . Hence dimF p h[b] E , [ A] E i = 2.
From the relation
[σc (C)] E = [C] E + [b] E ,
we see that [C] E is not in (W ∗ )σc . But we have h[b] E , [ A] E i ⊆ (W ∗ )σc . Hence
dimF p h[b] E , [ A] E , [C] E i = 3.
Observe that the element (σa − 1)(σc − 1) annihilates the F p [ G ]-module h[b] E , [ A] E , [C] E i,
while
(σa − 1)(σc − 1)[δ] E = σa ([ A] E ) − [ A] E = [b] E ,
we see that
dimF p W ∗ = dimF p h[b] E , [ A] E , [C] E , [δ] E i = 4.
Let H a,b = F(θa , θ A , θb ) and H b,c = F(θc , θC , θb ). Let
N := H a,b H b,c = F(θa , θc , θb , θ A , θC ) = E(θb , θ A , θC ).
CONSTRUCTION OF UNIPOTENT GALOIS EXTENSIONS AND MASSEY PRODUCTS
25
Then N/F is a Galois extension of degree p5 . This is because Gal( N/E) is dual to the
F p [ G ]-submodule h[b] E , [ A] E , [C] E i via Artin-Schreier theory, and the proof of the claim
above shows that dimF p h[b] E , [ A] E , [C] E i = 3. We have the following commutative diagram
Gal( N/F)
//
Gal( H a,b /F)
Gal( H b,c /F)
//
Gal( F(θb )/F).
So we have a homomorphism η from Gal( N/F) to the pull-back Gal( H b,c /F) ×Gal ( F (θb )/F )
Gal( H a,b /F):
η : Gal( N/F) −→ Gal( H b,c /F) ×Gal( F (θb )/F ) Gal( H a,b /F),
which make the obvious diagram commute. We claim that η is injective. Indeed, let σ
be an element in ker η. Then σ | H a,b = 1 in Gal( H a,b /F), and σ | H b,c = 1 in Gal( H b,c /F).
Since N is the compositum of H a,b and H b,c , this implies that σ = 1, as desired.
Since |Gal( H b,c /F) ×Gal( F (θb )/F ) Gal( H a,b /F)| = p5 = |Gal( N/F)|, we see that η is
actually an isomorphism. As in the proof of Proposition 4.1, we can choose an extension
σa ∈ Gal( H a,b /F) of σa ∈ Gal( F(θa , θb )/F) in such a way that
σa (θ A ) = θ A + θb .
Since the square commutative diagram above is a pull-back, we can choose an extension
σa ∈ Gal( N/F) of σa ∈ Gal( H a,b /F) in such a way that
σa | H b,c = 1.
Now we can choose any extension σa ∈ Gal( M/F) of σa ∈ Gal( N/F). Then we have
σa (θ A ) = θ A + θb and σa | H b,c = 1.
Similarly, we can choose an extension σc ∈ Gal( M/F) of σc ∈ Gal( F(θb , θc )/F) in
such a way that
σc (θC ) = θC + θb , and σc | H a,b = 1.
We define σb ∈ Gal( M/E) to be the element which is dual to [b] E via Artin-Schreier
theory. In other words, we require that
σb (θb ) = 1 + θb ,
and σb acts trivially on θ A , θC and θδ . We consider σb as an element in Gal( M/F), then
it is clear that σb is an extension of σb ∈ Gal( F(θa , θb , θc )/F). Let W = Gal( M/E), and
let H = Gal( M/F), then we have the following exact sequence
1 → W → H → G → 1.
By Artin-Schreier theory, it follows that W is dual to W ∗ , and hence W ≃ (Z/pZ )4 . In
particular, we have | H | = p6 .
JÁN MINÁČ AND NGUYỄN DUY TÂN
26
Recall that from [BD, Theorem 1], we know that the group U4 (F p ) has a presentation with generators s1 , s2 , s3 subject to the relations (R) displayed in the proof of Theorem 3.7. Note that |Gal( M/F)| = p6 . So in order to show that Gal( M/F) ≃ U4 (F p ),
we shall show that σa , σb and σc generate Gal( M/F) and they satisfy these relations (R).
The proofs of the following claims are similar to the proofs of analogous claims in the
proof of Theorem 3.7. Therefore we shall omit them.
Claim:
Claim:
Claim:
Claim:
Claim:
Claim:
Claim:
Claim:
The elements σa , σb and σc generate Gal( M/F).
The order of σa is p.
The order of σb is p.
The order of σc is p.
[σa , σc ] = 1.
[σa , [σa , σb ]] = [σb , [σa , σb ]] = 1.
[σb , [σb , σc ]] = [σc , [σb , σc ]] = 1.
[[σa , σb ], [σb , σc ]] = 1.
An explicit isomorphism ϕ : Gal( M/F) → U4 (F p ) may be defined as
1
0
σa 7→
0
0
1
1
0
0
0
0
1
0
0
0
, σb 7→
0
1
1
0
0
0
0
1
0
0
0
1
1
0
5. T RIPLE M ASSEY
0
0
, σc 7→
0
1
1
0
0
0
0
1
0
0
0
0
1
0
0
0
.
1
1
PRODUCTS
Let G be a profinite group and p a prime number. We consider the finite field F p
as a trivial discrete G-module. Let C • = (C• (G, F p ), ∂, ∪) be the differential graded
algebra of inhomogeneous continuous cochains of G with coefficients in F p (see [NSW,
Ch. I, §2] and [MT1, Section 3]). For each i = 0, 1, 2, . . ., we write H i (G, F p ) for the
corresponding cohomology group. We denote by Z1 (G, F p ) the subgroup of C1 (G, F p )
consisting of all 1-cocycles. Because we use trivial action on the coefficients F p , we have
Z1 (G, F p ) = H 1 (G, F p ) = Hom(G, F p ). Let x, y, z be elements in H 1 (G, F p ). Assume
that
x ∪ y = y ∪ z = 0 ∈ H 2 (G, F p ).
In this case we say that the triple Massey product h x, y, zi is defined. Then there exist
cochains a12 and a23 in C1 (G, F p ) such that
∂a12 = x ∪ y and ∂a23 = y ∪ z,
CONSTRUCTION OF UNIPOTENT GALOIS EXTENSIONS AND MASSEY PRODUCTS
27
in C2 (G, F p ). Then we say that D := { x, y, z, a12 , a23 } is a defining system for the triple
Massey product h x, y, zi. Observe that
∂( x ∪ a23 + a12 ∪ z) = ∂x ∪ a23 − x ∪ ∂a23 + ∂a12 ∪ z − a12 ∪ ∂z
= 0 ∪ a23 − x ∪ (y ∪ z) + ( x ∪ y) ∪ z − a12 ∪ 0
= 0 ∈ C2 (G, F p ).
Therefore x ∪ a23 + a12 ∪ z is a 2-cocycle. We define the value h x, y, ziD of the triple
Massey product h x, y, zi with respect to the defining system D to be the cohomology
class [ x ∪ a23 + a12 ∪ z] in H 2 (G, F p ). The set of all values h x, y, ziD when D runs over
the set of all defining systems, is called the triple Massey product h x, y, zi ⊆ H 2 (G, F p ).
Note that we always have
h x, y, zi = h x, y, ziD + x ∪ H 1 (G, F p ) + H 1 (G, F p ) ∪ z.
We also have the following result.
Lemma 5.1. If the triple Massey products h x, y, zi and h x, y′ , zi are defined, then the triple
Massey product h x, y + y′ , zi is defined, and
h x, y + y′ , zi = h x, y, zi + h x, y′ , zi.
′ , a′ }) be a defining system for h x, y, zi
Proof. Let { x, y, z, a12 , a23 } (respectively { x, y′ , z, a12
23
′ , a + a′ } is a defining system for
(respectively h x, y′ , zi). Then{ x, y + y′ , z, a12 + a12
23
23
′
h x, y + y , zi. We also have
h x, y, zi + h x, y′ , zi =[ x ∪ a23 + a12 ∪ z] + x ∪ H 1 (G, F p ) + H 1 (G, F p ) ∪ z
′
′
+ [ x ∪ a23
+ a12
∪ z] + x ∪ H 1 (G, F p ) + H 1 (G, F p ) ∪ z
′
′
=[ x ∪ (a23 + a23
) + (a12 + a12
) ∪ z] + x ∪ H 1 (G, F p ) + H 1 (G, F p ) ∪ z
as desired.
=h x, y + y′ , zi,
The following lemma is a special case of a well-known fact (see [Fe, Lemma 6.2.4 (ii)])
but for the sake of convenience we provide its proof.
Lemma 5.2. If the triple Massey product h x, y, ziis defined, then for any λ ∈ F p the triple
Massey product h x, λy, zi is defined, and
h x, λy, zi ⊇ λh x, y, zi.
Proof. Let D = { x, y, z, a12 , a23 } be any defining system for h x, y, zi. Clearly D ′ :=
{ x, λy, z, λa12 , λa23 } is a defining system for h x, λy, zi, and
λh x, y, ziD = λ[ x ∪ a23 + a12 ∪ z] = [ x ∪ (λa23 ) + (λa12 ) ∪ z] = h x, λy, ziD′ .
Therefore λh x, y, zi ⊆ h x, λy, zi.
28
JÁN MINÁČ AND NGUYỄN DUY TÂN
A direct consequence of Theorems 3.7, 3.9 and 4.3, is the following result which
roughly says that every "non-degenerate" triple Massey product vanishes whenever it
is defined.
Proposition 5.3. Let F be an arbitrary field. Let χ1 , χ2 , χ3 be elements in Hom(GF , F p ). We
assume that χ1 , χ2 , χ3 are F p -linearly independent. If the triple Massey product hχ1 , χ2 , χ3 i is
defined, then it contains 0.
Proof. Let L be the fixed field of ( F)s under the kernel of the surjection (χ1 , χ2 , χ3 ) : GF →
(F p )3 . Then Theorems 3.7, 3.9 and 4.3 imply that L/F can be embedded in a Galois
U4 (F p )-extension M/F. Moreover there exist σ1 , σ2 , σ3 in Gal( M/F) such that they generate Gal( M/F), and
χ1 (σ1 ) = 1, χ1 (σ2 ) = 0, χ1 (σ3 ) = 0;
χ2 (σ1 ) = 0, χ2 (σ2 ) = 1, χ2 (σ3 ) = 0;
χ3 (σ1 ) = 0, χ3 (σ2 ) = 0, χ3 (σ3 ) = 1.
(Note that for each i = 1, 2, 3, χi is trivial on Gal( M/M0 ), hence χi (σj ) makes sense for
every j = 1, 2, 3.) An explicit isomorphism ϕ : Gal( M/F) → U4 (F p ) can be defined as
1 0 0 0
1 0 0 0
1 1 0 0
0 1 0 0
, σ2 7→ 0 1 1 0 , σ3 7→ 0 1 0 0 .
σ1 7→
0 0 1 1
0 0 1 0
0 0 1 0
0 0 0 1
0 0 0 1
0 0 0 1
ϕ
Let ρ be the composite homomorphism ρ : Gal F → Gal( M/F) ≃ U4 (F p ). Then one can
check that
ρ12 = χ1 , ρ23 = χ2 , ρ34 = χ3 .
(Since all the maps ρ, χ1 , χ2 , χ3 factor through Gal( M/F), it is enough to check these
equalities on elements σ1 , σ2 , σ3 .) This implies that h−χ1 , −χ2 , −χ3 i contains 0 by [Dwy,
Theorem 2.4]. Hence hχ1 , χ2 , χ3 i also contains 0.
For the sake of completeness we include the following proposition, which together
with Proposition 5.3, immediately yields a full new proof for a result which was first
proved by E. Matzri [Ma]. Matzri’s result says that defined triple Massey products vanish over all fields containing a primitive p-th root of unity. Alternative cohomological
proofs for Matzri’s result are in [EMa2] and [MT5]. Our new proof given in this section
of the crucial "non-degenerate" part of this result (see Proposition 5.3), which relies on
explicit constructions of U4 (F p )-extensions, is a very natural proof because of Dwyer’s
result [Dwy, Theorem 2.4]. Observe that in [MT5] we extended this result to all fields.
Proposition 5.4. Assume that dimF p h[ a] F , [b] F , [c] F i ≤ 2. Then if the triple Massey product
hχa , χb , χc i is defined, then it contains 0.
Proof. We can also assume that a, b and c are not in ( F× ) p . The case that p = 2, was
treated in [MT1]. So we shall assume that p > 2.
CONSTRUCTION OF UNIPOTENT GALOIS EXTENSIONS AND MASSEY PRODUCTS
29
Case 1: Assume that a and c are linearly dependent modulo ( F× ) p . This case is considered in [MT5, Proof of Theorem 4.10]. We include a proof here for the convenience of
the reader. Let ϕ = { ϕ ab , ϕbc } be a defining system for hχ a , χb , χc i. We have
resker χ a (hχ a , χb , χc i ϕ ) = res ker χ a (χ a ∪ ϕbc + ϕ ab ∪ χc )
= res ker χa (χa ) ∪ res ker χa ( ϕbc ) + res ker χa ( ϕ ab ) ∪ res ker χa (χc )
= 0 ∪ res ker χa ( ϕbc ) + res ker χa ( ϕ ab ) ∪ 0
= 0.
Then [Se1, Chapter XIV, Proposition 2], hχ a , χb , χc i ϕ = χ a ∪ χ x for some x ∈ F× . This
implies that hχ a , χb , χc i contains 0.
Case 2: Assume that a and c are linearly independent. Then [b] F is in h[ a] F , [c] F i. Hence
there exist λ, µ ∈ F p such that
χb = λχ a + µχc .
Then we have
hχa , χb , χc i = hχa , λχa , χc i + hχa , µχc , χc i ⊇ λhχa , χa , χc i + µhχa , χc , χc i.
(The equality follows from Lemma 5.1 and the inequality follows from Lemma 5.2.) By
[MT5, Theorem 5.9] (see also [MT5, Proof of Theorem 4.10, Case 2]), hχ a , χ a , χc i and
hχa , χc , χc i both contain 0. Hence hχa , χb , χc i also contains 0.
Theorem 5.5. Let p be an arbitrary prime and F any field. Then the following statements are
equivalent.
(1) There exist χ1 , χ2 , χ3 in Hom(GF , F p ) such that they are F p -linearly independent, and
if charF 6= p then χ1 ∪ χ2 = χ2 ∪ χ3 = 0.
(2) There exists a Galois extension M/F such that Gal( M/F) ≃ U4 (F p ).
Moreover, assume that (1) holds, and let L be the fixed field of ( F)s under the kernel of the
surjection (χ1 , χ2 , χ3 ) : GF → (F p )3 . Then in (2) we can construct M/F explicitly such that
L is embedded in M.
If F contains a primitive p-th root of unity, then the two above conditions are also equivalent
to the following condition.
√ √
√
p
(3) There exist a, b, c ∈ F× such that [ F( p a, b, p c) : F] = p3 and (a, b) = (b, c) = 0.
If F of characteristic p, then the two above conditions (1)-(2) are also equivalent to the following
condition.
(3’) There exist a, b, c ∈ F× such that [ F(θa , θb , θc ) : F] = p3 .
Proof. The implication that (1) implies (2), follows from Theorems 3.7, 3.9 and 4.3.
Now assume that (2) holds. Let ρ be the composite ρ : GF ։ Gal( M/F) ≃ U4 (F p ).
Let χ1 := ρ12 , χ2 := ρ23 and χ3 := ρ34 . Then χ1 , χ2 , χ3 are elements in Hom(GF , F p ),
JÁN MINÁČ AND NGUYỄN DUY TÂN
30
and (χ1 , χ2 , χ3 ) : GF → (F p )3 is surjective. This implies that χ1 , χ2 , χ3 are F p -linearly
independent by [MT4, Lemma 2.6].
On the other hand, since ρ is a group homomorphism, we see that
Therefore (1) holds.
χ1 ∪ χ2 = χ2 ∪ χ3 = 0.
Now we assume that F contains a primitive p-th root of unity. Note that for any a, b ∈
F× , χ a ∪ χb = 0, if and only if (a, b) = 0 (see Subsection 2.1). Then (1)
is equivalent to (3)
√
√
√
p
p
by Kummer theory in conjunction with an observation that [ F( a, b, p c) : F] = p3 , if
and only if χ a , χb , χc are F p -linearly independent.
Now we assume that F of characteristic p > 0. Then (1) is equivalent to (3’) by ArtinSchreier theory in conjunction with an observation that [ F(θa , θb , θc ) : F] = p3 , if and
only if χ a , χb , χc are F p -linearly independent.
R EFERENCES
[Alb]
[AS]
[ASch]
[BD]
[BT1]
[BT2]
[Co]
[CEM]
[CMS]
[DGMS]
[DMSS]
[Dwy]
[Ef]
[EMa1]
A. A. Albert, Modern Higher Algebra, University of Chicago Press, Chicago, 1937.
S. A. Amitsur and D. Saltman, Generic abelian crossed products and p-algebras, J. Algebra 51 (1)
(1978), 76-87.
E. Artin and O. Schreier, Eine Kennzeichnung der reell abgeschlossenen K orper, Abh. Math. Sem.
Univ. Hamburg 5 (1927) pp. 225-231. Reprinted in: Artin’s Collected Papers (Eds. S. Lang
and J. Tate), Springer-Verlag, New York, 1965, pp. 289-295.
D. K. Biss and S. Dasgupta, A presentation for the unipotent group over rings with identity, J.
Algebra 237 (2001), 691-707.
F. Bogomolov and Y. Tschinkel, Reconstruction of higher-dimensional function fields, Moscow
Math. Journal 11, no. 2 (2011), 185-204.
F. Bogomolov and Y. Tschinkel, Introduction to birational anabelian geometry, Current developments in algebraic geometry, 17-63, Math. Sci. Res. Inst. Publ., 59, Cambridge Univ. Press,
Cambridge, 2012.
I. G. Connell, Elementary generalizations of Hilbert’s theorem 90, Canad. Math. Bull. 8 (1965),
749-757.
S. K. Chebolu, I. Efrat and J. Mináč, Quotients of absolute Galois groups which determine the entire
Galois cohomology, Math. Ann. 352 (2012), no. 1, 205-221.
S. K. Chebolu, J. Mináč and A. Schultz, Galois p-groups and Galois modules, to appear in Rocky
Mountain J. Math., arXiv:1411.6495.
P. Deligne, P. Griffiths, J. Morgan and D. Sullivan, Real homotopy theory of Kähler manifolds,
Invent. Math. 29 (1975), 245-274.
R. Dwilewicz, J. Mináč, A. Schultz and J. Swallow, Hilbert 90 for biquadratic extensions, American Mathematical Monthly 114 (7) (2007), 577-587.
W. G. Dwyer, Homology, Massey products and maps between groups, J. Pure Appl. Algebra 6
(1975), no. 2, 177-190.
I. Efrat, The Zassenhaus filtration, Massey products, and representations of profinite groups, Adv.
Math. 263 (2014), 389-411.
I. Efrat and E. Matzri, Vanishing of Massey products and Brauer groups, Can. Math. Bull. 58
(2015), 730-740.
CONSTRUCTION OF UNIPOTENT GALOIS EXTENSIONS AND MASSEY PRODUCTS
[EMa2]
[EM1]
[EM2]
[Fe]
[Ga]
[GLMS]
[GS]
[Ha]
[HW]
[Je]
[JLY]
[Ko]
[La]
[M]
[MNg]
[Ma]
[McL]
[Me]
[Mi1]
[Mi2]
[MZ]
[MSp]
[MS1]
[MS2]
[MSS]
[MT1]
31
I. Efrat and E. Matzri, Triple Massey products and absolute Galois groups, to appear in J. Eur.
Math. Soc., arXiv:1412.7265.
I. Efrat and J. Mináč, On the descending central sequence of absolute Galois groups, Amer. J. Math.
133 (2011), no. 6, 1503-1532.
I. Efrat and J. Mináč, Galois groups and cohomological functors, to appear in Trans. Amer. Math.
Soc., DOI:http://dx.doi.org/10.1090/tran/6724, arXiv:1103.1508v1.
R. Fenn, Techniques of Geometric Topology, London Math. Soc. Lect. Notes 57, Cambridge 1983.
J. Gärtner, Higher Massey products in the cohomology of mild pro-p-groups J. Algebra 422 (2015),
788-820.
W. Gao, D. Leep, J. Mináč and T. L. Smith, Galois groups over nonrigid fields, Proceedings of
the International Conference on Valuation Theory and its Applications, Vol. II (Saskatoon,
SK, 1999), 61-77, Fields Inst. Commun., 33, Amer. Math. Soc., Providence, RI, 2003.
H. G. Grundman and T. L. Smith, Galois realizability of groups of order 64, Cent. Eur. J. Math.
8(5) (2010), 846-854.
M. Hall, The Theory of Groups, The Macmillan Company, New York, 1963.
M. Hopkins and K. Wickelgren, Splitting varieties for triple Massey products, J. Pure Appl.
Algebra 219 (2015), 1304-1319.
C. U. Jensen, Finite groups as Galois groups over arbitrary fields, Proceedings of the International
Conference on Algebra, Part 2 (Novosibirsk, 1989) Contemp. Math., vol. 131, Amer. Math.
Soc., Providence, RI, 1992, pp. 435-448.
C. U. Jensen, A. Ledet and N. Yui, Generic Polynomials: Constructive Aspects of the Inverse Galois
Problem, MSRI Publications Volume 45, Cambridge University Press, 2002.
H. Koch, Galois theory of p-extensions, Springer Monographs in Mathematics (2001).
J. Labute, Linking Numbers and the Tame Fontaine-Mazur Conjecture, Ann. Math. Québec (2014),
38: 61-71.
R. Massy, Construction de p-extensions galoisiennes d’un corps de caractéristique différente de p, J.
Algebra 109 (1987), no. 2, 508-535.
R. Massy, and T. Nguyen-Quang-Do, Plongement d’une extension de degré p2 dans une surextension non abélienne de degré p3 : étude locale-globale, J. Reine Angew. Math. 291 (1977), 149-161.
E. Matzri, Triple Massey products in Galois cohomology, preprint (2014), arXiv:1411.4146.
C. McLeman, p-tower groups over quadratic imaginary number fields, Ann. Sci. Math. Québec 32
(2008), no. 2, 199-209.
A. S. Merkurjev, Certain K-cohomology groups of Severi-Brauer varieties, Proc. Symp. Pure Math.
58.2 (1995), 319-331.
I. M. Michailov, Four non-abelian groups of order p4 as Galois groups, J. Algebra 307 (2007),
287-299.
I. M. Michailov, Galois realizability of groups of orders p5 and p6 , Cent. Eur. J. Math. 11(5) (2013),
910-923
I. M. Michailov and N. P. Ziapkov, On realizability of p-groups as Galois groups, Serdica Math.
J. 37 (2011), 173-210.
J. Mináč and M. Spira, Witt rings and Galois groups, Ann. of Math. (2) 144 (1996), no. 1, 35-60.
J. Mináč and J. Swallow, Galois module structure of pth-power classes of extensions of degree p,
Israel J. Math. 138 (2003), 29-42.
J. Mináč and J. Swallow, Galois embedding problems with cyclic quotient of order p, Israel J. Math.
145 (2005), 93-112.
J. Mináč, A. Schultz and J. Swallow, Automatic realizations of Galois groups with cyclic quotient
of order pn , J. Théor. Nombres Bordeaux 20 (2008), no. 2, 419-430.
J. Mináč and N. D. Tân, Triple Massey products and Galois theory, to appear in J. Eur. Math.
Soc., arXiv:1307.6624.
JÁN MINÁČ AND NGUYỄN DUY TÂN
32
[MT2]
[MT3]
[MT4]
[MT5]
[NSW]
[NQD]
[Pop]
[Sch]
[Se1]
[Se2]
[Se3]
[Sha]
[Vi]
[Voe1]
[Voe2]
[Wat]
[Wh]
[Wi]
[Za]
J. Mináč and N. D. Tân, The Kernel Unipotent Conjecture and Massey products on an odd rigid
field (with an appendix by I. Efrat, J. Mináč and N. D. Tân), Adv. Math. 273 (2015), 242-270.
J. Mináč and N. D. Tân, Triple Massey products over global fields, Doc. Math 20 (2015) 1467-1480.
J. Mináč and N. D. Tân, Counting Galois U4 (F p )-extensions using Massey products,
preprint(2014), arXiv:1408.2586.
J. Mináč and N. D. Tân, Triple Massey products vanish over all fields, to appear in J. London
Math. Soc, arXiv:1412.7611.
J. Neukirch, A. Schmidt and K. Wingberg, Cohomology of number fields, Grundlehren der
Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], 323,
Springer-Verlag, Berlin, 2000.
T. Nguyen Quang Do, Étude Kummerienne de la q-suite centrale descendante d’un group de Galois,
Publ. Math. Besançon. Algèbre et théorie des nombres, vol. 2012/2, Presses Univ. FrancheComté, Besançon, 2012, 123-139.
F. Pop, On the birational anabelian program initiated by Bogomolov I, Invent. Math. 187 (2012),
no. 3, 511-533.
A. Schultz, Parameterizing solutions to any Galois embedding problem over Z/pn Z with elementary p-abelian kernel, J. Algebra 411 (2014), 50-91.
J.-P. Serre, Local fields, translated from the French by M. J. Greenberg, Graduate Texts in
Mathematics, 67, Springer-Verlag, New York-Berlin, 1979.
J.-P. Serre, Galois cohomology, translated from the French by P. Ion and revised by the author, corrected reprint of the 1997 English edition, Springer Monographs in Mathematics,
Springer-Verlag, Berlin, 2002.
J.-P. Serre, Topics in Galois Theory, Second edition, Research Notes in Mathematics, vol. 1, A
K Peters Ltd., Wellesley, MA, 2008. With notes by Henri Darmon.
R. Sharifi, Twisted Heisenberg representations and local conductors, Ph.D. thesis, The University
of Chicago (1999).
F. R. Villegas, Relations between quadratic forms and certain Galois extensions, a manuscript, Ohio
State University, 1988, http://www.math.utexas.edu/users/villegas/osu.pdf.
V. Voevodsky, Motivic cohomology with Z/2-coefficients, Publ. Math. Inst. Hautes Études Sci.
No. 98 (2003), 59-104.
V. Voevodsky, On motivic cohomology with Z/l-coefficients, Ann. of Math. (2) 174 (2011), no. 1,
401-438.
W. C. Waterhouse, The normal closures of certain Kummer extensions, Canad. Math. Bull. 37
(1994), no. 1, 133-139.
G. Whaples, Algebraic extensions of arbitrary fields, Duke Math. J. 24 (1957), 201- 204.
E. Witt, Konstruktion von galoisschen Körpern der Charakteristik p zu vorgegebener Gruppe der
Ordnung p f , J. Reine Angew. Math. 174 (1936), 237-245.
H. Zassenhaus, The Theory of Groups, Chelsea Publishing Company, 1949.
D EPARTMENT OF M ATHEMATICS , W ESTERN U NIVERSITY, L ONDON , O NTARIO , C ANADA N6A 5B7
E-mail address: [email protected]
D EPARTMENT OF M ATHEMATICS , W ESTERN U NIVERSITY,
AND I NSTITUTE OF M ATHEMATICS , V IETNAM A CADEMY OF
Q UOC V IET, 10307, H ANOI - V IETNAM
E-mail address: [email protected]
L ONDON , O NTARIO , C ANADA N6A 5B7
S CIENCE AND T ECHNOLOGY, 18 H OANG
| 4math.GR
|
F -SINGULARITIES IN FAMILIES
arXiv:1305.1646v3 [math.AG] 3 Oct 2017
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
Abstract. We study the behavior of test ideals and F -singularities in families. In particular, we
obtain generic (and non-generic) restriction theorems for test ideals and non-F -pure ideals, which
imply for example openness of most of the F -singularity classes when the relative canonical sheaf is
Q-Cartier. Additionally, we study the global behavior of certain canonical linear systems (induced
by Frobenius) associated to adjoint line bundles, in families. As a consequence, we obtain some
positivity results for pushforwards of some adjoint line bundles and for certain subsheaves of these.
Contents
1. Introduction
1.1. Organization
1.2. Acknowledgements
2. Notation and setup
2.13. Composing maps
2.16. Base change of ϕ
2.23. Passing to V ∞ and other perfect points
3. Relative non-F-pure ideals
3.11. Iterated non-F -pure ideals versus the absolute non-F -pure ideal
3.25. Application to sharply F -pure singularities and HSL numbers in families
4. Relative test ideals
4.10. Relative test ideals vs absolute test ideals
4.16. Applications to strong F -regularity and divisor pairs
5. Relative test submodules, F -rationality and F -injectivity
5.1. The definition and basic properties of σn (X/V, ωX/V )
5.6. The definition and basic properties of τn (X/V, ωX/V )
5.12. Applications to families of F -injective and F -rational singularities
6. Global applications
6.1. Basic definitions
6.2. Auxiliary definition and stabilization
6.3. Base-change
6.4. Uniform stabilization
6.5. Global generation and semi-positivity
6.6. Relation to global canonical systems
Appendix A. Relative Serre’s condition
References
2
5
5
5
9
11
14
14
19
27
28
32
33
35
35
37
38
39
39
41
45
47
50
52
56
58
2000 Mathematics Subject Classification. 14B05, 13A35, 14D05, 14F18, 13B40.
Key words and phrases. non-F -pure ideal, test ideal, F -pure, strongly F -regular, F -singularities, flat families.
The second author was partially supported by a Fellowship from the Sloan Foundation as well as NSF grant DMS
#1064485, NSF FRG Grant DMS #1265261/1501115 and NSF CAREER Grant DMS #1252860/1501102.
The third author was partially supported by the NSF grants DMS #1068946/#1247354 and #1606414.
The material is based upon work supported by the National Science Foundation under Grant No. 0932078 000,
while the second and third authors were in residence at the Mathematical Science Research Institute (MSRI) in
Berkeley, California, during the spring semester 2013.
1
2
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
1. Introduction
In [Kun69] Kunz proved that a scheme over a field of characteristic p > 0 is regular if and only if
the Frobenius (endo)morphism is flat, which initiated the study of singularities using the Frobenius
morphism. Classes of singularities defined via the Frobenius morphism are referred to usually as
F -singularities. Since Kunz’s work, the study of F -singularities has become an active research area,
cf. [HR76, GW77, Fed83, MR85, RR85, Sri91, Smi97, Har98, MS97, HY03, Tak04]. However, the
methods of F -singularities have been widely applied to global geometry over a field of positive
characteristic only recently [Sch11a, Mus11, Hac11, MS12, Zha12, CHMS12, Pat12, HX13, Tan13],
at least outside of special classes of varieties [MR85, BK05]. A large field within global geometry is
moduli theory, which requires understanding how varieties behave in flat families. In this direction,
there have not been any positive results on the behavior of test ideals in families. We fill this gap.
Previously F -rational, Cohen-Macaulay F -injective and (Gorenstein) F -pure singularities have
been studied in such a context [Has01, SZ09, Has10]. Additionally, [MY09, Example 4.7] showed
that the test ideal τ does not satisfy the generic restriction theorem, at least as stated for multiplier
ideals [Laz04b, Theorem 9.5.35]. We develop tools tackling this issue, obtaining generic (and nongeneric) restriction theorems for test ideals, see Theorem A below.
In a flat family f : X −
→ V the absolute Frobenius on X does not restrict to the Frobenius
→
morphism on each fiber. Thus we systematically study the relative Frobenius morphism1 X ′ −
X ×V V ′ where X ′ −
→ X and V ′ −
→ V are the Frobenius morphisms of X and V , respectively. Its
fibers over V are morphisms between thickenings of the fibers of f . However, its fibers over V ′ are
exactly Frobenius morphisms of fibers of f (at least over points with perfect residue field).
We begin by stating our main result in the local setting. We denote by X n and V n the domains
of n-iterated Frobenii of X and V respectively. Let τ and σ denote the test ideal2 [HH94] and
non-sharply-F -pure ideal [FST11, BB11], respectively. These ideals play key roles in the theory of
F -singularities, the test ideal τ being the unique smallest non-zero ideal fixed by p−e -linear maps
and σ being the unique largest. We define relative versions of these two ideals by using the relative
instead of absolute Frobenius. The iterated relative Frobenius targets different spaces X ×V V n ,
and so we actually define a sequence ideals, one on each of these spaces. We explain the setup.
Suppose that f : X −
→ V is a finite-type flat, equidimensional, reduced and S2 and G1 morphism of Noetherian F -finite schemes with V integral. Additionally suppose ∆ is a Q-divisor on X
satisfying suitable conditions which allow it to be restricted to fibers and such that KX/V + ∆ is
Q-Cartier with index not divisible by p > 0 so that (pe −1) KX/V + ∆ is Cartier (see Remark 2.11
for a precise statement). Then for each integer n > 0, divisible by e, we define ideals σn (X/V, ∆)
and τn (X/V, ∆) ⊆ OX×V V n called the relative non F -pure and relative test ideals (respectively).
Our main local theorem is that these ideals restrict to absolute non-F -pure and absolute test
ideals on all of the geometric fibers, and can be used to prove generic restriction theorems for the
usual absolute test ideals on X ×V V n at least if V is regular.
Theorem A. (Corollary 3.10, Theorem 3.23, Corollary 4.8, Corollary 4.15) With notation as above,
there exists an N > 0 such that for every perfect point3 s ∈ V , and every n ≥ N
σne (X/V, ∆) · OXsne = σ(Xs , ∆|Xs )
and
τne (X/V, ∆) · OXsne = τ (Xs , ∆|Xs )
where Xsne is the fiber of X ×V V ne −→ V ne over sne ∈ V ne , which is isomorphic to Xs since k(s)
is perfect. Additionally, both σn and τne map surjectively onto their arbitrary base changes.
1Also called the Radu-André morphism, especially in commutative algebra.
2Technically, we are working with what has recently become known as the big test ideal [Hoc07].
3By definition, a morphism Spec K −
→ V from a perfect field K, see Definition 2.2.
F -SINGULARITIES IN FAMILIES
3
Furthermore, if V is regular and N is sufficiently large, then for all perfect points s ∈ V we have
that the absolute non-F -pure ideal restricts to all of the fibers for n ≥ N
σ(X ×V V ne , ∆ ×V V ne ) · OXsne = σ(Xs , ∆|Xs )
and at least for an open dense set of the base U ⊆ V that the same holds for the absolute test ideal
τ (X ×V V ne , ∆ ×V V ne ) · OXsne = τ (Xs , ∆|Xs )
for all perfect points s ∈ U .
Additionally, we show that over a dense open subset U of V with W = f −1 (U ), τne (X/V, ∆)|W
coincides with the absolute test ideal τ (X ×V V ne , ∆ ×V V ne )|W and likewise σne (X/V, ∆)|W =
σ(X ×V V ne , ∆ ×V V ne )|W in Theorem 4.13 and Theorem 3.14 respectively.
The key method that allows to prove this result
(at least stated for σn ) is that if m > n, then
→ OX×V me ⊇ σme (X/V, ∆) and there is a dense open set
Im σne (X/V, ∆) ⊗OX×V ne OX×V me −
of the base V over which equality holds for all sufficiently large n, this is Proposition 3.3. This
stabilization result should be viewed as a relative version of [HS77, Proposition 1.11], [Lyu97,
Proposition 4.4], and [Gab04, Remark 13.6].
Remark 1.1. If one replaces the divisor ∆ with the language of principal Cartier algebras, then the
previous result still holds without the technical assumptions about divisors from Remark 2.11.
Remark 1.2. There are a number of subtle issues in the statement above that we are suppressing.
In particular, τn (X/V, ∆) depends on the choice of some ideal I contained within the test ideal of
every fiber, and shown to exist in Proposition 4.7.
A particularly important related question is that of deformation of sharply F -pure singularities
in flat families with Q-Cartier relative canonical divisors. This would be important for a positive
characteristic construction of the moduli of stable varieties, also known as the KSBA compactification. In characteristic zero, this is the moduli space given by the log-minimal model program. It
classifies log-canonical models, hence birational equivalence classes of varieties of general type, and
furthermore it contains some nodal varieties for the compactification. There is a conjectural framework of constructing this moduli space [Kol90]. One of the main ingredients in this framework is
to prove that log-canonical singularity deforms in flat families with Q-Cartier relative log-canonical
divisor. An important step in this direction in positive characteristic is the corresponding statement
for sharply F -pure singularities. It is also an important ingredient in an upcoming paper of the
first author where he is planning to address the question of the existence of an algebraic space
structure on the space of sharply F -pure stable varieties. In this paper, we handle the deformation
of F -pure and F -regular singularities. Indeed, openness of F -pure and F -regular singularities is a
direct consequence of Theorem A above.
Theorem B. (Deformation of F -pure and F -regular singularities: Corollary 3.31, Corollary 4.21)
Suppose f : X −→ V and ∆ is as in Theorem A and additionally assume that f is proper. If s ∈ V is
a perfect point and (Xs , ∆|Xs ) is sharply F -pure (respectively, strongly F -regular4) then there exists
an open set U ⊆ V such that for all u ∈ U that (Xu , ∆|Xu ) is also sharply F -pure (respectively,
strongly F -regular).
We also build relative test submodules and non-F -injective submodules of ωX×V V ne /V ne and
prove restriction theorems like Theorem A for them Corollary 5.5 and Corollary 5.10. As a consequence in Theorem 5.16 we reprove a result of M. Hashimoto [Has01], deformation for CohenMacaulay F -injectivity and F -rationality.
Furthermore we apply our setup to global questions. One of the reasons for the recent global
applications of F -singularity theorem is the lifting theorem shown by the second author in [Sch11a,
4In which case, you can even remove the index not divisible by p assumption from K
X/V + ∆, see Corollary 4.22.
4
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
Proposition 5.3]. This theorem can be used to replace some of the lifting arguments that use Kodaira
vanishing in characteristic zero. One of the fundamental ideas in [Sch11a] is to try to lift only a big
enough set of sections of adjoint bundles instead of all the sections. This canonical set of sections
for a pair (X, ∆) with (1 − pe )(KX + ∆) Cartier and for a line bundle M is defined as [Sch11a,
Definition 4.1]
\
S 0 X, σ(X, ∆) ⊗ M :=
im H 0 X, F∗me OX ((1 − pme )(KX + ∆)) ⊗ M −
→ H 0 X, M .
m>0
First, we investigate questions about how this canonical space of sections behaves in families:
semicontinuity, stabilization of the intersection, etc.
Theorem C. Let f : X −→ V and ∆ be as in Theorem A, with f projective and V regular. Further,
suppose that M is a line bundle on X such that M − KX/V − ∆ is f -ample (where by M we actually
mean the Cartier divisor corresponding to M ). Then the following statements hold.
(a) [Corollary 6.19, Example 6.22, Example 6.23, Example 6.25], cf. [Har98, Example 5.5], [Tan13,
Theorem 8.3]. The function
(1.2.1)
s 7→ dimk(s) S 0 (Xs , σ(Xs , ∆s ) ⊗ Ms )
is not semicontinuous (here s ∈ V is a perfect point) in either direction, however, there is a
dense open subset U ∈ V , such that the function (1.2.1) is constant on U .
(b) [Theorem 6.18]. There exists n > 0 so that for all integers m ≥ n and perfect points s ∈ V ,
im H 0 (Xs , F∗me OXs ((1 − pme )(KXs + ∆s )) ⊗ Ms ) −→ H 0 (Xs , Ms )
= S 0 (Xs , σ(Xs , ∆s ) ⊗ Ms ).
(c) [Theorem 6.20]. If there is a perfect point s0 ∈ V such that
H 0 (Xs0 , Ms0 ) = S 0 (Xs0 , σ(Xs0 , ∆s0 ) ⊗ Ms0 ),
then there is an open neighborhood U of s0 , such that f∗ M |U is locally free and compatible
with base change and
H 0 (Xs , Ms ) = S 0 (Xs , σ(Xs , ∆s ) ⊗ Ms )
for every perfect point s ∈ U . In particular, the function (1.2.1) is constant for s ∈ U .
We would like to also mention the following natural question left open by Theorem C.
Question 1. Can one remove the f -ampleness assumption from the statements of Theorem C?
By Theorem C, it does make sense to talk about general value of dimk(s) S 0 (Xs , σ(Xs , ∆s ) ⊗ Ms )
in the following theorem.
Theorem D. With assumptions as in Theorem C such that V is projective over a perfect field and
0
(pe − 1)(KX + ∆) is Cartier. Then for every n ≫ 0 there is a subsheaf S∆,ne
f∗ (M ) of (FVne )∗ (f∗ M ),
for which the following holds.
0
f∗ (M ) is the general value of dimk(s) S 0 (Xs , σ(Xs , ∆s ) ⊗
(a) [Corollary 6.19]. The rank of S∆,ne
Ms ).
0
f∗ (M ) is globally generated for
(b) [Proposition 6.26]. If M − KX/V − ∆ is ample, then S∆,ne
every n ≫ 0.
0
(c) [Theorem 6.31]. If M − KX/V − ∆ is nef, then S∆,ne
f∗ (M ) is weakly-positive for n ≫ 0.
l
(d) [Corollary 6.14]. If M = Q ⊗ P where Q is f -ample, then for all l ≫ 0 and nef line bundle
0
P , we have that S∆,ne
f∗ (M ) is contained in (fV ne )∗ (σne (X/V, ∆) ⊗ MV ne ) as subsheaves of
(fV ne )∗ MV ne , and furthermore, these two subsheaves are generically equal.
In fact points (a) and (d) are true without the projectivity assumption on V .
F -SINGULARITIES IN FAMILIES
5
Further, note that if V in Theorem D is a curve, M − KX/V − ∆ is f -ample and nef and there
is a s ∈ V , such that H 0 (Xs , Ms ) = S 0 (Xs , σ(Xs , ∆s )), then points Question 1 and Question 1
of Theorem D with point (1) of Theorem C imply that f∗ OX (M ) is a nef vector bundle. This
strengthens [Pat12, Proposition 3.6], it removes the cohomology vanishing assumption made there.
Of course our statements are considerably stronger. For example with the same assumptions except
replacing H 0 (Xs , Ms ) = S 0 (Xs , σ(Xs , ∆s )) with rk f∗ (σ(X, ∆) ⊗ M ) equals the general value of
S 0 (Xs , σ(Xs , ∆s )), we obtain that f∗ (σ(X, ∆) ⊗ M ) being nef. In fact, the following even stronger
statement can be made.
Theorem E. [Corollary 6.39] With assumptions as in Theorem C such that V is projective over
a prefect field, assume that M − KX/V − ∆ is nef and f -ample and that rk S 0 f∗ (σ(X, ∆) ⊗ M )
equals the general value of H 0 (Xs , σ(Xs , ∆s ) ⊗ Ms ). Then S 0 f∗ (σ(X, ∆) ⊗ M ) is weakly positive.
In particular, if V is a smooth curve then it is a nef vector bundle.
0
Further, we show how S∆,ne
f∗ (M ) relates to the other similar notion S 0 f∗ (σ(X, ∆) ⊗ M ) introduced in [HX13, Definition 2.14]. In particular we obtain that S 0 f∗ (σ(X, ∆) ⊗ M ) does not restrict
to S 0 (Xs , σ(Xs , ∆s ) ⊗ Ms ) in general for general s ∈ S. Intuitively, though S 0 f∗ (σ(X, ∆) ⊗ M ) is
a pushforward, it captures the global geometry of (X, ∆) rather then the geometry of the fibers.
In positive characteristic these two can differ considerably, essentially because the function field of
V is not perfect. Of course the relative and absolute S 0 f∗ are related. We study these similarities
and differences in Section 6.6.
1.1. Organization. In Section 2, we set up notation that we will follow throughout the paper,
explain the interplay between p−e -linear maps and Q-Cartier divisors, and discuss the behaviors of
such maps and divisors under base changes. In Sections 3, 4, and 5, we introduce the relative nonF -pure ideals, the relative test ideals and the relative test submodules, respectively. Applications
to F -singularities in families are discussed in these 3 sections. Section 6 is devoted to the behaviors
of S 0 (introduced in [Sch11a, Definition 4.1]) under base changes. Some semi-positivity results are
also proved in this section. Finally, in the Appendix, we collect some statements that we can not
find proper references for their generality but are needed in our paper.
1.2. Acknowledgements. The authors began working on this project while attending a workshop at the American Institute of Mathematics (AIM) titled “ACC for minimal log discrepancies
and termination of flips”, May 14th – 18th 2012, organized by Tommaso de Fernex and Christopher Hacon. We would like to thank the organizers and the American Institute of Mathematics.
Part of the work was done when the third author visited the Department of Mathematics at
Pennsylvania State University; he would like to thank them for their hospitality. Additionally,
the authors would like to thank Lawrence Ein, Christopher Hacon, János Kollár, Joseph Lipman,
Mircea Mustaţă, and Kevin Tucker for useful discussions. Especially, we would like to thank Brian
Conrad for pointing out compatibility of trace with base change. We would also like to thank
Hiromu Tanaka for comments on a previous draft. Additionally, the authors of the paper would
like to thank Leo Alonso, Damian Rössler, Graham Leuschke and Blup for providing answers
and commentary to the following questions http://mathoverflow.net/questions/120982/ and
http://mathoverflow.net/questions/120625/.
2. Notation and setup
Throughout this paper, all schemes are Noetherian and all maps of schemes are separated. We
fix the following notation, which is in effect for the entire paper. In particular, for simplicity we do
not state it in every statement even though it is assumed.
6
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
Notation 2.1. Suppose that f : X −
→ V is a flat, equidimensional and geometrically reduced
map5 of finite type from a scheme X to an excellent integral scheme V of equal characteristic p > 0
with a dualizing complex. We write F e = FVe : V = V e −
→ V as the absolute e-iterated Frobenius
e
on V , form the base change fV e : XV e = X ×V V −
→ V e and define FXe e /V e : X e −
→ XV e to be
the e-iterated relative Frobenius. Furthermore, we often assume that V is F -finite in which case we
q
automatically assume that Frobenius FV : V −
→ V satisfies the following identity6, (FV )! ωV ≃qis
q
ωV . If we say that V is a variety, it is always of finite type over a perfect field k.
Because we will be considering numerous different sheaves on the same topological space X =
X e = X ×V V e = etc., but with respect to different schemes, we will adopt the
followingsomeand
what nonstandard notation. We use these because otherwise writing numerous FXe e /V e
∗
−1
operations, which do nothing to the underlying space, is confusing. We would need
FXe e /V e
notations for projections of the form X e ×V e V e+d −
→ V e+d , X e in as well as more general relative
e
Frobenii F(X
: (X e )V e+d := X e ×V e V e+d −
→ XV e+d , and maps X e ×V e+d+c −
→ X e ×V e+d .
e)
/V e+d
V e+d
By using simply modules, this becomes more transparent.
(1) We will use R to denote OX .
(2) We will use A to denote f −1 OV .
e
(3) We will use A1/p to denote f −1 (FVe )∗ OV . We note that this is not an abuse of notation
since V is integral.
e
(4) We will use R1/p to denote (FXe )∗ R. This is a slight abuse of notation if X is not reduced.
e
(5) We will use RA1/pe to denote R ⊗A A1/p .
d
e+d
d
(6) We will use (RA1/pe )1/p to denote R1/p ⊗A1/pd A1/p . This may be a slight abuse of
notation if XV e is not reduced.
e
(7) Given R-module M , we will use M 1/p to denote F∗e M . This will generally not cause any
confusion because typically M will be locally free or even a line bundle.
(8) We use ωR to denote ωX and ωA to denote f −1 ωV .
(9) Etc.
Some of the main results of the paper concern restriction to fibers. These statements pertain only
to a special set of fibers of f , the fibers over perfect points (see definition below). For example, if
V is a curve over an algebraically closed field, then often we restrict to fibers over all closed points
and the perfect closure of the generic point of V , but not the generic point itself.
Definition 2.2 (Perfect points). A perfect point of V , s ∈ V is a morphism from the spectrum of
a perfect field k(s) to V , s = Spec k(s) −
→ V . It can also be viewed as a choice of a point v ∈ V
and a field extension k(s) = K ⊇ k(v) such that K is perfect. Finally, a neighborhood of a perfect
point s ∈ V is simply a neighborhood of the image v of s.
The fact that f : X −
→ V is of finite type implies that the relative Frobenius is a finite map. In
some cases, this will allow us to avoid assuming that V is F -finite.
Lemma 2.3. With notation as above, since f : X −→ V is of finite type, the relative Frobenius
map f : X e −→ XV e is a finite map.
e
Proof. We work locally on V and X. It is sufficient to show that R1/p is a finite RA1/pe -module.
e
→ S 1/p is a finite map.
Write S = A[x1 , . . . , xn ] and R = S/I. We first observe that SA1/pe −
e
However, xi11 · · · xinn for 0 ≤ ij < pe clearly form a generating set for S 1/p over SA1/pe .
5Meaning X = X × T is reduced for all T −
→ V with T integral.
T
V
6For some discussion of this identity, which always holds for varieties or schemes of essentially finite type over a
local ring with a dualizing complex, see [BSTZ10, Section 2].
F -SINGULARITIES IN FAMILIES
7
e
→ S 1/p with ⊗S (S/I) = ⊗S R. We obtain
Now, tensor the map SA1/pe −
e
e
e
e
→ (S/I) ⊗S S 1/p = (S/I [p ] )1/p −
→ (S/I)1/p
RA1/pe = R ⊗A A1/p = (S/I) ⊗S SA1/pe −
e
where the final map is the canonical surjection of rings. The map is finite since each part is.
We also recall the result of Radu and André.
Theorem 2.4. [Rad92, And93] Suppose that f : X −→ V is a flat map of Noetherian schemes. Then
e
e
f has geometrically regular fibers if and only if the relative Frobenius RA1/pe ∼
= A1/p ⊗A R −→ R1/p
is flat.
We immediately obtain the following corollary which will also be useful.
Corollary 2.5. Suppose that f : X −→ V is as in Notation 2.1 and additionally has geometrically
regular fibers. Then for any RA1/pe -module M , the natural evaluation-at-1 map below surjects
e
H omR
e
A1/p
(R1/p , M ) −→ M
e
Proof. R1/p is a locally free RA1/pe -module, and the result follows.
Now we state the object which we will study for the majority of the paper.
Definition 2.6 (ϕ). From here on, we will fix a line bundle L on X ∼
= X e and we will also fix a
e
1/p
−
→ RA1/pe .
(possibly zero) RA1/pe -linear map ϕ : L
Remark 2.7 (Reflexive sheaves for G1 and S2 Morphisms). Suppose that f : X −
→ V is G1 + S2.
Note that there exists an open set ι : U ֒→ X such that X \ U has codimension ≥ 2 along each
fiber and that f |U is a Gorenstein morphism.
Finally, suppose that M is any rank-1 reflexive R-module which is locally free on a set U as
e
e
above. Then M 1/p is not only reflexive as an R1/p -module, we claim it is also reflexive as an
RA1/pe -module. Since i∗ M |U = M already by Proposition A.7 because M is reflexive, it is sufficient
to replace X by U . Thus ωX/V and M are both locally free as R-modules and ωXV e is locally free
as an RA1/pe -module. We work locally so as to trivialize all these modules. Then
M 1/p
e
∼
=
=
∼
=
∼
=
1/pe
ωX/V
(FXe e /V e )∗ ωX/V
H omOXV e ((FXe e /V e )∗ OX , ωXVe /V e )
e
H omR 1/pe (R1/p , RA1/pe )
A
which is clearly reflexive (the second isomorphism follows from Grothendieck duality for the finite
relative Frobenius map).
Conversely, if M is any R-module which is locally free on a set U and reflexive as an RA1/pe e
module, then it is also reflexive as an R1/p -module. To see this, note that M |U is reflexive as an
R-module, and because M is reflexive as an RA1/pe -module it satisfies i∗ M |U = M .
8
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
Definition 2.8 (ϕ versus divisors). Observe the following identifications:
e
H omR 1/pe L1/p , RA1/pe
A
∗∗
e
∼
H omR 1/pe L1/p ⊗R 1/pe (ωR 1/pe /A1/pe ), ωR 1/pe /A1/pe
=
A
A
A
|
{z A
}
This
clearly
holds
on
U
,
note
both
sheaves
are
reflexive
and
use
Corollary
A.8.
∗∗
=
H omOXV e (FXe e /V e )∗ L ⊗ ((FXe e /V e )∗ ωXV e /V e ) , ωXV e /V e
{z
}
|
(2.8.1)
This just rewrites the
previous line using different notation. ∗∗
∼
(FXe e /V e )∗ H omOX e L ⊗ ((FXe e /V e )∗ ωXV e /V e ), ωX e /V e
=
|
{z
}
Grothendieck duality for a finite map
∗∗
−1
∼
(FXe e /V e )∗ L−1 ⊗ ωX e /V e ⊗ (FXe e /V e )∗ ωX
=
e
V e /V
|
{z
}
e
By Remark 2.7 and Corollary A.8 we may take reflexive hull as (FX
e /V e )∗ OX e -modules.
Now, observe that ωX/V is compatible with base change up to reflexification. In particular, if our
→ X as the
base change is the Frobenius FVe : V e −
→ V , then writing πV e : XV e = X ×V V e −
∗
∼
projection, we have ωXV e /V e = (πV e ωX/V )∗∗ since both sheaves are reflexive and they certainly
agree outside the non-relatively Cohen-Macaulay locus (which is of relative codimension at least 2)
by Corollary A.8. In particular
∗∗
∗∗
∗∗
e
∼
= (F e )∗ ωX/V )∗∗ = ω p
(F e e e )∗ ωX e /V e
.
= (F e e e )∗ π ∗ e ωX/V )
X /V
V
X /V
V
X
X/V
Plugging this into (2.8.1) we obtain
e
L1/p , RA1/pe
∗∗
1−pe
∼
= (FXe e /V e )∗ L−1 ⊗ ωX e /V e
H omR
e
A1/p
(2.8.2)
If additionally X is absolutely (instead of relatively) G1 and S2 (for example, if V is regular),
7
ϕ induces a non-zero, effective Weil divisorial sheaf Dϕ such that
then any choice
∗∗
of nondegenerate
e
1−p
−1
∼ L ⊗ω e e
by [Har94]. We would like to generalize this to the case that X −
OX (Dϕ ) =
→V
X /V
is relatively G1 and S2.
Definition 2.9. We say that ϕ is relatively divisorial if ϕ locally generates H omR
e
e
A1/p
e
(L1/p , RA1/pe )
as an R1/p -module at
(a) the generic points of each fiber and
(b) at the generic point of every codimension-1 singular point of every geometric fiber.
In this case, by removing a set of relative codimension 2 so that f is relatively Gorenstein, we
e
e
e
see that ϕ · R1/p ⊆ H omR 1/pe (L1/p , RA1/pe ) is a rank-1 free submodule of an invertible R1/p A
module. To be able to associate a divisor to this submodule in a sensible way, we should show that
it is the trivial (full) submodule at every singular codimension one point. Indeed, let ξ be a singular
codimension 1 point. Then one of the following cases hold.
◦ f (ξ) is a codimension 1 point. In this case, ξ is a general point of the fibers over f (ξ), hence
assumption (a) guarantees that ϕ generates a full submodule at ξ.
◦ f (ξ) is the general point of V . In this case ξ is a codimension 1 point of the fiber over
f (ξ), and it is not in the smooth locus of f . Therefore, assumption (b) shows that again ϕ
generates a full submodule at ξ.
e
Therefore, the submodule ϕ · R1/p determines a Cartier divisor and also an honest Weil divisor on
the original X. We denote this divisor Dϕ as well.
7Nonzero on any irreducible component.
F -SINGULARITIES IN FAMILIES
9
Definition 2.10 (ϕ as a divisor). If ϕ is relatively divisorial, we set ∆ϕ to be the Q-divisor pe1−1 Dϕ .
This makes sense because Dϕ is trivial along the codimension-1 components of the singular locus
of X and so we avoid the pathological issues which occur for Q-divisors on non-normal spaces.
We now explain how to recover ϕ from a Q-divisor.
Remark 2.11 (Obtaining ϕ from divisors). We work under the conventions of Definition 2.10. Untangling Definition 2.8 yields a method to obtain ϕ from a divisor ∆ ≥ 0 (which then coincides
with ∆ϕ ) under the following assumptions:
1
D for some Weil divisor D where p 6 | m.
(a) ∆ = m
(b) D is a Weil divisor on X which is Cartier in relative codimension 1.
(c) Supp D does not contain any general point or any singular codimension one point of any
geometric fiber.
∗∗
1−pe
= L is a line bundle.
⊗ OX ((1 − pe )∆)
(d) (pe − 1)/m ∈ Z and ωX/V
In such a case, the integral divisor (pe − 1)∆ (which is well defined since Supp ∆ does not contain
the codimension-1 components of the non-regular locus of X, cf. [Kc92, Pages 171–173] or [MS12,
Section 2.2]) induces an inclusion
1−pe ∗∗
) .
(FXe e /V e )∗ L ֒→ (FXe e /V e )∗ (ωX/V
This composed with the natural Grothendieck trace map
1−pe ∗∗
)
−
→ OXV e
(FXe e /V e )∗ (ωX/V
yields a map ϕ. It is easy to see that conditions (a)–(d) above guarantee that ϕ is relatively
divisorial. Furthermore, we also have that ∆ = ∆ϕ .
For future reference we make the following definition.
Definition 2.12. In the situation of Notation 2.1, (X, ∆) is a pair, if f is G1 + S2 and ∆ satisfies
the assumptions of Remark 2.11.
e
→ RA1/pe with
2.13. Composing maps. Given ϕ as in Definition 2.8, we can compose ϕ : L1/p −
itself (after twisting) similar to [BS13, Section 4] or [Sch09] and thus obtain new maps
1/p2e
2
(pe +1)
, RA1/p2e
ϕ ∈ HomR 1/p2e
L
A
and more generally
n
ϕ ∈ HomR
ne
A1/p
L
pne −1
pe −1
1/pne
, RA1/pne
!
.
We explain this construction.
Begin by tensoring ϕ by L over R, and then taking 1/pe th roots we obtain:
1/pe
1/pe
2e
e
e
e 1/p2e
e
= L1/p ⊗A1/pe A1/p
−
→ (R ⊗A A1/p ) ⊗R L
(2.13.1)
L1+p
= L1/p ⊗R L
On the other hand, we can also tensor ϕ by A1/p
(2.13.2)
e
2e
L1/p ⊗A1/pe A1/p
e
over A1/p to obtain:
2e
−
→ RA1/p2e
By composing (2.13.1) and (2.13.2) we obtain the desired map ϕ2 . We now define ϕn inductively
(n−1)e 1/p(n−1)e
p
−1
n−1
−
→ RA1/p(n−1)e tensor with L over R and then take
as follows. Given ϕ
: L pe −1
10
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
pe th roots which yields
=
L
pne −1
pe −1
L
1/pne
p(n−1)e −1
+p(n−1)e
pe −1
=
L ⊗R L
p(n−1)e −1
pe −1
1/p(n−1)e !1/pe
1/p(n−1)e !1/pe
1/pe
L ⊗R RA1/p(n−1)e
−
→
ne
e
= L1/p ⊗A1/pe A1/p .
We then apply ϕ to the first term in the final tensor product to obtain:
ne 1/pne
p −1
ne
ne ϕ⊗...
e
e
−
→ L1/p ⊗A1/pe A1/p −−−→ R ⊗A A1/p ⊗A1/pe A1/p ∼
L pe −1
= RA1/pne
which we take as the official definition of ϕn . On the other hand, for every 0 < m < n one can look
at the following composition of ϕm and ϕn−m
(2.13.3)
1
! (n−m)e
1
1
1
ne ne
me
ne
me me
p
p
p
p
p −1
p
−1
pne −pme
p
−1
p(n−m)e −1
∼
∼
γ : L pe −1
L pe −1
⊗R L pe −1
= L pe −1 ⊗R L pe −1
=
(ϕm ⊗idL··· )
−−−−−−−→
1
p(n−m)e
1
1
(n−m)e
(n−m)e (n−m)e
p(n−m)e −1
p
−1
1
p
p
me
e
e
∼
R ⊗A A p
⊗R L p −1
⊗
= L p −1
ϕn−m ⊗A··· A···
−−−−−−−−−→
R ⊗A A
1
p(n−m)e
1
⊗
A
1
p(n−m)e
1
A
1
p(n−m)e
A pne
1
A pne ∼
= R ⊗A A pne
In particular, taking m = 1 gives an a priori different map which we could also define as ϕn . We
now explain why this map is actually equal to the official ϕn .
Lemma 2.14. With notation as above, γ = ϕn .
Proof. The statement is local, and so we may suppose that L = R. With the (obscuring) powers of
L removed, ϕn is described as the following composition:
R1/p
ϕ1/p
(n−1)e
ne
(n−1)e
ne
⊗A1/p(n−1)e A1/p
−−−−−−−→ R1/p
(n−2)e
ne
(n−1)e
(n−2)e
ϕ1/p
⊗...
⊗A1/p(n−1)e A1/p
⊗A1/p(n−2)e A1/p
−−−−−−−−−→
R1/p
ϕ1/p
(n−3)e
⊗...
−−−−−−−−−→ · · ·
··· ···
e
ϕ1/p ⊗...
e
−−−−−−→ R1/p ⊗A1/pe . . . ⊗A1/p(n−1)e A1/p
ϕ⊗...
ne
−−−→ R ⊗A . . . ⊗A1/p(n−1)e A1/p .
ne
F -SINGULARITIES IN FAMILIES
11
1
The first m entries in the composition make up (ϕm ⊗ . . .) p(n−m)e in (2.13.3) and the last n − m
n
entries clearly yield ϕn−m ⊗A1/p(n−m)e A1/p as desired. The result is then obvious.
Lemma 2.15. With notation as in Definition 2.8 and additionally that ϕ is relatively divisorial,
then ϕn is relatively divisorial and ∆ϕ = ∆ϕn for every integer n ≥ 1.
Proof. For showing any of the two statements we may remove the non relatively Gorenstein locus.
That is, by possibly further restricting R we may assume that R is relatively Gorenstein over V
and that L is trivial (since our map f : X −
→ V is geometrically G1+S2). Hence by Remark 2.7,
HomR⊗
ie
AA
1/pie
ie
ie
(R1/p , R ⊗A A1/p ) ∼
= R1/p for all i.
From [Kun86, Appendix F], [Sch09, Lemma 3.9], or composition of Grothendieck trace, we have
ne !
e
ne
HomR1/pe ⊗ 1/pe A1/pne R1/p , R1/p ⊗A1/pe A1/p
A
(2.15.1)
ne
e
⊗R1/pe ⊗ 1/pe A1/pne HomR 1/pne R1/p ⊗A1/pe A1/p , RA1/pne
A
A
ne
∼
= HomR 1/pne R1/p , RA1/pne
A
via the homomorphism induced by composition. Let ϑe be the R
HomR⊗A A1/pe (R
1/pe
1/pe
, R ⊗A A
1/pe
-module generator of
)
for each e and let ϑne be defined as ϕn , but ϕ replaced by ϑ. Then by using (2.15.1) iteratively,
ϑne = ϑne up to multiplication by a unit.
e
Further let r ∈ R such that ϕ(−) = ϑe r 1/p · − . Then it is easy to verify that
ne 1/pne !
p −1
n
(2.15.2)
ϕ = ϑie
r pe −1
·
Then we see that if ϕ was generating at a point P ∈ X, or equivalently r is a unit at P , then so is
ϕn . This shows that ϕn is relatively divisorial. Furthermore (2.15.2) shows that ∆ϕi = ∆ϕ .
2.16. Base change of ϕ. Suppose that we are given g : T −
→ V any morphism of schemes such
that T is also excellent, integral and has a dualizing complex. For example, we could set T to be
a closed point of V and let g be the inclusion. Alternately, we could let g be a regular alteration
over some closed subscheme of V . We list the following maps:
(2.16.3)
p1
i
1/p
)
: X ×V T
−
→X
i
i
−
→ Xi
: X ×V i T
(p1
i
qi : X ×V T
−
→ X ×V V i
p1 = q 0 : X × V T
−
→ X.
(the projection).
(the projection for any i).
(base change)
These are pictured below.
1/pi
Xi
×V i
Ti
p1
/ Xi
Fi
(XT )i /T i
Fi
X i /V i
X ×V T i
qi
/ X ×V V i
e
→ RA1/pe = OXV e , we can form (qe )∗ ϕ which we denote by
Notice that given ϕ : L1/p −
1/pe
ϕT : LT
e
∼
→ qe∗ OXV e = OXT e .
= qe∗ L1/p −
12
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
e
e
1/p
We explain the isomorphism LT ∼
= qe∗ L1/p briefly. Working locally, let V = Spec A, T = Spec B
and X = Spec R. Then the map ϕT is identified with the map:
e
L1/p ⊗R
e
A1/p
e
e
e
e
e
→ ((R ⊗A A1/p ) ⊗A1/pe B 1/p ) ∼
RB 1/pe ∼
= R ⊗A B 1/p .
= L1/p ⊗A1/pe B 1/p −
The isomorphism in the definition of ϕT is now immediate.
The next lemma shows that base change of ϕ commutes with the self-composition defined in
Section 2.13.
Lemma 2.17. Suppose that g : T −→ V is as above. Then (ϕn )T = (ϕT )n .
Proof. It is sufficient to prove this in the affine case assuming our isomorphism is sufficiently
canonical (which will be clear). We notice that (ϕT )2 is the composition:
e
e
(Lp +1 )1/p
⊗A1/pe
e 1/pe
B 1/p
(ϕT ⊗R
B 1/p
e
LT )1/p
e
e
2e
e
−−−−−−−−−−−−−→ L1/p ⊗A1/pe B 1/p ⊗B 1/pe B 1/p
2e
ϕT ⊗ 1/pe B 1/p
2e
e
e
−−−−B−−−−−−−→
(R ⊗A A1/p ) ⊗A1/pe B 1/p ⊗B 1/pe B 1/p
Now unraveling the definitions we obtain
1/pe
2e
◦ (ϕT ⊗B 1/pe B 1/p )
(ϕT )2 =
ϕT ⊗R 1/pe LT
B
2e
e
e 1/pe
e
◦ (ϕ ⊗A1/pe B 1/p ⊗B 1/pe B 1/p )
=
(ϕ ⊗A1/pe B 1/p ) ⊗R 1/pe (L ⊗A1/pe B 1/p )
B
2e
e 1/pe
◦ (ϕ ⊗A1/pe B 1/p )
=
ϕ ⊗R 1/pe L ⊗A1/pe B 1/p
A
1/pe
2e
2e
2e
◦ (ϕ ⊗A1/pe A1/p ) ⊗A1/p2e B 1/p
=
⊗A1/p2e B 1/p
ϕ ⊗R 1/pe L
A
1/pe
2e
2e
=
◦ (ϕ ⊗A1/pe A1/p ) ⊗A1/p2e B 1/p
ϕ ⊗R 1/pe L
A
= (ϕ2 )T
as desired. The general nth self composition is similar.
Our next goal is to describe how divisors, corresponding to maps ϕ, fare under base change. We
thank Brian Conrad for pointing us in the right direction – [Con00, Theorem 3.61].
Lemma 2.18. Suppose that f : X −→ V is a finite type Cohen-Macaulay morphism over V an
excellent scheme of characteristic p > 0 and that g : T −→ V is as above. Then the Grothendiecktrace map
(FX e /V e )∗ ωX e /V e ∼
= H omOXV e ((FX e /V e )∗ OX e , ωXV e /V e ) −→ ωXV e /V e
is compatible with base change.
Proof. We clearly have a commutative diagram:
qe∗ H omOXV e ((FX e /V e )∗ OX e , ωXV e /V e )
/ q ∗ ωX e /V e
e
V
α
β
H omOXT e ((F(XT )e /T e )∗ O(XT )e , ωXT e /T e )
/ ωω
X
T e /T
e
and by [Con00, Theorem 3.6.1] the map β is an isomorphism. It is sufficient to verify that α
is an isomorphism as well. We work locally on some affine chart on X and hence assume that
XV ⊆ AN
V =: PV embeds as a closed subscheme since f is of finite type. The map α can then be
identified with
Q∗e E xtjOP
Ve
((FX e /V e )∗ OX e , ωPV e /V e ) −
→ E xtjOP e ((F(XT )e /T e )∗ O(XT )e , ωPT e /T e )
T
F -SINGULARITIES IN FAMILIES
13
where Qe : PT e −
→ PV e is the induced map and j = N − dim(X/V ) (we leave off the pushforward
for the inclusion i : XV −
→ PV above). This in turn can be identified with
Q∗e E xtjO(P
e
V)
→ E xtjO(P
(OX e , ω(PV )e /V e ) −
T)
e
(O(XT )e , ω(PT )e /T e )
This last map is exactly the bottom row of [Con00, Diagram (3.6.1) in Theorem 3.6.1], which is an
isomorphism, and hence the proof is complete.
Lemma 2.18 above allows us to show if the divisor Dϕ or ∆ϕ is trivial, then it stays trivial after
base change.
Lemma 2.19 (Divisors which are zero stay zero). Suppose that f : X −→ V is a G1 and S2
e
morphism and that ϕ : L1/p −→ RA1/pe is as above. Additionally suppose that g : T −→ V is any
base change. Finally suppose that the natural map
∗∗
e
e
1−pe
ϕ · R1/p −→ H omR 1/pe L1/p , RA1/pe ∼
= (FXe e /V e )∗ L−1 ⊗ ωX e /V e
A
e
1/pe
e
is an isomorphism of R -modules (here we define ϕ · R1/p to be the R1/p -submodule of the set
e
H omR 1/pe L1/p , RA1/pe generated by ϕ). Then
A
∗∗
e
1/pe
1−pe
−1
e
⊗
ω
L
)
ϕT · (RT )1/p −→ H omR 1/pe LT , RT 1/pe ∼
= (F(X
e
e
∗
T
(XT )e /T e
T ) /T
T
is also an isomorphism.
In particular, if ϕ is relatively divisorial (see Definition 2.9), then the following holds: If Dϕ is
zero, then so is DϕT .
Proof. The statement about divisors is trivial since it is easy to see that a divisor being zero
corresponds to the map above being an isomorphism. Thus we merely need to prove the assertion.
However, since f and all the sheaves involved are relatively S2, it suffices to prove the statement off
a set of relative codimension 2 by Corollary A.8. Therefore, by removing a set of codimension 2, we
can assume that f is a Gorenstein morphism. Thus, working locally if needed, ϕ can be identified
e
→ ωXV e /V e .
(up to multiplication by a unit in R1/p ) with the Grothendieck trace (FXe e /V e )∗ ωX e /V e −
Hence by Lemma 2.18, so can ϕT . The proof is complete.
We now move on to a discussion of base change with respect to divisors. First we observe that
by Lemma 2.19 if ϕ is relatively divisorial, then so is ϕT for every base change g : T −
→V.
Definition 2.20. Suppose that for some relatively divisorial ϕ we have ∆ϕ is as above and that
g :T −
→ V is a base change. Then we write ∆ϕ 8XT (respectively Dϕ 8XT ) to denote the divisor
∆ϕT (respectively DϕT ).
Lemma 2.21 (Pulling back ∆ϕ ). Suppose that f : X −→ V is G1 and S2 and that ϕ is relatively
1/pe
1/pe
divisorial. Then for any g : T −→ V we have ∆ϕ 8XT = (p1 )∗ ∆ϕ (recall p1 : X e ×V i T i −→ Xi
from (2.16.3)).
1/pe
Here, even though ∆ϕ is not necessarily Q-Cartier, (p1 )∗ ∆ϕ can be defined after removing a
set of relative codimension 2 outside of which it is Q-Cartier (since Dϕ is Cartier on such a set).
Proof. The statement is local on X and can be checked after removing a relative codimension
2 set and so we may assume that Dϕ is a Cartier divisor. Thus we assume that L is trivial on
the affine scheme X = Spec R and that f is a Gorenstein morphism. However then shrinking the
e
e
→ RA1/pe can be identified with s1/p · Tr
neighborhood further if necessary, the map ϕ : R1/p −
where Tr : (FXe e /V e ) ∗ ωX e /V e −
→ ωXV e /V e is the Grothendieck trace. The divisor Dϕ is then easily
seen to be the divisor divX (s). On the other hand, it is clear by Lemma 2.18 that then ϕT is trace
1/pe
on XT multiplied by s as well. In particular, it equals (p1 )∗ Dϕ as desired.
14
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
Corollary 2.22. Suppose that g : z −→ V is inclusion of a point. Suppose that ϕ is relatively
divisorial and corresponds to ∆ϕ . Then ∆ϕ |Xz = ∆ϕz .
Proof. The divisor ∆ϕz is determined in codimension 1 where ∆ϕ |Xz and ∆ϕz agree.
2.23. Passing to V ∞ and other perfect points. Suppose first that V is the spectrum of a
e
e
→ R since RA1/pe ∼
→ RA1/pe is also a map L1/p −
perfect field A. Then the map L1/p −
= R. In
such a case, we often also typically write the map as ψ to help distinguish how we are thinking
e
→ R can be composed with themselves as in [BS13,
about it. We note that maps such as ψ : L1/p −
Section 4.1] and furthermore this composition ψ n coincides with ϕn . For the rest of the subsection,
we discuss base change to perfect points. We first consider the perfection of the generic point of V .
1/p∞
Set V ∞ to denote the not-necessarily-Noetherian scheme Spec OV
and we set η∞ to be the
generic point of V ∞ with perfect fraction field k(V ∞ ) (note that this is a perfect point). We obtain
a map fk(V ∞ ) : Xk(V ∞ ) = X ×V Spec k(V ∞ ) −
→ Spec k(V ∞ ) now a scheme of finite type over a
perfect field.
As in Section 2.16, the map ϕ then induces a map
∞
e
.
→ Rk(V ∞ ) := R ⊗A k A1/p
ϕ∞ : L1/p ⊗k(V e ) k(V ∞ ) −
which can be identified with
ψ∞ : Lk(V ∞ )
1/pe
−
→ Rk(V ∞ )
since (FXe e /V e )∗ OX e ⊗V e k(V ∞ ) ∼
= F∗e OXk(V ∞ ) .
More generally, if s ∈ V is any perfect point, we can induce
e
e
e
e
ϕs : L1/p
→ R ⊗A k(s)1/p .
:= L1/p ⊗A1/pe k(s)1/p −
s
e
Since k(s) = k(s)1/p is perfect, we can identify this with a p−e -linear map:
e
ψs : L1/p
−
→ Rs .
s
As above, we see that the composition of ψs in the sense of [BS13, Section 4.1] coincides with the
composition ϕ as in Section 2.13. Finally, we study this process with respect to divisors.
Lemma 2.24. With notation as above, ∆ϕ |Xs = ∆ψs where ∆ϕ is as in Definition 2.10 and ∆ψs
is as in [BS13, Section 4.].
Proof. By Lemma 2.21, we simply must show that ∆ϕs coincides with ∆ψs . Working locally, and
e
removing a set of relative codimension 2, we may assume that Φ ∈ HomR 1/pe ((Rs )1/p , Rs1/pe )
e
e
s
generates the Hom as an (Rs )1/p -module and that ϕ( ) = Φ(z 1/p · ). Thus ∆ϕ = pe1−1 divX (z).
We then identify Rs1/pe with Rs and hence ϕ with ψ. Likewise, we can identify Φ with Ψ which
e
e
e
now generates HomRs ((Rs )1/p , Rs ) as an (Rs )1/p -module. Hence ψ( ) = Ψ(z 1/p · ) and so
∆ψ = pe1−1 divX (z) as desired.
3. Relative non-F-pure ideals
With notation as above, by the Hartshorne-Speiser-Lyubeznik-Gabber (HSLG-)Theorem [Gab04,
Lemma 13.1] cf. [HS77, Lyu97, Bli13], we know that the chain
Rk(V ∞ ) ⊇ ψ∞ (Lk(V ∞ ) )1/p
e
e
2
⊇ ψ∞
(Lpk(V+1∞ ) )1/p
eventually stabilizes. Say that it stabilizes at
(3.0.1)
2e
pne −1
e
p −1
1/p
n
⊇ · · · ⊇ ψ∞
(Lk(V
∞))
ne
n ≥ n0 .
For the rest of this section, we fix this integer n0 , and make the following definition.
⊇ ···
F -SINGULARITIES IN FAMILIES
15
Definition 3.1. With notation as above, we define the integer n0 to be the uniform integer for σ
over the generic point of V , and in general, it will be denoted by nσ(ϕ),k(V ) . We notice that for any
point η ∈ V , we can base change Spec k(η) −
→ V and form a corresponding integer nσ(ϕη ),k(η) .
On the other hand, without the passing to k(V ∞ ). We
e
a1 := ϕ1 L1/p
e
2e
a2 := ϕ2 (L(p +1) )1/p
...
pne −1
ne
an := ϕn (L pe −1 )1/p
...
have the images
⊆ RA1/pe ,
⊆ RA1/p2e ,
⊆ RA1/pne ,
These are ideals of different rings. However, we do have the following relation for any i > j
i
→ RA1/pi ⊇ ai .
Im aj ⊗A1/pj A1/p −
This is straightforward so we leave it to the reader to check (note that the image of (2.13.2) contains
e
the image of ϕ2 ). Additionally, observe that if A was regular, A1/p would be flat over A by [Kun69],
i
and so we could identify the tensor product aj ⊗A1/pj A1/p with its image in RA1/pi . Therefore, for
any integer n ≥ i, we set
n
→ RA1/pn
ai,n = Im ai ⊗A1/pi A1/p −
and consider the chain of ideals:
RA1/pne ⊇ a1,n ⊇ a2,n ⊇ · · · ⊇ an−1,n ⊇ an,n .
Definition 3.2 (nth relative non-F -pure ideal). For every integer n ≥ n0 = nσ(ϕ),k(V ) , we define
the nth limiting relative non-F -pure ideal to be an,n = an . It is denoted by σn (X/V, ϕ).
We now obtain a relative version of the Hartshorne-Speiser-Lyubeznik-Gabber theorem.
Proposition 3.3. Fix notation as above. Then there exists a nonempty open subset U ⊆ V of the
base scheme V satisfying the following for every integer m ≥ n ≥ n0
me
.
(3.3.1)
σm (X/V, ϕ)|f −1 (U ) = Im σn (X/V, ϕ) ⊗A1/pne A1/p −→ RA1/pme
−1
f
(U )
Proof. For any m ≥ n, we consider the containment:
an,m ⊇ am,m .
Fix k(V ) to be the residue field of the generic point η ∈ V and consider the induced containment:
m
m
an,m ⊗A k(V ) = an,m ⊗A1/pm k(V )1/p ⊇ am,m ⊗A1/pm k(V )1/p = am,m ⊗A k(V ).
since inverting an element is the same as inverting its pth power. We notice two identifications
pne −1 1/pne !
∞
pe −1
n
Lk(V
an,m ⊗A1/pm k(V )1/p = an,m ⊗A1/pm k(V ∞ ) = ψ∞
,
∞)
!
me 1/pme
am,m ⊗A1/pm k(V )1/p
Thus
∞
p
m
= am,m ⊗A1/pm k(V ∞ ) = ψ∞
∞
an,m ⊗A k(V )1/p = am,m ⊗A k(V )1/p
since m > n ≥ n0 . It immediately follows that
m
∞
−1
.
∞
an,m ⊗A1/pm k(V )1/p = am,m ⊗A1/pm k(V )1/p
m
e
p −1
Lk(V
∞)
m
since k(V )1/p ⊆ k(V )1/p is a faithfully flat extension. But now by generic freeness [Eis95, Theorem 14.4], for a fixed m, (3.3.1) follows for some open set Un,m (in the case that V is affine, which
we can certainly reduce to, we can invert a single element of A to form Un,m ).
16
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
We now vary m. Choose U = Un,n+1 that works for m = n + 1 and consider the diagram:
L
p(n+1)e −1
pe −1
L
p(n+2)e −1
pe −1
1
p(n+1)e
1
p(n+2)e
⊗A1/p(n+1)e A1/p
(n+2)e
e
ϕn+2
L
pne −1
pe −1
1
pne
(ϕn+1 )1/p ⊗...
⊗A1/pne A1/p
(n+2)e
e
(ϕn )1/p ⊗...
ϕn+1
···
ϕn
u t
e
L1/p ⊗A1/pe A1/p
(n+2)e
ϕ⊗...
* ))
RA1/p(n+2)e
e
e
where the maps labeled (ϕn )1/p ⊗ . . . and (ϕn+1 )1/p ⊗ . . . are induced as in (2.13.1). We know
e
e
that over U , ϕn+1 and ϕn have the same image. Therefore, so do (ϕn )1/p ⊗ . . . and (ϕn+1 )1/p ⊗ . . .
again over U (since tensor is right exact). But then, composing with (ϕ ⊗ . . .) one more time, we
e
e
know ϕn+1 = (ϕ ⊗ . . .) ◦ ((ϕn )1/p ⊗ . . .) and ϕn+2 = (ϕ ⊗ . . .) ◦ ((ϕn+1 )1/p ⊗ . . .) also have the same
image over U . Thus they also share the image with ϕn over U . Hence, if an,n+1 |U = an+1,n+1 |U ,
then an,n+2 |U = an+2,n+2 |U . We use this iteratedly to obtain that an,m |U = am,m |U , which is exactly
the statement of the proposition.
We give three examples of these σn and Un . In the first example, we show that it is possible that
the images σn (X/V, ϕ) never stabilize in the sense of Proposition 3.3 on all of V but only over an
open set. We do the same in the second example, but with respect to a more interesting choice of
X and ϕ. Finally, we give an example where stabilization occurs at n = 2 (instead of at n = 1).
Example 3.4. Fix k to be an algebraically closed field of characteristic p > 2, set A = k[t] and
set R = k[x, t] with the obvious map X −
→ V . Let ϕ : R1/p = k[x1/p , t1/p ] −
→ RA1/p = k[x, t1/p ]
be the composition of the local generator β ∈ HomR 1/p (R1/p , RA1/p ) with pre-multiplication by
A
1
t p (note since ϕ is k[t1/p ]-linear, this is also post-multiplication by t1/p , and it corresponds to the
1
{t = 0}). It easily follows that the image of ϕ is ht1/p i ⊆ k[x, t1/p ]. By tensoring
divisor ∆ϕ = p−1
2
2
with ⊗k[t1/p ] k[t1/p ], we obtain that a1,2 = ht1/p i = t1/p · k[x, t1/p ]. More generally, we see that
n
a1,b = t1/p · k[x, t1/p ]. Now we compute a2,2 as the image:
2
2
ϕ1/p
2
ϕ⊗...
2
k[x1/p , t1/p ] −−−→ k[x1/p , t1/p ] −−−→ k[x, t1/p ].
2
2
The image of ϕ1/p is t1/p and since the map ϕ ⊗ . . . is k[t1/p ]-linear, we see that the composition
2
2
has image ht1/p · t1/p i = ht(1+p)/p i = a2,2 . In general, we see that
an = an,n = ht(1+p+···+p
n−1 )/pn
i
F -SINGULARITIES IN FAMILIES
17
In particular, while we may take Ui = A1 \ {0} = Spec A \ hti, we see that an = σn (X/V, ϕ) never
stabilizes over all of V .
−
→ V := Spec(k[t]). Let the standard
Example 3.5. Let f be the morphism X := Spec (yk[x,y,t]
2 +x3 +t)
→ OXV 1 , for which ∆ϕ = 0. This map can be identified with the
trace map ϕ : (FX 1 /V 1 )∗ ωX 1 /V 1 −
descent of the following map to the quotient
k[x, y, t]
·(y 2 +x3 +t)p−1
/ k[x, y, t]
where Tr is a k-linear map, such that
( i+1−p j+1−p
p
y p tl
i j l
Tr(x y t ) = x
0
Tr
/ k[x, y, t]
if p|i + 1 and p|j + 1 .
otherwise
Then ϕn is given by
k[x, y, t]
·(y 2 +x3 +t)p
n −1
/ k[x, y, t]
where Trn is a k-linear map, such that
(
i+1−pn
j+1−pn
pn
pn
i j l
x
y
tl
(3.5.1)
Trn (x y t ) =
0
Trn
/ k[x, y, t]
if pn |i + 1 and pn |j + 1 .
otherwise
Assume that p is a prime such that p ≡ 1(mod 6) so that 2, 3|pn − 1 for all n > 0. Let us try to
compute now the following number.
d := min{c|tc ∈ im ϕn }
n
To have tc ∈ im ϕn it is necessary to have a polynomial in the ideal generated by (y 2 +x3 +t)p −1 one
n
n
n
of the non-zero monomials of which is xp −1 y p −1 tc . Therefore, (y 2 + x3 + t)p −1 itself has to have
n
n
n
a non-zero monomial which divides xp −1 y p −1 tc . That is, there have to be integers 0 ≤ a ≤ p 2−1
n
and 0 ≤ b ≤ p 3−1 , such that c = pn − 1 − a − b. So, we see that
d ≥ pn − 1 −
pn − 1 pn − 1
pn − 1
−
=
2
3
6
On the other hand we claim that Trn (y 2 + x3 + t)p
n −1
=t
pn −1
6
pn −1
, which will show that d =
pn −1
3
pn −1
6
pn −1
6
and
t
is the only monomial in
also that σn (X/V, ϕ) does not stabilize. First, note that x 2 y
n −1
2
3
p
the expansion of (y + x + t)
the image of which via Trn is not zero. Indeed, the expansion
n
can contain only monomials of the form y 2a x3b tp −1−a−b , where a, b ≥ 0 are integers. Hence, higher
powers (2pn − 1, 3pn − 1, etc) of x and y that do not go to zero by Trn cannot be obtained. Hence,
pn −1
pn −1
pn −1
as we stated x 2 y 3 t 6 is the only interesting monomial that can show up in the expansion.
Lastly we have to verify that the monomial it has non-zero coefficient in the expansion. That is,
(pn − 1)!
pn −1 pn −1 pn −1 6= 0
2 ! 3 ! 6 !
Equivalently, we have to show that the power of p in the prime factorization of the numerator is
the same as in the denominator. For an arbitrary number m this number for m! is
∞
X
m
.
pi
i=1
18
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
Now assume that m =
pn −1
r ,
where r|p − 1. Then the times p divides m! is
n−1
n−1
n−1
∞
X
X
X pn − 1 n−1
X
X p − 1
X
n−1
p
−
1
m
j−i
j
=
p
p
=
=
i
i
i
p
rp
rp
r
i=1
i=1
=
i=1
n−1
n−1
XX
pj−i
i=1 j=i
=
n−1
X
pi
i=0
i=1
j=0
j=0
n−1
n−1
n−1
i=1
i=1
i=1
p − 1 X pn−i − 1 p − 1 X pn−i − 1 X pi − 1
=
=
=
r
p−1
r
r
r
−1
=
r
pn
−1
r
Therefore, we see that p divides pn − 1 times both (pn − 1)! and
pn −1
6
pn −1 pn −1 pn −1
2 ! 3 ! 6 !,
as claimed earlier.
is the smallest power of t that is in σn (X/V, ϕ). Hence there is no stabilization
Summarizing, t
over all of V .
Based on this example, one might ask:
Question 3.6. Is there a relation between the asymptotics of the σn over the non-stable locus, and
the singularities of those fibers (for example, the F -pure threshold)?
We now give an example that stabilizes at the second step (over the entire base).
Example 3.7. (cf. [MY09, Example 4.7]) Fix k to be an algebraically closed field of characteristic
p > 2, set A = k[t] and set R = k[x, t] with the obvious map X −
→ V . Let ϕ : R1/p = k[x1/p , t1/p ] −
→
1/p
1/p
RA1/p = k[x, t ] be the composition of the local generator β ∈ HomR 1/p (R , RA1/p ) with
pre-multiplication by
p2
p−1 {x
=
2
λ1/p }
2
(xp
+ t)
1
p
= f
1
p
A
(which corresponds to ∆ϕ =
on the fiber over t = λ). Note that in RA1/p or
R1/p ,
f
1
p
1
p2
p−1 {x
= t} restricting to
can be written as xp + t1/p .
1
In particular, f p is already an element of RA1/p . Therefore, since ϕ is RA1/p -linear and β is clearly
surjective, the image of ϕ is just hf 1/p i = σ1 (X/V, ϕ) = xp + t1/p .
2
→
On the other hand, we now compose ϕ with itself as described above. In this case, ϕ2 : R1/p −
p+1
2
→ RA1/p2 and pre-multiplying by f p2 . Now,
RA1/p2 is induced by taking the generator β 2 : R1/p −
p+1
p+1
2
. This again is already an element of RA1/p2 , and so by the same argument
f p2 = x1 + t1/p
as above, we see that
E
E 1+p
D
D
2 p+1
2
σ2 (X/V, ϕ) = x + t1/p
= σ1 (X/V, ϕ) · x + t1/p = f p2 .
3
→ RA1/p3 and preNext we form ϕ3 . In this case, it is obtained by taking a generator R1/p −
1+p+p2
p3
1+p+p2
p3
. However, this time f
is not contained in RA1/p3 , and so we cannot
multiplying by f
argue as above. However, f still has a 1/p2 -root in RA1/p3 , and so
1+p+p2
E
E
D
D
1+p
3
3
3
ϕ
f p3
= f p2 ϕ3 f 1/p
.
= ϕ3 f p(1+p) f 1/p
3
3
3
3
→ k[x, t1/p ] we may take as the
So we must only compute ϕ3 (hf 1/p i). Now, ϕ3 : k[x1/p , t1/p ] −
p3 −1
p3
to 1 and other monomials
monomials form a basis for R1/p
map which sends x
Ein x to 0 (those
D
3
3
= k[x, t1/p ] = RA1/p3 . In particular,
over RA1/p3 ). It is thus obvious that ϕ3 f 1/p
1+p
3
σ3 (X/V, ϕ) = f p2 = σ2 (X/V, ϕ) ⊗A1/p2 A1/p .
3
F -SINGULARITIES IN FAMILIES
19
We now obtain a surprising base change statement.
Proposition 3.8. Suppose that T −→ V is a map from an excellent integral scheme with a dualizing
complex, then using the notation of Section 2.16
Im ((qne )∗ σn (X/V, ϕ) → OXT ne ) = σn (X/V, ϕ) · OXT ne = σn (XT /T, ϕT ).
Furthermore, if U satisfies condition (3.3.1) from Proposition 3.3, then
W = g−1 (U ) ⊆ T
satisfies the same condition for σn (XT /T, ϕT ).
Proof. Indeed, images of maps are compatible with arbitrary base change by the right exactness of
tensor. Thus the first statement follows immediately. The second statement follows from the first
since if the two images an,n and an−1,n are equal, they are also equal after base change.
Theorem 3.9 (Base change for σn ). There exists an integer N ≥ 0, such that for all points s ∈ V ,
N ≥ nσ(ϕs ),k(s) . In other words, we have both that σn (X/V, ϕ)·OXsne = σn (Xs /s, ϕs ) (which always
holds) and also that for all m ≥ n ≥ N that
m
σn (X/V, ϕ) ⊗V n k(s)1/p = σm (Xs /s, ϕs ).
Proof. Of course, the statement already holds on U with respect to some integer N0 . Set V1′ := V \U
to be the complement with reduced scheme structure and let
i1 : V1 = (V1′ )reg ֒→ V1′
be the regular locus. We notice that V1 has dimension strictly smaller than dim V . We consider the
base change XV1 −
→ V1 . Each fiber of XV1 −
→ V1 is isomorphic to a fiber of X −
→ V . Then choose
an open set U1 ⊆ V1 for which the statement holds for some integer N1 .
Now fix V2′ = V1′ \ U1 and i2 : V2 = (V2′ )reg ֒→ V2′ to be the regular locus and repeat. This process
terminates by Noetherian induction.
Setting N = max{N0 , N1 , N2 , . . . } completes the proof.
Suppose now that s ∈ V is a perfect point, see Definition 2.2. It follows that Xs /s is a variety
1/pe
over a perfect field and so since se ∼
ϕs : Ls
−
→ Rs e
= s, we have Xse ∼
= Xs . Thus we can identify
T
with a p−e -linear map ψs as in Section 2.23. Thus under these identifications n≥0 σn (Xs /s, ϕ) =
me
σ(Xs , ψs ). However, if we have the stabilization σn (Xs /s, ϕ) ∼
=
= σn (Xs /s, ϕ) ⊗k(s)1/pne k(s)1/p ∼
σm (Xs /s, ϕ) for every m ≥ n, the equality σn (Xs /s, ϕ) = σ(Xs , ψ) holds. Combining this with our
previous work we obtain the following corollary.
Corollary 3.10. With notation as above, there exists an integer N ≥ 0 such that
σn (X/V, ϕ) · OXsne = σ(Xs , ψs )
for all perfect points s ∈ V and n ≥ N .
3.11. Iterated non-F -pure ideals versus the absolute non-F -pure ideal. We now compare
the iterated non-F -pure ideal with the absolute non-F -pure ideal. First however, we make the
following assumption.
Convention 3.12. In this subsection, Section 3.11, we always assume that our base V is an F -pure
quasi-Gorenstein (i.e., ωV is a line bundle) F -finite integral scheme.
→ OV be a map generating H omV (F∗e OV ((1 − pe )KV ), OV ) as
Let ΨeV : F∗e OV ((1 − pe )KV ) −
ne
e
an F∗ OV -module. By taking p th roots, applying f −1 , and then writing the A-module M :=
f −1 (OV ((1 − pe )KV )), we obtain:
(ΦeA )1/p
ne
: M 1/p
(n+1)e
(n+1)e
= f −1 F∗
ne
(OV ((1 − pe )KV )) −
→ f −1 F∗ne OV = A1/p .
20
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
e
We tensor with L1/p ⊗A1/pe
and so obtain:
e
Φ′ : L1/p ⊗A1/pe M 1/p
(3.12.1)
(n+1)e
ne
e
−
→ L1/p ⊗A1/pe A1/p .
ne
Set N = L ⊗A M 1/p . We observe that N is a line bundle on XV ne . In fact, if q1 : XV ne =
X ×V V ne −
→ X is the first projection and q2 : XV ne −
→ V ne is the second projection, then
N = q1∗ L ⊗OXV ne q2∗ OV ne ((1 − pe )KV ne )
and so it follows for any integer l > 0 that
1/pne
∗ l
∗
e
l
∼
ne
ne
N = q1 L ⊗OXV ne q2 OV (l(1 − p )KV ) = L ⊗A M l
l
(3.12.2)
Finally, we compose (3.12.1) with ϕ′ := ϕ ⊗A1/pe A1/p
γ:
ne
to construct:
1/pe
N
(n+1)e
e
= L1/p ⊗A1/pe M 1/p
Φ′
e
−→ L1/p ⊗A1/pe A1/p
(3.12.3)
ϕ′
ne
−→ RA1/pe ⊗A1/pe A1/p
ne
= R ⊗A A1/p
= RA1/pne .
ne
This is a map from an invertible sheaf on XVe (n+1)e = (XV ne )e to the structure sheaf on XV ne . In
particular, it is a map such as one studied in [BS13, Section 4].
Let us point out an alternate way to construct γ. Indeed, take
ϕ′′ := ϕ ⊗A1/pe M 1/p
(n+1)e
e
: L1/p ⊗A1/pe M 1/p
(n+1)e
−
→ RA1/pe ⊗A1/pe M 1/p
We can then compose this with RA1/pe ⊗A1/pe (ΦeA )1/p
e
e
N 1/p = L1/p ⊗A1/pe M 1/p
(n+1)e
ne
ϕ′′
−
−
→ RA1/pe ⊗A1/pe M 1/p
(n+1)e
(n+1)e
∼
.
= R ⊗A M 1/p
= (ΦeA )′ to obtain
(n+1)e
(Φe )′
A
−−−
→ RA1/pe ⊗A1/pe A1/p
ne
= RA1/pne .
This composition is easily seen to coincide with γ.
We can then compose γ with itself m-times as in [BS13, Section 4] or [Sch09]. We recall this
construction for the benefit of the reader. To construct γ 2 , we tensor the map
e
→ R ⊗A A1/p
γ : N 1/p −
with the line bundle ⊗R
ne
= RA1/pne
N , take pe th roots and then compose it with γ to obtain:
e 1/p2e
e γ
ne
γ 2 : N 1+p
−
→ N 1/p −
→ R ⊗A A1/p
ne
A1/p
Recursively, we can construct:
γm
pme −1 1/pme
N pe −1
pme −1 1/p(n+m)e
pme −1 1/pme
⊗A1/pme M pe −1
=
L pe −1
−
→ RA1/pne
:
We are now in a position to relate the absolute σ(XV 1/pne , γ)
First we observe that for each m ≤ n, Φm
A induces maps
pme −1 1/pme
µm :
N pe −1
pme −1 1/pme
⊗A1/pme
=
L pe −1
ne
pme −1 1/pme
...⊗(Φm )1/p
−−−−−A−−−−→
L pe −1
⊗A1/pme
with the relative σn (X/V, ϕ).
M
pme −1
pe −1
ne
A1/p .
1/p(n+m)e
F -SINGULARITIES IN FAMILIES
21
Furthermore these maps are surjective since Φm
A is.
Lemma 3.13. Assuming Convention 3.12, we have the following factorization of γ n and γ n−1 as
indicated by the diagram below
p(n−1)e −1 1/p(n−1)e
N pe −1
pne −1 1/pne
N pe −1
pne −1 1/pne
pne −1 1/p(n+n)e
L pe −1
⊗A1/pne M pe −1
γ′
µn
p(n−1)e −1 1/p(n−1)e
⊗
L pe −1
pne −1 1/pne
L pe −1
(n−1)e
A1/p
p(n−1)e −1 1/p(n+n−1)e
M pe −1
µn−1
p(n−1)e −1 1/p(n−1)e
⊗
L pe −1
A1/p
A1/p
(n−1)e
ne
γ n−1
ϕn−1 ⊗...
ϕn
RA1/pne
γn
ne
where the arrow γ ′ is induced by γ and ϕn−1 ⊗ . . . is simply ϕn−1 ⊗A1/p(n−1)e A1/p . Furthermore,
since µn−1 is surjective we have
ne
Im(γ n−1 ) = an−1,n = Im(ϕn−1 ⊗A1/p(n−1)e A1/p ).
Likewise since µn is surjective, we have
Im(γ n ) = an,n = an = Im(ϕn ).
Proof. The two equalities are immediate from the surjectivities of µn−1 and µn and the commune
(n+1)e
−
→ A1/p
tativity of our diagram. The surjectivities of µn−1 and µn are clear since M 1/p
is surjective and tensor is right-exact. It remains to prove the commutativity of our diagram. It
suffices to prove the commutativity of the square:
pme −1 1/pme
pme −1 1/p(n+m)e
L pe −1
⊗A1/pme M pe −1
γ′
/
p(m−1)e −1 1/p(m−1)e
L pe −1
⊗
(m−1)e
A1/p
µm−1
µm
pme −1 1/pme
ne
L pe −1
⊗A1/pme A1/p
p(m−1)e −1 1/p(n+m−1)e
pe −1
M
/
L
p(m−1)e −1
pe −1
1/p(m−1)e
⊗
A1/p
(m−1)e
A1/p
ne
and we will prove it by induction on m. It is clear that, if the above square is commutative for
(n+1)e
and taking pe th roots, we will have the
m, then, by tensoring it with L ⊗R (−) ⊗A1/pne M 1/p
commutativity for m + 1. Hence it remains to prove the commutativity when m = 2. To this end,
ne
(n+1)e
−
→ A1/p by α. Then α induces a map
we will denote the map M 1/p
e
p2e −1 1/p(n+2)e
(n+1)e 1/p
ne
e
= M 1/p ⊗A1/pne M 1/p
α′ = (id ⊗ α)1/p : M pe −1
−
→ M 1/p
e
ne
⊗A1/pne A1/p
ne
= M 1/p
(n+1)e
e
→ R ⊗A A1/p induces
It is clear that α ◦ α′ = α2 . Likewise, the map ϕ : L1/p −
p2e −1 1/p2e
e
2e
e
e
2e 1/p
e
−
→ L1/p ⊗A1/pe A1/p
ϕ′ = (id ⊗ ϕ)1/p : L pe −1
= L1/p ⊗R1/pe L1/p
We will analyze the 4 maps involved in the square. The left vertical map
p2e −1 1/p2e
p2e −1 1/p(n+2)e
p2e −1 1/p2e
⊗A1/p2e M pe −1
−
→ L pe −1
µ2 : L pe −1
22
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
is given by id ⊗ (α ◦ α′ ); similarly the right vertical map µ1 is id ⊗ α. The top horizontal map is
given by
p2e −1 1/p(n+2)e ϕ′ ⊗α′
p2e −1 1/p2e
(n+1)e
2e
e
⊗A1/p2e M pe −1
−−−−→ L1/p ⊗A1/pe A1/p ⊗A1/p2e M 1/p
L pe −1
∼
e
−
→ L1/p ⊗A1/pe M 1/p
(n+1)e
And the bottom horizontal map is give by
p2e −1 1/p2e
ne
2e
ne ϕ′ ⊗id
e
L pe −1
⊗A1/p2e A1/p −−−→ L1/p ⊗A1/pe A1/p ⊗A1/p2e A1/p
∼
ne
e
−
→ L1/p ⊗A1/pe A1/p .
p2e −1 1/p2e
p2e −1 1/p(n+2)e
P
Now given an arbitrary x ⊗ z ∈ L pe −1
⊗A1/p2e M pe −1
, write ϕ′ (x) = i xi ⊗ yi
2e
e
with xi ∈ L1/p and yi ∈ A1/p . Then, on one hand, if we follow the top horizontal map and then
the right vertical map, we will have
X
X
X
X
x ⊗ z 7→ (
xi ⊗ yi ) ⊗ α′ (z) 7→
xi ⊗ yi α′ (z) 7→
xi α(yi α′ (z)) =
xi ⊗ yi α(α′ (z)),
i
i
i
i
2e
ne
where the last equality holds since α is A1/p -linear and hence A1/p -linear.
On the other hand, if we follow the other path (i.e. the left vertical map first and then the
bottom horizontal map), we will have
X
X
x ⊗ z 7→ x ⊗ α′ (z) 7→ x ⊗ α(α′ (z)) 7→ (
xi ⊗ yi ) ⊗ α(α′ (z)) 7→
xi ⊗ yi α(α′ (z)).
i
i
This proves that the square is indeed commutative when m = 2 and concludes the proof of our
Lemma.
Theorem 3.14. With notation as in Convention 3.12 and below, choose n > n0 = nσ(ϕ),k(V ) . Then
there exists a dense open set U ⊆ V ∼
= V e with W = f −1 (U ) ⊆ X such that
σ(XV 1/pne , γ)|W = σn (X/V, ϕ)|W .
Furthermore, shrinking U further if necessary we can require for all perfect points u ∈ U , that
σ(XV 1/pne , γ) · OXu = σn (Xu /u, ϕu ) = σ(Xu , ψu )
where ψu is ϕu viewed as an absolute p−e -linear map.
Proof. The second statement is immediate from the first by Proposition 3.8, so we need only prove
the first statement. By the diagram in Lemma 3.13 and applying Proposition 3.3, we see that
Im(γ n )|W = an,n |W = an−1,n |W = Im(γ n−1 )|W
for all n > n0 . Hence Im(γ n )|W = σ(XV 1/pne , γ)|W and the result follows.
Convention 3.15. Assuming everything in Convention 3.12, we additionally assume that ϕ is relatively divisorial.
Before we proceed, we prove a lemma about interpreting maps as divisors in the relative vs
absolute setting.
Lemma 3.16. With notation as in Convention 3.15, if ∆ϕ is the divisor corresponding to ϕ as in
Definition 2.8. Then ∆γ , the divisor on XV ne corresponding to γ as in [BS13, Section 4], is equal
to h∗ ∆ where h : XV ne = X ×V V ne −→ X is the projection.
F -SINGULARITIES IN FAMILIES
23
Proof. Working off a set of relative codimension 2, and then working locally on X and V , we
e
can assume that L and M are trivial, and further assume that ϕ( ) = Φ(z 1/p · ) where Φ
e
e
generates H omR 1/pe (R1/p , RA1/pe ) as an R1/p -module. Then ∆ϕ is the divisor corresponding to
A
1
div
(z).
On
the other hand, since the map Γ corresponding to Φ is a composition of generating
e
X
p −1
maps, Γ generates H omR
ne
A1/p
(R
1/pe
A1/p
(n+1)e
, RA1/pne ), see for example [Kun86, Appendix F], and
e
so corresponds to the zero divisor. Furthermore, we see that γ( ) = Γ(z̄ 1/p · ) where z̄ is the
element corresponding to z ⊗ 1 ∈ RA1/pne . Hence ∆γ = pe1−1 divXV ne (z̄) and the claim follows.
We now need a Lemma which says roughly that in an F -pure F -finite ring A with maximal ideal
e
→ A such that the closed point V (m) is
m, then at least locally, one can always choose a α : A1/p −
an F -pure center of ∆α . To this end, we introduce the following.
e
e
Me,A := {ϕ ∈ HomA (A1/p , A)|ϕ(m1/p ) ⊆ m}
and
e
e
Ne,A := {ϕ ∈ HomA (A1/p , A)|ϕ(A1/p ) ⊆ m}.
e
e
It is straightforward to check that both Me,A and Ne,A are A1/p -submodules of HomA (A1/p , A)
and that Ne,A ⊆ Me,A . We now observe that:
Lemma 3.17. For any F -finite reduced G1 and S2 ring A with maximal ideal m, the formation of
Me,A and Ne,A commute with localization and completion completion, i.e. if W is a multiplicative
system and ˆ denotes completion along m then
W −1 Me,A ∼
= (W −1 M )e,W −1 A and W −1 Ne,A ∼
= (W −1 N )e,W −1A and
Me,A ⊗A Â ∼
and Ne,A ⊗A Â ∼
=M
=N
e,Â
e,Â
The following proof was suggested to us by Manuel Blickle and Kevin Tucker. We believe there
are more general proofs that work without the G1 and S2 hypothesis (but this proof is short).
e
e
e
Proof. We have a natural injective map HomA (m1/p , m) ֒→ HomA (m1/p , A), but HomA (m1/p , A) ∼
=
e
HomA (A1/p , A) since both modules are S2 (since a reflexive module is S2 in a G1 and S2 ring,
e
e
[Har94]). Thus we have HomA (m1/p , m) ֒→ HomA (A1/p , A), the image of which is Me,A . Therefore Me,A clearly commutes with localization and completion since the formation of both Hom
sets commutes with localization and completion. Similarly, the formation of Ne,A commutes with
localization and completion.
Lemma 3.18. Suppose that A is an F -pure F -finite G1 and S2 ring with maximal ideal m. Then
e
e
there exists an A-linear map αA : A1/p −→ A such that αA is surjective and αA (m1/p ) ⊆ m.
Proof. We first assume that A is a local ring. In this case, it suffices to show that Ne,A 6= Me,A .
By Lemma 3.17, we may assume that A is complete since  is faithfully flat over A. By the Cohen
Structure Theorem, there is a ring of formal power series S = kJx1 , . . . , xn K over a coefficient field k
e
e
of A with surjection S−
→
→A. Write A = S/I. Recall that HomS (S 1/p , S) is generated (as an S 1/p e
pe −1
→ S that sends (x1 · · · xn ) pe to 1 and other basis monomials (including
module) by g : S 1/p −
e
1) to zero. By Fedder’s lemma [Fed83, Lemma 1.6], each element of HomA (A1/p , A) has the form
e
e
e
e
f (u1/p −) with u ∈ (I [p ] : I). Since A is F -pure, (I [p ] : I) 6⊂ n[p ] where n = (x1 , . . . , xn ). Let
e
e
u be an element of (I [p ] : I)\n[p ] . Then u must have a monomial term cxa11 · · · xann with c ∈ k
and ai < pe for each i. Choose such a monomial appearing in u with the least degree and still
e
e
e
e
denote it by cxa11 · · · xann . Then it is clear that (xp1 −1−a1 · · · xpn −1−an )u − c(x1 · · · xn )p −1 ∈ n[p ] .
e
e
e
e
→ A by α(−) = g((x1p −1−a1 · · · xpn −1−an u)1/p −). Then α is in Me,A \Ne,A .
Now define α : A1/p −
This completes the local case by Lemma 3.17.
24
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
For the non-local case, we have the map Me,A −
→ A which is evaluation at 1. Since the formation
of Me,A commutes with localization, this map is surjective if and only if it is surjective locally. The
e
above work proves the result after localizing at m. However, clearly (Me,A )n = HomA (A1/p , A) n
for prime ideals not equal to m. The result follows.
For the moment, we work sufficiently locally so that X = Spec R and V = Spec A are affine
and that L and M are isomorphic to R and A respectively. Continue to assume that A is F pure and quasi-Gorenstein, and fix a point s ∈ Spec A. Shrinking V if necessary, choose a map
e
1/pe
αs ∈ HomA (A1/p , A) which is surjective and which satisfies α(ms ) ⊆ ms (whose existence is
e
guaranteed by Lemma 3.18). We write αs ( ) = ΦeA (z 1/p · ) for some z ∈ A. Note we have an
e
induced map αs : (A/ms )1/p −
→ A/ms .
e
→ RA1/pne defined by the rule β( ) =
We thus induce the following map β : (RA1/pne )1/p −
(n+1)e
1/p
· ). Certainly using the notation of (3.12.3),
γ(z
1/p(n+1)e
β(ms
=
·R
1/pe
(n+1)e
A1/p
(n+1)e
1/p(n+1)e
ms
γ(z 1/p
·R
)
1/pe
(n+1)e )
A1/p
(n+1)e
e
ne
1/p(n+1)e
1/pe
ms
· R 1/p(n+1)e
z 1/p
=
ϕ′ ◦ (ΦeA )1/p ⊗A1/pne R1/p
A
e
ne
1/p(n+1)e
1/pe
1/p
′
1/p
ne
ms
· R 1/p(n+1)e
⊗A1/p R
=
ϕ ◦ (αs )
1/pne
⊆ ϕ′ (ms
1/pne
⊆ ms
A
1/pe
· RA1/pne )
RA1/pne .
And hence we induce a map
β:R
1/pe
(A/ms )1/p
(n+1)e
−
→ R(A/ms )1/pne .
Lemma 3.19. With notation as above, ∆β = ∆γ |Xsne where the divisors are induced as in [BS13,
Section 4].
Proof. Since ϕ is relatively divisorial, by working in relative codimension 1, it suffices to show the
result in the case that ∆γ (and so ∆ϕ ) is trivial (since locally ϕ or ψ is a generator pre-multiplied
by the element which determines ∆ϕ or ∆ψ ). But now consider the composition constructing β.
R
1/pe
(n+1)e
A1/p
·z 1/p
(n+1)e
−−−−−−−→ R
Φ′
1/pe
(n+1)e
A1/p
The composition of the first two maps sends m1/p
(n+1)e
ϕ′
1/pe
−→ RA1/pne −→ RA1/pne
·R
1/pe
out by these ideals induces a generating map for HomR1/pe
ne
(A/m)1/p
1/p(n+1)e
1/pe
ne
(n+1)e
A1/p
(R
to m1/p · RA1/pne and modding
1/pe
1/pe
(n+1)e
(A/m)1/p
, R(A/m)1/pne ) via the
ne
−
→ (A/m)1/p is a generating map
fact that αs is surjective and so the induced map (A/m)
ne
′
as well. On the other hand, if ϕ is a generating map, so is ϕ ⊗A1/pne (A/m)1/p by Lemma 2.19.
Since the composition of two generating maps is a generating map, we are done.
We now explain how the absolute σ behaves when restricting to fibers. First we need two lemmas.
Lemma 3.20. With notation as above, and v ∈ V a closed point.
σ(XV ne , ∆γ ) · OXvne = σ(XV ne , γ) · OXvne ⊇ σ(Xvne , β) = σ(Xvne , ∆β ) = σ(Xvne , ∆γ |Xvne ).
Proof. Obviously σ(XV ne , γ) ⊇ σ(XV ne , β). On the other hand, it follows easily that σ(XV ne , β)
restricts to σ(Xvne , β) by F -adjunction, see for instance [FST11].
The next lemma is a generalization of the fact that for an ideal in a regular ring, I ⊆ (I [1/p] )[p]
using the •[1/p] notation from [BMS08].
F -SINGULARITIES IN FAMILIES
25
Lemma 3.21. Let B be an F -finite regular ring such that B 1/p is a free B-module and that
a
HomB (B 1/p , B) ∼
= B 1/p (for example this happens if B is local). Further suppose that Q is a B 1/p
a
submodule of B 1/p ⊗B P for some B-module P ,
a
Q ⊆ B 1/p ⊗B P.
a
a
Then Q ⊆ B 1/p ⊗B (ϑ ⊗ idP )(Q) where ϑ ∈ HomB (B 1/p , B) is the generator.
any element
Proof. The strategy is similar to [BMS08,
P Proposition 2.5]. By [Kun69, Theorem 2.1],
a
1/p
over B and
of Q can be written as a finite sum
λi ⊗ mi , where {λi } form a basis for B
a
mi ∈ P . Now since ϑ is a local generating map of HomB (B 1/p , B), the projection maps onto
a
the λi are multiples of ϑ. In other words, for each λi there exists a ui ∈ B 1/p such that both
ϑ(ui λi ) = 1 ∈ R and ϑ(ui λj ) = 0 ∈ R for j 6= i.
P
We observe that each mj is then in (ϑ ⊗ idP )(Q) since ϑ(uj λi ⊗ mi ) = mj . It follows then
P
a
that
λi ⊗ mi ∈ B 1/p ⊗B (ϑ ⊗ idP )(Q) .
Lemma 3.22. Suppose that A = k is an F -finite field and K ⊇ k is a field extension of k such that
K is perfect. Choose X = Spec R −→ V = Spec A = Spec k a flat map of finite type with X −→ V
e
possessing geometrically reduced, G1 and S2 fibers. Additionally suppose that γ : R1/p −→ R is any
e
R-linear map which is a local generator of HomR (R1/p , R) at the generic points of the codimension1 components of the non-smooth locus of X. Then, setting ∆γ as in [BS13, Section 4].
σ(X, ∆γ ) ⊗k K ⊇ σ(XK , (∆γ ) ×k K)
ne
ne
ne
Proof. First choose n > 0 such that σ(X, ∆γ ) = γ n (R1/p ). Since Homk (k1/p , k) is a free k1/p module of rank 1, by [Kun86, Appendix F] we can factor γ n as
R1/p
ne
β
/ R 1/pne
k
ϑ
/
: Rk
γn
with ϑ a local generator of H omRk (Rk1/pne , Rk ) and so that ∆β coincides with ∆γ . To see this last
ne
e
point, notice that if γ is a local generator of HomR (R1/p , R), then so is γ n of HomR (R1/p , R)
ne
and hence β is also a local generator of HomR 1/pne (R1/p , RA1/pne ). More generally then, if γ n
A
ne
ne
is a local generator pre-multiplied by some g1/p ∈ R1/p (working in codimension 1) then β is
ne
also a generator times g 1/p (possibly times a unit depending on our choice of ϑ). This proves that
∆β = ∆γ as asserted.
ne
ne
ne
ne
ne
We now observe that γ n (R1/p ) ⊗k k1/p = ϑ(β(R1/p ) ⊗k k1/p ⊇ β(R1/p ) by Lemma 3.21
ne
ne
(here B = k and P = β(R1/p )). After embedding k1/p ⊆ K we tensor β by K to obtain a map
βK : R1/p
ne
→ Rk1/pne ⊗k1/pne K.
⊗k1/pne K −
We see immediately that ∆βK = (∆γ ) ×k K (again by an argument about local generators). Then,
by Lemma 3.21,
ne
ne
ne
σ(X, γ) ⊗k K ∼
= γ n (R1/p ) ⊗k k1/p ⊗ 1/pne K ⊇ β(R1/p ) ⊗ 1/pne K ⊇ σ(X, βK ).
k
This completes the proof.
k
Theorem 3.23. With notation as in Convention 3.15, there exists an N > 0 such that for all
n > N we have σ(XV 1/pne , ∆γ ) · OXsne = σ(Xsne , (∆γ )|Xsne ) for all perfect points s ∈ V .
Proof. The statement is local, so we may assume X and V are affine and that L and M are trivial
just as above in Lemma 3.19. We know
σ(XV 1/pne , γ n ) ⊆ Im(γ n ) = σn (X/V, ϕ)
26
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
by Lemma 3.13. On the other hand, by Theorem 3.9, we have σn (X/V, ϕ) · OXs = σ(Xs , ϕs ).
Therefore, it is sufficient to show that σ(XV 1/pne , γ) · OXsne ⊇ σ(Xsne , ψsne ).
For this we may assume that V is the spectrum of an F -Finite F -pure local ring and that
Spec K = s 7→ V has image the closed point Spec k = v ∈ V . We first see that σ(XV ne , ∆γ )·OXvne ⊇
σ(Xvne , ∆β ) = σ(Xvne , ∆γ |Xvne ) by and using the notation of Lemma 3.20. On the other hand, by
Lemma 3.22
σ(Xvne , ∆γ |Xvne ) · OXsne = σ(Xvne , ∆β ) · OXsne ⊇ σ(Xsne , ∆β ×vne sne ) = σ(Xsne , ∆γ |Xsne ).
This completes the proof.
Using the same method as Lemma 3.20, we also obtain the following result which may also be
of independent interest.
Proposition 3.24. Assume that f : X −→ V is a flat finite type reduced G1 and S2 morphism
e
and that V is regular. Additionally assume that γ : R1/p −→ R is an R-linear map which is a local
generator at the generic points of the codimension-1 components of the non-smooth locus of f for
each fiber of f . Then for any a > 0 we have
a
σ(X, ∆γ ) ⊗A A1/p ⊇ σ(XV a , ∆γ ×V V a ).
e
Proof. The statement is local over the base and so we may assume that KV ∼ 0 and that A1/p is
a free A-module since V is regular [Kun69]. Note then that KX/V ∼
= KX . Choose ne ≥ a > 0 such
ne
a
ne
that σ(X, ∆γ ) = γ n (R1/p ). Since HomA (A1/p , A) is a free A1/p -module of rank 1, by [Kun86,
Appendix F] we can factor γ n as
R1/p
ne
β
/ R 1/pa
A
ϑ
/
;R
γn
ne
where ϑ is a local generator of H omR (RA1/pa , R). Note β ∈ HomR 1/pa (R1/p , RA1/pa ) which is
A
identified with
a
a
ne
H omR 1/pa R1/p ⊗R 1/pa (A1/p ⊗A ωR/A ), (A1/p ⊗A ωR/A )
A
A
ne
∼
= H omRA1/pa (OX (pne KX/V ))1/p , ωR 1/pa /A1/pa
A
ne
∼
= H omRA1/pa (OX (pne KX ))1/p , ωRA1/pa
∼ H om 1/pne (OX (pne KX ))1/pne , ω 1/pne
=
R
R
ne
∼
= (OX ((1 − pne )KX ))1/p .
Note that since the base is regular, all these sheaves are automatically reflexive and so there is no
need to double dualize as we did before in Definition 2.8. By dividing by (pne − 1) we see that β
coincides with a divisor ∆β such that (1 − pne )(KX + ∆β ) ∼ 0. Furthermore, it is easy to see that
∆β coincides with ∆γ just as in Lemma 3.20 since ϑ is a local generator.
ne
ne
R1/p ⊗
a 1/p
ne+a ...⊗(ΦA )
−−−−−−−−→
A1/p
By composing β with a local generating (and surjective) map
ne
ne
ne
−
→ RA1/pa satisfying the
R1/p we then obtain a map γ ′ : (RA1/pa )1/p ∼
= R1/p ⊗A1/pne A
a
′
′
condition ∆γ = ∆γ ×V V . We also notice that γ has the same image as β.
ne
A1/p
1/pne+a
F -SINGULARITIES IN FAMILIES
a
ne
a
ne
27
ne
Now, we observe that γ n (R1/p ) ⊗A A1/p = ϑ(β(R1/p )) ⊗A A1/p ⊇ β(R1/p ) by Lemma 3.21
ne
setting B = A and P = β(R1/p ). The remainder of the proof is easy since
a
=
⊇
=
⊇
=
σ(X, ∆γ ) ⊗A A1/p
a
ne
γ n (R1/p ) ⊗A A1/p
ne
β(R1/p )
ne
γ ′ (RA1/pa )1/p
σ(XV a , ∆γ ′ )
σ(XV a , ∆γ ×V V a ).
3.25. Application to sharply F -pure singularities and HSL numbers in families. We
e
observe that Corollary 3.10 has an immediate application. Recall that if γ : L1/p −
→ R is an Rn
n+1
linear map, then the HSL number is the first integer n such that Im(γ ) = Im(γ
) = σ(R, γ), cf.
[Sha07].
Corollary 3.26 (Uniform behavior of HSL numbers). Given a flat family f : X −→ V over a
e
excellent integral scheme V with a dualizing complex of characteristic p > 0 and ϕ : L1/p −→ OXV e
as before, there exists an integer N > 0 such that gives an upper bound on the HSL number of
1/pe
(Xs , ϕs : Ls
−→ OXs ) for every perfect point s ∈ V .
Proof. The statement immediately follows from Theorem 3.9 since clearly σm (Xs /s, ϕs ) agrees with
σ(Xs , γs = ϕs ) for m ≫ n.
Now we study the deformation of sharp F -purity, a question which has been studied before in
[SZ09, Has10]. We believe that the hypothesis that f is proper is necessary in Theorem 3.30.
e
→ OX = R) is called sharply F -pure if σ(X, ψ) =
Definition 3.27. Recall that a pair (X, ψ : L1/p −
e
→ RA1/pe ), we define (X/S, ϕ) to be relatively sharply F -pure if
OX . Given (X/S, ϕ : L1/p −
σn (X/S, ϕ) = RA1/pne for some n > 0, cf. [Has10].
Lemma 3.28. If σn (X/S, ϕ) = RA1/pne for some n > 0, then the same holds for all n > 0.
Proof. Suppose σn (X/S, ϕ) = RA1/pne for some n > 0. Choose m < n to start. Then we know
n
→ RA1/pn
am,n ⊇ an,n = σn (X/S, ϕ) and so am,n = RA1/pne . But am,n = Im am ⊗A1/pm A1/p −
which is just the extension of σm (X/S, ϕ) ⊆ RA1/pme to RA1/pne . Additionally, RA1/pme ⊆ RA1/pne
is an integral extension and hence we must have σm (X/S, ϕ) = RA1/pme as well.
We finish the proof by showing σ2n (X/S, ϕ) = RA1/p2ne . This follows quickly since σ2n (X/S, ϕ)
factors as a composition of two surjective maps (as ⊗ is right exact) as in (2.13.1) and (2.13.1).
Remark 3.29. If S is a point, then (X/S, ϕ) being relatively sharply F -pure is equivalent to (X/S, ϕ)
being geometrically sharply F -pure (i.e., that (Xt , γt ) is sharply F -pure for every geometric point
t−
→ S where γt is induced from ϕt as in Section 3.11). To see this, certainly observe that if (X/S, ϕ)
is relatively sharply F -pure then so is any base change (Xt /t, ϕt ). But then (Xt , γ) is sharply F -pure
by Theorem 3.14. Conversely, if (Xt , γt ) is sharply F -pure, then certainly (Xt /t, ϕt ) is relatively
sharply F -pure by Lemma 3.13. But then so is (X/S, ϕ) near that t by Nakayama.
Theorem 3.30 (Openness of sharp F -purity). With notation as before, assume that f : X −→ V
is proper. Assume that s ∈ V is a point and that (Xs /s, ϕs ) is relatively sharply F -pure (in other
words, geometrically F -pure). Then there exists a dense open set U ⊆ V containing s such that
(Xu /u, ϕu ) is relatively sharply F -pure for all u ∈ U (in particular, (Xu , ϕu ) is sharply F -pure for
all perfect points u ∈ U ).
28
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
Proof. Choose n ≫ 0 such that σn (Xs /s, ϕs ) = RO1/pne since (Xs /s, ϕs ) is sharply F -pure. By
s
Theorem 3.9, we know that σn (X/V, ϕ) = RA1/pne in a neighborhood W ⊆ X of Xs . Let Z =
X \ W ⊆ X be the complement of that neighborhood. Since f is proper, f (Z) is closed, and also
does not contain s. Set U = V \ f (Z). Then σn (XU /U, ϕU ) = RO1/pne . It follows from Theorem 3.9
U
that all the fibers (Xu /u, ϕu ) are relatively sharply F -pure for perfect u ∈ U as desired.
We now state the above theorem in the context of divisors.
Corollary 3.31 (Openness of sharp F -purity for divisor pairs). Suppose that f : X −→ V is a
proper G1 and S2 morphism as before and now additionally suppose that ∆ is a Q-divisor satisfying
conditions (a)–(d) from Remark 2.11. Additionally suppose that s ∈ V is a perfect point and that
(Xs , ∆|Xs ) is sharply F -pure (or more generally, if s is not perfect, that (Xse , ∆|Xse ) is sharply
F -pure for some/all e > 0). Then there exists an open set U ⊆ V containing s such that for all
u ∈ U we have (Xu , ∆|Xu ) is sharply F -pure (respectively, we have (Xue , ∆|Xue ) is sharply F -pure
for some/all e > 0 and all u ∈ U ).
e
→ RA1/pe as in Remark 2.11. The result then
Proof. Associate to ∆ a relatively divisorial ϕ : L1/p −
follows immediately from Theorem 3.30 and Corollary 2.22.
4. Relative test ideals
In this section we define a notion of a relative test ideal. We construct it using the following
method. We first find some ideal I which when restricted to every geometric fiber is contained in
ne
the test ideal of that fiber. We then sum up the image of I 1/p under ϕn (the same method is used
to construct the test ideal in the absolute case). This sounds simple enough, but it actually leads
to a somewhat disappointing construction. The problem is that the ideal we pick is in no sense
canonical and we do not see a way to make it both canonical and stable under base change. In the
case that ϕ corresponds to a trivial divisor, we can make our choices canonical by setting I to be
the Jacobian ideal (this is done later when we discuss relative F -rationality in Section 5), but for
more general ϕ we don’t know of an analogous choice.
e
→ RA1/pe ). We begin
We fix the notation of the previous sections (in particular, we fix ϕ : L1/p −
∞
by choosing an arbitrary ideal I ⊆ OX . After base change to k(V ) as before, setting I∞ = I · R∞ ,
we can form the following sum:
τ (Rk(V ∞ ) , ψ∞ I∞ ) :=
∞
X
i=0
i
ψ∞
(I∞ · L
pie −1
pe −1
k(V ∞ )
)1/p
ie
.
If it happens that I∞ is contained in the test ideal of (Rk(V ∞ ) , ϕ∞ ) and non-zero on any component
of X∞ , then we see that the absolute test ideal τ (Rk(V ∞ ) , ψ∞ (I∞ )) is equal to τ (Rk(V ∞ ) , ψ∞ ) for
example by [ST12, Section 7].
We can also describe this as follows, first essentially observed in [Kat08]. Fix ideals
b∞
0
b∞
1
1/pe
∞
:= b∞
1 + ψ∞ (b1 · Lk(V ∞ ) )
1/pe · L
1/pe
= b∞
k(V ∞ ) )
1 + ψ∞ (I∞ + ψ∞ (I∞ · Lk(V ∞ ) )
pie −1
P2
pe −1
i
1/pie
=
i=0 ψ∞ (I∞ · Lk(V ∞ ) )
... ...
pie −1
Pn
e
pe −1
1/pie
1/p
i
∞
∞
∞
= i=0 ψ∞ (I∞ · Lk(V
bn := bn−1 + ψ∞ (bn−1 · Lk(V ∞ ) )
∞))
b∞
2
(4.0.1)
0 (I )
:= I∞ = ψ∞
∞
1/pe
1/pe = I + ψ
:= b0 + ψ∞ (b∞
∞
∞ (I∞ · Lk(V ∞ ) )
0 · Lk(V ∞ ) )
F -SINGULARITIES IN FAMILIES
29
∞
And we notice that this ascending chain stabilizes say at t. Also note that the first time b∞
t = bt+1 ,
∞
∞
∞
we then have bt = bt+1 = bt+2 = . . .. We fix this integer t.
Definition 4.1. With notation as above, we define the integer t to be the uniform integer for τ
and I over the generic point of V , and in general, it will be denoted by nτ (ϕI),k(V ) . We notice that
for any point η ∈ V , by base change we can replace V by Spec k(η) and form a corresponding
integer nτ (ϕη Iη ),k(η) .
On the other hand, without the passing to k(V ∞ ), we have the images
e
e
→ RA1/pe + ϕ((I · L)1/p ) ⊆ RA1/pe ,
b1 := Im I ⊗A A1/p −
pie −1
P2
1/p2e −
i ((I · L pe −1 )1/pie ) ⊗
→
R
⊆ RA1/p2e ,
b2 :=
Im
ϕ
2e
ie A
1/p
1/p
i=0
A
A
...
pie −1
Pn
ie
i
1/pne −
pe −1 )1/p ) ⊗
bn :=
⊆ RA1/pne ,
→
R
ie A
1/pne
1/p
A
i=0 Im ϕ ((I · L
A
...
2e
→ RA1/p2e ⊆ b2 and more generally for j > i that Im bi ⊗A1/pie
Notice that Im b1 ⊗A1/pe A1/p −
je
→ RA1/pne ⊆ bj . Also observe that this is the opposite containment compared to what we
A1/p −
had in Section 3.
By the same argument as in Section 3, we know that there exists an open set U ⊆ V with
W = f −1 (U ) such that
Im(bt ⊗A1/pte A1/p
(4.1.1)
(t+1)e
−
→ RA1/p(t+1)e )|W = bt+1 |W .
As before in Proposition 3.3, we claim that:
Lemma 4.2. With notation as above, Im(bn ⊗A1/pne A1/p
n ≥ t.
(n+1)e
−→ RA1/p(n+1)e )|W = bn+1 |W for all
Proof. Replacing V by U and X by W , we may assume that
(t+1)e
bt+1 = Im bt ⊗A1/pte A1/p
−
→ RA1/p(t+1)e
By induction on n, it suffices to show that, if bn+1 = Im(bn ⊗A1/pne A1/p
bn+2 = Im(bn+1 ⊗A1/p(n+1)e
RA1/p(n+1)e ) if and only if
(n+2)e
A1/p
(n+1)e
−
→ RA1/p(n+1)e ), then
−
→ RA1/p(n+2)e ). Note that bn+1 = Im(bn ⊗A1/pne A1/p
p(n+1)e −1
(n+1)e
−
→
(n+1)e
)−
→ RA1/p(n+1)e )
Im(ϕn+1 ((I · L pe −1 )1/p
n
X
pie −1
(n+1)e
ie
Im ϕi ((I · L pe −1 )1/p ) ⊗A1/pie A1/p
−
→ RA1/p(n+1)e
⊆
i=0
Tensoring with ⊗R L, taking 1/pe th roots, and applying ϕ yields
Im(ϕn+2 ((I · L
⊆
n+1
X
i=1
Hence,
p(n+2)e −1
pe −1
)1/p
(n+2)e
)−
→ RA1/p(n+2)e )
pie −1
(n+2)e
ie
−
→ RA1/p(n+2)e
Im ϕi ((I · L pe −1 )1/p ) ⊗A1/pie A1/p
bn+2 = Im(bn+1 ⊗A1/p(n+1)e A1/p
This finishes the proof.
(n+2)e
−
→ RA1/p(n+2)e ).
30
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
As before, we define relative test ideals as follows.
Definition 4.3. The nth limiting relative test ideal with respect to I (of the pair (X/V, ϕ)) is
defined to be bn ⊆ RA1/pne and is denoted by τn (X/V, ϕI).
Before proceeding to various properties of relative test ideals, we give one example of relative
test ideals.
Example 4.4. (Example 3.7 revisited) Fix k to be an algebraically closed field of characteristic
p > 2, set A = k[t] and set R = k[x, t] with the obvious map X = Spec(R) −
→ V = Spec(A).
1/p
1/p
1/p
1/p
Let ϕ : R
= k[x , t ] −
→ RA1/p = k[x, t ] be the composition of the local generator β ∈
HomR
A1/p
2
1
1
(R1/p , RA1/p ) with pre-multiplication by (xp + t) p = f p . Set I = hxp + ti. Then
b1 = Im(I⊗A A1/p −
→ RA1/p )+ϕ(I 1/p ) = hxp +ti+h(xp +t1/p )(x+t1/p )i = hxp +t, (xp +t1/p )(x+t1/p )i.
2
2
To calculate b2 , it suffices to calculate Im(ϕ2 (I 1/p )) = Im(β 2 ((xp +t)
RA1/p2 , we can see that Im(ϕ2 (I
check that β 2 ((x
p2
+ t)
1+p
p2
x
1/p2
p2 −p−1
p2
)) ⊆ h(x
p2
1/p2
(xp + t)
+ t)
1+p
p2
) = (x
p2
1+p
p2
2
2
I 1/p )). Since (xp +t)
1+p
p2
∈
i. On the other hand, it is straightforward to
+ t)
1+p
p2
2
= (x + t1/p )(xp + t1/p ). Therefore,
2
b2 = hxp + t, (xp + t1/p )(x + t1/p ), (x + t1/p )(xp + t1/p )i.
1+p+···+pi
2
i+1
For each i ≥ 2, to calculate bi+1 , it suffices to calculate Im(β i+1 ((xp + t) pi+1 I 1/p )). It is
2
straightforward to check that it is contained in h(xp + t1/p )(x + t1/p )i and hence is contained in b2 .
Therefore, bi+1 = b2 as well. So we have
(
hxp + t, (xp + t1/p )(x + t1/p )i
when n = 1
bn =
2
p
1/p
p
p
1/p
1/p
1/p
)(x + t )i n ≥ 2.
b2 = hx + t, (x + t )(x + t ), (x + t
We now prove a base change statement for relative test ideals. Compare with Proposition 3.8.
As before, we fix qi : X ×V T i −
→ X ×V V i to be the natural map.
Theorem 4.5 (Relative test ideals and base change). Fix g : T −→ V to be any morphism with T
excellent, integral and admitting a dualizing complex. Then
Im ((qne )∗ τn (X/V, ϕI) ֒→ OXT ne ) := τn (X/V, ϕI) · OXT ne = τn (XT /T, ϕT IT ).
Furthermore, if U = Un satisfies the condition of Lemma 4.2, then W = g−1 (U ) ⊆ T satisfies the
same condition for τn (XT /T, ϕT IT ).
Proof. We write B = OT and work locally. By construction and right exactness of tensor:
τn (X/V, ϕ) · OXT ne = bn ⊗A1/pne B 1/p
ne
=
n
X
pie −1
ie
ϕi ((I · L pe −1 )1/p ) ⊗A1/pie B 1/p
i=0
But by Lemma 2.17, we have
n
X
pie −1
i=0
=
n
X
i=0
ϕiT ((IT
pie −1
pe −1
· LT
= τn (XT /T, ϕT IT )
which completes the proof.
ie
ϕi ((I · L pe −1 )1/p ) ⊗A1/pie B 1/p
ie
ne
)1/p ) ⊗B 1/pie B 1/p
ne
ne
⊆ RB 1/pne .
F -SINGULARITIES IN FAMILIES
31
Theorem 4.6 (Restriction of τn to fibers). With notation as above, there exists an integer N ≥ 0,
such that for all points s ∈ V , N ≥ nτ (ϕs I),k(s) . In other words, we have both that τn (X/V, ϕ) ·
OXsne = τn (Xs /s, ϕs ) (which always holds) and also that for all m ≥ n ≥ N that
m
τn (X/V, ϕI) ⊗V n k(s)1/p = τm (Xs /s, ϕs Is ).
Proof. The idea is the same as for Theorem 3.9 and we only sketch it here. We stratify V as follows.
The result holds on a dense open subset of U0 ⊆ V . Let V1′ be the complement and let V1 denote
the regular locus of V1′ . Base change to V1 , and then repeat. This procedure stops after finitely
many iterations by Noetherian induction and we choose an N that works for them all.
We now construct I ⊆ R whose restriction is contained in the test ideal of every geometric fiber.
Proposition 4.7 (The existence of relative test elements). With notation as above suppose additionally that f : X −→ V is relatively G1 and S2. Then there exists an ideal I ⊂ OX such that
for every point s ∈ V and every perfect extension K ⊇ k(s), we have that IK ⊆ τ (XK , ψK ) where
again ψK is simply ϕK interpreted as in [BS13]. Additionally, we can assume that I = OX at all
e
points of X such that both X/V is smooth and that ϕ locally generates H omOX (L1/p , RA1/pe ).
Furthermore, if V is strongly F -regular and quasi-Gorenstein, then I can additionally be chosen so
that IV e is within the absolute test ideal of (XV e , γ) (where γ is as in Section 3.11).
We caution the reader that it is possible that ψs (and thus ψK ) could be the zero-map, and
hence IK the zero ideal.
e
e
Proof. First let Z1 ⊆ X denote the locus where ϕR1/p ⊆ H omR 1/pe (L1/p , RA1/pe ) is not an
A
isomorphism. Additionally let Z2 denote the locus where X is not smooth over V . Set Z = Z1 ∪ Z2 .
Set W = X \ Z (and note that we can also view this as a subset of XV e for any e) and notice that
on W we know that ϕ can be identified with the trace map up to multiplication by a unit.
Since f |W is smooth, we observe that ϕ (which is identified with trace) is surjective when restricted to W by Corollary 2.5. Now choose an exponent m > 0 such that
IZm · RA1/pe ⊆ a1,1 = σ1 (X/S, ϕ)
Now choose an integer l > 0 such that IZml is locally generated by cubes of elements of IZm (this
only depends on the number of local generators of IZ ). Then the formation of IZml is obviously
compatible with base change (in that the extension of IZml to the base change will also satisfy the
same containment condition.). Thus set I = IZml . The first result follows immediately from [Hoc07,
Theorem on Page 90] or [Sch11b, Proof of Proposition 3.21].
e
For the second result, we notice that γ locally generates H omR 1/pe (N 1/p , RA1/pe ) and that IZm
A
is also in the image of γ. Therefore, I ml works by the above references.
Therefore we obtain:
Corollary 4.8. Using the notation of Theorem 4.6 assume further that I satisfies the condition of
Proposition 4.7. If s is a perfect point then for all n ≥ N as in Theorem 3.9
m
τn (X/V, ϕI) ⊗V n k(s)1/p = τ (Xs , ψs )
where ψs is ϕs interpreted as a p−e -linear map.
Proof. Simply observe that in general the ascending τj (Xs /s, ϕs Is ) ⊆ τj+1 (Xs /s, ϕs Is ) ⊆ . . . stabilize to τ (Xs , ψs ). Furthermore, by Theorem 4.6 this ascending chain stabilizes.
Corollary 4.9. Using the notation of Theorem 4.6 and assume further that I satisfies the condition
of Proposition 4.7. If the perfect closure of the generic fiber of f : X −→ V is strongly F -regular,
i.e. τ (Rk(V ∞ ) , ψ∞ I∞ ) = Rk(V ∞ ) , then there exists an open subset U ⊆ V such that (Xs , ψs ) is also
strongly F -regular for each perfect point s ∈ U .
Proof. It follows immediately from the proof of Proposition 3.3 and Corollary 4.8.
32
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
4.10. Relative test ideals vs absolute test ideals. We make the following assumption.
Convention 4.11. For the rest of Section 4.10, we assume that V is regular and F -finite.
It is natural to try to relate the relative test ideal τn (X/V, ϕI) with the absolute test ideal
τ (XV ne , γ n ) as in Section 3.11. Indeed, if we construct I as we did in the proof of Proposition 4.7
then it follows easily that I · RA1/pne ⊆ τ (XV ne , γ n ). Thus fix such an I.
We now recall the following diagram from Lemma 3.13
L
pne −1
pe −1
1/pne
⊗A1/pne
pne −1 1/p(n+n)e
M pe −1
γ′
µn
L
pne −1
pe −1
p(n−1)e −1 1/p(n−1)e
⊗
L pe −1
A1/pke
denote the extension of I 1/p
µj ’s domain) back onto I
We then observe that:
1/pje
M
p(n−1)e −1
pe −1
A1/p
(n−1)e
1/p(n+n−1)e
A1/p
ne
γ n−1
ϕn−1 ⊗...
RA1/pne
γn
We let I
µn−1
p(n−1)e −1 1/p(n−1)e
⊗
L pe −1
1/pne
ϕn
1/pje
(n−1)e
A1/p
je
to R
1/pje
A1/pke
. Then notice that µj sends I
1/pje
A1/p
(j+n)e
(times
since each µj is surjective.
Lemma 4.12. With notation as above and assuming Convention 4.11, then τn (X/V, ϕI) ⊆ τ (XV ne , γ).
Proof. We know τn (X/V, ϕI) is the sum
!
n
pje −1 1/pje
X
ne
1/pje
1/pne
j
1/p
e
· L p −1
I
ϕ ⊗A1/pje A
(4.12.1)
⊗A1/pje A
j=0
On the other hand, τ (X, γ) is the sum
∞
pje −1 1/p(j+n)e
pje −1 1/pje
X
1/pje
⊗A1/pje M pe −1
γ j I 1/p(j+n)e · L pe −1
(4.12.2)
j=0
A
By our above observations about the surjectivities of the µj above, the sum of the 0th through jth
terms of (4.12.2) is equal to the sum (4.12.1). The result follows.
Additionally, at least over an open subset of the base, we actually have that the relative test
ideal and absolute test ideal agree, cf. Theorem 3.14.
Theorem 4.13. With notation as above and assuming Convention 4.11, choose n > t = nτ (ϕI),k(V ) .
Then there exists a dense open set U ⊆ V ∼
= V e with W = f −1 (U ) ⊆ X such that τ (XV 1/pne , γ)|W =
τn (X/V, ϕI)|W . Furthermore, shrinking U further if necessary and possibly increasing n, we can
require for all perfect points u ∈ U , that
τ (XV 1/pne , γ) · OXu = τn (Xu /u, ϕu Iu ) = τ (Xu , ψu ).
where again ψu is ϕu viewed as a p−e -linear map.
Proof. First we observe that by taking U as in (4.1.1), we can assume that bn−1,n = bn,n . But by
the diagram above, we see that the n − 1st partial sum defining τ (XV 1/pne , γ) in (4.12.2) is equal
to bn−1,n . Likewise the nth partial sum is equal to bn,n . However, once two adjacent partial sums
defining τ (XV 1/pne , γ) coincide, the sum stabilizes for further powers by the computation we made
when defining the b∞
n in (4.0.1). The second statement follows from the first statement and from
Theorem 4.5, cf. the argument of Corollary 4.8.
When dealing with relatively non-F -pure ideals σ, we actually obtained the above restriction
theorem without shrinking X to U in Theorem 3.23. The difference is that for σ we have the easy
F -SINGULARITIES IN FAMILIES
33
containment σ(XV 1/pne , γ n ) ⊆ σn (X/V, ϕ) by Lemma 3.13. For τ however, the easy containment is
reversed. This leads us to the following question:
Question 4.14. Is it true that τ (XV 1/pne , γ) · OXsne = τ (Xs , ψs ) for all perfect points s ∈ V at least
when n ≫ 0?
Rephrasing Theorem 4.13 for divisors we obtain:
Corollary 4.15. Suppose that f : X −→ V is a flat finite map to an excellent regular scheme V .
Additionally suppose that f is relatively G1 and S2 and I is chosen as in Proposition 4.7. Choose
∆ satisfying conditions (a)–(d) of Remark 2.11. Then there exists an open dense set U ⊆ V such
that
τ (XV 1/pne , ∆) · OX 1/pne = τ (Xu , ∆|Xu )
u
for all perfect points u ∈ U .
e
→ RA1/pe .
Proof. Using Lemma 2.24 and Lemma 3.16, we see that ∆ corresponds to some ϕ : L1/p −
Thus we simply apply Corollary 2.22 and Theorem 4.13.
4.16. Applications to strong F -regularity and divisor pairs.
e
→ OX = R) is called strongly F -regular if
Definition 4.17. Recall that a pair (X, ψ : L1/p −
e
→ RA1/pe ) and some I ⊆ R satisfying the condition of
τ (X, ψ) = OX . Given (X/S, ϕ : L1/p −
Proposition 4.7, we say that (X/S, ϕI) is relatively strongly F -regular if we have τn (X/S, ϕI) =
RA1/pne for some n > 0 (equivalently, all n ≫ 0 since the τn ascend), cf. [Has10].
Remark 4.18. Suppose that S is a point and I is chosen as in Proposition 4.7, then (X/S, ϕI) is
relatively strongly F -regular if and only if it is geometrically strongly F -regular (i.e., (Xt , γt ) is
strongly F -regular for every point t −
→ S). The argument is the same as in Remark 3.29
The same ideas imply the strongly F -regular locus of a proper map is open.
Corollary 4.19 (Openness of the strongly F -regular locus). With notation as before, assume
additionally that f : X −→ V is proper and that I satisfies the condition of Proposition 4.7. Assume
that s ∈ V is a point and that (Xs /s, ϕs Is ) is relatively strongly F -regular (for example, if s is a
perfect point, this just means it is strongly F -regular and is independent of I). Then there exists a
dense open set U ⊆ V containing s such that (Xu /u, ϕu Iu ) is relatively strongly F -regular for all
u ∈ U (in particular, (Xu , ϕu ) is strongly F -regular for all perfect points u ∈ U ).
Proof. Choose n ≫ 0 such that τn (Xs /s, ϕs Is ) = RO1/pne since (Xs /s, ϕs Is ) is strongly F -regular.
s
By Theorem 4.6, we know that τn (X/V, ϕI) = RA1/pne in a neighborhood W ⊆ X of Xs . Let
Z = X \ W ⊆ X be the complement of that neighborhood. Since f is proper, f (Z) is closed,
and also doesn’t contain s. Set U = V \ f (Z). Then τn (XU /U, ϕU I|U ) = RO1/pne . It follows from
U
Theorem 4.6 that all the fibers (Xu /u, ϕu Iu ) are relatively strongly F -regular as desired.
At least for proper maps, we also obtain that the definition of relative strongly F -regularity
Definition 4.17 is independent of the choice of I.
Lemma 4.20. Suppose that f : X −→ V is proper and that (X/V, ϕI) is relatively strongly F regular for some I satisfying the condition of Proposition 4.7. Then for all J ⊆ R such that Js
is non-zero on every component of every fiber Xs , we have that τn (X/S, ϕJ) = RA1/pne for some
n > 0 (equivalently, all n ≫ 0).
In particular in Definition 4.17, it would be equivalent to require τn (X/S, ϕJ) = RA1/pne for all
such J and some n.
34
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
Proof. Choose a closed point s ∈ V and a perfect extension K ⊇ k(s). Then τm (XK /K, ϕK IK ) =
RK 1/pne for some m > 0 by Theorem 4.6. But since K is a perfect field extension, this implies that
τm (XK , ψK IK ) = RK 1/pme as well. On the other hand, since (XK , ψK ) is strongly F -regular, we
see that τn (XK /K, ϕK JK ) = RK 1/pne for some n > 0. Hence since k(s) ⊆ K is faithfully flat, we
see that τn (Xs /s, ϕs Js ) = Rk(s)1/pne . But then since τn (X/V, ϕJ) restricts to τn (Xs /s, ϕs Js ) we
observe that τn (X/V, ϕJ) = RA1/pne at least in a neighborhood of s using that f is proper and the
same argument we made in Corollary 4.19. But we can find such an n for every s ∈ V , and so the
lemma holds by the quasi-compactness of V .
We now state our result in the divisorial case.
Corollary 4.21. With notation as above, suppose that f : X −→ V is a proper map and that ∆
is a Q-divisor satisfying conditions (a)–(d)8 of Remark 2.11. Additionally suppose that for some
perfect point s ∈ V , the fiber (Xs , ∆|Xs ) is strongly F -regular. Then there exists a dense open set
U ⊆ V containing s such that (Xu , ∆|Xu ) is strongly F -regular for all perfect u ∈ U .
e
→ RA1/pe corresponding
Proof. Using Remark 2.11 we construct a relatively divisorial ϕ : L1/p −
to ∆. Choose now I satisfying the condition of Proposition 4.7. It follows that (Xs , ψs ) is strongly
F -regular and hence that (Xs /s, ϕs Is ) is relatively strongly F -regular since s is a perfect point.
Then Corollary 4.19 and Corollary 2.22 complete the proof.
By a perturbation trick we can also handle the case that the index of KX + ∆ is divisible by
p > 0, at least over curves.
Corollary 4.22. With notation as above, suppose that f : X −→ V is a projective map to a regular
1-dimensional base, and that ∆ is a Q-divisor satisfying the following 4 conditions:
1
D for some Weil divisor D.
(a′ ) ∆′ = m
(b) D is a Weil divisor on X which is Cartier in codimension 1 and Cartier at every codimension
1 point of every fiber.
(c) D is trivial along the codimension-1 components of the non-smooth locus of X −→ V and
the codimension-1 components of
∗∗the non-smooth locus of every fiber.
l
is a line bundle.
(d′ ) l/m ∈ Z and ωX/V
⊗ OX (l∆′ )
Additionally suppose that for some perfect point s ∈ V , the fiber (Xs , ∆|Xs ) is strongly F -regular.
Then there exists a dense open set U ⊆ V containing s such that (Xu , ∆|Xu ) is strongly F -regular
for all perfect u ∈ U .
Proof. We may certainly suppose that V is affine. Since the fiber Xs is normal, and hence geometrically normal, we may also assume that the nearby fibers satisfy the same condition (using that
f is proper). Thus we may assume that all the fibers are geometrically normal and in fact that
the map f : X −
→ V is geometrically integral. Without loss of generality, by base change we can
assume that V is normal and hence X is normal itself. Thus we may assume that X is normal.
Now, since the base is 1-dimensional, we claim that can assume that KX/V doesn’t contain any
fiber. The only fiber we must worry about is Xs (as the others can be handled by shrinking V ).
We argue as follows: note first that KX/V can be viewed as an honest Weil divisor since we already
assumed that X is normal. On the other hand, each fiber is a Cartier divisor. Hence, if KX/V is
non-trivial along the generic point of Xs , by twisting by the pullback of s, we can assume that
KX/V does not contain Xs .
Write l∆′ = cpe0 ∆′ where p does not divide c. Let E be an effective Cartier divisor on X not
containing any fiber such that KX/V + E is effective. We also observe that KX/V + E satisfies
8For example, these conditions hold if V is regular, f : X −
→ V is geometrically normal, KX + ∆ is Q-Cartier
with index not divisible by p and ∆ does not contain any fiber of f in its support.
F -SINGULARITIES IN FAMILIES
conditions (b) and (c) above. Let Γ =
E+KX/V +∆′
pe −1
35
for some e ≫ e0 . Then consider
∆ := ∆′ + Γ.
Note that
c(pe − 1)(KX/V + ∆)
= c(pe − 1)(KX/V + ∆′ + Γ)
= c(pe − 1)(KX/V + ∆′ ) + c(E + KX/V + ∆′ )
= cpe (KX/V + ∆′ ) + cE
which is certainly Cartier. Hence KX/V + ∆ satisfies conditions (a)–(d) from Remark 2.11. On the
other hand, since e ≫ 0, we know that (Xs , ∆|Xs ) = (Xs , ∆′ |Xs + Γ|Xs ) is still strongly F -regular
and so by Corollary 4.21 there exists an open set U such that (Xu , ∆|Xu ) is strongly F -regular for
all u ∈ U . But ∆ ≥ ∆′ and so the result follows for ∆′ as well.
Remark 4.23. In the case of a normal X and in the non-relative case, we know that τ (X; ∆) =
τ (ωX , KX + ∆). Furthermore, we then know that
Tre (F∗e τ (ωX , KX + ∆)) = τ (ωX ,
1
1
(KX + ∆)) = τ (X, e (KX + ∆) − KX ).
e
p
p
Reversing this process gives us a nice means to compute τ (X, ∆) when the index of KX + ∆ is
divisible by p. It would be natural try to prove a relative version of this, which may yield a suitable
definition of relative test ideals for KX + ∆ of any index. We won’t attempt this here.
5. Relative test submodules, F -rationality and F -injectivity
Our goal in this (somewhat shorter) section is to explore relative test submodules and non-F injective modules. Throughout this section, we assume that f : X −
→ V is a Cohen-Macaulay
morphism. This provides us with base change for relative canonical sheaves ωX/V [Con00].
We first define relative non-F -injective modules σn (X/V, ωX/V ) and relative test submodules
τn (X/V, ωX/V ). Recall the trace map, Lemma 2.18
n
ne
1/p
→ ωXV n /V n ∼
ΦX/V,n : ωX/V = ωX n /V n −
= ωX/V ⊗A A1/p .
Here the final isomorphism follows from [Con00, Theorem 3.6.1] since f is a Cohen-Macaulay
morphism. We will form σ and τ relative to these maps, instead of relative to the map ϕ discussed
previously. One key point to remember is that by Lemma 2.18, the map ΦX/V,n is compatible with
arbitrary base change (since in this section, f is a Cohen-Macaulay morphism). We also observe
that the composition of trace maps
1/pm−n
(5.0.1)
ΦX/V,m−n
ωX m /V m −−−−−−→ ω
XVn m /V m
ΦX/V,n ⊗A1/pn A1/p
m
m
−−−−−−−−−−−−−→ ωX/V ⊗A A1/p ∼
= ωXV m /V m .
can be identified with ΦX/V,m .
5.1. The definition and basic properties of σn (X/V, ωX/V ). For each integer n > 0, define
n
cn := Im(ΦX/V,n ) ⊆ ωXV n /V n ∼
= ωX/V ⊗A A1/p . Furthermore, for each m ≥ n, using the factorization in (5.0.1) it is easy to see that
m
→ ωR 1/pm
(5.1.2)
cm ⊆ Im cn ⊗A1/pn A1/p −
A
just as we observed in Section 3.
Definition 5.2. With notation as above, we define the nth relative non-F -injective submodule,
denoted σn (X/V, ωX/V ) to be cn ⊆ ωXV n /V n .
36
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
As before, we will prove a stabilization statement for cn and cm , in particular that the containment
(5.1.2) is an equality for all m > n ≫ 0. By base changing with k(V ∞ ), the perfection of the residue
field of the generic point of V , we can again find an integer n0 > 0 and an open set U ⊆ V such
that
n0 +1
cn0 +1 |f −1 U = Im cn0 ⊗A1/pn0 A1/p
−
→ ωR 1/pn0 +1 |f −1 U .
A
Again observe that for
U inside the regular locus of V , we can identify the image above with
n
n +1
m
cn0 ⊗A1/pn0 A1/p 0 |U by the flatness of A1/p over A1/p . We then obtain:
Proposition 5.3. Fix notation as above. For every integer n ≥ n0 , there exists a nonempty open
subset Un ⊆ V of the base scheme V satisfying the following condition. If one sets Xn = f −1 (Un ),
then we have that for every m ≥ n
m
(5.3.1)
σm (X/V, ωX/V )|Xn = Im σn (X/V, ωR/A ) · OXn ⊗A1/pn A1/p −→ ωR 1/pm .
A
Furthermore we may assume that Un0 ⊆ Un0 +1 ⊆ · · · ⊆ Un ⊆ Un+1 ⊆ · · · form an ascending chain
of open sets.
Proof. The proof is identical to that of Proposition 3.3 and so we omit it.
We now point out that relative non-F -injective modules behave well with respect to base change.
Recall that if g : T −
→ V is a map, then qn : XT n −
→ XV n is the induced map.
Proposition 5.4 (Base change for σn (X/V, ωX/V )). Suppose that g : T −→ V is a map from an
excellent scheme with a dualizing complex, then using the notation of Section 2.16
Im (qn )∗ σn (X/V, ωX/V ) → ωXT n /T n = σn (XT /T, ωXT /T ).
Furthermore, if U = Un satisfies condition (5.3.1) from Proposition 5.3, then W = g−1 (U ) ⊆ T
satisfies the same condition for σn (XT /T, ωXT /T ).
Proof. The first statement is an immediate consequence of base change relative canonical sheaves
n+1
and ΦX/V,n+1 have the same image
and trace Lemma 2.18. For the second, if ΦX/V,n ⊗A1/pn A1/p
in ωXV n /V n , then its easy to see that the base changed maps also have the same images.
Tre
Recall for any F -finite scheme Y with canonical module ωY , then σ(ωY ) is equal to ℑ(F∗e ωY −−→
ωY ) for any e ≫ 0.
Corollary 5.5 (Restriction theorem for σn (X/V, ωX/V )). With notation as above, there exists an
integer N > 0 such that for every perfect point s ∈ V , we have
ne
σ(ωXs ) = Im σn (X/V, ωX/V ) ⊗A1/pne k(s)1/p −→ ωXs /s .
for all n > N .
Proof. Taking g : s −
→ V in the previous theorem and so obtain that the image is equal to
σn (Xs /s, ωXs /s ). Furthermore, since k(s) is perfect, we have containments . . . ⊇ σn (Xs /s, ωXs /s ) ⊇
σn+1 (Xs /s, ωXs /s ) ⊇ . . . with a descending intersection that coincides with σ(ωXs ). Furthermore,
by Proposition 5.4, over a dense open set U ⊆ V and some N > 0 we have σn (Xs /s, ωXs /s ) =
σn+1 (Xs /s, ωXs /s ) and hence σn (Xs /s, ωXs /s ) = σ(ωXs ) by the construction of σ(ωXs ) for all n ≥
N0 . Let V1 = V \ U , base change with V1 and obtain the result over a dense open subset U1 of V1
(for some N1 ). By Noetherian induction, this process terminates.
F -SINGULARITIES IN FAMILIES
37
5.6. The definition and basic properties of τn (X/V, ωX/V ). The goal of this section is to
develop the basics of a relative theory of test submodules. We fix the notation of the previous
sections and additionally assume that X is geometrically normal over V which we now also assume
is regular. We let J = JX/V ⊆ R be the Jacobian ideal sheaf of X over V . We observe that the
formation of JX/V commutes with base change in the following sense: for any T −
→ V , we have
JX/V · OXT = JXT /T . To see this, just note that JX/V can be defined as a Fitting ideal of ΩX/V
([HS06, Discussion 4.4.7]) and that the formation of ΩX/V ([Eis95, Proposition 16.4]) and Fitting
ideals ([HS06, Discussion 4.4.7]) commutes with arbitrary base change. Note that J is nonzero at
any generic point of X since X is geometrically reduced. Furthermore, on every perfect fiber, the
Jacobian ideal is contained in the (big) test ideal [Hoc07, Theorem on Page 213].
We have the images
→ ωR 1/p /A1/p + ΦX/V,1 ((J · ωR )1/p ) ⊆ ωR 1/p ,
d1 := Im J · ωR 1/p /A1/p −
A
A
A
P2
2
i
1/p
1/p
−
→ ωR 1/p2 ⊆ ωR 1/p2 ,
) ⊗A1/pi A
d2 :=
i=0 Im (ΦX/V,i (J · ωR/A )
A
A
...
Pn
1/pn −
1/pi ) ⊗
→
ω
⊆ ωR 1/pn ,
dn :=
Im
(Φ
(J
·
ω
)
i A
R
n
1/p
X/V,i
R/A
i=0
1/p
A
A
A
...
2
→ ωR 1/p2 ⊆ d2 and more generally for j > i that
Notice that d1,2 := Im d1 ⊗A1/p A1/p −
A
j
1/p
−
→ ωR 1/pn ⊆ dj .
di,j := Im di ⊗A1/pi A
A
By the same argument as in Section 3, we know that there exists an open set U ⊆ V with
(t+1)e
−
→ ωR 1/pt+1 |W = dt+1 |W .
W = f −1 (U ) such that Im dt ⊗A1/pte A1/p
A
Lemma 5.7. With notation as above, Im dn ⊗A1/pne A1/p
all n > t.
(n+1)e
−→ ωR
(n+1)e
A1/p
|W = dn+1 |W for
Proof. The proof is the same as the proof of Lemma 4.2 and so we omit it.
Definition 5.8 (Relative test submodules). With notation as above (in particular, ωX/V is still
compatible with base change), we define the nth iterated relative test submodule to be dn and denote
it by τn (X/V, ωX/V ).
We now discuss base change for relative test ideals.
Proposition 5.9 (Base change for τn (X/V, ωX/V )). Suppose that g : T −→ V is a map from a
excellent scheme with a dualizing complex, then using the notation of Section 2.16
Im (qn )∗ τn (X/V, ωX/V ) → ωXT n /T n = τn (XT /T, ωX/T ).
Furthermore, if U = Un satisfies condition from Lemma 5.7, then Y = g −1 (U ) ⊆ T satisfies the
same condition for τn (XT /T, ωXT /T ).
Proof. It is just as before since ωX/V and J are compatible with arbitrary base change.
Corollary 5.10 (Restriction theorem for τn (X/V, ωX/V )). With notation as above, there exists an
integer N > 0 such that for every perfect point s ∈ V , we have
ne
Im τn (X/V, ωX/V ) ⊗A1/pne k(s)1/p −→ ωXs /s = τ (ωXs ).
for all n > N .
38
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
Proof. Taking g : s −
→ V in Proposition 5.9 and so obtain that the image is equal to τn (Xs /s, ωXs /s ).
Furthermore, since k(s) is perfect, we can make identifications so as to have containments . . . ⊆
τn (Xs /s, ωXs /s ) ⊆ τn+1 (Xs /s, ωXs /s ) ⊆ . . . with a ascending union that coincides with τ (ωXs ). Furthermore, by Proposition 5.4, over a dense open set U ⊆ V and some N > 0 we have τn (Xs /s, ωXs /s ) =
τn+1 (Xs /s, ωXs /s ) for all n ≥ N0 and hence τn (Xs /s, ωXs /s ) = τ (ωXs ) by the computation of (4.0.1).
Let V1 = V \ U , base change with V1 and obtain the result over a dense open subset U1 of V1 (for
some N1 ). By Noetherian induction, this process must terminate.
Remark 5.11 (Relative versus absolute τ (ω)). It would be natural to relate the relative σ(ω) and
τ (ω) with the absolute σ(ω) and τ (ω). While the authors believe that this is possible along the
lines of Section 3.11 or Section 4.10, we won’t work out the details here.
5.12. Applications to families of F -injective and F -rational singularities. In this section,
we obtain new proofs of results of M. Hashimoto [Has01], deformation of F -injectivity and F rationality in proper flat families. Note Hashimoto proved the local version of these results (and in
the F -rationality case, over a variety). However, the local results generalize to the non-local case
via straightforward computations. We begin with a definition essentially first made by Hashimoto.
Definition 5.13 (cf. [Has01, Has10]). We say that X/V (which is still assumed to be relatively
Cohen-Macaulay) is relatively F -injective if for some n > 0 we have σn (X/V, ωX/V ) = ωXV n /V n .
Likewise, X/V is relatively F -rational if for some n > 0 we have τn (X/V, ωX/V ) = ωXV n /V n .
Remark 5.14. If V is the spectrum of a perfect field k, it is easy to see that f : X −
→ V is
relatively F -injective (respectively relatively F -rational) if and only if X is F -injective (respectively
F -rational) in the usual sense [ST12, Section 8] via an identification of k ∼
= k1/p . Note we are
implicitly assuming that X is Cohen-Macaulay in this case.
Lemma 5.15. (cf. [Has01, Proposition 5.5]) With notation as above, if σn (X/V, ωX/V ) = ωXV n /V n
for some n > 0, then σm (X/V, ωX/V ) = ωXV m /V m for all m > 0 divisible by n. Furthermore, if V
is regular, then the result holds for all m ≥ n. Additionally, if τn (X/V, ωX/V ) = ωXV n /V n for some
n > 0, then τm (X/V, ωX/V ) = ωXV m /V m for all m ≫ 0.
Proof. We begin with σ. We notice that
2n
→ ωXV 2n /V 2n
σ2n (X/V, ωX/V ) = Im Im ωX n /V n −
→ ωXVn /V n ⊗A1/pn A1/p −
and our hypothesis σn (X/V, ωX/V ) = ωXV n /V n implies that the inner map is surjective. But then
the outer map is surjective too by right exactness of tensor and so σ2n (X/V, ωX/V ) = ωXV 2n /V 2n .
The general case repeats this process and so follows similarly.
Now we assume that V is regular for the second statement about σn . We fix an n > 0 such that
σn (X/V, ωX/V ) = ωXV n /V n . Then for all m ≤ n, we have
n
n
σm (X/V, ωX/V ) ⊗A1/pm A1/p ⊇ σn (X/V, ωX/V ) = ωXV n /V n ∼
= ωXV m /V m ⊗A1/pm A1/p
n
n
and so σm (X/V, ωX/V ) ⊗A1/pm A1/p = ωXV m /V m ⊗A1/pm A1/p . On the other hand, if we know
n
σm (X/V, ωX/V ) ( ωXV m /V m then the previous equality is impossible since A1/p faithfully flat over
A [Kun69] (since V is regular).
Handling τ is easy. Again use that ωX/V is compatible with base change, since
m
→ ωXVm /V m
ωXVm /V m = Im ωX/V ⊗A1/pn A1/p −
m
→ ωXVm /V m
= Im τn (X/V, ωX/V ) ⊗A1/pn A1/p −
= dn,m
⊆ dm
= τn (X/V, ωX/V )
⊆ ωXVm /V m
F -SINGULARITIES IN FAMILIES
39
Theorem 5.16. (Deformation of F -rationality and F -injectivity, cf. [Has01, Theorem 5.8, Remark
6.7]) Suppose that f : X −→ V is a proper flat finite type equidimensional reduced Cohen-Macaulay
morphism to an excellent integral scheme V with a dualizing complex. Suppose that for some point
s ∈ V , the fiber Xs /s is relatively F -injective (respectively, F -rational). Then there exists an open
neighborhood U ⊆ V containing s such that Xu −→ u is relatively F -injective (respectively, F rational) for all u ∈ U .
Proof. We first show that σn (ωX/V ) = ωXV ne /V ne at each point of the fiber Xs . Indeed, let z ∈ X
be a point on Xs and let I denote the ideal sheaf of Xs . We observe that for some n, the natural
map σn (ωX/V )/(I · σn (ωX/V )) −
→ ωX/V /(I · ωX/V ) = ωXs /Vs is surjective. This is preserved after
localizing at z, and so since I ⊆ mz , we see that the generators of the stalk (σn (ωX/V ))z generate
(ωX/V )z by Nakayama’s lemma. Hence (σn (ωX/V ))z = (ωX/V )z . Since this holds at every point
z ∈ Xs ⊆ X, it holds in a neighborhood of Xs .
As before, let Z denote the locus where σn (ωX/V ) 6= ωXV ne /V ne . This is closed and since f is
proper, its image f (Z) is closed too. But f (Z) is then a closed set not containing s. The result
follows for F -injectivity. The proof for F -rationality is the same.
6. Global applications
The purpose of this section is to develop a global theory of the previous sections for a projective
family f : X −
→ V . In the last few years, there has been a new push to use Frobenius and the
trace map to replace the Kodaira vanishing theorem. In this section we extend some of these ideas
to families. We study how the canonical linear subsystems S 0 (Xs , σ(Xs , ∆s ) ⊗ Ms ) ⊆ H 0 (Xs , Ms ),
introduced in [Sch11a], behave as we vary s ∈ V . Furthermore, as mentioned in the introduction
we also obtain some global generation and semi-positivity statements.
6.1. Basic definitions. We use the following setup throughout Section 6.
→ V is projective and
Notation 6.1. In the situation of Notation 2.1, assume also that f : X −
additionally that V is regular (which implies that FVe : V e −
→ V is flat). Furthermore, fix a line
bundle M on X. Sometimes we also assume the following (in which case we write Notation 6.1* ):
(*): there is an integer N ≥ 0, such that for every integer m ≥ N ,
σm (X/V, ϕ) = σN (X/V, ϕ) ⊗
1
A
1
pNe
A pme .
In this situation we denote σN (X/V, ϕ) by σN . We notice that this condition (*) always holds over
a dense open set of the base by Proposition 3.3.
Remark 6.2. Note that since X (resp. V ) is topologically isomorphic to XVneme (resp. V me ) for every
→ V me .
m ≥ n, f∗ can be identified with g∗ , where g is any of the induced morphisms XVneme −
ne
→ V me
Hence, we use only f∗ for all purposes, even when g∗ for one of the above maps g : XV me −
would be more natural. The downside of this notation is that it does not show if a sheaf has a
me
A1/p structure, and consequently its pushforward by f a OV me structure. We decided to still use
it, because it simplifies greatly the notations.
Summarizing: when reading the following arguments it is important to trace through the space
XVneme on which the adequate sheaves live. Then f∗ of these sheaves will live on V me .
Notation 6.3. When dealing with sheaves G on V ne , we will frequently pull them back to V me
1/pme
⊗O1/pne G, we will write
for m ≥ n. When doing this, instead of writing OV me ⊗OV ne G or OV
V
me
simply V
×V ne G or GV me . We trust this abuse of notation will cause no confusion, as it helps
compactify the notation substantially.
40
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
Our approach to understanding how the canonical linear systems of [Sch11a] behave in families
is to define a relative version of them, and then show certain base-change properties. These objects
will be relative versions of σ for global sections, and they play the same role for σn that T 0 and S 0
plays for τ and σ in [BST11] and [Sch11a] respectively.
Definition 6.4. In the situation of Notation 6.1, define
Sϕ0 n f∗ (M )
:= im f∗
L
pne −1
pe −1
1
pne
⊗R M
!
!
1
ne
.
−−−−−−−−−→ f∗ A p ⊗A M
f∗ (ϕn ⊗R idM )
Note that Sϕ0 n f∗ (M ) is a sheaf on V ne and it is a subsheaf of (f∗ (M ))V ne by flat base-change. In
0
case (X, ∆) is a pair we define S∆,ne
f∗ (M ) := Sϕ0 n f∗ (M ) (assuming (pne − 1)(KX + ∆) is Cartier
∆
0 f (M ) for S 0
and ϕ is the corresponding map). If ∆ = 0, then we write Sne
∗
∆,ne f∗ (M ).
Since the image of ϕn ⊗R idM in the above definition is σn (X/V, ϕ)⊗R M the following proposition
is immediate.
Proposition 6.5. In the situation of Notation 6.1,
Sϕ0 n f∗ (M ) ⊆ f∗ (σn (X/V, ϕ) ⊗R M ).
Our first goal is to show that the images in Definition 6.4 descend (up to appropriate base change
by Frobenius). Compare with the containments ai,n ⊇ ai+1,n ⊇ · · · ⊇ an,n of Section 3.
Proposition 6.6. For all integers m ≥ n ≥ 0,
(6.6.1)
V me ×V ne Sϕ0 n f∗ (M ) ⊇ Sϕ0 m f∗ (M )
as subsheaves of V me ×V f∗ (M ).
Remark 6.7. To be precise the left and right hand side of (6.6.1) are subsheaves of
1
1
V me ×V ne f∗ (A pne ⊗A M ) and of f∗ (A pme ⊗A M ),
1
1
respectively. However, both V me ×V ne f∗ (A pne ⊗A M ) and f∗ (A pme ⊗A M ) are canonically isomorphic
to V me ×V f∗ (M ) via flat base-change (since V is regular).
F -SINGULARITIES IN FAMILIES
41
Proof of Proposition 6.6. The following commutative diagram shows that f∗ (ϕm ⊗R idM ) factors
through V me ×V ne f∗ (ϕn ⊗R idM ).
V me ×V ne f∗
L
pne −1
pe −1
1
pne
⊗R M
!
V me ×V ne f∗ (ϕn ⊗R idM )
∼
=
1
f∗ A pme ⊗
A
1
pne
1
f∗ A pme ⊗
A
1
pne
(ϕn ⊗R idM )
1
1
/ f∗ A pme ⊗A M
R
A
1
p(m−n)e
⊗R L
O
pne −1
pe −1
1
pne
⊗R M
f∗ ϕm−n ⊗R id
!
pne −1
e
L p −1
/ f∗ A pme ⊗A M
O
!
1
pne
⊗R idM
1
! pne
1
(m−n)e (m−n)e
ne −1
p
p
−1
p
⊗R L pe −1
⊗R M
f∗
L pe −1
=
∼
=
f∗
flat basechange
∼
=
flat base-change [Har77,
Proposition III.9.3] of f∗
by V me −→ V ne
!
1
ne ne
p
p −1
L pe −1
⊗R M
1
/ V me ×V ne f∗ A pne ⊗A M
f∗ (ϕm ⊗R idM )
∼
=
projection formula
f∗
L
pme −1
pe −1
1
pme
⊗R M
!
BC
Hence, the statement of the proposition holds by the following computation.
Sϕ0 m f∗ (M ) = im (f∗ (ϕm ⊗R idM ))
⊆ im (V me ×V ne f∗ (ϕn ⊗R idM ))
= V me ×V ne im (f∗ (ϕn ⊗R idM )) .
|
{z
}
=V
V ne −→ V me is flat
me
×V ne Sϕ0 n f∗ (M )
6.2. Auxiliary definition and stabilization. We would now like to obtain a global result similar
to Proposition 3.3. In particular, we’d like the containments of Proposition 6.6 to be equality over
a dense open subset subset of the base V . There is a complicating factor however, while we can
still find an open set U of V such that
V (n+1)e ×V ne Sϕ0 n f∗ (M )
= Sϕ0 n+1 f∗ (M )
U
U
we do not see how to use this to show that we have the n + 1 to n + 2 equality without additional
assumptions. The issue is that in the proof of Proposition 3.3, twisting by line bundles is exact.
For Sϕ0 n however we also push forward. Therefore, in order to obtain our stabilization over a dense
open set of the base we need additional positivity assumptions on M and L.
42
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
Furthermore, we need an auxiliary version of Sϕ0 which involves the σN := σN (X/V, ϕ) from
pme −1
Notation 6.1*. To do this, first observe that since f is flat, the tensor product L pe −1 ⊗R σN is
pme −1
pme −1
Ne
naturally identified with a subsheaf of L pe −1 ⊗R RA1/pNe ∼
= L pe −1 ⊗A A1/p . In order to motivate
this auxiliary definition, we make the following observation:
Lemma 6.8. In the situation of Notation 6.1* , we have that the image of the natural map
1
1
me
me
me
me
(N+m)e
p
p
ϕm ⊗A1/pme A1/p
p
−1
p
−1
e −1
e −1
p
p
−−−−−−−−−−−−−−−→ RA1/p(N+m)e
⊗ R σN
⊗R RA1/pNe
֒→ L
αN +m : L
is equal to σN ⊗A1/pNe A1/p
(N+m)e
∼
= σm+N (X/V, ϕ).
pme −1
Proof. The first term is the image of the pme th root of L pe −1 ⊗R ϕN . Therefore the image of the
composition also equals σm+N (X/V, ϕ) by our construction of ϕj in Section 2.13.
This stabilization suggests it is reasonable to make the following definition.
Definition 6.9. In the situation of Notation 6.1* ,
!
1
me
me
1
p
p
−1
f∗ (αN+m ⊗R M )
pe −1 ⊗
p(n+N)e ⊗
Sϕ0 n ,σN f∗ (M ) := im
M
σ
f
⊗
M
L
−
−
−
−
−
−
−
−
−
−
→
f
A
A
R
N
∗
R
∗
where αN +m is as in Lemma 6.8.
Consider the following proposition relating the various S 0 f∗ objects so far described.
Proposition 6.10. In the situation of Notation 6.1* , for every integer m ≥ 0,
V (N +m)e ×V me Sϕ0 m f∗ (M ) ⊇ Sϕ0 m ,σN f∗ (M ) ⊇ Sϕ0 m+N f∗ (M )
(Here all sheaves are regarded as subsheaves of V (N +m)e ×V f∗ (M ) via flat base-change.)
Proof. This follows directly from the definitions. The first containment is trivial, and the second
follows from a factorization similar to the one from Proposition 6.6.
Proposition 6.11. In the situation of Notation 6.1* ,
(a) for all integers m ≥ n ≥ 0,
V (N +m)e ×V (N+n)e Sϕ0 n ,σN f∗ (M ) ⊇ Sϕ0 m ,σN f∗ (M ),
as subsheaves of V (N +m)e ×V f∗ (M ), and
e
(b) if furthermore L ⊗ M p −1 is f -ample, then there is an integer n > 0, such that for every
integer m ≥ n, the above inclusion is equality.
(c) if furthermore M = Ql ⊗ P , where Q is f -ample, then there is an integer l0 > 0, such that
for every integer l ≥ l0 , m ≥ n ≥ 0 and nef line bundle P the above inclusion is equality.
Proof. Define the coherent sheaf Bϕ on XV (N+1)e as the kernel of the top horizontal map in the
following commutative diagram.
(6.11.1)
Bϕ
1
/ (L ⊗R σN ) pe
OO
ϕ⊗
1
/ / A p(N+1)e ⊗ 1 σN
pNe
OO A
1
(idL ⊗R ϕN ) pe
1
1 ! pe
Ne Ne
p
−1
p
L ⊗ L pe −1
ϕN+1
∼
=
L
p(N+1)e −1
pe −1
1
p(N+1)e
F -SINGULARITIES IN FAMILIES
43
1
Note here we used that the image of ϕN +1 is A p(N+1)e ⊗
1
A pNe
σN by the assumptions made in
Notation 6.1* . Also, the horizontal arrow is surjective by either Lemma 6.8 or from the diagram
!
1
me
me
p
p
−1
(which is how we proved Lemma 6.8). Then one can apply f∗
L pe −1 ⊗R
⊗R M
to the top row of (6.11.1), which is shown in the following commutative diagram. We have also
included important isomorphisms to the different terms of the exact sequence.
f∗
L
pme −1
pe −1
⊗ R Bϕ
_
1
pme
⊗R M
f∗
L
pme −1
pe −1
⊗R (L ⊗ σN )
1
pe
1
pme
!
⊗R M
!
∼
=
f∗
L
p(m+1)e −1
pe −1
⊗ R σN
1
p(m+1)e
ν
f∗
L
pme −1
pe −1
⊗ R σN ⊗
1
A pNe
A
1
p(N+1)e
R1 f
f∗
L
pme −1
pe −1
∗
L
⊗ R σN
pme −1
pe −1
1
pme
⊗ R Bϕ
⊗
Ap
1
pme
1
(N+m)e
1
pme
k
⊗R M
A
⊗R M
!
!
1
p(N+m+1)e
∼
=
!s
⊗R M
!
∼
=
by flat base change
V
(N +m+1)e
×V (N+m)e f∗
V
(N +m+1)e
L
pme −1
pe −1
×V (N+m)e
⊗ R σN
1
pme
Sϕ0 m ,σN f∗ (M )
⊗R M
!
Sϕ0 m+1 ,σN f∗ (M )
o
Both V (N +m+1)e ×V (N+m)e Sϕ0 m ,σN f∗ (M ) and Sϕ0 m+1 ,σN f∗ (M ) can be regarded as subsheaves of
V (N +m+1)e ×V f∗ (M ) via flat base change. Furthermore, the bottom horizontal arrow becomes a
map of subsheaves, i.e., an injection, this way. This shows point (a).
e
To prove point (b), we are supposed to prove that whenever L ⊗ M p −1 is f -ample, this arrow
is also surjective for m ≫ 0. This would follow if the third vertical arrow (from above), labeled ν,
was surjective. Therefore, it is sufficient to show that for m ≫ 0,
0 = R 1 f∗
L
pme −1
pe −1
⊗ R Bϕ
1
pme
⊗R M
!
= R 1 f∗
= R 1 f∗
L
pme −1
pe −1
⊗ R Bϕ ⊗ R M
L⊗M
pe −1
−1
ppme
e −1
pme
1
pme
!
⊗ R Bϕ ⊗ R M
1
pme
!
.
⊗R M
!
44
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
1
Note that since applying ( ) pme does not change the sheaf of abelian groups structure, this is
equivalent to showing
pme −1
e
0 = R1 f∗ L ⊗ M p −1 pe −1 ⊗R Bϕ ⊗R M
−1
pme
e −1
1
p
e
⊗R
(Bϕ ⊗R M ) .
= R1 f∗ L ⊗ M p −1 ⊗A A pN+1
(N+1)e
A1/p
e
1
Furthermore, L ⊗ M p −1 ⊗A A pN+1 is a relatively ample line bundle by the assumption of point (b)
and Bϕ ⊗R M is a coherent sheaf on XV (N+1)e . Therefore, relative Serre vanishing concludes our
proof.
Point (c) follows immediately from the above argument. Indeed, if M = Ql ⊗ P , then by relative
Fujita vanishing [Kee03, Theorem 1.5] there is an integer l0 > 0, such that the above vanishing
holds for every m ≥ 0, l ≥ l0 and nef line bundle P .
e
Corollary 6.12. In the situation of Notation 6.1* , if L⊗M p −1 is f -ample, then there is an integer
n > 0, such that for every integer m ≥ n
= Sϕ0 m−N ,σN f∗ (M )
V me ×V ne Sϕ0 n f∗ (M ) = Sϕ0 m f∗ (M )
as subsheaves of V me ×V f∗ (M ). Furthermore in the situation of point (c) of Proposition 6.11 where
M = Ql ⊗ P = (ample)l ⊗(nef) with l ≫ 0, we can pick n = N .
Proof. By Proposition 6.10,
(6.12.1) V (m+2N )e ×V (m+N)e Sϕ0 m ,σN f∗ (M ) ⊇ V (m+2N )e ×V (m+N)e Sϕ0 m+N f∗ (M ) ⊇ Sϕ0 m+N ,σN f∗ (M ).
Furthermore, by Proposition 6.11.b, we know the statement holds when Sϕ0 m f∗ (M ) is replaced by
Sϕ0 m ,σN f∗ (M ). Hence the inclusion of the right end of (6.12.1) to the left end is an equality for m ≫ 0.
However, then all inclusions in (6.12.1) are equalities. In particular from the first equality of (6.12.1),
using that V (m+2N )e −
→ V (m+N )e is faithfully flat, we obtain that Sϕ0 m ,σN f∗ (M ) = Sϕm+N f∗ (M )
for m ≫ 0. Using Proposition 6.11.b once more concludes our proof of the main statement. The
final statement is similar.
We can now obtain our promised analog of Proposition 3.3.
e
Theorem 6.13. In the situation of Notation 6.1, if L ⊗ M p −1 is ample, there exists an integer
n0 and a dense open set U ⊆ V such that for all m ≥ n ≥ n0 we have that
V me ×V ne Sϕ0 n f∗ (M )
= Sϕ0 m f∗ (M )
.
U
U
Proof. By shrinking V and applying Proposition 3.3 we can reduce to the case of Notation 6.1* .
The result follows directly from Corollary 6.12.
Corollary 6.14. In the situation of Notation 6.1, there is a natural inclusion
Sϕ0 n f∗ (M ) ⊆ f∗ (σn (X/V, ϕ) ⊗R M ).
as subsheaves of V ne ×V f∗ (M ).
Further, in the situation of Notation 6.1* , if M = Ql ⊗ P where Q is f -ample, then there is
an integer l0 > 0, such that for every integer l ≥ l0 , n ≥ N and f -nef line bundle P , the above
inclusion is equality.
F -SINGULARITIES IN FAMILIES
45
Proof. Consider the surjection.
ξ:
L
pne −1
pe −1
1
pne
ϕn
/ / σn (X/V, ϕ)
Then Sϕ0 n f∗ (M ) = im f∗ (ξ ⊗R idM ), which is a subsheaf of f∗ (σn (X/V, ϕ) ⊗R M ).
We prove the addendum by induction on n. If n = N ,
Sϕ0 N f∗ (M ) = Sϕ0 0 ,σN f∗ (M ) = f∗ (σN ⊗R M ) .
{z
}
|
{z
} |
by Definition 6.9
by Corollary 6.12
Let us assume then that n > N . By flat base change and the assumption of Notation 6.1*
f∗ (σn (X/V ) ⊗R M ) = f∗ (σN (X/V ) ⊗R M ) ×V Ne V ne .
Hence, using the inductional hypothesis, it is enough to see that for every n ≥ N ,
Sϕ0 n f∗ (M ) = Sϕ0 n−1 f∗ (M ).
However, this follows from point (c) of Proposition 6.11 and Corollary 6.12
6.3. Base-change. We now prove base change for Sϕ0 n f∗ .
Notation 6.15. In the situation of Notation 6.1, choose
◦ T a regular integral excellent scheme with a dualizing complex.
◦ T −
→ V a morphism.
We indicate base-change via this morphism by T in subscript. Define the following sheaves on XT .
(a) B := (fT )−1 OT
(b) Q := OXT
1
→ X i.
Denote by (p1 ) pi the natural projection morphism X i ×V i T i −
e
Proposition 6.16. In the situation of Notation 6.15, if L ⊗ M p −1 is f -ample then for n ≫ 0 the
natural base change morphism
1
1
f∗ A pne ⊗A M ×V ne T ne −→ (fT )∗ B pne ⊗B MT
induces a surjective morphism on subsheaves
(6.16.1)
Sϕ0 n f∗ (M ) ×V ne T ne ։ Sϕ0 n (fT )∗ (MT ).
T
Furthermore,
(a) the lower bound on the n for which the above statement holds depends only on f and M . In
particular, it is independent of T .
(b) If M = Ql ⊗ P for some f -ample line bundle Q, nef line bundle P and integer l > 0, then
there is a uniform lower bound on n independent of l and P .
e
(c) Even if L ⊗ M p −1 is not assumed to be f -ample, (6.16.1) is an isomorphism if T −→ V is
flat.
(d) There is a dense open set U ⊆ V , such that if the image of T −→ V is contained in U , then
(6.16.1) is an isomorphism.
Proof of Proposition 6.16. First, note that
! ne 1
1
1 !
ne
∗ pne −1
∗ pne −1 pne
p −1
pne
p
1
1
ne
e
∼
⊗Q MT .
L pe −1 ⊗R M p
L pe −1
⊗R M ∼
(p1 ) pne
= (p1 ) pne
= LTp −1
46
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
Hence, there is a natural base-change morphism below in (6.16.2). We will show that it is an
isomorphism for n ≫ 0. Furthermore, this n can be chosen independently of T .
!
!!
1
1
ne ne
pne −1 pne
p
p −1
e −1
p
ne
e
(6.16.2)
f∗
L p −1
⊗Q MT
⊗R M
×V ne T −
→ (fT )∗
LT
Indeed by cohomology and base-change [Sta, Lemma 25.20.1] it is enough to show that for every
integer i > 0,
!
1
ne ne
p
p −1
i
e
R f∗
⊗R M = 0.
L p −1
Since V ne −
→ V is affine, this is equivalent to showing that for i > 0,
!
1
1 !
ne
ne
ne ne
p
p
p −1
p −1
ne
,
⊗R M = Ri (fV ne )∗
L pe −1 ⊗R M p
0 = Ri (fV ne )∗
L pe −1
which is further equivalent to showing that for i > 0,
ne
p −1
i
pne
e −1
p
(6.16.3)
0 = R f∗ L
⊗R M
.
However, the last vanishing holds for n ≫ 0 by relative Serre vanishing, independently of T . Hence
the base change homomorphism of (6.16.2) is isomorphism indeed for n ≫ 0, which can be chosen
independently of T . In particular, for n ≫ 0 there is a commutative base change diagram as follows,
which implies the statement of the proposition together with addendum (a).
(6.16.4)
f∗
L
pne −1
pe −1
1
pne
⊗R M
!
×V ne T ne
1
/ f∗ A pne ⊗A M
×V ne T ne
∼
=
(fT )∗
pne −1
pe −1
LT
1
pne
⊗Q MT
!
1
/ (fT ) B pne ⊗B MT
∗
→ V is flat, then both vertical morphisms in the above diagram
For addendum (c) note that if T −
e
are isomorphisms even if L ⊗ M p −1 is not assumed to be f -ample. Furthermore, images via flat
pullbacks are pullbacks of images, which concludes the proof of (c).
For addendum (b), note that if M = Ql ⊗ P for an f -ample line bundle Q and a nef line bundle
P , then (6.16.3) holds for every l > 0 uniformly, by relative Fujita vanishing [Kee03].
For addendum (d) note that by [Har77, Theorem 12.8 and Corollary 12.9] there is a dense open
set U ⊆ V (i.e., the open set where h0 (Xs , Ms ) is constant), such that if T maps into U then the
right vertical arrow in (6.16.4) is an isomorphism. In particular then the homomorphism (6.16.1)
is injective, because it is a homomorphism between subsheaves of the two sheaves involved in the
above vertical map. Therefore, by the already proven surjectivity (6.16.1) is an isomorphism. This
concludes addendum (d).
Remark 6.17. In particular if in Notation 6.15, T = s is a perfect point of V (that is, T = Spec (K)
for some perfect field K), then Sϕ0 n (fT )∗ (MT ) can be interpreted as follows. Since, the natural
i
T
→ XT is an isomorphism.
map K −
→ (K)1/p is an isomorphism, the natural projection XT i −
i
0
→ K 1/p . Thus
Furthermore (fT )∗ (MT i ) can be identified with (fT )∗ (MT ) = H (XT , MT ) via K −
F -SINGULARITIES IN FAMILIES
47
Sϕ0 n (fT )∗ (MT ) is identified with
T
(6.17.1)
pme −1
pe −1
0
Ss,m
:= im H 0 Xs , Ls
1
pme
⊗ Ms
!
!
−
→ H 0 (Xs , Ms ) ,
In particular, for m ≫ 0, (6.17.1) can be identified with S 0 (Xs , σ(Xs , ϕs ) ⊗ Ms ).
6.4. Uniform stabilization.
e
Theorem 6.18. In the situation of Notation 6.1, if L ⊗ M p −1 is f -ample, then there is an integer
n > 0, such that for all integers m ≥ n and perfect points s ∈ V ,
1
0
= S 0 (Xs , σ(Xs , ψs ) ⊗ Ms ).
im Sϕ0 m f∗ (M ) ⊗OV me k(s) pme −→ H 0 (Xs , Ms ) = Ss,m
0
(Note ψs is defined in Section 2.23 and Ss,m
in Remark 6.17.)
Proof. We show the statement by induction on the dimension of V . There is nothing to prove
if dim V = 0. Hence, we may proceed to the inductional step and assume that dim V > 0. By
Proposition 6.16, there is an n > 0 such that for every perfect point s ∈ V and every m ≥ n
1
0
im Sϕ0 m f∗ (M ) ⊗OV me k(s) pme −
→ H 0 (Xs , Ms ) = Ss,m
.
According to Lemma 2.17, by possibly increasing n, we may find a non-empty, dense open set U
such that for all m ≥ n,
σm (X/V, ϕ)|U = σn (X/V, ϕ) ×V ne V me |U .
Therefore, applying Corollary 6.12, yields (also by possibly increasing n) that for all m ≥ n,
V me ×V ne Sϕ0 n f∗ (M )
U
= Sϕ0 m f∗ (M )
U
.
It follows then that for every perfect point s ∈ U ,
1
→ H 0 (Xs , M )
im Sϕ0 m f∗ (M ) ⊗OV me k(s) pme −
0
is stabilized for all perfect s ∈ U for values of m ≥ n.
is the same for all m ≥ n. Therefore, Ss,m
0
Hence it is equal to S (Xs , σ(Xs , ψs ) ⊗ Ms ). This shows the statement of the proposition for all
perfect s ∈ U . We fix the n0 = n used above for future use.
`
On the other hand, consider the reduced scheme V1 = V \ U . Write V1 = V1,i as a disjoint
union of regular locally closed integral subschemes. Each V1,i has dimension smaller than V . In
particular, there exists an ni such that the statement holds for fi : XVi −
→ Vi . Letting n be the
maximum of n0 and the ni we obtain our result since every perfect point of X factors through a
point of U or of one of the Vi .
Corollary 6.19. In the situation of Notation 6.1, if L ⊗ M p
open set U ⊆ V , such that for every perfect point s ∈ U ,
e −1
is f -ample, then there is a dense
1
Sϕ0 m f∗ (M ) ⊗OV me k(s) pme = S 0 (Xs , σ(Xs , ψs ) ⊗ Ms ).
In particular, dimk(s) S 0 (Xs , σ(Xs , ψs ) ⊗ Ms ) is constant on an open set, and the rank of Sϕ0 m f∗ (m)
equals this dimension.
Proof. The main statement follows immediately from Theorem 6.18 and point (d) of Proposition 6.16.
The corollary follows from basic properties of coherent sheaves [Har77, Exercise II.5.8]
If there is a point s0 ∈ V , such that in Xs0 we have H 0 = S 0 , then we can say more than the
above corollary: it turns out that it can be assumed that s0 ∈ U , and further that U has other
useful properties. This is proved in the following theorem.
48
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
e
Theorem 6.20. In the situation of Notation 6.1, if L⊗M p −1 is f -ample and there is a perfect point
s0 ∈ V such that H 0 (Xs0 , Ms0 ) = S 0 (Xs0 , σ(Xs0 , ψs0 ) ⊗ Ms0 ), then there is an open neighborhood
U of s0 , such that
(a) f∗ M |U is locally free and compatible with base change and
(b) H 0 (Xv , Mv ) = S 0 (Xv , σ(Xv , ψv ) ⊗ Mv ) for every perfect point v ∈ U .
In particular, dim S 0 (Xv , σ(Xv , ψv ) ⊗ Mv ) is constant for v ∈ U .
Proof. By Theorem 6.18, there is an n, such that for every v ∈ V and integer m ≥ n, the natural
base-change homomorphism induces a surjection as follows
1
(6.20.1)
Sϕ0 m f∗ (M ) ⊗OV me k(v) pme
/ / S 0 (Xv , σ(Xv , ψv ) ⊗ Mv )
_
1
1
me
p
M
⊗
A
f∗
⊗OV me k(v) pme
A
/ H 0 (Xv , Mv ).
Furthermore, the right vertical arrow is surjective for v = s0 . It follows that so is the bottom
horizontal arrow then. Using [Har77, Corollary III.12.11] concludes (a).
To prove point (b), let us consider (6.20.1) for v = s0 again. Now we know that the bottom horizontal arrow is isomorphism. However, then the left vertical arrow is also surjective. By
Nakayama’s-lemma, there is an equality (Sϕ0 n f∗ (M ))v = (f∗ (M ))v of stalks, which means that by
possibly restricting U , for every v ∈ U , the left vertical arrow of (6.20.1) is surjective. However
then, since the bottom horizontal arrow is isomorphism for every v ∈ U , point (b) follows.
The surjection in Proposition 6.16 is not an isomorphism in general by the following two examples. In the first example the global geometry of the smooth fibers causes the anomaly, while in
the second, the degeneration of F -pure singularities to non F -pure singularities is the main culprit.
However, we first show a lemma.
Lemma 6.21. Let Y be a smooth curve over k and G an effective divisor on Y . Define BY1 as the
cokernel of OX −→ F∗ OX . Then
S 0 (Y, σ(X, 0) ⊗ ωY (G)) = H 0 (Y, ωY (G)) ⇔ H 0 (Y, BY1 (−pn G)) = 0
(∀n > 0).
Proof. Consider the following exact sequence.
0
/ OY
/ F∗ OY
/ B1
Y
/0
Twist this sequence by −G, apply cohomology and use that since G is effective, either
(a) G > 0 and hence H 0 (Y, (F∗ OY )(−G)) ∼
= H 0 (OY (−F ∗ G)) = 0, or
(b) G = 0 and then H 0 (Y, OY ) −
→ H 0 (Y, F∗ OY ) is an isomorphism.
In either case, we have another exact sequence:
0
/ H 0 (Y, B 1 (−G))
Y
/ H 1 (Y, OY (−G))
/ H 1 (Y, F∗ OY (−F ∗ G)) ,
where the last map is the Serre dual to H 0 (Y, (F∗ ωY )(G)) −
→ H 0 (Y, ωY (G)). Furthermore, observe
0
n
0
n+1
→ H (Y, (F∗ ωY )(G)) is harmlessly identified with the map
that the map H (Y, (F∗ ωY )(G)) −
H 0 (Y, (F∗ ωY )(pn G)) −
→ H 0 (Y, (ωY (pn G))). Therefore, combining the previous statements, we see
that H 0 (Y, (F∗n+1 ωY )(G)) −
→ H 0 (Y, (F∗n ωY )(G)) surjects if and only if H 0 (Y, BY1 (−pn G)) = 0.
Example 6.22. We choose two smooth projective curves C and D of genus 3 over an algebraically
closed field k of prime characteristic with two ample line bundles NC and ND (of degree 1), such
that S 0 (C, ωC ⊗NC ) 6= H 0 (C, ωC ⊗NC ), but S 0 (D, ωD ⊗ND ) = H 0 (D, ωD ⊗ND ). Since the relative
Picard scheme over the moduli stack of smooth curves of genus g is smooth and irreducible, there is
a (possibly reducible) curve connecting (C, NC ) and (D, ND ) in the above moduli space. However,
F -SINGULARITIES IN FAMILIES
49
then by possibly replacing (C, NC ) and (D, ND ) we may find also an irreducible curve connecting
them. Therefore, by passing to the normalization of this curve, we may assume that there is a family
f :X −
→ V , a line bundle N on X and two points c, d ∈ C, such that if we set (C, NC ) := (Xc , N |Xc )
and (C, NC ) := (Xc , N |Xc ), then
(a) V is a smooth curve,
(b) X is a family of smooth curves of genus 3,
(c) degX/V N = 1,
(d) S 0 (C, ωC ⊗ NC ) 6= H 0 (C, ωC ⊗ NC ) and
(e) S 0 (D, ωD ⊗ ND ) = H 0 (D, ωD ⊗ ND ).
1
1−p
p
−
→ OXV 1 the map with Dϕ = 0 (see Definition 2.8
and ϕ : ωX/V
Furthermore, fix e = 1, L := ωX/V
for the definition of Dϕ ). Note that by Lemma 2.24, Dϕs = 0 as well. Set M := ωX/V ⊗ N .
By assumption (c), for every v ∈ V and i > 0, H i (Xv , Nv ) = 0. Therefore, f∗ N is a vector
bundle of rank 3, and its formation is compatible with arbitrary base-change. In particular, by
Theorem 6.20 and assumption (e), Sϕn f∗ (ωX/V ⊗M ) ⊆ f∗ (ωX/V ⊗M ) is a subsheaf that is isomorphic
generically to f∗ (ωX/V ⊗ M ). Furthermore, by being a subsheaf of f∗ (ωX/V ⊗ M ) it is torsion free,
and by V being a curve, it is locally free of rank 3. Therefore,
dimk Sϕn f∗ (ωX/V ⊗ M ) ⊗ k(c) = 3
However,
dimk S 0 (C, ωC ⊗ MC ) < dimk H 0 (C, ωC ⊗ MC ) = 3
This shows that the base change morphism
→ S 0 (C, ωC ⊗ MC )
Sϕ0 n f∗ (ωX/V ⊗ M ) ⊗ k(c) −
cannot be isomorphism.
We are left to give the polarized curves (C, MC ) and (D, MD ). For that consider the situation
of Lemma 6.21. By [Tan72, Lemma 12], H 0 (Y, B 1 (−pn G)) = 0 if pn G ≥ n(Y ) where n(Y ) is a
numerical invariant of the curve, which is at most 1 for genus 3 curves [Tan72, Lemma 10]. Hence
H 0 (Y, B 1 (−pn G)) = 0 if deg G ≥ 1 and n > 1. In particular, in our special case (i.e., if Y is of
genus 3 and deg G ≥ 1) Lemma 6.21 states that
S 0 (Y, σ(Y, 0) ⊗ ωY (G)) = H 0 (Y, ωY (G)) ⇔ H 0 (Y, BY1 (−G)) = 0.
Finding a curve where H 0 (Y, BY1 (−G)) = 0 is quite easy, because again using [Tan72, Lemma 12]
yields that if n(Y ) = 1, then there is a degree one divisor G for which H 0 (Y, BY1 (−G)) 6= 0, and
then by passing to a linearly equivalent divisor we may also assume that G is effective. Further,
[Tan72, Example 1] gives a curve ( x3 y + y 3 z + z 3 x = 0) for which n(Y ) = 1. Therefore we have
found C and NC .
To find D and ND assume further that k = F3 . Then by [Mil72, Kob75] a general genus
three curve curve is ordinary, or equivalently S 0 (Y, σ(Y, 0) ⊗ ωY ) = H 0 (Y, ωY ). Therefore then
H 0 (Y, BY1 ) = 0 for such a Y . Take now an arbitrary effective degree 1 divisor G. Then BY1 (−G)
embeds into BY1 and hence H 0 (Y, BY1 (−G)) = 0 as well. In particular, we can choose D to be a
generic curve and ND to be an arbitrary degree one line bundle on D.
Example 6.23. In the following example for an f -ample line bundle Q the surjection
1
l
0
ne
p
k(s)
։ S 0 (Xs , σ(Xs , ϕs ) ⊗ Qls ).
f
(Q
)
⊗
(6.23.1)
S ϕn ∗
OV ne
is not an isomorphism for any integer n ≥ n0 and l > 0. Therefore, the isomorphism in Proposition 6.16
cannot be obtained by stronger positivity assumptions.
Let C be the projective cone over a supersingular elliptic curve, and let D be a non-singular
cubic surface. Then these can be put into a family f : X −
→ V as above. More precisely, we may
find a family f : X −
→ V , and a point c ∈ V , such that C := Xc , and hence the following hold
50
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
(a)
(b)
(c)
(d)
V is a smooth curve,
X is a a flat family of normal surfaces,
Xc is not sharply F -pure at on point P ∈ Xc and
Xs is sharply F -pure for every s ∈ V \ {c}.
1−p
and ϕ := ϕ0 as in the previous example. Then by Corollary 3.10 and Nakayama
Let L := ωX/V
lemma, for every n ≫ 0, σn (X/V, ϕ)|V \{c} ∼
= Of −1 (V \{c}) . Choose now an arbitrary sufficiently
f -ample line bundle M . By Corollary 6.14, for every n ≫ 0, Sϕ0 n f∗ (M l )|V \{c} ∼
= (f∗ (M ))V ne \{c} for
l
l
0
every integer l > 0. In particular the rk Sϕn f∗ (M ) = rk f∗ (M ) for every n ≫ 0 and l > 0 (where
n does not depend on l). Consequently,
(6.23.2)
dim Sϕ0 n f∗ (M l ) ⊗ k(c) ≥ rk f∗ (M l ) =
dimk H 0 (Xc , Mcl )
{z
}
|
M is relatively ample enough
On the other hand for all n ≫ 0 (independent of l)
(6.23.3)
Sϕ0 n f∗ (M l ) ⊗ k(c) ։
Sϕ0 nc (fc )∗ (Mcl )
=
{z
}
|
by point (b) of Proposition 6.16
H 0 (Xc , σ(Xc , 0) ⊗ Mcl )
|
{z
}
sine Mc is ample enough ([Pat12]) and Lemma 2.24
However, by the relative ampleness assumption, Mc is globally generated, hence dimk H 0 (Xc , σ(Xc , 0)⊗
Mcl ) < dimk H 0 (Xc , Mcl ) (because σ(Xc , 0) ( OXc ). Therefore, (6.23.2) and (6.23.3) implies that
indeed the surjection (6.23.1) is not an isomorphism for any l > 0 and any n ≫ 0 (independently
bounded of l).
Remark 6.24. The fundamental reason for the above example is that σn (X/V, ϕ) is NOT flat in
general.
The above two examples show that dimk (S 0 (Xv , σ(Xv , ϕv ) ⊗ Mv ) is not upper semicontinuous.
One might guess then it is lower semicontinuous. The next example shows that that is also not the
case (this can also be deduced from [Har98, Example 5.5] and [Tan13, Theorem 8.3] as pointed out
to us by Tanaka). So, dimk S 0 (Xv , σ(Xv , ϕv ) ⊗ Mv ) is not semicontinuous in either direction.
Example 6.25. Take a flat family f : X −
→ V of ordinary elliptic curves, and a line bundle M on
∼
X of relative degree 0, such that MXv0 = OXv0 for a special v0 ∈ V , and MXv ∼
6 OXv for generic
=
v ∈ V . Then, by the ordinarity of the fibers, for all v ∈ V ,
H 0 (Xv , Mv ) = S 0 (Xv , σ(Xv , 0) ⊗ Mv ).
However, dim H 0 (Xv , Mv ) = 1 for the special fiber and 0 for the generic one. So, dim S 0 (Xv , σ(Xv , 0)⊗
Mv ) is not lower semicontinuous in this example. One can also easily modify this example by taking
an M with higher relative degree to obtain higher values of dimension for the generic fiber.
6.5. Global generation and semi-positivity. Suppose now V is a projective variety over a
e
prefect field k. In this section, we explore global generation results if L ⊗ M p −1 is ample (instead
0
of just relatively ample). In particular, Sϕn f∗ (M ) is globally generated for all large n. This should
not be surprising since Sϕ0 n f∗ (M ) lives on V ne where ampleness is amplified.
Proposition 6.26. In the situation of Notation 6.1, if V is projective over a perfect field k and
e
L ⊗ M p −1 is ample, then Sϕ0 n f∗ (M ) is globally generated for every n ≫ 0.
Proof. Choose a globally generated ample divisor H on V and let d := dim V . Since Sϕ0 n f∗ (M ) is
defined as the image of
!
1
ne ne
p
p −1
⊗R M ,
f∗
L pe −1
F -SINGULARITIES IN FAMILIES
51
it is enough to show that this sheaf is globally generated as an OV ne -module. By Mumford’s criterion
[Laz04a, Theorem 1.8.5], it is enough to show that for every i > 0 and for the divisor Hne on V ne
identified with H via the isomorphism V ne ∼
=V,
!!
1
ne ne
p
p −1
ne
i
e
⊗R M
= 0.
(6.26.1)
H V , OV ne (−iHne ) ⊗OV ne f∗
L p −1
By relative Serre vanishing for i > 0,
i
R f∗
L
pne −1
pe −1
1
pne
⊗R M
!
= 0.
In particular, then to prove (6.26.1) it is enough to show
pne −1
i
pne
∗
e −1
p
H X, L
⊗ M (−if H) = 0.
However this is equivalent to showing that,
!
pne −1
∗
(pe −1) pe −1
i
⊗ M (−if H) = 0,
H X, L ⊗ M
which holds by Serre-vanishing for n ≫ 0 and i > 0.
We recall the following definition.
Definition 6.27 (Definition 2.11 of [Vie95]). Let F be a sheaf on a normal, quasi-projective (over
a perfect field k) variety V and U ⊆ V a dense open set. Let Ulf be the open locus of V where
F ′ := F /(torsion) is locally free. Then F is weakly positive over U , if for a fixed (or equivalently
every: [Vie95, Lemma 2.14.a]) ample line bundle H for every a > 0 there is a b > 0 such that
S habi (F )⊗H b is globally generated over U ∩Ulf (here S habi (F ) denotes the abth symmetric reflexive
power of F ). We say that F is weakly-positive if it is weakly-positive over some dense open set.
Lemma 6.28. If g : Y −→ Z is a finite morphism of normal varieties, quasi-projective over k,
U ⊆ Z a dense open set and F a sheaf on Z, then F is weakly positive over U if and only if g ∗ F
is weakly positive over g −1 (U ).
Proof. The only if direction is shown in [Vie95, Lemma 2.15.1]. For the other direction, according
to Definition 6.27 by throwing out codimension two subset we may assume that F is locally free
and Y is flat over Z. In particular then g ∗ F is also locally free. Choose a very ample divisor H
on Z, such that
(a) H om(g∗ OY , H ) is globally generated and
(b) g∗ OY ⊗ H is globally generated.
Fix then an a > 0. By the weak positivity of g∗ F , there is a b > 0, such that there is a homo→ S habi (g∗ F ) ⊗ g∗ H b surjective over g∗ U . Consider then for every choice of
morphism α : OY⊕N −
s ∈ Hom(g∗ OY , H ) and s′ ∈ H 0 (H b−1 ) the following composition.
γ:=g∗ α
/ g (S habi (g ∗ F ) ⊗ g ∗ H b ) ∼
(g∗ OY )⊕N ❳❳
= S habi (F ) ⊗ H b ⊗ g∗ OY
❳❳❳❳❳ ∗
❳❳❳
❳❳❳❳❳
❳❳❳❳❳
❳❳❳❳+
βs
δs,s′ :=idS habi (F )⊗H b ⊗s⊗s′
S habi (F ) ⊗ H 2b
Choose a point P ∈ U and an element f of the fiber S habi (F ) ⊗ k(P ) over P . Then by assumption
(b) for any preimage Q of P there is a section t′ ∈ OY⊕N , such that α(t′ )Q = g ∗ f × ”generator“.
The section t′ descends to a section t := g∗ (t′ ) ∈ (g∗ OY )⊕N , such that γ(t)P = f × ”generator“ × h
52
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
for some h ∈ (g∗ OY )P . However then by (a) and the very ampleness of H for a suitable choice of
s and s′ , δs,s′ (γ(t)))P = f × “generator” is not zero at P . Therefore, for every point P in U and
every element f in the fiber of S habi (F ) ⊗ H 2b at that point we find a section of S habi (F ) ⊗ H 2b
whose image in the fiber is f up to multiplication by unit. This finishes our proof.
Corollary 6.29. In the situation of Notation 6.1, let V be projective over a prefect field k, L ⊗
e
M p −1 ample and n > 0 be an integer such that there is an open set U ⊆ V for which
Sϕ0 n f∗ (M ) ×U ne U me = Sϕ0 m f∗ (M )|U me ,
for every integer m ≥ n. Then Sϕ0 n f∗ (M ) is weakly-positive. Note that an integer as above always
exists by Corollary 6.12 and Proposition 5.3.
Proof. By the assumption, there is an embedding
Sϕ0 m f∗ (M ) ֒→ Sϕ0 n f∗ (M ) ×V ne V me ,
which is generically isomorphism. Since the left sheaf is globally generated and hence weaklypositive, so is the right one. However then by Lemma 6.28, so is Sϕ0 n f∗ (M ).
Lemma 6.30. If F is a coherent sheaf on a normal variety V over a perfect field k, then F is
weakly-positive if and only if for every ample line bundle H and every integer p ∤ r > 0 there is a
finite morphism τ : T −→ V , such that τ ∗ H ∼
= (H ′ )r for some line bundle H ′ and τ ∗ F ⊗ H ′ is
weakly-positive.
Proof. The proof is identical to the (b) ⇒ (a) part of [Vie95, Lemma 2.15.1].
Theorem 6.31. In the situation of Notation 6.1, if V is projective and L ⊗ M p
f -ample, then Sϕ0 n f∗ (M ) is weakly positive for n ≫ 0.
e −1
is a nef and
Proof. Choose an integer n > 0, as in Corollary 6.29. Fix an ample line bundle H on V , an integer
p ∤ r > 0 and a finite morphism τ : T −
→ V , such that τ ∗ H ∼
= (H ′ )r for some line bundle H ′ .
Such a morphism exits by [Vie95, Lemma 2.1]. By Lemma 6.30, we are supposed to prove that
1
∗
1
τ pne Sϕn f∗ (M ) ⊗OV ne (H ′ ) pne is weakly positive. By disregarding codimension two closed sets,
we may assume that T is regular as well. Then by point (c) of Proposition 6.16, we see that n
satisfies also the assumptions of Corollary 6.29 but for f and M replaced by fT and MT ⊗ fT∗ HT′ ,
respectively. In particular, since MT ⊗ fT∗ HT′ is ample,
Sϕ0 nT (fT )∗ (MT ⊗ fT∗ HT′ ) ∼
= HT′ ⊗ Sϕ0 nT (fT )∗ (MT )
is weakly-positive over T ne .
6.6. Relation to global canonical systems. There has been another subsheaf S 0 f∗ (σ(X, ∆) ⊗
M ) of f∗ (M ) introduced in [HX13, Definition 2.14] with a definition similar to that of Sϕ0 n f∗ (M ). In
this section we show some of the similarities and differences between the two sheaves. The advantage
of S 0 f∗ (σ(X, ∆) ⊗ M ) over Sϕ0 n f∗ (M ) is that it lives on one V , there is no involvement of V ne at
all. On the other hand we show that contrary to Sϕ0 n f∗ (M ) it does not restrict even generically to
S 0 (Xs , σ(Xs , ∆s ) ⊗ Ms ).
Throughout the section we use the divisorial language since the presentation seems to be more
straightforward this way, hence ∆ satisfies the conditions of Remark 2.11. Recall that we defined
the notion of pair in our setting in Definition 2.12. Whenever we say pair, we mean everything
assumed there. In particular, all the assumptions on f : X −
→ V from Notation 2.1: flatness,
equidimensionality, etc.
First, we recall the definition of S 0 f∗ (M ).
F -SINGULARITIES IN FAMILIES
53
Definition 6.32. [HX13, Definition 2.14] Let ∆ be an effective Q-divisor, such that (X, ∆) is a
pair and assume that f : X −
→ V is projective. Let e be the smallest9 positive integer such that
(pe − 1)(KX + ∆) is Cartier. Then given a line bundle M on X, we define
\
S 0 f∗ (σ(X, ∆) ⊗ M ) :=
im (f∗ (F∗ne OX ((1 − pne )(KX + ∆)) ⊗ M ) −
→ f∗ (M )) .
n>0
S 0f
We say that
∗ (σ(X, ∆) ⊗ M ) stabilizes if the above intersection stabilizes. Note that in that
case S 0 f∗ (σ(X, ∆) ⊗ M ) is a coherent sheaf.
Recall that S 0 f∗ (σ(X, ∆)⊗M ) stabilizes if M −KX −∆ is relatively ample, see [HX13, Proposition
2.15] for a proof.
Proposition 6.33. Let ∆ be an effective Q-divisor, such that (X, ∆) is a pair and assume that
f : X −→ V is projective. Let M be a line bundle on X such that S 0 f∗ (σ(X, ∆) ⊗ M ) stabilizes.
Further assume that V is regular. Then for every n ≫ 0
0
S∆,ne
f∗ (M ) ⊆ (S 0 f∗ (σ(X, ∆) ⊗ M ))V ne
0
as subsheaves of (f∗ (M ))V ne , where S∆,ne
f∗ (M ) is defined in Definition 6.4.
Proof. The statement is local over the base hence we can assume that V is affine, KV ∼ 0, and
that H omOV (F∗i OV , OV ) ∼
= F∗i OV is a free OV -module for all i ≥ 0. Note then also we can identify
KX/V with KX . We introduce the notation
L := OX ((1 − pne )(KX + ∆)).
→ OV which we identify with
We have the evaluation-at-1 map F∗ne OV ∼
= H omOV (F∗ne OV , OV ) −
the trace of V based on our previous assumption ωV ∼
= OV . This map pulls back to X to provide
∗ ne
∼
us with a map which we also denote as Trne
→ OX ⊗OX f ∗ OV ∼
= OX .
FV : OXV ne = OX ⊗ f F∗ OV −
Then there is a diagram as follows, which is commutative up to multiplication by a unit cf. [Sch09,
Lemma 3.9].
F∗ne L
TrFX
/ F ne OX ((1 − pne )KX )
∗
/ OX
O
Trne
F
V
F∗ne OX ((1
−
pne )(KX/V
+ ∆))
/ F ne OX ((1 − pne )KX/V )
∗
TrF ne
/ OX ⊗O f ∗ F ne OV
∗
X
X/V
Applying the functor f∗ ( ⊗OX M ) to the above diagram, we obtain the following.
f∗ ((F∗ne L) ⊗OX M )
/ f∗ (M )
O
κ
f∗ ((F∗ne OX ((1
−
pne )(KX/V
+ ∆))) ⊗OX M )
/ f∗ (M ⊗O f ∗ F ne OV )
∗
X
Note that for every n ≫ 0 the image of the top horizontal row is S 0 f∗ (σ(X, ∆) ⊗ M ) and for every
0 f (M ). Hence for every n ≫ 0,
n > 0 the image of the bottom horizontal row is S∆,n
∗
0
0
S f∗ (σ(X, ∆) ⊗ M ) = im S∆,n f∗ (M ) −
→ f∗ (M ) ,
where the above map is induced by TrFV .
Since we want containment for subsheaves of (f∗ (M ))V ne , we localize at a point of V . By setting
ne (S 0 f (M )), we are in the following situation: if
B := Γ(V, OV ), P := f∗ (M ) and Q := FV,∗
∆,n ∗
ne
ne
Q is a B 1/p submodule of B 1/p ⊗B P for some B module P , we would like to show that
ne
ne
Q ⊆ ϑ(Q) ⊗B B 1/p where ϑ ∈ HomB (B 1/p , B) is the generator. This is simply Lemma 3.21.
9The intersection below is easily seen to be descending, hence if one chooses another e, the same object results.
54
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
Example 6.34. We provide an example where the two sheaves of Proposition 6.33 have different
ranks for every n ≫ 0. Fix an algebraically closed field k of characteristic p > 0. Let X := P1 × A1 ,
1
U = A1 × A1 = Spec k[y, x] ⊆ X and V := A1 := Spec k[x]. Define the divisor ∆ := p−1
V (y p − x)
which is a priori a divisor in U , but it happens to have support which is a closed subvariety of X
as well. Note that V (y p − x) is a subvariety of X isomorphic to A1 . In particular, (X, ∆) is sharply
F -pure. Choose now a line bundle N on X which is relatively ample over V and let M := N l for
some l ≫ 0. Then by [HX13, Lemma 2.19] S 0 f∗ (σ(X, ∆) ⊗ M ) = f∗ (M ). Therefore,
(S 0 f∗ (σ(X, ∆) ⊗ M ))V ne = (f∗ (M ))V ne
(6.34.1)
0 f (M ) ⊆ f (σ (X/V, ∆) ⊗ M ). By Corollary 3.10 and
On the other hand by Proposition 6.5, S∆,n
∗
∗ n
Lemma 2.24, for every n ≫ 0 and s ∈ V , σn (X/V, ∆)|Xs = σ(Xs , ∆). However, ∆|Xs is one point
p
, and hence not sharply F -pure. In particular, σ(Xs , ∆) 6= OXs for every
with multiplicity p−1
s ∈ S. Hence for every n ≫ 0, σn (X/V, ∆)|Xs ( OXs . It follows by the relative ampleness of M
0 f (M ) has also smaller rank
that f∗ (σn (X/V, ∆) ⊗ M ) has smaller rank than f∗ (M ). Then S∆,n
∗
than (S 0 f∗ (σ(X, ∆) ⊗ M ))V ne by (6.34.1).
The above example has an important corollary, which follows immediately from (d) of Proposition 6.16
and Theorem 6.18.
Corollary 6.35. In general S 0 f∗ (σ(X, ∆)⊗M )⊗k(s) is not isomorphic to S 0 (Xs , σ(Xs , ∆s )⊗Ms ).
In fact there are examples when the former has strictly bigger dimension than the latter for every
closed point s ∈ V .
Compare the following proposition with Proposition 3.24.
Proposition 6.36. Let ∆ be an effective Q-divisor, such that (X, ∆) is a pair and assume that
f : X −→ V is projective. Let M be a line bundle on X and assume that V is regular. Then for
every n ≥ m ≥ 0,
S 0 fV ne ,∗ (σ(XV ne , ∆V ne ) ⊗ MV ne ) ⊆ (S 0 fV me ,∗ (σ(XV me , ∆V me ) ⊗ MV me ))V ne
as subsheaves of (f∗ (M ))V ne . Furthermore, if the above two sheaves stabilize then
(n−m)e 0
S fV ne ,∗ (σ(XV ne , ∆V ne ) ⊗ f ∗ OV ((1 − p(n−m)e )KV ne )) ⊗ MV ne )
κ FV,∗
= S 0 fV me ,∗ (σ(XV me , ∆V me ) ⊗ MV me )
(n−m)e
where κ is induced by TrV
.
Proof. Note that by replacing f : X −
→ V by fV me : XV me −
→ V me , we may assume that m = 0,
and then by replacing e by (n − m)e that n = 1. As in the previous proof, let us assume that
V is affine, KV ∼ 0, and that H omOV (F∗e OV , OV ) ∼
= F∗e OV is a free OV -module. Consider the
following commutative diagram.
ν:=(FVe )X e
(XV e )e ❲❲❲
❲❲❲❲❲
/ X e ❱❱
❱❱❱❱
❱❱❱❱FXe
❲❲❲❲❲
❱❱❱❱
❲❲❲❲❲
❱❱❱❱
e
FX
❱❱❱❱
❲
❲
❲+
Ve
/+ X
e
XV
e
fV e
V
η:=(FV )X
e
FVe
f
/V
We fix the notations
L∆′ := OXV e ((1 − pe )(KXV e + η ∗ ∆)),
L∆ := OX ((1 − pe )(KX + ∆)).
F -SINGULARITIES IN FAMILIES
55
Then we have
(1 − pe )KXV e = (1 − pe )(η ∗ KX/V + fV∗ e KV e ), and (1 − pe )KX = (1 − pe )(KX/V + f ∗ KV ).
Therefore,
◦ applying pullback of TrFVe yields a homomorphism ν∗ L∆′ −
→ L∆ , which then induces a
e ν L′ −
e
homomorphism η∗ FXe V e ,∗ L∆′ ∼
= FX,∗
∗ ∆ → FX,∗ L∆ ,
◦ applying TrFXe e yields a homomorphism FXe V e ,∗ L∆′ −
→ OXV e ,
V
e L −
◦ applying TrFXe yields a homomorphism FX,∗
→
O
∆
X,
e O −
e
◦ by the assumption KV ∼ 0, TrFV corresponds to a homomorphism FV,∗
V → OV generating
e
HomOV (FV,∗ OV , OV ) and
◦ the previous homomorphism also induces a pullback homomorphism η∗ OXV e −
→ OX .
Furthermore, the above homomorphisms fit into the following diagram
η∗ FXe V e ,∗ L∆′
/ F e L∆
X,∗
❑❑
PPP
❑❑
PPP
❑❑
PPP
❑%
(
/ OX
η∗ OX e
V
The diagram is a composition of trace maps (restricted to smaller domains determined by ∆), hence
it is commutative. Applying now f∗ ( ⊗ M ) to the above diagram and using the projection formula
we obtain the following commutative diagram
f∗ η∗ (FXe V e ,∗ L∆′ ⊗ MV e )
/ f∗ (F e L∆ ⊗ M )
X,∗
❘❘❘
❱❱❱❱
❘❘❘
❱❱❱❱
❘❘❘
❱❱❱❱
❘(
❱❱*
/ f∗ (M ).
f∗ η∗ MV e
e f e yields another commutative diagram
Observing that f∗ η∗ = FV,∗
V ,∗
/ f∗ (F e L∆ ⊗ M )
X,∗
❱❱❱❱
◗◗◗
❱❱❱❱❱
◗◗◗
◗◗◗
❱❱❱❱
❱❱+
◗(
e
/ f∗ (M ).
F fV e ,∗ MV e
e f e (F e
′
FV,∗
V ,∗ XV e ,∗ L∆ ⊗ MV e )
V,∗
Note now that the top horizontal arrow is split, because it is induced from ν∗ L∆′ −
→ L∆ which is
split as well. Therefore, if we define P := f∗ (M ) and
e
S := im f∗ (FX,∗
L∆ ⊗ M ) −
→ f∗ (M )
Q := im fV e ,∗ (FXe V e ,∗ L∆′ ⊗ MV e ) −
→ fV e ,∗ MV e ,
e
then we have Q ⊆ P ⊗A A1/p , S ⊆ P and (idP ⊗ TrFVe )(Q) = S. Therefore, by Lemma 3.21,
e
Q ⊆ S ⊗A A1/p . Since this holds for any n, the statement of the proposition follows.
Proposition 6.37. Let ∆ be an effective Q-divisor, such that (X, ∆) is a pair and assume that
f : X −→ V is projective and V is regular. Let M be a line bundle on X. Then for every n ≫ 0,
0
S 0 fV ne ,∗ (σ(XV ne , ∆V ne ) ⊗ MV ne ) ⊆ S∆,ne
f∗ (M )
as subsheaves of (f∗ (M ))V ne .
Proof. As before, let us assume that V is affine, KV ∼ 0, and that H omOV (F∗i OV , OV ) ∼
= F∗i OV is
a free OV -module for all i ≥ 0. By Lemma 3.13 and Lemma 3.16 there is a commutative diagram
/ F nene ne OX (1 − pne ) K
FXneV ne ,∗ OXV ne ((1 − pne ) (KXV ne + ∆V ne ))
X/V + ∆
X /V ,∗
❨❨❨❨❨❨
❨❨❨❨❨❨
❨❨❨❨❨❨
❨❨❨❨❨❨
❨❨❨❨❨❨
,
OXV ne .
56
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
Furthermore, the top horizontal arrow in the above diagram is split surjective. Therefore, after
applying fV ne ,∗ ( ⊗ MV ne ) the image of the vertical map still agrees with the image of the diagonal
0 f (M ), while the latter contains S 0 f ne (σ(X, (∆) ne )⊗ M ), since
map. The former is exactly S∆,n
∗
V ,∗
V
it is one of the terms in the intersection defining S 0 fV ne ,∗ (σ(X, (∆)V ne ) ⊗ M ).
Remark 6.38. Assuming that f : X −
→ V is projective and M −KX −∆ is ample, it would be natural
0
to ask whether Im S fV ne ,∗ (σ(XV ne ,∆V ne ⊗MV ne )⊗MV ne ) −
→ fV ne ,∗ Ms equals S 0 (Xs , σ(Xs , ∆s )⊗
Ms ) for all perfect points s ∈ V , in analogy with Theorem 3.23. We suspect this is true but will
not try to prove it here.
Corollary 6.39. Let f : (X, ∆) −→ V be a projective morphism from a pair with V regular
and projective over a perfect field k. Further, suppose that M is a line bundle on X such that
M − KX/V − ∆ is nef and f -ample (here M denotes a Cartier divisor corresponding to M ) and that
rk S 0 f∗ (σ(X, ∆)⊗M ) equals the general value of H 0 (Xs , σ(Xs , ∆s )⊗Ms ). Then S 0 f∗ (σ(X, ∆)⊗M )
is weakly positive. In particular, if V is a smooth curve then it is a nef vector bundle.
Proof. From the assumption rk S 0 f∗ (σ(X, ∆) ⊗ M ) equals the general value of H 0 (Xs , σ(Xs , ∆s ) ⊗
Ms ) follows that the inclusion of Proposition 6.33 is generically an isomorphism for every n ≫ 0.
However, since Sϕ0 n f∗ (M ) is weakly-positive for every n ≫ 0 by Theorem 6.31, we obtain by the
above generically isomorphic inclusion that (S 0 f∗ (σ(X, ∆) ⊗ M ))V ne is also weakly positive. Then
weak-positivity of S 0 f∗ (σ(X, ∆) ⊗ M ) follows from Lemma 6.28.
Appendix A. Relative Serre’s condition
Here we collect the statements of [HK04] and other sources, that are important for the current
paper for ease of reference. In some cases, we also state them in the greater generality that we need.
All schemes are Noetherian, excellent and possess dualizing complexes and all maps are separable.
Definition A.1. Let r > 0 be an integer. A coherent sheaf E on a Noetherian scheme X is Sr , if
for every x ∈ X,
depthOX,x Ex ≥ min{r, dimOX,x Ex }.
The sheaf E is said to have full support, if Supp E = X. It is reflexive if the natural map E −
→
E ∗∗ := H omOX (H omOX (E , OX ), OX ) is an isomorphism.
Definition A.2. If f : X −
→ V is a morphism of Noetherian schemes and E is a coherent sheaf on
X flat over V , then E is Sr over V if E |Xv is Sr for every v ∈ V . That is, for every x ∈ X,
o
n
depthOX (E |Xf (x) )x ≥ min r, dimOX (E |Xf (x) )x
f (x)
f (x)
We now recall two results from EGA about how depth and dimension behave in families.
Proposition A.3. [Gro67, Proposition 6.3.1] Let ϕ : A −→ B be a local homomorphism of Noetherian local rings, k the residue field of A and M and N are finite A and B modules respectively.
If N is a flat non-zero A-module, then
depthB (M ⊗A N ) = depthA (M ) + depthB⊗A k (N ⊗A k).
Proposition A.4. [Gro67, Corollaire 6.1.2] Let ϕ : A −→ B be a local homomorphism of Noetherian
local rings, k the residue field of A and M and N are non-zero finite A and B modules respectively.
If N is a flat A-module, then
dimB (M ⊗A N ) = dimA (M ) + dimB⊗A k (N ⊗A k).
From these we obtain the following corollary.
F -SINGULARITIES IN FAMILIES
57
Corollary A.5. Let f : X −→ V be a morphism of Noetherian schemes and E a coherent sheaf on
X flat over V , then E is Sr over V if and only if for every x ∈ X,
min{i|Hxi (E ) 6= 0} ≥ depthOV,v OV,v + min{r, dim Ex − dim OV,v }
= depthOV,v OV,v + min {r, dim (E |Xv )x }
where v := f (x).
Proof. To say that E |Xv has depth equal to t at x is equivalent to asserting that depthx E =
(depthOV,v OV,v ) + t by Proposition A.3. The result follows.
The following vanishing allows us to extend sections over sets of relative codimension 2.
Proposition A.6. [HK04, Proposition 3.3] Let f : X −→ V be a morphism of Noetherian schemes
and E a coherent sheaf on X flat and Sr over V . Let Z ⊆ X be a closed subscheme of X such that
we have codimSupp Ev (Supp Ev ∩ Zv ) ≥ r for every v ∈ V . Then for each 0 ≤ i < r, HZi (E ) = 0.
Proof. We follow the proof in [HK04, Proposition 3.3]. First, we certainly have a spectral sequence
E2p,q := Hxp ◦ HZq ⇒ Hxp+q . We then induct on r, the base case of r = 0 being trivial. Now assume
that HZ0 (E ) = . . . = HZr−2 (E ) = 0 but HZr−1 (E ) 6= 0. Then for any x ∈ Z which is a generic
point10 of the support of HZr−1 (E ), we have that Hx0 (HZr−1 (E )) 6= 0 (note here that x is probably
not a closed point). Additionally, since x is a point of Z ∩ Supp(E ), we have that dim(E |Xv )x ≥ r.
On the other hand, for all i > 0 we have Hxi (HZr−1−i (E )) = 0, and so by the spectral sequence
Hxr−1 (E ) ∼
= Hx0 (HZr−1 (E )) 6= 0. This contradicts Corollary A.5.
We now obtain a relative version of Hartog’s phenomena, just as in [HK04].
Proposition A.7. [HK04, Proposition 3.5, 3.6.1] Let f : X −→ V be a morphism of Noetherian
schemes and that E a coherent sheaf on X which satisfies one of the following two conditions:
(i) E is reflexive and f is flat and relatively S2 , or
(ii) E is of full support, flat and S2 over V .
Let j : U ֒→ X be an open set such that for Z := X \ U , codimXv Zv ≥ 2 for every v ∈ V . Then
the following natural map is an isomorphism:
E −→ j∗ E |U .
Proof. First, assume that E is reflexive. Consider a presentation of E ∗ by locally free sheaves
E2
/ E∗
/ E1
/ 0 and the dual 0
/E
/ E∗
1
/ E∗
2
Then by applying Proposition A.6 to OX and the applying local cohomology to the above exact
sequence yields that
HZ1 (E ) = HZ0 (E ) = 0.
(A.7.1)
If instead, E is flat, S2 and has full support, then we still have the vanishing (A.7.1) by Proposition A.6.
Consider then the exact sequence
0
/ H 0 (E )
Z
/E
Applying (A.7.1) concludes our proof.
/ j∗ E
/ H 1 (E ).
Z
Corollary A.8. [HK04, Proposition 3.6.2] Let f : X −→ V be a flat, S2 morphism of Noetherian
schemes, E a coherent X satisfying either Proposition A.7(i) or (ii). Let F be another coherent
sheaf satisfying Proposition A.7(i) or (ii). Let j : U ֒→ X be an open set, such that for Z := X \ U ,
codimXv Zv ≥ 2 for every v ∈ V . Assume also that E |U ∼
= F |U . Then E ∼
= F.
10a minimal prime in the language of commutative algebra
58
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
Proof. In any case, j∗ E |U = E and j∗ F |U = F and so the result follows.
Proposition A.9. [HK04, Corollary 3.7] Let f : X −→ V be a flat, S2 morphism of Noetherian
schemes and j : U ֒→ X an open set, such that for Z := X \ U , codimXv Zv ≥ 2 for every v ∈ V .
Assume also that E is a reflexive, coherent sheaf on U . Then j∗ E is reflexive.
Proof. Choose a coherent subsheaf F ⊆ j∗ E such that F |U = E [Har77, Ex II.5.15]. Then j∗ E is
reflexive by the following computation.
F ∗∗ ∼
j∗ E
= j∗ (F ∗∗ |U ) ∼
=
= j∗ E ∗∗ ∼
|{z}
| {z }
| {z }
by Proposition A.7
E∼
=F |U
This completes the proof.
E is reflexive
Proposition A.10. If f : X −→ Y is a projective, flat, relatively S2 and G1, equidimensional
morphism, then ωX/Y is reflexive.
Proof. The proof given in [PS12, Lemma 4.8] works.
References
M. André: Homomorphismes réguliers en caractéristique p, C. R. Acad. Sci. Paris Sér. I Math. 316
(1993), no. 7, 643–646. 1214408 (94e:13026)
[Bli13]
M. Blickle: Test ideals via algebras of p−e -liner maps, J. Algebraic Geom. 22 (2013), no. 1, 49–83.
[BB11]
M. Blickle and G. Böckle: Cartier modules: finiteness results, J. Reine Angew. Math. 661 (2011),
85–123. 2863904
[BMS08] M. Blickle, M. Mustaţǎ, and K. E. Smith: Discreteness and rationality of F -thresholds, Michigan
Math. J. 57 (2008), 43–61, Special volume in honor of Melvin Hochster. 2492440 (2010c:13003)
[BS13]
M. Blickle and K. Schwede: p−1 -linear maps in algebra and geometry, Commutative Algebra: Expository Papers Dedicated to David Eisenbud on the Occasion of His 65th Birthday (I. Peeva, ed.), Springer,
Berlin, 2013, pp. 123–205.
[BSTZ10] M. Blickle, K. Schwede, S. Takagi, and W. Zhang: Discreteness and rationality of F -jumping
numbers on singular varieties, Math. Ann. 347 (2010), no. 4, 917–949. 2658149
[BST11]
M. Blickle, K. Schwede, and K. Tucker: F -singularities via alterations, arXiv:1107.3807.
[BK05]
M. Brion and S. Kumar: Frobenius splitting methods in geometry and representation theory, Progress
in Mathematics, vol. 231, Birkhäuser Boston Inc., Boston, MA, 2005. MR2107324 (2005k:14104)
[CHMS12] P. Cascini, C. Hacon, M. Mustaţă, and K. Schwede: On the numerical dimension of pseudo-effective
divisors in positive characteristic, arXiv:1206.6521, to appear in the American Journal of Mathematics.
[Con00]
B. Conrad: Grothendieck duality and base change, Lecture Notes in Mathematics, vol. 1750, SpringerVerlag, Berlin, 2000. MR1804902 (2002d:14025)
[Eis95]
D. Eisenbud: Commutative algebra, Graduate Texts in Mathematics, vol. 150, Springer-Verlag, New York,
1995, With a view toward algebraic geometry. 1322960 (97a:13001)
[Fed83]
R. Fedder: F -purity and rational singularity, Trans. Amer. Math. Soc. 278 (1983), no. 2, 461–480.
[And93]
MR701505 (84h:13031)
[FST11]
[Gab04]
[GW77]
[Gro67]
[Hac11]
[HX13]
[Har98]
[HY03]
O. Fujino, K. Schwede, and S. Takagi: Supplements to non-lc ideal sheaves, Higher Dimensional
Algebraic Geometry, RIMS Kôkyûroku Bessatsu, B24, Res. Inst. Math. Sci. (RIMS), Kyoto, 2011, pp. 1–
47.
O. Gabber: Notes on some t-structures, Geometric aspects of Dwork theory. Vol. I, II, Walter de Gruyter
GmbH & Co. KG, Berlin, 2004, pp. 711–734.
S. Goto and K. Watanabe: The structure of one-dimensional F -pure rings, J. Algebra 49 (1977), no. 2,
415–421. MR0453729 (56 #11989)
A. Grothendieck: Éléments de géométrie algébrique. IV. Étude locale des schémas et des morphismes
de schémas IV, Inst. Hautes Études Sci. Publ. Math. (1967), no. 32, 361. MR0238860 (39 #220)
C. D. Hacon: Singularities of pluri-theta divisors in char p > 0, arXiv:1112.2219.
C. D. Hacon and C. Xu: On the three dimensional minimal model program in positive characteristic,
arXiv:1302.0298.
N. Hara: A characterization of rational singularities in terms of injectivity of Frobenius maps, Amer. J.
Math. 120 (1998), no. 5, 981–996. MR1646049 (99h:13005)
N. Hara and K.-I. Yoshida: A generalization of tight closure and multiplier ideals, Trans. Amer. Math.
Soc. 355 (2003), no. 8, 3143–3174 (electronic). MR1974679 (2004i:13003)
F -SINGULARITIES IN FAMILIES
[Har77]
[Har94]
59
R. Hartshorne: Algebraic geometry, Springer-Verlag, New York, 1977, Graduate Texts in Mathematics,
No. 52. MR0463157 (57 #3116)
R. Hartshorne: Generalized divisors on Gorenstein schemes, Proceedings of Conference on Algebraic
Geometry and Ring Theory in honor of Michael Artin, Part III (Antwerp, 1992), vol. 8, 1994, pp. 287–339.
MR1291023 (95k:14008)
[HS77]
[Kat08]
R. Hartshorne and R. Speiser: Local cohomological dimension in characteristic p, Ann. of Math. (2)
105 (1977), no. 1, 45–79. MR0441962 (56 #353)
M. Hashimoto: Cohen-Macaulay F-injective homomorphisms, Geometric and combinatorial aspects of
commutative algebra (Messina, 1999), Lecture Notes in Pure and Appl. Math., vol. 217, Dekker, New
York, 2001, pp. 231–244. 1824233 (2002d:13007)
M. Hashimoto: F -pure homomorphisms, strong F -regularity, and F -injectivity, Comm. Algebra 38
(2010), no. 12, 4569–4596. 2764840 (2012a:13007)
B. Hassett and S. J. Kovács: Reflexive pull-backs and base extension, J. Algebraic Geom. 13 (2004),
no. 2, 233–247. MR2047697 (2005b:14028)
M. Hochster: Foundations of tight closure theory, lecture notes from a course taught on the University
of Michigan Fall 2007 (2007).
M. Hochster and C. Huneke: F -regularity, test elements, and smooth base change, Trans. Amer. Math.
Soc. 346 (1994), no. 1, 1–62. MR1273534 (95d:13007)
M. Hochster and J. L. Roberts: The purity of the Frobenius and local cohomology, Advances in Math.
21 (1976), no. 2, 117–172. MR0417172 (54 #5230)
C. Huneke and I. Swanson: Integral closure of ideals, rings, and modules, London Mathematical Society
Lecture Note Series, vol. 336, Cambridge University Press, Cambridge, 2006. MR2266432 (2008m:13013)
M. Katzman: Parameter-test-ideals of Cohen-Macaulay rings, Compos. Math. 144 (2008), no. 4, 933–948.
[Kee03]
D. S. Keeler: Ample filters of invertible sheaves, J. Algebra 259 (2003), no. 1, 243–283. 1953719
[Kob75]
N. Koblitz: p-adic variation of the zeta-function over families of varieties defined over finite fields,
Compositio Math. 31 (1975), no. 2, 119–218. 0414557 (54 #2658)
J. Kollár: Projectivity of complete moduli, J. Differential Geom. 32 (1990), no. 1, 235–268. 1064874
[Has01]
[Has10]
[HK04]
[Hoc07]
[HH94]
[HR76]
[HS06]
MR2441251 (2009d:13030)
(2003m:14026)
[Kol90]
(92e:14008)
[Kc92]
[Kun69]
J. Kollár and 14 coauthors: Flips and abundance for algebraic threefolds, Société Mathématique
de France, Paris, 1992, Papers from the Second Summer Seminar on Algebraic Geometry held at the
University of Utah, Salt Lake City, Utah, August 1991, Astérisque No. 211 (1992). MR1225842 (94f:14013)
E. Kunz: Characterizations of regular local rings for characteristic p, Amer. J. Math. 91 (1969), 772–784.
MR0252389 (40 #5609)
[Kun86]
[Laz04a]
[Laz04b]
[Lyu97]
[MR85]
[MS97]
[MS12]
[Mil72]
[Mus11]
[MS12]
[MY09]
E. Kunz: Kähler differentials, Advanced Lectures in Mathematics, Friedr. Vieweg & Sohn, Braunschweig,
1986. MR864975 (88e:14025)
R. Lazarsfeld: Positivity in algebraic geometry. I, Ergebnisse der Mathematik und ihrer Grenzgebiete. 3.
Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics and Related Areas. 3rd Series.
A Series of Modern Surveys in Mathematics], vol. 48, Springer-Verlag, Berlin, 2004, Classical setting: line
bundles and linear series. MR2095471 (2005k:14001a)
R. Lazarsfeld: Positivity in algebraic geometry. II, Ergebnisse der Mathematik und ihrer Grenzgebiete.
3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics and Related Areas. 3rd
Series. A Series of Modern Surveys in Mathematics], vol. 49, Springer-Verlag, Berlin, 2004, Positivity for
vector bundles, and multiplier ideals. MR2095472 (2005k:14001b)
G. Lyubeznik: F -modules: applications to local cohomology and D-modules in characteristic p > 0, J.
Reine Angew. Math. 491 (1997), 65–130. MR1476089 (99c:13005)
V. B. Mehta and A. Ramanathan: Frobenius splitting and cohomology vanishing for Schubert varieties,
Ann. of Math. (2) 122 (1985), no. 1, 27–40. MR799251 (86k:14038)
V. B. Mehta and V. Srinivas: A characterization of rational singularities, Asian J. Math. 1 (1997),
no. 2, 249–271. MR1491985 (99e:13009)
L. E. Miller and K. Schwede: Semi-log canonical vs F -pure singularities, J. Algebra 349 (2012),
150–164. 2853631
L. Miller: Curves with invertible Hasse-Witt-matrix, Math. Ann. 197 (1972), 123–127. 0314849 (47 #3399)
M. Mustaţǎ: The non-nef locus in positive characteristic, arXiv:1109.3825, to appear in the volume
dedicated to Joe Harris on the occasion of his 60th birthday.
M. Mustaţǎ and K. Schwede: A Frobenius variant of Seshadri constants, arXiv:1203.1081.
M. Mustaţă and K.-I. Yoshida: Test ideals vs. multiplier ideals, Nagoya Math. J. 193 (2009), 111–128.
MR2502910
60
ZSOLT PATAKFALVI, KARL SCHWEDE, AND WENLIANG ZHANG
[Pat12]
[PS12]
[Rad92]
Zs. Patakfalvi: Semi-positivity in positive characteristics, arXiv:1208.5391.
Zs. Patakfalvi and K. Schwede: Depth of F -singularities and base change of relative canonical sheaves,
arXiv:1207.1910, to appear in the Journal of the Institute of Mathematics of Jussieu.
N. Radu: Une classe d’anneaux noethériens, Rev. Roumaine Math. Pures Appl. 37 (1992), no. 1, 79–82.
1172271 (93g:13014)
[RR85]
[Sch09]
[Sch11a]
[Sch11b]
[ST12]
[Sha07]
[SZ09]
[Smi97]
S. Ramanan and A. Ramanathan: Projective normality of flag varieties and Schubert varieties, Invent.
Math. 79 (1985), no. 2, 217–224. MR778124 (86j:14051)
K. Schwede: F -adjunction, Algebra Number Theory 3 (2009), no. 8, 907–950. 2587408 (2011b:14006)
K. Schwede: A canonical linear system associated to adjoint divisors in characteristic p > 0,
arXiv:1107.3833, to appear in the Journal für die reine und angewandte Mathematik.
K. Schwede: Test ideals in non-Q-Gorenstein rings, Trans. Amer. Math. Soc. 363 (2011), no. 11, 5925–
5941.
K. Schwede and K. Tucker: A survey of test ideals, Progress in Commutative Algebra 2. Closures,
Finiteness and Factorization (C. Francisco, L. C. Klinger, S. M. Sather-Wagstaff, and J. C. Vassilev, eds.),
Walter de Gruyter GmbH & Co. KG, Berlin, 2012, pp. 39–99.
R. Y. Sharp: On the Hartshorne-Speiser-Lyubeznik theorem about Artinian modules with a Frobenius
action, Proc. Amer. Math. Soc. 135 (2007), no. 3, 665–670 (electronic). 2262861 (2007f:13008)
K. Shimomoto and W. Zhang: On the localization theorem for F -pure rings, J. Pure Appl. Algebra 213
(2009), no. 6, 1133–1139. 2498803 (2010d:13006)
K. E. Smith: F -rational rings have rational singularities, Amer. J. Math. 119 (1997), no. 1, 159–180.
MR1428062 (97k:13004)
[Sri91]
V. Srinivas: Normal surface singularities of F -pure type, J. Algebra 142 (1991), no. 2, 348–359. MR1127067
(93b:14060)
[Sta]
[Tak04]
[Tan13]
[Tan72]
[Vie95]
T. Stacks Project Authors: Stacks Project.
S. Takagi: An interpretation of multiplier ideals via tight closure, J. Algebraic Geom. 13 (2004), no. 2,
393–415. MR2047704 (2005c:13002)
H. Tanaka: The trace map of frobenius and extending sections for threefolds, arXiv:1302.3134.
H. Tango: On the behavior of extensions of vector bundles under the Frobenius map, Nagoya Math. J.
48 (1972), 73–89. 0314851 (47 #3401)
E. Viehweg: Quasi-projective moduli for polarized manifolds, Ergebnisse der Mathematik und ihrer Grenzgebiete (3) [Results in Mathematics and Related Areas (3)], vol. 30, Springer-Verlag, Berlin, 1995.
MR1368632 (97j:14001)
[Zha12]
Y. Zhang: Pluri-canonical map of varieties of maximal albanese dimension in positive characteristic,
arXiv:1203.1360.
EPFL, SB MATHGEOM CAG, MA B3 635 (Bâtiment MA), Station 8, CH-1015 Lausanne
E-mail address: [email protected]
Department of Mathematics, University of Utah, Salt Lake City, UT 84112
E-mail address: [email protected]
Department of Mathematics, University of Illinois at Chicago, Chicago, IL 60607
E-mail address: [email protected]
| 0math.AC
|
Understanding Boltzmann Machine and Deep Learning via
A Confident Information First Principle
Xiaozhao Zhao
[email protected]
School of Computer Science and Technology, Tianjin University
Tianjin, 300072, China
arXiv:1302.3931v7 [cs.NE] 9 Oct 2013
Yuexian Hou
[email protected]
School of Computer Science and Technology, Tianjin University
Tianjin, 300072, China
Qian Yu
[email protected]
School of Computer Software, Tianjin University
Tianjin, 300072, China
Dawei Song
[email protected]
School of Computer Science and Technology, Tianjin University
Tianjin, 300072, China
and Department of Computing, The Open University
Milton Keynes, MK7 6AA, UK
Wenjie Li
[email protected]
Department of Computing, The Hong Kong Polytechnic University
Hung Hom, Kowloon, Hong Kong, China
Editor:
Abstract
Typical dimensionality reduction methods focus on directly reducing the number of random variables while retaining maximal variations in the data. In this paper, we consider
the dimensionality reduction in parameter spaces of binary multivariate distributions. We
propose a general Confident-Information-First (CIF) principle to maximally preserve parameters with confident estimates and rule out unreliable or noisy parameters. Formally,
the confidence of a parameter can be assessed by its Fisher information, which establishes
a connection with the inverse variance of any unbiased estimate for the parameter via the
Cramér-Rao bound. We then revisit Boltzmann machines (BM) and theoretically show
that both single-layer BM without hidden units (SBM) and restricted BM (RBM) can be
solidly derived using the CIF principle. This can not only help us uncover and formalize
the essential parts of the target density that SBM and RBM capture, but also suggest
that the deep neural network consisting of several layers of RBM can be seen as the layerwise application of CIF. Guided by the theoretical analysis, we develop a sample-specific
CIF-based contrastive divergence (CD-CIF) algorithm for SBM and a CIF-based iterative
projection procedure (IP) for RBM. Both CD-CIF and IP are studied in a series of density
estimation experiments.
Keywords: Boltzmann Machine, Deep Learning, Information Geometry, Fisher Information, Parametric Reduction
1
1. Introduction
Recently, deep learning models (e.g., Deep Belief Networks (DBN)(Hinton and Salakhutdinov, 2006), Stacked Denoising Auto-encoder (Ranzato et al., 2006), Deep Boltzmann
Machine (DBM) (Salakhutdinov and Hinton, 2012) and etc.) have drawn increasing attention due to their impressive results in various application areas, such as computer vision
(Bengio et al., 2006; Ranzato et al., 2006; Osindero and Hinton, 2007), natural language processing (Collobert and Weston, 2008) and information retrieval (Salakhutdinov and Hinton,
2007a,b). Despite of these practical successes, there have been debates on the fundamental
principle of the design and training of those deep architectures. One important problem is
the unsupervised pre-training, which would fit the parameters to better capture the structure
in the input distribution (Erhan et al., 2010). From the probabilistic modeling perspective,
this process can be interpreted as an attempt to recover a set of model parameters for a
generative neural network that would well describe the distribution underlying the observed
high-dimensional data. In general, to sufficiently depict the original high-dimensional data
requires a high-dimensional parameter space. However, overfitting usually occur when the
model is excessively complex. On the other hand, Dauphin and Bengio (2013) empirically
shows the failure of some big neural networks in leveraging the added capacity to reduce
underfitting.
Hence it is important to uncover and understand what the first principle would be on
reducing the dimensionality of the parameter space concerning the sample size. This question is also recognized by Erhan et al. (2010). They empirically show that the unsupervised
pre-training acts as a regularization on parameters in a way that the parameters are set in a
region, from which better basins of attraction can be reached. The regularization on parameters could be seen as a kind of dimensionality reduction procedure on parameter spaces
by restricting those parameters in a desired region. However, the intrinsic mechanisms
behind the regularization process are still unclear. Thus, further theoretical justifications
are needed in order to formally analyze what the essential parts of the target density that
the neural networks can capture. An indepth investigation into this question will lead to
two significant results: 1) a formal explanation on what exactly the neural networks would
perform in the pre-training process; 2) some theoretical insights on how to do it better.
Duin and Peȩkalska (2006) empirically show that the sampling density of a given dataset
and the resulting complexity of a learning problem are closely interrelated. If the initial
sampling density is insufficient, this may result in a preferred model of a lower complexity,
so that we have a satisfactory sampling to estimate model parameters. On the other hand,
if the number of samples is originally abundant, the preferred model may become more
complex so that we could have represented the dataset in more details. Moreover, this
connection becomes more complicated if the observed samples contain noises. Now the
obstacle is how to incorporate this relationship between the dataset and the preferred model
into the learning process. In this paper, we mainly focus on analyzing the Boltzmann
machines, the main building blocks of many neural networks, from a novel information
geometry (IG)(Amari and Nagaoka, 1993) perspective.
Assuming there exists an ideal parametric model S that is general enough to represent
all system phenomena, our goal of the parametric reduction is to derive a lower-dimensional
sub-model M for a given dataset (usually insufficient or perturbed by noises) by reducing
2
the number of free parameters in S. In this paper, we propose a Confident-InformationFirst (CIF) principle to maximally preserve the parameters with highly confident estimates
and rule out the unreliable or noisy parameters with respect to the density estimation of
binary multivariate distributions. From the IG point of view, the confidence 1 of a parameter
can be assessed by its Fisher information (Amari and Nagaoka, 1993), which establishes a
connection with the inverse variance of any unbiased estimate for the considered parameter
via the Cramér-Rao bound (see Rao, 1945). It is worth emphasizing that the proposed CIF
as a principle of parametric reduction is fundamentally different from the traditional feature
reduction (or feature extraction) methods (Fodor, 2002; Lee and Verleysen, 2007). The
latter focus on directly reducing the dimensionality on feature space by retaining maximal
variations in the data, e.g., Principle Components Analysis (PCA) (Jolliffe, 2002), while CIF
offers a principled method to deal with high-dimensional data in the parameter spaces by
a strategy that is universally derived from the first principle, independent of the geometric
metric in the feature spaces.
The CIF takes an information-oriented viewpoint of statistical machine learning. The
information 2 is rooted on the variations (or fluctuations) in the imperfect observations
(due to insufficient sampling, noise and intrinsic limitations of the “observer”) and transmits throughout the whole learning process. This idea is also well recognized in modern
physics, as stated in Wheeler (1994): “All things physical are information-theoretic in origin and this is a participatory universe...Observer participancy gives rise to information;
and information gives rise to physics.”. Following this viewpoint, Frieden (2004) unifies
the derivation of physical laws in major fields of physics, from the Dirac equation to the
Maxwell-Boltzmann velocity dispersion law, using the extreme physical information principle (EPI). The information used in this unification is exactly the Fisher information (Rao,
1945), which measures the quality of any measurement(s).
In terms of statistical machine learning, the aim of this paper is of two folds: a) to
incorporate the Fisher information into the modelling of the intrinsic variations in the data
that give rise to the desired model, by using the IG framework (Amari and Nagaoka, 1993);
b) to show by examples that some existing probabilistic models, e.g., SBM and RBM,
comply with the CIF principle and can be derived from it. The main contributions are:
1. We propose a general CIF principle for parameter reduction to maximally preserve
the parameters of high confidence and eliminate the unreliable parameters of binary
multivariate distributions in the framework of IG. We also give a geometric interpretation of CIF by showing that it can maximally preserve the expected information
distance.
2. The implementation of CIF, that is, the derivation of probabilistic models, is illustrated by revisiting two widely used Boltzmann machines, i.e., Single layer BM without hidden units (SBM) and restricted BM (RBM). And the deep neural network
consisting of several layers of RBM can be seen as the layer-by-layer application of
CIF.
1. Note that, in this paper, the meaning of confidence is different from the common concept degree of
confidence in statistics.
2. There are many kinds of “information” defined for a probability distribution p(x), e.g., entropy, Fisher
information. The entropy is a global measure of smoothness in p(x). Fisher information is a local
measure that is sensitive to local rearrangement of points (x, p(x)).
3
3. Based on the above theoretical analysis, a CIF-based iterative projection procedure
(IP) inherent in the learning of RBM is uncovered. And traditional gradient-based
methods, e.g., maximum-likelihood (ML) and contrastive divergence (CD) (CarreiraPerpinan and Hinton, 2005), can be seen as approximations of IP. Experimental results
indicate that IP is more robust against sampling biases, due to its separation of the
positive sampling process and the gradient estimation.
4. Beyond the general CIF, we propose a sample-specific CIF and integrate it into the CD
algorithm for SBM to confine the parameter space to a confident parameter region
indicated by samples. It leads to a significant improvement in a series of density
estimation experiments, when the sampling is insufficient.
The rest of the paper is organized as follows: Section 2 introduces some preliminaries of
IG. Then the general CIF principle is proposed in Section 3. In Section 4, we analyze two
implementations of CIF using the BM with and without hidden units, i.e., SBM and RBM.
After that, a sample-specific CIF-based CD learning method (CD-CIF) for SBM and a CIFbased iterative projection procedure for RBM are proposed and experimentally studied in
Section 5. Finally, we draw conclusions and highlight some future research directions in
Section 6.
2. Theoretical Foundations of Information Geometry
In this section, we introduce and develop the theoretical foundations of Information Geometry (IG) (Amari and Nagaoka, 1993) for the manifold S of binary multivariate distributions
with a given number of variables n, i.e., the open simplex of all probability distributions
over binary vector x ∈ {0, 1}n . This will lay the foundation for our theoretical deviation of
the general CIF.
2.1 Notations for Manifold S
In IG, a family of probability distributions is considered as a differentiable manifold with
certain parametric coordinate systems. In the case of binary multivariate distributions,
four basic coordinate systems are often used: p-coordinates, η-coordinates, θ-coordinates
and mixed-coordinates (Amari and Nagaoka, 1993; Hou et al., 2013). Mixed-coordinates is
of vital importance for our analysis.
For the p-coordinates [p] with n binary variables, the probability distribution over 2n
states of x can be completely specified by any 2n − 1 positive numbers indicating the
probability of the corresponding exclusive states on n binary variables. For example, the
p-coordinates of n = 2 variables could be [p] = (p01 , p10 , p11 ). Note that IG requires all
probability terms to be positive.
For simplicity, we use the capital letters I, J, . . . to index the coordinate parameters
of probabilistic distribution. An index I can be regarded as a subset of {1, 2, . . . , n}.
And pI stands for the probability that all variables indicated by I equal to one and
the complemented variables are zero. For example, if I = {1, 2, 4} and n = 4, then
pI = p1101 = P rob(x1 = 1, x2 = 1, x3 = 0, x4 = 1). Note that the null set can also be
a legal index of the p-coordinates, which indicates the probability that all variables are
zero, denoted as p0...0 .
4
Another coordinate system often used in IG is η-coordinates, which is defined by:
Y
ηI = E[XI ] = P rob{ xi = 1}
(1)
i∈I
Q
where the value of XI is given by i∈I xi and the expectation is taken with respect to
the probability distribution over x. Grouping the coordinates by their orders, η-coordinate
2 , . . . , ηn
system is denoted as [η] = (ηi1 , ηij
1,2...n ), where the superscript indicates the order
2 denotes the set of all η parameters
number of the corresponding parameter. For example, ηij
with the order number 2.
The θ-coordinates (natural coordinates) is defined by:
X
θI XI − ψ
(2)
log p(x) =
I⊆{1,2,...,n},I6=N ullSet
where ψ = − log P rob{xi = 0, ∀i ∈ {1, 2, ..., n}}. The θ-coordinate is denoted as [θ] =
(θ1i , θ2ij , . . . , θn1,...,n ), where the subscript indicates the order number of the corresponding
parameter. Note that the order indices locate at different positions in [η] and [θ] following
the convention in Amari et al. (1992).
The relation between coordinate systems [η] and [θ] is bijective (Amari et al., 1992).
More formally, they are connected by the Legendre transformation:
θI =
∂φ(η)
∂ψ(θ)
, ηI =
∂ηI
∂θI
(3)
where ψ(θ) and φ(η) meet the following identity
X
ψ(θ) + φ(η) −
θI ηI = 0
(4)
The function ψ(θ) is introduced in Eq. (2):
X
X
ψ(θ) = log(
exp{
θI XI (x)})
(5)
x
I
and hence φ(η) is the negative of entropy:
X
φ(η) =
p(x; θ(η)) log p(x; θ(η))
(6)
x
Next we introduce mixed-coordinates, which is important for our derivation of CIF. In
general, the manifold S of probability distributions could be represented by the l-mixedcoordinates (Amari et al., 1992):
i,j,...,k
2
l
[ζ]l = (ηi1 , ηij
, . . . , ηi,j,...,k
, θl+1
, . . . , θn1,...,n )
(7)
where the first part consists of η-coordinates with order less or equal to l (denoted by [η l− ])
and the second part consists of θ-coordinates with order greater than l (denoted by [θl+ ]),
l ∈ {1, ..., n}.
5
2.2 Fisher Information Matrix for Parametric Coordinates
For a general coordinate system [ξ], the ith-row and jth-col element of the Fisher information matrix for [ξ] (denoted by Gξ ) is defined as the covariance of the scores of [ξi ] and [ξj ]
(Rao, 1945), i.e.,
∂ log p(x; ξ) ∂ log p(x; ξ)
gij = E[
·
]
∂ξi
∂ξj
The Fisher information measures the amount of information in the data that a statistic
carries about the unknown parameter (Kass, 1989). The Fisher information matrix is of
vital importance to our analysis because the inverse of Fisher information matrix gives an
asymptotically tight lower bound of the covariance matrix of any unbiased estimate for the
considered parameters (Rao, 1945). Moreover, the Fisher information matrix, as a distance
metric, is invariant to re-parameterization (Rao, 1945), and can be proved to be the unique
metric that is invariant to the map of random variables corresponding to a sufficient statistic
(Amari et al., 1992; Chentsov, 1982).
Another important concept related to our analysis is the orthogonality defined by Fisher
information. Two coordinate parameters ξi and ξj are called orthogonal if and only if their
Fisher information vanishes, i.e., gij = 0, meaning that their influences to the log likelihood
function are uncorrelated. A more technical meaning of orthogonality is that the maximum
likelihood estimates (MLE) of orthogonal parameters can be independently performed.
∂ 2 ψ(θ)
The Fisher information for [θ] can be rewritten as gIJ = ∂θ
I ∂θ J , and for [η] it is
2
∂ φ(η)
g IJ = ∂η
(Amari and Nagaoka, 1993). Let Gθ = (gIJ ) and Gη = (g IJ ) be the Fisher
I ∂ηJ
information matrices for [θ]
It can be shown that Gθ and Gη are mutuPand [η] respectively.
I , where δ I = 1 if I = K and 0 otherwise (Amari
ally inverse matrices, i.e., J g IJ gJK = δK
K
and Nagaoka, 1993).
In order to generally compute Gθ and Gη , we develop the following Proposition 1 and
2. Note that Proposition 1 is a generalization of Theorem 2 in Amari et al. (1992).
Proposition 1 The Fisher information between two parameters θI and θJ in [θ], is given
by
gIJ (θ) = ηI S J − ηI ηJ
(8)
Proof in Appendix A.1
Proposition 2 The Fisher information between two parameters ηI and ηJ in [η], is given
by
X
1
g IJ (η) =
(−1)|I−K|+|J−K| ·
(9)
pK
K⊆I∩J
where | · | denotes the cardinality operator.
Proof in Appendix A.2
For example, let [p] = (p001 , p010 , p011 , p100 , p101 , p110 , p111 ) be the p-coordinates of a probability distribution with three variables. Then, the Fisher information of ηI and ηJ can be
1
1
+ p010
, and if
calculated based on Equation (9): if I = {1, 2} and J = {2, 3}, g IJ = p000
1
1
1
1
IJ
I = {1, 2} and J = {1, 2, 3}, g = −( p000 + p010 + p100 + p110 ), etc.
6
3. The General CIF Principle For Parametric Reduction
As described in Section 2.1, the general manifold S of all probability distributions over
binary vector x ∈ {0, 1}n , could be exactly represented using the parametric coordinate
systems of dimensionality 2n − 1. However, given the limited samples generated from a target distribution, it is almost infeasible to determine its coordinates in such high-dimensional
parameter space with acceptable accuracy and in reasonable time. Given a target distribution q(x) on the general manifold S, we consider the problem of realizing it by a lowerdimensionality submanifold. This is defined as the problem of parametric reduction for
multivariate binary distributions.
3.1 The General CIF Principle
In this section, we will formally illuminate the general CIF for parametric reduction. Intuitively, if we can construct a coordinate system so that the confidences (measured by Fisher
information) of its parameters entail a natural hierarchy, in which high confident parameters can be significantly distinguished from low confident ones, then the general CIF can be
conveniently implemented by keeping the high confident parameters unchanging and setting
the lowly confident parameters to neutral values. However, the choice of coordinates (or
equivalently, parameters) in CIF is crucial to its usage. This strategy is infeasible in terms
of p-coordinates, η-coordinates or θ-coordinates, since it is easy to see that the hierarchies
of confidences in these coordinate systems are far from significant, as shown by an example
later in this section.
The following propositions show that mixed-coordinates meet the requirement realizing
the general CIF. Let [ζ]l be the mixed-coordinates defined in Section 2.1. Proposition 3
gives a closed form for calculating the Fisher information matrix Gζ .
Proposition 3 The Fisher information matrix of the l-mixed-coordinates [ζ]l is given by:
A 0
(10)
Gζ =
0 B
−1
−1
−1
where A = ((G−1
η )Iη ) , B = ((Gθ )Jθ ) , Gη and Gθ are the Fisher information matrices
of [θ] and [ζ]l , respectively, and Iη is the index of the parameters shared by [η] and [ζ]l ,
l }, and J is the index set of the parameters shared by [θ] and [ζ] , i.e.,
i.e., {ηi1 , ..., ηij
θ
l
i,j,...,k
{θl+1
, . . . , θn1,...,n }.
Proof in Appendix A.3.
Proposition 4 The diagonal of A are lower bounded by 1, and that of B are upper bounded
by 1.
Proof in Appendix A.4.
According to Proposition 3 and Proposition 4, the confidences of coordinate parameters
in [ζ]l entail a natural hierarchy: the first part of high confident parameters [η l− ] are separated from the second part of low confident parameters [θl+ ], which has a neutral value
(zero). Moreover, the parameters in [η l− ] is orthogonal to the ones in [θl+ ], indicating that
7
Figure 1: By projecting a point q(x) on S to a submanifold M , the l-tailored-mixedcoordinates [ζ]lt gives a desirable M that maximally preserve the expected Fisher
information distance when projecting a ε-neighborhood centered at q(x) onto M .
we could estimate these two parts independently (Hou et al., 2013). Hence we can implement the general CIF principle for parametric reduction in [ζ]l by replacing low confident
parameters with neutral value zeros and reconstructing the resulting distribution. It turns
l
, 0, . . . , 0). We call
out that the submanifold tailored by CIF becomes [ζ]lt = (ηi1 , ..., ηij...k
[ζ]lt the l-tailored-mixed-coordinates.
To grasp an intuitive picture for the general CIF strategy and its significance w.r.t
mixed-coordinates, let us consider an example with [p] = (p001 = 0.15, p010 = 0.1, p011 =
0.05, p100 = 0.2, p101 = 0.1, p110 = 0.05, p111 = 0.3). Then the confidences for coordinates
in [η], [θ] and [ζ]2 are given by the diagonal elements of the corresponding Fisher information matrices. Applying the 2-tailored CIF in mixed-coordinates, the loss ratios of Fisher
information is 0.001%, and the ratio of the Fisher information of the tailored parameter
(θ3123 ) to the remaining η parameter with the smallest Fisher information is 0.06%. On the
other hand, the above two ratios become 7.58% and 94.45% (in η-coordinates) or 12.94%
and 92.31% (in θ-coordinates), respectively.
Next, we will restate the CIF principle in terms of the geometric perspective of submanifold projection. Let M be a smooth submanifold in S. Given a point q(x) ∈ S, the
projection of q(x) to M is the point p(x) that belongs to M and is closest to q(x) in the
sense of the Kullback-Leibler divergence (K-L divergence) from the distribution q(x) to p(x)
(Amari et al., 1992):
X
q(x)
D(q(x), p(x)) =
q(x) log
(11)
p(x)
x
Alternatively, since K-L divergence is not symmetric, the projection of q(x) to M can also
be defined as the point p(x) ∈ M that minimizes the K-L divergence from M to q(x). In
the rest of this paper, the direction of the K-L divergence used in a particular projection is
explicitly specified when there is an ambiguity.
8
The CIF entails a submanifold of S via the l-tailored-mixed-coordinates [ζ]lt . However,
there exist many different submanifolds of S. Now our question is: does there exist a
general criterion to distinguish which projection is best? If such principle does exist, is CIF
the right one? The following proposition shows that the general CIF entails a geometric
interpretation illuminated in Figure 1, which would lead to an optimal submanifold M .3
Proposition 5 Given a statistical manifold S in l-mixed-coordinates [ζ]l , let the corresponding l-tailored-mixed-coordinates [ζ]lt has k free parameters. Then, among all k-dimensional
submanifolds of S, the submanifold determined by [ζ]lt can maximally preserve the expected
information distance induced by Fisher-Rao metric.
Proof in Appendix A.5.
4. Two Implementations of CIF using Boltzmann Machine
In previous section, a general CIF is uncovered in the [ζ]l coordinates for multivariate binary
distributions. Now we consider the implementations of CIF when l equals to 2 using the
Boltzmann machines (BM). More specifically, we show that two kinds of BMs, i.e., the
single layer BM without hidden units (SBM) and the restricted BM (RBM), are indeed
instances following the general CIF principle. For each case, the application of CIF can be
interpreted in two perspectives: an algebraic and geometric interpretation.
4.1 Neural Networks as Parametric Reduction Model
Many neural networks with fixed architecture, such as SBM, RBM, high-order BM (Albizuri
et al., 1995), deep belief networks (Hinton and Salakhutdinov, 2006), have been proposed to
approximately realize the underlying distributions in different application scenarios. Those
neural networks are designed to fulfill the parametric reduction for certain tasks by specifying the number of adjustable parameters, namely the number of connection weights and the
number of biases. We believe that there exists a general criterion to design the structure of
neural submanifolds for the particular application in hand, and the problem of parametric
reduction is equivalent to the choice of submanifolds. Next, we will briefly introduce the
general BM and the gradient-based learning algorithm.
4.1.1 Introduction To The Boltzmann Machines
In general, a BM (Ackley et al., 1985) is defined as a stochastic neural network consisting of
visible units x ∈ {0, 1}nx and hidden units h ∈ {0, 1}nh , where each unit fires stochastically
depending on the weighted sum of its inputs. The energy function is defined as follows:
1
1
EBM (x, h; ξ) = − xT U x − hT V h − xT W h − bT x − dT h
2
2
(12)
3. Note that the CIF is related to but fundamentally different from the m-projection in Amari et al. (1992).
Amari et al. (1992) focuses on the problem of projecting a point Q on S to the submanifold of BM and
shows that m-projection is the point on BM that is closest to Q. Actually, the m-projection is a special
case of our [ζ]lt -projection when l is 2. In the present paper, we focus on the problem of developing a
general criterion that could help us find the optimal submanifold to project on.
9
where ξ = {U, V, W, b, d} are the parameters: visible-visible interactions (U ), hidden-hidden
interactions (V ), visible-hidden interactions (W ), visible self-connections (b) and hidden selfconnections (d). The diagonals of U and V are set to zero. We can express the Boltzmann
distribution over the joint space of x and h as below:
p(x, h; ξ) =
1
exp{−EBM (x, h; ξ)}
Z
(13)
where Z is a normalization factor.
Let B be the set of Boltzmann distributions realized by BM. Actually, B is a submanifold
of the general manifold Sxh over {x, h}. From Equation (13) and (12), we can see that
ξ = {U, V, W, b, d} plays the role of B’s coordinates in θ-coordinates (Equation 2) as follows:
h
θ1 : θ1xi = bxi , θ1 j = dhj (∀xi ∈ x, hj ∈ h)
x xj
θ2 : θ2 i
x hj
= Uxi ,xj , θ2 i
x ...xj hu ...hv
θ2+ : θmi
h hj
= Wxi ,hj , θ2 i
= Vhi ,hj (∀xi , xj ∈ x; hi , hj ∈ h)
= 0, m > 2, (∀xi , . . . , xj ∈ x; hu , . . . , hv ∈ h)
(14)
So the θ-coordinates for BM is given by:
h
xh
xx
hh
[θ]BM = (θ1xi , θ1 j , θ2 i j , θ2 i j , θ2 i j , 0, . . . , 0).
{z
} | {z }
| {z } |
1−order
(15)
orders>2
2−order
The SBM and RBM are special cases of the general BM. Since SBM has nh = 0 and all
the visible units are connected to each other, the parameters of SBM are ξsbm = {U, b} and
{V, W, d} are all set to zero. For RBM, it has connections only between hidden and visible
units. Thus, the parameters of RBM are ξrbm = {W, b, d} and {U, V } are set to zero.
4.1.2 Formulation on the Gradient-based Learning of BM
Given the sample x that generated from the underlying distribution, the maximum-likelihood
(ML) is commonly used gradient ascent method for training BM in order to maximize the
log-likelihood log p(x; ξ) of the parameters ξ (Carreira-Perpinan and Hinton, 2005). Based
on Equation (13), the log-likelihood is given as follows:
X
X
0 0
log p(x; ξ) = log
e−E(x,h;ξ) − log
e−E(x ,h ;ξ)
x0 ,h0
h
Differentiating the log-likelihood, the gradient vector with respect to ξ is as follows:
∂ log p(x; ξ) X
∂[−E(x, h; ξ)] X
∂[−E(x0 , h0 ; ξ)]
p(h0 |x0 ; ξ)
=
p(h|x; ξ)
−
∂ξ
∂ξ
∂ξ
0 0
h
(16)
x ,h
The ∂E(x,h;ξ)
can be easily calculated from Equation (12). Then we can obtain the stochastic
∂ξ
gradient using Gibbs sampling (Gilks et al., 1996) in two phases: sample h given x for the
first term, called the positive phase, and sample (x0 , h0 ) from the stationary distribution
p(x0 , h0 ; ξ) for the second term, called the negative phase. Now with the resulting stochastic
gradient estimation, the learning rule is to adjust ξ by:
∆ξ = ε ·
∂ log p(x; ξ)
∂E(x, h; ξ)
∂E(x0 , h0 ; ξ)
= ε · (−h
i0 + h
i∞ )
∂ξ
∂ξ
∂ξ
10
(17)
where ε is the learning rate, h·i0 denotes the average using the sample data and h·i∞ denotes
the average with respect to the stationary distribution p(x, h; ξ) after the corresponding
Gibbs sampling phases.4
In following sections, we will revisit two special BM, namely SBM and RBM, and theoretically show that both SBM and RBM can be derived using the CIF principle. This helps
us formalize what essential parts of the target density the SBM and RBM capture.
4.2 The CIF-based Derivation of Boltzmann Machine without Hidden Units
Given any underlying probability distribution q(x) on the general manifold S over {x}, the
logarithm of q(x) can be represented by a linear decomposition of θ-coordinates as shown
in Equation (2). Since it is impractical to recognize all coordinates for the target distribution, we would like to only approximate part of them and end up with a k-dimensional
submanifold M of S, where k ( 2nx − 1) is the number of free parameters. Here, we set
k to be the same dimensionality as SBM, i.e., k = nx (n2x +1) , so that all candidate submanifolds are comparable to the submanifold endowed by SBM (denoted as Msbm ). Next, the
rationale underlying the design of Msbm can be illuminated using the general CIF from two
perspectives, algebraically and geometrically.
4.2.1 SBM as 2-tailored-mixed-coordinates
2 , θ i,j,k , . . . , θ 1,...,nx ). Applying
Let the 2-mixed-coordinates of q(x) on S be [ζ]2 = (ηi1 , ηij
nx
3
the general CIF on [ζ]2 , our parametric reduction rule is to preserve the high confident
part parameters [η 2− ] and replace low confident parameters [θ2+ ] with a fixed neutral value
2 , 0, . . . , 0), as the
zero. Thus we derive the 2-tailored-mixed-coordinates: [ζ]2t = (ηi1 , ηij
optimal approximation of q(x) by the k-dimensional submanifolds. On the other hand,
given the 2-mixed-coordinates of q(x), the projection p(x) ∈ Msbm of q(x) is proved to be
2 , 0, . . . , 0) (Amari et al., 1992). Thus, SBM defines a probabilistic parameter
[ζ]p = (ηi1 , ηij
space that is exactly derived from CIF.
4.2.2 SBM as Maximal Information Distance Projection
Next corollary, following Proposition 5, shows a geometric derivation of SBM. We make it
explicit that the projection on Msbm could maximally preserve the expected information
distance comparing to other tailored submanifolds of S with the same dimensionality k.
Corollary 6 Given the general manifold S in 2-mixed-coordinates [ζ]2 , SBM defines an kdimensional submanifold of S that can maximally preserve the expected information distance
induced by Fisher-Rao metric.
Proof in Appendix A.6.
From the CIF-based derivation, we can see that SBM confines the statistical manifold in
the parameter subspace spanned by those directions with high confidences, which is proved
to maximally preserve the expected information distance.
4. Both the two special BMs, i.e., SBM and RBM, can be trained using the ML method. Note that SBM
has no hidden units and hence no positive sampling is needed in training SBM.
11
4.2.3 The Relation Between [ζ]2t and the ML Learning of SBM
To learn such [ζ]2t , we need to learn the parameters ξ of SBM such that its stationary
distribution preserves the same coordinates [η 2− ] as target distribution q(x). Actually, this
is exactly what traditional gradient-based learning algorithms intend to do to train SBM.
Next proposition shows that the ML method for training SBM is equivalent to learn the
tailored 2-mixed coordinates [ζ]2t .
Proposition 7 Given the target distribution q(x) with 2-mixed coordinates:
2
[ζ]2 = (ηi1 , ηij
, θ2+ ),
the coordinates of the SBM with stationary distribution q(x; ξ), learnt by ML, are uniquely
2 ,θ
given by [ζ]2t = (ηi1 , ηij
2+ = 0)
Proof in Appendix A.7.
4.3 The CIF-based Derivation of Restricted Boltzmann Machine
In previous section, the general CIF uncovers why SBM uses the coordinates up to 2nd order, i.e., preserves the η-coordinates of the 1st -order and 2nd -order. In this section, we
will investigate the cases where hidden units are introduced. Particularly, one of the fundamental problem in neural network research is the unsupervised representation learning
(Bengio et al., 2013), which attempts to characterize the underlying distribution through
the discovery of a set of latent variables (or features). Many algorithmic models have been
proposed, such as restricted Boltzmann machine (RBM) (Hinton and Salakhutdinov, 2006)
and auto-encoders (Rifai et al., 2011; Vincent et al., 2010), for learning one level of feature
extraction. Then, in deep learning models, the representation learnt at one level is used as
input for learning the next level, etc. However, some important questions remain to be clarified: Do these algorithms implicitly learn about the whole density or only some aspects?
If they capture the essence of the target density, then can we formalize the link between the
essential part and omitted part? This section will try to answer these questions using CIF.
4.3.1 Two Necessary Conditions for Representation Learning
In terms of one level feature extraction, there are two main principles that guide a good
representation learning:
• Compactness of representation: minimize the redundancy between hidden variables
in the representation 5 .
• Completeness of reconstruction: the learnt representation captures sufficient information in the input, and could completely reconstruct input distribution, in a statistical
sense.
5. The concept of compactness in the neural network is of two-folds. 1) model-scale compactness: a
restriction on the number of hidden units in order to give a parsimonious representation w.r.t underlying
distribution; 2) structural compactness: a restriction on how hidden units are connected such that the
redundancy in the hidden representation is minimized. In this paper, we mainly focus on the latter case.
12
Let Sxh be the general manifold of probability distributions over the joint space of
visible units x and hidden units h, and Sx be the general manifold over visible units x.
Given any observation distribution q(x) ∈ Sx , our problem is to find the p(x, h) ∈ Sxh
with the marginal distribution p(x) that best approximates q(x), while p(x, h) is consistent
with the compactness and completeness conditions. Here, the K-L divergence, defined in
Equation (11), is used as the criterion of approximation.
First, we will investigate the submanifold of joint distributions p(x, h) ∈ Sxh that fulfill
the above two conditions. Let us denote this submanifold as Mcc . Extending Equation (2)
to manifold Sxh , p(x, h) has the θ-coordinates defined by:
X
θI XI − ψ
(18)
log p(x, h) =
I⊆{x,h}&I6=N ullSet
For the completeness requirement, it is easy to prove that the probability of any input
variable xi can be fully determined only by the given hidden representation and independent
with remainng input variables xj (j 6= i), if and only if θI = 0 for any I that contains two
or more input variables in Equation (18). Similarly, the compactness corresponds to the
extraction of statistically independent hidden variables given the input, i.e., θI = 0 for any
I that contains two or more hidden variables in Equation (18). Then Mcc is given by the
following coordinate system:
h
xh
xx
h hj
[θ]cc = (θ1xi , θ1 j , θ2 i j = 0, θ2 i j , θ2 i
| {z } |
{z
1−order
= 0, 0, . . . , 0).
} | {z }
(19)
orders>2
2−order
Then our problem is restated as to find the p(x, h) ∈ Mcc with the marginal distribution
p(x) that best approximates q(x).
4.3.2 The Equivalence between RBM and Mcc
RBM is a special kind of BM that restricts the interactions in Equation (12) only to those
between hidden and visible units, i.e., Uxi ,xj = 0, Vhi ,hj = 0 ∀xi , xj ∈ x; hi , hj ∈ h. Let
ξrbm = {W, b, d} denotes the set of parameters in RBM. Thus, the θ-coordinates for RBM
can be derived directly from Equation (15):
h
xx
xh
h hj
[θ]RBM = (θ1xi , θ1 j , θ2 i j = 0, θ2 i j , θ2 i
| {z } |
{z
1−order
2−order
= 0, 0, . . . , 0)
} | {z }
(20)
orders>2
Comparing Equation (19) to (20), the submanifold Mrbm defined by RBM is equivalent
with Mcc since they share exactly the same coordinate system. This indicates that the
compactness and completeness conditions is indeed realized by RBM. We use a simpler
notation B to denote Mrbm . Next, we will show how to use CIF to interpret the training
process of RBM.
4.3.3 The CIF-based Interpretation on the Learning of RBM
A RBM produces a stationary distribution p(x, h) ∈ Sxh over {x, h}. However, given the
target distribution q(x), only the marginal distribution of RBM over the visible units are
13
specified by q(x), leaving the distributions on hidden units vary freely. Let Hq be the
set of probability distributions q(x, h) ∈ Sxh that have the same marginal distribution
q(x) and the conditional distributions q(hj |x) of each hidden unit hj is realized by the
RBM’s activation function with parameter ξrbm (that is the logistic sigmoid activation:
1
f (hj |x; ξrbm ) = 1+exp{− P
Wij xi −dj } ):
i∈{1,...,nx }
Hq = {q(x, h) ∈ Sxh |∃ξrbm ,
X
q(x, h) = q(x), and q(h|x; ξrbm ) =
Y
f (hj |x; ξrbm )}
(21)
hj ∈h
h
Then our problem in Section 4.3.1 is restated with respect to Sxh : search for a RBM in B
that minimizes the divergence from Hq to B 6 .
Given p(x, h; ξp ), its best approximation on Hq is defined by the projection ΓH (p), which
gives the minimum K-L divergence from Hq to p(x, h; ξp ). Next proposition shows how the
projection ΓH (p) is obtained.
Proposition 8 Given a distribution p(x, h; ξp ) ∈ B, the projection ΓH (p) ∈ Hq that gives
the minimum divergence D(Hq , p(x, h; ξp )) from Hq to p(x, h; ξp ) is the q(x, h; ξq ) ∈ Hq that
satisfies ξq = ξp .
Proof in Appendix A.8.
On the other hand, given q(x, h; ξq ) ∈ Hq , the best approximation on B is the projection
ΓB (q) of q to B. In order to obtain an explicit expression of ΓB (q), we introduce the following
fractional mixed coordinates [ζ xh ] 7 for the general manifold Sxh :
hh
xx
[ζ xh ] = (ηx1i , ηh1j , θ2 i j , ηx2i hj , θ2 i j , θ2+ )
| {z } |
{z
} |{z}
1−order
2−order
(22)
orders>2
The [ζ xh ] is a valid coordinate system, that is, the relation between [θ] and [ζ xh ] is bijective.
This is shown in the next proposition.
Proposition 9 The relation between the two coordinate systems [θ] and [ζ xh ] is bijective.
Proof in Appendix A.9.
The next proposition gives an explicit expression of the coordinates for the projection
ΓB (q) learnt by RBM using the fractional mixed coordinates in Equation (22).
6. This restated problem directly follows from the fact that: the minimum divergence D(Hq , B) in the
whole manifold Sxh is equal to the minimum divergence D[q(x), Bx ] in the visible manifold Sx , shown
in Theorem 7 in Amari et al. (1992).
7. Note that both the fractional mixed coordinates [ζ xh ] and 2-mixed coordinates [ζ] are mixtures of ηcoordinates and θ-coordinates. In [ζ], coordinates of the same order are taken from either [η] or [θ].
h h
x x
However, in [ζ xh ], the 2nd -order coordinates consist of the {ηx2i hj } from [η] and {θ2 i j , θ2 i j } from [θ],
that is why the term “fractional” is used.
14
Figure 2: The iterative learning for RBM: in searching for the minimum divergence between
Hq and B, we first choose an initial RBM p0 and then perform projections ΓH (p)
and ΓB (q) iteratively, until the fixed points of the projections p∗ and q ∗ are
reached. With different initializations, the iterative projection algorithm may
end up with different local minima on Hq and B, respectively.
Proposition 10 Given q(x, h; ξq ) ∈ Hq with fractional mixed coordinates:
xx
hh
[ζ xh ]q = (ηx1i , ηh1j , θ2 i j , ηx2i hj , θ2 i j , θ2+ ),
the coordinates of the learnt projection ΓB (q) of q(x, h; ξq ) on the submanifold B are uniquely
given by:
hh
xx
(23)
[ζ xh ]ΓB (q) = (ηx1i , ηh1j , θ2 i j = 0, ηx2i hj , θ2 i j = 0, θ2+ = 0)
Proof This proof comes in three parts:
1. the projection ΓB (q) of q(x, h) on B is unique;
2. this unique projection ΓB (q) can be achieved by minimizing the divergence D[q(x, h), B]
using gradient descent method;
3. The fractional mixed coordinates of ΓB (q) is exactly the one given in Equation (23).
See Appendix A.10 for the detailed proof.
Back to the problem of obtaining the best approximation to the given target q(x), the
learning of RBM can be implemented by the following iterative projection process8 :
Let p0 (x, h; ξp0 ) be the initial RBM. For i = 0, 1, 2, . . . ,
1. Put qi+1 (x, h) = ΓH (pi (x, h; ξpi ))
8. Amari et al. (1992) proposed a similar iterative algorithm framework for the fully-connected BM. In
the present paper, we reformulate this iterative algorithm for the learning of RBM and give explicit
expressions of how the projections are achieved.
15
2. Put pi+1 (x, h; ξpi+1 ) = ΓB (qi+1 (x, h))
where ΓH (p) denotes the projection of p(x, h; ξp ) to Hq , and ΓB (q) denotes the projection
of q(x, h) to B. The iteration ends when we reach the fixed points of the projections p∗
and q ∗ , that is ΓH (p∗ ) = q ∗ and ΓB (q ∗ ) = p∗ . The iterative projection process of RBM is
illustrated in Figure 2. The convergence property of this iterative algorithm is guaranteed
using the following proposition:
Proposition 11 The monotonic relation holds in the iterative learning algorithm:
D[qi+1 , pi ] ≥ D[qi+1 , pi+1 ] ≥ D[qi+2 , pi+1 ], ∀i = {0, 1, 2, . . . }
(24)
where the equality holds only for the fixed points of the projections.
Proof in Appendix A.11.
The CIF-based iterative projection procedure (IP) for RBM gives us an alternative way
to investigate the learning process of RBM. The invariance in the learning of RBM is the
CIF: in the ith iteration, given qi ∈ Hq with fractional mixed coordinates:
hh
xx
[ζ xh ]qi = (ηx1i , ηh1j , θ2 i j , ηx2i hj , θ2 i j , θ2+ )
then the ordinates of the projection pi on B, i.e., ΓB (qi ), is given by Equation (23):
x xj
[ζ xh ]pi = (ηx1i , ηh1j , θ2 i
h hj
= 0, ηx2i hj , θ2 i
= 0, θ2+ = 0)
Now we will show that the process of the projection ΓB (qi ) can be derived from CIF, i.e.,
highly confident coordinates [ηx1i , ηh1j , ηx2i hj ] of qi are preserved while lowly confident coordinates [θ2+ ] are set to neutral value zero. For the fractional mix coordinates system
[ζ xh ], the closed form of its Fisher information matrix does not have the good expression
formula like Proposition 3 which are possessed by the mixed-coordinate system [ζ]. Next,
we will show that the fractional mix-coordinates of ΓB (qi ) can be derived in three steps by
jointly applying the CIF and the completeness and compactness conditions. Let the corresponding 2-mixed ζ-coordinates for qi be [ζ]2,qi = (ηx1i , ηh1j , ηx2i xj , ηx2i hj , ηh2i hj , θ2+ ). First,
we apply the general CIF for parametric reduction in [ζ]2,qi by replacing lowly confident
coordinates [θ2+ ] with neutral value zeros and preserving the remaining coordinates, resulting in the tailored mix-coordinates [ζ]2t ,qi = (ηx1i , ηh1j , ηx2i xj , ηx2i hj , ηh2i hj , θ2+ = 0), as
described in Section 3. Then, we transmit [ζ]2t ,qi into the fractional coordinate system, i.e.,
xx
hh
[ζ xh ]qi = (ηx1i , ηh1j , θ2 i j , ηx2i hj , θ2 i j , θ2+ = 0). Finally, the completeness and compactness
xx
hh
conditions require that the [θ2 i j , θ2 i j ] in [ζ xh ]qi are also set to neutral value zeros. Hence,
we can see that the coordinates of the projection ΓB (qi ) is exactly the one given by Equation
(23).
4.3.4 Comparison of The Iterative Projection and Gradient-based Methods
Given current parameters ξ i of RBM and samples x that generated from the underlying
distribution q(x), IP could be implemented in two phases:
16
1. In the first phase, we generate samples for the projection ΓH (pi ) of the stationary
distribution pi (x, h; ξ i ) on Hq . This is done by sampling h from RBM’s conditional
distribution pi (h|x; ξ i ) given x, and hence (x, h) ∼ ΓH (pi (x, h; ξ i ));
2. In the second phase, we train a new RBM with those generated samples (x, h), and
then update the RBM’s parameters to be the newly trained ones, denoted as ξ i+1 .
Note that in the second phase all hidden units in RBM are visible in samples (x, h). Thus
this sub-learning task is similar with training a BM without hidden units, which can be
implemented by traditional gradient-based methods.
Given current parameters ξ i of RBM (with stationary distribution pi ) and samples
x ∼ q(x), we can see that both ML and IP share the same sampling process, sampling
(x, h) in the ΓH (pi ) projection phase of IP and the positive phase of ML. In terms of the
quality of the sampling process, if x is sufficient and so is (x, h), both the ΓH (pi ) of IP and
positive phase of ML can achieve an accurate estimation of q(x, h; ξ i ). On the other hand,
if x is insufficient (it is usually true in real-world applications), sampling biases with respect
to q(x, h; ξ i ) may be introduced in the sampling process so that the accurate estimation can
not be guaranteed.
However, the updating rule is different: IP realizes the parameter updating by using a
sub-learning task (fitting a new RBM to the generated sample (x, h)), while ML adjusts the
parameters directly using Equation (17). Let qi+1 denote the distribution of (x, h), and pi+1
and p0i+1 denote the stationary distributions of RBM after the parameter updating using
IP and ML respectively. With a proper learning rate λ, the parameter updating phase of
ML would lead to the decrease of divergence, that is, D[qi+1 , pi ] ≥ D[qi+1 , p0i+1 ]. Since
pi+1 is the projection of qi+1 on B, then we have D[qi+1 , p0i+1 ] ≥ D[qi+1 , pi+1 ]. Therefore,
ML can be seen as an “unmature projection” of qi+1 on B, and it does not guarantee
that the theoretical projection ΓB (qi+1 ) is reached. To achieve the same projection point
ΓB (qi+1 ) as IP, ML needs multiple updating iterations, where each iteration moves the
current distribution towards ΓB (qi+1 ) in the gradient direction by some oracle step size
(controlled by the learning rate).
Another big difference is that IP separates the positive sampling process and the gradient
estimation in two phases: ΓH and ΓB , meaning that there is no positive sampling in the
sub-learning of ΓB . However, ML needs to constantly adjust the gradient direction with
respect to certain learning rate immediately after each sampling process. Later experiments
in Section 5.2 indicate that this may give IP the advantage of robustness against sampling
biases, especially when the gradient is too small to be distinguishable from these biases in
the learning process.
4.3.5 Discussions on deep Boltzmann machine
For deep Boltzmann machine (DBM) (Salakhutdinov and Hinton, 2012), several layers
of RBM compose a deep architecture in order to achieve a representation at a sufficient
abstraction level, where the hidden units are trained to capture the dependencies of units
at the lower layers, as shown in Figure 3. In this section, we give a discussion on some
theoretical insights on the deep architectures, in terms of the CIF principle.
In Section 4.3.2, we have shown that the structure of RBM implies the compactness and
completeness conditions, which could guide the learning of a good representation. Thus
17
Figure 3: A multi-layer BM with visible units x and hidden layers h(1) , h(2) and h(3) . The
greedy layer-wise training of deep architecture is to maximally preserve the confident information layer by layer. Note that the prohibition sign indicates that
the Fisher information on lowly confident coordinates is not preserved.
DBM can be seen as the composition of a series of representation learning stages. Then,
an immediate question is: what kind of representation of the data should be generated as
the output of each stage? From an information abstraction point of view, each stage of
the deep architecture could build up more abstract features by using the highly confident
information on parameters (or coordinates) that is transmitted from less abstract features in
lower layers. Those more abstract features would potentially have a greater representation
power (Bengio et al., 2013). The CIF principle describes how the information flows in those
representation transformations, as illustrated in Figure 3. We propose that each layer of
DBM determines a submanifold M of S, where M could maximally preserve the highly
confident information on parameters, as shown in Section 4.3.3. Then the whole DBM
can be seen as the process of repeatedly applying CIF in each layer, achieving the tradeoff
between the abstractness of representation features and the intrinsic information confidence
preserved on parameters.
Once a good representation has been found at each level by layer-wise application of
unsupervised greedy pre-training, it can be used to initialize and train the deep neural
networks through supervised learning (Hinton and Salakhutdinov, 2006; Erhan et al., 2010).
Recall that the straightforward application of gradient-based methods to train all layers of a
DBM simultaneously tends to fall into a poor local minima (Younes, 1998; Desjardins et al.,
2012). Now, our next question is: why can the layer-wise pre-training give a more reasonable
parameter initialisation? Empirically, Erhan et al. (2010) shows that the unsupervised pretraining acts as a regularization on parameters in a way that the parameters are set into a
region, from which better basins of attraction can be reached. Theoretically, by using the
fractional mixed coordinates, it can be shown that this regularized region is actually the
layer-wise restriction using CIF, i.e., the highly confident coordinates are preserved with
18
respect to the target density and all lowly confident coordinates are set to a neutral value of
zero, as illustrated in Figure 3. Effectively, the parameter space is regularized to fall into a
region where the parameters can be confidently estimated based on the given data. Under
this CIF-based regularization, the pre-training can be seen as searching for a reasonable
parameter setting, from which a good representation of the input data can be generated in
each layer.
5. Experimental Study
In this section, we will empirically investigate the CIF principle in density estimation tasks
on two types of Boltzmann machines, i.e., SBM and RBM. More specifically, for SBM, we
will investigate how to use CIF to take effect on the learning trajectory with respect to the
specific sample, and hence further confine the parameter space to the region corresponding to
the most confident information contained in given data. For RBM, it is inconvenient to use
a sample-specific strategy since the information of hidden variables is missed. Alternatively
we investigate the potential of the iterative projection procedure proposed previously. For
both SBM and RBM, two baseline learning methods, i.e., the contractive divergence (CD)
(Hinton, 2002; Carreira-Perpinan and Hinton, 2005) and maximum-likelihood (ML) (Ackley
et al., 1985), are adopted.
The ML learning is described in Section 4.1.2. The CD can be seen as an approximation
of ML. Let q(x) be the underlying probability distribution from which sample x are generated independently. Then our goal is to train a BM (with stationary probability p(x, h; ξ)
in Equation 13) based on x that realizes q(x) as faithfully as possible. Comparing to ML,
the CD learning realizes the gradient descend of a different objective function to avoid the
difficulty of computing the log-likelihood gradient in ML, shown as follows:
∆ξ = −ε ·
∂(D(q0 , p) − D(pm , p))
∂E(x, h; ξ)
∂E(x, h; ξ)
= ε · (−h
i0 + h
im )
∂ξ
∂ξ
∂ξ
(25)
where q0 is the sample distribution, pm is the distribution by starting the Markov chain
with the data and running m steps, h·i0 and h·im denote the averages with respect to the
distribution p0 and pm , and D(·, ·) denotes the K-L divergence.
In these experiments, two kinds of binary datasets are used:
1. The artificial binary dataset: we first randomly select the target distribution q(x),
which is chosen uniformly from the open probability simplex over the n random variables. Then, the dataset with N samples are generated from q(x).
2. The 20 News Groups binary dataset: 20 News Groups is a collection of approximately
20,000 newsgroup documents, partitioned evenly across 20 different newsgroups 9 . The
collection is preprocessed using porter stemmer and stop-word removal. We select top
100 terms with highest frequency in the collection. Each document is represented as
a 100-dimensional binary vector, where each element indicates whether certain term
occurs in current document or not.
9. The 20 News Group dataset is freely downloadable from http://qwone.com/∼jason/20Newsgroups/
19
5.1 Experiments with SBM
From the perspective of IG, we can see that ML/CD learning is to update parameters in
SBM so that its corresponding coordinates [η 2− ] are getting closer to the data distribution.
This is consistent with our theoretical analysis in Section 3 and Section 4.2 that SBM uses
the most confident information (i.e., [η 2− ]) for approximating an arbitrary distribution in
an expected sense. But, for the distribution with specific samples, can CIF further recognize
less-confident parameters in SBM and reduce them properly?
Our solution here is to apply CIF to take effect on the learning trajectory with respect to
specific samples, and hence further confine the parameter space to the region that indicated
by the most confident information contained in the samples. This experiment shows that
given specific samples we need to preserve the confident parameters to certain extend, and
there should exist some golden ratio that would produce best performance on average.
5.1.1 A Sample-specific CIF-based CD Learning
The main modification of our CIF-based CD algorithm (CD-CIF for short) is that we
generate the samples for pm (x) based on those parameters with confident information,
where the confident information carried by certain parameter is inherited from the sample
and could be assessed using its Fisher information computed in terms of the sample.
For CD-1 (i.e., m=1), the firing probability for the ith neuron after one-step transition
(0) (0)
(0)
from the initial state x(0) = {x1 , x2 , . . . , xn }) is:
(1)
p(m=1) (xi
1
= 1|x(0) ) =
1 + exp{−
(0)
j6=i Uij xj
P
(26)
− bi }
For CD-CIF, the firing probability in Equation 26 is modified as follows:
1
(1)
p0(m=1) (xi = 1|x(0) ) =
P
(0)
1 + exp{− (j6=i)&(F (Uij )>τ ) Uij xj − bi }
(27)
where τ is a pre-selected threshold, F (Uij ) = Eq0 [xi xj ]−Eq0 [xi xj ]2 is the Fisher information
of Uij (see Equation 8) and the expectations are estimated from the given sample x. We
can see that those weights whose Fisher information are less than τ are considered to be
unreliable w.r.t. x. In practice, we could setup τ by the ratio r to specify the remaining
proportion of the total Fisher information TF I of all parameters, i.e., τ = r ∗ TF I .
In summary, CD-CIF is realized in two phases. In the first phase, we initially “guess”
whether certain parameter could be faithfully estimated based on the finite sample. In the
second phase, we approximate the gradient using the CD scheme, except for the CIF-based
firing function is used.
5.1.2 Results and Discussions on Artificial Dataset
In this section, we empirically investigate our justifications for the CIF principle, especially
how the sample-specific CIF-based CD learning works in the context of density estimation.
Experimental Setup and Evaluation Metric: For computation simplicity, the artificial dataset is set to be 10-dimensional. Three learning algorithms are investigated: ML,
CD-1 and our CD-CIF. K-L divergence is used to evaluate the goodness-of-fit of the SBM
20
Density Estimation Performance for CD−CIF
Performance Change with varying amount of Confident Information (1200)
3
Performance Change with varying amount of Confident Information (100)
0.55
0.4
0.35
0.3
0.25
0.2
0
200
400
600
800
Number of samples N
(a)
1000
1200
CD−1
ML
CD−CIF
0.52
K−L Divergence (Background vs SBMs)
0.45
K−L Divergence (Background vs SBMs)
K−L Divergence (Background vs SBMs)
0.54
CD−1
ML
CD−CIF
0.5
0.5
0.48
0.46
0.44
0.42
0.4
0.38
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Learning Trajectory using 400 samples
0.36
True Distribution
0.35
0.34
2
0.33
0.14
1.5
0.135
1
0.13
0.94
0.32
0.96
0.98
1
0.31
0.5
0
0.1
1
CD−1
ML
CD−CIF
2.5
Fixed Point
of CD−CIF
Fixed Point of CD−1
Fixed Point of ML
Training Start Point
0.3
0.2
0.3
0.4
0.5
0.6
r
r
(b)
(c)
0.7
0.8
0.9
1
0.29
0.93
0.94
0.95
0.96
0.97
0.98
0.99
(d)
Figure 4: (a): the performances of CD-CIF on different sample sizes; (b) and (c): the
performances of CD-CIF with various values of r on two typical sample sizes, i.e.,
100 and 1200; (d): learning trajectories of last 100 steps for ML (squares), CD-1
(triangles) and CD-CIF (circles).
trained by various algorithms. For sample size N , we run 100 instances (20 randomly generated distributions × 5 randomly running) and report the averaged K-L divergences. Note
that we focus on the case that the variable number is relatively small (n = 10) in order to
analytically evaluate the K-L divergence and give a detailed study on algorithms. Changing the number of variables only offers a trivial influence for experimental results since we
obtained qualitatively similar observations on various variable numbers (not reported here).
Automatically Adjusting r for Different Sample Sizes: The Fisher information
is additive for i.i.d. sampling. When sample size N changes, it is naturally to require that
the total amount of Fisher information contained in all tailored parameters is steady. Hence
we have α = (1 − r)N , where α indicates the amount of Fisher information and becomes a
constant when the learning model and the underlying distribution family are given. It turns
out that we can first identify α using the optimal r w.r.t. several distributions generated
from the underlying distribution family, and then determine the optimal r for various sample
sizes using: r = 1 − α/N . In our experiments, we set α = 35.
Density Estimation Performance: The averaged K-L divergences between SBM
(learned by ML, CD-1 and CD-CIF with the r automatically determined) and the underlying
distribution are shown in Figure 4(a). In the case of relatively small samples (N ≤ 500)
in Figure 4(a), our CD-CIF method shows significant improvements over ML (from 10.3%
to 16.0%) and CD-1(from 11.0% to 21.0%). This is because we could not expect to have
reliable identifications for all model parameters from insufficient samples, and hence CD-CIF
gains its advantages by using parameters that could be confidently estimated. This result is
consistent with our previous theoretical insight that Fisher information gives a reasonable
guidance for parametric reduction via the confidence criterion. As the sample size increases
(N ≥ 600), CD-CIF, ML and CD-1 tend to have similar performances. Since with relatively
large samples most model parameters can be reasonably estimated, and hence the effect of
parameter reduction using CIF gradually becomes marginal. In Figure 4(b), Figure 4(c), we
show how sample size affects the interval of r that achieves improvements over CD-1. For
N = 100, CD-CIF achieves significantly better performances for a wide range of r. While,
for N = 1200, CD-CIF can only marginally outperform baselines for a narrow range of r.
Effects on Learning Trajectories: We use the 2D visualizing technology SNE to
investigate learning trajectories and dynamical behaviors of three comparative algorithms
21
Density estimation performance on 20 News Groups
22
21.5
Hamming distance
21
20.5
CD−1
ML
CD−CIF
20
19.5
19
18.5
18
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
r
Figure 5: the performances of CD-CIF with various values of r on 20 news group dataset
(Carreira-Perpinan and Hinton, 2005). We start three methods with the same parameter initialization. Then each intermediate state is represented by a 55-dimensional vector
formed by its current parameter values. From Figure 4(d), we can see that: 1) In the final
100 steps, three methods seem to end up with staying in different regions of the parameter
space, and CD-CIF confines the parameter in a relatively thinner region compared to ML
and CD-1; 2) The true distribution is usually located on the side of CD-CIF, indicating its
potential for converging to the optimal solution. Note that the above claims are based on
general observations and Figure 4(d) is shown as an illustration. Hence we may conclude
that CD-CIF regularizes the learning trajectories in a desired region of the parameter space
using the sample-specific CIF principle.
5.1.3 Results and Discussions on Real Textual Dataset
In this section, we empirically investigate how the sample-specific CIF-based CD learning
works on real-world datasets in the context of density estimation. In particular, we use
the SBM to learn the underlying probability density over 100 terms of the 20 News Groups
binary dataset. The learning rate for CD-1, ML and CD-CIF are manually tuned in order to
converge properly and all set to 0.001. Since it is infeasible to compute the K-L devergence
due to the high dimensionality, the averaged Hamming distance between the samples in the
dataset and those generated from the SBM is used to evaluate the goodness-of-fit of the
SBM’s trained by various algorithms. Let D = {d1 , d2 , . . . , dN } denote the dataset of N
documents, where each document di is a 100-dimensional binary vector. To evaluate a SBM
with parameter ξsbm , we first randomly generate N samples from the stationary distribution
p(x; ξsbm ), denoted as V = {v1 , v2 , . . . , vN }. Then the averaged hamming distance Dham is
calculated as follows:
P
di (minvj (Ham[di , vj ])
Dham [D, V ] =
N
where Ham[di , vj ] is the number of positions at which the corresponding values are different.
22
The result is shown in Figure 5. Our CD-CIF method shows maximal improvements
over ML (12.15%) and CD-1(15.05%) at r = 0.92. We can also see that CD-CIF achieves
significantly better performances for a wide range of r ∈ [0.5, 0.96], which is consistent with
our observations with the experiments on artificial datasets when the samples is insufficient.
5.2 Experiments with RBM
The RBM is practically more interesting than SBM, since it has higher representational
power. In this section, we will compare three different learning algorithms for RBM: CD-1,
ML and IP. In Carreira-Perpinan and Hinton (2005), CD is shown to be biased with respect
to ML for almost all data distributions. In Section 4.3.4, we have compared ML and IP
theoretically. In this section, an empirical study on the three algorithms is conducted.
5.2.1 Results and Discussion on Artificial Datasets
Experimental Setup and Evaluation Metric: For computational simplicity, the artificial dataset is of 5 dimensionality, and the number of hidden units in RBM is set to
5. Three learning algorithms are investigated: ML, CD-1 and IP. K-L divergence is used
to evaluate the goodness-of-fit of the RBM’s trained by various algorithms. Six different
sample sizes N are tested, namely 50, 100, 200, 500, 5000 and 50000. For sample size N ,
the learning rates for CD-1 and ML are set to be ε = 0.5/N , and we observe that they
could converge properly. For the sub-learning phase ΓB of IP, we adopt the CD algorithm
for the training of BM without hidden units, whose learning rate is also set to ε = 0.5/N .
We need to scan the dataset multiple times in order to iteratively update parameters, and
each full scan of the whole dataset is called an epoch. In CD-1 and ML, we set the maximal
number of epoches to 8000. We run the IP for 40 iterations, and each iteration is a CD
sub-training with the maximal number of epoches setting to 200. We adopt CD-1 as the
baseline method.
Results and Analysis: The on-average performances of the three methods on dataset
of different scales are shown in Table 1 in the context of density estimation. In order to study
the behavior of IP, we plot the sequences of K-L divergence between target distribution and
RBM in each iteration along the whole learning trajectory, shown in Figure 6. Comparing
CD-1 with ML, we can see that the K-L divergences of both CD-1 and ML decrease in a
similar way, converging at the same rate, taking the same number of iterations to converge
to a given tolerance, which is consistent with the conclusion in Hinton (2002).
From the convergence behavior of IP shown in Figure 6, we can see that the general trend
is that the K-L divergence decreases steadily with small fluctuations. Since there are only
few iterations in IP, it is reasonable for us to select the best performance that IP has reached
in the whole learning process, called the best IP. Here the K-L divergence between the sample
distribution and RBM’s stationary distribution is adopted as the selection metric. Thus, in
addition to the converging performance, we also show the best performance selected among
all 40 iterations for IP in Table 1. Note that the best performances for CD-1 and ML are
not reported, since their converging performances are often approximately the best ones.
For small sample size (e.g., 50 and 100), we can see that the converging performance of
IP is comparable with respect to CD and ML. As the sample size increases, we can see that
IP gradually outperforms CD-1 and ML, and shows significant improvement on large sample
23
sample size = 50000
sample size = 100
1.4
1.4
ML
CD−1
IP
1.2
1.2
1
K−L divergence
1
K−L divergence
ML
CD−1
IP
0.8
0.6
epoch=200
epoch=2000
0.8
0.6
0.4
0.4
epoch=200
epoch=4000
epoch=4000
0.2
0.2
0
epoch=2000
0
0.5
1
1.5
2
2.5
3
3.5
0
4
0
0.5
1
1.5
2
Iteration 10x
Iteration 10x
(a)
(b)
2.5
3
3.5
4
Figure 6: (a) and (b) illustrate the averaged learning curves for CD-1, ML and IP for 30
randomly chosen data distributions, with sample size 100 and 50000 respectively.
The x-axis is in log10 scale. To compare the efficiency of different algorithms
along the time-line (the time unit is an epoch, that is, the time used for the
parameter updates in each full scan of the whole dataset), we plot three time
stamps (epoch=100, 1000 and 4000). Note that each iteration in IP contains a
sub-training task, which is trained using 200 epoch in this experiment.
size (e.g., 5000 and 50000). For the best IP, its performance is significantly better than CD1 and ML for all sample sizes, indicating that IP has the potential to further improve its
performance by using some suitable selection metric. Comparing to ML, IP takes much
shorter iterations as expected to achieve a performance threshold. Theoretically, IP can
converge to the local minimum of RBM based on our theoretical analysis in Section 4.3.4.
If a proper learning rate is selected, ML can also converge to a local minimum. However,
one interesting result is that sometimes there is a big difference between the convergence
points of IP and ML, even in cases that the sampling is sufficient (e.g., sample size equals
50000).
This can be explained as follows. In practice, ML needs to constantly do positive
and negative sampling in each updating, which may produce much sample biases. As the
gradient decreasing to a small value, the correct gradient direction may be fluctuated by
the sample biases. Thus, instead of converging to the local minimum, ML might fluctuate
around some sub-optimal region. Actually, the ΓH in IP is a sampling process that may also
introduce sample biases. That is why the IP is fluctuated when the sample is insufficient, as
shown in Figure 6(a). Given sufficient samples, the sampling biases in ΓH decreases and the
fluctuation of IP declines, as shown in Figure 6(b). For ML, though the sampling biases for
the whole dataset becomes small given sufficient samples, the gradient estimation for each
sample is still closely intergraded with the sample bias of certain sample, meaning that
this inseparable coupling relationship results in a biased gradient estimation. The main
advantage of the IP procedure over traditional gradient-based methods is the separation
24
Sample Size
50
100
200
500
5000
50000
CD-1
0.0970
0.0786
0.0672
0.0621
0.0532
0.0497
ML
0.0971
0.0793
0.0678
0.0621
0.0524
0.0475
Converge IP (-chg%)
0.0978 (+0.81%)
0.0787 (+0.08%)
0.0640 (-4.75%)
0.0594 (-4.25%)
0.0468 (-12.00%)
0.0411 (-17.37%)
Best IP (-chg%)
0.0774 (-20.21%)
0.0637 (-18.92%)
0.0575 (-14.45%)
0.0567 (-8.61%)
0.0437 (-17.81%)
0.0332 (-33.08%)
Table 1: Performance comparison of CD-1, ML and IP in the density estimation task. The
change of IP with respect to the baseline method (CD-1) is reported.
Density estimation performance for 20 News Groups
18
Hamming distance
17
CD−1
ML
IP
16
15
14
13
12
10
20
30
40
50
60
70
Number of hidden units
80
90
100
Figure 7: Performances of IP with various number of hidden units on 20 News Group
of the positive sampling process and the gradient estimation. We conjecture that this
independent design gives IP the potential to achieve optimal solutions that CD-1 and ML
can not reach.
5.2.2 Results and Discussion on Real Textual Dataset
In this section, we empirically investigate how the IP works for RBM on real-world datasets
in the context of density estimation. We use the RBM to learn the probability density
over 100 terms on the 20 News Groups binary dataset. The number of hidden units nh in
RBM is set to [10, 20, . . . , 100]. The learning rate for CD-1and ML are manually tuned in
order to converge properly and all set to 0.001. The learning rate for the CD sub-learning
task in IP is also set to 0.001. We run IP for 9 iterations for all settings of nh . Similar
with the experiments in Section 5.1.3, the averaged Hamming distance is used to evaluate
the goodness-of-fit of the RBM’s trained by various algorithms. The average Hamming
distances for ML, CD-1 and IP are shown in Figure 7. We can see that IP achieves better
25
performances on all settings of nh . And the hamming distance for IP drops dramatically
as nh increases. This trend can be explained as follows. As nh grows, the sampling biases
increases and the interference of sampling biases with respect to the gradient estimation
becomes more and more serious. This limits the actual performance of RBMs learnt by
CD-1 and ML, with respect to the growing modelling power gained by increasing nh . As
shown in Section 5.2.1, the IP procedure separates the positive sampling process and the
gradient estimation in two phases: ΓH and ΓB . This result shows that IP has the potential
to achieve optimal solutions that CD-1 and ML can not reach in real-world applications.
6. Conclusions
The CIF principle proposed in this paper tackles the problem of dimensionality reduction in
parameter space by preserving the parameters with highly confident estimates and tailoring the less confident parameters. It provides a strategy for the derivation of probabilistic
models. The SBM and RBM are specific examples in this regard. They have been theoretically shown to achieve a reliable representation in parameter spaces by exactly using the
general CIF principle. CIF gives us a principled and context-independent way to address
the questions on what we should do for parameter reduction (regularization) and how to do
it. Based on CIF, we also show that the deep neural networks consisting of several layers of
RBM can be seen as the layer-wise usage of CIF, leading to some theoretical interpretations
of the rationale behind deep learning models.
One interesting result shown in our experiments is that: although CD-CIF is a biased algorithm, it could significantly outperform ML when the sample is insufficient. This suggests
that CIF gives us a reasonable criterion for recognizing and utilizing confident information
from the underlying data while ML fails to do so. Another interesting observation is that
ML and the CIF-based IP lead to different convergence points in the training of RBM.
Our experimental results indicate that IP has the advantage of robustness against sampling
biases, due to the separation of the positive sampling process and the gradient estimation.
In the future, we will further develop the formal justification of CIF w.r.t various contexts
(e.g., distribution families or models). We will also conduct more extensive experiments on
real world applications, such as document classification and handwritten digit recognition,
to further justify the properties of IP. We will also extend the IP to train deep neural
networks.
Acknowledgments
This work is partially supported by the Chinese National Program on Key Basic Research
Project (973 Program, grant no. 2013CB329304 and 2014CB744604), the Natural Science
Foundation of China (grants no. 61070044, 61111130190, 61272265 and 61105072), and
the European Union Framework 7 Marie-Curie International Research Staff Exchange Programme (grant no. 247590).
26
Appendix A.
A.1 Proof of Proposition 1
Proof By definition, we have:
gIJ =
∂ 2 ψ(θ)
∂θI ∂θJ
where ψ(θ) is defined by Equation (4). Hence, we have:
gIJ
P
∂ 2 ( I θI ηI − φ(η))
∂ηI
= J
=
∂θI ∂θJ
∂θ
By differentiating ηI , defined by Equation (1), with respect to θJ , we have:
gIJ
P
P
∂ x XI (x)(exp{ I θI XI (x) − ψ(θ)})
∂ηI
=
=
∂θJ
∂θJ
X
=
XI (x)[XJ (x) − ηJ ]p(x; θ) = ηI S J − ηI ηJ
x
This completes the proof.
A.2 Proof of Proposition 2
Proof By definition, we have:
g IJ =
∂ 2 φ(η)
∂ηI ∂ηJ
where φ(η) is defined by Equation (4). Hence we have:
g
IJ
=
P
∂ 2 ( J θJ ηJ − ψ(θ))
∂θI
=
∂ηI ∂ηJ
∂ηJ
Based on Equation (2) and (1), the θI and pK could be calculated by solving a linear
equation system of [p] and [η] respectively. Hence we have:
θI =
X
X
(−1)|I−K| log(pK ); pK =
K⊆I
(−1)|J−K| ηJ
K⊆J
Therefore, the partial derivation of θI with respect to ηJ is:
g IJ =
∂θI
∂ηJ
=
X ∂θI ∂pK
X
1
(−1)|I−K|+|J−K| ·
·
=
∂pK ∂ηJ
pK
K
K⊆I∩J
This completes the proof.
27
A.3 Proof of Proposition 3
Proof
The
Fisher information matrix of [ζ] could be partitioned into four parts: Gζ =
A C
. It can be verified that in the mixed coordinate, the θ-coordinate of order k
D B
is orthogonal to any η-coordinate less than k-order, impling the corresponding element of
Fisher information matrix is zero (C = D = 0) (Nakahara and Amari, 2002). Hence, Gζ is
a block diagonal matrix.
According to Cramér-Rao bound (Rao, 1945), a parameter (or a pair of parameters)
has a unique asymptotically tight lower bound of the variance (or covariance) of unbiased
estimate, which is given by the corresponding element of the inverse of Fisher information
matrix involving this parameter (or this pair of parameters). Recall that Iη is the index
set of the parameters shared by [η] and [ζ]l and Jθ is the index set of the parameters
−1
−1
−1
−1
=
shared by [θ] and [ζ]l , we have (G−1
ζ )Iζ = (Gη )Iη and (Gζ )Jζ = (Gθ )Jθ , i.e., Gζ
−1
(Gη )Iη
0
. Since Gζ is a block tridiagonal matrix, the proposition follows.
0
(G−1
θ )Jθ
A.4 Proof of Proposition 4
U X
, which is parXT V
titioned based on Iη and Jθ . Based on Proposition 3, we have A = U −1 . Obviously, the
diagonal elements of U are all smaller than one. According to the succeeding Lemma 12,
we can see that the diagonal elements of A (i.e., U −1 ) are greater than 1.
Next we need to show that the diagonal elements of B are smaller than 1. Using the
−1
Schur complement of Gθ , the bottom-right block of G−1
θ , i.e., (Gθ )Jθ , equals to (V −
X T U −1 X)−1 . Thus the diagonal elements of B: Bjj = (V − X T U −1 X)jj < Vjj < 1. Hence
we complete the proof.
Proof Assume the Fisher information matrix of [θ] be Gθ =
Lemma 12 With a l × l positive definite matrix H, if Hii < 1, then (H −1 )ii > 1, ∀i ∈
{1, 2, . . . , l}.
Proof Since H is positive definite, it is a Gramian matrix of l linearly independent vectors
v1 , v2 , . . . , vl , i.e., Hij = hvi , vj i (h·, ·i denotes the inner product). Similarly, H −1 is the
Gramian matrix of l linearly independent vectors w1 , w2 , . . . , wl and (H −1 )ij = hwi , wj i.
It is easy to verify
that hwi , vi i = 1, ∀i ∈ {1, 2, . . . , l}. If Hii < 1, we can see that the
√
norm kvi k = Hii < 1. Since kwi k × kvi k ≥ hwi , vi i = 1, we have kwi k > 1. Hence,
(H −1 )ii = hwi , wi i = kwi k2 > 1.
A.5 Proof of Proposition 5
Proof Let Bq be a ε-ball surface centered at q(x) on manifold S, i.e., Bq = {ζ ∈ S|kζ −
ζq k2 = ε}, where k · k2 denotes the Euclid norm and ζq is the coordinates of q(x). Let
q(x) + dq be a neighbor of q(x) uniformly sampled on Bq and ζq(x)+dq be its corresponding
coordinates. For a small ε, we can calculate the expected information distance between
q(x) and q(x) + dq as follows:
Z
1
EBq = [(ζq(x)+dq − ζq )T Gζ (ζq(x)+dq − ζq )] 2 dBq
(28)
28
where Gζ is the Fisher information matrix at q(x).
Since Fisher information matrix Gζ is both positive definite and symmetric, there exists
a singular value decomposition Gζ = U T ΛU where U is an orthogonal matrix and Λ is a
diagonal matrix with diagonal entries equal to the eigenvalues of Gζ (all ≥ 0).
Apply the singular value decomposition into Equation (28), the distance becomes:
Z
1
EBq = [(ζq(x)+dq − ζq )T U T ΛU (ζq(x)+dq − ζq )] 2 dBq
(29)
Note that U is an orthogonal matrix, and the transformation U (ζq(x)+dq − ζq ) is a normpreserving rotation.
Now we need to show that among all tailored k-dimensional submanifolds of S, [ζ]lt is
the one that preserves maximum information distance. Assume IT = {i1 , i2 , . . . , ik } is the
index of k coordinates that we choose to form the tailored submanifold T in the mixedcoordinates [ζ]. Based on Equation (29), the expected information distance EBq for T is
proportional to the sum of eigenvalues of the sub-matrix (Gζ )IT , where the sum equals to
the trace of (Gζ )IT .
Next we show that the sub-matrix of Gζ specified by [ζ]lt gives maximum trace. Based
on Proposition 4, the elements on the main diagonal of the sub-matrix A are lower bounded
by one, and those of B upper bounded by one. Therefore, [ζ]lt gives maximum trace among
all sub-matrices of Gζ . This completes the proof.
A.6 Proof of Proposition 6
Proof Let Msbm be the set of all probability distributions realized by SBM. Amari
et al. (1992) proves that the mixed-coordinates of the resulting projection P on Msbm
2 , 0, . . . , 0), given the 2-mixed-coordinates of q(x). M
is [ζ]P = (ηi1 , ηij
sbm is equivalent to the
submanifold tailored by CIF, i.e. [ζ]2t . The corollary follows from Proposition 5.
A.7 Proof of Proposition 7
Proof Based on Equation 15, the coordinates [θ2+ ] for SBM is zero: θ2+ = 0. Next, we
2 ] with q(x).
show that the stationary distribution p(x; ξ) learnt by ML has the same [ηi1 , ηij
For SBM, the
∂E(x;ξ)
∂ξ
can be easily calculated from Equation (12):
∂E(x;ξ) = xi xj ,
∂Ux x
f or Uxi xj ∈ ξ;
∂E(x;ξ)
∂bxi
f or bxi ∈ ξ.
i j
= xi ,
Thus, based on Equation 17, the gradients for Uxi xj , bxi ∈ ξ are as follows:
∂ log p(x;ξ) = hxi xj i0 − hxi xj i∞ = η 2 (q(x)) − η 2 (p(x; ξ)), f or Uxi xj ∈ ξ;
ij
ij
∂Ux ,x
i
j
∂ log p(x;ξ) = hxi i0 − hxi i∞ = ηi1 (q(x)) − ηi1 (p(x; ξ)),
∂bx
i
f or bxi ∈ ξ.
where h·i0 denotes the average using the sample data and h·i∞ denotes the average with
respect to the stationary distribution p(x; ξ).
29
Since SBM defines an e-flat submanifold Msbm of S (Amari et al., 1992), then ML
converges to the unique solution that gives the best approximation p(x; ξ) ∈ Msbm of q(x).
p(x;ξ)
When ML converges, we have ∆ξ → 0 and hence ∂ log∂ξ
→ 0. Thus, we can see that ML
2 ] of q(x). This
converges to stationary distribution p(x; ξ) that preserves coordinates [ηi1 , ηij
completes the proof.
A.8 Proof of Proposition 8
Proof Based on the definition of divergence in Equation (11), the following relation holds:
D[q(x, h), p(x, h)] = D[q(x)q(h|x), p(x)p(h|x)]
q(x)
q(h|x)
= Eq(x,h) [log
+ log
]
p(x)
p(h|x)
= D[q(x), p(x)] + Eq(x) [D[q(h|x), p(h|x)]]
where Eq(x,h) [·] and Eq(x) [·] are the expectations taken over q(x, h) and q(x) respectively.
Therefore, the minimum divergence between p(x, h; ξp ) and Hq is given as:
D(Hq , p(x, h; ξp )) =
min
q(x,h;ξq )∈Hq
D[q(x, h; ξq ), p(x, h; ξp )]
= min{D[q(x), p(x)] + Eq (x)[D[q(h|x; ξq ), p(h|x; ξp )]]}
ξq
= D[q(x), p(x)] + min{Eq(x) [D[q(h|x; ξq ), p(h|x; ξp )]]}
ξq
= D[q(x), p(x)]
In the last equality, the expected divergence between q(h|x; ξq ) and p(h|x; ξp ) vanishes if
and only if ξq = ξp . This completes the proof.10
A.9 Proof of Proposition 9
Proof First, we show that [ζ xh ] is determined given [θ]. Since there is a one-to-one
correspondence between coordinates [θ] and [p], [ζ xh ] can be directly calculated from the
p-coordinates corresponding to [θ] based on Equation (1) and (2).
hh
xx
Second, [θ] is determined by knowing [ζ xh ]. The {θ2 i j , θ2 i j , θ2+ } part of [θ] are set to be
xx
hh
h
xh
equal to those in [ζ xh ]. By fixing {θ2 i j , θ2 i j , θ2+ } and setting {θ1xi , θ1 j , θ2 i j } free, we now
have an e-flat smooth submanifold R. Assume that there exist two different distributions
P1 and P2 with coordinates [θ]1 and [θ]2 that have the same mixed coordinates [ζ xh ]. Thus
both P1 and P2 belong to R and share the same value of {ηx1i , ηh1j , ηx2i hj }. Let Q ∈ Sxh be a
distribution whose projection on R is P1 . Based on the Projection Theorem in Amari and
Nagaoka (1993), P1 is the unique closest point on R to Q. Considering the minimization of
the divergence D[Q, PR ] between PR ∈ R and Q, the gradient vector of D[Q, PR ] over the
h
xh
free parameters {θ1xi , θ1 j , θ2 i j } at P1 , that is {ηx1i , ηh1j , ηx2i hj }PR −{ηx1i , ηh1j , ηx2i hj }Q , equals to
zero vector. Then, P2 also has a zero-gradient vector and hence is the projection point of Q,
since P2 has the same {ηx1i , ηh1j , ηx2i hj } with P1 . However, since R is e-flat 11 , the projection
10. Note that a similar path of proof is also used in Theorem 7 of Amari et al. (1992), which is for the
fully-connected BM. Here, Here, we reformulate the proof for RBM to derive the projection ΓH (p(x, h)).
11. For more information about the concept of flatness, please refer to the book Amari and Nagaoka (1993).
30
of Q on R is unique, meaning that P1 and P2 are the same point. Therefore, there does
not exist two different distributions P1 and P2 that have the same mixed coordinates [ζ xh ].
This completes the proof.
A.10 Proof of Proposition 10
Proof First, we prove the uniqueness of the projection ΓB (q). From the [θ] of RBM in
Equation (20), B is an e-flat smooth submanifold of Sxh . Thus the projection is unique.
Note that
Second, in order to find the p(x, h; ξp ) ∈ B with parameter ξp that minimizes the
divergence between q(x, h; ξq ) and B, the gradient descent method iteratively adjusts ξp
in the negative gradient direction that the divergence D[q, p(ξp )] decreases fastest:
4ξp = −λ
∂D[q, p(ξp )]
∂ξp
where D[q, p(ξp )] is treated as a function of RBM’s parameters ξp and λ is the learning rate.
As shown in Albizuri et al. (1995), the gradient descent method converges to the minimum
of the divergence with proper choices of λ, and hence achieves the projection point ΓB (q).
Last, we show that the fractional mixed coordinates [ζ xh ]ΓB (q) in Equation (23) is exactly
the convergence point of the learning for RBM . We calculate the first-order derivative of
D[q, p(ξp )], w.r.t ξp , where p(x, h; ξp ) = Z1 exp{−E(x, h; ξp )} is given in Equation (13).
For Wxi ,hj in ξp (denoted as Wij ), we have:
X q(x, h) ∂p(x, h; ξp )
∂D[q, p(ξp )]
=−
∂Wij
p(x, h; ξp )
∂Wij
(30)
x,h
where the
∂p(x,h)
∂Wij
is calculated as follows:
∂p(x, h; ξp )
∂Wij
= Z −1 exp{−E(x, h)}{
∂(−E(x, h)) X
∂(−E(x, h))
−
p(x, h)
}
∂Wij
∂Wij
x,h
∂(−E(x, h))
∂(−E(x, h))
− p(x, h)
p(x, h) ·
∂Wij
∂Wij
x,h
X
= p(x, h) · xi hj − p(x, h) ·
p(x, h)xi hj
X
= p(x, h) ·
(31)
x,h
Combining Equation (30) and (31), we have:
X
X
∂D[q, p(ξp )]
=−
q(x, h)xi hj +
p(x, h)xi hj = ηx2i hj (p) − ηx2i hj (q)
∂Wij
x,h
(32)
x,h
where ηx2i hj (p) and ηx2i hj (q) denotes the 2nd -order η-coordinates of p and q respectively.
Similarly, the first-order derivatives for biases bxi and dhj can be proved to be:
∂D[q, p(ξp )]
= ηx1i (p) − ηx1i (q)
∂bxi
31
(33)
∂D[q, p(ξp )]
= ηh1j (p) − ηh1j (q)
∂dhj
(34)
Summarizing Equation (32), (33) and (34), the first-order derivatives of ξpI , where the
indexing I = {xi } or {hj } or {xi , hj } (∀ xi ∈ x, hj ∈ h), can be calculated in the same
way, that is subtracting p’s and q’s corresponding η-coordinates ηI :
∂D[q, p(ξp )]
= ηI (p) − ηI (q)
∂ξpI
When converging, we have
∂D[q,p(ξp )]
∂ξpI
→ 0. Hence, the gradient descent method con-
verges to the projection point ΓB (q) with a stationary distribution p(x, h; ξp ) that preserves
coordinates [ηx1i , ηh1j , ηx2i hj ] of q(x, h; ξq ). This completes the proof.
A.11 Proof of Proposition 11
Proof Since pi ∈ B and pi+1 ∈ B is the projection of qi+1 , then D[qi+1 , pi ] ≥ D[qi+1 , pi+1 ].
Similarly, qi+1 ∈ Hq and qi+2 ∈ Hq is the projection of pi+1 , thus D[qi+1 , pi+1 ] ≥ D[qi+2 , pi+1 ].
This completes the proof.
References
D. H. Ackley, G. E. Hinton, and T. J. Sejnowski. A learning algorithm for Boltzmann
machines. Cognitive Science, 9:147–169, 1985.
F. X. Albizuri, A. d’Anjou, M. Grana, J. Torrealdea, and M. C. Hernandez. The highorder boltzmann machine: learned distribution and topology. Neural Networks, IEEE
Transactions on, 6(3):767–770, 1995.
S. Amari and H. Nagaoka. Methods of Information Geometry. Translations of Mathematical
Monographs. Oxford University Press, 1993.
S. Amari, K. Kurata, and H. Nagaoka. Information geometry of boltzmann machines. IEEE
Transactions on Neural Networks, 3(2):260–271, 1992.
Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle. Greedy layer-wise training of deep
networks. In NIPS’06, pages 153–160, Vancouver, British Columbia, Canada, 2006.
Y. Bengio, A. C. Courville, and P. Vincent. Representation learning: A review and new
perspectives. IEEE Trans. Pattern Anal. Mach. Intell., 35(8):1798–1828, 2013.
M. A. Carreira-Perpinan and G. E. Hinton. On contrastive divergence learning. Artificial
Intelligence and Statistics, pages 17–24, 2005.
N. N. Chentsov. Statistical Decision Rules and Optimal Inference. Translations of mathematical monographs, 53:477–493, 1982.
R. Collobert and J. Weston. A unified architecture for natural language processing: Deep
neural networks with multitask learning. In ICML’08, pages 160–167, 2008.
32
Y. Dauphin and Y. Bengio.
abs/1301.3583, 2013.
Big neural networks waste capacity.
arXiv CoRR,
G. Desjardins, A. C. Courville, and Y. Bengio. On training deep boltzmann machines.
CoRR, abs/1203.4416, 2012.
R. W. Duin and E. Peȩkalska. Object representation, sample size, and data set complexity.
In Data Complexity in Pattern Recognition, pages 25–58. Springer London, 2006.
D. Erhan, Y. Bengio, A. Courville, P-A. Manzagol, P. Vincent, and S. Bengio. Why does
unsupervised pre-training help deep learning? Journal of Machine Learning Research,
11:625–660, 2010.
I. Fodor. A survey of dimension reduction techniques. Technical report, Center for Applied
Scientic Computing, Lawrence Livermore National Laboratory, United States, 2002.
B. R. Frieden. Science from Fisher Information: A Unification. Cambridge University
Press, 2004.
W. R. Gilks, S. Richardson, and D.J. Spiegelhalter. Markov Chain Monte Carlo in Practice.
Chapman and Hall/CRC, 1996.
G. E. Hinton. Training products of experts by minimizing contrastive divergence. Neural
Comput., 14(8):1771–1800, 2002.
G. E. Hinton and R. R. Salakhutdinov. Reducing the dimensionality of data with neural
networks. Science, 313(5786):504–507, 2006.
Y. Hou, X. Zhao, D. Song, and W. Li. Mining pure high-order word associations via
information geometry for information retrieval. ACM TOIS, 31(3), 2013.
I. T. Jolliffe. Principal component analysis. Springer Series in Statistics, New York, US,
2002.
R. E. Kass. The geometry of asymptotic inference. Statistical Science, 4(3):188–219, 1989.
John A. Lee and Michel Verleysen, editors. Nonlinear Dimensionality Reduction. Springer,
New York, US, 2007.
H. Nakahara and S. Amari. Information geometric measure for neural spikes. Neural
Computation, 14:2269–2316, 2002.
S. Osindero and G. E. Hinton. Modeling image patches with a directed hierarchy of markov
random field. In NIPS’07, pages 1121–1128, 2007.
M. Ranzato, C. Poultney, S. Chopra, and Y. LeCun. Efficient learning of sparse representations with an energy-based model. In NIPS’06, pages 1137–1144, 2006.
C. R. Rao. Information and the accuracy attainable in the estimation of statistical parameters. Bulletin of calcutta mathematics society, 37:81–89, 1945.
33
S. Rifai, P. Vincent, X. Muller, X. Glorot, and Y. Bengio. Contractive auto-encoders:
Explicit invariance during feature extraction. In ICML’11, pages 833–840, 2011.
R. Salakhutdinov and G. E. Hinton. Using deep belief nets to learn covariance kernels for
gaussian processes. In NIPS’07, pages 1249–1256, 2007a.
R. Salakhutdinov and G. E. Hinton. Semantic hashing. In Workshop SIGIR’07, 2007b.
R. Salakhutdinov and G. E. Hinton. An efficient learning procedure for deep boltzmann
machines. Neural Computing, 24(8):1967–2006, 2012.
P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, and P-A. Manzagol. Stacked denoising
autoencoders: Learning useful representations in a deep network with a local denoising
criterion. J. Mach. Learn. Res., 11:3371–3408, 2010.
J. A. Wheeler. Time today. In Physical Origins of Time Asymmetry, pages 1–29. Cambridge
University Press, 1994.
L. Younes. On the convergence of markovian stochastic algorithms with rapidly decreasing
ergodicity rates. In Stochastics and Stochastics Models, pages 177–228, 1998.
34
| 9cs.NE
|
1
Resource Optimization with Load Coupling
in Multi-cell NOMA
Lei You∗ , Di Yuan∗ , Lei Lei† , Sumei Sun‡ , Symeon Chatzinotas† , and Björn Ottersten†
arXiv:1708.05281v3 [cs.IT] 7 Jan 2018
∗ Department of Information Technology, Uppsala University, Sweden
{lei.you; di.yuan}@it.uu.se
† Interdisciplinary Centre for Security, Reliability and Trust, Luxembourg University, Luxembourg
{lei.lei; symeon.chatzinotas; bjorn.ottersten}@uni.lu
‡ Institute for Infocomm Research, A*STAR, Singapore
[email protected]
Abstract—Optimizing non-orthogonal multiple access (NOMA)
in multi-cell scenarios is much more challenging than the singlecell case because inter-cell interference must be considered. Most
papers addressing NOMA consider a single cell. We take a
significant step of analyzing NOMA in multi-cell scenarios. We
explore the potential of NOMA networks in achieving optimal
resource utilization with arbitrary topologies. Towards this goal,
we investigate a broad class of problems consisting in optimizing
power allocation and user pairing for any cost function that is
monotonically increasing in time-frequency resource consumption. We propose an algorithm that achieves global optimality for
this problem class. The basic idea is to prove that solving the joint
optimization problem of power allocation, user pair selection,
and time-frequency resource allocation amounts to solving a
so-called iterated function without a closed form. We prove
that the algorithm approaches optimality with fast convergence.
Numerically, we evaluate and demonstrate the performance of
NOMA for multi-cell scenarios in terms of resource efficiency
and load balancing.
Index Terms—NOMA, multi-cell, resource allocation
I. I NTRODUCTION
N
ON-ORTHOGONAL multiple access (NOMA) is considered as a promising technique for enhancing resource
efficiency [2]–[13]. In two recent surveys [2], [3], the authors
pointed out that resource allocation in multi-cell NOMA
poses much more research challenges compared to the singlecell case, because optimizing NOMA with multiple cells
has to model the interplay between successive interference
cancellation (SIC) and inter-cell interference. As one step
forward, the investigations in [2], [3] have addressed two-cell
scenarios. In [6], the authors proposed two coordinated NOMA
beamforming methods for two-cell scenarios. Reference [8]
uses stochastic geometry to model the inter-cell interference in
NOMA. Hence the results do not apply for analyzing network
with specific given network topology. Reference [9] optimizes
energy efficiency in multi-cell NOMA with downlink power
control. However, the aspect of determining which users share
resource by SIC, i.e., user pairing, is not considered. To the
best of our knowledge, finding optimal power allocation and
user pairing simultaneously for enhancing network resource
efficiency in multi-cell NOMA without restrictions on network
topology has not been addressed yet.
Part of this paper has been presented at IEEE GLOBECOM, Singapore,
Dec. 2017 [1].
The crucial aspect of multi-cell NOMA consists of capturing the mutual interference among cells; This is a key
consideration in SIC of NOMA. Therefore, the cells cannot
be optimized independently. For orthogonal multiple access
(OMA) networks, a modeling approach had been proposed
that characterizes the inter-cell interference via capturing the
mutual influence among the cells’ resource allocations [14]–
[36]. The model, named load-coupling, refers to the timefrequency resource consumption in each cell is referred to
as the cell load. However, the model does not allow SIC. In
our recent work [1], we addressed resource optimization in
multi-cell NOMA. However, the system model is constrained
by fixed power allocation. How to model joint optimization
of power allocation and user pairing and how to solve the
resulting problems to optimality have remained open so far.
II. M AIN R ESULTS
We demonstrate how NOMA can be modeled in multi-cell
scenarios by significantly extending the approaches in [14]–
[36], with joint optimization of power allocation and user
pairing. One fundamental result under such type of models
in OMA is the existence of the equilibrium for resource allocation. However, this modeling approach in NOMA leads to
non-closed form formulation for cell load coupling, unlike the
case of OMA. The fact poses significant challenges in analysis
and problem solving. As one of our main results, we prove that
such an equilibrium for resource allocation in NOMA exists
as well and propose an efficient algorithm for obtaining the
equilibrium. Furthermore, we prove that the equilibrium is the
global optimum for resource optimization in multi-cell NOMA
and thus a wide class of resource optimization problems can be
optimally solved by our algorithm. Because of our analytical
results, previous works about OMA with load coupling is
a special case of ours, namely, the algorithmic notions and
mathematical tools being used in those works of classic multicell power control or OMA load coupling thus directly apply
to the analysis multi-cell NOMA, suggesting future works on
this topic.
To the best of our knowledge, this is the first work investigating how to optimally utilize power and time-frequency
resources jointly in multi-cell NOMA. As a key strength of
our modeling approach, it enables to formulate and optimize
an entire class of resource optimization problems. Namely,
2
III. C ELL L OAD M ODELING
Denote by C = {1, 2, . . . , n} and J = {1, 2, . . . , m} the sets
of cells and user equipments (UEs), respectively. Denote by
Ji the set of UEs served by cell i (i ∈ C). When using j to
refer to one UE in J, i by default indicates j’s serving cell,
unless stated otherwise. Downlink is considered in our model.
A. Resource Sharing in NOMA
The resource in time-frequency domain is divided into resource blocks (RBs). In OMA, one RB can be accessed by only
one UE. In NOMA, multiple UEs can be clustered together to
access the same RB by SIC. Increasing the number of UEs in
SIC, however, leads to fast growing decoding complexity [2],
[3]. In previous works, it has been demonstrated that most
of the possible performance improvement by SIC is reached
by pairing as few as two UEs [2]–[7]. Pairing two UEs for
resource sharing is illustrated in Figure 1. UEs within one
pair share the same RB and the RBs allocated to different
pairs do not overlap. We use u as a generic notation for a user
pair (referred to as “pair” for simplicity). For cell i (i ∈ C),
denote by Ui the set of candidate pairs. Suppose
there are in
total mi UEs in cell i. Then |Ui | is up to m2i . Denote
by Vj
S
(j ∈ J) the set of pairs
containing
UE
j.
Let
U
=
U
i∈C i (or
S
equivalently U = j∈J Vj ) be the set of candidate pairs of all
cells. Let s = |U|. If there is a need to differentiate between
pairs, we put indices on u, i.e., U = {u1 , u2 , u3 , . . . us }. Finally,
in our model, for UEs we allow for both OMA and NOMA
RB
b
9
b
b
4
1
10
b
b
8
5
b
7
b
b
b
b
3
Frequency
as long as the cost function is monotonically increasing in
the cells’ time-frequency resource consumption, our proposed
framework in multi-cell NOMA applies. Specifically, for solving this class of problems optimally, we derive a polynomialtime algorithm S-C ELL that gives the optimal power allocation
and user pairing, for any given input of inter-cell interference.
To address the dynamic coupling of inter-cell interference, we
derive a unified algorithmic framework M-C ELL that solves
the multi-cell resource optimization problems optimally. The
algorithm S-C ELL serves as a sub-routine and is iteratively
called by M-C ELL. We demonstrate theoretically the linear
convergence of this process.
The fundamental differences between our investigated problems and single-cell NOMA are summarized as follows. For
multiple cells, the resource allocation in one cell affects the
interference that the cell generates to other cells. The amounts
of required resource to meet the demand for all cells are
coupled together, rather than being independent to each other.
Optimizing resource allocation within one cell leads to a chain
reaction among all other cells. Individual optimization for the
cells results in sub-optimality and very inaccurate performance
analysis. For multi-cell NOMA, not only the time-frequency
resource allocation but also the power splits and user pairings
in all cells are coupled together for the same reason. Therefore,
joint optimization in NOMA leads to a rather complex problem
for analysis.
By numerical experiments, optimizing resource utilization
by our algorithm enlightens how much we can gain from
NOMA in terms of resource efficiency and load balancing.
2
6
Time
User Pairing
Resource Allocation
Figure 1. This figure illustrates user pairing and time-frequency resource
sharing. There are 10 UEs in one cell. Eight form four user pairs {1, 5},
{2, 7}, {3, 6}, {8, 9}, and the other two UEs 4 and 10 are unpaired. The
UEs within one pair share the same time-frequency resource as indicated by
the colors.
with SIC. For each UE, that which mode is used (or both can
be used) is determined by optimization. In the following, we
refer to these two modes as orthogonal RB allocation and nonorthogonal RB allocation, respectively. In general NOMA, we
include both modes.
B. NOMA Downlink
We first consider orthogonal RB allocation in NOMA. Let
pi be the transmission power per RB in cell i (i ∈ C). Denote
by gij the channel coefficient from cell i to UE j. The signalto-interference-and-noise ratio (SINR) is:
pi gij
.
(1)
γj = P
2
k∈C\{i} Ikj + σ
The term Ikj denotes the inter-cell interference from cell k to
UE j, and is possibly zero. This generic notation is used for the
sake of presentation. Later, we use the load-coupling model,
where the cell load that reflects the usage of RBs governs the
amount of interference. The term σ2 is the noise power.
We then consider non-orthogonal RB allocation in NOMA.
In [37] (Chapter 6.2.2, pp. 238) it is shown that, with superposition coding, one UE of pair u (u ∈ U) can decode the other
by SIC. When there is need to consider the decoding order
in u, to be intuitive, we use ⊕ to denote the UE that applies
interference cancellation, followed by decoding its own signal.
And denotes the UE that only decodes its own signal. Note
that both ⊕ and are generic notations and refer to the two
different users in any pair u (u ∈ U) in consideration. For any
pair u, pi is divided to q⊕u and q u (q⊕u + q u = pi ), with
q⊕u and q u being allocated to ⊕ and , respectively. (The
generic notation qju (j ∈ u) denotes the power allocated to UE
j.) We remark that ⊕ decodes ’s signal first and hence ’s
signal does not compose the interference for ⊕. The SINR of
⊕ is computed by (2).
q⊕u gi⊕
.
(2)
γ⊕u = P
2
k∈C\{i} Ik⊕ + σ
The UE
is subject to intra-cell interference from ⊕, i.e.,
q u gi
X
γ u=
.
(3)
q⊕u gi +
Ik +σ2
| {z }
k∈C\{i}
intra-cell
| {z }
inter-cell
3
Denote by q the power allocation of all candidate pairs:
q
q⊕u2 · · · q⊕us
q = ⊕u1
.
q u1 q u2 · · · q us
We use qu to represent the column of pair u (u ∈ U) in q,
named power split for u. We remark that it is not necessary to
use all the pairs in U for resource sharing. Whether or not a
pair would be put in use and allocated with RBs is determined
by optimization, discussed later in Section III-E. In addition,
we remark that the decoding order is not constrained by the
power split [37], even though by our numerical results, more
power is always allocated to in optimal solutions. The issue
of the influence of inter-cell interference on the decoding order
is addressed later in Section III-D.
C. Inter-cell Interference Modeling
The basic idea is to use the cells’ RB consumption levels to
characterize respectively the cell’s likelihood of interfering to
the others. The approach is specified as follows. Denote by ρk
the proportion of RBs allocated for serving UEs in cell k. The
intuition behind the model is partially explained by the two
extreme cases ρk = 1 and ρk = 0. If cell k is fully loaded,
meaning that all RBs are allocated, then ρk = 1. In the other
extreme case, cell k is idle and accordingly ρk = 0. Consider
any UE j served by cell i. The interference j receives from
cell k is Ikj = pk gkj or Ikj = 0 in the two cases, respectively.
In general, ρk serves as a scaling parameter for interference,
see (4). By the interference modeling approach, the cell load
directly translates to the scaling effect of interference and
therefore the same notation is used for both.
Ikj = pk gkj ρk .
(4)
Intuitively, ρk reflects the likelihood that a UE outside cell
k receives interference from k. Note that ρk in fact is the
amount of time-frequency resource consumption of cell k and
hence is referred to as the load of cell k.
We remark that this type of interference modeling approach
is a suitable approximation for network-level performance
analysis, which enables study of inter-cell interference in
large-scale multi-cell networks without having to modeling
micro-level interference. Detailed system-level simulations
(e.g. [24] and [32]) have shown that this type of modeling has
sufficient accuracy for cell-level interference characterization.
This approach has been widely used and is getting increasingly
popular [14]–[36], which however, to our best knowledge, are
limited to OMA. We provide analytical results in order to
extend the modeling approach to NOMA.
for h, followed by decoding its own signal, and, user h does
not apply SIC. That is, the decoding order ⊕ = j and = h
always hold for pair independent of interference.
Proof. Denote by γhj and γhh respectively in (5) and (6) the
SINRs at users j and h for the downlink signal of h.
q g
P hu ij
.
(5)
γhj =
qju gij + k∈C\{i} pk gkj ρk + σ2
γhh =
qju gih +
Lemma 1. For u = {j, h} (gij > gih ) in cell i, if gij /gih >
gkj /gkh (k ∈ C\{i}), then SIC at j decodes first the signal
qhu gih
.
2
k∈C\{i} pk gkh ρk + σ
(6)
With superposition coding, j cancels the interference from h if
j can decode any data that h can decode [37], i.e. γhj > γhh ,
which reads:
X
qju gij gih + gij
pk gkh ρk + gij σ2
k∈C\{i}
> qju gij gih + gih
X
pk gkj ρk + gih σ2 .
k∈C\{i}
Further, γhj > γhh if and only if:
X pk ρk
(gih gkj − gij gkh ) 6 (gij − gih ).
σ2
(7)
k∈C\{i}
Recall that gij > gih , and therefore the right-hand side of (7)
is non-negative. Because of the condition gij /gih > gkj /gkh
for all k ∈ C\{i} in the statement of the lemma, the left-hand
side is non-positive. Hence Lemma 1.
The result of Lemma 1 is coherent with the previous
observations that two UEs with large difference in channel
conditions are preferred to be paired [4], [7]. If gij gih ,
then most likely the condition in Lemma 1 holds, as the large
scale path-loss from other cells, tends not to differ as much
as from the serving cell i in this case. Besides, the large scale
path-loss is a practically reasonable factor for ranking the
decoding order [38], [39]. In Section VII, numerical results
further show that considering the UE pairs as defined by
Lemma 1 virtually does not lead to any loss of performance.
Lemma 1 is used to filter the candidate pairs set U (i.e. to drop
some candidate pairs from U) so as to reduce computational
complexity. From now on, we let Ui be composed of pairs
satisfying Lemma 1.
E. RB Allocation
If UE j (j ∈ J) is using orthogonal RB allocation, then the
achievable capacity1 of j is (8), with γj being (1).
cj = log(1 + γj ).
D. Determining Decoding Order
Inter-cell interference affects the decoding order in NOMA,
and thus how to model the load-coupling in NOMA is significantly more challenging than OMA. Lemma 1 below resolves
this issue by identifying pairs for which the decoding order
can be determined independently of interference. As another
benefit, it significantly reduces the set of candidate pairs.
P
(8)
For non-orthogonal RB allocation, the achievable capacity
for j and u (j ∈ u) is computed by cju = MB log (1 + γju )
with γju being (2) or (3). Therefore,
cju =
log (1 + γju )
0
j∈u
.
j∈
/u
(9)
1 For the sake of presentation, we use the natural logarithm throughout the
paper. We remark that all conclusions hold for the logarithm to base 2.
4
For UE j (j ∈ J), we use xj to denote the proportion of RBs
with orthogonal RB allocation to j. For any pair u (u ∈ U),
denote by xu the non-orthogonal RB allocation for the two UEs
in pair u. We use the vector x to represent the RB allocation
for all the UEs, i.e.,
x = [ x1 , x2 , . . . , xm ,
|
{z
}
xu , xu , . . . , xus ].
| 1 2{z
}
one can observe that increasing xu (or xj ) for some pair u
(or some UE j) may enhance the throughput of the UEs of u
(or UE j). However, since xu (or xj ) appears in the inter-cell
interference term (see (4) and (10)), the increase of xu (or xj )
results in less available resources for other UEs and leads to
more interference. The user pairing selection is not given a
priori but is determined by optimization.
Orthogonal RB allocation Non-orthogonal RB allocation
For any UE j, xj = 0 means that UE j does not use orthogonal
RB allocation. Similarly, for any pair u, xu = 0 means that
pair u is not put in use. For any UE j, if xu = 0 for all
u ∈ Vj , then it means that UE j only uses orthogonal RB
allocation. Resources used by different pairs are orthogonal
such that there is no interference among pairs. Denote by ρ̄
the cell load limit. By constraining that the sum of them which
equals to the load of cell i does not exceed ρ̄, the amounts
represented by xj (j ∈ Ji ) and xu (u ∈ Ui ) do not overlap.
Orthogonal RB allocation is considered among the pairs in
one cell, meaning that the pairs do not have interference with
each other.
X
X
xu 6 ρ̄.
(10)
ρi =
xj +
↑
↑
u∈Ui
j∈Ji
| {z }
| {z }
Cell
Load
load
Orthogonal
RB proportion
Non-orthogonal
RB proportion
ρ = [ρ1 , ρ2 , . . . , ρn ].
Similarly, we use vector ρ̄ to denote the load limits of all cells.
The term cj xj computes the bits delivered to UE j with
orthogonal RB allocation, because cj is the achievable capacity
of UE j on all RBs and xj is the proportion of RBs with
orthogonal RB allocation. Similarly, the term cju xu is the bits
delivered to UE j by non-orthogonal RB allocation for pair u.
Denote by dj the bits demand of UE j. The quality-of-service
(QoS) requirement is:
X
cj xj +
cju xu > dj .
(11)
|{z}
u∈V
↑
j
Bits delivered
| {z }
by orthogonal
Bits
Bits delivered by
non-orthogonal
RB allocation
The models proposed for OMA in [14]–[20] are inherently
a special case of our NOMA model. The former is obtained by
setting U = φ. Then, the terms for non-orthogonal RB allocation disappear in (10) and (11) and x is therefore eliminated
in (8)–(11). Also, there is no power split in OMA. Hence (8)–
(11) form a non-linear system only in terms of ρ. This system
falls into the analytical framework of standard interference
function (SIF) [40], which enables the computation of the
optimal network load settings via fixed-point iterations [17].
However, for the general NOMA case, the resource allocation
is not at UE-level. One needs to split a UE’s demand between
orthogonal and non-orthogonal RB allocations, which results
in a new dimension of complexity.
limit
We use ρ to represent the vector of network load, i.e.,
RB allocation
F. Comparison to OMA Modeling
IV. P ROBLEM F ORMULATION
By successively plugging (1) and (4) into (8), we obtain a
function cj in load ρ, i.e., cj (ρ). Similarly, we obtain cju (q, ρ)
from (2), (3), (4), and (9). For pair u (u ∈ U), we use a binary
variable yu to indicate whether or not the pair u is selected.
Define y as
y = [yu1 , yu2 , . . . , yus ].
We minimize a generic cost function F(ρ) that is monotonically (but not necessarily strictly monotonically) increasing in
each element of ρ. M IN F is given below.
[M IN F]
s.t.
min F(ρ)
ρi 6 ρ̄, i ∈ C
X
X
ρi =
xj +
xu , i ∈ C
demand
We remark that dj is normalized by the RB spectral bandwidth
and the total number of RBs, for the sake of presentation. Note
that a user can use orthogonal RB allocation individually, or
non-orthogonal RB allocation with the other user in the pair, or
both, which is subject to optimization. The amount of allocated
RBs to a user in OMA or a pair adopting NOMA, is subject
to optimization, under the constraint that the overall allocated
resource does not exceed limit.
We remark that there is an implicit pair selection problem
in the above expressions. Note that |U| increases fast with |J|.
It is therefore impractical to simultaneously use all pairs in U.
To deal with this issue, each UE is allowed to use up to one
pair in U for optimization, as formulated later in Section IV,
though our system model is not limited by this. The problem
of pairing and resource allocation is challenging: First, UEs
of the same pair are coupled in resource allocation. Second,
(12a)
ρ,q,x,y
j∈Ji
cj (ρ)xj +
X
X
(12b)
(12c)
u∈Ui
cju (q, ρ)xu > dj , j ∈ J
(12d)
u∈Vj
qju = pi , u ∈ Ui , i ∈ C
(12e)
j∈u
xu 6 yu , u ∈ U
X
yu 6 1, j ∈ J
(12f)
(12g)
u∈Vj
ρ, q, x > 0
(12h)
yu ∈ {0, 1}, u ∈ U
(12i)
Constraints (12b) guarantee that the cell load complies
to the load limit ρ̄. Constraints (12c) state the relationship
between RB allocation and cell load. Constraints (12d) and
(12e) are for QoS and power, respectively. Constraints (12f)
guarantee that RB allocation occurs only for selected pairs.
By constraints (12g), each UE belongs up to one pair such
5
that the selected pairs are mutually exclusive. The variables
are cell load ρ, power allocation q, RB allocation x, and
pair selection y. The variable domains are imposed by (12h)
and (12i). Throughout this paper, we use 0 to represent zero
vector/matrix. For simplicity, the dimension(s) of 0 is not
explicitly stated.
V. O PTIMIZATION WITHIN A C ELL
In multi-cell NOMA, due to the interference among cells,
one cell’s pair selection may affect the other cells’ power
splits, and vice versa. Let us consider a simple case in this
section. Suppose we optimize the load of one cell i, and the
cell load levels of C\{i} are temporarily fixed. This optimization step is a module for solving M IN F later in Section VI.
We respectively use qi , xi , yi to denote the corresponding
variable elements for power allocation, RB allocation, and pair
selection. Vector ρ−i is composed of all elements but ρi of
ρ. We minimize ρi under fixed ρ−i , as formulated below.
min
ρi ,qi ,xi ,yi
ρi s.t. (12c)–(12i) of cell i, with fixed ρ−i . (13)
Since ρ−i is fixed, cj is a constant and cju is a function in qi
only.
The optimization is not straightforward even under fixed
inter-cell interference. The optimal power split for one pair
is up to how much time-frequency resource is allocated to
this pair. In other words, for one pair u, if the amount of
RBs allocated to u changes, the optimal power split for u
before this change loses its optimality. So the power split q
and the resource allocation x are coupled together. In addition,
the pair selection is a combinatorial problem. Therefore, the
power split q, the time-frequency resource allocation x, and
the user pair selection y, must be optimized jointly.
Lemma 2. All constraints of (12d) in (13) hold as equalities
at any optimum.
Proof. Denote the optimal objective value of (13) by ρi0 and
the optimal orthogonal RB allocation of j (j ∈ Ji ) by xj0 .
Suppose strict inequality holds for some j. If xj0 > 0, by fixing
all other variables except for xj in (13), one can verify that
the solution xj0 − ( > 0) is still feasible to (13) as long
as is sufficiently small. In addition, xj0 − leads to a lower
objective value ρi0 − , which contradicts our assumption that
ρi0 is the optimal objective value. If xj0 = 0, then j’s demand
has to be satisfied by non-orthogonal RB allocation and the
same argument applies to variable xu (j ∈ u).
The first analytical result is that, the optimal power split is
independent of pair selection, in Theorem 1. Denote by Yu the
set of all possible pairing solutions of (13) that includes pair
u, i.e.,
X
Yu = {yi |yu = 1,
yu 6 1, j ∈ Ji }.
u∈Vj
^ i (y
^ i ∈ Yu ), the optimal
Definition 1. Given a pair selection y
^ u , is the column
power split for pair u (u ∈ Ui ), denoted by q
^ i , where q
^ i is obtained by optimally solving
for pair u in q
^ i.
(13) for y
Theorem 1. Consider u (u ∈ Ui ). The optimal power split
^ i (y
^ i ∈ Yu ) is also optimal for y
^ i0 (y
^ i0 ∈ Yu ).
for any y
^ u and q
^ u0 the optimal power splits for
Proof. Denote by q
0
^ i and y
^ i , respectively. Suppose q
^ u is not optimal for y
^ i0 .
y
0
^ u ) = cju (q
^ u ) (j ∈ u); 2)
There are two possibilities: 1) cju (q
^ u ) 6= cju (q
^ u0 ) for at least one j in u.
cju (q
^ u and q
^ u0 result in the same xj and xu for
For 1), q
satisfying (12d) and are equally good for (13), which conflicts
^ u is optimal for y
^ i0 . We then consider
our assumption. Thus q
0
^ u ) > cju (q
^ u ). By Lemma 2, q
^ u0 makes
2) and assume cju (q
0
0
^ i . Replacing q
^ u by q
^ u leads to
(12d) become equality under y
some slack in (12d) and hence the objective can be improved.
^ u0 is optimal for yi0 . The same proof
This contradicts that q
^ u ) < cju (q
^ u0 ). Hence the conclusion.
applies to cju (q
By Theorem 1, the optimal power split is decoupled from
pair selection. Next we analytically prove how to find the
optimal power split for any pair.
A. Finding Optimal Power Split
Under fixed yi (yi ∈ Yu , u ∈ Ui ), constraints (12g)
are removed. Constraints (12f), and (12i) of (13) for all u
with yu = 0 in yi are removed. Therefore, for each pair
u = {⊕, }, we can formulate a problem in (14). Solving this
problem yields the optimal power split. In (14), x⊕ and x
are the orthogonal RB allocation for ⊕ and , respectively.
The variable xu denotes the amount of non-orthogonal RB
allocation for u.
x⊕ + x + xu
(14a)
c⊕ (ρ−i )x⊕ + c⊕u (qu , ρ−i )xu > d⊕
(14b)
min
x⊕ ,x ,xu >0
qu >0
s.t.
c (ρ−i )x + c
q⊕u + q
u
u (qu , ρ−i )xu
>d
= pi
(14c)
(14d)
For deriving solution method for (14), define function wj
(j = ⊕ or j = ) of ρ−i as follows.
,
X
wj (ρ−i ) =
pk gkj ρk + σ2 gij .
(15)
k∈C\{i}
For q⊕u , one can derive from (2) and (9):
q⊕u = (ec⊕u − 1)w⊕ (ρ−i ).
(16)
Combining (16) with (14d), q⊕u and q u can be eliminated,
giving (17) below. Formulation (17) is equivalent to (14).
Given c⊕u and c u , the corresponding q⊕u and q u can be
obtained from c⊕u and c u by (14d) and (16).
min
x⊕ ,x ,xu >0
c⊕u ,c u >0
x⊕ + x + xu
s.t. c⊕ (ρ−i )x⊕ + c⊕u xu > d⊕
c (ρ−i )x + c
Cvu (c⊕u , c
u xu > d
u , ρ−i )
60
(17a)
(17b)
(17c)
(17d)
6
In (17), the function Cvu is defined in (32) in the Appendix.
One can easily verify that Cvu is convex in c⊕u and c u (with
gi⊕ > gi ). The difficulty of (17) is on the two bi-linear
constraints (17b) and (17c). However, they become linear with
fixed xu . To ease the presentation, we define the function
below.
Zu (xu , ρ−i ) = xu +
min
x⊕ ,x >0
c⊕u ,c u >0
x⊕ + x
c⊖u
K
K= [cK
⊕u , c⊖u ]
s.t. (17b)–(17d). (18)
Solving (17) (and equivalently (14)) is to find the minimum
of Zu (xu , ρ−i ). The following theorem shows the uniqueness
of the minimum of Zu (xu , ρ−i ).
[0, 0]
Theorem 2. Zu (xu , ρ−i ) has unique minimum in xu .
Proof. Since the first term xu in Zu (xu , ρ−i ) is strictly monotonically increasing, to prove that Zu (xu , ρ−i ) has unique
minimum, we only need to prove that the remaining part
of Zu (xu , ρ−i ) is monotonically (but not necessarily strictly
monotonically) decreasing in xu . For this part, at the optimum (17b) and (17c) hold as equalities because of Lemma 2.
Hence, reformulating the problem by replacing the inequalities
in (17b) and (17c) with equalities does not lose optimality.
With equalities, the variables x⊕ and x can be represented
by c⊕u and c u :
(d − c u xu )
(d⊕ − c⊕u xu )
, x =
.
x⊕ =
c⊕ (ρ−i )
c (ρ−i )
(19)
Therefore x⊕ and x can be eliminated from the objective
function. The minimization is thus equivalent to maximizing
c⊕u /c⊕ (ρ−i ) + c u /c (ρ−i ). We formulate this maximization problem below.
c⊕u
c u
max
+
(20a)
c⊕u ,c u >0 c⊕ (ρ−i )
c (ρ−i )
s.t. c⊕u xu 6 d⊕
(20b)
c
u xu
6d
Cvu (c⊕u , c
(20c)
u , ρ−i )
60
(20d)
Constraints (20b) and (20c) originate from the nonnegativity requirement of x⊕ and x . Note that (20) is convex.
In addition, the feasible region shrinks with the increase of xu .
Then the optimum of (20) monotonically decreases with xu .
Hence the theorem.
Because of Theorem 2, bi-section search of xu reaches the
minimum of Zu (xu , ρ−i ). Note that for any xu , computing
Zu (xu , ρ−i ) needs to solve (20). In the following, we prove
how this can be done much more efficiently than employing
standard convex optimization.
We remark that (20) is a two-dimensional optimization
problem with respect to c⊕u and c u . Constraints (20b)
and (20c) are defined by two hyperplanes c⊕u = d⊕ /xu and
c u = d /xu , respectively. Due to the convexity of the function Cvu in c⊕u and c u , the curve Cvu (c⊕u , c u , ρ−i ) = 0
in (20d) along with c⊕u = 0 and c u = 0 forms a convex
set of [c⊕u , c u ]. The optimum of (20) depends on whether
the two hyperplanes intersect with the curve and how they
intersect. This leads to three possible cases to be considered,
h
1
1
c⊕ (ρ−i ) , c⊖ (ρ−i )
i
c⊕u
Figure 2. This figure shows the feasible region of (20) of c⊕u and c u
for different xu . The shadowed area below the curve is (20d). The vertical
and horizontal dashed lines are the hyperplanes defined by (20b) and (20c),
respectively. The hyperplanes are shown by red, blue, and green dashed lines
for Case 1, Case 2, and Case 3, respectively. The point K is optimal for Case
1. The blue and green circles represent the optimal solutions for Case 2 and
Case 3, respectively. The black dot is the point on the curve where the two
hyperplanes intersect with each other (see footnote 3 for the existence of this
point.).
named Case 1, Case 2, and Case 3, respectively. In Case 1,
constraints (20b) and (20c) are redundant, and the optimum is
determined by the coefficients 1/c⊕ (ρ−i ) and 1/c (ρ−i ) in
the objective function and the curve Cvu (c⊕u , c u , ρ−i ) = 0.
In Case 2, the optimum is defined by one of the hyperplanes
and the curve. In Case 3, constraint (20d) is redundant, and
the optimum is determined by the two hyperplanes. With xu
increasing from 0 to ∞, Case 1, Case 2, and Case 3 happen
sequentially, and all happen eventually. The three cases are
illustrated in Figure 2 with colors. Below we respectively show
how to compute the optimum for each case.
In the following, we first compute the optimum in Case 1,
K
represented by K = [cK
⊕u , c u ], which is the intersection of the
vector [1/c⊕ (ρ−i ), 1/c (ρ−i )] and the curve. The point also
leads to a closed-form solution for the optima of all the three
cases. Mathematically, point K is solved by applying bi-section
search to (21) below.
Cvu (c⊕u , c u , ρ−i ) = 0
c⊕u c⊕ (ρ−i ) = c u c (ρ−i ).
(21)
For solving (21), we can first eliminate c⊕u or c u in the first
equation. This can be done by representing one of c⊕u and
c u with the other by the second equation. Since there is only
one variable in the first equation, one can use bi-section search
to find its solution2 . Then, we compute the value of xu when
at least one hyperplane goes through K, denoted by xK
u in (22).
K
K
xK
u = min{d⊕ /c⊕u , d /c u }.
(22)
The three cases, indicated by colors in Figure 2, are as follows.
2 The solution is guaranteed to be unique and hence bi-section search
applies. This is because, by representing one of c⊕u and c u by the other by
function Cvu , one variable is monotonically decreasing in the other, resulting
in a unique zero point.
7
Case 1 (xu 6 minj∈u dj /cK
ju ): Point K is the optimum of
(20), because (20b) and (20c) are redundant. This happens
when xu is sufficiently small (or 0), as shown in Figure 2.
Case 2 (xu > minj∈u dj /cK
ju and Cvu (d⊕ /xu , d /xu ) > 0):
There exists one point on the curve where both two hyperplanes intersect3 . We represent this point by the black dot
on the curve in Figure 2. In Case 2, one hyperplane intersects
with the curve at some point between K and the black dot, and
intersects with the other hyperplane on some point above the
curve, see Figure 2. The intersection point of the curve and the
hyperplane is the optimum of (20). Without loss of generality,
we assume K violates (20c), meaning that the hyperplane of
(20c) goes through the optimum, as shown by Figure 2. By
plugging the equation of the hyperplane into that of the curve,
the optimal c⊕u is a function of xu . Similarly, if K violates
(20c) instead, then c⊕u = d⊕ /xu and the optimal c u is
a function of xu . To know which hyperplane goes through
K
the optimum, one only needs to check which of cK
ju xju > dj
(j = ⊕ or j = ) holds. Note that exactly one of the two holds
in Case 2. The optimal c⊕u and c u are computed respectively
by (33) and (34) defined in the Appendix.
Case 3 (xu > minj∈u dj /cK
ju and Cvu (d⊕ /xu , d /xu ) 6 0):
Constraint (20d) is redundant, as shown in Figure 2. The
optimum is the intersection point of the two hyperplanes,
computed by cju = dj /xu (j = ⊕ or j = ).
In summary, the optimal solution of (20) is computed by
(23) below (j = ⊕ or j = ) in closed form, with Hju
being (33) or (34) in the Appendix.
K
Case 1
cju
Cju (xu , ρ−i ) =
(23)
H (x , ρ ) Case 2
ju u −i
dj /xu
Case 3
The function Zu (xu , ρ−i ) computes the amount of resource
used for both orthogonal and non-orthogonal RB allocations
for the UEs in u. It is optimal to serve the two UEs only
by orthogonal RB allocation, if the minimum of Zu (xu , ρ−i )
occurs at xu = 0. In all other cases, minxu Zu (xu , ρ−i ) yields
the optimal power split for non-orthogonal RB allocations. The
algorithm optimally solving (14), named S PLIT, is as follows.
S PLIT(u, ρ−i )
1 x∗u = arg minxu Zu (xu , ρ−i ) // Bi-section search
2 Compute hc∗⊕u , c∗ u i by (23) // With x∗u , ρ−i
3 Compute hx∗⊕ , x∗ i by (19) // With c∗⊕u , c∗ u , ρ−i
4 Convert hc∗⊕u , c∗ u i to hq∗⊕u , q∗ u i // By (14d), (16)
5 return hq∗⊕u , q∗ u , x∗⊕ , x∗ , x∗u i
B. Optimal Pairing
Denote by Yi the set of all candidate pair selections:
Yi = (∪u∈Ui Yu ) ∪ {0}.
3 The existence of this point is guaranteed: With the increase of x , both
u
hyperplanes will eventually intersect with the curve with two intersection
points. By increasing xu , the distance between the two intersections keeps
being smaller. The two intersections will eventually overlap.
0.18
1
0.02
∆
0.14
0.14
0.16 3
2
0.07
0.06
0.40
4
5
0.12
Figure 3. The figure shows an example of one cell i with five UEs,
i.e., Ji = {1, 2, 3, 4, 5}. Assume the candidate pair set is Ui =
{{1, 2}, {2, 4}, {4, 5}, {3, 5}}. The blue edges are type-1. The red edges
are type-2. A matching is a set of edges without common vertices (also called
independent edge set) and is a pair selection solution. The maximum matching,
as highlighted in the figure, is {{1, 2}, {3, 5}, {4, ∆}}.
By obtaining minxu Zu (xu , ρ−i ) for all u ∈ Ui as shown
earlier in Section V-A, enumerating all yi in Yi gives the
optimal solution to (13). This exhaustive search however does
not scale, as |Yi | is exponential in the number of UEs. By
the following derivation, we are able to obtain the optimum
of (13) in polynomial time.
Theorem 3. The optimum of (13) is computed by finding the
maximum weighted matching in an undirected graph.
Proof. To prove the conclusion, an undirected weighted graph
Gi is constructed and explained below.
Gi =
hJi , Ui , wi
hJi ∪ {∆}, Ui ∪ {{j, ∆}|j ∈ Ji }, wi
|Ji | even
|Ji | odd.
(24)
In (24), the graph is represented by a 3-tuple, with the first
element being the vertex set, the second element being the
edge set, and the third element being the weight vector.
Parameter ∆ is an auxiliary vertex for odd |Ji |. Without loss
of generality, below we focus on odd |Ji |. (All conclusions
naturally hold for |Ji | being even.) By the definition in (24),
each UE is corresponding to a vertex. For each pair u in Ui ,
there is one edge connecting the two UEs in u, associated
with weight wu . We name these as type-1 edges. Besides, for
each UE j in Ji , there is one extra edge connecting j and the
auxiliary vertex ∆, associated with weight wj . We name these
as type-2 edges. An illustration is given in Figure 3.
The weight w is defined as follows, where T is a positive
value keeping all weights being positive.
Type-1 edge
Type-2 edge
wu = T − minxu Zu (xu , ρ−i ) (u ∈ Ui )
wj = T − dj /cj (ρ−i ) (j ∈ Ji )
First, we remark that any yi is feasible to (13) if and only
if all the pairs u with yu = 1 (u ∈ Ui ) form a matching (or
an
P empty edge set) in Gi . Otherwise, there exists j such that
u∈Vj yu > 2, and (12g) would be violated. Then, by the
definition of weights, minimizing the load ρi becomes finding
a maximum weighted matching.
The algorithm S-C ELL solving (13) exactly is as follows.
Lines 1–8 compute the edge weights of the graph to be
8
S-C ELL(ρ−i )
1 for u ∈ Ui
2
hq⊕u , q u , x⊕ , x , xu i = S PLIT(u, ρ−i )
3
wu = T − (x⊕ + x + xu ) // wu > 0
4 end for
5 if |Ji | is odd
6
for j ∈ Ji
7
xj = dj /cj (ρ−i )
8
wj = T − xj // wj > 0
9
end for
10 end if
11 Construct Gi by (24)
12 U∗i = M AXIMUM -W EIGHTED -M ATCHING(Gi )
13 x∗i = 0, q∗i = 0, y∗i = 0
14 for u ∈ U∗i ∩ Ui
15
x∗u = xu , y∗u = 1
16
for j ∈ u
17
q∗ju = qju , x∗j = xj
18
end for
19 end for
20 if |Ji | is odd
21
Find the {j, ∆} in U∗i and let x∗j = xj
22 end ifP
P
23 ρi = j∈Ji x∗j + u∈Ui x∗u
24 return hρi , q∗i , x∗i , y∗i i
To prove (13) is feasible, we prove the remaining problem is
always feasible. Note that, with yi and qu being fixed, (13)
becomes a linear programming (LP) problem, which is stated
below (the equalities are by Lemma 2).
X
X
X
min
xj +
xu , s.t. cj xj +
cju xu = dj . (25)
x>0
u∈Ui
j∈Ji
u∈Vj
By Farkas’ lemma, a group of linear constraints in standard
form, i.e. Ax = b (b > 0), is feasible with x > 0 if and only
if there does not exist v such that v> A > 0> and v> b < 0.
Obviously, there is no v with d > 0 satisfying (26).
X
vj > 0 (j ∈ Ji ) and
v j dj < 0
(26)
j∈Ji
Hence (25) is feasible, and the conclusion holds.
Let λi = |Yi |. For each yi in Yi , we use an integer in
[1, λi ] to uniquely index yi . We refer to all the pair selection
solutions in Yi as pairing 1, pairing 2, . . . , pairing λi . Denote
by fik (ρ−i ) the optimum of (13) under pairing k (1 6 k 6
λi ), i.e.,
fik (ρ−i ) =
min ρi s.t. (12c)–(12f) and (12h) of cell i.
ρi ,qi ,xi
Let fi (ρ−i ) be the optimum5 of (13). Then we have:
fi (ρ−i ) =
min
k=1,2,...,λi
fik (ρ−i ).
Network-wisely, we have:
f(ρ) = [f1 (ρ−1 ), f2 (ρ−2 ), . . . , fn (ρ−n )].
constructed. Then we construct the graph in Line 11 and
compute the maximum matching4 U∗i in Line 12, which by
Theorem 3 is the optimal pair selection in cell i. Lines 13–22
assign the obtained solutions to hq∗⊕u , q∗ u , x∗⊕ , x∗ , x∗u i for the
pairs in U∗i . The other pairs are not selected and hence their
values in x∗i , q∗i , and y∗i are zeros.
We remark that in the matching process, if the number
of nodes is odd, for the unpaired UE, it is allocated with
orthogonal RBs. For two UEs that are paired in the solution of
matching, they are not necessarily in non-orthogonal allocation
but is up to optimization.
VI. M ULTI - CELL L OAD O PTIMIZATION
This section proposes the algorithmic framework M-C ELL
for deriving the optimum of M IN F, by analyzing sufficientand-necessary conditions of optimality and feasibility.
A. Revisiting Single-cell Load Minimization
Recall that for single cell optimization, the optimum of (13)
of cell i (i ∈ C) is a function of the load of other cells ρ−i .
By Lemma 3 below, this function is well-defined for any nonnegative ρ−i .
best known algorithm
[41] runs on Gi in O((|Ui | + |Ji |)
p
(odd |Ji |) or O(|Ui | |Ji |) (even |Ji |).
4 The
Theorem 4. f(ρ) is an SIF, i.e. the following properties hold:
1) (Scalability) αf(ρ) > f(αρ), ρ > 0, α > 1.
2) (Monotonicity) f(ρ) > f(ρ 0 ), ρ > ρ 0 , ρ, ρ 0 > 0.
Proof. We first prove monotonicity and scalability for
fik (ρ−i ) (i ∈ C, k = 1, 2, . . . , λi ). For monotonicity, we
0
0
prove that fik (ρ−i
) 6 fik (ρ−i ) for ρ−i
6 ρ−i as follows.
0
Given any non-negative ρ−i , we replace ρ−i with ρ−i
. Note
0
that cju (ρ−i ) > cju (ρ−i ). Thus the replacement makes the
0
solution space of (12d) larger, and the optimum with ρ−i
is no
0
larger than that with ρ−i . Therefore fik (ρ−i ) 6 fik (ρ−i ). For
scalability, we prove that fik (αρ−i ) 6 αfik (ρ−i ) for α > 1
and non-negative ρ−i as follows. Denote the optimal solution
of fik (ρ−i ) by hρi00 , qi00 , xi00 i. We have fik (ρ−i ) = ρi00 . Due
to that 1/cju (qi , ρ−i ) and 1/cj (ρ−i ) are strictly concave in
ρ−i , the two inequalities
α
1
α
1
<
,
<
(29)
cj (αρ−i )
cj (ρ−i ) cju (qi , αρ−i )
cju (qi , ρ−i )
hold for α > 1. Consider the following minimization problem 30, with yi being fixed to pairing k.
min
ρi ,qi ,xi >0
p
|Ji |)
(28)
The following theorem reveals a key property of f(ρ).
Lemma 3. The problem in (13) is always feasible.
Proof. We select some yi in Yi and fix it in (13) (Yi 6= φ
by definition). For each pair u, we fix qu to [pi /2, pi /2]> .
(27)
s.t.
ρi
(30a)
(12c), (12e) and (12f) of cell i
(30b)
X
cj (ρ−i )xj +
cju (q, ρ−i )xu > αdj , j ∈ Ji (30c)
u∈Vj
5 Therefore,
fi (ρ−i ) equals the ρi obtained from S-C ELL(ρ−i ).
9
Note that hαρi00 , qi00 , αxi00 i is feasible to (30), with the objective
value being αfik (ρ−i ). Hence the optimum of (30) is no more
than αfik (ρ−i ). For fik (αρ−i ), note that the corresponding
optimization problem only differs with (30) in (30c). Instead
of (30c), in fik (αρ−i ) we have:
X
cj (αρ−i )xj +
cju (q, αρ−i ) > dj , j ∈ Ji
(31)
u∈Vj
By Lemma 2, (30c) is equality at the optimum. Then by (29),
for any solution of (30), using it for the optimization problem
associated with fik (αρ−i ) makes (31) an inequality. Therefore the problem for fik (αρ−i ) has a lower optimum than
(30). Further, the optimum is lower than αfik (ρ−i ). Hence
fik (αρ−i ) < αfik (ρ−i ).
We then allow k to be variable and consider fi (ρ−i ) (i ∈ C).
0
For ρ−i
6 ρ−i we have
0
0
fi (ρ−i
) = min fik (ρ−i
) 6 min fik (ρ−i ) = fi (ρ−i )
k
k
and for α > 1 we have
fi (αρ−i ) = min fik (αρ−i ) < α min fik (ρ−i ) = αfi (ρ−i )
k
k
Hence the conclusion.
Given ρ, denote by fk (k > 1) the function composition of
f(fk−1 (ρ)) (with f0 (ρ) = ρ). Lemma 4 holds by [40].
Lemma 4. If limk→∞ fk (ρ) exists, then it exists uniquely for
any ρ > 0.
B. Optimality and Feasibility
Based on Theorem 4, we derive sufficient-and-necessary
conditions for M IN F in terms of its feasibility and optimality.
For any load ρ, we say that a load ρ is achievable if and only
if there exist q, x, and y such that the solution hρ, q, x, yi is
feasible to M IN F.
Lemma 5. For any ρ > 0, if there exists i ∈ C such that
ρi < fi (ρ−i ), then ρ is not achievable in M IN F.
Proof. Let ρi0 = fi (ρ−i ). By the definition of fi , ρi0 is the
minimum value satisfying (12c)–(12i) under ρ−i . Therefore
any ρi with ρi < ρi0 is not achievable with constraints (12c)–
(12i). Hence the conclusion.
Theorem 5. In M IN F, ρ (ρ 6 ρ̄) is achievable if and only if
f(ρ) is achievable and ρ > f(ρ).
Proof. By the inverse proposition of Lemma 5, an achievable
ρ always satisfies ρ > f(ρ). The necessity is proved as
follows. Suppose ρ is achievable for M IN F. Consider using
f(ρ) as another solution (together with the hq, x, yi obtained
when computing f(ρ)). Then f(ρ) satisfies (12b). Also, f(ρ)
together with its hq, x, yi fulfills (12c)–(12i) by the definition
of f(ρ). Thus, f(ρ) is achievable.
For the sufficiency, note that the achievability of f(ρ)
implies that ρ along with hq, x, yi obtained by solving f(ρ)
satisfies (12c)–(12i). Combined with the precondition ρi 6 ρ̄
(i ∈ C), the load ρ is feasible to (12b)–(12i) (and thus
achievable in M IN F). Hence the conclusion.
Theorem 5 provides an effective method for improving any
sub-optimal solution to M IN F. For any achievable ρ, evaluating f(ρ) always yields a better solution6 . This conclusion
is based on Theorem 5: Suppose ρ (ρ > 0) is the current
cell load, and let ρ 0 be the function value evaluated at ρ, i.e.
ρ 0 = f(ρ). By Theorem 5, we always have ρ 0 6 ρ.
Recall that F(ρ) is the objective function of the problem
M IN F. Theorem 6 below states that, the fixed point of f(ρ)
(along with hq, x, yi obtained when computing f(ρ)) is optimal to M IN F.
Theorem 6. Load ρ∗ is the optimum of M IN F if (and only if
when F(ρ) is strictly monotonic) ρ∗ = f(ρ∗ ) 6 ρ̄.
Proof. (Necessity) If ρ∗ is optimal (and thus feasible), then
obviously we have ρ∗ 6 ρ̄. By Theorem 5, f(ρ∗ ) is also
feasible and f(ρ∗ ) 6 ρ∗ . By successively applying Theorem 5,
fk (ρ∗ ) for any k > 1 is a feasible solution and fk (ρ∗ ) 6
fk−1 (ρ∗ ). Let ρ 0 = limk→∞ fk (ρ∗ ). Then ρ 0 6 ρ∗ holds by
the above derivation. In addition, note that ρ 0 is a feasible
solution as well. By that ρ∗ is optimal for M IN F, we have
ρ 0 = ρ∗ , otherwise ρ 0 would lead to a better objective value in
M IN F than ρ∗ . Hence ρ∗ = limk→∞ fk (ρ∗ ), i.e. ρ∗ = f(ρ∗ ).
(Sufficiency) By Theorem 5, for any feasible ρ,
limk→∞ fk (ρ) is feasible and limk→∞ fk (ρ) 6 ρ holds.
By Lemma 4, the limit remains for any ρ > 0, and thus
limk→∞ fk (ρ) = limk→∞ fk (ρ∗ ). Since ρ∗ = f(ρ∗ ), we
have ρ∗ = limk→∞ fk (ρ∗ ). Thus ρ∗ 6 ρ for any feasible
ρ, meaning that ρ∗ is optimal for M IN F.
Hence the conclusion.
C. The Algorithmic Framework
Starting from any non-negative ρ(0) , we compute
limk→∞ fk (ρ) iteratively. During each iteration, n problems
in (13) for i ∈ C are solved. The convergence is guaranteed by
Lemma 4. At the convergence, by Theorem 6, the optimum is
reached. Note that once ρ(k) is feasible for any k > 0, then by
Theorem 5, all ρ(k+1) , ρ(k+2) , . . . are feasible as well. One
can terminate prematurely to obtain a sub-optimal solution
with less computation. M-C ELL is outlined below.
M-C ELL(ρ(0) , )
1 k=0
2 repeat
3
k = k+1
4
for i ∈ C
(k)
(k) (k)
(k)
(k−1)
5
hρi , qi , xi , yi i = S-C ELL(ρ−i )
6
end for
7 until kρ(k) − ρ(k−1) k∞ 6
(k)
8 if ρi > ρ̄ for some i (i ∈ C)
9
M IN F is infeasible
10 return hρ(k) , q(k) , x(k) , y(k) i
M-C ELL applies fixed point iterations using f(ρ). The
convergence of fixed point iterations on f(ρ) is linear [42]. The
6 Rigorously, Theorem 5 implies that the new solution is not worse. In fact
it is guaranteed to be strictly better (with strictly monotonic F(ρ)) unless the
old one is already optimal. A proof can be easily derived based on Theorem 6.
10
Total load
15
called “asynchronous fixed-point iterations” [40] can be used.
The asynchronous fixed-point iterations converge to the fixed
point that is the same as obtained by its synchronized version.
Intuitively, the fixed point is unique, regardless of how we
reach it.
OMA(Opt)
SP(Uni)
SP(FTPC)
NOMA(Opt)
10
VII. P ERFORMANCE E VALUATION
5
0
0.2
0.4
0.6
0.8
1
Normalized demand d
Max load
1
OMA(Opt)
SP(Uni)
SP(FTPC)
NOMA(Opt)
We use a cellular network of 19 cells. To eliminate edge
effects, wrap-around technique [43] is applied. Inside each
cell, 30 UEs are randomly and uniformly distributed. In each
cell, there are in total 30
= 435 possible choices for user
2
pairing in NOMA. User demands are set to be a uniform
value d. In the simulations, d is normalized by M × B in (8)
and (9), and belongs to (0, 1]. The network in OMA reaches
the resource limit at d = 1.0, i.e., any d > 1.0 leads to at
least one cell being overload in OMA. Other parameters are
given in Table I.
Table I
S IMULATION PARAMETERS .
0.5
0
0.2
0.4
0.6
0.8
1
Normalized demand d
Figure 4. This figure illustrates the total and maximum load in function of
normalized demand. At d = 1.0, the network reaches its resource limit
such that any larger demand cannot be satisfied by OMA(Opt) . SP(Uni) and
SP(FTPC) are two sub-optimal NOMA power allocation schemes, for which,
pair selections are optimally computed.
feasibility check is done by Lines 8 and 9. The infeasibility
of M IN F implies that at least one cell will be overloaded for
meeting user demands. If this happens, we know for sure
that the user demands cannot be satisfied. We remark that
all the conclusions derived in this section are independent
of the implementation of S-C ELL in Line 5. As long as the
sub-routine S-C ELL yields the optimal solution to (13), MC ELL achieves the optimum of M IN F7 . Besides, M-C ELL
possesses the optimality for M IN F with any objective function
that is monotonically (but not necessarily strictly monotonic)
increasing in each element of ρ. These two properties make
M-C ELL an algorithmic framework. To our knowledge, the
most efficient S-C ELL is what we derived in Section V.
For a cell i (i ∈ C), given the information of other cells’ load
ρ−i , solving fi (ρ) is based on local information, making MC ELL suitable to run in a distributed manner. A cell can maintain the information of a subset of cells (e.g., the surrounding
cells) having major significance in terms of interference, and
exchange the information with other cells periodically, which
can be implemented via the LTE X2 interface. The technique
7 With filtered U, the proposed M-C ELL is proved to converge to the global
optimum of M IN F. Without filtered U (or for any possible candidate pairs
set U), M-C ELL is still applicable to M IN F though there is no theoretical
guarantee of convergence or optimality, as the decoding order for each pair
may change in the iteration process.
Parameter
Cell radius
Carrier frequency
Total bandwidth
Cell load limit ρ̄
Path loss model
Shadowing (Log-normal)
Fading
Noise power spectral density
RB power pi (i ∈ C)
Convergence tolerance ()
Value
500 m
2 GHz
20 MHz
1.0
COST-231-HATA
6 dB standard deviation
Rayleigh flat fading
−173 dBm/Hz
800 mW
10−4
We consider two objectives for performance evaluation:
resource efficiency and load balancing.
P For resource efficiency,
the objective function is F(ρ) = i∈C ρi , i.e., to minimize
the total network time-frequency resource consumption (or
cells’ average resource consumption if divided by n). For load
balancing, we adopt min-max fairness and the objective function is F(ρ) = maxi∈C ρi . Section VII-A and Section VII-B
provide results for power allocation and user pairing, respectively. The optimal OMA, named OMA(Opt) , is obtained by
fixing y to 0 in M IN F and solving the remaining problem to
optimality8 . The proposed optimal NOMA solution is named
NOMA(Opt) in the remaining context.
A. Power Allocation
We use OMA(Opt) as baseline. As for NOMA, the pairing
candidate set U initially covers all pairs of UEs in each cell.
Then, those pairs not fulfilling Lemma 1 are dropped from
U. We then use M-C ELL to compute NOMA(Opt) . Besides the
optimal NOMA, we implement two other sub-optimal NOMA
power split schemes for comparison. One is named “SP(Uni) ”,
in which the power pi splits equally between q⊕u and q u for
any pair u = {⊕, } (u ∈ U). The other is “fractional transmit
power control” (FTPC), named SP(FTPC) , using a parameter
8 With y being fixed to 0 in M IN F, the variables q and x disappear. Then
P
(k)
we modify Line 5 of M-C ELL to be “ρi =
j∈Ji dj /cj (ρ−i )” and
Line 10 to be “return ρ(k) ”. The modified M-C ELL gives the optimal load
for OMA (see [17] for further details).
11
Best
Best
(a) PA(B-W)
Best
Best
(c) Optimal pairing in filtered U
(b) PA(B-SB)
(d) Pairing in non-filtered U
Figure 5. This figure illustrates pair selection in a typical cell of 30 UEs. The UEs are represented by the vertices on a circle. The UE marked “Best” at the
top position has the best channel condition. The UEs are arranged clock-wisely in the descending order of channel conditions. The edges are selected pairs.
Each subfigure represents one pairing method. Figure 5(a) and Figure 5(b) show PA(B-W) and PA(B-SB) , respectively. Figure 5(c) shows optimal pairing after
filtered U. Figure 5(d) shows the pairing solution obtained with non-filtered U.
Cell load
0.7
0.6
|U| =
30
2 × 19
= 8265
|U| = 5779 (U filtered by Lemma 1)
0.5
0.4
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
Cell index
Figure 6. This figure shows the optimized
load levels of all 19 cells. The cells are numbered in an ascending order of loads. The blue bars show the computed
cell load under U composed of all 30
2 × 19 pairs. The red bars show the minimum cell load with pairs satisfying Lemma 1.
0.6
0.4
0.8
OMA(Opt)
PA(B-W) -SP(Uni)
PA(B-W) -SP(FTPC)
PA(B-W) -SP(Opt)
NOMA(Opt)
Average load
Average load
0.8
0.2
0.6
0.4
OMA(Opt)
PA(B-SB) -SP(Uni)
PA(B-SB) -SP(FTPC)
PA(B-SB) -SP(Opt)
NOMA(Opt)
0.2
0
0
0.2
0.4
0.6
0.8
1.0
Normalized demand d
0.2
0.4
0.6
0.8
1.0
Normalized demand d
Figure 7. This figure evaluates PA(B-W) , combined with three power split
schemes, SP(Uni) , SP(FTPC) , and SP(Opt) . In SP(Opt) , the power split is optimal
for each pair. OMA(Opt) and NOMA(Opt) are baselines.
Figure 8. This figure evaluates PA(B-SB) , combined with three power split
schemes, SP(Uni) , SP(FTPC) , and SP(Opt) . In SP(Opt) , the power split is optimal
for each pair. OMA(Opt) and NOMA(Opt) are baselines.
to control the fairness for power split. We set this parameter
to be 0.4 as recommended in [44]. Under both SP(Uni) and
SP(FTPC) , we use the method in Section V-B to compute the
optimal pair selection. Both two power split schemes are easily
accommodated by M-C ELL.
Figure 4 shows the total load and the maximum load in
function of normalized demand. As expected, the cell load
levels monotonically increase with user demand. At high user
demand, NOMA(Opt) dramatically improves the load performance. For d = 1.0, it achieves 31% better performance
than OMA(Opt) for both total load and maximum load. The
two sub-optimal solutions SP(Uni) and SP(FTPC) also result in
load improvement than OMA(Opt) . Compared to the two suboptimal solutions, the improvement achieved by NOMA(Opt)
over OMA(Opt) is doubled or more. On average, by using the
same amount of time-frequency resource, NOMA(Opt) delivers
33% more bits demand than OMA(Opt) . Besides, SP(FTPC)
achieves better performance than SP(Uni) , as the former takes
into account the channel conditions in power split. Generally,
in SP(FTPC) , UE with worse channel is allocated with more
12
B. User Pairing
We study the influence of user pairing by considering
two sub-optimal ones [4], named “PA(B-W) ” and “PA(B-SB) ”,
respectively. Suppose we sort the UEs in descending order
of their channel conditions. In PA(B-W) , the UE with the best
channel condition is paired with the UE with the worst, and
the UE with the second best is paired with one with the
second worst, and so on. In PA(B-SB) , the UE with the best
channel condition is paired with the one with the second
best, and so on. See Figures 5(a) and 5(b) for an illustration.
In addition, we examine to what extend pair filtering (by
Lemma 1) affects performance. For filtered U, optimal pair
selection is done by Section V-B. For non-filtered U, we
apply M-C ELL even though there is no theoretical guarantee
on optimality. Convergence, however, is observed for all the
instances we considered. Figures 5(c) and 5(d) illustrated the
resulted selection patterns.
In Figure 6, we show the load levels of all 19 cells with
d = 1.0, under both filtered and non-filtered
U. In this
specific scenario, |U| is reduced from 30
2 × 19 = 8265 to
5779 after being filtered by Lemma 1. We choose d = 1.0
because the performance difference among the solutions is
the largest. There is very slight difference in cell load levels
between the two cases. Numerically, the differences between
them are only 0.1% and 0.5% for average and maximum cell
load, respectively. This result is coherent with Figure 5(c) and
Figure 5(d). One can see that the patterns of the two pair
selection solutions are almost identical. Thus, pair filtering by
Lemma 1 is effective in reducing the number of candidate
pairs, with virtually no impact on performance.
In Figure 7 and Figure 8, we respectively evaluate PA(B-W)
and PA(B-SB) , combined with three power split schemes SP(Uni) ,
SP(FTPC) , and SP(Opt) . In SP(Opt) , we use the algorithm S PLIT to
compute the optimal power split for each pair. All of SP(Uni) ,
SP(FTPC) , and SP(Opt) are put into the framework of M-C ELL
but with fixed pair selection PA(B-W) or PA(B-SB) . In addition,
OMA(Opt) and NOMA(Opt) are also included for comparison as
baselines.
One can see that all the NOMA schemes outperform
OMA(Opt) . In Figure 7, with PA(B-W) , SP(FTPC) outperforms
SP(Uni) . SP(Opt) beats the other two. On one hand, there
is non-negligible gap in load performance between SP(Opt)
and NOMA(Opt) , even though in SP(Opt) , the power split is
optimal for the PA(B-W) pairing. Hence pair selection plays an
important role for NOMA performance. On the other hand,
SP(Opt) yields significantly load improvement compared to
OMA(Opt) , and we conclude that PA(B-W) is a good sub-optimal
pair selection for NOMA. Indeed, PA(B-W) pairs the UEs in
a greedy way, aiming at maximizing the diversity of channel
conditions of paired UEs. As shown in Figure 5(c), the optimal
pair selection has a similar trend. The difference is that optimal
pairing has a more “global view” than PA(B-W) . In Figure 8,
under PA(B-SB) , SP(Uni) , SP(FTPC) , and SP(Opt) improve the load
very slightly. All of the three are far from the global optimum
and the gap is large under high user demands. We conclude
that PA(B-SB) is not as effective as PA(B-W) in terms of network
load optimization.
As the overall conclusion, jointly optimizing power allocation and user pairing is important for the performance of
NOMA.
C. Convergence Analysis
We show the convergence performance of M-C ELL in Figure 9, for demands 0.3, 0.5, 0.7, and 1.0, respectively. Initially,
(0)
ρi = 1 (i ∈ C). We observe that M-C ELL converges very
fast. With higher demand, the convergence becomes slightly
faster. High accuracy is reached after a very few iterations. For
all the demands consider in the figure, even if we terminate
M-C ELL after a very few iterations, the obtained solution is
close to the optimum.
10−1
kρ(k) − ρ(k−1) k∞
power.
In summary, power allocation has considerably large influence on NOMA. Even if the UE pairs are optimally selected,
sub-optimal power allocations in NOMA have significant
deviation from optimal NOMA.
d = 1.0
d = 0.7
d = 0.4
d = 0.1
10−5
10−9
10−13
2
3
4
5
6
7
8
9
10
Iteration k in M-C ELL
Figure 9. This figure shows the norm k·k∞ in function of iteration k in
M-C ELL, under the uniform demands 0.1, 0.4, 0.7, and 1.0, respectively.
VIII. C ONCLUSIONS
This paper has investigated optimal resource management
in multi-cell NOMA, with power allocation and user pairing
being considered simultaneously. Joint optimization of both is
shown to be very important for NOMA performance. The proposed system model admits a mixed use of OMA and NOMA
for the users. Therefore, network architectures that support
various multiple access techniques can be analyzed under this
model. Finally, as for future work, the paper suggests that
mathematical tools in SIF are useful for analyzing multi-cell
NOMA. In summary, NOMA is a promising technique for
spectrum efficiency enhancement and cell load balancing.
13
A PPENDIX
We remark that the function (32) is first referred to in Section V-A and is used in the formulation (17). The two functions (33)
and (34) are first referred to in Section V-A and are used in the algorithm S PLIT.
w⊕ (ρ−i )e(c⊕u +c u ) + (w (ρ−i ) − w⊕ (ρ−i ))ec u
Cvu (c⊕u , c u , ρ−i ) = log
.
(32)
pi + w (ρ−i )
K
d⊕ /x
cK
u
⊕u xu > d⊕
h
i
d
/x
u
H⊕u (xu , ρ−i ) =
(33)
(pi +w (ρ−i ))/e ⊕
+w⊕ (ρ−i )−w (ρ−i )
log
Otherwise
w⊕ (ρ−i )
K K
d /x
hu
i c u xu > d
H u (xu , ρ−i ) =
(34)
pi +w (ρ−i )
Otherwise
log
d /xu
w⊕ (ρ−i )e
R EFERENCES
[1] L. You, L. Lei, D. Yuan, S. Sun, S. Chatzinotas, and B. Ottersten, “A
framework for optimizing multi-cell NOMA: Delivering demand with
less resource,” in 2017 IEEE GLOBECOM, 2017.
[2] S. M. R. Islam, N. Avazov, O. A. Dobre, and K. S. Kwak, “Powerdomain non-orthogonal multiple access (NOMA) in 5G systems: Potentials and challenges,” IEEE Communications Surveys Tutorials, vol. 19,
no. 2, pp. 721–742, 2017.
[3] W. Shin, M. Vaezi, B. Lee, D. J. Love, J. Lee, and H. V.
Poor, “Non-orthogonal multiple access in multi-cell networks: Theory,
performance, and practical challenges,” arXiv.org, 2016. [Online].
Available: https://arxiv.org/pdf/1611.01607.pdf
[4] Z. Ding, P. Fan, and H. V. Poor, “Impact of user pairing on 5G
nonorthogonal multiple-access downlink transmissions,” IEEE Transactions on Vehicular Technology, vol. 65, no. 8, pp. 6010–6023, 2016.
[5] “Evaluation methodologies for downlink multiuser superposition transmissions,” 3GPP, Tech. Rep. R1-153332, 2014.
[6] W. Shin, M. Vaezi, B. Lee, D. J. Love, J. Lee, and H. V. Poor,
“Coordinated beamforming for multi-cell MIMO-NOMA,” IEEE Communications Letters, vol. 21, no. 1, pp. 84–87, 2017.
[7] J. Kim, J. Koh, J. Kang, K. Lee, and J. Kang, “Design of user clustering
and precoding for downlink non-orthogonal multiple access (NOMA),”
in 2015 IEEE MILCOM, 2015, pp. 1170–1175.
[8] H. Tabassum, E. Hossain, and M. J. Hossain, “Modeling and analysis of
uplink non-orthogonal multiple access (NOMA) in large-scale cellular
networks using poisson cluster processes,” arXiv.org, 2016. [Online].
Available: http://arxiv.org/abs/1610.06995.pdf
[9] Y. Fu, Y. Chen, and C. W. Sung, “Distributed power control for the
downlink of multi-cell NOMA systems,” IEEE Transactions on Wireless
Communications, to appear.
[10] L. P. Qian, Y. Wu, H. Zhou, and X. Shen, “Joint uplink base station association and power control for small-cell networks with nonorthogonal multiple access,” IEEE Transactions on Wireless Communications, vol. 16, no. 9, pp. 5567–5582, 2017.
[11] B. Di, L. Song, Y. Li, and G. Y. Li, “Non-orthogonal multiple access
for high-reliable and low-latency V2X communications in 5G systems,”
IEEE Journal on Selected Areas in Communications, vol. 35, no. 10,
pp. 2383–2397, 2017.
[12] L. P. Qian, Y. Wu, H. Zhou, and X. Shen, “Dynamic cell association
for non-orthogonal multiple-access V2S networks,” IEEE Journal on
Selected Areas in Communications, vol. 35, no. 10, pp. 2342–2356,
2017.
[13] J. Zhu, J. Wang, Y. Huang, S. He, X. You, and L. Yang, “On optimal
power allocation for downlink non-orthogonal multiple access systems,”
IEEE Journal on Selected Areas in Communications, to appear.
[14] L. You and D. Yuan, “Joint CoMP-cell selection and resource allocation
in fronthaul-constrained C-RAN,” in 2017 WiOpt Workshop, 2017, pp.
1–6.
[15] L. Lei, D. Yuan, C. K. Ho, and S. Sun, “Optimal cell clustering and
activation for energy saving in load-coupled wireless networks,” IEEE
Transactions on Wireless Communications, vol. 14, no. 11, pp. 6150–
6163, 2015.
[16] I. Viering, M. Dottling, and A. Lobinger, “A mathematical perspective
of self-optimizing wireless networks,” in 2009 IEEE ICC, 2009, pp. 1–6.
[17] I. Siomina and D. Yuan, “Analysis of cell load coupling for LTE
network planning and optimization,” IEEE Transactions on Wireless
Communications, vol. 11, no. 6, pp. 2287–2297, 2012.
+w (ρ−i )−w⊕ (ρ−i )
[18] A. J. Fehske, I. Viering, J. Voigt, C. Sartori, S. Redana, and G. P.
Fettweis, “Small-cell self-organizing wireless networks,” Proceedings
of the IEEE, vol. 102, no. 3, pp. 334–350, 2014.
[19] L. You, D. Yuan, N. Pappas, and P. Värbrand, “Energy-aware wireless
relay selection in load-coupled OFDMA cellular networks,” IEEE Communications Letters, vol. 21, no. 1, pp. 144–147, 2017.
[20] L. You and D. Yuan, “Load optimization with user association in
cooperative and load-coupled LTE networks,” IEEE Transactions on
Wireless Communications, vol. 16, no. 5, pp. 3218–3231, 2017.
[21] I. Siomina, A. Furuskr, and G. Fodor, “A mathematical framework for
statistical QoS and capacity studies in OFDM networks,” in 2009 IEEE
PIMRC, 2009, pp. 2772–2776.
[22] K. Majewski and M. Koonert, “Conservative cell load approximation
for radio networks with Shannon channels and its application to LTE
network planning,” in 2010 Sixth Advanced International Conference on
Telecommunications, 2010, pp. 219–225.
[23] E. Pollakis, R. L. G. Cavalcante, and S. Staczak, “Base station selection for energy efficient network operation with the majorizationminimization algorithm,” in 2012 IEEE SPAWC, 2012, pp. 219–223.
[24] A. J. Fehske and G. P. Fettweis, “Aggregation of variables in load models
for interference-coupled cellular data networks,” in 2012 IEEE ICC,
2012, pp. 5102–5107.
[25] A. J. Fehske, H. Klessig, J. Voigt, and G. P. Fettweis, “Concurrent loadaware adjustment of user association and antenna tilts in self-organizing
radio networks,” IEEE Transactions on Vehicular Technology, vol. 62,
no. 5, pp. 1974–1988, 2013.
[26] C. K. Ho, D. Yuan, and S. Sun, “Data offloading in load coupled
networks: A utility maximization framework,” IEEE Transactions on
Wireless Communications, vol. 13, no. 4, pp. 1921–1931, April 2014.
[27] R. L. G. Cavalcante, S. Stanczak, M. Schubert, A. Eisenblaetter, and
U. Tuerke, “Toward energy-efficient 5g wireless communications technologies: Tools for decoupling the scaling of networks from the growth
of operating power,” IEEE Signal Processing Magazine, vol. 31, no. 6,
pp. 24–34, 2014.
[28] S. Tombaz, S. w. Han, K. W. Sung, and J. Zander, “Energy efficient
network deployment with cell DTX,” IEEE Communications Letters,
vol. 18, no. 6, pp. 977–980, 2014.
[29] B. Baszczyszyn, M. Jovanovic, and M. K. Karray, “Performance laws
of large heterogeneous cellular networks,” in 2015 WiOpt, 2015, pp.
597–604.
[30] C. K. Ho, D. Yuan, L. Lei, and S. Sun, “Power and load coupling
in cellular networks for energy optimization,” IEEE Transactions on
Wireless Communications, vol. 14, no. 1, pp. 509–519, 2015.
[31] R. L. G. Cavalcante, S. Staczak, J. Zhang, and H. Zhuang, “Low
complexity iterative algorithms for power estimation in ultra-dense load
coupled networks,” IEEE Transactions on Signal Processing, vol. 64,
no. 22, pp. 6058–6070, 2016.
[32] H. Klessig, D. hmann, A. J. Fehske, and G. P. Fettweis, “A performance
evaluation framework for interference-coupled cellular data networks,”
IEEE Transactions on Wireless Communications, vol. 15, no. 2, pp. 938–
950, 2016.
[33] R. L. G. Cavalcante, Y. Shen, and S. Staczak, “Elementary properties of
positive concave mappings with applications to network planning and
optimization,” IEEE Transactions on Signal Processing, vol. 64, no. 7,
pp. 1774–1783, 2016.
[34] Q. Liao, “Dynamic uplink/downlink resource management in flexible
duplex-enabled wireless networks,” in 2017 ICC Workshops, 2017, pp.
625–631.
14
[35] R. L. G. Cavalcante, M. Kasparick, and S. Staczak, “Max-min utility
optimization in load coupled interference networks,” IEEE Transactions
on Wireless Communications, vol. 16, no. 2, pp. 705–716, 2017.
[36] D. A. Awan, R. L. G. Cavalcante, and S. Stanczak, “A robust machine
learning method for cell-load approximation in wireless networks,”
arXiv.org, 2017. [Online]. Available: http://arxiv.org/abs/1710.09318
[37] D. Tse and P. Viswanath, Fundamentals of Wireless Communication.
Cambridge university press, 2005.
[38] G. Geraci, M. Wildemeersch, and T. Q. S. Quek, “Energy efficiency
of distributed signal processing in wireless networks: A cross-layer
analysis,” IEEE Transactions on Signal Processing, vol. 64, no. 4, pp.
1034–1047, 2016.
[39] M. Wildemeersch, T. Q. S. Quek, M. Kountouris, A. Rabbachin, and
C. H. Slump, “Successive interference cancellation in heterogeneous
networks,” IEEE Transactions on Communications, vol. 62, no. 12, pp.
4440–4453, 2014.
[40] R. D. Yates, “A framework for uplink power control in cellular radio
systems,” IEEE Journal on Selected Areas in Communications, vol. 13,
no. 7, pp. 1341–1347, 1995.
√
[41] S. Micali and V. V. Vazirani, “An O( V · |E|) algoithm for finding
maximum matching in general graphs,” in 21st Annual Symposium on
Foundations of Computer Science, 1980, pp. 17–27.
[42] H. R. Feyzmahdavian, M. Johansson, and T. Charalambous, “Contractive
interference functions and rates of convergence of distributed power
control laws,” IEEE Transactions on Wireless Communications, vol. 11,
no. 12, pp. 4494–4502, 2012.
[43] D. Huo, “Clarification on the wrap-around hexagon network structure,”
IEEE, Tech. Rep. C802.20-05/15, 2005.
[44] Y. Saito, A. Benjebbour, Y. Kishiyama, and T. Nakamura, “System-level
performance evaluation of downlink non-orthogonal multiple access
(NOMA),” in 2013 IEEE PIMRC, 2013, pp. 611–615.
| 7cs.IT
|
arXiv:1802.10427v1 [math.GR] 28 Feb 2018
A few remarks on invariable generation in infinite groups
Gil Goffer
Gennady A.Noskov
The Weizmann Institute of Science
The Sobolev Institute of Mathematics
76100, Rehovot, Israel
644043, Omsk, Russia
∗
February 12, 2018
Abstract
A subset S of a group G invariably generates G if G is generated by {sg (s)|s ∈ S} for any choice
of g(s) ∈ G, s ∈ S. In case G is topological one defines similarly the notion of topological invariable
generation. A topological group G is said to be T IG if it is topologically invariably generated by
some subset S ⊆ G. In this paper we study the problem of (topological) invariable generation for
linear groups and for automorphism groups of trees. Our main results show that the Lie group SL2 (R)
and the automorphism group of a regular tree are T IG , and that the groups P SLm (K), m ≥ 2 are
not IG for certain countable fields of infinite transcendence degree over the prime field.
Contents
1 Introduction
2
2 Preliminary results
3
3 Parabolics and stabilizers as Wiegold subgroups
6
4 Borel versus Wiegold
9
5 Compact Lie groups and their generalizations
10
6 SL2 (R) is T IG
10
6.1
Conjugacy classes in SL2 (R) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10
6.2
One-parameter subgroups of SL2 (R) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11
6.3
One-dimensional Lie subgroups of SL2 (R) and their spectra . . . . . . . . . . . . . . . . .
12
∗ The
authors thank the support of Tsachik Gelander and Uri Bader
1
6.4
2-dimensional Lie subgroups of SL2 (R) and their spectra . . . . . . . . . . . . . . . . . . .
13
6.5
Main lemma and theorem 14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
7 Aut(T ) is T IG
14
7.1
Elements classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
14
7.2
The local action by permutation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
7.3
Conjugacy classes in Aut(T ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
7.3.1
Elements of type (n, P) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
16
7.3.2
Spherically-transitive elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
17
7.4
A set of generators for Aut(T ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
17
7.5
Proof of theorem 20 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
8 Uncountably many countable non-IG -groups
19
8.1
Laws in groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
19
8.2
Rational points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
8.3
Building up free tuples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
8.4
When is PSLm (K) an IG ?
22
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9 Open problems
1
23
Introduction
The notion of invariable generation of groups goes back at least as far as Jordan’s paper of 1872.
Theorem 1 ([Jo, Se]) . Assume that a finite group G acts transitively on a set X of cardinality at least
2. Then there exists g ∈ G displacing every element of X, i.e. g has no fixed point in X.
(We define g ∈ G to be active or fixed point free if it satisfies the conclusion of Jordan’s theorem.)
Independently, but more than hundred years later (in 1976), J.Wiegold introduced the class X of groups
G with the property that whenever G acts transitively on a set X of cardinality at least 2, there is an
active element g ∈ G, see [Wi76]. Thus Jordan’s theorem above says that every finite group belongs
to the class X. Not only Wiegold failed to mention Jordan, but also he claims indirectly that Jordan’s
theorem is obvious. “Observe that, in particular, finite groups and soluble groups have III – obvious
facts once noted, though I know of no reference to them in the literature” see [Wi76]. Wiegold gave two
more characterizations of class X. Namely, G ∈ X if and only if for every proper subgroup H ≤ G the
union ∪{H g : g ∈ G} does not cover the whole G. (Here, we use the common notation H g to denote
the conjugate g −1 Hg of H.) Secondly, G ∈ X if and only if
representative of every conjugacy class of G, generates G.
2
every subset, containing at least one
These characterizations lead naturally to the following definitions (of our own, though the concepts
implicitly occurred in above conversation). We call a proper subgroup W of a group G a Wiegold subgroup
of G if G is covered by conjugates of W , i.e. G = ∪g∈G W g . Alternative terms might be “parabolic
subgroup” or “Borel subgroup” but this could lead to confusion. Yet another alternative term might
be “abnormal subgroup” but again it could conflict with existing terminology. A subset S of a group
G is conjugation complete
(in G) if it meets every non-trivial conjugacy class in G. In these terms
it is easy to see that the class X consists of those groups that do not have Wiegold subgroups. Now
Jordan’s theorem becomes really “obvious” (but modulo Wiegold’s characterizations) for finite groups
by counting argument. Indeed, let G be finite and suppose H was a Weigold subgroup. For every g ∈ G
S
S
we have H g = H Hg , hence the union g∈G H g can be written as the union Hg∈H\G H Hg . There are
|G : H| members in the last union, each member is of cardinality |H| and all members contain the identity
element, hence the cardinality of the union is strictly less than |G : H| · |H| = |G|, contradicting the
definition of Wiegold’s subgroup.
Independently of Wiegold, but somewhat later, J.D.Dixon (1992) invented the notion of invariable
generation for finite groups. A subset S of a finite group G invariably generates G if G is generated by
{sg (s)|s ∈ S} for any choice of g(s) ∈ G, s ∈ S. This notion was motivated by Chebotarev’s Density
Theorem [Ch]. In case G is a (Hausdorff) topological group one defines similarly the notion of topological
invariable generation. A topological group G is said to be T IG if it is topologically invariably generated
by some subset S ⊆ G. In this paper we study the problem of (topological) invariable generation for
linear groups and for automorphism groups of trees. Our main results show that the Lie group SL2 (R)
and the group of tree automorphisms Aut(T ) are T IG , and that the groups P SLm (K), m ≥ 2 are not
IG for certain countable fields of infinite transcendence degree over the prime field.
2
Preliminary results
We shall formulate now the invariable generation problem in topological setup. The (left) action G y X
of a topological group G on a topological space X is continuous if the map (g, x) 7→ gx from G × X to
X is continuous. We call a proper subgroup W of a topological group G the Wiegold subgroup of G if G
is covered by conjugates of W i.e. G = ∪g∈G W g . A subset S of G is conjugation complete
(in G) if it
meets every non-trivial conjugacy class in G. A transitive action G y X of a group G on the set X has
fixed-point property (FP) if every g ∈ G has a fixed point.
Lemma 2 The following conditions on a topological group G are equivalent:
I.
Whenever an action G y X of G on a topological space X of cardinality at least 2 is continuous and
transitive, there is an active element g ∈ G (i.e. g has no fixed point in X).
3
II.
There is no closed Wiegold subgroup in G, i.e. G is not covered by the conjugates of every proper
closed subgroup H < G.
III. Every conjugation complete
subset S of G topologically generates G, i.e. the closure of subgroup
hSi generated by S equals G.
IY. There is a family (Ci ) of conjugacy classes of G (not necessarily distinct) invariably and topologically
generating G in the sense that each family (ci ) of elements ci ∈ Ci topologically generate G.
Proof.
I ⇒ II. Suppose that I holds. Arguing by contradiction suppose that there exists a closed Wiegold
−1
subgroup W in G. Fix g ∈ G. Since W is Wiegold we have g ∈ W h
for some h ∈ W which is
equivalent to the equality ghW = hW which in turn means that g has a fixed point hW in the
topological quotient G/W . The natural action G y (G/W ) is transitive and since W is proper,
the cardinality of G/W is at least 2. This contradicts the assumption that condition I holds with
X = G/W .
II ⇒ I. Suppose that II holds. Assume, to get a contradiction, that the action G y X is transitive,
|X| ≥ 2 and every g has a fixed point in X. It follows that for every x ∈ X we have a covering
∪g∈G Ggx = G, since every h ∈ G fixes some gx. But Ggx = gGx g −1 , thus G is covered by conjugates
gGx g −1 and hence Gx is Wiegold. This contradicts the assumption that II holds.
II ⇒ III. Suppose that II holds. Assume, to obtain a contradiction, that there is a conjugation complete subset S of G that S does not topologically generates G, i.e. the closure hSi is proper in G.
Each h ∈ G equals to some sg ∈ S, hence h ∈ S g ⊆ ∪g∈G S g . Then hSi is Wiegold, contradiction
with III.
III ⇒ II. Suppose that III holds. Again arguing by contradiction suppose that there is a conjugation
complete subset S that does not generate G topologically. It follows that hSi < G and S meets every
g
non-trivial conjugacy class, which yields ∪g∈G hSi = G hence W = hSi is an Wiegold subgroup,
contradicting to III.
III ⇒ IY. Suppose that III holds. Let (Ci ) be the family of all nontrivial conjugacy classes of G. Then
every family ci of elements ci ∈ Ci is conjugation complete and hence it topologically generate G.
IY ⇒ III. Suppose that IY holds. Let (Ci ) be a family of conjugacy classes of G that invariably and
topologically generate G in the sense that each family ci of elements ci ∈ Ci topologically generate
G. Let S be an arbitrary conjugation complete
subset of G. For each i there is si ∈ S such that
si ∈ Ci . Hence the family (si ) topologically generates G, and this is true all the more so for the set
S.
4
We say that the topological group G is topologically invariably generated (T IG ) if the equivalent
conditions I-IY are satisfied. In case G is discrete conditions I-III are introduced in [Wi76] and in case G
is finite condition IY is introduced in [Di]. We skip the adverb “topologically” in case the topology on G
is discrete and say about invariable generation property IG . Clearly, if G is IG then G is T IG relative
to every group topology on G. The converse is not true – for instance the free group F2 of rank 2 is
T IG relative to profinite topology but it is non non-IG ([Wi77]).
Example 3 If G is an abelian group then the conjugacy classes are just 1-element subsets of G. So the
conjugacy complete sets are those containing G − {1}. Thus every conjugacy complete set generate G.
Hence G has no Wiegold subgroup and G is IG .
Example 4 Consider the extreme case when G contains precisely two conjugacy classes: {1} and
G − {1}. Let us show that G is IG iff G ' Z/2. Clearly, G = Z/2 is IG . Conversely, suppose that G is
IG . Every 1-element set {g} ⊂ (G − {1}) is conjugacy complete, hence G is generated by g and therefore
G is cyclic. It is easily verified that the only cyclic group that has precisely two conjugacy classes is
Z/2. Thus G = Z/2. We conclude that every group G 6= Z/2 with a single non-trivial conjugacy class is
non-IG . It is proved by D.Osin that any countable torsion free group can be embedded into a finitely
generated group with exactly two conjugacy classes [Os10].
These examples suggest, that intuitively the group G is IG if it contains “many conjugacy classes”.
Lemma 5 ([Wi76, Wi77]) i The class IG is closed under extensions.
ii The class IG is closed under restricted direct products with arbitrarily many factors.
iii The class IG not closed under unrestricted direct products and not closed under passage to subgroups.
iv The class IG contains all finite groups and all solvable groups.
In [Wi77] an IG -group is presented, whose commutator subgroup is outside IG . In the present paper
it is shown that SL2 (R) is T IG , whereas it contains a discrete free group F2 , which is not IG .
Slightly modifying Wiegold’s arguments we obtain the following properties of the class T IG .
Lemma 6 The class T IG contains all profinite groups and all topological solvable groups.
Not so many classes of groups are known to be IG or T IG:
1. Virtually solvable groups (including finite groups) are T IG [Wi76].
5
2. The pro-finite completions of all arithmetic groups with the Congruence Subgroup Property are
T IG , see [KLS2].
This list is extended in the present paper by the group SL2 (R) which is shown to be in T IG .
The list of known non-IG groups is lengthy but rather chaotic:
1. Groups with one single non-trivial conjugacy class, except Z/2.
2. The finite index net subgroups in groups SO(Q, Z) where Q is a rational quadratic form of signature
(2, 2), see [GM].
3. SO(3), and, more generally, every nonabelian connected compact Lie group G since every maximal
abelian subgroup is Wiegold , see [Ge].
4. SLn (R) for n > 2, see [KLS2] .
5. Any non-virtually-solvable linear algebraic group over an algebraically closed field, for instance
SLn (C) for n ≥ 2, see [KLS2].
6. Non-elementary convergence groups (including Gromov hyperbolic groups) [Ge].
7. The Thompson groups T and V , see [GGJ].
3
Parabolics and stabilizers as Wiegold subgroups
Let GLn (K) be a general linear group of degree n ≥ 2 over the (topological) field K. Consider the
standard left action GLn (K) y K n of GLn (K) on K n . Let Gr(n, r) denote the Grassmannian of all
r-dimensional linear subspaces of K n . In the next proposition we give various conditions on the field and
the action of a group G on Grassmanian that provide the Wiegold property for certain stabilizers.
Proposition 7 Let G be a subgroup of GLn (K) and let X ⊆ Gr(n, r), n ≥ r > 1 be a G-invariant subset
with more than one element.
1. Suppose that the action G y X is transitive and satisfies fixed-point property FP. Then every
stabilizer GP (P ∈ X) is a Wiegold subgroup in G.
2. In case of algebraically closed K we can drop the fixed-point assumption at the expense of assuming
that X = Gr(n, r). Namely, let K be an algebraically closed field and suppose a subgroup G ≤
GLn (K) acts transitively on Gr(n, r) (1 ≤ r < n). Then for every r-dimensional linear subspace L
of K n the stabilizer GL is Wiegold. In particular GLn (K) and SLn (K) are not IG .
6
3. Suppose a subgroup G ≤ GLn (R) acts transitively on Grassmanian of planes Gr(n, 2) (n ≥ 3). Then
for every plane L of Rn the stabilizer GL is Wiegold. In particular GLn (R), SLn (R), On (R), SOn (R)
are not IG .
Proof. (1) is just a direct consequence of lemma 2 in terms of action of G on Grassmanian.
(2) Indeed, to apply assertion (1) it is sufficiently to check that every g ∈ GLn (K) fixes some rdimensional linear subspace L of K n . This is true for r = 1 since K is algebraically closed and for
arbitrary r an easy induction does the job.
(3) It is well known, via complexification, that every g ∈ GLn (R) fixes subspace of dimension 1 or 2,
see [Ax, p. 279],[KM, Prop.12.16]. Moreover, to each nonreal eigenvalue corresponds an invariant plane.
It follows that if n ≥ 2 then every g ∈ GLn (R) fixes subspace of dimension 2. Then for every plane L of
Rn the stabilizer GL is Wiegold. Thus GLn (R) is not IG
and moreover every subgroup G ≤ GLn (R)
which acts transitively on Gr(n, 2) turns out to be non-IG .
Symplectic groups. The symplectic group Sp2n (F ) is defined as the set of linear transformations
of a 2n-dimensional vector space F 2n over the field F that preserve a non-degenerate, skew-symmetric,
F -bilinear form ω on F 2n . A subspace L of a symplectic vector space is called Lagrangian if L = L⊥ where
L⊥ is an orthogonal complement of L relative to ω. A subspace L is Lagrangian if and only if dim L = n
and L is isotropic i.e. ω|L is identically zero. The set of Lagrangian subspaces of symplectic space
(F 2n , ω) is called the Lagrangian Grassmannian and denoted L(F 2n ). The group Sp2n (F ) acts transitively
on (F 2n ). Indeed by Witt’s theorem (1937) every linear isomorphism f : L → L0 of Lagrangian subspaces
can be extended to a symplectic automorphism of F 2n . Moreover, the following is true:
Theorem 8 ([Du], Theorem 3.4.3) Let (E, ω) be a symplectic vector space over a field k and let A
be a symplectic mapping f : (E, ω) → (E, ω) such that all its eigenvalues are contained in k. Then there
exists an f −invariant Lagrangian subspace L of (E, ω). Moreover, if k is algebraically closed then every
symplectic automorphism fixes some Lagrangian subspace.
It follows that Sp2n (F )L is Wiegold for every Lagrangian subspace L.
Wiegold subgroups over real closed fields. These fields have the same first-order properties as
R. A field K is said to be real if −1 is not a sum of squares in K, see [La, Corollary 9.3, Chapter VI][La].
A field K said to be real closed if it is real, and if any algebraic extension of K which is real must be
equal to K. By a real closure we shall mean a real closed field L which is algebraic over K. Every real
field admits a real closure. If K is real closed then every polynomial of odd degree in K[X] has a root in
√
K and K := K( −1) is the algebraic closure of K.
Proposition 9 Let K be a real closed field. Suppose a subgroup G ≤ GLn (K) acts transitively on
Grassmanian of planes Gr(n, 2) (n ≥ 3). Then for every plane L of K n the stabilizer GL is Wiegold. In
particular GLn (K), SLn (K), On (K), SOn (K) are not IG .
7
√
Proof. We assert that every g ∈ GLn (K)) (n ≥ 2) has invariant plane in column space K( −1)n . If all
eigenvalues of g are real, i.e. belong to K, then g is triangulizable over K and so has an invariant plane.
√
√
n
Now suppose that there is a non-real eigenvalue a + b −1 ∈ K, (a, b ∈ K)b 6= 0. Let w = u + −1v ∈ K
be an eigenvector corresponding to g. Then
gw = gu +
√
√
√
√
−1gv = (a + b −1)(u + −1v) = (au − bv) + −1(av + bu)
(1)
from which it follows that the space Ku + Kv is g-invariant. In fact this space is 2-dimensional. Indeed,
let say u = cv, 0 6= c ∈ K, then substituting u to equation 1 we obtain that c2 = −1 contradicting to
assumption that K is real. To complete the proof we apply now lemma 7 and obtain that the action
G y Gr(n, 2) satisfies F P and is transitive by assumption so GL is Wiegold.
In attempt to generalize proposition 1 to the subspaces of arbitrary dimension r we came to
Lemma 10 Let K be an algebraically closed field of degree d ≥ 2 over its subfield k and suppose that
the extension K|k is separable. Then every g ∈ GLn (k) fixes linear subspace in k n of dimension smaller
than or equal d.
Proof. By Primitive Element Theorem there exists a primitive element ζ ∈ K i.e. K = k(ζ). It follows
from finiteness assumption that ζ is algebraic over k. Then K = k(ζ) = k[ζ], and k(ζ)) is finite over
k [La, Proposition 1.4., Chapter V]. Moreover, the degree d = [k(ζ) : k] is equal to the degree of the
irreducible polynomial p(X) of ζ over k (loc.cit). The proof of [La, Proposition 1.4., Chapter V] shows
that the powers 1, ζ, . . . , ζ d−1 form a basis the space K over k. Let V = K ⊗k k n be an extension of
scalars from k to K. This is an n-dimensional space over K. The action of g ∈ GLn (k) on k n extends
naturally to the action on K ⊗k k n . Since K is algebraically closed there is an eigenvector v ∈ V with
P
P
eigenvalue λ ∈ K. Write uniquely λ = i li ζ i , v = j ζ j ⊗ vj , where li ∈ k, vi ∈ k n . We assert that g
P
fixes the space i kvi (whose dimension is clearly at most r). Since gv = λv we have
gv =
X
X
X
X
ζ j ⊗ gvj = λv = (
li ζ i )(
ζ j ⊗ vj ) =
li ζ i+j ⊗ vj .
j
i
j
i,j
Rewriting ζ i+j as a k-linear combinations of 1, ζ, . . . , ζ r−1 we obtain that
X
li ζ i+j ⊗ vj =
i,j
X
ζ k ⊗ vk0
k
where vk0 are k-linear combinations of vj . It follows from above that gvj = vj0 , and the lemma follows.
Unfortunately, the applicability of the above lemma is very much restricted by the Artin-Schreier
theorem which asserts that if a field K is algebraically closed with k a subfield such that 1 < [K : k] < ∞
then K = k(i) where i2 = −1, and k has characteristic 0 [La, Corollary 9.3, Chapter VI].
8
4
Borel versus Wiegold
The following is a slight variation on the result in [KLS2].
Theorem 11 Let G be a linear algebraic group over an algebraically closed field k and let B be its
Borel subgroup (=maximal connected closed solvable subgroup). 1) If G is not virtually solvable then the
normalizer NG (B) is a Wiegold subgroup of G. 2) If furthermore G is connected then B is itself Wiegold.
Proof. Firstly let us prove that NG (B) is a proper subgroup of G. Suppose to the contrary that
NG (B) = G, i.e. B is normal in G. Note that, being connected, B is contained in the connected
component G0 of G. Since B is normal in G, it also normal in G0 . Note that B is proper in G0 since
otherwise G would be a finite extension of B = G0 , i.e. virtually solvable. Finally by Chevalley result
[Bo2, theorem 11.16] B is self-normalizing in G0 , contradicting the normality of B. Hence NG (B) is a
proper subgroup of G.
By a theorem of Steinberg (see Theorem 7.2 of [St]), every automorphism of G fixes some Borel
subgroup of G. This implies that if g is any element of G, then g −1 B 0 g = B 0 for some Borel subgroup
B 0 of G. Thus the union of the normalizers NG (B 0 ) over the Borel subgroups B 0 of G equals G. Since
the Borel subgroups are all conjugate, it follows that ∪g (NG (B))g = G, thus NG (B) is Wiegold.
2) In this case we make use Grothendiecks Covering Theorem [Co, Theorem 4.11]: Let G be a
connected linear algebraic group over algebraically closed field k and let B ≤ G be a Borel subgroup then
[
G(k) =
g −1 B(k)g
g∈G
.
This immediately implies that B is Wiegold.
2-dimensional case. The standard Borel subgroup B of GL2 (K) is the group of all invertible up
per triangle matrices a0 db . For the Borel subgroup of SL2 (K) we require in addition that ad = 1.
The following exercise provides a supply of Wiegold subgroups in 2-dimensional case, cf. [La, Exercise
23,p.547].
Exercise 12 Let K be a field in which every quadratic polynomial has a root. Let B be the Borel subgroup
of GL2 (K) or of SL2 (K). Show that B is Wiegold.
Solution. We have to show that every matrix A = ac db ∈ GL2 (K) is conjugate to a matrix in
0 1
X
B. If b = 0 we conjugate A by −1
by the matrix X = ( x1 10 ) gives
0 . If b 6= 0 the conjugation A
(AX )2,1 = bx2 + (d − a)x − c which can be made zero by assumption. We conclude that B is a Wiegold
subgroup (and so GL2 (K) is not IG ).
An example of countable K satisfying condition of the above exercise is the ”quadratic closure” of Q
in C. How many such fields are there?
9
5
Compact Lie groups and their generalizations
As we mentioned above every nonabelian connected compact Lie group is not IG . Here we generalize
this remark to definably compact groups according to [Be].
We consider groups definable in an o-minimal expansion M = (R, <, +, . . .) of a real closed field
(R, <, +, . . .). Classical examples of such groups are the (real) algebraic subgroups of the general linear
group GLn (R), for instance the orthogonal group SOn (R). Less classical examples of definable groups
can be found in [PS] or in [Str].
Theorem 13 If G is a nonabelian definably connected definably compact group in an o-minimal expansion
R of a real closed field R, then for any maximal definable abelian subgroup T of G, G is the union of the
conjugates of T i.e. T is Wiegold subgroup in G. In particular, G is not IG .
This is immediately follows from the main result of [Be].
SL2 (R) is T IG
6
Theorem 14 The group SL2 (R) satisfies the topological invariable generation property relative to Euclidean topology.
Structure of the proof: The theorem will immediately follow from the key lemma 19 in which we prove
that if a closed subgroup G of SL2 (R) is conjugation complete then G = SL2 (R). To prove this lemma we
calculate the eigen-spectrum of SL2 (R) in section 6.1, relying on the description of conjugacy classes of
SL2 (R) given in [Co]. Then we describe the 1-parameter subgroups of SL2 (R) in section 6.2. In section 6.3
we describe spectra of 1-dimensional closed subgroups of SL2 (R) and show that their spectra are (much)
smaller than that of SL2 (R). In section 6.4 we do the same in case of 2-dimensional subgroups. In section
6.5 we prove the main lemma and theorem 14.
6.1
Conjugacy classes in SL2 (R)
Following to [Co] we remark that the conjugacy in SL2 (R) should not be confused with conjugacy in the
larger group GL2 (R).
Theorem 15 ([Co]) Let g ∈ SL2 (R).
1. If (Trg)2 > 4 then g is conjugate to a unique matrix of the form
λ
0
0 1/λ
with |λ| > 1.
10
2. If (Trg)2 = 4 then g is conjugate to exactly one of
1
± I2 , ±
0
1
1
,±
−1
0
1
.
−1
(2)
3. If (Trg)2 < 4 then g is conjugate to a unique matrix of the form
cos θ − sin θ
sin θ cos θ
other than ±I2 .
For a set of (complex) matrices M let Sp(M ) to denote the eigenvalue spectrum of M , i.e. the set
of all eigenvalues of all matrices in M (counting without multiplicity). The following is an immediate
consequence of the above result.
Corollary 16 Sp (SL2 (R)) = R× ∪ S, where R× = R − {0} and S = {z ∈ C : |z| = 1}.
6.2
One-parameter subgroups of SL2 (R)
Recall that a one-parameter subgroup in GLn (R) is a homomorphism R → GLn (R) and it is necessarily of
the form f (t) = eAt where A is the matrix f 0 (0). We describe here the conjugacy classes of one-parameter
subgroups in SL2 (R). For that we need to know the orbits of SL2 (R)-action on its Lie algebra. We set
b : a, b, c ∈ R}
sl2 (R) = {X ∈ M2 (R) | TrX = 0} = { ac −a
b ) ∈ sl (R)
to denote the Lie algebra of the Lie group SL2 (R). For a nonzero 2 × 2 real matrix X = ( ac −a
2
there are three cases:
1) X has nonzero real eigenvalues ±t.
0
Then X is conjugate by some g ∈ GL2 (R) to diagonal matrix ( 0t −t
.) To see that X is also conjugate
−1
t 0
det(g)
0
to ( 0 −t ) by the smaller group SL2 (R) we replace g by
g.
0
1
0
0
Matrices of the form ( 0t −t
) constitute a 1-dimensional Lie subalgebra a = R( 10 −1
) of sl2 (R) called
the split Cartan subalgebra. Exponentiating elements of a we obtain the split Cartan subgroup
et
0
:t∈R ,
A=
0 e−t
which is a 1-parameter subgroup of SL2 (R) corresponding to subalgebra a.
2) X has zero eigenvalue of multiplicity 2.
Then X is conjugate by some g ∈ GL2 (R) to nilpotent matrix of the form ( 00 d0 ), 0 6= d ∈ R. Again we
−1
0 .
can make g to be in SL2 (R) replacing g by matrix det(g)
0
1
11
Matrices of the form ( 00 0t ) constitute a 1-dimensional Lie subalgebra n = R( 00 10 ) of sl2 (R) called
nilpotent subalgebra n. Exponentiating elements of n we obtain the unipotent subgroup
1 R
U =
0 1
which is a 1-parameter subgroup of SL2 (R) corresponding to subalgebra n.
3) X has nonzero purely imaginary conjugate eigenvalues ±θi (θ > 0).
√
This happens if and only if a2 + bc < 0 and then easy calculation shows that θ = −a2 − bc. We
0 1
claim that X is conjugate to θ −1
by a matrix from SL2 (R). There are real linearly independent
0
column vectors u, v such that X(u + iv) = θi(u + iv) = θ(−v + iu) = Xu + iXv, from what we obtain
Xu = −θv, Xv = θu. It follows that a 2-by-2 matrix (u, v) ∈ M2 (R) with columns u, v is invertible. We
0 1
have X(u, v) = (Xu, Xv) = θ(−v, b) = θ(u, v) −1
0 .
It follows that
(u, v)−1 X(u, v) =
0
−1
1
.
0
(3)
We wish to modify g = (u, v) so that it became unimodular. For this purpose we first multiply g by
p
−1
|g|
– then 6.2 remains true; so we may as well assume det(u, v) = ±1. If det(u, v) = −1 then we
0
.
replace g by g 10 −1
0 t
0 t
Matrices of the form ( −t
0 ) constitute a 1-dimensional Lie subalgebra k = R( −t 0 ) of sl2 (R) called
non-split Cartan subalgebra. Exponentiating elements of k we obtain a 1-parameter subgroup K = SO2 (R)
of rotations called the non-split Cartan subgroup:
cos t sin t
:t∈R .
K=
− sin t cos t
Clearly SO2 (R) is a compact subgroup of SL2 (R). In fact, it is a maximal closed subgroup of SL2 (R)
[Gl, Ch.VI, section 1].
6.3
One-dimensional Lie subgroups of SL2 (R) and their spectra
Lemma 17 The eigen-spectrum of every closed 1-dimensional subgroup in SL2 (R) is a proper subset in
Sp(SL2 (R)).
Proof.
Suppose that G is a Lie subgroup of SL2 (R) such that its component G0 is 1-dimensional, so is a
1-parameter subgroup of one of the above types A, U, K.
12
1) In case G0 = A we have
A = G0 E G E N (A),
0 1
where N (A) is the normalizer of A in SL2 (R) and w = ( −1
0 ) is the Weyl element. By Bruhat
decomposition N (A) = A ∪ wA, see [La, p.539]. We conclude that
Sp (G) ≤ Sp (N (A)) = Sp(A) ∪ Sp(wA).
Note that (surprisingly but luckily) Sp(wA) equals
−t
0
e
, t ∈ R = {±i}
Sp
−et
0
and indeed Sp(G) = R0 ∪ {i} is a proper subset of Sp(SL2 (R)).
2) If G0 = U is unipotent, then as easy to see N (U ) = B, where B =
n t
± e0
s
e
−t
o
; t, s ∈ R . Hence
G0 E G E N (U ) = B and clearly Sp (G) ≤ Sp (B) = R0 .
3) Finally, in case G0 = K we have N (K) = K [Gl, Ch.VI, sec. 1] hence Sp (G) ≤ Sp (K) = S.
6.4
2-dimensional Lie subgroups of SL2 (R) and their spectra
Two-dimensional Lie algebra over R is either abelian or solvable. Furthermore a solvable Lie algebra s2
has generators e, f and defining relation [ef ] = e, see [Ja]. Using the above classification of 1-dimensional
subalgebras it is easy to conclude that there no two-dimensional abelian subalgebras in sl2 (R). On the
other hand there is only one (up to conjugacy) subalgebra isomorphic to s2 , namely the Borel subalgebra
o
n t s
; t, s ∈ R .
b=
0 −t
The (connected) component B 0 of the Borel subgroup
et
s
; t, s ∈ R .
B = ±
0 e−t
has b as its Lie algebra. It follows that B 0 is the only (up to conjugacy) 2-dimensional connected subgroup
of SL2 (R). We conclude with
Lemma 18 Let G be a Lie subgroup of SL2 (R) such that its component G0 is a component B 0 of the
Borel subgroup B. Then G ≤ B and Sp (G) = R0 .
13
6.5
Main lemma and theorem 14
Lemma 19 If a closed subgroup G of SL2 (R) is conjugation complete then G = SL2 (R).
Proof. Suppose, on the contrary, that there exists a proper closed and conjugacy complete subgroup G
in SL2 (R). It follows from conjugation completeness that Sp(G) = Sp(SL2 (R)).
By Cartan’s theorem any closed subgroup of GLn (R) is a Lie group. (A short elegant proof can be
found in [Ad], pages 17-19). Hence G is a Lie subgroup of SL2 (R). The connected component G0 of G is
a proper closed normal Lie subgroup of G. Since G is a proper subgroup of 3-dimensional connected Lie
group SL2 (R), we conclude that dim G0 = 0, 1 or 2.
If dim G0 = 0 then G is at most countable discrete subgroup of SL2 (R) and thus its spectrum is at
most countable, so Sp (G) is properly contained in R× ∪ S = Sp (SL2 (R)). Contradiction.
If dim G0 = 1 then G0 is a 1-parameter subgroup of SL2 (R) and by lemma 17 the eigen-spectrum of
G is a proper subset in Sp(SL2 (R)). Contradiction.
Finally if dim G0 = 2 then G0 is the connected component of the Borel subgroup and by lemma 18
the eigen-spectrum of G is a proper subset in Sp(SL2 (R)). Contradiction.
The lemma and thereby theorem 14 are proved.
7
Aut(T ) is T IG
A (simplicial) graph Γ consists of a set of vertices V and a family E of edges which are 2-element subsets
of V . A (simple) path from vertex x to vertex y is a sequence of vertices πx = x0 , x1 , . . . , . . . , xn = y such
that every subset xi , xi+1 is an edge and every two subsequent edges are distinct. The number n ≥ 0 is
the length of π. Define the distance d(x, y) between vertices x and y to be the minimum of the lengths of
simple paths joining them. A loop is a path of positive length such that x0 = xn . A tree is a (simplicial)
graph without loops. The valence or degree of a vertex in a tree T is the number of edges that contain
it. A regular m-tree is a tree where every vertex has fixed valence m. Consider the group Aut(T ) of
automorphisms (=isometries) of T , equipped with the standard pointwise convergence topology. then
Aut(T ) turns into a totally disconnected, locally compact, unimodular topological group. In this section
we prove the following theorem:
Theorem 20 The automorphism group Aut(T ) of an m-regular tree is T IG for every m ≥ 2.
7.1
Elements classification
Let us begin by classifying a single automorphism g of a m-regular tree T . According to [Ti] there are
three disjoint cases to consider:
14
• g is elliptic, i.e. g has a fixed point in V (T ),
• g is inversion, i.e. there is an edge which is flipped by g,
• g is hyperbolic with an axis Ag ; that is to say Ag is a subtree with all vertices of degree 2 and g acts
on Ag by translation by a positive amount |g|. The number |g| is called the translation length of G.
We will denote by E , I , H the classes of elliptics, inversions and hyperbolics, respectively. It is easy
to see that they are all nonempty and pairwise disjoint. Moreover each of these classes is invariant
under conjugation. A useful remark is that conjugation transfers the minimal displacement set, i.e. if
g, h ∈ Aut(T ) and h is hyperbolic with axis Ah then Ag−1 hg = g.Ah . If h is elliptic fixing v ∈ V (T ) then
g −1 hg fixes g.v, and if h is an inversion flipping the edge e ∈ E(T ) then g −1 hg flips g.e.
7.2
The local action by permutation
We denote the pointwise stabilizer in Aut(T ) of a subset M ⊂ T by StabAut(T ) (M ), or simply by Stab(M ).
When v ∈ V (T ), we may simply write Stab(v) to mean Stab({v}). When K ≤ Aut(T ) is a subgroup,
we denote by StabK (M ) the stabilizer of M inside K. For a vertex v ∈ V (T ) and n ≥ 0 we denote by
B(v, n), S(v, n) ⊂ T the ball and the sphere of radius n about v in T respectively.
Let T be a regular tree. Following Burger and Mozes (see [BM]), in order to be able to talk about the
action of Aut(T ) locally, we present the following maps, that take elements in Aut(T ) to permutations
in the groups Sd and Sd−1 , representing the local action of an element g on a given set of vertices
that is preserved by g. Fix a legal coloring c of the edges of T by d colors, that is, c is a function
c : E(T ) → {1, . . . , d} satisfying the condition that for every vertex v ∈ T the colors of the edges adjacent
to v are distinct. For every i ∈ {1 . . . d} fix an isomorphism ψi between the subgroup StabSd (i) ≤ Sd and
the group Sd−1 .
Let n = 1 and let v ∈ V (T ) be any vertex. We define ϕv,1 to be the homomorphism ϕv,1 : Stab(v) → Sd
sending an element g ∈ Stab(v) to the permutation it induces on the colors of the d edges adjacent to v.
Observe that ϕv,1 is onto, and that ker(ϕv,1 ) = Stab(B(v, 1)).
For n > 1, v ∈ V (T ) and u ∈ S(v, n − 1) we define ϕv,n,u : Stab(B(v, n − 1)) → Sd to be the
homomorphism sending g to the permutation it induces on the colors of the edges adjacent to u. Let
e be the edge connecting u to the sphere S(v, n − 2) and let i = c(e). Since e is stabilized by any g ∈
Stab(B(v, n − 1)), the image of ϕv,n,u is the subgroup StabSd (i) ≤ Sd . Composing with the isomorphism
ψi : StabSd (i) → Sd−1 , we can think of ϕv,n,u as a surjective map ϕv,n,u : Stab(B(v, n − 1)) → Sd−1 .
Q
Let ϕv,n : Stab(B(v, n − 1)) → S(v,n−1) Sd−1 be the map g 7→ (ϕv,n,u (g))u∈S(v,n−1) . Observe that
ϕv,n is onto, and that ker(ϕv,n ) = Stab(B(v, n)).
15
7.3
Conjugacy classes in Aut(T )
Let T be a d-regular tree. We follow [GNS] for the description of the conjugacy classes of Aut(T ). For
g ∈ Aut(T ) we define the orbital type of g to be the marked graph of orbits. That is: the orbits space
hgi\T together with a function on the vertices m : V (hgi\T ) → N∪{∞} giving to any orbit its cardinality.
For elliptic elements and inversions, the orbital type is always a marked tree. Two such orbital types are
called equivalent if they are isomorphic as marked trees.
Any automorphisms conjugate to an elliptic element (inversion, hyperbolic) is clearly elliptic (resp.
inversion, hyperbolic) as well. Therefore, conjugacy classes in Aut(T ) can be described separately in each
of these cases.
Lemma 21 Elliptic elements g, g 0 ∈ Aut(T ) are conjugate to each other iff their orbital types are equivalent;
Inversions i, i0 ∈ Aut(T ) are conjugate to each other iff their orbital types are equivalent;
Hyperbolic elements h, h0 ∈ Aut(T ) are conjugate to each other iff they have equal translation length
|h| = |h0 |.
For the full proof of this lemma see [GNS].
7.3.1
Elements of type (n, P)
A partition P is called non-trivial if it contains a set of size larger than 1. Let n = 1 and let P =
{P1 , . . . , Pk } be a non-trivial partition of {1, . . . , d}. We say an elliptic element g ∈ Aut(T ) has type
(1, P) about v ∈ V (T ) if g stabilizes v, and acts on the neighbors {u1 , . . . , ud } of v with orbits sizes
correspond to the partition P.
Let n > 1 and let P = {P1 , . . . , Pk } be a non-trivial partition of {1, . . . , d − 1}. We say an elliptic
element g ∈ Aut(T ) has type (n, P) about v ∈ V (T ) if g stabilizes v, and there exits a vertex u ∈ S(v, n−1)
with neighbors {u1 , . . . , ud−1 } ∈ S(v, n) such that:
• g stabilizes (pointwise) all vertices in B(v, n)\{u1 , . . . , ud−1 }
• g acts on {u1 , . . . , ud−1 } with orbits sizes correspond to the partition P.
An elliptic element g ∈ Aut(T ) is said to be of type (n, P) if it is of type (n, P) about v for some
vertex v ∈ V (T ).
In particular, being of type (n, P) is determined by the existence of a radius-n marked ball in the orbital
type, suitable for n and P. Therefore if type(g) = (n, P) and g, g 0 are conjugate, then type(g 0 ) = (n, P).
Notice further that for any choice of n ≥ 1 and a partition P of {1, . . . , d} (if n = 1) or {1, . . . d − 1} (if
n > 1) there exists an elliptic element g ∈ Aut(T ) of type (n, P).
16
Figure 1: (1, {1, 2, 2})-type element (left); (2, {1, 3})-type element (right)
We will denote by P1 , Pn the automorphisms of type (1, P) for any non-trivial partition P of
{1, . . . , d}, and, respectively, the automorphisms of type (n, P) for any n > 1 and any non-trivial partition
P of {1, . . . , d − 1}.
7.3.2
Spherically-transitive elements
Let s ∈ Aut(T ) be an elliptic element stabilizing a vertex v ∈ V (T ) and assume that the action of hsi
is transitive on all spheres about v. We call such an element v-spherically transitive. An elliptic element
s ∈ Aut(T ) is called spherically transitive if it is v-spherically transitive for some v.
For any v ∈ V (T ) there exists a v-spherically transitive element in Aut(T ). If s is v-spherically
transitive, it follows that the orbital type of s is an infinite ray (x0 , x1 , x2 , . . . ) with marks f (x0 ) = 1,
f (xn ) = d(d − 1)n−1 , n ≥ 1 (that is, the size of the nth -sphere about v). Hence, all spherically transitive
elements in Aut(T ) are conjugate. We will denote by T the set of all spherically transitive automorphisms
of T .
Corollary 22 The sets T , H , P1 , Pn , n ≥ 2 are pairwise disjoint and are invariant under conjugation
in Aut(T ). Let H ≤ Aut(T ) be a conjugation complete subgroup, then H meets each of the following sets
T , H , P1 , Pn . Moreover, H ∩ H contains at least one hyperbolic automorphism of displacement length
1.
7.4
A set of generators for Aut(T )
Lemma 23 Let h be a hyperbolic element of length |h| = 1 and let s be a spherically transitive element.
Then hh, si acts transitively on the vertices of T .
Proof. Let v ∈ V (T ) be a vertex such that s is v-spherically transitive. Denote by Ah the axis stabilized
2
by h, and let d = d(v, Ah ). The elements sh , sh are both spherically transitive. Let v1 , v2 ∈ V (T ) be
2
the fixed points of sh , sh respectively. Notice that v1 = h.v, v2 = h2 .v, hence d(v, v2 ) = 2d + 2 and
17
Figure 2: Vertex transitivity via spherical one.
d(v1 , v2 ) = 2d + 1. Denote by u the neighbor of v having d(u, v2 ) = 2d + 1, then since both u, v1 are of
2
same distance from v2 , there exists a power n such that (sh )n .v1 = u. Denote by s0 the conjugation of
2
s by (sh )n .
It follows that s0 ∈ hs, hi is u-spherically transitive. Let e = (v, u) and denote by Tv , Tu the connected
components in T \{e} of v, u respectively. Next, we show that hs, s0 , hi = hh, si acts transitively on the
vertices of T . Let x ∈ V (T ), it is enough to show that ∃g ∈ hs, s0 , hi such that g.v = x. Denote by
dx = d(x, v). Without loss of generality, can assume that dx is even, otherwise replace x by h.x. We
construct by induction a sequence {gn }n≥0 ∈ hs, hi such that gn .v ∈ Tv and d(v, gn .v) = 2n, see Figure
2. Clearly we need g0 = e. Given gn−1 , observe that d(u, gn−1 .v) = 2n − 1. Use u-sphere transitivity
of s0 to get a power m such that (s0 )m gn−1 .v is in Tu . Since d(u, (s0 )m gn−1 .v) = 2n − 1, we have that
d(v, (s0 )m gn−1 .v) = 2n. Now use v-sphere transitivity of s to get a power k such that sk (s0 )m gn−1 .v is
again in Tv . Notice sk didn’t change the distance of (s0 )m gn−1 .v from v, so we have d(v, sk (s0 )m gn−1 .v) =
2n. Set gn = sk (s0 )m gn−1 . In order to get g.v = x, choose n such that d(v, gn .v) = d(v, x), and use
v-sphere transitivity of s to find a power l such that sl gn .v = x.
Lemma 24 Let v ∈ V (T ) and let K ≤ Stab(v) be a closed subgroup. Suppose that K contains an
element gn,P of type (n, P) about v for every pair (n, P) ∈ {1} × {partitions of {1, . . . d − 1}} ∪ N>1 ×
{partitions of {1, . . . d}}. Then K = Stab(v).
Proof. Let k ∈ Stab(v), we use the fact that the finite permutation group Sd is IG
in order to
construct by induction a sequence kn ∈ H converging to k. Since K is closed, the result will follow. As
S
T = n B(v, n), it is enough to show that for all n kn B(v,n) = k B(v,n) .
Let n = 1, and denote by {v1 , . . . vd } the elements in S(v, 1) (that is, neighbors of v). Since K contains
g1,P for all partitions P of {1, . . . , d}, the subgroup Γ ≤ Sym{v1 , . . . vd } = Sd which is the image of K
under the map ϕv,1 is conjugation complete. As the permutations group Sd is finite, it is in particular
IG , therefore Γ = Sd , so K contains an element acting like k on S(v, 1). Take k1 to be this element.
18
Let n > 1. By induction hypothesis kn−1 ∈ K has that kn−1 B(v,n−1) = k B(v,n−1) .
Enumerate the vertices on the sphere S(v, n − 1) by u1 , . . . um (m = |S(v, n − 1)|). Recall that
each element gn,P of type (n, P) about v in K, stabilizes the entire sphere S(v, n) except the neighbors
{uj1 , . . . , ujd−1 } of some uj ∈ S(v, n − 1). Conjugating by powers of s, which is v-spherically transitive,
j
j
we can get elements gn,P
∈ K (j = 1, . . . m) such that gn,P
is of type (n, P), it stabilizes the entire
sphere S(v, n) except the neighbors of uj , and its orbits among the neighbors of uj have sizes according
j
to P. Let uj ∈ S(v, n − 1), since K contains an element gn,P
for every partition P of {1, . . . , d − 1},
we have that the subgroup Γ ≤ Sd−1 which is the image of K under the map ϕv,n,uj is conjugation
complete. Again, as the permutation group Sd−1 is finite, it is IG and therefore Γ = Sd−1 . In particular,
we can find {gj }j=1...m in K such that g j stabilizes the entire ball B(v, n) except the neighbors of uj ,
and g j {uj ...uj
1
d−1 }
−1
= kn−1
k {uj ...uj
1
d−1 }
. Observe that {g j } pairwise commute. Define kn be the product
kn = kn−1 g 1 g 2 . . . g m , then kn B(v,n) = k B(v,n) .
7.5
Proof of theorem 20
We prove the following rephrasement of theorem 20, based on lemma 23 and lemma 24
Theorem 25 Every closed conjugation complete
subgroup H of the automorphism group Aut(T ) of an
m-regular tree T , (m ≥ 2), coincides with Aut(T ).
Proof. Suppose the assertion of the theorem is false and let H < Aut(T ) be a closed Wiegold subgroup.
By corollary 22, H contains at least one hyperbolic element h of translation length |h| = 1, and at least
one elliptic element s which is v-spherically transitive element for a vertex v ∈ V (T ). Therefore by lemma
23 H is vertex-transitive. Using again corollary 22, we conclude that H contains elliptic elements gn,P of
all possible types (n, P). Using the transitivity of H, we can conjugate each element gn,P in H such that
it becomes of type (n, P) about v. We will denote StabH (v) ≤ StabAut (T )(v) briefly by K. The subgroup
K is closed, since it is equal to the intersection H ∩ StabAut(T ) (v). Hence K satisfies the assumptions of
lemma 24 and therefore K = StabAut(T ) (v). We have got that H is vertex transitive and contains the
full vertex-stabilizer group StabAut(T ) (v). It follows that H = Aut(T ).
8
8.1
Uncountably many countable non-IG -groups
Laws in groups
±1
A non-empty word w = w (x1 , . . . , xn ) (= string of symbols in alphabet x±1
1 , . . . , xn ) is reduced if it
does not contain adjacent symbols of the form xεi x−ε
i , ε = ±1. A non-empty word w = w (x1 , . . . , xn )
is reduced on the tuple of elements (g1 , . . . , gn ) of a group G if w is reduced and every word of the form
19
xm
i does not occur as a subword in w when |m| ≥ |gi |, where |gi | denotes the order of gi ( we of course
assume that |gi | = ∞ when giZ is infinite cyclic). A tuple of elements (g1 , . . . , gn ) of a group G is free
if the group hg1 , . . . , gn i generated by them is a free product of cyclics giZ , i.e. the group hg1 , . . . , gn i,
generated by g1 , . . . , gn , is isomorphic to the free product g1Z ∗ · · · ∗ gnZ under the natural homomorphism
from g1Z ∗ · · · ∗ gnZ to hg1 , . . . , gn i. A word w is said to be a law in a group G if w becomes trivial whatever
values the arguments xi are assigned from G; i.e. w (g1 , . . . , gn ) = 1 for all gi ∈ G.
The proof of the following lemma is routine and is left to the reader.
±1
Lemma 26 A tuple (g1 , . . . , gn ) over a group G is free iff every non-empty word w x±1
1 , . . . , xn , which
is reduced over the tuple (g1 , . . . , gn ), does not vanish on (g1 , . . . , gn ), i.e. w (g1 , . . . , gn ) 6= e. In other
words a tuple (g1 , . . . , gn ) is free iff there are no of non-trivial relations between the elements g1 , . . . , gn .
More generally, consider a free product G[x] = G ∗ xZ of G and the infinite cyclic group xZ . An
element w ∈ G[x] will be called a monomial in variable x over the group G. In case w 6∈ G we say that w is
non-constant. A monomial w induces a map w : G → G, where w(g) is the image of w under the natural
group homomorphism
G[x] → G,
taking x to g and fixing G pointwise.
The subset V (w) = {(g∈ G : w(g) = 1} is called the prin-
cipal algebraic set of G corresponding to w. Every monomial w ∈ G[x] can be written as a word
w = a0 xl1 a1 xl2 · · · am−1 xlm am with ai -s in G. We say that this word expression is reduced over G if
it does not contain a subword x±1 ax∓1 with a in the center of G. If a is central then replacing x±1 ax∓1
by a does not change the principal algebraic set V (w). After finitely many such replacements any word w
can be brought to reduced form (which, of course, could occur to be empty or constant). A (non-empty)
reduced non-constant word w (x) = a0 xl1 a1 xl2 · · · am−1 xlm am over G is said to be a generalized law in a
group G if w becomes trivial whatever values the arguments x are assigned from G; i.e. w (g) = 1 for all
g ∈ G.
The following is the simplified version of the result by Mikhalev and Golubchik, showing that there
is no generalized law in GLm (K) , (m ≥ 2) in case of infinite field K.
Theorem 27 ([GM]) Let K be an infinite field and w (x) = g0 xi0 g1 xi1 · · · gm xim gm+1 be a monomial
over the group GLm (K) , (m ≥ 2). If g1 , . . . , gm are all non-central in GLm (K) then the principal algebraic set V (w) = {g ∈ GLm (K) : w (g) = 1} is a proper subset of GLm (K).
The same result (with the same proof) is true for the groups SLm (K) and PSLm (K).
20
8.2
Rational points
As in [Bo1] we identify implicitly an algebraic group G with the group G (Ω) of its points in a ”universal
domain” Ω, i.e. an algebraically closed extension of infinite transcendence degree over a prime field
(recall that the prime field is either the rational number field Q, or a finite field of prime order Fp ).
Assume now that G is defined over a field K of infinite transcendence degree over a prime field k.
We shall need the following lemma which guaranties that, under some natural conditions, given a family
of proper algebraic subsets Vi (i ∈ N) of G, the set of K-rational points outside of a countable union
∪Vi (i ∈ N) is nonempty. (This is kind of an algebraic analog of Baire category theorem).
Lemma 28 ([Bo1], §2,Lemma 2) Let K be a field of infinite transcendence degree over a prime field.
Let X be an irreducible unirational K-variety. Let L be a finitely generated subfield of K containing a
field of definition of X, and Vi (i ∈ N) a family of proper algebraic subsets of X defined over an algebraic
closure L of L. Then X (K) is not contained in the union of the Vi (i ∈ N).
8.3
Building up free tuples
Lemma 29 (main) Let K be a field of infinite transcendence degree over a prime field k. Let C1 , . . . , Cn
be non-trivial distinct conjugacy classes in the group G = PSLm (K) , where n, m ≥ 2. Suppose that
c1 ∈ C1 , . . . , cn−1 ∈ Cn−1 are such that (c1 , . . . , cn−1 ) is a free tuple. Then there is a representative
cn ∈ Cn such that the tuple (c1 , . . . , cn ) is free.
Proof.
Suppose that c1 , . . . , cn−1 are given and we have to construct cn . Fix some c ∈ Cn . We
seek g ∈ PSLm (K) such that the n-tuple (c1 , . . . , cn ) is free with cn = g −1 cg. We will first find out g
in a larger group G (Ω) = PSLm (Ω) and then apply Borel’s lemma. Let W be the set of all non-empty
words w (x1 , . . . , xn ) which are reduced on the tuple (c1 , . . . , cn−1 , c). Thus W = {w = w (x1 , . . . , xn ) :
m
w is reduced, xm
i does not occur in w when |m| ≥ |ci | and xn does not occur in w when |m| ≥ |c| }.
Let w = w (x1 , . . . , xn ) ∈ W . The replacement x1 7→ c1 , . . . , xn−1 →
7 cn−1 , xn 7→ x−1 cx gives rise
to the monomial w c1 , . . . , cn−1 , x−1 cx over G in variable x, which is not necessarily reduced over G.
−1
Indeed, for every maximal occurrence of xm
cx contains a
n in w the result of substitution xn 7→ x
m
m
monomial x−1 cx
which is clearly not reduced. The replacement x−1 cx
by x−1 cm x is a non-empty
monomial w c1 , . . . , cn−1 , x−1 cx red with variable x over G and this monomial reduced over G. The set
G(w)
g ∈ G (Ω) : w c1 , . . . , cn−1 , g −1 cg = 1
=
g ∈ G (Ω) : w c1 , . . . , cn−1 , g −1 cg red = 1
=
(4)
(5)
is an algebraic subset of G (Ω) given by w c1 , . . . , cn−1 , x−1 cx red = 1. It is defined over the field L =
k (c1 , . . . , cn−1 , c) which is finitely generated over the prime subfield k and contains the field of definition
21
k of G. The irreducible components G (w)1 , . . . , G (w)nw of G (w) are defined over the separable closure
Ls [Bo1, §12.3]. Each G (w) is a proper subset of G (Ω) since otherwise w c1 , . . . , cn−1 , x−1 cx red = 1 is
a non-empty generalized law on G = PSLm (Ω), which contradicts Theorem 27. The family of algebraic
subsets {G (w)i : w ∈ W } is countable. Finally, the group G, being reductive, is unirational over k [Bo2,
section 18.2]. All above means that we can apply Borel’s lemma 28 with X = G, L = k (c1 , . . . , cn−1 , c)
and (Vi ) the family of all irreducible components of all G (w) , w ∈ W .
We conclude that G (K) is not contained in the union ∪w∈W G (w). The complement
G (Ω) − ∪ G(w)
w∈W
(6)
consists of g ∈ G (Ω) such that the n-tuple c1 , . . . , cn−1 , g −1 cg is free. Hence G (K) contains a point g
outside of union ∪w∈W G (w), so the corresponding tuple c1 , . . . , cn−1 , g −1 cg is free and K-rational.
8.4
When is PSLm (K) an IG ?
Theorem 30 If a field K is a countable and has infinite transcendence degree over the prime field, then
the group PSLm (K) is not IG for m ≥ 2.
Proof. The group G = PSLm (K) is countable. Then G − {e} = ∪Ci , (i ∈ N) , where Ci are all nontrivial distinct conjugacy classes. Inductively applying Lemma 29, we find out representatives ci so that
the subgroup hci : i ∈ Ni is isomorphic to the free product C = ∗ cZ
i : i ∈ N . By construction, C meets
all conjugacy classes and is non-trivial. The set S = {ci } is conjugacy complete. Moreover, hci : i ∈ Ni is
Q
N
a proper subgroup of G because its abelianization is C ab = i (Z/ |ci |) (where |ci | = 0 if ci is of infinite
order ) is non-trivial whereas G is a perfect group, even simple.
Theorem 31 There are uncountably many isomorphism types of non-IG groups of the form PSLm (K),
where m ≥ 2 and K is a countable field of infinite transcendence degree over Q.
Proof. Let P denote the set of all primes in N and let R ⊆ P be any subset. Consider the fields Q
√
R
generated by square roots of all the primes in R. Each of these fields is countable, and it’s not too hard to
√
see that they are pairwise non-isomorphic. Now adjoin to each Q
R countably many indeterminates
√
X = {Xi : i ∈ N} , i.e. consider the field of rational functions Q
R, X in indeterminates X. The fields
√
Q
R, X are all countable, pairwise non-isomorphic and have infinite transcendence degree over the
prime field Q.
By the result of Schreier and van der Waerden [SW] the groups PSLm (K) , PSLm0 (K 0 ) , (m, m0 ≥ 2)
can be isomorphic only when m = m0 , excluding the case of PSL2 (F7 ) ' PSL3 (F2 ). Furthermore,
if m = m0 > 2, the isomorphism is possible only when the fields K, K 0 are isomorphic. The same
22
is true when m = m0 = 2 , excluding the case {K, K 0 } = {F4 , F5 }. We conclude that the groups
√
R, X ,m ≥ 2, R ⊆ P , are pairwise non-isomorphic.
PSLm Q
Similar construction can be carried over when the prime field is finite.
9
Open problems
There are several open problems, concerning IG .
1. Conjecture: A real semi-simple Lie group G is T IG when it is of rank 1 and G is not T IG otherwise.
2. : Is an infinite compact group T IG if and only if it is connected abelian by profinite [KLS2]?
3. : Is SLn (Z), (n ≥ 3) IG [KLS2]?
4. : Is SLn (Q), (n ≥ 2) IG [KLS2]?
5. : Is SL2 (R) IG (T.Gelander)?
6. : Is every IG -linear group virtually solvable [KLS2]?
7. Is a finite index subgroup of an IG group necessarily IG (Wiegold)?
8. “I think it is worth noting that every group outside IG that I know satisfies no non-trivial law,
and it would be nice to have an example with a non-trivial law. For example, can a Tarski like
group of finite exponent be outside IG ?” [Wi77].
References
[Ad] J.F. Adams. Lectures on Lie groups. Benjamin, 1969. Also available as Midway reprint, University
of Chicago, 1982.
[Ax] Sheldon Axler, Linear Algebra Done Right, Springer, 2015.
[Be]
A.Berarducci, Zero-Groups And Maximal Tori, Logic Colloquium 2004, Lecture Notes in Logic 29,
pages 33–47, Edited By A. Andretta, K.Kearnes, and D. Zambella, Cambridge University Press
City: Year: 2008.
[BM] Mark Burger and Shahar Mozes, Groups acting on trees: from local to global structure, Publications
mathematiques de l’I.H.E.S, tome 92 (2000), p.113-150.
[Bo1] Borel A., On free subgroups of semisimple groups, Enseign. Math. (2) 29 (1983), 151–164.
23
[Bo2] Borel, Armand (1991) [1969], Linear algebraic groups, Graduate Texts in Mathematics, 126 (2nd
ed.), Berlin, New York: Springer-Verlag.
[Ch] N. Chebotarev, Die Bestimmung der Dichtigkeit einer Menge von Primzahlen, welche zu einer
gegebenen Substitutionsklasse gehören, Math. Ann. 95, 1926, 191 228.
[Co] Keith
Conrad,
Decomposing
SL2 (R),
http://www.math.uconn.edu/kconrad/blurbs/
grouptheory/SL(2,R).pdf.
[Co] Brian Conrad, Reductive groups over fields (version of august 11, 2017), Lectures by Brian Conrad,
notes by Tony Feng, http://math.stanford.edu/~conrad/249BW16Page/handouts/249B_2016.
pdf.
[Di]
J.D. Dixon, Random sets which invariably generate the symmetric group, Discrete Math. 105 (1992)
25–39.
[Du] J. J. Duistermaat, Fourier integral operators, Modern Birkhäuser Classics, Birkhäuser/Springer,
New York, 1996.
[Ge] T. Gelander, Convergence groups are not invariably generated, Int. Math. Res. Not. 2015, no. 19,
9806–9814.
[GM] T. Gelander, C. Meiri, The Congruence Subgroup Property does not imply Invariable Generation,
Int. Math. Res. Not., to appear arXiv:1505.06881.
[GGJ] Tsachik Gelander, Gili Golan, and Kate Ushchenko Invariable generation of Thompson groups,
arXiv:1611.08264v2 [math.GR], 28 Nov 2016.
[Gl]
Shmuel Glasner, Proximal flows, (Lecture notes in mathematics, 517) Springer, 1976.
[GNS] P. Gawron, V. V. Nekrashevich, and V. I. Sushchanskii, Mathematical Notes, Vol. 65, No. 6, 1999.
[GM] I. Z. Golubchik and A. V. Mikhalev, Generalized group identities in classical groups, Journal of
Soviet Mathematics, 1984, 27:4, 2904–2918.
[Ja]
Nathan Jacobson, Lie algebras, Interscience Tracts on Pure and. Applied Mathematics, no. 10.
Interscience Publishers, New York, 1962. 9+331 pp.
[Jo]
C. Jordan, Recherches sur les substitutions, J. Liouville 17 (1872), 351–367 (= Oe. I. 52).
[KLS21] William M.Kantor, Alexander Lubotzky and Aner Shalev, Invariable generation and the Chebotarev invariant of a finite group, Journal of Algebra, Volume 348, Issue 1, 15 December 2011,
Pages 302–314.
24
[KLS2] William M.Kantor, Alexander Lubotzky and Aner Shalev,
Invariable generation of infinite
groups. J. of Algebra 421(2015), 296–310.
[KM] Alexei I. Kostrikin, Yu. I. Manin, Linear Algebra and Geometry, (Algebra, Logic and Applications),
CRC Press, 1989.
[KZ] Emmanuel Kowalski and David Zywina, The Chebotarev Invariant of a Finite Group, Experimental
Mathematics, 21(1):38–56, 2012, DOI: 10.1080/10586458.2011.565261
[Lu]
A. Lucchini, The Chebotarev invariant of a finite group: A conjecture of Kowalski and Zywina,
arXiv:1509.05859v2.
[Lu]
Andrea Lucchini and Gareth Tracey, An upper bound on the Chebotarev invariant of a finite group
Israel Journal of Mathematics, April 2017, Volume 219, Issue 1, pp 449–467.
[La]
Serge Lang, Algebra, revised third edition, Springer, New York, 2002.
[Os10] Osin, Denis, Small cancellations over relatively hyperbolic groups and embedding theorems, Annals
of Mathematics, 172 (2010), 1–39.
[PS] Y. Peterzil and C. Steinhorn, Definable compactness and definable subgroups of O-minimal groups,
J. London Math. Soc 59 (1999), 769–786.
[Se]
Jean-Pierre Serre, On a Theorem Of Jordan, BULLETIN (New Series) Of The American Mathematical Society Volume 40, Number 4, Pages 429–440, (2003).
[SW] Schreier O., van der Waerden B. L., Die Automorphismen der projektiven Gruppen, Abh. Math.
Sem. Hamburg Univ., B (1928), 303–322.
[St]
R. Steinberg, Endomorphisms of Linear Algebraic Groups, Mem. Amer. Math. Soc., vol.80, Amer.
Math. Soc., Providence, RI, 1968.
[Str] A. Strzebonski, Euler characteristic in semialgebraic and other O-minimal structures, J. Pure and
Applked Algebra, 86 (1994), 173–201.
[Ti]
J. Tits, Sur le groupe d’automorphismes d’un arbre, in ”Essays on Topology and Related Topics“,
Springer Berlin Heidelberg (1970), 188211.
[Wi76] J. Wiegold, Transitive groups with fixed-point-free permutations, Arch. Math. (Basel), 27, (1976)
473–475.
[Wi77] J. Wiegold, Transitive groups with fixed-point-free permutations.II, Arch. Math. (Basel) 29 (1977)
571–573.
25
| 4math.GR
|
arXiv:1803.07177v1 [math.GR] 19 Mar 2018
Flat manifolds with homogeneous holonomy
representation
R. Lutowski
Abstract
We show that a rational holonomy representation of any flat manifold except torus must have at least two non-equivalent irreducible
subrepresentations.
1
Introduction
Let n ∈ N. Let Γ ⊂ Isom(Rn ) be a crystallographic group, i.e. a discrete and
co-compact subgroup if isometries of the n-dimensional euclidean space Rn .
By Bieberbach theorems Γ fits into the following short exact sequence
π
0 −→ M −→ Γ −→ G −→ 1
(1.1)
where G is a finite group and M is a maximal abelian normal subgroup of
Γ. Moreover, M is free abelian group of rank n and it admits a structure of
a faithful G-module, defined by the conjugations in Γ. The representation
ϕ : G → GL(M) which corresponds to this module is called the integral
holonomy representation of Γ.
If Γ is torsion-free we call it a Bieberbach group. In this case the orbit
space X = Rn /Γ is a closed Riemannian manifold with sectional curvature
equal to zero, a flat manifold for short. Moreover Γ, which is isomorphic to
the fundamental group of X, determines the manifold up to affine equivalence. It is worth to notice that the holonomy representation of X is equivalent to ϕ over the field of real numbers.
We will say that G-module M is homogeneous if QG-module Q ⊗Z M has
only one homogeneous component, i.e. all of its irreducible submodules are
isomorphic. In the case when Γ is torsion-free and M is homogeneous, we
will call X a homogeneous flat manifold.
In this paper we prove the following theorem:
Theorem. The only homogeneous flat manifolds are flat tori.
1
The theorem is a generalization of the result presented in [4] where the
authors consider the case of Q ⊗Z M being irreducible.
2
Lattices
Remark 2.1. Since we consider tensor product over integers only, we will
omit the subscript Z from now on.
Let G be a finite group. For the convenience of our further considerations
we introduce a notion of a G-lattice, i.e. a G-module which is a free Z-module
of finite rank. A G-submodule M ′ of M is called a sublattice if it is pure as Zsubmodule. As a consequence of this definition, we will call M an irreducible
lattice if its only submodule of lower rank is 0 or, equivalently, Q ⊗ M is a
simple QG-module. This approach allows us to build a descending chain
M = Mk ⊃ Mk−1 ⊃ . . . ⊃ M0 = 0
of G-modules such that Mi−1 is a maximal sublattice of Mi and hence
Mi /Mi−1 is an irreducible G-lattice, for i = 1, . . . , k. Hence we can speak
about composition series and composition factors for lattices (see [2, §73] for
more detailed description).
3
Reduction
As in the previous section, let G be a finite group. For any G-lattice M
denote by Irr(G, M) the set of irreducible characters arising from constituents
of C ⊗ M. Note that if M is a homogeneous G-lattice and S is a composition
factor of M then Irr(G, M) = Irr(G, S). We immediately get the following
generalization of [4, Lemma 2.1]:
Lemma 3.1. Let M be a homogeneous G-lattice. Let p be a prime and Zp
– the ring of p-adic integers. Suppose that Zp ⊗ M contains an indecomposable direct summand in the principal Zp G-block. Then every irreducible
constituent of C ⊗ M is in the principal p-block of G.
Let Γ be a crystallographic group which fits into the sequence (1.1) and
let α ∈ H 2 (G, M) be the cohomology class corresponding to this extension.
Then Γ is torsion-free if and only if α is special, i.e. the restriction of α to
every subgroup of G of prime order is non-zero (see [1, Theorem III.2.1] for
example).
Note that in the [4, Lemma 2.2] the irreducibility of a lattice was only
needed to use [4, Lemma 2.1]. Hence we can generalize it to the homogeneous
case:
2
Lemma 3.2. Let M be a homogeneous G-lattice such that H 2 (G, M) contains
a special element. Let S denote a simple component in the socle Soc(G) (the
product of minimal normal subgroups) of G. Then we have:
(a) If ϑ ∈ Irr(G, M), then ϑ is in the principal p-block for every prime p
dividing |G|.
(b) If ψ ∈ Irr(S, M), then ψ is in the principal p-block for every prime p
dividing |G|.
(c) Let p be a prime dividing |S| such that a Sylow p-subgroup of S is
cyclic. Then there is Θ ∈ Irr(S, M) which has the following position on
the oriented Brauer tree:
(3.1)
1S
Θ
Denote by S(G) the set of those complex irreducible characters χ of G
such that for every prime p dividing |G|:
(1) χ is in the principal p-block;
(2) If Sylow p-subgroup of G is cyclic, then χ has the position given by the
position of Θ on the Brauer tree (3.1) of the principal p-block.
The base finding of [4], giving its main result, may be stated as follows:
Proposition 3.3 ([4, Section 3]). Let S be a non-abelian finite simple group.
Then S(S) = ∅.
Since for every finite group its socle is a product of simple groups, we
immediately get the following corollary.
Corollary 3.4. Let M be a homogeneous G-lattice such that H 2 (G, M) contains a special element. Then Soc(G) is a product of elementary abelian
groups.
3
4
Proof of the Theorem
In the case of non p-groups of the proof we follow [4, Section 4].
Let n ∈ N, Γ ⊂ Isom(Rn ) be a Bieberbach group which fits into the short
exact sequence (1.1) and X = Rn /Γ be a homogeneous flat manifold, i.e. M
is a homogeneous G-lattice. Denote by χ its character.
Assume that G is a non-trivial group. Then Soc(G) is non-trivial. Now
take any prime p dividing | Soc(G)| and a Sylow p-subgroup N of Soc(G).
By Corollary 3.4 N is characteristic in the socle, hence it is normal in G.
If q 6= p is another prime dividing the order of G then, by Lemma 3.2,
every ϑ ∈ Irr(G, M) is in the principal q-block and so, by [3, Lemma IV.4.12],
has N ⊂ Oq′ (G) in its kernel. But if k is the composition length of M then
X
χ = k
ϑ
ϑ∈Irr(G,M )
has N in its kernel and M is not a faithful G-module.
On the other hand, if G is a p-group and C is a cyclic and normal subgroup
of G then, by the Clifford’s theorem (see [2, Theorem 49.2]), the restriction
χ′ of χ to C equals
χ′ = ϕ1 + ϕ2 + . . . + ϕl
where ϕ1 , . . . , ϕl are characters arising from irreducible components of the
QC-module Q ⊗ M. Moreover, they are pairwise conjugated, i.e.
∀1≤i,j≤l ∃g∈G ∀c∈C ϕi (c) = ϕj (gcg −1).
(4.1)
By [5, Theorem 7.1] if H 2 (C, M) 6= 0 then there exists 1 ≤ i ≤ l s.t. ϕi is
the trivial character. But then the formula (4.1) shows that ϕi is trivial for
every 1 ≤ i ≤ l. And again we get that M is not a faithful G-module.
The above considerations show that G must be the trivial group and
hence X is a flat n-dimensional torus.
Acknowledgement
The author would like to thank Gerhard Hiss and Andrzej Szczepański for
helpful discussions.
References
[1] L.S. Charlap. Bieberbach groups and flat manifolds.
Springer-Verlag, New York, 1986.
4
Universitext.
[2] C.W. Curtis and I. Reiner. Representation theory of finite groups and
associative algebras. Pure and Applied Mathematics, Vol. XI. Interscience
Publishers, a division of John Wiley & Sons, New York-London, 1962.
[3] Walter Feit. The representation theory of finite groups, volume 25 of
North-Holland Mathematical Library. North-Holland Publishing Co.,
Amsterdam-New York, 1982.
[4] Gerhard Hiss and Andrzej Szczepański. On torsion free crystallographic
groups. J. Pure Appl. Algebra, 74(1):39–56, 1991.
[5] Saunders MacLane. Homology. Classics in Mathematics. Springer-Verlag
Berlin Heidelberg, reprint of the 1975 edition, 1995.
5
| 4math.GR
|
On The Complexity of Bounded Time and
Precision Reachability for Piecewise Affine
Systems?
Hugo Bazille3 , Olivier Bournez1 , Walid Gomaa2,4 , and Amaury Pouly1
arXiv:1601.05353v3 [cs.CC] 17 Jan 2017
1
2
École Polytechnique, LIX, 91128 Palaiseau Cedex, France
Egypt Japan University of Science and Technology, CSE, Alexandria, Egypt
3
ENS Cachan/Bretagne et Université Rennes 1, France
4
Faculty of Engineering, Alexandria University, Alexandria, Egypt
Abstract. Reachability for piecewise affine systems is known to be undecidable, starting from dimension 2. In this paper we investigate the
exact complexity of several decidable variants of reachability and control questions for piecewise affine systems. We show in particular that
the region-to-region bounded time versions leads to NP-complete or coNP-complete problems, starting from dimension 2. We also prove that a
bounded precision version leads to P SP ACE-complete problems.
1
Introduction
A (discrete time) dynamical system H is given by some space X and a function f : X → X. A trajectory of the system starting from x0 is a sequence
x0 , x1 , x2 , . . . etc., with xi+1 = f (xi ) = f [i+1] (x0 ) where f [i] stands for ith iterate of f . A crucial problem in such systems is the reachability question: given a
system H and R0 , R ⊆ X, determine if there is a trajectory starting from a point
of R0 that falls in R. Reachabilty is known to be undecidable for very simple
functions f . Indeed, it is well-known that various types of dynamical systems,
such as hybrid systems, piecewise affine systems, or saturated linear systems,
can simulate Turing machines, see e.g., [12,9,14,15].
This question is at the heart of verification of systems. Indeed, a safety property corresponds to the determination if there is a trajectory starting from some
set R0 of possible initial states to the set R of bad states. The industrial and
economical impact of having efficient computer tools, that are able to guarantee that a given system does satisfy its specification, have indeed generated very
important literature. Particularly, many undecidability and complexity-theoretic
results about the hardness of verification of safety properties have been obtained
in the model checking community. However, as far as we know, the exact complexity of natural restrictions of the reachability question for systems as simple
as piecewise affine maps are not known, despite their practical interest.
Indeed, existing results mainly focus on the frontier between decidability
and undecidability. For example, it is known that reachability is undecidable
?
This work was partially supported by DGA Project CALCULS.
for piecewise constant derivative systems of dimension 3, whereas it is decidable
for dimension 2 [1]. It is known that piecewise affine maps of dimension 2 can
simulate Turing machines [13], whereas the question for dimension 1 is still open
and can be related to other natural problems [2,3,5]. Variations of such problems
over the integers have recently been investigated [6].
Some complexity facts follow immediately from these (un)computability results: for example, point-to-point bounded time reachability for piecewise affine
maps is P -complete as it corresponds to configuration to configuration reachability for Turing machines.
However, there remain many natural variants of reachability questions for
whose complexity have not yet been established.
For example, in the context of verification, proving the safety of a system can
often be reduced to a reachability question of the form point-to-region or regionto-region reachability. These variants are more general questions than point-topoint reachability. Their complexities do not follow from existing results.
In this paper we choose to restrict to the case of piecewise affine maps and
we consider the following natural variant of the problem.
BOUNDED TIME: we want to know if region R is reached in less than some
prescribed time T , with f assumed to be continuous.
FIXED PRECISION: given some precision ε = 2−n , we want to know if region
R is reached by f˜ε such that f˜ε (x) = b f (x)
ε cε. This corresponds to truncating f
at precision 2−n on each coordinate, or equivalently assuming that computations
happen at some precision not more that 2−n , for some given n. In other words,
one wants to know if R is reached by the dynamics where precision operation is
applied at each iteration.
Remark 1. We consider a version where everything is rounded downwards to a
multiple of epsilon. Variants could also be considered, with for example closest
or upper rounding. This would not change the complexity.
Remark 2. A variant could also be chain reachability: Deciding the existence of
sequence xn from initial region to target region such that at any intermediate
step i, ||xi+1 − f (xi )|| ≤ . We do not know the complexity of variants based on
this idea.
Remark 3. We consider piecewise affine maps over the domain [0, 1]d , The case
of integer domains has been studied in [6] and turns out to be quite different. We
also assume f to be continuous. This makes the hardness result more natural.
In an orthogonal way, control of systems or constructions of controllers for
systems often yield to dual questions. Instead of asking if some trajectory reaches
region R, one wants to know if all trajectories reach R. The questions of stability,
mortality, or nilpotence for piecewise affine maps and saturated linear systems
have been established in [7]. Still in this context, the complexity of the problem
when restricting to bounded time or fixed precision is not known.
This paper provides an exact characterization of the algorithmic complexity
of those two types of reachability for discrete time dynamical systems. Let P AFd
denote the set of piecewise-affine continuous functions over [0, 1]d . At the end
we get the following picture.
Remark 4. Notice that we expect in several statements time to be given in unary,
in order to get a completeness result. We do not know about the complexity of
the variants where time would be given in binary.
Problem: REACH-REGION
Inputs: a continuous P AFd f and two regions R0 and R in dom(f )
Output: ∃x0 ∈ R0 , t ∈ N, f [t] (x0 ) ∈ R?
Theorem 5 ([13]). Problem REACH-REGION is undecidable (and is recursively
enumerable-complete).
Problem: CONTROL-REGION
Inputs: a continuous P AFd f and two regions R0 and R in dom(f )
Output: ∀x0 ∈ R0 , ∃t ∈ N, f [t] (x0 ) ∈ R?
Theorem 6 ([7]). Problem CONTROL-REGION is undecidable (and is co-recursively
enumerable complete) for d > 2.
Problem: REACH-REGION-TIME
Inputs: a time T ∈ N in unary, a continuous P AFd f and two regions R0 and
R in dom(f )
Output: ∃x0 ∈ R0 , ∃t 6 T, f [t] (x0 ) ∈ R?
Theorem 7 (Theorems 31 and 36). REACH-REGION-TIME is NP-complete for
d > 2.
Problem: CONTROL-REGION-TIME
Inputs: a time T ∈ N in unary, a continuous P AFd f and two regions R0 and
R in dom(f )
Output: ∀x0 ∈ R0 , ∃t 6 T, f [t] (x0 ) ∈ R?
Theorem 8 (Theorems 37 and 38). CONTROL-REGION-TIME is coNP-complete
for d > 2.
Problem: REACH-REGION-PRECISION
Inputs: a continuous P AFd f , two regions R0 and R in dom(f ) and ε = 2−n ,
n given in unary
Output: ∃x0 ∈ R0 , t ∈ N, (f˜ε )[t] (x0 ) ∈ R?
Problem: CONTROL-REGION-PRECISION
Inputs: a continuous P AFd f , two regions R0 and R in dom(f ) and ε = 2−n ,
n given in unary
Output: ∀x0 ∈ R0 , ∃t ∈ N, (f˜ε )[t] (x0 ) ∈ R?
Theorem 9 (Theorems 40 and 40). REACH-REGION-PRECISION is P SP ACEcomplete for d > 2.
Theorem 10 (Theorems 40 and 40). CONTROL-REGION-PRECISION is P SP ACEcomplete for d > 2.
All our problems are region-to-region reachability questions, and requires
new proof techniques.
Indeed, classical tricks to simulate a Turing machine using a piecewise affine
maps encode a Turing machine configuration by a point, and assume that all
the points of the trajectories encode (possibly ultimately) valid Turing machines
configurations.
This is not a problem in the context of point-to-point reachability, but this
can not be extended to region-to-region reachability. Indeed, a (non-trivial) region consists mostly in invalid points: almost all points do not correspond to
encoding of Turing machines for all the encodings considered in [13,7].
In order to establish hardness results, the trajectories of all (valid and invalid)
points must be carefully controlled. This turns out not to be easily possible using
the classical encodings.
Let us insist on the fact that we restrict our results to continuous dynamics.
In this context, this is an additional source of difficulties: Dealing with points
and trajectories not corresponding to valid configurations or evolutions.
A short version of this paper has been presented at the conference “Reachability Problems 2014” [4]. The current journal version contains full proofs for all
statements, and is also providing new results: bounded precision variants (problems REACH-REGION-PRECISION and CONTROL-REGION-PRECISION) were not considered in short version [4].
2
2.1
Preliminaries
Notations
The set of non-negative integers is denoted N and the set of the first n naturals
is denoted Nn = {0, 1, . . . , n − 1}. For any finite set Σ, let Σ ∗ denote the set of
finite words over Σ. For any word w ∈ Σ ∗ , let |w| denote the length of w. Finally,
let λ denote the empty word. If w is a word, let w1 denote its first character,
w2 the second one and so on. For any i, j ∈ N, let wi...j denote the subword
wi wi+1 . . . wj . For any σ ∈ Σ, and k ∈ N, let σ k denote the word of length k
where all symbols are σ. For any function f , let f E denote the restriction of
f to E and let dom(f ) denote the domain of definition of f .
2.2
Piecewise affine functions
Let I denote the unit interval [0, 1]. Let d ∈ N. A convex closed polyhedron in
the space I d is the solution set of some linear system of inequalities:
Ax ≤ b
(1)
with coefficient matrix A and offset vector b. Let P AFd denote the set of
piecewise-affine continuous functions over I d : That is to say, any f : I d → I d
in P AFd , f satisfies:
• f is continuous,
• there exists a sequence (Pi )1≤i≤p of convex closed polyhedra with nonempty
Sp
interior such that fi = f Pi is affine, I d = i=1 Pi and P̊i ∩ P˚j = ∅ for
i 6= j, where P̊ denotes the interior of P .
In the following discussion we will always assume that any polyhedron P can
be defined by a finite set of linear inequalities, where all the elements of A and
b in (1) are all rationals. A polyhedron over which f is affine will also be called
a region.
2.3
Decision problems
In this paper, we will show hardness results by reduction from known hard
problems. We give the statement of these latter problems in the following.
Problem: SUBSET-SUM
Inputs: a goal B ∈ N andP
integers A1 , . . . , An ∈ N.
Output: ∃I ⊆ {1, . . . , n}, i∈I Ai = B?
Theorem 11 ([8]). SUBSET-SUM is NP-complete.
Problem: NOSUBSET-SUM
Inputs: a witness B ∈ N and
P integers A1 , . . . , An ∈ N.
Output: ∀I ⊆ {1, . . . , n}, i∈I Ai 6= B?
Theorem 12. NOSUBSET-SUM is coNP-complete.
Proof. Basically the same proof as Theorem 11 [8].
t
u
Problem: LINSPACE-WORD
Inputs: A Linear Bounded Automaton (i.e. a one-tape TM that does not use
any space besides the input) M and a word w ∈ Σ ∗ .
Output: does M accept w?
Theorem 13 (see e.g. [10]). LINSPACE-WORD is PSPACE-complete.
3
Hardness of Bounded Time Reachability
In this section, we will show that REACH-REGION-TIME is an NP-hard problem
by reducing SUBSET-SUM to it.
3.1
Solving SUBSET-SUM by iteration
We will now show how to solve the SUBSET-SUM problem by iterating a function. Consider an instance I = (B, A1 , . . . , An ) of SUBSET-SUM. We will need to
introduce some notions before defining our piecewise affine function. Our first
notion is that of configurations, representing partial summation of the number
for a given choice of I.
Remark 14. Without loss of generality, we will only consider instances where
Ai 6 B, for all i. Indeed, if Ai > B, it will never be an element of a solution to
the instance and so we can simply remove this variable from the problem. This
ensures that Ai < B + 1 in everything that follows.
Definition 15 (Configuration). A configuration of I is a tuple (i, σ, εi , . . . , εn )
where i ∈ {1, . . . , n + 1}, σ ∈ {0, . . . , B + 1}, εj ∈ {0, 1} for all i ≤ j. Let CI be
the set of all configurations of I.
The intuitive understanding of a configuration, made formal in the next definition, is the following: (i, σ, εi , . . . , εn ) represents a situation where after having
summed a subset of {A1 , . . . , Ai−1 }, we got a sum σ and εj is 1 if and only if we
are to pick Aj in the future.
Definition 16 (Transition function). The transition function TI : CI → CI ,
is defined as follows:
(
(i, σ)
if i = n + 1
TI (i, σ, εi , . . . , εn ) =
(i + 1, min (B + 1, σ + εi Ai ) , εi+1 , . . . , εn ) otherwise
It should be clear, by definition of a subset sum that we have the following
simulation result.
Lemma 17. For any configuration c = (i, σ, εi , . . . , εn ) and k ∈ {0, . . . , n + 1 −
i},
[k]
i+k−1
εj Aj , εi+k , . . . , εn )
TI (c) = (i + k, min B + 1, σ + Σj=i
t
u
Proof. By induction.
A consequence of this simulation is that we can reformulate satisfiability in
terms of reachability.
Lemma 18. I is a satisfiable instance ( i.e., admits a subset sum equal to the
target value) if and only if there exists a configuration c = (1, 0, ε1 , . . . , εn ) ∈ CI
[n]
such that TI (c) = (n + 1, B).
Proof. TheP
“only if” direction is the simplest: assume there exists I ⊆ {1, . . . , n}
such
that
i∈I Ai = B. Define εi = 1 if i ∈ I and 0 otherwise. We get that
Pn
ε
A
=
B. Apply Lemma 17 to get that:
i
i
i=1
!
n
X
[n]
TI (1, 0, ε1 , . . . , εn ) = (n + 1, min B + 1, 0 +
ε i Ai )
i=1
= (n + 1, min(B + 1, B)) = (n + 1, B)
The “if” direction is very similar: assume that there exists c = (1, 0, ε1 , . . . , εn )
[n]
such that TI (c) = (n + 1, B). Lemma 17 gives:
[n]
TI (1, 0, ε1 , . . . , εn )
= (n + 1, min B + 1, 0 +
n
X
!
εi Ai )
i=1
Pn
We can easily
Pconclude that i=1 εi Ai = B and thus by defining I = {i | εi =
1} we get that i∈I Ai = B. Hence, I is satisfiable.
t
u
3.2
Solving SUBSET-SUM with a piecewise affine function
In this section, we explain how to simulate the function TI using a piecewise affine function and some encoding of the configurations for a given I =
(B, A1 , . . . , An ). Since the reduction is quite technical, we start with some intuitions. In order to simulate the function TI , we first need to encode configurations
with real numbers. Let c = (i, σ, εi , . . . , εn ) be a configuration, we encode it using two real numbers in [0, 1]: the first one encodes i and σ and the second one
encodes εi , . . . , εn . A simple approach is to encode them as digits of dyadics
numbers, as depicted below:
hci =
i
.
0. 0 . . εi
0.
σ
. . . εn
!
i2−p + σ2−q
.
εi 2−1 + εi+1 2−2 + · · ·
=
In the above encoding, we allocate p bits to i and q − p bits to σ. We simply
need to choose them large enough to accomodate the largest possible value. The
rationale behind this encoding is that it is easy to implement the transition
function fI with a linear function. Note that for technical reasons explained
later, we need to encode the second coordinate in basis β = 5 instead of 2. We
now encode 0 as 0? = 1 and 1 as 1? = 4. Graphically, the action of f is very
simple:
?
if εi = 0 then fI
?
if εi = 1 then fI
0.
i
0. 0 . . . εi
σ
εi+1 . . . ε
n
!
i
.
0. 0 . . εi
σ
εi+1 . . . ε
n
!
0.
=
0.
i+1
0. 0 . . . 0
σ
εi+1 . . . ε
n
!
i+1
0. 0 . . . 0
σ + Ai
εi+1 . . . ε
n
!
0.
=
,
.
For technical reasons, we will in fact split the case εi = 1? into two, depending
on whether σ + Ai becomes greater then B + 1 or not. This is similar to what
we did in Definition 16 with min(B + 1, σ + Ai ):
if σ + Ai > B then fI
σ
i
?
0. 0 . . . 1 εi+1 . . . εn
0.
!
=
i+1
0. 0 . . . 0
0.
B+1
εi+1 . . . ε
n
!
.
We can formulate the “if εi = α” by using regions. Indeed, the set of encodings such that εi = α is
(
i
.
0. 0 . . α
0.
σ
εi+1 . . . ε
n
!
)
: σ ∈ N, εj ∈ {0, 1}
−p −p
i2 , i2 + 2−p−1
.
× αβ −i , (α + 1)β −i
⊂
It is crucial to note that the statement of the problems only allows for polyhedral
regions. This is why in the above equation, we had to overapproximate the
region by intervals on each coordinate. This overapproximation is the root of all
difficulties. Indeed, the region now contains many points that do not correspond
to encodings anymore. We now go to the details of the construction.
Definition 19 (Encoding). Define p = dlog2 (n + 2)e, ω = dlog2 (B + 2)e,
q = p + ω + 1 and β = 5. Also define 0? = 1 and 1? = 4. For any configuration
c = (i, σ, εi , . . . , εn ), define the encoding of c as follows:
hci = i2−p + σ2−q , 0? β −n−1 +
n
X
ε?j β −j
j=i
Also define the following regions for any i ∈ {1, . . . , n+1} and α ∈ {0, . . . , β −1}:
R0 = [0, 2−p−1 ] × [0, 1]
Ri = [i2−p , i2−p + 2−p−1 ] × [0, β −i+1 ]
Ri,α = i2−p , i2−p + 2−p−1 × αβ −i , (α + 1)β −i
−p −p
lin
Ri,1
, i2 + (B + 1 − Ai )2−q × 1? β −i , 5β −i
? = i2
(i > 1)
Ri = ∪α∈Nβ Ri,α
sat
lin
Ri,1
? = Ri,1? \ Ri,1?
As noted before, we use basis β = 5 on the second component to get some
“space” between consecutive encodings. The choice of the value 1 and 4 for the
encoding of 0 and 1, although not crucial, has been made to simplify the proof
as much as possible.
The region R0 is for initialization purposes and is defined differently from the
other Ri . The regions Ri correspond to the different values of i in the configuration (the current number). Each Ri is further divided into the Ri,α corresponding
to all the possible values of the next ε variable (recall that it is encoded in basis
β). In the special case of ε = 1, we cut the region Ri,1? into a linear part and a
saturated part. This is needed to emulate the min(σ +Ai , B +1) in Definition 16:
the linear part corresponds to σ + Ai and the saturated part to B + 1. Figure 1
and Figure 2 give a graphical representation of the regions.
Lemma 20. For any configuration c = (i, σ, εi , . . . , εn ), if i = n + 1 then hci ∈
Rn+1,0? , otherwise hci ∈ Ri,ε?i . Furthermore if εi = 1 and σ + Ai 6 B + 1, then
lin
sat
hci ∈ Ri,1
? , otherwise hci ∈ Ri,1? .
1
R0
R1
R2
0
0
1
Rn+1
B=2
n=2
p=2
ω=2
β=5
Fig. 1. Graphical representation of the regions
Proof. Recall that ω = dlog2 (B + 2)e so B + 1 < 2ω , and q = p + ω + 1. Since
σ 6 B + 1 by definition, (n + 1)2−p 6 hci1 6 (n + 1)2−p + (B + 1)2−p−1−ω 6
(n + 1)2−p + 2−p−1 . This shows the result for the first component. In the case
where σ + Ai 6 B + 1 then σ2−q 6 (B + 1 − Ai )2−q yielding the result for the
second part of the result for the first component.
If i = n + 1, then hci2 = 0? β −p−1 and it trivially belongs to [0? β −n−1 , (0? +
1)β −n−1 ]. Otherwise,
ε?i β −i 6 hci2 6 ε?i β −i +
n+1
X
1? β −j 6 ε?i β −i + 1? β −i−1
j=i+1
6 ε?i β −i + 4β −i−1
1 − β n−i
1 − β −1
β
6 ε?i β −i + β −i 6 (ε?i + 1)β −i .
β−1
This shows the result when i < n+1, for the second component of the result.
t
u
We can now define a piecewise affine function that will mimic the behavior
of T I . The region R0 is here to ensure that we start from a “clean” value on the
first coordinate.
Definition 21 (Piecewise affine simulation).
(2−p , b)
(a, b)
fI (a, b) = (a + 2−p , b − 0? β −i )
(a + 2−p + Ai 2−q , b − 1? β −i )
((i + 1)2−p + (B + 1)2−q , b − 1? β −i )
if
if
if
if
if
(a, b) ∈ R0
(a, b) ∈ Rn+1
(a, b) ∈ Ri,0?
lin
(a, b) ∈ Ri,1
?
sat
(a, b) ∈ Ri,1
?
Lemma 22 (Simulation is correct). For any configuration c ∈ CI , hTI (c)i =
fI (hci).
Proof. Let c = (i, σ, εi , . . . , εn ). There are several cases to consider: if i = n + 1
then TI (c) = c, also by Lemma 20, hci ∈ Rn+1,0? . Thus by definition of f ,
fI (hci) = hci = hT (c)i and this shows the result. If i < n + 1, we have three
more cases to consider: the case where we don’t take the value (εi = 0) and the
two cases where we take it (εi = 1) with and without saturation.
– If εi = 0 then TI (c) = (i + 1, P
σ, εi+1 , . . . , εn ). On the other hand, hci =
n
(a, b) = (i2−p + σ2−q , 0? β −i + j=i+1 εj β −j + 0? β −n−1 ). By Lemma 20,
hci ∈ Ri,0? so by definition of f :
fI (hci) = (a + 2−p , b − 0? β −i )
= ((i + 1)2−p + σ2−q ,
n
X
εj β −j + 0? β −n−1 )
j=i+1
= h(i + 1, σ, εi+1 , . . . , εn )i = hTI (c)i
– If εi = 1 and σ + Ai 6 B + 1 then TI (c) = (i + 1, σP
+ Ai , εi+1 , . . . , εn ). On the
n
other hand, hci = (a, b) = (i2−p + σ2−q , 1? β −i + j=i+1 εj β −j + 0? β −n−1 ).
lin
By Lemma 20, hci ∈ Ri,1? so by definition of f :
fI (hci) = (a + 2−p + Ai 2−q , b − 1? β −i )
n
X
= ((i + 1)2−p + (σ + Ai )2−q ,
εj β −j + 0? β −n−1 )
j=i+1
= h(i + 1, σ + Ai , εi+1 , . . . , εn )i = hTI (c)i
– If εi = 1 and σ + Ai > B + 1 then TI (c) = (i + 1, B + 1, εi+1 , . . . , εn ). By
sat
Lemma 20, hci ∈ Ri,1
? so by definition of f :
fI (hci) = ((i + 1)2−p + (B + 1)2−q , b − 1? β −i )
= h(i + 1, B + 1, εi+1 , . . . , εn )i = hTI (c)i
t
u
3.3
Making the simulation stable
In the previous section, that we have defined fI over a subset of the entire space
and it is clear that this subspace is not stable in any way5 . In order to match
the definition of a piecewise affine function, we need to define f over the entire
space or a stable subspace (containing the initial region). We follow this second
approach and extend the definition of f on some more regions. More precisely,
we need to define f over Ri = Ri,0 ∪ Ri,1 ∪ Ri,2 ∪ Ri,3 ∪ Ri,41 and at the moment
we have only defined f over Ri,1 = Ri,0? and Ri,4 = Ri,1? . Also note that
5
For example R1,1 ⊆ f (R0 ) but f is not defined over R1,1 .
lin
sat
Ri,4 = Ri,4
∪ Ri,4
and we define f separately on those two subregions. In order
to correctly and continuously extend f , we will need to further split the region
lin
sat
Ri,3 into linear and saturated parts Ri,3
and Ri,3
: see Figure 2.
Before jumping into the technical details of the extension, we start with the
intuition. First, it is crucial to understand that our main constraint is continutity:
since we already defined f over Ri,1 and Ri,4 , our extension need to agree with
f on the borders of those regions. Furthermore, f still needs to be affine, leaving
us with little flexibility. Second, the behaviour of f on those regions needs to
be carefully chosen. Indeed, as we mentioned before, we had to overapproximate
regions in several places and our simulation now includes extra points. We do not
want these extra points to have completely unpredictable trajectories, otherwise
they might reach the final region by chance and break the reduction. Therefore,
our strategy is to define f in such a way that its behaviour on those “wrong
points” still has a valid interpretation in the original SUBSET-SUM problem.
We detail this idea for the various region right after.
Let (a, b) ∈ Ri,0 ∪ Ri,2 ∪ Ri,3 : intuitively, this point corresponds to a configuration c = (i, σ, εi , . . . , ε) where εi ∈
/ {0? , 1? }. We know by construction that
this point does not correspond to a proper trajectory so it is tempting to simply
discard it. A very simple way of discarding point is to send them to a point x
that is (i) stable by f (f (x) = x) and (ii) not in the accepting region. That way
we trap the trajectory of invalid points into a useless region of the space. Let us
illustrate this on Ri,0 : let (a, b) ∈ Ri,0 and take fI (a, b) = (a, b). It is now trivial
that the point is stuck in Ri,0 . Unfortunately, fI is not continuous anymore.
Indeed, for (a, b) = (a, β −i ) ∈ Ri,0 ∩ Ri,1 we have a discontinuity on the first
coordinate. Indeed, f1 (a, b) = a on one side but f1 (a, b) = a + 2−p on the other.
A simple fix is to take f (a, b) = (a + 2−p , 0), this corresponds to:
!
!
σ
σ
i
0. i + 1
0.
=
.
if εi = 0 then fI
0. 0 . . . εi εi+1 . . . εn
0. 0 . . . 0 0 . . . 0
One can check that f is also continuous on the second coordinate. The reason
why this choice is clever is because f (Ri,0 ) ⊆ Ri+1,0 . Although the invalid points
are not stable, they are now stuck in the bottom regions Rj,0 for the rest of the
simulation.
Unfortunately, this trick now does not work on Ri,2 because of the continuity requirement on the second coordinate. Indeed, on Ri,1 ∩ Ri,2 we have that
f (a, b) = (a + 2−p , β −i ). To understand what it means, imagine a configuration
where εi = 2 and all the remaining εj are 0, then its image by f corresponds
to a configuration
were all the remaining εj are 4 = 1? plus an error. Indeed,
P∞
−n
−i
β = n=i+1 4β . In other words, we have:
!
!
σ
σ
i
0.
0. i + 1
if εi = 2 then fI
=
.
0. 0 . . . εi 0 . . . 0
0. 0 . . . 0 4 . . . 4 4 . . .
Furthermore, thinking about the future a bit, on Ri,3 ∩Ri,4 we have that f (a, b) =
(a + 2−p , 0). In other words, somewhere on Ri,2 ∪ Ri,3 , the second coordinate
has to go from β −i to 0 in a continuous fashion. This is where the clever tricks
comes in: we can continuously change the second coordinate from 1 to 0 such
that the action of f looks like all εj were “flipped”: 0 is exchange with 4 and 1
with 2. To visualise how this possible, simply think about the configuration as
having an infinite number of εj (although we use a finite amount of them) that
gets turned into an infinite number of µj where µj = 4 − εj :
!
!
σ
σ
0. i + 1
0. i + 1
=
.
fI
0. 0 . . . 2 εi+1 . . . εn εn+1. . .
0. 0 . . . 0 µi+1 . . . µn µn+1. . .
This might seem convoluted at first until one realises why this is helpful. Imagine
an extended SUBSET-SUM simulation where we now have three actions instead
of two:
– εi = 0: go to next number,
– εi = 1: add Ai and go to next number,
– εi = 2: flip all remaining εj and go to next number.
Our simulation corresponds to this extended problem, which is still NP-complete.
The remaining region Ri,3 follows the same principle as Ri,0 , with a slight complication because of the saturated sum. Figure 3 gives the interpretation of the
definition of f on each subregion.
Definition 23 (Extended region splitting). For i ∈ {1, . . . , n} and α ∈
{0, . . . , β − 1}, define:
2−p−1 + i2−p − a
lin
sat
lin
Ri,3
= Ri,3 ∩ (a, b) bβ i − 3 6 −p−1
Ri,3
= Ri,3 \Ri,3
2
− (B + 1 − Ai )2−q
sat
lin
sat
It should be clear by definition that Ri,3
= Ri,3
∪ Ri,3
and that the two
subregions are disjoint except on the border.
Definition 24 (Extended piecewise affine simulation).
(a + 2−p , 0)
(a + 2−p , 3β −i − b)
fI (a, b) =
(a + 2−p + Ai 2−q (bβ i − 3), 0)
((i + 23 )2−p − (bβ i − 3)(2−p−1 − (B + 1)2−q ), 0)
if
if
if
if
(a, b) ∈ Ri,0
(a, b) ∈ Ri,2
lin
(a, b) ∈ Ri,3
sat
(a, b) ∈ Ri,3
This extension was carefully chosen for its properties. In particular, we will
see that f is still continuous, Also, the domain of definition of f is f -stable (i.e.
f (dom f ) ⊆ dom f ). And finally, we will see that f is somehow “reversible”.
Lemma 25 (Simulation is continuous). For any i ∈ {1, . . . , n}, fI (Ri ) is
well-defined and continuous over Ri .
Proof. As outlined on Figure 2, we need to check that the definitions of f match
at the borders of each subregions of Ri . More precisely, we need to check that
Definition 21 and Definition 24 agree on all borders.
β −i+1
lin
−p
Ri,1
+ Ai 2−q , b − 1? β −i )
? : (a + 2
sat
Ri,1
? :
((i + 1)2−p + (B + 1)2−q ,
b − 1? β −i )
4β −i
sat
Ri,3
: (?)
lin
Ri,3
: (a + 2−p + Ai 2−q (bβ i − 3), 0)
3β −i
Ri,2 : (a + 2−p , 3β −i − b)
2β −i
Ri,0? : (a + 2−p , b − 0? β −i )
β −i
Ri,0 : (a + 2−p , 0)
0
i2−p
i2−p + (B + 1 − Ai )2−q
i2−p + 2−p−1
(?) : ((i + 1)2−p + 2−p−1 − (bβ i − 3)(2−p−1 − (B + 1)2−q ), 0)
Fig. 2. Zoom on one Ri with the subregions and formulas.
β −i+1
lin
Ri,1
? : add Ai
sat
Ri,1
? : set sum to B + 1
4β −i
lin
Ri,3
:
sat
Ri,3
: (?)
set all remaining εj to 0
add a funny amount†
3β −i
Ri,2 : flip all remaining εj
2β
−i
Ri,0? : go to next number
β −i
Ri,0 : set all remaining εj to 0
0
i2−p
i2−p + (B + 1 − Ai )2−q
(?) :
set all remaining εj to 0
add a funny amount†
i2−p + 2−p−1
†
The amount depends on (a, b) and
has no simple interpretation.
Fig. 3. Interpretation of the behaviour of f on each suregion. See also Figure 2 and
Section 3.2. All regions include an implicit “go to next number”.
– (a, b) ∈ Ri,0 ∩ Ri,0? : the first component is computed using the same formula
so is clearly continuous, the second component is always 0 on both side of
the border because b − 0? β −i = 0 for b = β −i
– (a, b) ∈ Ri,0? ∩ Ri,2 : the first component is computed using the same formula
so is clearly continuous, the second component is always β −i on both side of
the border because b − 0? β −i = 3β −i − b = β −i for b = 2β −i
lin
– (a, b) ∈ Ri,2 ∩ Ri,3
: the first component is a + 2−p and the second component
is 0 on both side of the border because 3β −i − b = bβ i − 3 = 0 for b = 3β −i
−p
lin
lin
+ Ai 2−q and the second
– (a, b) ∈ Ri,3
∩ Ri,1
? : the first component is a + 2
component is 0 on both side of the border because bβ i −3 = 1 and b−1? β −i =
0 for b = 4β −i
lin
sat
– (a, b) ∈ Ri,3
∩ Ri,3
: the second component is always 0 on both regions so is
Y
holds on
clearly continuous. From Definition 23 one can see that bβ i −3 = X
−p−1
−p
−p−1
the border where Y = 2
+ i2 − a and X = 2
− (B + 1 − Ai )2−q .
Consequently, if we compute the difference between the two expression at
the borders, we get:
(i + 1)2−p + 2−p−1 − (bβ i − 3)(2−p−1 − (B + 1)2−q )
− a + 2−p + Ai 2−q (bβ i − 3)
Y
= i2−p + 2−p−1 − a − (2−p−1 − (B + 1)2−q + Ai 2−q )
X
Y
=Y − X=0
X
this proves that they are equal.
sat
sat
−p
– (a, b) ∈ Ri,3
∩ Ri,1
+ (B + 1)2−q and the
? : the first component is (i + 1)2
second component is 0 on both side of the border because bβ i − 3 = 1 and
b − 1? β −i = 0 for β = 4β −i
lin
sat
−p
– (a, b) ∈ Ri,1
+ (B + 1)2−q on
? ∩ Ri,1? : the first component is (i + 1)2
−p
−q
both side of the border because a = i2 + (B + 1 − Ai )2 , and the second
component is computed using the same formula so is clearly continuous
t
u
Lemma 26 (Simulation is stable). For any i ∈ {1, . . . , n}, fI (Ri ) ⊆ Ri+1 .
Furthermore, f (R0 ) ⊆ R1 and f (Rn+1 ) ⊆ Rn+1 .
Proof. We need to examinate all possible cases for (a, b) ∈ Ri . Since Ri =
S4
lin
sat
α=0 Ri,α and that Ri,α = Ri,α ∪ Ri,α we indeed cover all cases.
– If (a, b) ∈ R0 : then fI (a, b) = (a+2−p , b) so fI (R0 ) = fI ([0, 2−p−1 ]×[0, 1]) =
[2−p , 2−p + 2−p−1 ] × [0, 1] = R1 .
– If (a, b) ∈ Rn+1 : then fI (a, b) = (a, b) so fI (Rn+1 ) = Rn+1 .
– If (a, b) ∈ Ri,0 : then fI (a, b) = (a + 2−p , 0) so fI (Ri,0 ) = fI ([i2−p , i2−p +
2−p−1 ] × [0, β −i ]) = [(i + 1)2−p , (i + 1)2−p + 2−p−1 ] × {0} ⊆ Ri+1 .
– If (a, b) ∈ Ri,1 = Ri,0? : then fI (a, b) = (a + 2−p , b − 0? β −i ) so fI (Ri,1 ) =
fI ([i2−p , i2−p + 2−p−1 ] × [β −i , 2β −i ]) = [(i + 1)2−p , (i + 1)2−p + 2−p−1 ] ×
[0, β −i ] = Ri+1 .
– If (a, b) ∈ Ri,2 : then fI (a, b) = (a+2−p , 3β −i −b) so fI (Ri,2 ) = fI ([i2−p , i2−p +
2−p−1 ] × [2β −i , 3β −i ]) = [(i + 1)2−p , (i + 1)2−p + 2−p−1 ] × [0, β −i ] = Ri+1 .
lin
: the image of the second component is always 0 so it’s
– If (a, b) ∈ Ri,3
easy for this one, also from Definition 23, bβ i − 3 6
2−p−1 +i2−p −a
Ai 2−q
−p−1
−p−1
−q
2−p−1 +i2−p −a
2−p−1 −(B+1−Ai )2−q
−q
ω −q
6
2−p−1 +i2−p −a
2−p−1 −(B+1−Ai )2−q
>
because 2
− (B + 1)2 > 0 since (B + 1)2 6 2 2 6
2
. Consequently, for the first coordinate we get that fI (a, b)1 6 a +
−p−1
−p
−a
2−p + Ai 2−q 2 A+i2
6 (i + 1)2−p + 2−p−1 . Also, since i2−p 6 a 6
−q
i2
−p
−p−1
lin
i2 + 2
, it is clear that fI (a, b)1 > (i + 1)2−p . So finally, fI (Ri,3
)⊆
[(i + 1)2−p , (i + 1)2−p + 2−p−1 ] × {0} ⊂ Ri+1 .
sat
– If (a, b) ∈ Ri,3
: the image of the second component is always 0 so it’s
easy for this one, also from Definition 23, bβ i − 3 >
2−p−1 +i2−p −a
2−p−1 −(B+1)2−q
because Ai > 0. Consequently, for the first coordinate we get
−p−1
−p
2
+i2 −a
that fI (a, b)1 6 (i + 1)2−p + 2−p−1 − (2−p−1 − (B + 1)2−q ) 2−p−1
−(B+1)2−q 6
−p
−p−1
−p
−p−1
−p
−p−1
(i + 1)2 + 2
+ i2 + 2
− a 6 (i + 1)2 + 2
since a 6
i2−p + 2−p−1 . Also since bβ i − 3 6 1 we get that fI (a, b)1 > (i + 1)2−p +
2−p−1 − (2−p−1 − (B + 1)2−q ) × 1 > (i + 1)2−p + (B + 1)2−q . So finally,
sat
fI (Ri,3
) ⊆ [(i + 1)2−p + (B + 1)2−q , (i + 1)2−p + 2−p−1 ] × {0} ⊂ Ri+1 .
lin
lin
−p
– If (a, b) ∈ Ri,4
= Ri,1
+ Ai 2−q , b − 1? β −i ) so
? : then fI (a, b) = (a + 2
lin
−p
−p
−q
−i
fI (Ri,4 ) = fI ([i2 , i2 + (B + 1 − Ai )2 ] × [4β , 5β −i ]) = [(i + 1)2−p +
Ai 2−q , (i+1)2−p +(B +1)2−q ]×[0, β −i ] ⊆ Ri+1 because (B +1)2−q 6 2−p−1 .
sat
sat
– If (a, b) ∈ Ri,4
: then fI (a, b) = ((i + 1)2−p + (B + 1)2−q , 0) so fI (Ri,4
)=
−p
−q
{(i + 1)2 + (B + 1)2 } × {0} ⊆ Ri+1 .
t
u
We now get to the core lemma of the simulation. Up to this point, we were
only interested in forward simulation: that is given a point, what are the iterates
of x. In order to prove the NP-hardness result, we need a backward result:
given a point, what are the possible preimages of it. To this end, we introduce
new subregions Riunsat of the Ri , that we call unsaturated. Intuitively, Riunsat
corresponds to the encodings where σ 6 B, that is the sum did not saturate at
B + 1. We also introduce the Rf in region, that will be the region to reach. We
will be interested in the preimages of Rf in .
Definition 27 (Unsaturated regions). For i ∈ {1, . . . , n + 1}, define
Riunsat = [i2−p , i2−p + B2−q ] × [β −n−1 , β −i+1 − β −n−1 ]
Rf in = [(n + 1)2−p + B2−q − 2−q−1 , (n + 1)2−p + B2−q ] × [β −n−1 , 2β −n−1 ]
Lemma 28 (Simulation is reversible). Let i ∈ {2, . . . , n} and (a, b) ∈ Riunsat
Then the only points x such that fI (x) = (a, b) are:
unsat
– x = (a − 2−p , b + 0? β −i+1 ) ∈ Ri−1,0? ∩ Ri−1
−p
i
? −i+1
unsat
– x = (a − 2 , β − b + 0 β
) ∈ Ri−1,2 ∩ Ri−1
lin
unsat
– x = (a − 2−p − Ai 2−q , b + 1? β −i+1 ) ∈ Ri−1,1
(only if a > 2−p +
? ∩ Ri−1
−q
Ai 2 )
Proof. First notice that since fI (Ri ) ⊆ Ri+1 for all i ∈ {0, . . . , n}, the only
candidates for x must belong to Ri−1 . Furthermore, on each affine region, there
can only be one candidate except if the function is trivial.
A close look at the proof of Lemma 26 reveals that:
– fI (Ri−1,0 ) ⊆ [(i + 1)2−p , (i + 1)2−p + 2−p−1 ] × {0} shareing no point with
Riunsat so there is no possible candidate
– fI (Ri−1,1 ) = Ri and there is only one possible candidate
– fI (Ri−1,2 ) = Ri and there is only one possible candidate
– fI (Ri−1,3 ) ⊆ [(i + 1)2−p , (i + 1)2−p + 2−p−1 ] × {0} so like Ri−1,0 there is no
possible candidate
lin
– fI (Ri−1,4
) ⊆ Ri and there is only one possible candidate
sat
– fI (Ri−1,4 ) ⊆ [(i + 1)2−p + (B + 1)2−q , (i + 1)2−p + 2−p−1 ] × [0, β −i ] sharing
no point with Riunsat so there is no possible candidate
It is then only a matter of checking that the claimed formulas work and they
lin
trivially do except for the case of Ri−1,4
where we need the potential candidate
to belong to the region.
t
u
The goal of those results is to show that if there is a point in Rf in that is
reachable from R0 then we can extract, from its trajectory, a configuration that
also reaches Rf in . Furthermore, we arranged so that Rf in contains the encoding
of only one configuration: (n + 1, B) (see Lemma 18).
Lemma 29 (Backward-forward identity). For any point x ∈ Rf in , if there
[k]
exists a point y ∈ R0 and an integer k such that fI (y) = x then there exists a
[k]
configuration c = (1, 0, ε1 , . . . , εn ) such that fI (hci) ∈ Rf in .
Proof. Define y0 = y and yi+1 = fI (yi ) for all i ∈ {0, k − 1}. Since y0 ∈ R0 ,
we immediately get that yi ∈ Ri using Lemma 26 and in particular, k > n + 1
because yk = x ∈ Rn+1,0? .
unsat
Now apply Lemma 28 starting from yn+1 ∈ Rn+1
: we conclude that for
lin
all i > 1, yi ∈ (Ri,1 ∪ Ri,2 ∪ Ri,4
) ∩ Rnunsat . Define εi = 0 if yi ∈ Ri,1 ∪ Ri,2
lin
and 1 if yi ∈ Ri,4
. Write yi = (ai , bi ). Again using Lemma 26 we get that
−p
ai−1 = ai −2 −εi Ai 2−q (just check all three cases). Also since x = yn+1 ∈ Rf in
then an+1 ∈ [(n + 1)2−p + B2−q − 2−q−1 , (n + 1)2−p + B2−q ]. Finally, y0 ∈ R0 so
fI (a0 , b0 ) = (2−p , 0) = (a1 , b1 ). We conclude that a1 = 2−p . Putting everything
together we get:
Pn
an+1 = (n + 1)2−p + 2−q i=1 εi Ai
an+1 ∈ [(n + 1)2−p + B2−q − 2−q−1 , (n + 1)2−p + B2−q ]
Since the Ai , B are integers and εi ∈ {0, 1}, we get that B =
Lemma 22 on the configuration to conclude.
Pn
i=1
Ai εi . Apply
t
u
Lemma 30 (Final region is accepting). For any configuration c, if hci ∈
Rf in then c = (n + 1, B).
Pn
Proof. Write c = (i, σ, εi , . . . , εn ), thenhci = i2−p + σ2−q , j=i ε?i β −i + 0? β −n−1 .
It implies that i2−p + σ2−q ∈ [(n + 1)2−p + B2−q − 2−q−1 , (n + 1)2−p + B2−q ]
and because i is an integer in range [0, n + 1] and σ an integer in range [0, B + 1],
necessarily i = n + 1 and σ = B.
t
u
3.4
Complexity result
We now have all the tools to show that REACH-REGION-TIME is an NP-hard
problem.
Theorem 31. REACH-REGION-TIME is NP-hard for d > 2.
Proof. Let I = (B, A1 , . . . , An ) be an instance of SUBSET-SUM. We consider the
instance J of REACH-REGION-TIME defined in the previous section with maximum number of iterations set to n (the number of Ai ), the initial region set
to R0 and the final region set to Rf in . One easily checks that this instance has
polynomial size in the size of I. The two directions of the proof are:
– If I is satisfiable then use Lemma 17 and Lemma 22 to conclude that there
[n]
is a point x ∈ R0 in the initial region such that fI (x) ∈ Rf in so J is
satisfiable.
[k]
– If J is satisfiable then there exists x ∈ R0 and k 6 n such that fI (x) ∈
Rf in . Use Lemma 29 and Lemma 22 to conclude that there exists a config[k]
[k]
uration c = (1, 0, ε1 , . . . , εn ) such that hTI (c)i = fI (hci) ∈ Rf in . Apply Lemma 30 and use the injectivity of the encoding to conclude that
[k]
TI (c) = (n + 1, B) and Lemma 18 to get that I is satisfiable.
t
u
4
Bounded Time Reachability is in NP
In the previous section we have shown that the REACH-REGION-TIME problem
is NP-hard. We now give a more precise characterization of the complexity of
this problem, by proving that it is NP-complete. Since we have shown its NPhardness, the only thing that remains to be shown is that REACH-REGION-TIME
belongs to NP. This is done in this section.
4.1
Notations and definitions
For any i ∈ {1, . . . , d}, let πid : Rd → R denote the ith projection function, that
is, π(x1 , . . . , xd ) = xi . Let gd : Rd+1 → Rd be defined by gd (x1 , . . . , xd+1 ) =
(x1 , . . . , xd ). For a square matrix A of size (d + 1) × (d + 1) define the following
pair of projection functions. The first function h1,d takes as input a square matrix
A of size (d + 1) × (d + 1) and returns a square matrix of size d × d that is the
upper-left block of A. The second function h2,d takes as input a square matrix A
of size (d+1)×(d+1) and returns the vector of size d given by [a1,d+1 · · · ad,d+1 ]T
(the last column of A minus the last element).
Let s denote the size function, its domain of objects will be overloaded and
understood from the context. For x ∈ Z, s(x) is the length of the encoding of
x in base 2. For x ∈ Q with x = pq with p and q coprime, we have s(x) =
max(s(p), s(q)). For an affine function f we define the size of f (x) = Ax + b
(where all entries of A and b are rationals) as: s(f ) = max(maxi,j (s(ai,j )), max(s(bi ))).
We define the size of a polyhedron r defined by Ax 6 b as: s(r) = max(s(A), s(b)).
We define the size of a piecewise affine function f as: s(f ) = maxi (s(fi ), s(ri ))
where fi denotes the restriction of f to ri the ith region.
We define the signature of a point x as the sequence of indices of the regions
traversed by the iterates of f on x (that is, the region trajectory).
4.2
REACH-REGION-TIME is in NP
In order to solve a reachability problem, we will formulate it with linear algebra.
However a crucial issue here is that of the size of the numbers, especially when
computing powers of matrices. Indeed, if taking the nth power of A yields a
representation of exponential size, no matter how fast our algorithm is, it will
run on exponentially large instances and thus be slow.
First off, we show how to move to homogeneous coordinates so that f becomes
piecewise linear instead of piecewise affine.
Lemma 32. Assume that f (x) = Ax
+ b with
A = (ai,j )16i,j6d and let y =
Ab
0
T
0
A (x, 1) where A is the block matrix
. Then f (x) = gd (A0 (x, 1)T ).
0 1
Remark 33. Notice that this lemma extends nicely to the composition of affine
functions: if f (x) = Ax + b and h(x) = Cx + d then h(f (x)) = gd (C 0 A0 (x, 1)T ).
We can now state the main lemma, namely that the size of the iterates of f
vary linearly in the number of iterates, assuming that f is piecewise affine.
Lemma 34. Let d > 2 and f ∈ P AFd . Assume that all the coefficients of f
on all regions are rationals. Then for all t ∈ N, s(f [t] ) 6 (d + 1)2 s(f )pt + (t −
1)dlog2 (d + 1)e where p is the number of regions of f . This inequality holds even
if all rationals are taken to have the same denominator.
Proof. Using Lemma 32, we get that f [t] (x) = gd (h[t] ([x 1]T )), where h is a
piecewise linear function in dimension d + 1 such that s(h) = s(f ). We show this
result by induction on t for h. The result then follows for f . In all cases we take
all rationals to have the same denominator.
In the case t = 1, it suffices to see that taking all rationals to have the same
denominator involves multiplying the numerator and denominators by at most
2
the lowest common multiple of all numbers, and hence is at most 2s(f )(p(d+1) ) .
Indeed the greatest number is 2s(h) by definition, and there are (d + 1)2 numbers
per region (the entries of the matrix).
Assume the result is true for t ∈ N. Let y ∈ Qd+1 . Then h[t+1] (y) =
Bt+1 · · · B1 y, where Bi ’s are the matrices corresponding to some regions of
h. In particular, s(Bi ) 6 s(h). From the induction hypothesis we can assume
that all rationals have the same denominator and we get that s(Bt · · · B1 ) 6
(d + 1)2 s(h)pt + (t − 1)dlog2 (d + 1)e. It follows6 that for any 1 6 i, j 6 d + 1:
s((Bt+1 · · · B1 )i,j ) = s
d+1
X
!
(Bt+1 )i,k (Bt · · · B1 )k,j
k=1
6 dlog2 (d + 1)e + s(Bt+1 ) + s(Bt · · · B1 )
6 dlog2 (d + 1)e + s(h) + (d + 1)2 s(h)pt + (t − 1)dlog2 (d + 1)e
6 (d + 1)2 s(h)p(t + 1) + tdlog2 (d + 1)e
This shows the result for the particular region where y belongs. Since the
bound does not depend on y and h[t+1] has finitely many regions, it is true for
all regions of h[t+1] .
t
u
Finally, we need some result about the size of solutions to systems of linear
inequalities. Indeed, if we are going to quantify over the existence of a solution
of polynomial size, we must ensure that the size constraints do not change the
satisfiability of the system.
Lemma 35 ([11]). Let A be a N × d integer matrix and b an integer vector.
If the Ax 6 b system admits a solution, then there exists a rational solution xs
such that s(xs ) 6 (d + 1)L + (2d + 1) log2 (2d + 1) where L = max(s(A), s(b)).
Proof. See Theorem 5 of [11]: s(xs ) 6 s (2d + 1)!2L(2d+1) .
t
u
Putting everything together, we obtain a fast nondeterministic algorithm to
solve REACH-REGION-TIME. The nondeterminism allows us to choose a signature
for the solution. Once the signature is fixed, we can write it as a linear program
of reasonable size using Lemma 34 and solve it. The remaining issue is the one
of the size of solution but fortunately Lemma 35 ensures us that there is a small
solution that can be found quickly.
Theorem 36. REACH-REGION-TIME is in NP.
Proof. The idea of the proof is to nondeterministically choose a signature for a
solution, that is a sequence of regions for the iterates of the solution. We then
build a system of linear inequalities stating that a point x belongs to the initial
region and that the iterates match the signature chosen and finally that the
iterates reach the final region. Using the results of the previous section, we can
build this system in polynomial time and solve it in non-deterministic polynomial
time. Here is an outline of the algorithm:
6
Use elementary properties of the size function: s(xy) 6 s(x)+s(y), s(x1 +· · ·+xk ) 6
s(k) + maxk s(xk )
–
–
–
–
–
Non-deterministically choose t 6 T
Non-deterministically choose regions r1 , . . . , rt−1 regions of f
Define r0 = R0 the initial region and rt = R the final region
Build (S) the system Ax 6 b stating that the signature of x matches r
Non-deterministically choose xs a rational of polynomial size in the size of
(S)
– Accept if Axs 6 b
We have two things to prove. First we need to show that this algorithm
indeed has non-deterministic polynomial running time. Second we need to show
that it is correct. Recall that T is a unary input of the problem.
The complexity of the algorithm is clear, assuming that (S) is of polynomial
size. Indeed verifying that a rational point satisfies a system of linear inequalities
with rationals coefficients can be done in polynomial time.
We build (S) this way: (S) = ∪ti=1 (Si ) where (Si ) states that f [i] (x) ∈
ri . Since we choose a signature of x we know that if x satisfies the system
then from Lemma 32 f [i] (x) = gd A0i−1 · · · A01 (x, 1)T where A0j is the matrix
corresponding to the region rj . Write Ci = A0i−1 · · · A01 and define (Si ) by the
system gd Ci (x, 1)T ∈ ri . Since ri is a polyhedron, (Si ) is indeed a system of
linear inequalities7 .
We can now see that S is of polynomial size using Lemma 34. Indeed,
s(Ci ) 6 s(f [i] ) 6 poly(s(f ), i), thus s((Si )) 6 s(Ci ) + s(ri ) 6 poly(s(f ), i)
because the description of the regions is part of the size of f . And finally
s((S)) 6 poly(s(f ), t).
The correctness follows from the construction of the system and Lemma 35.
More precisely we show that x ∈ (S) if and only if ∀i ∈ {0, . . . , t}, f [i] (x) ∈ ri .
Indeed, (S) ⇔ ∀i ∈ {0, . . . , t}, x ∈ (Si ) and by definition (Si ) ⇔ f [i] (x) ∈ ri
since gd (Ci (x, 1)T ) = f [i] (x). Then by Lemma 35, we get that ∃x ∈ (S) ⇔ ∃x ∈
(S) and s(x) 6 poly(s((S))).
t
u
5
Other Bounded Time Results
In this section, we give succinct proofs of the other result mentioned in the introduction about CONTROL-REGION-TIME. The proof is based on the same arguments
as before.
Theorem 37. Problem CONTROL-REGION-TIME is coNP-hard for d > 2.
Proof. The proof is exactly the same except for two details:
– we modify f over Rn+1 as follows: divide Rn+1 in three regions: Rlow that
is below Rf in , Rf in and Rhigh that is above Rf in . Then build f such that
f (Rlow ) ⊆ Rlow , f (Rf in ) ⊆ Rf in and f (Rhigh ) ⊆ Rlow .
– we choose a new final region Rf0 in = Rlow .
7
More precisely if ri is defined by Pi (x, 1)T 6 0 then (Si ) is the system Pi Ci (x, 1)T 6 0
Let I = (B, A1 , . . . , An ) be an instance of NOSUBSET-SUM, let J be the
corresponding instance of CONTROL-REGION-TIME we just built. We have
to show that I has no subset sum if and only if J is “controlled”. This is the
same as showing that I has a subset sum if and only if J has points never
reaching Rf0 in .
Now assume for a moment that the instance is in SUBSET-SUM (as opposed
to NOSUBSET-SUM), then by the same reasoning as the previous proof, there
will be a point that reaches the old Rf in region (and is disjoint from Rf0 in ). And
since Rf in is a f -stable region, this point will never reach Rf0 in .
And conversely, if the control problem is not satisfied, necessarily there is
a point whose trajectory went through the old Rf in (otherwise if would have
reached either Rlow = Rf0 in or Rhigh but f (Rhigh ) ⊆ Rlow ). Now we proceed as
in the proof of Theorem 31 to conclude that there is a subset that sums to B,
and thus I is satisfiable.
t
u
Theorem 38. Problem CONTROL-REGION-TIME is in coNP for d > 2.
Proof. Again the proof is very similar to that of Theorem 36: we have to build
a non-deterministic machine that accepts the “no” instances. The algorithm is
exactly the same except that we only choose signatures that avoid the final
region (as opposed to ending in the final region) and are of maximum length
(that is t = T as opposed to t 6 T ). Indeed, if there is a such a trajectory, the
problem is not satisfied. And for the same reasons as Theorem 36, it runs in
non-deterministic polynomial time.
t
u
6
Fixed Precision Results
Theorem 39. REACH-REGION-PRECISION and CONTROL-REGION-PRECISION are
P SP ACE-hard.
Proof. Consider a polynomial space Turing machine M = (Q, Σ, B, q0 , F, δ).
Without loss of generality, we can assume that F = {qf } ⊂ Q (there is a single
accepting state) and that the working alphabet is Σ = {0, 1, 2, . . . , β}, assuming
that B = 0 is the blank character, and δ : Q×Σ → Q×Σ ×{/, , .} is complete.
We also assume the set of internal states to be such that Q ⊆ Σ m for some m.
Let c be an instantaneous configuration (sometimes also called ID for Instataneous Description) of M. We write c = (x, σ, q, y) where x (resp. y) is encoding
the part of the tape on the left (resp. right) of the head, σ is the symbol under
the head, and q is the state of the machine. Specifically, if the non-blank part of
the tape is s−n . . . s−1 σs1 s2 . . . sm , with the head in front of σ, then x is encoded
as word s−1 s−2 . . . s−n , and y as word s1 s2 . . . sm .
Define the encoding of configuration c as hci = (hxi, hqσyi) where for any
P|w|
word w ∈ Σ ∗ , hwi = i=1 2wi (2γ )−i , where γ is such that 2β + 1 6 2γ . Define
regions Rα,q,σ = [hαi] × [hqσi] where [hwi] is a shortcut for [hwi] = [hwi, hwi +
(2γ )−|w| ]. Intuitively, Rα,q,σ contains all configurations in state q, with symbol σ
under the head and symbol α immediately at the left of the head. By construction
hci ∈ Rx1 ,q,σ with the above notations. Finally, for α, σ ∈ Σ, q ∈ Q, define f on
region Rα,q,σ by:
−γ
0
γ
0
a2 + hσ i, (b − hqσi)2 + hq i)
f (a, b) = (a, b − hqσi + hq 0 σ 0 i)
(a − hαi)2γ , (b − hqσi)2−γ + hq 0 ασ 0 i)
if δ(q, σ) = (q 0 , σ 0 , .)
if δ(q, σ) = (q 0 , σ 0 , )
if δ(q, σ) = (q 0 , σ 0 , /)
It is clear from the definition that f is piecewise affine over its domain of
definition. Let T be the function corresponding to one step of computation of M:
T is acting on configurations and maps any configuration c to the corresponding
next configuration according to the program of M. A simple case analysis shows
that for any configuration c, hT (c)i = f (hci), using the fact that the blank
character B was chosen to be 0.
Observe furthermore that by the choice of the encoding, all the regions Rα,q,σ
are closed and at positive distance from each other: Rα,q,σ ∩ Rα0 ,q0 ,σ0 = ∅ whenever (α, q, σ) 6= (α0 , q 0 , σ 0 ). It follows that f can be easily extended to a continuous piecewise linear function defined over the whole domain [0, 1] (similar
arguments are used in [12]). By construction it will still satisfy that for any
configuration c, hT (c)i = f (hci).
We can now state the reduction from problem LINSPACE-WORD: consider an
instance (M, w) of this decision problem. Define ε = (2γ )−(|w|+m+1) where the
choice of m was explained above. Then for any configuration c reachable from the
initial configuration c0 = (ε, q0 , w1 , w2 · · · w|w| ), we have the stronger property
that hci = b hci
ε cε. Indeed, by assumption the machine never uses more space
than the size of the input, thus the left and right part of tape are always smaller
than |w| at any point during the computation, and we simply need an extra m+1
space to store the current state of the machine. In other words, rounding to ε
does not perturbate the simulation. Consequently, we get that for any reachable
configuration c, hT (c)i = fε (hci).
Define R0 = {hc0 i} and R = [0, 1] × [hqf i] that are convex regions. Then the
instance (f, R0 , R, ε) of REACH-REGION-PRECISION is satisfiable if and only if
(M, w) belongs to problem LINSPACE-WORD. One easily checks that (f, R0 , R, ε)
has polynomial size in the size of (M, w).
The same instance also works for CONTROL-REGION-PRECISION. If we want to
make R0 a region with non-empty interior, just take a ball of radius smaller that
ε(2γ )−2 around hc0 i so that any input error is removed after the first application
of the function fε .
t
u
Theorem 40. REACH-REGION-PRECISION and CONTROL-REGION-PRECISION are
in P SP ACE.
Proof. Let N = bε−1 c and consider the graph G = (V, E) where
V = {0, . . . , N − 1}d
Cα =
d
Y
k=1
αk ε, (αk + 1)ε
(α ∈ V )
E = {(α, β) | f (Cα ) ∩ Cβ 6= ∅}
S = {α | f (R0 ) ∩ Cα 6= ∅}
T = {α | R ∩ Cα 6= ∅}
We can now restate our reachability problem in the graph G as an accessibility
problem from S to T . This can be done in space logarithmic in the size of the
graph, using the fact that accessibility in a graph with N vertices can be done
in non-deterministic space O(log N ), and using the fact that N SP ACE(N ) =
2
SP ACE(N
) (Savitch’s Theorem) [16,
Theorem 8.5].
Since the
graph is of size
d
O N , this requires space O log2 N = O − log2 ε = O n2 if ε = 2−n . Also
note that computing the transitions of the graph is fast since f is a piecewise
affine function. The same proof applies to CONTROL-REGION-PRECISION except
we now want to know if for every vertex in S there is a path to T .
t
u
References
1. Asarin, E., Maler, O., Pnueli, A.: Reachability analysis of dynamical systems having piecewise-constant derivatives. Theoretical Computer Science 138(1), 35–65
(Feb 1995)
2. Asarin, E., Schneider, G.: Widening the boundary between decidable and undecidable hybrid systems. In: Brim, L., Jancar, P., Kretı́nský, M., Kucera, A.
(eds.) CONCUR 2002 - Concurrency Theory, 13th International Conference, Brno,
Czech Republic, August 20-23, 2002, Proceedings. Lecture Notes in Computer Science, vol. 2421, pp. 193–208. Springer (2002), http://link.springer.de/link/
service/series/0558/bibs/2421/24210193.htm
3. Asarin, E., Schneider, G., Yovine, S.: On the decidability of the reachability
problem for planar differential inclusions. In: Benedetto, M.D.D., SangiovanniVincentelli, A.L. (eds.) Hybrid Systems: Computation and Control, 4th International Workshop, HSCC 2001, Rome, Italy, March 28-30, 2001, Proceedings. Lecture Notes in Computer Science, vol. 2034, pp. 89–104. Springer (2001), http:
//link.springer.de/link/service/series/0558/bibs/2034/20340089.htm
4. Bazille, H., Bournez, O., Gomaa, W., Pouly, A.: On the complexity of bounded
time reachability for piecewise affine systems. In: Ouaknine, J., Potapov, I., Worrell, J. (eds.) Reachability Problems, Lecture Notes in Computer Science, vol.
8762, pp. 20–31. Springer International Publishing (2014), http://dx.doi.org/
10.1007/978-3-319-11439-2_2
5. Bell, P., Chen, S.: Reachability problems for hierarchical piecewise constant derivative systems. In: Abdulla, P., Potapov, I. (eds.) Reachability Problems, Lecture
Notes in Computer Science, vol. 8169, pp. 46–58. Springer Berlin Heidelberg (2013),
http://dx.doi.org/10.1007/978-3-642-41036-9_6
6. Ben-Amram, A.M.: Mortality of iterated piecewise affine functions over the integers: Decidability and complexity. In: STACS. pp. 514–525 (2013)
7. Blondel, V.D., Bournez, O., Koiran, P., Tsitsiklis, J.: The stability of saturated
linear dynamical systems is undecidable. Journal of Computer and System Science
62(3), 442–462 (May 2001), http://dx.doi.org/10.1006/jcss.2000.1737
8. Garey, M.R., Johnson, D.S.: Computers and Intractability. W. H. Freeman and Co
(1979)
9. Henzinger, T.A., Kopke, P.W., Puri, A., Varaiya, P.: What’s decidable about hybrid automata? Journal of Computer and System Sciences 57(1), 94–124 (Aug
1998)
10. Karp, R.M.: Reducibility among combinatorial problems. Springer (1972)
11. Koiran, P.: Computing over the reals with addition and order. Theor. Comput.
Sci. 133(1), 35–47 (1994)
12. Koiran, P., Cosnard, M., Garzon, M.: Computability with low-dimensional dynamical systems. Theoretical Computer Science 132(1-2), 113–128 (Sep 1994)
13. Koiran, P., Cosnard, M., Garzon, M.: Computability with Low-Dimensional Dynamical Systems. Theoretical Computer Science 132, 113–128 (1994)
14. Moore, C.: Generalized shifts: unpredictability and undecidability in dynamical
systems. Nonlinearity 4(3), 199–230 (1991)
15. Siegelmann, H.T., Sontag, E.D.: On the computational power of neural nets. Journal of Computer and System Sciences 50(1), 132–150 (Feb 1995)
16. Sipser, M.: Introduction to the Theory of Computation. PWS Publishing Company
(1997)
| 3cs.SY
|
A New Optimization Approach Based on
Rotational Mutation and Crossover Operator
Masoumeh Vali
Department of Mathematics, Dolatabad Branch, Islamic Azad University, Isfahan, Iran
E-mail: [email protected]
Abstract
Evaluating a global optimal point in many global optimization problems in large space is required
to more calculations. In this paper, there is presented a new approach for the continuous functions
optimization with rotational mutation and crossover operator.
This proposed method (RMC) starts from the point which has best fitness value by elitism
mechanism and after that rotational mutation and crossover operator are used to reach optimal
point. RMC method is implemented by GA (Briefly RMCGA) and is compared with other wellknown algorithms such as: DE, PGA, Grefensstette and Eshelman[15,16] and numerical and
simulating results show that RMCGA achieve global optimal point with more decision by smaller
generations.
Keywords: genetic algorithm (GA), rotational mutation, crossover operator, optimal point.
1. Introduction
This paper concerned with the simple-bounded continuous optimization problem as follows:
f ( , , ..., ) where each is a real parameter so that ≤ ≤ for some constants and
and = 1, 2, … , .
This problem has widespread applications including optimization simulating models, fitting
nonlinear curve to data, solving system of nonlinear, engineering design and control problem, and
setting weights on neural networks.
Since GA provides a comprehensive search methodology for optimization, this problem is
implemented by GA for more optimal performance. In global optimization scenarios, GAs often
manifests their strengths: efficiency, parallelizable search; the ability to evolve solutions with
different dimension; and a characterized and controllable process of innovation.
1
In this paper is presented a new approach for simple-bounded continuous optimization problem
that is called RMC. In other words, RMC is a new approach for finding global optimal point
(max/min) by simple algebra, without derivation and based on search.
This method, at first, search the point has the best fitness in its search space and after that by using
rotational mutation and crossover operator finds the global optimal point.
This paper starts with the description of related work in section 2. Section 3 gives the model and
problem definition of RMC. In section 4, text problems of RMC method is implemented. In
section 5, Schemata Analysis for RMCGA is present. Evaluation by De Jong's functions and the
compression of RMCGA with the other methods (DE, PGA, Grefensstette and Eshelman) for De
Jong Functions are shown in sections 6. The discussion ends with a conclusion and future trend.
2. Related Work
In the early 1960s and 1970s, new search algorithms were initially proposed by Holland, his
colleagues and his students at the University of Michigan. These search algorithms which are
based on nature and mimic the mechanism of natural selection were known as Genetic
Algorithms (GAs) [1, 2, 3, 4, 5, 6, 7]. Holland in his book “Adaptation in Natural and Artificial
Systems” [1] initiated this area of study. Theoretical foundations besides exploring applications
were also presented.
As a matter of fact, “Genetic algorithms’ functionality is based upon Darwin's theory of evolution
through natural and sexual selection.” [6], they mimic biological organisms [3]. In GAs a solution
to the problem is represented as a genome (or chromosome) [1, 3, 4].
Pratibha Bajpal and Manojkumar [9] have proposed an approach to solve Global Optimization
Problems by GA and obtained that GA is applicable to both continuous and discrete optimization
problems.
Hayes and Gedeon [10] considered infinite population model for GA where the generation of the
algorithm corresponds to a generation of a map. They showed that for a typical mixing operator all
the fixed points are hyperbolic.
Gedeon et al. [11] showed that for an arbitrary selection mechanism and a typical mixing operator,
their composition has finitely many fixed points.
Qian et al.[12] proposed a GA to treat with such constrained integer programming problem for the
sake of efficiency. Then, the fixed-point evolved (E)-UTRA PRACH detector was presented,
which further underlines the feasibility and convenience of applying this methodology to practice.
Devis Karaboga and Selcuk. [15] is proposed new heuristic approach with deferential evaluation
(DE) for finding a true global minimum regardless of the initial parameter values, fast convergence
using similar operators crossover, mutation and selection.
2
3. Model and Problem Definition of RMC
In this section, at first, model of RMC is expressed.
Suppose f , , … . . , with constraint ≤ ≤ for i=1, 2,…, n. The serial algorithm
RMC is as follows:
Step 1: Draw the diagrams for = and = for i=1, 2,…, n.
Step 2: Consider the vertex of this polytope which has best fitness among other vertexes and call it
S.
Step 3: Put S = f(S) , =(1,1,…,1),
(1,-1,1,…,1),…, or (-1,-1,…,-1) and
= ,
= 0.1" , n=1,2,…,10.
Step 4: Move the point ‘S’ with length of α in direction of vector (notice: the direction of
vector must be inside the search space, also the α measurement depends on problem precision).
The endpoint of vector is called P.
Step 5: If f (P) better than S , then put S=P and go to step 8 else go to step 6.
Step 6: Put # = 0.1, 0.25, …, % = 1,1, … ,1, 1, −1,1, … ,1, … , and −1, −1, … , −1,P = #% .
Step 7: If f (P) is better than S , then S=P else go to step 8.
Step 8: If P is in search space, go to step 9 else go to stop.
Step 9: Halve the adjacent sides of polytope. The result is the production of new points.
Step 10: Select point has best fitness among the new points produced from step 9 and point ‘S’.
Then, put S the point which has the best value and S' fitness of point S then go to step 3 and repeat
this trend.
4. Implementing Test problems by RMC
4.1.Test problem1:
The minimization problem is formulated as follows:
min fx = x + x − 18 cos x − 18 cos x
* ∈ℝ-
−1 ≤ 5 ≤ 1,
5
= 1,2
Its global minimum is equal to -2 and the minimum point is at (0,0).
The process of achieve to optimal global point by RMC is shown following figure.
3
1
S
Q
-1
Q''
S Q'
1
-1
Figure 1: The Performance RMC
Note that f(1,1) = f(-1,1) = f(1,-1) = f(-1,-1) so it's not important that from which side start. As you
can see, the first two mutations improve the fitness of problem- The fitness value of mutated
points are better than the fitness value of points produced by crossover - but the fitness of the third
mutation cannot improve the fitness of problem so use crossover operator and produce Q and Q'
points.Then among S, Q and Q' select S which has best fitness and after that rotational mutation
and crossover operator is used for achieve global optimal point.
4.2.Test problem2:
The Goldstein-Price function [GP71] is a global optimization test function.
function definition:
6789: , ;1 0 0 0 1 . 19 & 14 0 3 & 14 0 6 0 3 @.
;30 0 2 & 3 . 18 & 32 0 12 0 48 & 36 0 27 @
&2 2,
1,2
Global minimum:
6 ,
3; ,
0, &1
-2
Figure 2: Figure of the Goldstein-Price
4
2
Q
2
-2
Q'
-2
S
Figure 3: The Performance RMC
As regards the fitness value of points (2,-2) and (-2,-2) is equal and better than the fitness value of
points (2, 2) and (-2, 2) so we start from point (2,-2). As you can see, the first third mutations
improve the fitness of problem- The fitness value of mutated points are better than the fitness
value of points produced by crossover - but the fitness of the forth mutation cannot improve the
fitness of problem so use crossover operator and produce Q and Q' points and select Q' which has
the best fitness. After that use rotational mutation to achieve point which the improvement is better
but this point didn't find so point Q' is chosen as the global point optimal.
5. Schema Analysis RMCGA
This schema (Figure 4) starts from the vertex (offspring) which has best fitness value that is
called ‘S’. Then it sets initial point of
vector at ‘S’ and mutates offspring ‘S’ in direction
of
vector with length of mutation α. ( notice that the direction of vector must be inside
the search space, also the α measurment depends on problem precision.) . This mutated
offspring is calledP.
If fitness value of offspring P was better than the fitness value of offspring S we
would use crossover operator for adjacent edges of offspringP. Otherwise by using
rotational mutation by %
vector, we would search an offspring-with better fitness value in
comparison with‘S’. Then we make a crossover. After that we select an offspring with the
better fitness value –between the mutated offspring and offspring's which were generated
crossover operator. We mutate and make a crossover on it again. We repeat this action while
the mutated offspring doesn’t get out of search space. Eventually, we would select the last
5
new produced offspring- inside the search space- as the global optimization point which would
be fixed point of our question.
start
, ≤ ≤ C ,
…, ≤ ≤ D
E = , , F = , C ,
…,G = , D
S= vertex of the polytope which has
best fitness and put S’= f(S)
=(1,1,…,1),(1,-1,1,…,1),…,or
(-1,-1,…,-1)
Set initial point of
vector at H
=
, = 0.1"
n=1,2,…,10, So that be
in search
space Y
Put P=end of point of vector
# = 0.1, 0.25, …
% =(1,1,…,1),(1,-1,1,…,1),…, and
(-1,-1,…,-1)
= #%" ,p=end of point of
vector
Is fitness value of
No
I better than S’?
N
o
Is fitness value of p
better than S’?
S=p
Yes
N
e
Is p in search
space?
No
Yes
Acts Crossover operator on the
adjacent sides P
Put S’= best fitness of offsprings (vertexes)
among S and resulting offspring of
crossover operator
Print 'S' point
Figure 4: The improved GA
6
End
6. Evaluation
In this section, Definition of De Jong's Functions, Numerical results of RMCGA on De Jongs'
functions and finally simulating results and the compression of RMCGA with the other methods
such as DE, PGA, Grefensstette and Eshelman[15, 16] are shown.
6.1 De Jong's Functions
In this section, Definition of De Jong's Functions (F1 to F5) and initial population RMCGA which
depends on the dimensional space are shown in Table 1.
Table 1: De Jong's Functions
Function
Number
F1
F2
Function
F5
Dim.
&5.12 ≤ ≤ 5.12
Initial
Population
3
8
−2.048 ≤ ≤ 2.048
2
4
−5.12 ≤ ≤ 5.12
5
32
-1.28≤ ≤ 1.28
30
−65.536 ≤ ≤ 65.536
2
J
K
100. − + 1 −
O
F3
F4
Limits
30. + JLM N
K
J; C . +PQ 0,1@
K
0.002 +
1
∑C
K
D∑Y
TZ[;ST UVWT @
4
X
6.2. Numerical Results
In this section, the experimental results of RMCGA on the five problems of De Jong (F1 to F5)
[De Jong, 1975] in Table 2 are shown. Furthermore, results were saved for the best performance
(BP) which BP is the best fitness of the objective function obtained over all function evaluations.
At last, standard deviation (SD) is calculated and measured with final answer of De Jong function.
The following parameters were evaluated in the following table.
rotational mutation size (RMS), number of iterations rotational mutation (TRM), number of
iterations crossover operator (TC).
7
Table 2: The average number of generations.
Algorithm
De Jong’s
Function
F1
RMCGA
F2
RMCGA
F3
RMCGA
RMCGA
RMCGA
F4
F5
RMS
TRM
TC
Best Point
BP
SD
0.1
55
102
(0,0,0)
0
0
0.1
15
20
(1,1)
0
0
0.1
4
1
(-5.12,-5.12,-5.12, 5.12, -5.12)
0
0
0.1
15
21
0
30
Depend
on \
(-32,-32)
0
0
(0,0,…,0)
0.1
340
670
6.3. Simulating Results
Dervis Karaboga and Selcuk O^kdema15b introduced new method DE for finding global
optimization problems. They compared DE method with other methods: PGA, Grefensstette and
Eshelman [16] and showed that DE algorithm works much better than other methods[16].
In order to obtain the average results, PGA, Grefensstette and Eshelman algorithms were run 50
times; the DE algorithm was run 1000 times and RMCGA 200 times for each function.
Table 3: The average number of generations
Algorithms
F1
F2
F3
F4
F5
PGAc = d
1170
1235
3481
3194
1256
PGAc = e
1526
1671
3634
5243
2076
Grefensstette
2210
14229
2259
3070
4334
Eshelman
1538
9477
1740
4137
3004
DE(F: RandomValues)
260
670
125
2300
1200
RMCGA
157
35
5
36
1010
PNG
1.656
19.142
25
63.888
1.188
8
As you can see in Table 3, the convergence speed RMCGA achieves the optimal point with more
decision by smaller generation. Furthermore, the most significant improvement is with F4 since
the proportion of the number of generations (PNG) about f ≅ 64 times smaller than the
average of the number of generations DE algorithm.
7. Conclusion
RMCGA is a new method for finding the true optimal global optimization is based on rotational
mutation and crossover operator. In this work, the performance of the RMCGA has been compared
to that of some other well known GAs. From the simulation studies, it was observed that RMCGA
achieve the optimal point with more decision by smaller generation. Therefore, RMCGA seems to
be a promising approach for engineering optimization problems.
References
Holland J.H., "Adaptation in Natural and Artificial Systems: An Introductory Analysis with
Applications to Biology, Control, and Artificial Intelligence", University of Michigan Press,
Ann Arbor, Mich, USA, 1975.
[2]
J. H. Holland, Adaptation in Natural and Artificial Systems: An Introductory Analysis with
Applications to Biology, Control, and Artificial Intelligence, Cambridge, USA: The MIT
Press, 1992.
[3]
Y. Liao, and C.T. Sun, “An Educational Genetic Algorithms Learning Tool” ©2001 IEEE.
[4]
N. A. AL-Madi, A. T. Khader, “A SOCIAL-BASED MODEL FOR GENETIC
ALGORITHMS”. Proceedings of the third International Conference on Information
[5]
Technology (ICIT), AL- Zaytoonah Univ., Amman, Jordan, 2007.
[6]
J. Dalton, “Genetic Algorithms” Newcastle Engineering Design Centre,
http://www.edc.ncl.ac.uk. 2007.
[7]
E. Bagheri, H. Deldari, 2006. “ Dejong Function Optimization by means of a Parallel
Approach to Fuzzified Genetic Algorithm” Proceedings of the 11th IEEE Symposium on
Computers and Communications (ISCC'06) 0-7695-2588-1/06 $20.00 © 2006 IEEE.
[8]
T. Back. “Evolutionary Algorithms in theory and practice Evolution Strategies, Evolutionary
Programming, Genetic Algorithms”. Accessed 2008.
[9]
Pratibha Bajpal and Manojkumar, Genetic Algorithm – an Approach to Solve Global
Optimization Problems,Pratibha Bajpai et al. / Indian Journal of Computer Science and
Engineering Vol 1 No 3 199-206.
[10] Zhang J., Dong Y., Gao R., and Shang Y.," An Improve Genetic Algorithm Based on Fixed
Point Theory for Function Optimization", IEEE, International Conference on Computer
Engineering and Technology, pp. 481-484, 2009.
[11] Gedeon T., Hayes C., and Swanson R., " Genericity of the Fixed Point set for the Infinite
Population Genetic Algorithm" Foundations of Genetic Algorithms Lecture Notes in
Computer Science, Volume 4436, pp. 97-109,2007.
[1]
9
[12]
[13]
[14]
[15]
[16]
Qian R., Peng T., Qi Y., and Wang W. " Genetic Algorithm-Aided Fixed Point Design of
E-UtraPrach Detector on Multi-Core DSP",In the proceeding of 17th European Signal
Processing Conference ,Glasgow, Scotland, pp. 1007-1011, 2009.
Rainer Storn and Kenneth Price, Differential Evolution - A simple and efficient adaptive
scheme for global optimization over continuous spaces, TR-95-012March 1995.
H. M¨uhlenbein, M. Schomisch, J.Born, \The Parallel Genetic Algorithm as Function
Optimizer", Proceedings of the Fourth International Conference on Genetic Algorithms,
University of California, San diego, pp. 270-278, 1991.
Dervis Karaboga and Selcuk Okdem, A Simple and Global Optimization Algorithm for
Engineering Problems: Differential Evolution Algorithm, Turk J Elec Engin, VOL.12, NO.1
2004.
H. MQ^ hlenbein, M. Schomisch, J. Born, "The Parallel Genetic Algorithm as Function
Optimizer",Proceedings of The Fourth International Conference on Genetic Algorithms,
University of California, San diego, pp. 270-278, 1991.
10
| 9cs.NE
|
NIELSEN EQUIVALENCE IN A CLASS OF RANDOM GROUPS
arXiv:1309.7458v2 [math.GR] 12 Jan 2016
ILYA KAPOVICH AND RICHARD WEIDMANN
Abstract. We show that for every n ≥ 2 there exists a torsion-free one-ended
word-hyperbolic group G of rank n admitting generating n-tuples (a1 , . . . , an )
and (b1 , . . . , bn ) such that the (2n − 1)-tuples
(a1 , . . . , an , 1, . . . , 1 ) and (b1 , . . . , bn , 1, . . . , 1 )
| {z }
| {z }
n−1 times
n−1 times
are not Nielsen equivalent in G. The group G is produced via a probabilistic
construction.
Introduction
The notion of Nielsen equivalence goes back to the origins of geometric group
theory and the work of Jakob Nielsen in nineteen twenties [N1, N2]. If G is a group,
n ≥ 1 and τ = (g1 , . . . , gn ) is an n-tuple of elements of G, an elementary Nielsen
transformation on τ is one of the following three types of moves:
(1) For some i ∈ {1, . . . , n} replace gi in τ by gi−1 .
(2) For some i 6= j, i, j ∈ {1, . . . , n} interchange gi and gj in τ .
(3) For some i 6= j, i, j ∈ {1, . . . , n} replace gi by gi gj±1 in τ .
Two n-tuples τ = (g1 , . . . , gn ) and τ = (g10 , . . . , gn0 ) of elements of G are called
Nielsen equivalent, denoted τ ∼N E τ 0 , if there exists a finite chain of elementary
Nielsen transformations taking τ to τ 0 . Since elementary Nielsen transformations
are invertible it follows that Nielsen equivalence is an equivalence relation on the
set Gn of n-tuples of elements of G for every n ≥ 1.
Nielsen showed [N1, N2] that for the free group Fn := F (x1 , . . . , xn ) any generating n-tuple of Fn is Nielsen equivalent to the standard basis (x1 , . . . , xn ). This
fact implies, in particular, that the automorphism group of a finitely generated
free group is generated by so called “Nielsen automorphisms” that correspond to
the elementary equivalences discussed above. It follows that for any group G two
n-tuples τ = (g1 , . . . , gn ), τ 0 = (g10 , . . . , gn0 ) ∈ Gn are Nielsen equivalent in G if and
only if there exist ϕ ∈ Hom(Fn , G) and α ∈ Aut(Fn ) such that the following hold:
(1) gi = ϕ(xi ) for 1 ≤ i ≤ n.
(2) gi0 = ϕ ◦ α(xi ) for 1 ≤ i ≤ n.
Thus Nielsen equivalence classes of n-tuples in G correspond to orbits in Hom(Fn , G)
under the natural right action of Aut(Fn ). In the case of generating tuples these
are orbits in Epi(Fn , G) under the action of Aut(Fn ).
If hg1 , . . . , gn i = G then an arbitrary h ∈ G can be expressed as a product
of elements of {g1 , . . . , gn }±1 . Therefore the (n + 1)-tuples (g1 , . . . , gn , 1) and
(g1 , . . . , gn , h) are Nielsen-equivalent. For similar reasons, if k ≥ 1 and h1 , . . . , hk ∈
G are arbitrary, then (g1 , . . . , gn , 1, . . . , 1) ∼N E (g1 , . . . , gn , h1 , . . . , hk ). It follows
| {z }
k times
2000 Mathematics Subject Classification. Primary 20F65, Secondary 20F67, 57M07.
The first author was supported by he Collaboration Grant no. 279836 from the Simons Foundation, and by the National Science Foundation grant DMS-1405146.
1
2
ILYA KAPOVICH AND RICHARD WEIDMANN
that if (h1 , . . . , hn ) ∈ G is another generating tuple for G then
(g1 , . . . , gn , 1, . . . , 1) ∼N E (g1 , . . . , gn , h1 , . . . , hn ) ∼N E (h1 , . . . , hn , 1, . . . , 1)
| {z }
| {z }
n times
in G.
n times
Recall that for a finitely generated group G the rank of G, denoted rank(G), is
the smallest number of elements in a generating set of G. A generating n-tuple of
G is called minimal if n = rank(G).
The operation of transforming an n-tuple of elements of G into an (n + 1)tuple by appending the identity element as the (n + 1)-st entry is sometimes called
a stabilization move. As we observed above, any two generating n-tuples of a
group G become Nielsen equivalent after n stabilization moves. The main goal of
this paper is to show that there exist groups that have two generating n-tuples
that do not become Nielsen equivalent after n − 1 stabilizations. Evans previously
showed [E2, E3] that for any k there exists a metabelian group G and generating
n-tuples that do not become Nielsen equivalent after k stabilizations. However, in
these example n is much larger than k and the generating tuples are not minimal.
The main result of this paper is:
Theorem A. Let n ≥ 2 be an arbitrary integer. Then there exists a generic set R
of 2n-tuples
τ = (v1 , . . . , vn , u1 , . . . , un )
where for each τ ∈ R every vi is a freely reduced word in F (a1 , . . . , an ), every ui is
a freely reduced word in F (b1 , . . . , bn ), where |v1 | = · · · = |vn | = |u1 | = · · · = |un |
and such that for each τ = (v1 , . . . , vn , u1 , . . . , un ) ∈ R the following holds:
Let G be a group given by the presentation
(*)
G = ha1 , . . . , an , b1 , . . . , bn |ai = ui (b), bi = vi (a), for i = 1, . . . , ni.
Then G is a torsion-free one-ended word-hyperbolic group with rank(G) = n, and
the (2n − 1)-tuples (a1 , . . . , an , 1, . . . , 1 ) and (b1 , . . . , bn , 1, . . . , 1 ) are not Nielsen| {z }
| {z }
n−1 times
n−1 times
equivalent in G.
The precise meaning of R being generic is explained in Section 5. However, the
main point of Theorem A is to prove the existence of groups G with the above
properties. Genericity considerations are used as a tool for justifying existence
of tuples τ such that G given by presentation (*) satisfies the conclusion of Theorem A. In an earlier paper [KW3] we proved, using rather different arguments
from those deployed in the present article, a weaker related statement. Namely,
we proved that if G is given by presentation (*) corresponding to a generic tuple
τ = (v1 , . . . , vn , u1 , . . . , un ), then the (n+1)-tuples (a1 , . . . , an , 1) and (b1 , . . . , bn , 1)
are not Nielsen equivalent in G. In [KW3] we also conjectured that Theorem A
should hold.
It is in general quite difficult to distinguish Nielsen equivalence classes of tuples
of elements. Even in the algorithmically nice setting of torsion-free word-hyperbolic
groups the problem of deciding if two tuples are Nielsen equivalent is algorithmically
undecidable. Indeed, the subgroup membership problem is a special case of this
problem, as two tuples (g1 , . . . , gn , h) and (g1 , . . . , gn , 1) are Nielsen equivalent if and
only if h ∈ hg1 , . . . , gn i. Therefore Nielsen equivalence is not decidable for finitely
presented torsion-free small cancellation groups, since in general, by a result of
Rips [R], such groups have undecidable subgroup membership problem. Only in
the case of 2-tuples there is a test that works reasonably well for distinguishing
Nielsen equivalence: If two pairs (g1 , g2 ) and (h1 , h2 ) are Nielsen equivalent in a
group G, then, by a result of Nielsen [N1], the commutator [g1 , g2 ] is conjuate
NIELSEN EQUIVALENCE IN A CLASS OF RANDOM GROUPS
3
to [h1 , h2 ] or two [h1 , h2 ]−1 . Thus having a solution to the conjugacy problem
often gives a way to distinguish Nielsen classes of pairs. This test is particularly
important in the theory of finite groups.
As noted above, there are no general methods that allow us to distinguish Nielsen
classes. However, Lustig and Moriah [LM1, LM2, LM3] and Evans [E1, E2, E3]
produced algebraic methods that work in many special situations. While the methods of Lustig and Moriah only work in the context of minimal generating tuples,
the methods of Evans also work for non-minimal and stabilized generating tuples;
see the discussion before the statement of Theorem A. The methods of Moriah and
Lustig are K-theoretic in nature. The approach of Evans generalizes some linear
algebra considerations in the context of solvable groups. Our approach, sketched
in the next section, is essentially geometric. Negatively curved features of groups
G, given by presentation (∗) in Theorem A, play a crucial role in the proof of this
theorem.
We are grateful to the two referees of this paper for useful suggestions.
1. Outline of the proof
In this section we outline the idea of the proof of Theorem A. First, we pass
from presentation (*) of G to the presentation
G = ha1 , . . . , an |U1 , . . . , Un i
(**)
where Ui is obtained by freely and cyclically reducing the word ui (v1 , . . . , vn )a−1
in
i
F (A), where A = {a1 , . . . , an }. Genericity of (*) implies that presentation (**) satisfies an arbitrarily strong small cancellation condition C 0 (λ), where λ ∈ (0, 1) is an
arbitrarily small fixed number. Suppose that the (2n−1)-tuples (a1 , . . . , an , 1, . . . , 1 )
| {z }
n−1 times
and (b1 , . . . , bn , 1, . . . , 1 ) are Nielsen-equivalent in G. Then in the free group F (A)
| {z }
n−1 times
the (2n − 1)-tuple (a1 , . . . , an , 1, . . . , 1 ) is Nielsen equivalent to a tuple of freely
| {z }
n−1 times
reduced F (A)-words (x1 , . . . , x2n−1 ) such that xi =G bi for i = 1, . . . , n. Among
all such tuples (x1 , . . . , x2n−1 ) we choose one of minimal complexity, in the appropriate sense. We then consider the labelled graph Γ given by a wedge of 2n − 1
loops labelled by the words x1 , . . . , x2n−1 . The fact that (a1 , . . . , an , 1, . . . , 1 ) ∼N E
| {z }
n−1 times
(x1 , . . . , x2n−1 ) in F (A) implies that the graph Γ admits a finite sequence of Stallings
folds taking it to the graph Rn which is the wedge of n loop-edges labelled a1 , . . . , an .
In this folding sequence we look at the last term ∆ such that ∆ does not contain
a copy of Rn as a subgraph. Then Γ folds onto ∆ and ∆ contains a subgraph Ψ
consisting of ≤ n + 2 edges such that Ψ transforms into a graph containing Rn
by a single fold. By construction, the first Betti number of ∆ is ≤ 2n − 1, which
puts combinatorial constrains on the topology of ∆ which we later exploit. The
F (A)-words x1 , . . . xn are readable in Γ and therefore they are also readable in ∆.
In Section 3 and Section 4 we study how generic words can be read in labeled
graphs of Betti number at most 2n − 1, in particular we show that a path reading
such a word only travels few edges more than once unless the graph contains Rn or
a 2-sheeted cover of Rn , see Theorem 4.3. This gives us control over how generic
words are read in ∆.
Using genericity and small cancellation theory considerations and the fact that
xi =G bi for 1 ≤ i ≤ n, we conclude (see Section 5) that generically one of the
following holds:
4
ILYA KAPOVICH AND RICHARD WEIDMANN
Γ
x2n−1
∆
x1
x3
x2
Rn
Figure 1. ∆ is the last graph in the folding sequence that doesn’t
contain a copy of Rn . Ψ is the fat subgraph of ∆.
(1) xi = vi for 1 ≤ i ≤ n.
(2) For some i ∈ {1, . . . , n} the word xi contains a subword Z which is almost
all of a cyclic permutation U∗ of some Uj±1 .
In Section 7 we show that both cases yield a contradiction to either the genericity
of the presentation or the minimality assumption mentioned above.
In case (1), genericity implies that for each i the path γvi in ∆, that reads the
word vi , runs over a long (subdivided) segment of ∆\Ψ that is not visited by γvj
for j 6= i. Therefore the Betti number of ∆/Ψ ist at least n, which is impossible as
b(∆) ≤ 2n − 1 and b(Ψ) ≥ n.
In case (2), the subword Z in the label of the i-th petal of Γ projects to a path
sZ with label Z in ∆. Since words readable in Ψ are highly non-generic, we exploit
the genericity properties of the presentation to conclude that there is some ”long”
(comprising a definite portion of the length of Uj ) arc Q0 in ∆\Ψ with label W such
that sZ runs over Q0 exactly once. We then perform a surgery on ∆ by replacing
Q0 in ∆ by an arc labelled by the word complimentary to W in U∗ word V (so that
U∗ ≡ W V −1 in F (A)), to obtain a graph ∆0 . Genericity considerations allow us to
conclude that the full pre-image of Q0 in Γ consists of a disjoint collections of arcs
labelled W inside the petals of Γ. We replace each such arc by an arc labelled V to
obtain a graph Γ0 which is a wedge of 2n − 1 circles labeled z1 , . . . , z2n−1 . Let x0i be
the word obtained from zi by free reduction. By construction the graph Γ0 still folds
onto ∆0 . Since ∆0 still contains the subgraph Ψ, it follows that ∆0 folds onto Rn
and hence Γ0 folds onto Rn as well. Therefore (x01 , . . . , x02n−1 ) is Nielsen equivalent
in F (A) to (a1 , . . . , an , 1, . . . , 1 ). We then use results established in Section 6 to
| {z }
n−1 times
show that the tuple (x01 , . . . , x02n−1 ) has strictly smaller complexity than the tuple
(x1 , . . . , x2n−1 ), yielding a contradiction to the minimal choice of (x1 , . . . , x2n−1 ).
2. Preliminaries
2.1. Conventions regarding graphs. By a graph Γ we mean a 1-dimensional
cell complex, equipped with the weak topology. We refer to 0-cells of Γ as vertices
and to open 1-cells of Γ as topological edges. The set of all vertices of Γ is denoted
NIELSEN EQUIVALENCE IN A CLASS OF RANDOM GROUPS
5
V Γ and the set of topological edges of Γ is denoted Etop Γ. For each topological
edge e of Γ we fix a characteristic map τe : [0, 1] → Γ which is the attaching map
for e. Thus τe is a continuous map whose restriction to (0, 1) is a homeomorphism
between (0, 1) and the open 1-cell e, and the points τe (0), τe (1) are (not necessarily
distinct) vertices of Γ.
Each topological edge e is homeomorphic to (0, 1) and thus, being a connected
1-manifold, it admits exactly two possible orientations. An oriented edge or just
edge of Γ is a topological edge together with the choice of an orientation on it. We
denote by EΓ the set of all oriented edges of Γ. If e ∈ EΓ is an oriented edge,
we denote by e−1 ∈ EΓ the same topological edge with the opposite orientation.
Thus −1 : EΓ → EΓ is a fixed-point-free involution on EΓ. For every oriented edge
e ∈ EΓ the attaching map for the underlying topological edge naturally defines the
initial vertex α(e) ∈ V Γ and the terminal vertex ω(e) ∈ V Γ of Γ. Note that for
every e ∈ EΓ we have α(e) = ω(e−1 ) and ω(e) = α(e−1 ).
For a vertex v ∈ V Γ the degree deg(v) of v in Γ is the cardinality of the set
{e ∈ EΓ|α(e) = v}.
Let Γ and Γ0 be graphs. A morphism or a graph-map f : Γ → Γ0 is a continuous
map f such that f (V Γ) ⊆ V Γ0 and such that the restriction of f to any topological
edge (that is, an open 1-cell) e of Γ is a homeomorphism between e and some
topological edge e0 of Γ0 . Thus a morphism f : Γ → Γ naturally defines a function
(still denoted by f ) f : EΓ → EΓ0 . Moreover, for any e ∈ EΓ we have f (e−1 ) =
(f (e))−1 ∈ EΓ0 and α(f (e)) = f (α(e)), ω(f (e)) = f (ω(e)).
2.2. Paths, coverings and path-surjective maps. A combinatorial edge-path
or just an edge-path of length n ≥ 1 in a graph Γ is a sequence γ = e1 , . . . , en where
ei ∈ EΓ for i = 1, . . . , n and such that for all 1 ≤ i < n we have ω(ei ) = α(ei+1 ). We
denote the length n of γ as n = |γ| and also denote α(γ) := α(e1 ) and ω(γ) := ω(en ).
Also, for any vertex v ∈ V Γ we regard γ = v as an edge-path of length |γ| = 0 with
α(γ) = ω(γ) = v. An edge-path γ in Γ is called reduced if γ does not contain a
subpath of the form e, e−1 where e ∈ EΓ.
For a finite graph Γ (not necessarily connected), we denote by b(Γ) the first betti
number of Γ. If Γ is a graph and Γ0 ⊆ Γ is a subgraph, we denote by Γ/Γ0 the
graph obtained from Γ by collapsing every connected component of Γ0 to a vertex.
We will need the following elementary but useful lemma whose proof is left to
the reader:
Lemma 2.1. Let Γ be a finite graph and let Γ0 ⊆ Γ be a subgraph (not necessarily
connected). Then b(Γ) = b(Γ0 ) + b(Γ/Γ0 ).
If Γ is a connected non-contractible graph, the core of Γ, denoted by Core(Γ), is
the smallest connected subgraph Γ0 of Γ such that the inclusion map ι : Γ0 → Γ is a
homotopy equivalence. A connected graph Γ is called a core graph if Γ = Core(Γ).
Note that if Γ is a finite connected non-contractible graph then Γ is a core graph
if and only if for every v ∈ V Γ we have deg(v) ≥ 2. If Γ is a graph and v ∈ V Γ
a vertex then we call the pair (Γ, v) a core pair or a core graph with respect to v
if Γ contains no proper subgraph containing v such that the inclusion map is a
homotopy equivalence. Thus if (Γ, v) is a core graph then either Γ is a core graph
or Γ is obtained from a core graph Γ̄ by adding segment joining v and Γ̄.
Note that if f : ∆ → Γ is a morphism of graphs and if β = e1 , . . . , en is an
edge-path in ∆ of length n > 0 then f (β) := f (e1 ), . . . f (en ) is an edge-path in Γ
with |γ| = |β|. In this case we call β a lift of f (β) to ∆. Also, if u ∈ V ∆ and
v = f (u) ∈ V Γ then we call the length-0 path β = u a lift of the length-0 path
γ = v.
6
ILYA KAPOVICH AND RICHARD WEIDMANN
It is not hard to see that a graph morphism f : ∆ → Γ is a covering map (in
the standard topological sense) if and only if f is surjective and locally bijective.
It follows from the definitions that a surjective morphism f : ∆ → Γ is a covering
map if and only if for every vertex v ∈ V Γ, every edge-path γ in Γ with α(γ) = v
and every vertex u ∈ V ∆ with f (u) = v there exists a unique lift γ
e of γ such that
α(e
γ ) = u.
We will need to investigate the following weaker version of being a covering map:
Definition 2.2 (Path-surjective map). Let f : ∆ → Γ be a surjective morphism
of graphs. We say that f is path-surjective if for every reduced edge-path γ in Γ
there exists a lift γ
e of γ in ∆.
Let ∆ be a finite graph which contains at least one edge. A subgraph Λ of ∆
is an arc if ∆ is the image of a simple (possibly closed) edge-path γ = e1 , . . . , en ,
where n ≥ 1, such that for all 1 ≤ i < n the vertex ω(ei ) has degree 2 in ∆. If, in
addition, deg∆ (α(e1 )) 6= 2 and deg∆ (ω(en )) 6= 2, then Λ is called a maximal arc in
∆. An arc Λ in ∆ is semi-maximal if for the simple path γ = e1 , . . . , en defining Λ
either deg∆ (α(e1 )) 6= 2, deg∆ (ω(en )) = 2 or deg∆ (α(e1 )) = 2, deg∆ (ω(en )) 6= 2.
If ∆ is a finite connected graph with at least one edge and such that ∆ is
not a simplicially subdivided circle then every edge of ∆ is contained in a unique
maximal arc of ∆. Moreover in this case for every arc Λ of ∆ the simple edge-path
γ = e1 , . . . , en , whose image is equal to Λ, is unique up to inversion and the set of
its endpoints {α(e1 ), ω(en )} is uniquely determined by Λ.
3. A characterization of 2-sheeted covers of a rose
Definition 3.1 (Standard n-rose). For an integer n ≥ 1 we denote by Rn the
rose with n petals, i.e. the graph with a single vertex v0 and edge set ERn =
±1
{e±1
1 , . . . , en }. For 0 ≤ m < n we consider Rm as a subgraph of Rn in the obvious
way.
In this paper we will utilize some of the basic results about Stallings folds. We
refer the reader to [Sta, KM] for the relevant background information.
It is easy to explicitly characterize all the connected 2-fold covers of Rn . Namely,
a morphism p : Γ → Rn (where Γ is a connected graph) is a 2-fold cover if and only
if Γ has two distinct vertices v and w and for each i = 1, . . . , n there either exist
loop edges at v and w that map to ei or there exist two edges from v to w that
map to ei and e−1
i , respectively. The connectedness of Γ implies that for at least
one i the second option occurs.
e2
e1
e3
e1
e1
e3
p
e3
e2
e2
Figure 2. A 2-sheeted cover of R3
Note that if p : Γ → Rn is a morphism of graphs then p(V Γ) = {v0 } since Rn
is a graph with a single vertex v0 . Thus any such morphism p : Γ → Rn can be
NIELSEN EQUIVALENCE IN A CLASS OF RANDOM GROUPS
7
uniquely encoded by specifying for every edge e ∈ EΓ its “label” p(e) ∈ ERn =
±1
{e±1
1 , . . . , en }. Moreover, a labelling of EΓ by elements of ERn determines a
morphism if and only if this labeling respects edge inversion.
The main purpose of this section is to characterize 2-fold covers of Rn in terms
of path-surjective maps:
Theorem 3.2. Let Γ be a connected finite core graph, n ≥ 2 and p : Γ → Rn be a
path surjective morphism such that the following hold:
(1) Rn does not lift to Γ, i.e. there exists no morphism f : Rn → Γ such that
p ◦ f = idRn .
(2) b(Γ) ≤ 2n − 1.
Then p is a 2-sheeted covering map.
Note that it is easy to find a finite connected core graph Γ and a path-surjective
map p : Γ → Rn such that Γ does not contain a finite-sheeted cover of Rn . Indeed
taking Γ to be rose with (2n)(2n − 1) loops of length 2 and choosing f to be the
map that maps the (2n)(2n − 1) loops of Γ to the (2n)(2n − 1) distinct reduced
paths of length 2 in Rn does the trick. Thus the hypothesis on the Betti number
in Theorem 3.2 cannot be dropped. However, we do not know whether 2n − 1 is
the best possible lower bound.
Convention 3.3. For the remainder of this section we assume that n ≥ 2 and
p : Γ → Rn is as in Theorem 3.2. For any nonempty subset S ⊂ ERn let ΓS be the
subgraph of Γ formed by the edges that get mapped by p to some edge in S ∪ S −1 .
For S = {e} we will write Γe instead of Γ{e} . Note that for every e ∈ ERn we have
Γe = Γe−1 = Γ{e,e−1 } .
Lemma 3.4. The following hold:
(1) For every e ∈ {e1 , . . . , en } we have b(Γe ) ≥ 1.
(2) There exists e ∈ {e1 , . . . , en } such that b(Γe ) = 1.
Proof. To see that (1) holds let e ∈ ERn be arbitrary. Put k = |EΓ| + 1 and
consider the path γ := ek = e, . . . , e in Rn . Since p is path-surjective, there exists
| {z }
k times
a lift γ
e of γ to Γ, and, by definition of Γe , this path γ
e is contained in Γe . By the
choice of k the path γ
e contains at least two occurrences of the same oriented edge,
and hence γ
e contains a simple circuit embedded in Γe . Therefore b(Γe ) ≥ 1, as
claimed. Thus part (1) of the lemma is established.
Suppose now that part (2) fails. By part (1) this means that b(Γe ) ≥ 2 for every
e ∈ {e1 , . . . , en }. Since Γ = ∪ni=1 Γei , this implies that b(Γ) ≥ 2n, contrary to the
assumptions about Γ in Theorem 3.2. Thus part (2) of the lemma also holds.
In view of Lemma 3.4, after possibly permuting the indexing of the edges of Rn
we may assume that b(Γe1 ) = 1.
Lemma 3.5. There exists a permutation of {e1 , . . . , en } such that, after this permutation, for all k ∈ {1, . . . , n} we have
b(ΓSk ) ≤ 2k − 1
where Sk = {e1 , . . . , ek }.
Proof. The proof is by induction on k. As noted above, in view of Lemma 3.4 we
may assume that b(Γe1 ) = 1.
Suppose now that 2 ≤ k ≤ n and that we have already re-indexed the edges of
Rn so that b(ΓSi ) ≤ 2i − 1 for i = 1, . . . , k − 1. In particular, b(ΓSk−1 ) ≤ 2k − 3.
For 1 ≤ i < j ≤ n put Sij := {1, . . . , i} ∪ {j}.
8
ILYA KAPOVICH AND RICHARD WEIDMANN
Choose j ∈ {k, . . . , n} so that b(ΓS j
k−1
i
) = min{b(ΓSk−1
)|k ≤ i ≤ n}. Then re-
i
)
) ≤ b(ΓSk−1
index the edges ek , . . . , en so that j = k. Thus now b(ΓSk ) = b(ΓSk−1
k
for i = k, . . . , n.
By Lemma 2.1, we have
b(ΓSk ) − b(ΓSk−1 ) = b(ΓSk /ΓSk−1 ).
If b(ΓSk /ΓSk−1 ) ≤ 2 then b(Sk ) ≤ (2k − 3) + 2 = 2k − 1 and Sk has the properties
required by the lemma, completing the inductive step.
Thus suppose that b(ΓSk /ΓSk−1 ) ≥ 3. The minimality assumption in the choice
i
/ΓSk−1 ) ≥ 3. Thus for
of ek implies that for all i = k, . . . , n we have b(ΓSk−1
i = k, . . . , n the graph Γ/ΓSk−1 contains a graph of Betti number at least 3 all of
whose edges map to e±1
in Rn . Therefore
i
b(ΓSk /ΓSk−1 ) ≤ b(Γ/ΓSk−1 ) − 3(n − k).
Recall that by assumptions on Γ in Theorem 3.2 we have b(Γ) ≤ 2n − 1. Hence
b(ΓSk ) = b(ΓSk /ΓSk−1 ) + b(ΓSk−1 ) ≤ b(Γ/ΓSk−1 ) − 3(n − k) + b(ΓSk−1 ) =
= b(Γ) − 3(n − k) ≤ (2n − 1) − 3(n − k) ≤ 2k − 1,
as required. This completed the inductive step and the proof of the lemma.
Convention 3.6. For the remainder of the section we assume that the edges of
Rn are indexed so that b(Γe1 ) = 1 and that for every k = 1, . . . , n we have 1 ≤
b(ΓSk ) ≤ 2k − 1. Denote by Ce1 the unique embedded circuit in Γe1 and let Γ0e1 be
the component of Γe1 containing Ce1 .
For k = 1, . . . , n let ∆k be the core of the component of ΓSk containing Ce1 .
Thus b(∆k ) ≤ b(ΓSk ) ≤ 2k − 1.
Lemma 3.7. For every k = 1, . . . , n every reduced path γ in Rk lifts to a path in
∆k .
Proof. Let 1 ≤ k ≤ n and let γ be any reduced path in Rk . Choose , δ ∈ {−1, 1}
mδ
so that the path e1 γeδ1 is reduced. Put m = |EΓ| + 1. Thus γm = em
is a
1 γe1
reduced path in Rk . Since p : Γ → Rn is path-surjective, the path γm lifts to some
path γ
em in Γ. The choice of m implies that γ
em is contained in the component of
ΓSk containing Ce1 , that is, the component Y of ΓSk containing ∆k . Note that
∆k = core(Y ) and Ce1 ⊆ ∆k . Each of the two subpaths of γ
em corresponding to
mδ
the lifts of em
and
e
overlaps
C
and
hence
overlaps
core(Y
). Therefore the
e
1
1
1
portion of γ
em corresponding to the lift of γ is contained in core(Y ) = ∆k . Thus γ
lifts to ∆k , as claimed.
We now examine the properties of the graph ∆2 .
Proposition 3.8. One of the following holds:
(1) The restriction of p to ∆2 is a 2-sheeted cover onto R2 .
(2) R2 lifts to ∆2 .
Proof. Recall that Ce1 is the unique embedded circuit in Γe1 and that Γ0e1 is the
component of Γe1 containing Ce1 . Thus Ce1 is labelled by a power of e1 .
Note that there must exist a circuit Ce2 labeled by a power of e2 that lies in a
component Γ0e2 of Γe2 that intersects Γ0e1 in some (not necessarily unique) vertex v
as any path of type en1 en2 is assumed to lift to ΓS2 ⊂ Γ. We distinguish a number
of subcases, which cover all possibilities:
Subcase A: b(Γe2 ) = b(Γ0e2 ) = 2. In this case we have ∆2 ⊂ Γ0e1 ∪ Γ0e2 and
Γ0e1 ∩ Γ0e2 = {v} as ∆2 is a core graph with b(∆2 ) ≤ 3. It follows that R2 lifts to
NIELSEN EQUIVALENCE IN A CLASS OF RANDOM GROUPS
9
∆2 as by Lemma 3.7 the path e1 , e2 , e1 , e2 must have a lift which implies that for
i = 1, 2 there must be a loop edge based at v that is mapped to ei .
Subcase B: b(Γe2 ) = 2 and b(Γ0e2 ) = 1. Again we must have Γ0e1 ∩ Γ0e2 = {v} for
Betti number reasons. Moreover there exists a component Γ1e2 of Γe2 with Γ1e2 6= Γ0e2
and b(Γ1e2 ) = 1.
If Γ0e1 ∩ Γ1e2 = ∅ then any path of type en1 , en2 , en1 , en2 with n ≥ |EΓ| must lift to
∪ Γ0e2 as lifts are connected and any subpath eni must lift to a component of
Γei that contains a non-trivial circuit. Thus any path eni must lift to a closed path
in Γ0ei based at v for all n ≥ |EΓ|. As b(Γ0e1 ) = b(Γ0e2 ) = 1 this implies that Γ0ei
contains a loop edge based at v that is mapped to ei for i = 1, 2, thus R2 lifts to
∆2 .
If Γ0e1 ∩ Γ1e2 6= ∅ then Γ0e1 ∩ Γ1e2 must consist of a single vertex w for Betti
number reasons. Moreover we have ∆2 ⊂ Γ0e1 ∪ Γ0e2 ∪ Γ1e2 . By Lemma 3.7 the path
e1 , e2 , e1 , e2 , e1 , e2 , e1 lifts to ∆2 and the lifts of the occurrences of e2 must be loop
edges based at v or w. If the lifts of two consecutive occurrence of e2 are identical
then there must also be a loop edge mapped to e1 at the same base vertex and it
follows that R2 lifts. Otherwise there must be an edge from v to w and also an
edge from w to v that is mapped to e1 in which case the restriction of p to ∆2 is a
2-sheeted cover onto R2 .
Γ0e1
Subcase C: b(Γe2 ) = b(Γ0e2 ) = 1 and Γ0e1 ∩ Γ0e2 = {v}. In this case any path of
type en1 , en2 , en1 , en2 with n ≥ |EΓ| must lift to Γ0e1 ∪ Γ0e2 and we argue as in Subcase
B that R2 lifts.
Subcase D: b(Γe2 ) = b(Γ0e2 ) = 1 and Γ0e1 ∩ Γ0e2 = {v, w}.
If ∆2 contains no loop edge at v or w then there must be edges from v to w that
−1
−1 −1
map to e1 , e2 , e−1
1 and e2 as, by Lemma 3.7, the path e1 , e2 , e1 , e2 , e1 , e2 lifts
to ∆2 , and the lift must alternate between v and w. It follows that the restriction
of p to ∆2 is a 2-sheeted cover of R2 .
In the remaining case we assume by symmetry that Γ0e1 has a loop edge based
at v. We may assume that Γ0e2 does not have a loop edge based at v, otherwise
R2 lifts to ∆2 . Let now f be the unique edge of Γ0e1 ∩ ∆2 with α(f ) = w. Choose
ε ∈ {−1, 1} such that p(f ) = eε1 . It follows that any lift of eεn
1 for n ≥ |EΓ| to ∆2
terminates at v. As eεn
e
e
lifts
to
∆
and
as
there
is
no
loop
edge in Γ0e2 based
2
1 2 1
at v it follows that any lift of eεn
e
to
∆
must
terminate
at
w.
It
follows that the
2
1 2
−ε
path eεn
e
e
does
not
lift
to
∆
,
thus
this
case
does
not
occur.
2
1 2 1
Proof of Theorem 3.2. We claim that for 2 ≤ k ≤ n one of the following holds:
(1) The restriction of p to ∆k is a 2-sheeted cover onto Rk .
(2) Rk lifts to ∆k .
For k = n the above claim implies the statement of Theorem 3.2 since Γ =
core(Γ) = ∆n . Note that the claim above is false for k = 1.
We establish the claim by induction on k. The base case, k = 2, is exactly the
conclusion of Proposition 3.8.
Suppose now that 3 ≤ k ≤ n and that the claim has been verified for k − 1. Thus
we know that ∆k−1 is either a 2-sheeted cover of Rk−1 or that ∆k−1 contains a lift
of Rk−1 . In the first case we may moreover assume that there is an edge with label
e−1
1 and an edge with label e1 connecting the two vertices of ∆k−1 .
Subcase A: ∆k−1 contains a lift R̃k−1 of Rk−1 . If ∆k contains a loop edge at the
base vertex of R̃k−1 that is mapped to ek then ∆k contains a lift of Rk and there
is nothing to show. Thus we may assume that such a loop edge does not exist. We
will show that this yields a contradiction.
10
ILYA KAPOVICH AND RICHARD WEIDMANN
Let Γ1k be the core of the component of the subgraph Γ{e1 ,ek } containing the
unique loop edge mapped to e1 . Moreover let R1k be the subgraph of Rn consisting
±1
of the edge pairs e±1
1 and ek . By an argument similar to the proof of Lemma 3.7,
any path γ in R1k must lift to Γ1k . It now follows from the case k = 2 and the
fact that R1k does not lift to Γ that Γ1k is either a 2-sheeted cover of R1k or that
b(Γ1k ) ≥ 4. As Γ1k does not contain a second loop edge mapped to e1 it cannot be
a 2-sheeted cover, thus b(Γ1k ) ≥ 4. Let now Γ0 := Γ1k ∪ R̃k−1 . It follows that
b(Γ0 ) ≥ 4 + (k − 2) = k + 2.
Claim: For 2 ≤ i ≤ k − 1 either ∆k contains a circuit that maps to ei and that is
distinct from the loop in R̃k−1 or there exist two edges in E∆k − EΓ0 originating
at Γ0 that map to ei and e−1
i , respectively.
Proof of Claim: Suppose that no such circuit exists for some i ∈ {2, . . . k − 1}. We
need to show that for ε ∈ {−1, 1} there exists an edge in E∆k − EΓ0 originating
at Γ0 that maps to eεi . Suppose that ε ∈ {−1, 1} and that no such edge exists. It
ε·|EΓ|
follows that any lift of the path ei
must terminate at the base vertex v0 of R̃k−1
ε·|EΓ|
and any lift of ei
ek must terminate at a vertex distinct from v0 . This implies
ε·|EΓ|
that the path ei
ek eεi does not lift to ∆k , a contradiction. The claim follows.2
Now each additional circuit gives a contribution of at least 1 to the Betti number
and two edges originating at Γ0 do the same as we are looking at a core graph, thus
the loose ends must close up. It follows that
b(Γk ) ≥ b(Γ0 ) + (k − 2) = (k + 2) + (k − 2) = 2k
which gives the desired contradiction.
Subcase B: ∆k−1 is a 2-sheeted cover of Rk−1 and the edges with label e1 span
a circuit of length 2. It follows in particular that b(Γk−1 ) = 2k − 3.
Define Γ1k as in Subcase A, again we apply the case k = 2 to this graph. As Γ
contains no loop edge with label e1 and as for Betti number reasons b(Γ1k ) ≤ 3 it
follows that Γ1k is a 2-sheeted cover of R1k . It follows that Γ1k consists either of
a circuit of length two with label e1 , e1 and two loop edges with label ek or of two
circuits of length 2 with label e1 , e1 and ek , ek with common vertices. In both cases
it is immediate that ∆k−1 ∪ Γ1k is a 2-sheeted cover of Rk .
4. Lifting random paths in Rn
In this section all edges are geometric edges, i.e. edge pairs. Theorem 3.2 is only
used in this section, specifically in the proof of Theorem 4.3.
Definition 4.1. Let α ∈ [0, 1]. We say that a path γ in some graph Γ is α-injective
if γ crosses at least α · |γ| distinct topological edges. Thus γ is 1-injective if and
only if γ travels no topological edge twice. In other words, γ is α-injective if its
image is of volume at least α · |γ|.
Definition 4.2. Let Γ be a graph and let γ = e1 , . . . , ek be an edge-path in Γ.
We say that the edge ei is γ-injective if the topological edge of Γ corresponding
to ei is traversed by γ exactly once. Similarly, a subpath γ 0 = eq . . . er (where
1 ≤ q ≤ r ≤ k) of γ is γ-injective if for every i = q, . . . , r the edge ei is γ-injective.
Similarly, let Υ be a graph which is a topological segment subdivided into finitely
many edges, let Γ be a graph and let f : Υ → Γ be a graph map. We say that
an edge e of Υ is Υ-injective if for the edge-path γ determined by Υ the edge e is
γ-injective. An arc of Υ is Υ-injective if every edge of this arc is Υ-injective.
NIELSEN EQUIVALENCE IN A CLASS OF RANDOM GROUPS
11
In the following we denote the set of all reduced paths in Rn by Ω and the set
of all reduced paths of length N by ΩN . We further call a subset S ⊂ Ω generic if
N|
lim |S∩Ω
|ΩN | = 1. This definition of genericity agrees with the more detailed notions
N →∞
of genericity defined in Section 5 below.
Throughout this section we assume that n ≥ 2.
The purpose of this section is to establish the following theorem:
Theorem 4.3. Let α ∈ [0, 1). The set Ω of all reduced paths in Rn contains a
generic subset S such that the following holds:
Let s ∈ S and (Γ, v0 ) be a connected core graph with b(Γ) ≤ 2n − 1. Suppose
that f : Γ → Rn is a morphism such that Γ does not contain a finite subgraph Γ0
such that p|Γ0 : Γ0 → Rn is a covering (of degree 1 or 2). Then any lift s̃ of s is
α-injective.
Before we give the proof Theorem 4.3 we establish Proposition 4.4 and Proposition 4.8 which follow from elementary probabilistic considerations.
Proposition 4.4. For any ε > 0 there exists a constant L = L(ε, n) and a generic
subset S ⊂ Ω such that the following holds:
Suppose that s ∈ S and that γ is a path in Rn of length at least L such that γ
occurs as least twice as a subpath of s such that these two subpaths do not overlap.
Then the total length of any collection of non-overlapping subpaths of s that coincide
with γ is bounded from above by ε|s|.
For the remainder of this section, unless specified otherwise, we fix the following
conventions and notations.
Convention 4.5.
(1) Denote by ΩN the set of of all reduced paths of length N , endowed with the
uniform distribution. Let γ be a fixed path in Rn and put k := |γ|. Note that in
Rn there are exactly 2n(2n − 1)k−1 reduced paths of length k.
Now consider the random variable Y = Yγ,N : ΩN → R defined by
Yγ,N (s) := |γ| · max{l | s has l disjoint subpaths that coincide with γ}.
Thus Y assigns to any path s ∈ ΩN the total length of a maximal portion of s that
is covered by a collection of disjoint copies of γ. To prove Proposition 4.4 we will
need to estimate P (Y ≥ N ) from above. Our estimate will not depend on γ but
only on k = |γ|.
1
(2) Put p := (2n−1)
k . Note that for n ≥ 2 and k ≥ 1 we have 0 < 2p < 1. We
will also consider a random variable X = Xk that is binomially distributed with
parameters N and 2p.
In addition, let Ω0N = {0, 1}N with the distribution corresponding to performing
N independent tosses of a coin such that for a single toss the probability of the coin
landing ”heads” is 2p and of it landing ”tails” is 1 − 2p. Thus for a binary string
w ∈ Ω0N we have P (w) = (2p)t (1 − 2p)N −t where t is the number of 1’s in the string
w. Hence for every integer t ∈ [0, N ] P (Xk = t) is equal to the probability of the
event that a randomly chosen string w ∈ Ω0N contains exactly t entries 1.
(3) Consider the finite state Markov chain M generating freely reduced words in
Fn = F (a1 , . . . , an ). Thus the state set of M is Q = {a1 , . . . , an }±1 with transition
1
probabilities PM (a, b) = 2n−1
if a, b ∈ Q, b 6= a−1 and PM (a, b) = 0 if a, b ∈ Q
−1
and b = a . Note that M is an irreducible finite state Markov chain with the
stationary distribution being the uniform distribution on Q. Since n ≥ 2, whenever
a, b ∈ Q satisfy PM (a, b) > 0 then
1
4 1
3 1
(♣)
·
≤ PM (a, b) =
≤ ·
.
4 2n
2n − 1
3 2n
12
ILYA KAPOVICH AND RICHARD WEIDMANN
Lemma 4.6. There exists a constant C1 = C1 (n, ) ≥ 1 such that if k = |γ| ≥ C1
then for every N ≥ 1 and every integer t ∈ [0, N ] satisfying t ≥ N/k we have
(4/3)t pt ≤ (2p)t (1 − 2p)N .
Proof. Let n ≥ 2, N ≥ 1, k ≥ 1 be arbitrary integers and let > 0. Recall that
p = 1/(2n − 1)k and that 0 < 1 − 2p < 1.
Let t ∈ [0, N ] be an integer satisfying t ≥ N/k, clearly t ≥ 1. As t > 0 the
inequality (4/3)t pt ≤ (2p)t (1 − 2p)N is equivalent to (4/3)p ≤ 2p(1 − 2p)N/t which
in turn is equivalent to
(∗)
2/3 ≤ (1 − 2p)N/t .
Thus we need to show that (∗) holds provided k is sufficiently large. Since
t ≥ N/k, we have N/t ≤ k/. As 0 < 1 − 2p < 1, it follows that (1 − 2p)k/ ≤
(1 − 2p)N/t .
For fixed n and we have
k/
2
k→∞
(1 − 2p)k/ = 1 −
−−−−→ 1.
(2n − 1)k
Hence there exists a constant C1 = C1 (n, ) ≥ 1 such that for every k ≥ C1 we have
(1 − 2p)k/ ≥ 2/3. Then for any k ≥ C1
(1 − 2p)N/t ≥ (1 − 2p)k/ ≥ 2/3,
as required. Thus the lemma is proven.
Lemma 4.7. Let C1 = C1 (n, ) ≥ 1 be the constant provided by Lemma 4.6.
Suppose that k = |γ| ≥ C1 .
Then for every integer t ∈ [0, N ] such that t ≥ N/k we have
P (Yγ,N ≥ kt) ≤ P (Xk ≥ t).
Proof. For any integers 1 ≤ i1 < · · · < it ≤ N let V (i1 , . . . , it ) ⊆ ΩN be the event
that for s ∈ ΩN for each j = 1, . . . t the subpaths of s of length k starting in
positions i1 , . . . , it are equal to γ and that these subpaths do not overlap in s.
We first claim that for every tuple 1 ≤ i1 < · · · < it ≤ N we have
(♥)
P (V (i1 , . . . , it )) ≤ (4/3)t pt .
Note that the uniform distribution on ΩN is the same as the distribution on ΩN
obtained by taking the uniform distribution on Q, considered as giving the initial
letter of a word of length N , followed by applying the Markov chain M exactly
N − 1 times.
Choose a specific t-tuple 1 ≤ i1 < · · · < it ≤ N .
If N − it < k − 1, then the terminal segment of s starting with the letter in
position it has length < k and hence P (V (i1 , . . . , it )) = 0 ≤ (4/3)t pt . Similarly, if
there exists j such that ij − ij−1 ≤ k − 1 then the subpaths of s of length k starting
in positions ij−1 and ij overlap, and hence P (V (i1 , . . . , it )) = 0 ≤ (4/3)t pt . Thus
we assume that N − it ≥ k − 1 and that ij − ij−1 ≥ k. For similar reasons, we may
assume that whenever ij − ij−1 = k then the first letter of γ is not the inverse of
the last letter of γ since otherwise we again have P (V (i1 , . . . , it )) = 0.
By symmetry, the probability that the i1 -th letter of a random s ∈ ΩN is the
1
. Then, applying the Markov chain M defining
same as the first letter of γ is 2n
above we see that the probability of reading γ starting at letter number i1 in s is
1
1
1
equal to 2n(2n−1)
k−1 , so that P (V (i1 )) = 2n(2n−1)k−1 ≤ (2n−1)k = p.
Given that the event V (i1 ) occurred, consider the conditional distribution ν1 on
Q corresponding to the letter in s in the position i1 + k, i.e. the first letter in s
immediately after the last letter of the γ-subword that started in position i1 . By
NIELSEN EQUIVALENCE IN A CLASS OF RANDOM GROUPS
13
1
(♣) it follows that for every a ∈ Q we have ν1 (a) ≤ (4/3) 2n
. For each a ∈ Q let
Ha,i1 +k be the event that for an element of ΩN the letter in position i1 + k is a.
Let P (V (i2 )|Ha,i1 +k ) be the conditional probability of V (i2 ) given Ha,i1 +k . Then
the conditional probability of V (i1 , i2 ) given that V (i1 ) occurred can be computed
as
X
P (V (i1 , i2 )|V (i1 )) =
ν1 (a)P (V (i2 )|Ha,i1 +k ).
a∈Q
Also,
the unconditional probability P (V (i2 )) can be computed as P (V (i2 )) =
P
1
a∈Q 2n P (V (i2 )|Ha,i1 +k ). The same argument as the argument above for com1
1
puting P (V (i1 )) shows that P (V (i2 )) = 2n(2n−1)
k−1 ≤ (2n−1)k = p. Since for every
1
a ∈ Q we have ν1 (a) ≤ (4/3) 2n
, it follows that
X
P (V (i1 , i2 )|V (i1 )) =
ν1 (a)P (V (i2 )|Ha,i1 +k ) ≤
a∈Q
X 1
≤ (4/3)
P (V (i2 )|Ha,i1 +k ) = (4/3)P (V (i2 )) ≤ (4/3)p.
2n
a∈Q
Similarly, P (V (i1 , i2 , i3 )|V (i1 , i2 )) ≤ (4/3)p and, so on, up to
P (V (i1 , i2 , i3 , . . . , it )|V (i1 , i2 , . . . , it−1 )) ≤ (4/3)p.
Hence
P (V (i1 , . . . , it )) =
P (V (i1 ))P (V (i1 , i2 )|V (i1 )) . . . P (V (i1 , i2 , i3 , . . . , it )|V (i1 , i2 , . . . , it−1 ))
≤ (4/3)t−1 pt ≤ (4/3)t pt
as required. Thus (♥) is verified.
For 1 ≤ i1 < · · · < it ≤ N let W (i1 , . . . , it ) ⊆ Ω0N be the event that for a
binary string w ∈ Ω0N the digits in the positions i1 , . . . , it are equal to 1 and for
all j < it , j 6= i1 , . . . , it the digit in position j in w is 0. Then, by independence,
P (W (i1 , . . . , it )) = (2p)t (1 − 2p)it −t . Hence, by Lemma 4.6 and by (♥), we have
P (V (i1 , . . . , it )) ≤ (4/3)t pt ≤ (2p)t (1−2p)N ≤ (2p)t (1−2p)it −t = P (W (i1 , . . . , it )).
The event ”Yγ,N ≥ kt” is the (non-disjoint) union of events V (i1 , . . . , it ) taken
over all t-tuples 1 ≤ i1 < · · · < it ≤ N . Also, the event that w ∈ Ω0N has at
least t digits 1 is the disjoint union of W (i1 , . . . , it ), again taken over all t-tuples
1 ≤ i1 < · · · < it ≤ N . Therefore
X
P (Yγ,N ≥ kt) ≤
P (V (i1 , . . . , it )) ≤
1≤i1 <···<it ≤N
X
P (W (i1 , . . . , it )) = P (Xk ≥ t),
1≤i1 <···<it ≤N
as required. This completes the proof of Lemma 4.7.
Having established Lemma 4.6 and Lemma 4.7, we can now prove Proposition 4.4.
Proof of Proposition 4.4. Mean and variance for kXk are given by
2
µ(kXk ) = k · µ(Xk ) = k · N · 2p = k · N ·
(2n − 1)k
and
var(kXk ) = k 2 · var(Xk ) = k 2 · N · 2p · (1 − 2p) =
2
2
2
= k2 · N ·
· (1 −
) < k2 · N ·
.
k
k
(2n − 1)
(2n − 1)
(2n − 1)k
14
ILYA KAPOVICH AND RICHARD WEIDMANN
Choose L > C1 (n, ) such that µ(kXk ) ≤ 2 · N for all k ≥ L. It now follows
from Lemma 4.7 and the Chebyshev inequality (as N > 2 N ≥ µ(kXk )) that for
any reduced path γ with |γ| = k ≥ L we have
P (Yγ,N ≥ N ) ≤ P (Xk ≥
var(kXk )
N
≤
) = P (kXk ≥ N ) ≤
k
(N − µ(kXk ))2
8
k 2 · N · (2n−1)
k
4 · var(kXk )
var(kXk )
k2
8
=
≤
≤
=
.
·
(N − 2 N )2
(N )2
2 N 2
2 N (2n − 1)k
By Proposition 2.2 of [M] we know that we can assume that no path γ of length
11
occurs twice in a random word of length
greater than C0 ln(N ) with C0 = ln(2n−1)
N . Thus we only need to sum up the probabilities for all paths γ of length up to
C0 ln(N ). Thus, taking into account that for each k there are only 2n(2n − 1)k−1
words to consider, the probability that a path of length at least L and at most
C0 ln(N ) covers the portion · N of a path of ΩN is bounded from above by
C0 ln(N )
X
2n(2n − 1)k−1 ·
k=L
16n
8
k2
=
·
2
k
N (2n − 1)
(2n − 1)2 N
C0 ln(N )
X
k2 ≤
k=L
16n
· (C0 ln(N ))3 .
≤
(2n − 1)2 N
Now this number converges to 0 as N tends to infinity which proves the proposition.
Proposition 4.8. Let α ∈ (0, 1] and γ be a reduced path in Rn . There exists a
constant δ > 0 and a generic subset Sγ ⊂ Ω such that the following hold:
If s ∈ Sγ can be written as a reduced product s = s0 t1 s1 · . . . · tq sq such that
(1) ti does not contain γ as a subpath for 1 ≤ i ≤ q and
q
P
(2)
|ti | ≥ α · |s|.
i=1
Then q ≥ δ · |s|
Proof. The idea of the proof is simple: As an average subpath of a generic path
that doesn’t contain γ as a subpath must be short, i.e. bounded in terms of n and
γ, and as the ti cover a definite portion of s it follows that q must be large is |s| is
large.
We sketch a proof but omit some of the details for brevity. Suppose that γ is
1
a reduced path of length k in Rn . Let p = 2n(2n−1)
k−1 . For any j ∈ N≥1 let f (j)
be the probability that in a random semi-infinite reduced path in Rn , a subpath
between two consecutive non-overlapping occurrences of γ is of length j.
This defines a probability distribution f on N≥1 . This distribution is essentially
geometric as the probability f (l + k − 1) is approximately (1 − p)l . Note that this
calculation ignores the fact that the events that distinct subpaths are of type γ are
(slightly) correlated. It follows in particular that the probability distribution f has
∞
∞
P
P
finite mean, i.e that the series
j · f (j) converges. Put µ :=
j · f (j) and choose
j=1
j0 ∈ N with j0 > 100k such that
∞
X
j=j0
j=1
j · f (j) ≤
α
µ.
10
Let now s ∈ Ω. We consider the collection of subpaths that make up the complement of the union of all subpath of s that are of type γ; we call these subpaths γcomplementary. Then there exists a generic subset Ω0 ⊆ Ω such that for any s ∈ Ω0
the collection of γ-complementary subsets is non-empty. For each s ∈ Ω0 we obtain
NIELSEN EQUIVALENCE IN A CLASS OF RANDOM GROUPS
15
a probability distribution fs on N≥1 where fs (j) is the number of γ-complementary
subpaths of length j in s divided by the total number of γ-complementary subpaths
ns , i.e. fs is the relative frequency of γ-complementary subpaths of length j in s.
Then we have
∞
X
ns ·
j · fs (j) ≤ |s|
j=1
as the sum on the left is just the number of edges of s that lie in the γ-complementary
subpaths of s.
Claim: There exists some generic subset S ⊆ Ω0 ⊆ Ω such that for every s ∈ S we
have
∞
∞
X
α X
j · fs (j).
ns ·
j · fs (j) ≤ ns · ·
5 j=1
j=j
0
To verify the claim note first that it can be shown using the law of large numbers,
that there exists β ∈ [0, 1) such that for any > 0 there exists a generic set S such
that the following hold:
(1) For any s ∈ S
∞
X
(β − )|s| ≤ ns ·
j · fs (j) ≤ (β + )|s|.
j=1
(2) For any j ∈ {1, . . . j0 − 1} and s ∈ S we have
j · f (j)
j · f (j)
· |s| ≤ ns · j · fs (j) ≤ (β + ε) ·
· |s|.
µ
µ
Indeed, the first assertion states for a generic path s the γ-complementary paths
of s cover roughly a fixed portion β of s, i.e. their total length is roughly β|s|. The
second assertion states that out of this γ-complementary portion of total length β|s|,
the portion that consists of γ-complementary path of length j is roughly j·fµ(j) , i.e.
(β − ε) ·
that total length of the γ-complementary paths of length j is roughly
This conclusion follows from the definition of f and µ.
Using (2) we get that for any s ∈ S we have
|s| ·
jX
0 −1
(β − ε) ·
j=1
j·f (j)
µ
· |s|.
jX
jX
0 −1
0 −1
j · f (j)
j · f (j)
≤
ns · j · fs (j) ≤ |s| ·
(β + ε) ·
.
µ
µ
j=1
j=1
It follows that
ns ·
∞
X
j=j0
j · fs (j) = ns ·
∞
X
j=1
j · fs (j) − ns ·
jX
0 −1
j · fs (j) ≤
j=1
j0 −1
β− X
β−
α
≤ |s| · (β + ) − |s| ·
·
j · f (j) ≤ |s| β + −
(1 − ) · µ ≤
µ
µ
10
j=1
αβ
α
≤ |s| ·
+ 2−
.
10
10
For sufficiently small we further have
∞
X
αβ
α
α
α
|s| ·
+ 2−
≤ (β − ) · |s| ≤ · ns ·
j · fs (j)
10
10
5
5
j=1
which proves the claim.
Let now s ∈ S and suppose that s = s0 t1 s1 · . . . · tq sq is as in the formulation of
the lemma. Note that for 1 ≤ i ≤ q either |ti | ≤ 2k − 2 or ti contains a subpath
t0i of length at least |ti | − 2k + 2 such that t0i is contained in a γ-complementary
16
ILYA KAPOVICH AND RICHARD WEIDMANN
subpath of s. It follows from the claim that the total length of all γ-complementary
subpaths of s of length greater than j0 is bounded from above by
ns ·
∞
X
j=j0
j · fs (j) ≤ ns ·
∞
α X
α
·
j · fs (j) ≤ · |s|.
5 j=1
5
Thus the sum of the lengths of those ti that contain a subpath of a γ-complementary
subpath of length at least j0 is bounded by
102k α
α
j0 + 2k α
· · |s| ≤
· · |s| ≤ · |s|.
j0
5
100k 5
4
It follows the sum of the length of those ti that are of length at most j0 + 2k
must be at least 3α
4 · |s| thus there must be at least
1
3α
3α
·
· |s| =
· |s|
j0 + 2k 4
4j0 + 8k
such edges. Thus the conclusion of the lemma holds for δ :=
3α
4j0 +8k .
We can now give the proof of the main result of this section.
Proof of Theorem 4.3. Put β := 1 − α. Recall that in Subsection 2.2 we introduced
the notion of arcs, maximal arcs and semi-maximal arcs in a finite graph. Note
that it suffices to consider core pairs (Γ, v0 ) with b(Γ) = 2n − 1 as we can otherwise
embed (Γ, v0 ) into a core pair (Γ0 , v0 ) with b(Γ0 ) = 2n − 1 that still satisfies the
hypothesis of Theorem 4.3 and the conclusion for (Γ0 , v0 ) then implies the conclusion
for (Γ, v0 ).
Note that if ∆ is a finite connected core graph with first Betti number m ≥ 2,
then ∆ has no degree-1 vertices and ∆ is the union of at most 3m − 3 maximal
arcs (which are obtained but cutting ∆ open at all vertices of degree ≥ 3). Also, if
(∆, v0 ) is a finite connected core pair with first Betti number m ≥ 2, then ∆ is the
union of at most 3m − 1 maximal arcs.
Now let (Γ, v0 ) be as in Theorem 4.3 with b(Γ) = 2n−1 ≥ 3. As (Γ, v0 ) is a finite
connected core graph the above remark implies that Γ decomposes as the union of at
most 3(2n − 1) − 1 = 6n − 4 distinct maximal arcs. If γ is a non-degenerate reduced
edge-path in Γ then after possibly subdividing the maximal arcs of Γ containing
endpoints of γ along those endpoints, we obtain a decomposition of Γ as the union
of distinct pairwise non-overlapping arcs Λ1 , . . . Λt such that t ≤ 6n − 2 < 6n and
such that γ admits a unique decomposition as a concatenation of simple (and hence
reduced) edge-paths
γ = γ0 γ1 . . . γr
such that for each i = 0, . . . , r there exists y(i) ≤ t such that the image of γi is equal
to the arc Λy(i) . Note that since the arcs Λ1 , . . . Λt are non-overlapping, for every
i, j ∈ {0, . . . , r} either γi = γj±1 or the paths γi and γi have no to topological edges
in common. We call the above factorization of γ the canonical arc decomposition
of γ. We also call the arcs Λ1 , . . . Λt the arc components of Γ adapted to γ.
Claim A. We will show that there exist constants `0 , `1 , . . . , `2n and generic sets
S2n ⊆ S2n−1 · · · ⊆ S1 ⊆ S0 ⊆ Ω
that only depend on n and α such that the following hold:
If k ∈ {0, 1, . . . , 2n}, s ∈ Sk , (Γ, v0 ) and f : Γ → Rn as in the statement of
the theorem with b(Γ) = 2n − 1 then for any non-α-injective lift s̃ of s there exist
connected subgraphs Γ0 ⊂ Γ1 ⊂ . . . ⊂ Γk of Γ with the following properties:
(1) The volume of Γi is bounded by above by `i for 0 ≤ i ≤ k.
(2) b(Γi ) ≥ i for 0 ≤ i ≤ k.
NIELSEN EQUIVALENCE IN A CLASS OF RANDOM GROUPS
(3) If s̃ = e1 . . . eN then #{i|ei ∈ EΓ0 } ≥
β·N
6n
=
17
β·|s|
6n .
Note that for k = 2n this claim implies the theorem. Indeed it implies that the
generic set S := S2n has the property that for any s ∈ S any lift s̃ is α-injective as
a non-α-injective lift would imply the existence of a subgraphs of Γ of Betti number
2n which is impossible.
The proof of Claim A is by induction on k.
β
β
and let S and L = L(ε, n) = L( 6n
, n) be as in
The case k = 0. Let ε = 6n
the conclusion of Proposition 4.4. We will show that the claim holds for `0 := L
and S0 := S. Let s ∈ S0 , f : Γ → Rn as in the statement of Theorem 4.3, and let
s̃ a non-α-injective lift of s. As s̃ is not α-injective, it traverses fewer than α · |s|
distinct topological edges. This implies that there is a collection of subpaths of s̃
of total length at least β · |s| such that the image of these subpaths consists of all
the topological edges visited at least twice by s̃.
Since s is reduced and s̃ is a lift of s, the path s̃ is reduced as well. Let Λ1 , . . . , Λt ,
where t ≤ 6n, the the arc components of Γ adapted to s̃ and consider the canonical
arc decomposition of s̃. Then the union of topological edges visited at least twice by
s̃ is the union of some of the arcs Λi . Since t ≤ 6n, there exists i0 ≤ t such that the
subpaths of s̃ that map to Λi0 are of total length at least β·|s|
6n and that the number
of such subpaths is at least 2. It follows from the conclusion of Proposition 4.4 that
Λi0 has length at most L. Thus Claim A follows by choosing Γ0 = Λi0 and putting
`0 = L. Note that this choice of `0 is independent of the map f : Γ → Rn and s.
The case k ≥ 1. By induction we already have the constants `0 , . . . , `k−1 and
a generic set Sk−1 such that for any s ∈ Sk−1 and any non-α-injective lift s̃ of
s there exist subgraphs Γ0 ⊂ Γ1 ⊂ . . . ⊂ Γk−1 with the properties specified in
Claim A. Note that since the volume of Γk−1 is ≤ `k−1 , there are only finitely
many possibilities for Γk−1 and for the map f |Γk−1 : Γk−1 → Rn .
Claim B. We will now show that for any pair P := (Θ, pΘ ) consisting of a
graph Θ and a morphism pΘ : Θ → Rn there exists a constant `P and a generic set
SP ⊂ Sk−1 such that the following hold:
Suppose that f : Γ → Rn is as above, s ∈ SP , s̃ is a non-α-injective lift. Let Γk−1
be the subgraph of Γ whose existence is guaranteed by the induction hypothesis. If
P = (Γk−1 , f |Γk−1 ) then there exists a subgraph Γk of Γ containing Γk−1 such that
b(Γk ) > b(Γk−1 ) and that the volume of Γk is bounded by `P .
As there are only finitely many possibilities for P and as the intersection of
finitely many generic sets is generic, the conclusion of the inductive step for Claim A
follows from Claim B by taking `k to be the maximum of all occurring `P and taking
Sk to be the intersection of all SP .
To establish Claim B we need to show the existence of SP and `P for a fixed
pair P . Thus assume that f : Γ → Rn is as above, s ∈ Sk−1 , s̃ is a non-α-injective
lift and Γk−1 is as in the conclusion of the claim in the case k − 1 for s such that
P = (Γk−1 , f |Γk−1 ). By Theorem 3.2 there is a path γP in Rn that does not lift to
Γk−1 . Now write s̃ as a product
s0 t1 s1 · . . . · tq sq
where the ti are the path travelled in Γk−1 and where for 0 < i < q si is a nontrivial
(reduced) edge-path which does not pass through any edges of Γk−1 . Then the paths
ti do not contain a subpath that maps to γP as we assume that γP does not lift to
18
ILYA KAPOVICH AND RICHARD WEIDMANN
Γk−1 . The choice of Γ0 (see condition (3) above) implies that
q
X
i=1
|ti | ≥
β
· |s|.
6n
Proposition 4.8 implies that if s lies in the generic set SγP then q ≥ δ · |s| where
δ only depends on γP , n and α. By removing a finite subset from SγP (which
does not affect genericity of SγP ) we may assume that for every s ∈ SγP we have
δ|s| − 1 ≥ 1.
Then it follows that for the above decomposition of s̃ there exists i ∈ {1, . . . , q−1}
such that |si | ≤ |s|/(δ|s|−1). Indeed, if for all i = 1, . . . , q−1 |si | > |s|/(δ|s|−1) > 0
then
|s| ≥
q−1
X
|si | > (q − 1)|s|/(δ|s| − 1) ≥ (δ|s| − 1)|s|/(δ|s| − 1) = |s|
i=1
which is a contradiction.
Thus we can find i0 ∈ {1, . . . , q − 1} such that |si0 | ≤ |s|/(δ|s| − 1). Since
limN →∞ N/(δN − 1) = 1/δ > 0, by removing a finite set from SγP we may assume
that for every s ∈ SγP we have |s|/(δ|s|−1) ≤ 2/δ. Thus |si0 | ≤ |s|/(δ|s|−1) ≤ 2/δ.
By construction, the initial and terminal vertices of si0 are in Γk−1 = Θ, but
si0 does not traverse any edges of Γk−1 . Therefore for Γk := Γk−1 ∪ si0 we have
b(Γk ) ≥ b(Γk−1 ) + 1 ≥ (k − 1) + 1 = k. Thus the claim follows by taking `P :=
`k−1 + 2δ and SP = Sk−1 ∩ SγP . Thus Claim B is verified, which completes the
proof of Theorem 4.3.
5. Words representing bi .
Recall that we consider groups given by presentations of type
G = ha1 , . . . , an , b1 , . . . , bn |ai = ui (b), bi = vi (a)i
(*)
where n ≥ 2 and the ui and vi are freely reduced words of length N for some (large)
±1
by
N . If we now take the word/relator a−1
i ui (b) and replace all occurrences of bi
±1
vi (a) we obtain a relator in the ai . We denote by Ui the word obtained from this
word by free and cyclic cancellation. A simple application of Tietze transformations
shows that
G = ha1 , . . . , an | U1 , . . . , Un i.
(**)
We need an appropriate notion of genericity when working with groups given by
presentation (*). For A = {a1 , . . . , an }, with n ≥ 2, we think of F (A) as the set of
all freely reduced words over A±1 and we use the same convention for F (B) where
B = {b1 , . . . , bn }. Thus for N ≥ 1 there are exactly 2n(2n − 1)N −1 elements of
length N in F (A) (and same for F (B)). For N ≥ 1 we denote by T (A, N ) the set of
all n-tuples (v1 , . . . , vn ) ∈ F (A) such that |v1 | = · · · = |vn | = N . Similarly, T (B, N )
denotes the set of all n-tuples (u1 , . . . , un ) ∈ F (B) such that |u1 | = · · · = |un | = N .
Thus for every N ≥ 1 we have
n
#T (A, N ) = #T (B, N ) = 2n(2n − 1)N −1 = (2n)n (2n − 1)n(N −1) .
We denote by T (A, B, N ) the set of all 2n-tuples (v1 , . . . , vn , u1 , . . . , un ) such
that (v1 , . . . , vn ) ∈ T (A, N ) and (u1 , . . . , un ) ∈ T (B, N ). We also set T (A) =
∞
∞
∪∞
N =1 T (A, N ), T (B) = ∪N =1 T (B, N ) and T (A, B) = ∪N =1 T (A, B, N ).
Definition 5.1 (Generic sets). A subset S ⊆ F (A) is generic in F (A) if
lim
N →∞
#{v ∈ S : |v| = N }
#{v ∈ S : |v| = N }
= lim
=1
N
→∞
#{v ∈ F (A) : |v| = N }
2n(2n − 1)N −1
NIELSEN EQUIVALENCE IN A CLASS OF RANDOM GROUPS
19
A property of elements of F (A) is said to hold generically if the set of all v ∈ F (A)
satisfying this property is a generic subset of F (A).
A subset Y ⊆ T (A) is called generic in T (A) if
lim
N →∞
#(Y ∩ T (A, N ))
#(Y ∩ T (A, N ))
= lim
=1
N
→∞
#T (A, N )
(2n)n (2n − 1)n(N −1)
The notions of genericity for subsets of F (B) and of T (B) are defined similarly.
A subset T ⊆ T (A, B) is called generic if
#(T ∩ T (A, B, N ))
#(T ∩ T (A, B, N ))
= 1.
= lim
N →∞
N →∞ (2n)2n (2n − 1)2n(N −1)
#T (A, B, N )
lim
Definition 5.2 (Genericity for presentations).
We say that some property P for groups given by presentation (*) holds generically if the set TP of all (v1 , . . . , vn , u1 , . . . , un ) ∈ T (A, B), such that the group
G = ha1 , . . . , an , b1 , . . . , bn |ai = ui , bi = vi , for i = 1, . . . , ni satisfies P, is a generic
subset of T (A, B).
The following fact is a straightforward consequence of known results about genericity. Lemma 5.3 below is essentially a reformulation of the fact that a generic presentation with a fixed number of generators and a fixed number of defining relations
satisfies arbitrarily strong small C 0 (λ) cancellation condition, see [AO, A, KS, M].
Lemma 5.3. Let 0 < α < 1 be arbitrary. Then there is a generic subset T ⊆
T (A, B) such that for every τ = (v1 , . . . , vn , u1 , . . . , un ) ∈ T with |v1 | = · · · =
|vn | = |u1 | = · · · = |vn | = N the following hold:
For every freely reduced z ∈ F (A) with |z| ≥ αN there exists at most one i ∈
{1, . . . , n} and at most one ∈ {1, −1} such that z is a subword of vi . Moreover,
in this case there is at most one occurrence of z inside vi . Also, the same property
holds for every freely reduced z 0 ∈ F (B) with |z 0 | ≥ αN with respect to the words
u1 , . . . , un .
Lemma 5.3 implies in particular that generically there is very little cancellation
between the vi (a)±1 , and thus Ui is essentially just the concatenation of the vi (a)±1
introduced. The following statement is a straightforward consequence of standard
small cancellation arguments; see [AO, A, KS, M] for similar arguments.
Lemma 5.4. Generically the following hold:
1
(1) |Ui | ≥ N 2 (1 − 1010
n ) for 1 ≤ i ≤ n.
2ε1
1
(2) If W is a subword of length N 2 · 1010
and Uj2ε2 whose initial
n of both Ui
2ε1
2ε2
letter lies in the first half of Ui , respectively Uj , then i = j, ε1 = ε2
and the two occurrences of W in Uiε1 coincide, i.e. start at the same letter
of Uiε1 .
We refer the reader to [LS, O, Str] for basic results of small cancellation theory
and basic properties of small cancellation groups.
Lemma 5.3 implies that for any fixed L ≥ 1, generically, the above presentation
(*) is a C 0 ( L1 )-small cancellation presentation (note that the defining relations in
(*) are already cyclically reduced). This implies that for any α ∈ [0, 1) we may
assume that for any word in the a±1
and b±1
that represents the trivial element
i
i
in G there exists a subword R such that |R| ≥ α|RT −1 | where RT −1 is a cyclic
conjugate of a defining relator. In this case we call the process that replaces the
subword R by (its complementary word) T an α-small cancellation move (relative
to the relator RT −1 ). We will also consider α-small cancellation moves relative
to relators that are not cyclic conjugates of defining relators of (*); in particular
relative to the Ui , the defining relators of (**).
20
ILYA KAPOVICH AND RICHARD WEIDMANN
Lemma 5.5. Let α ∈ [0, 1). Let w be a reduced word in the a±1
such that w =G bi .
i
Then for generic G either w = vi or w admits a α-small cancellation move relative
to some cyclic conjugate of some Ui±1 .
In particular w can be transformed into vi by applying finitely many α-small
cancellation moves relative to cyclic conjugates of the relators Ui±1 and by free
cancellation.
Proof. This lemma can also be proven using the small cancellation properties of
(**); however we argue using the presentation (*).
The above discussion implies that the word b−1
i w can be transformed into the
trivial word by replacing subwords that are almost whole relators by the complementary word of the relator. If w 6= vi then this process can initially only involve
small cancellation moves relative to the relators b−1
i vi (a) as w is a word in the ai
and therefore b−1
w
does
not
contain
long
subwords
of the relators a−1
i
i ui (b).
±1
These initial moves introduce letters of type bj and possible remove the initial
±1
letter b−1
that were introi . Subsequent move of that type cannot remove the bj
duced as otherwise the original word w would have been non-reduced contradicting
the hypothesis.
We perform these moves as long as possible, in the end we must have a long word
in the bi with at most one ai in between that allows for a small cancellation move
with respect to some relation a−1
i ui (b) but then before the original moves almost
all of a cyclic conjugate of Ui must have occurred as a subword, indeed the origianl
subword can be obtainded by resubstituing the vi (a) for the bi , which gives almost
all of some Ui by definition of the Ui .
We also need to establish several basic algebraic properties of groups G given by
generic presentation (*). The proof of the following theorem follows the argument
of Arzhantseva and Ol’shanskii [AO, A] (see also [KS]).
Theorem 5.6. Let n ≥ 2. Then for the group G given by a generic presentation
(*) the following hold:
(1) Every subgroup of G, generated by ≤ n − 1 elements of G, is free.
(2) rank(G) = n.
(3) G is torsion-free, word-hyperbolic and one-ended.
Proof. By genericity of (*), given any 0 < λ < 1 we may assume that (∗) satisfies
the C 0 (λ) small cancellation condition.
(1) Choose a subgroup H ≤ G of rank ≤ n − 1. Thus H = hh1 , . . . , hk i ≤ G for
some h1 , . . . , hk ∈ G such that k ≤ n − 1.
Therefore there exists a finite connected graph Γ (e.g. a wedge of k circles labelled
by reduced words in A ∪ B representing h1 , . . . , hk ) with positively oriented edges
labelled by elements of A ∪ B such that the first Betti number of Γ is ≤ k and
such that for the natural “labeling homomorphism” µ : π1 (Γ) → G the subgroup
µ(π1 (Γ)) ≤ G is conjugate to H in G. Among all such graphs choose the graph Γ
with the smallest number of edges. Then, by minimality Γ is a folded core graph.
Since b(Γ) ≤ n − 1, Theorem 4.3 is applicable to Γ.
If µ : π1 (Γ) → G is injective, then H ∼
= π1 (Γ) is free, as required. Suppose
therefore that µ is non-injective. Then there exists a nontrivial closed circuit in Γ
whose label is a cyclically reduced word equal to 1 in G. Therefore there exists a
path γ̂ in Γ such that the label ẑ of γ̂ is a subword of some a cyclic permutation of
some defining relation r of (*) and that |ẑ| ≥ (1 − 3λ)|r| ≥ (1 − 3λ)N .
Without loss of generality, and up to a possible re-indexing, we may assume that
r is a cyclic permutation of b−1
1 v1 (a). Then ẑ contains a subword z such that z is
NIELSEN EQUIVALENCE IN A CLASS OF RANDOM GROUPS
21
also a subword of v1 and that |z| ≥ ( 12 − 32 λ)|r| − 1. Thus we may assume that
9
|z| ≥ 20
|r|.
Let γ be the subpath of γ̂ corresponding to the occurrence of z in ẑ. This implies
9
|r|, so that |r| ≤ 20
that |γ| = |z| ≥ 20
9 |γ|.
By genericity of (*) and by Theorem 4.3 we may assume, given an arbitrary
0 < α0 < 1, that γ̂ is α0 -injective. Since γ comprises almost a half of γ̂, it follows
that almost all edges of γ are γ̂-injective. Therefore, for any fixed 0 < α < 1,
no matter how close to 1, we may assume that the portion of γ corresponding to
γ̂-injective edges of γ covers the length ≥ α|γ|. We choose 0 < α < 1 and 0 < λ < 1
so that (9n − 12)λ 20
9 < α.
There are at most 3(n − 1) − 3 = 3n − 6 maximal arcs of Γ. These maximal
0
arcs subdivide γ as a concatenation of non-degenerate arcs γ = γ00 γ10 . . . γm
where
0
0
for i 6= 0, m γi is a maximal arc of Γ and for i = 0, m γi is either maximal or
0
semi-maximal arc in Γ. The paths γ00 and γm
may, a priori, overlap in Γ which
0
happens if the initial vertex of γ is contained in the interior of γm
and the terminal
0
0
0
vertex of γ is contained in the interior of γ0 . By subdividing γ0 and γm
along the
initial and terminal vertices of γ if needed, we obtain a decomposition of γ as a
concatenation of nondegenerate arcs of Γ
γ = γ 0 . . . γk
such that for 2 ≤ i ≤ k − 3 the path γi is a maximal arc of Γ and such that for i 6= j
either γi = γj±1 or γi and γj have no topological edges in common. Let τ1 , . . . , τq be
all the distinct arcs in Γ given by the paths γ0 . . . γk Since Γ has ≤ 3n − 6 maximal
arcs, we have q ≤ 3n − 4. Recall that the sum of the lengths of γi such that the arc
γi is γ̂-injective, is ≥ α|γ|.
Suppose first that there exists i0 such that γi0 is γ̂-injective and that |γi0 | > 3λ|r|.
Then we remove the arc corresponding to γi0 from Γ and add an arc τ from the
terminal vertex of γ̂ to the initial vertex of γ̂ with label ŷ such that ẑ ŷ is a cyclic
permutation of r. Thus |ŷ| = |τ | < 3λ|r|. Denote the new graph by Γ0 . Then the
image of the labelling homomorphism π1 (Γ0 ) → G is still conjugate to H, but the
number of edges in Γ0 is smaller than the number of edges in Γ. This contradicts
the minimal choice of Γ. Therefore no i0 as above exists.
Thus for every i such that γi is a γ̂-injective arc we have |γi | ≤ 3λ|r|. Hence the
total length covered by the γ̂-injective γi is
≤ (3n − 4) · 3λ|r| ≤ (9n − 12)λ
20
|γ| < α|γ|,
9
yielding a contradiction.
Therefore µ : π1 (Γ) → G is injective, and H ∼
= π1 (Γ) is free, as required.
(2) From presentation(*) we see that G = ha1 , . . . , an i. Thus rank(G) ≤ n.
Suppose that k := rank(G) < n. Then, by part (1), G is free of rank k < n.
Hence there exists a graph Γ with edges labelled by elements of A ∪ B such
that the first betti number of Γ is ≤ k and such that the image of the labelling
homomorphism µ : π1 (Γ) → G is equal to G. Among all such graphs choose Γ with
the smallest number of edges. By minimality, Γ is a folded core graph.
Since the first betti number of Γ is ≤ k < n, there exists a letter x ∈ A ∪ B
such that Γ has no loop-edge labelled by x. Without loss of generality we may
assume that x = a1 . Since the image of µ is equal to G, there exists a reduced
path γ̃ in Γ of length > 2 whose label z̃ is a freely reduced word equal to a1 in
−1
G. Thus z̃a−1
is freely reduced, it follows
1 =G 1. Whether or not the word z̃a1
that γ̃ contains a subpath γ with label z such that z is also a subpath of a cyclic
permutation of some defining relation r of (*) with |γ| = |z| ≥ (1 − 3λ)|r| − 1. Then
22
ILYA KAPOVICH AND RICHARD WEIDMANN
exactly the same argument as in the proof of (1) yields a contradiction with the
minimal choice of Γ.
Thus rank(G) < n is impossible, so that rank(G) ≥ n. Since we have already
seen that rank(G) ≤ n, it follows that rank(G) = n, as required.
(3) The group G is given by presentation (**) as a proper quotient of Fn . Since
Fn is Hopfian, it follows that G is not isomorphic to Fn . The fact that rank(G) = n
implies that G is not isomorphic to Fr for any 1 ≤ r ≤ n − 1. Thus G is not free.
Also, since (*) is a C 0 (1/6) small cancellation presentation for G where the
defining relators are not proper powers, basic small cancellation theory [LS] implies
that G is a torsion-free non-elementary word-hyperbolic group.
We claim that G is one-ended. Indeed, suppose not. Since G is torsion free and
not cyclic, Stalling’s theorem then implies that G is freely decomposable, that is G =
G1 ∗ G2 where both G1 and G2 are nontrivial. Therefore by Grushko’s Theorem,
n = rank(G) = rank(G1 ) + rank(G2 ) and hence 1 ≤ rank(G1 ), rank(G2 ) ≤ n − 1.
Part (1) of the theorem then implies that G1 and G2 are free. Hence G = G1 ∗ G2 is
also free, which contradicts part (2). Thus G is indeed one-ended, as claimed.
Remark 5.7. The proof of Theorem 5.6 can be elaborated further, following
a similar argument to that of [KS], to prove the following: For G given by a
generic presentation (*), if g1 , . . . , gn ∈ G are such that hg1 , . . . , gn i = G then
(g1 , . . . , gn ) ∼N E (a1 , . . . , an ) or (g1 , . . . , gn ) ∼N E (b1 , . . . , bn ) in G.
6. Reduction moves
Recall that for A = {a1 , . . . , an } we think of F (A) as the set of all freely reduced
words over A±1 .
We say that a word z ∈ F (A) is a U -word if z is a subword of of a power of some
Ui±1 (where Ui are the defining relators for the presentation (**)). Recall that by
1
Lemma 5.4 any U -word of length at least N 2 · 1010
n is a subword of a power of only
±1
one Ui . We say that a word V is U -complementary to a U -word W if W V −1 is a
power of a cyclic conjugate of some Ui±1 . Note that V is then also a U -word. Note
also that for a given U -word W , its U -complementary word is not uniquely defined.
Thus if W is an initial segment of U∗k , where k 6= 0 and U∗ is a cyclic permutation
of Ui±1 , and V is U -complimentary to W then for every integer m such that mk < 0
the word U∗m V is also U -complimentary to W .
We now introduce a partial order on F (A). First, for w ∈ F (A) we define c1 (w)
to be the minimum k such that w can be written as a reduced product of type
w1 · . . . · wk where each wi is a U -word. We call any such decomposition of w as
a product of c1 (w) U -words an admissible decomposition and we call the words
w1 , . . . , wk the U -factors of this admissible decomposition.
The number c1 (w) can be interpreted geometrically in the Cayley complex CC of
the presentation (**): If k = c1 (w), w1 · . . . · wk is as above and γw is a path reading
w then each wi is read by a subpath of γw that travels (possibly with multiplicity)
along the boundary of a 2-cell of CC.
Note that admissible decompositions are not unique. However the following
statement shows that the degree of non-uniqueness for admissible decompositions
can be controlled:
Lemma 6.1. Let w ∈ F (A), k = c1 (w) and let w = w1 · . . . · wk = w10 · . . . · wk0 be
admissible decompositions of w. Then the following hold:
(1) The subwords wi and wi0 overlap non-trivially in w for all i ∈ {1, . . . , k}.
NIELSEN EQUIVALENCE IN A CLASS OF RANDOM GROUPS
23
(2) If V is the minimal subword of w that contains both wi and wi0 then both
wi and wi0 contain the subword V 0 obtained from V by removing the initial
1
2
and terminal subword of length 1010
nN .
(3) If wi is subword of a power of Ujε of length at least 1015 n N 2 then wi0 is
also a subword of a power of Ujε and the common subword of wi and wi0
differs from wi and wi0 by at most an initial and terminal segment of length
1
2
1010 n N .
Proof. (1) Suppose that this does not hold, i.e. that for some i ∈ {1, . . . , k} the
words wi and wi0 do not overlap. Without loss of generality we assume that wi comes
0
before wi0 when reading w. Therefore wi0 wi+1
·. . . wk0 is a subword of wi+1 wi+2 ·. . .·wk ,
0 0
0
and, in particular, wi wi+1 · . . . wk is a product of k − i U -words. It then follows
that w is product of (i − 1) + (k − i) = k − 1 U -words, contradicting the definition
of c1 (w).
(2) Suppose that the conclusion of part (2) of the lemma does not hold. Without
loss of generality we may assume that w1 · . . . · wi−1 has a suffix w̄i of length at
1
2
0
least 1010
n N that is a prefix of wi . The subword w̄i is in fact a suffix of wi−1 as
0
otherwise wi−1 and wi−1 do not overlap, contradicting (1).
Thus wi−1 = uw̄i for some word u and wi0 = w̄i v for some word v. Now as
1
2
wi−1 and wi0 are U -words and w̄i is of length at least 1010
n N , it follows from
Lemma 5.4(2) that z = uw̄i v is a U -word. It follows that
0
w = w1 · . . . wi−2 zwi+1
· . . . wk0 ,
contradicting the assumption that k = c1 (w).
(3) This part follows immediately from the arguments used to prove (1) and (2)
together with Lemma 5.4(2).
Now let w be a freely reduced word over {a1 , . . . , an }±1 with c1 (w) = k. For
1 ≤ i ≤ k we will define a number `w (i) that essentially measures the length of the
i-th factor in an admissible decomposition of w, provided this factor is almost of
length N 2 . The precise definition of `w (i) is independent of the chosen admissible
decomposition of w and is more robust under certain modifications of w that we
will introduce later on.
Lemma 6.2. Let w ∈ F (A) and k = c1 (w). Fix i ∈ {1, . . . , k}. Choose an
admissible decomposition w = w1 · . . . · wk of w. Suppose that for some j 6= i the
subword wj is a maximal U -subword of w with |wj | ≥ 1013 n N 2 . Choose a word
w̄j that is U -complementary to wj that is also of length at least 1013 n N 2 and let
w̄ := w1 · . . . wj−1 w̄j wj+1 · . . . · wk .
Then the following hold:
(1) w̄ is freely reduced.
(2) w̄j is a maximal U -subword of w̄.
(3) c1 (w̄) = c1 (w).
Proof. (1) and (2) follow from the maximality of the U -subword wj and the fact
that w is reduced and that wj w̄j−1 is cyclically reduced. (3) follows from (1) and (2)
as the assumption on the length of wj and w̄j together with Lemma 6.1(3) imply
that we can find admissible decompositions of w and w̄ such that wj , respectively
w̄j , occur as factors.
For every k ≥ 1 denote by F (A; k) the set of all w ∈ F (A) with c1 (w) = k.
Definition 6.3 (i-equivalence). If w ∈ F (A; k), i ∈ {1, . . . , k} and w̄ is obtained
from w as in Lemma 6.2, we say that w̄ is elementary i-equivalent to w. Lemma 6.2
implies that w̄ ∈ F (A; k).
24
ILYA KAPOVICH AND RICHARD WEIDMANN
We further say that w ∈ F (A; k) is i-equivalent to a word w0 ∈ F (A) if w can be
transformed into w0 by applying finitely many elementary i-equivalences (note that
in this definition the value of the index i is fixed). Thus for every i ∈ {1, . . . , k}
i-equivalence is equivalence relation on F (A; k).
Lemma 6.4. Suppose that w, w0 ∈ F (A; k) are i-equvalent for some 1 ≤ i ≤
c1 (w) = c1 (w0 ) and that w = w1 · . . . · wk and w0 = w10 · . . . · wk0 are admissible
decompositions of w, respectively w0 . Then the following hold:
(1) There exists a word V ∈ F (A) such that wi = xV y and wi0 = x0 V y 0 where
x, y, x0 , y 0 are words of length at most 10110 · N 2 .
(2) The lengths of wi and wi0 differ at most by 10210 · N 2 .
Proof. This statement follows from the fact that an elementary i-equivalence that
modifies a wj only affects the adjacent factors and increases or decreases their length
1
2
by at most 1010
n N by Lemma 6.1, and that another such elementary i-equivalence
applied to the j-th factor reverses theses changes.
Definition 6.5 (Complexity of a word in F (A)). Let w ∈ F (A), k = c1 (w). For
each i ∈ {1, . . . , k} we define `ˆw (i) to be the maximum over all |wi | where wi is the
i-th factor in some admissible decomposition of a word w0 that is i-equivalent to w.
Note that part (2) of Lemma 6.4 implies that `ˆw (i) < ∞.
We further define `w (i) = 0 if `ˆw (i) < (1 − 1013 n )N 2 and `w (i) = `ˆw (i) otherwise.
Note that `w (i) = `w̄ (i) if w and w̄ are i-equivalent. We now define
c2 (w) := `w (1) + . . . + `w (c1 (w)).
For w ∈ F (A) we define the complexity of w as c(w) = (c1 (w), c2 (w)) and order
complexities lexicographically. This order is a well-ordering on F (A)
It is not hard to verify that an α-small cancellation move relative to Uj for α
close to 1 decreases the complexity. We will need the following generalization of
this fact:
Lemma 6.6. Let w ∈ F (A) be a reduced word and let W ∈ F (A) be a U -word of
1
N 2 . Suppose that w has (as a word) a reduced product decomposition of
length 100n
type
w = x0 W ε1 x1 W ε2 · . . . W εq xq
with εt ∈ {−1, 1} for 1 ≤ t ≤ q. Choose V such that W V −1 is a cyclic conjugate
of some Uj±1 and let w0 be the word obtained from the word x0 V ε1 x1 V ε2 · . . . V εq xq
by free reduction. Then the following hold:
(1) c(w0 ) ≤ c(w).
(2) If one of the subwords W εi is contained in a U -subword of w of length at
least (1 − 1014 n )N 2 , then c(w0 ) < c(w).
Proof. We will first observe that we may choose an admissible decomposition w =
w1 · . . . · wk (with k = c1 (w)) such that any subword W εt of w occurs inside some
factor wj of w. Indeed in any admissible decomposition of w the subword W εt can
overlap with at most 3 U -factors of w, as otherwise an admissible decomposition
of w with fewer U -factors can be found. Thus W εt overlaps with some factor in
1
N 2 . Because of Lemma 6.1, we can expand this
a subword of length at least 300n
εt
factor to contain W . The claim follows by making this modifications for all W εt .
Let now w = w1 · . . . · wk be an admissible decomposition such that any subword
W εi occurs inside some factor wj of w. Note that possibly several W εt occur inside
the same wj . Note also that if some W εt is contained in a U -subword x of length at
least (1 − 1014 n )N 2 then it is a subword of some wj of length at least (1 − 1024 n )N 2 .
NIELSEN EQUIVALENCE IN A CLASS OF RANDOM GROUPS
25
This conclusion can be established by first observing that by the argument before
another admissible decomposition w = z1 · . . . · zk (with U -factors zi ) can be chosen
such that x is contained in some zj , and then applying Lemma 6.1.
Replacing the W εt by the V εt inside each wj and performing free reduction
yields a new word wj0 , which is again a U -word. The word w0 is then obtained
from w10 · . . . · wk0 by free reduction. This construction implies, in particular, that
c1 (w0 ) ≤ c1 (w).
Thus we may assume that c1 (w) = c1 (w0 ) = k as there is nothing to show
otherwise. To prove the lemma it suffices to show that `w (i) ≥ `w0 (i) for all
1 ≤ i ≤ k and that `w (i) > `w0 (i) for some 1 ≤ i ≤ k if condition (2) is satisfied.
If condition (2) of the lemma is satisfied, there exists some factor wi of length
at least (1 − 1024 n )N 2 that contains one of the W εt . Therefore it is sufficient to
consider the following two cases:
Case 1: Suppose that wi contains at least one of the words W εt . Note that
`ˆw (i) = `w (i) ≥ N 2 if wi contains more than one such word.
Note first that `w0 (i) = 0 if `w (i) = 0, since the word obtained from wi be
replacing W ±1 with V ±1 followed by free cancellation is of length at most (1 −
2
1
2
2
100n )N and since i-equivalence cannot increase this length by more than 1010 n N
by Lemma 6.4.
If `w (i) 6= 0 then the replacements of W ±1 with V ±1 decrease the length of wi
2
2
0
by more than 1010
n N . Thus it follows from Lemma 6.4 that `w (i) < `w (i). This
holds in particular if wi is of length at least (1 − 1024 n )N 2 , which (by the above
discussion) is guaranteed to occur if condition (2) is satisfied.
Case 2:
Suppose that wi contains no subword W εt . To conclude the proof,
we need to show that `w (i) ≥ `w0 (i). Note that for any admissible decomposition
z1 · . . . · zk the words wi and zi have a non-trivial common subword that is not
touched by the replacement and free cancellations when going from w to w0 , since
otherwise c1 (w0 ) < k; this follows as in the proof of Lemma 6.1(1).
If w is i-equivalent to w0 then there is nothing to show as `w (i) = `w0 (i) by the
definition of `(i). However it may happen that w and w0 are not i-equivalent as
some of the U -words wj may be replaced by U -words that are too short for the
replacements to qualify as i-equivalences. This is the reason why the argument for
treating Case 2 is slightly more subtle.
Choose a word w00 that is i-equivalent to w0 such that w00 realizes `ˆw0 (i), i.e.
that w00 has an admissible decomposition w100 w200 · . . . · wk00 such that |wi00 | = `ˆw0 (i). It
suffices to establish:
Claim: The word w is i-equivalent to a word w̃ that has an admissible decomposition w̃1 w̃2 · . . . · w̃k such that |w̃i | = |wi00 |.
Indeed, the existence of such a word w̃ implies that
`ˆw (i) ≥ |w̃i | = |wi00 | = `ˆw0 (i)
which implies that `w (i) ≥ `w0 (i).
To verify the Claim, we first construct a word ẇ that is i-equivalent to w in the
following way: Start with w and perform the same replacements as when going from
w to w0 , but whenever one of the replacements results in a maximal U -subword v
of length less than 13 N 2 , multiply it with a cyclic conjugate of the appropriate Uj±1
to obtain a U -word longer than 13 N 2 that has v as both initial and terminal word
and such that the replacement is an elementary i-equivalence.
Call the resulting word ẇ. Note that ẇ is i-equivalent to w and all words that
occur as the i-th factors in any admissible decomposition of w0 also occur as i-th
factors in an admissible decomposition of ẇ. Now all elementary i-equivalences that
26
ILYA KAPOVICH AND RICHARD WEIDMANN
need to be applied to w0 to obtain w00 can be applied to ẇ. We obtain a word that
is i-equivalent to ẇ and that has wi00 as the i-th U -word. This proves the Claim
and completes the proof of the lemma.
7. Proof of the Main Theorem
In this section we establish Theorem A from the Introduction:
Theorem 7.1. Let n ≥ 2 be arbitrary. Then for the group G given by a generic presentation (*), the group G is torsion-free, word-hyperbolic, one-ended of rank(G) =
n, and the (2n−1)-tuples (a1 , . . . , an , 1, . . . 1) and (b1 , . . . , bn , 1, . . . 1) are not Nielsenequivalent in G.
Let G be given by a generic presentation (*). By Theorem 5.6 we know that G
is torsion-free, word-hyperbolic, one-ended and that rank(G) = n.
For a (2n − 1)-tuple τ = (w1 , . . . , w2n−1 ) of freely reduced words in F (A) define
the complexity c(τ ) of τ as
c(τ ) := (c(w1 ), . . . , c(wn )).
We order complexities lexicographically. Note that in this definition we completely
disregard the words wn+1 , . . . , w2n−1 .
We argue by contradiction. Suppose that the conclusion of Theorem 7.1 fails
for G, so that the (2n − 1)-tuples (a1 , . . . , an , 1, . . . 1) and (b1 , . . . , bn , 1, . . . 1) are
Nielsen-equivalent in G.
Then there exist elements x1 , . . . , x2n−1 in F (A) such that the (2n − 1)-tuple
(x1 , . . . , x2n−1 ) is Nielsen equivalent in F (A) to the (2n−1)-tuple (a1 , . . . , an , 1, . . . , 1)
and such that for some permutation σ ∈ Sn we have xi =G bσ(i) for 1 ≤ i ≤ n.
Minimality assumption: Among all (2n − 1)-tuples (x1 , . . . , x2n−1 ) of freely
reduced words in F (A) with the above property choose a tuple (x1 , . . . , x2n−1 ) of
minimal complexity.
This minimality assumption implies that c(xi ) ≤ c(xi+1 ) for i = 1, . . . , n − 1.
After a permutation of the bi we may moreover assume that σ = id. We fix
this minimal tuple (x1 , . . . , x2n−1 ) for the remainder of the proof. The minimality
assumption plays a crucial role in the argument and ultimately it will allow us to
derive a contradiction.
We choose k maximal such that xi = vi in F (A) for 1 ≤ i ≤ k. By convention if
x1 =
6 v1 , we set k = 0.
Note that if k < n (which we will show below) then it follows from Lemma 5.5
that xk+1 admits an α-small cancellation move relative to some Uj for α close to 1.
1
This implies in particular that we may assume that xk+1 is of length at least 99n
N 2.
Let now Γ be the wedge of 2n − 1 circles, wedged at a vertex pΓ , such that the
i-the circle is labeled by the word xi for 1 ≤ i ≤ 2n − 1. By assumption the xi form
a generating set of F (A), thus the natural morphism f from Γ to the rose Rn is
π1 -surjective as the loops of Γ map to a generating set of π1 (Rn ). It follows that Γ
folds onto Rn by a finite sequence of Stallings folds. Thus there exists a sequence
of graphs
Γ = Γ0 , Γ1 , . . . , Γl = Rn
and Stallings folds fi : Γi−1 → Γi for 1 ≤ i ≤ l such that f = fl ◦ . . . ◦ f1 . We may
choose this folding sequence such that we do not apply a fold that produces a lift
of Rn in Γi if another fold is possible. Choosing ∆ to be the last graph in such
a folding sequence that does not contain a lift of Rn it follows that there exists a
graph ∆ with the following properties:
(1) There exists a map f∆ : Γ → ∆ such that f : Γ → Rn factors through f∆ .
NIELSEN EQUIVALENCE IN A CLASS OF RANDOM GROUPS
27
(2) ∆ does not contain a lift of Rn .
(3) Any fold applicable to ∆ produces a lift of Rn , thus ∆ contains a connected
subgraph Ψ with at most n + 2 edges that folds onto Rn .
Put p∆ := f∆ (pΓ ). Thus the petal-circles of Γ get mapped to closed paths based
at p∆ .
As the Betti number of ∆ is bounded by 2n − 1 and as ∆ folds onto Rn it follows
that ∆ does not contain a 2-sheeted cover of Rn . Indeed such a 2-sheeted cover
would be of Betti number 2n − 1 and therefore be the core of ∆ which implies that
∆ does not fold onto Rn . It follows that the hypothesis of Theorem 4.3 is satisfied
for ∆. Thus for α arbitrarily close to 1 we may assume that any lift of of the path
γvi in Rn with label vi = vi (a) to ∆ is α-injective.
Lemma 7.2. We have k < n.
Proof. For i ∈ {1, . . . , k} the i-th loop of Γ maps to the path γvi in Rn . Thus the
i-th loop of Γ maps to a path in ∆ that is a lift of γvi , thus by Theorem 4.3 we can
assume that this image in ∆ is α-injective for α arbitrarily close to 1.
This implies that for 1 ≤ i ≤ k the image of the i-th loop of Γ in ∆ contains an
1
1
|vi | = 100n
N as the image of σi is the union of at
arc σi of ∆ of length at least 100n
most 6n arcs of ∆. By genericity these arcs are travelled only once by the first k
loops, in particular they are pairwise distinct. Moreover they are disjoint from Ψ.
It follows that there exists a connected subgraph Ψ̄ of ∆ containing Ψ such that
the arcs σi meet Ψ̄ only in their endpoints. As the Betti number of Ψ is at least n
it follows that the the subgraph of ∆ spanned by Ψ̄ and the σi has Betti number
n + k. As ∆ has Betti number at most 2n − 1 this implies that k ≤ n − 1 < n.
We now continue with the proof of Theorem 7.1.
Recall that Ui was obtained from the generic word a−1
i ui (b) by replacing each bj
by the respective vj (a). The genericity of the vj implies that there is very little cancellation in a freely reduced word in the vj , thus Ui is essentially the concatenations
of a−1
and the words vj that were introduced to replace the letters b1 , . . . , bn .
i
For each vj let vj0 be the subword of vj obtained by chopping-off small initial
and terminal segments of vi and such that for every occurrence of bj in a−1
i ui (b)
the word vj0 survives without cancellation when we replace b1 , .., bn by v1 , .., vn in
ai−1 ui (b) and then freely and cyclically reduce to obtain Ui .
As the cancellation is small relative to the length of the vj it follows that vj0
is almost all of vj . Moreover, we can choose vj0 so that |v10 | = · · · = |vn0 | = N 0 .
Choosing the vj from the appropriate generic set, we may assume that N ≥ N 0 ≥
N (1 − 0 ) for any fixed 0 < 0 < 1.
By Theorem 4.3 and by the genericity of the presentation (*), for any 0 < α0 < 1
we may assume that the conclusion of Theorem 4.3 holds for the words vj . It
follows that for any given 0 < α < 1 we may also assume that the words vj0
satisfy the conclusion of Theorem 4.3 as well. Indeed if we choose α0 = 12 (1 + α)
and 0 = 12 (1 − α) then any image of vj0 in some graph as in the hypothesis of
Theorem 4.3 contains at least α0 · N − 0 · N = α · N ≥ α · |vi0 | = αN 0 distinct
topological edges, so that the path corresponding to vj0 in the graph in question is
α-injective.
It follows from Lemma 5.5 that xk+1 , the label of the (k + 1)-st loop, admits an
α-small cancellation move with respect to some U
i with α close to 1. Thus xk+1
contains a subword Z of length at least 1 − 1014 n N 2 that is a subword of a cyclic
conjugate of some Ui . After dropping a short suffix and prefix we can assume that Z
is a reduced product of remnants of some vj±1 in Ui , and each such remnant contains
the corresponding (vj0 )±1 . We refer to the sub-words (vj0 )±1 of Z (understood both
28
ILYA KAPOVICH AND RICHARD WEIDMANN
as words and as positions in Z where they occur) arising in this way as v-syllables
of Z. Denote the arc in the (k + 1)-st loop of Γ that corresponds to Z by s̃Z . Note
that Z is readable in ∆ since Z is the label of the path sZ := f∆ (s̃Z ).
Convention 7.3. We fix the word Z, the arc s̃Z in Γ and its image sZ in ∆ for
the remainder of the proof of Theorem 7.1.
In the following two lemmas we will establish some important properties of the
word Z and of the path sZ .
Lemma 7.4. For 1 ≤ j ≤ n there exists a non-trivial distinguished subword ṽj of
vj0 (this means both the abstract word and the position in vj0 where it occurs) such
that the following hold:
(1) For any v-syllable (vj0 ) of Z with corresponding subword ṽj the subpath of
sZ corresponding to ṽj is contained in an arc of ∆\Ψ. (This means the
word ṽj is read on an arc of ∆\Ψ).
(2) If 1 ≤ j1 , j2 ≤ n, j1 6= j2 and (vj0 1 ) and (vj0 2 )δ are two v-syllables of Z,
then the subpaths of sZ corresponding to their distinguished subwords ṽj1
and ṽjδ2 have no topological edges in common.
(3) Similarly, if 1 ≤ j ≤ n and v 0 , v 00 are two distinct (i.e. appearing in
different positions in Z) v-syllables of Z both representing (vj0 ) and (vj0 )δ ,
then the subpath of sZ corresponding to the distinguished subword (ṽj ) ,
(ṽj )δ are either disjoint or coincide if = δ and are inverses if = −δ.
Proof. For a finite graph Λ let f (Λ) be the number of paths e1 , . . . , ek in Λ such
that {ei , e−1
6 {ej , e−1
6 {1, k}. Thus f (Λ) is
i } =
j } for i 6= j ∈ {1, . . . , k} and {i, j} =
the number of path that travel no topological edge twice except possibly that the
first and the last edges may agree. Let f (n) be the maximal f (Λ) where Λ is a
graph with the following properties:
(1) There exists a vertex v ∈ V Λ such that (Λ, v) is a core graph with respect
to v of Betti number at most 2n − 1.
(2) Λ has no valence-2 vertex.
Choose α = α(n) < 1 (close to 1) and d = d(n) > 0 (close to 0) such that
1
.
2
As noted above, by genericity we may assume that any path corresponding to vj0 is
α-injective. Moreover by Lemma 5.3 we may assume that vj0 does not contain any
word of length d(n) · |vj0 | twice. Let Ψ0 be the union of all maximal arcs of ∆ of
length less than d(n)|vj0 | = d(n)N 0 . Note that Ψ0 contains Ψ.
For j = 1, . . . , n we say that a reduced decomposition
f (n) · ((1 − α) + 6 · n · d(n)) ≤
(!)
is basic
hold:
(1)
(2)
(3)
vj0 = y0 p1 y1 . . . pm ym
if m ≥ 1 and there exists a way to read vj0 in ∆ such that the following
|pt | ≥ d(n) · |vj0 | = d(n)N 0 for 1 ≤ i ≤ m and pt is read on an arc of ∆ \ Ψ0 .
For 1 ≤ t ≤ m − 1 each yt is read Ψ0 .
y0 = y00 y000 where y000 is read in Ψ0 , y00 is read on an arc (possibly degenerate)
of ∆ \ Ψ0 and |y00 | < d(n)N 0 .
0 00
0
00
(4) ym = ym
ym where ym
is read in Ψ0 , ym
is read on an arc of ∆ \ Ψ0 and
00
0
|ym | < d(n)N .
For a basic decomposition (!) of vj0 , we say that a path β in ∆ corresponding to
reading vj0 in ∆ as in the above definition, is a basic path representing (!).
NIELSEN EQUIVALENCE IN A CLASS OF RANDOM GROUPS
29
Note that the yt may or may not be degenerate, i.e. may be single vertices of
∆. Note that whenever vj0 can be read in ∆ then the corresponding path is a basic
path corresponding to a basic decomposition. This implies that any subpath of sZ
corresponding to a subword of type vj0±1 is a basic path corresponding to a basic
decomposition.
Note that (2) implies that that the word pt is read along a maximal arc of ∆ \ Ψ0
if 2 ≤ t ≤ m − 1.
Moreover for any t1 6= t2 and {t1 , t2 } =
6 {1, m} the subpaths of β corresponding
to pt1 and pt2 have no topological edges in common by the choice of d(n). The
subpath corresponding to pm can overlap the subpath corresponding to p1 in a
subpath of length at most d(n)N 0 if they lie on the same arc of ∆ \ Ψ0 .
ˆ be the graph obtained from ∆ by collapsing all connected components
Let now ∆
of Ψ0 to points. If (!) is a basic decomposition of Vj0 and β is a basic path in ∆
00
ˆ is a path β̂ with label y 0 · p1 · . . . · pm · ym
representing (!), then the image of β in ∆
.
0
0
Moreover, the path β̂ is (d(n) · N )-almost edge-simple, that is, it travels at most
ˆ more than once. Note also that b1 (∆)
ˆ ≤ n − 1. Since
d(n) · N 0 topological edge of ∆
ˆ
ˆ
ˆ
(∆, p∆
ˆ ) is a core pair where p∆
ˆ is the image of p∆ in ∆, it now follows that ∆ has
at most 3n − 1 maximal arcs. Therefore for any basic decomposition (!) we have
m ≤ 3n as distinct pi lie on distinct maximal arcs, with the possible exception of
p1 and pm lying on the same maximal arc.
The number of tuples (p1 , . . . , pm ) (with each pt is understood as a subword
of vj0 occurring in a specific position in vj0 ) arising in basic decompositions of a
˜ be the
given vj0 , is bounded above by f (n). This can been as follows: Let ∆
ˆ
˜
graph obtained from ∆ by replacing maximal arcs by edges, ∆ has no vertex of
˜ where ei
valence 2. Now any tuple (p1 , . . . , pm ) gives rise to a path e1 , . . . , em in ∆
is the edge obtained from the arc on which pi was read by the basic path. This path
satisfies the condition given in the definition of f (n), and therefore there are at most
˜ ≤ f (n) different paths that occur. Each such path (e1 , . . . , em ) determines a
f (∆)
unique tuple (p1 , . . . , pm ): That pi is determined by ei is obvious for 2 ≤ i ≤ m − 1
as pi is simply the label of the maximal arc of ∆ corresponding to ei . For i = 1, m
this follows from the fact that subwords of length at least d(n)|vi0 | occur in only
one position of the word vi0 . This shows that at most f (n) tuples (p1 , . . . , pm ) can
occur.
We now show that the total length of all yt occurring in any fixed basic decompositions of vj0 is at most ((1 − α) + 6nd(n)) · |vj0 |.
Choose a fixed basic decomposition (!) of vj0 and let β be a basic path representing
(!). Each yt decomposes as a concatenation of arc subwords that are read along
00
maximal arcs in Ψ0 or the paths corresponding to y00 or ym
as β is read in ∆. Note
further that the sum of the lengths of y0 , . . . , ym exceeds the number of edges in
their image by at most (1 − α)|vj0 | as vj0 is α-injective. Each maximal arc in Ψ0
00
has length < d(n)|vj0 | and so do the paths corresponding to y00 and ym
. Thus the
number of edges in the image of the arcs corresponding to y0 , . . . , ym is bounded
00
from above by 6nd(n)|vj0 |, since ∆ has at most 6n − 1 maximal arcs and if y00 or ym
0
are non-degenerate then at most 6n − 2 of these arcs lie in Ψ . Therefore
m
X
|yt | ≤ ((1 − α) + 6nd(n)) · |vj0 |.
t=0
By definition of f (n), it follows that the total length of all yt occurring in all
possible basic decompositions of a given vj0 is at most
f (n) · ((1 − α) + 6nd(n)) · |vi0 | ≤
1 0
|v |.
2 j
30
ILYA KAPOVICH AND RICHARD WEIDMANN
Thus there exists a non-trivial subword ṽj of vj0 (we could actually choose ṽj
to be a 1-letter subword) such that in no basic decomposition of vj0 this subword
ṽj overlaps with any yt . Then the subwords ṽ1 , . . . , ṽn satisfy the requirements of
Lemma 7.4
Recall that in Definition 4.2 we defined the notions of an edge/arc of a graph
Υ being Υ-injective with respect to a graph map from Υ to another graph. Recall
also that in our situation the map f : Γ → Rn factors through f∆ : Γ → ∆. In the
following Υ-injectivity of some arc Υ in Γ refers to Υ-injectivity with respect to the
map f∆ . Recall that p∆ = f∆ (pΓ ) is the image of the base-vertex of Γ in ∆.
Recall that, as specified in Convention 7.3, the (k + 1)-st loop of Γ contains an
arc s̃Z labelled by a U -word Z of length at least (1 − 1014 n )N 2 , and that the path
sZ in ∆ is the image of s̃Z under p∆ .
1
Lemma 7.5. The arc s̃Z contains a s̃Z -injective subarc Q of length 100n
N 2 that
0
0
is mapped to an arc Q in ∆ such that Q does not contain the vertex p∆ in its
interior.
Proof. Let ṽ1 , . . . , ṽn be distinguished subwords of v10 , . . . vn0 provide by Lemma 7.4.
Let now Θ be the subgraph of ∆ consisting of the edges traversed by sZ .
We now construct a graph Θ̂ from Θ as follows:
First, collapse all edges of Θ that do not lie on the arcs that are images of the
subpaths of sZ corresponding to distinguished subwords ṽj of the v-syllables of Z.
The resulting graph Θ0 is the union of arcs labeled with the ṽj and two such arcs
intersect at most in the endpoints.
Now in Θ0 replace any such arc with label ṽj with an edge with label bj . We
±1
thus obtain a graph Θ̂ with edges labelled by letters of {b±1
1 , . . . , bn }. Note that Θ̂
has first Betti number at most n − 1, since all edges of Θ that lie in Ψ are collapsed
when constructing Θ0 .
Since Z is subword of length at least 1 − 1014 n N 2 of some cyclic permutation
of some Ui±1 , the word Z contains a subword Z 0 with the following properties:
49
(1) |Z 0 | ≥ 100
N 2.
0
(2) Z consists of the remnants of the vi after substituting the bi in ui .
Let sZ 0 be the subpath of sZ corresponding to the subword Z 0 of Z. Similarly
let s̃Z 0 be the subpath of s̃Z corresponding to Z 0 .
Now for j = 1, . . . , n replace in Z 0 any remnant of vjε by bεj . We thus obtain a
±1
±1
freely reduced word Ẑ in the {b±1
1 , . . . , bn } that is a subword of ui . Note that
49
|Ẑ| ≥ 100 N .
The word Ẑ is readable in Θ̂ by construction. Hence, by Theorem 4.3 and by
making the appropriate genericity assumptions on the ui , we may assume that the
corresponding path sẐ in Θ̂ is α-injective for some α > 1000n−1
1000n . It follows from
α-injectivity of sẐ that any subpath of sẐ of length
2(1 − α) · |Ẑ| + 1 ≤ (1 − α) · 3N
contains an sẐ -injective edge. This implies that any subpath of sZ 0 of length
1
1
((1 − α) · 3N + 1)N ≤ (1 − α) · 4N 2 =
N2 ≤
N2
250n
100n
contains an sZ 0 -injective edge.
Now write sZ 0 as the the product of 20n − 1 subpaths s1 , . . . , s20n−1 of length
1
49
greater or equal than 50n
N 2 . This can be done since |sZ 0 | ≥ 100
N 2 . By the
argument given above, it follows that every sj contains an sZ 0 -injective edge.
Since ∆ has at most 3(2n − 1) − 3 = 6n − 6 maximal arcs, it follows that the the
collection of all sZ 0 -injective edges in ∆ is the union of at most 6n − 5 sZ 0 -injective
NIELSEN EQUIVALENCE IN A CLASS OF RANDOM GROUPS
31
arcs of ∆. At most one of these sZ 0 -injective arcs of ∆ can contain p∆ in its interior.
After possibly subdividing this arc by p∆ in two, it follows that the collection of
all sZ 0 -injective edges in ∆ is the union of pairwise non-overlapping arcs t1 , . . . , tk ,
with k ≤ 6n and such that p∆ does not lie in the interior of any ti .
Since each subpath si of sZ 0 contains an sZ 0 -injective edge, it follows that each
si nontrivially overlaps some tj . We claim that there exist k0 and j such that
each of sk0 −1 , sk0 , sk0 +1 nontrivially overlaps tj . Otherwise each tj would overlap
at most two of the paths s1 , . . . , s20n−1 , and hence the paths t1 , . . . , tk overall at
most 2 · 6n = 12n < 20n − 1 of the paths s1 , . . . , s20n−1 . This contradicts the fact
that each si intersects some tj . Thus indeed there exist k0 and j such that each of
sk0 −1 , sk0 , sk0 +1 overlaps tj , so that tj contains sk0 . Note that sk0 does not contain
the vertex p∆ in its interior, since p∆ does not belong to the interior of tj .
1
N 2.
Thus sk0 is an arc of ∆ which is sZ 0 -injective and which has length ≥ 50n
By genericity of presentation (∗) we may assume that there is no word W of
1
N 2 such that W ±1 occurs more than once in Z. Since sk0 is an arc of
length 500n
1
length ≥ 50n N 2 and since Z is freely reduced, the path sZ cannot run over the
entire arc sk0 more than once. We also know that sk0 does occur as a subpath of
sZ , so that at some point the path sZ does run over the entire arc sk0 . It may also
happen that, in addition, the initial and/or terminal segment of sZ overlaps the
1
N 2 . Therefore
arc sk0 nontrivially; however such overlaps must have length ≤ 500n
1
sk0 contains a sub arc Q0 of length 100n
N 2 such that Q0 is sZ -injective.
1
0
We put Q be the lift of Q to Γ. Since |Q| = |Q0 | = 100n
N 2 , the conclusion of
0
Lemma 7.5 holds with these choices of Q and Q . Thus Lemma 7.5 is established.
Concluding the proof of Theorem 7.1. We now have all tools to conclude the proof
of the Theorem 7.1. Recall that we argue by contradiction and that we have
assumed that the 2n − 1-tuples (a1 , . . . , an , 1 . . . , 1) and (b1 , . . . , bn , 1 . . . , 1) are
Nielsen-equivalent in G.
We will see that we can apply Lemma 6.6 to reduce the complexity of xk+1 while
maintaining the complexity of xi for 1 ≤ i ≤ k and preserving the fact that xi =G bi
for 1 ≤ i ≤ n. This will yield is a contradiction to the minimality assumption about
(x1 , . . . , x2n−1 ) made at the beginning of the proof.
Recall that k was maximal such that xi = vi in F (A) for all 1 ≤ i ≤ k and that
by Lemma 7.2 we know that k ≤ n − 1. By convention if x1 6= v1 , we set k = 0.
Thus we always have 1 ≤ k + 1 ≤ n.
1
Let W be the word that is the label of the arc Q of length 100n
N 2 provided
1
2
by Lemma 7.5. Thus W is of length 100nN . Note that W is a subword of the
U -subword Z of length at least 1 − 1014 n N 2 . Recall that Q is a sub-arc of the
(k + 1)-st petal of Γ and that under the map Γ → ∆ the arc Q maps injectively to
an arc Q0 of ∆.
By Lemma 7.5 the base-vertex p∆ = f∆ (pΓ ) does not belong to the interior of
Q0 . Since the petals of Γ are labelled by freely reduced words, it now follows that
the full preimage of Q0 in Γ under f∆ is the union of a collection of pairwise nonoverlapping arcs Q1 , . . . , Qm in Γ such that Q = Q1 and such that each of the arcs
Q1 , . . . , Qm maps bijectively to Q0 .
For 1 ≤ i ≤ 2n − 1 write xi as a reduced product
(‡)
x0,i W ε1,i x1,i W ε2,i · . . . W εqi ,i xqi ,i
with εj,i ∈ {−1, 1} where the W εj,i are the occurrences of W ±1 in xi corresponding
to all those of the arcs Q1 , . . . , Qm that are contained in the i-th petal-loop of Γ.
32
ILYA KAPOVICH AND RICHARD WEIDMANN
Note that for 1 ≤ i ≤ k the length of xi = vi is too short to contain W as a
subword. Therefore for 1 ≤ i ≤ k we have xi = x0,i and qi = 0.
We now perform the change described in Lemma 6.6 simultaneously for all xi ,
i = 1, . . . , 2n − 1. That is, we pick a word V ∈ F (A) such that W V −1 is a cyclic
permutation of one of the defining relations Us±1 of presentation (**). Then for
each xi we replace each W j,i in the decomposition (‡) of xi by V j,i . Denote
the resulting word by zi and denote the freely reduced form of zi by x0i , where
i = 1, . . . , 2n − 1.
By construction, xi is unchanged for 1 ≤ i ≤ k, that is xi = x0i for 1 ≤ i ≤ k.
Moreover the complexity of xk+1 decreases as the presence of subword Z of xk+1
triggers clause (2) of Lemma 6.6. Thus we have c(xi ) = c(x0i ) for 1 ≤ i ≤ k and
c(x0k+1 ) < c(xk+1 ).
Since W =G V , we have xi =G x0i for 1 ≤ i ≤ 2n − 1. In particular we have
0
xi =G bi for 1 ≤ i ≤ n. Thus to get the desired contradition to the minimal choice
of (x1 , . . . , x2n−1 ) we only need to verify that the tuple (x01 , . . . , x02n−1 ) is Nielsenequivalent to (a1 , . . . , an , 1, . . . , 1) in F (A). This fact follows from the following
considerations:
In the graph ∆ we replace the arc Q0 by an arc Y with label V and call the
resulting graph ∆0 . The graph ∆0 still folds onto Rn , since ∆0 contains Ψ.
By Lemma 7.5 the vertex p∆ = f∆ (pΓ ) does not belong to the interior of Q0 .
Since the petals of Γ are labelled by freely reduced words, it now follows that the
full preimage of Q0 in Γ under f∆ is the union of a collection of pairwise nonoverlapping arcs Q1 , . . . , Qm in Γ such that Q = Q1 and such that each of the arcs
Q1 , . . . , Qm maps bijectively to Q0 . Note also that each of the arcs Q1 , . . . , Qm is
labelled by W amd that the base-vertex pΓ of Γ does not belong to the interior of
any of Q1 , . . . , Qm .
We obtain a new graph Γ0 from Γ by replacing each Qj by an arc Yj labelled
by V . Then there is a natural label-preserving graph map f∆0 : Γ0 → ∆0 which
sends each arc Qj bijectively to Y , and which agrees with f∆ on the complement
of Y1 ∪ · · · ∪ Ym in Γ0 . The map f∆ : Γ → ∆ arose from a sequence of folds and
therefore f∆ is π1 -surjective.
Topologically f∆0 can be viewed as the same as the map f∆ and hence the map
f∆0 is π1 -surjective. As we noted earlier, the graph ∆0 folds onto Rn , and hence
the map ∆0 → Rn is π1 -surjective. By composing this map with f∆0 , we conclude
that the canonical map Γ0 → Rn preserving edge-labels is π1 -surjective.
The graph Γ0 is a wedge of 2n−1 circles with labels z1 , . . . , z2n−1 . As noted above,
the canonical map Γ0 → Rn is π1 -surjective. Since π1 (Γ0 ) is generated by the looppetals with labels z1 , . . . , z2n−1 and since each zi freely reduces to x0i , it follows that
the elements x01 , . . . , x02n−1 of F (A) generate the entire group F (A) = F (a1 , . . . , an ).
Therefore the (2n − 1)-tuple (x01 , . . . , x02n−1 ) is Nielsen-equivalent in F (A) to the
(2n − 1)-tuple (a1 , . . . , an , 1, . . . , 1).
Recall that xi = vi = x0i for i = 1, . . . , k (where k < n) and that c(x0k+1 ) < c(xk ).
Therefore c(x01 , . . . , x02n−1 ) < c(x1 , . . . , x2n−1 ), which contradicts the minimality
assumption on (x1 , . . . , x2n−1 ).
Hence the 2n − 1-tuples (a1 , . . . , an , 1, . . . , 1) and (b1 , . . . , bn , 1, . . . , 1) are not
Nielsen-equivalent in G, and Theorem 7.1 is proved.
References
[A]
G. Arzhantseva, On groups in which subgroups with a fixed number of generators are
free, Fundam. Prikl. Mat. 3 (1997), no. 3, 675–683.
NIELSEN EQUIVALENCE IN A CLASS OF RANDOM GROUPS
[AO]
[D]
[E1]
[E2]
[E3]
[G]
[KM]
[KS]
[KW1]
[KW2]
[KW3]
[LGM]
[L]
[LP]
[LM1]
[LM2]
[LM3]
[LS]
[No]
[M]
[N1]
[N2]
[O]
[Pa]
[PRZ]
[R]
[Sta]
[Str]
[W]
33
G. Arzhantseva and A. Yu. Ol’shanskii, Generality of the class of groups in which
subgroups with a lesser number of generators are free. Mat. Zametki 59 (1996), no. 4,
489–496
T. Delzant Sous-groupes a deux generateurs des groupes hyperboliques. E. Ghys,
A. Haefliger and A. Verjovsky, (editors), Group theory from a geometrical viewpoint.
Proceedings of a workshop, held at the International Centre for Theoretical Physics in
Trieste, Italy, 26 March to 6 April 1990. Singapore, World Scientific, 1991, 177–189.
M.J. Evans Presentations of Groups involving more Generators than are neceessary,
Proc. London Math. Soc. 67 (1993), 106–126.
M.J. Evans Presentations of Groups involving more Generators than are neceessary
II, Cont. Math. 421(2006), 101–112.
M.J. Evans Nielsen equivalence classes and stability graphs of finitely generated groups,
Ischia Group theory 2006, World Scientific, 2007.
I. Grushko, On the bases of a free product of groups Mat. Sbornik 8 (1940), 169–182.
I. Kapovich, and A. Myasnikov, Stallings foldings and subgroups of free groups. J.
Algebra 248 (2002), no. 2, 608–668
I. Kapovich, and P. Schupp, Genericity, the Arzhantseva-Olshanskii method and the
isomorphism problem for one-relator groups. Math. Ann. 331 (2005), no. 1, 1–19
I. Kapovich and R. Weidmann, Freely indecomposable groups acting on hyperbolic
spaces, Internat. Jour. Algebra Comput. 14 (2004), 115–171
I. Kapovich and R. Weidmann, Kleinian groups and the rank problem, Geometry &
Topology 9 (2005), 375–402
I. Kapovich and R. Weidmann, Nielsen equivalence in small cancellation groups,
preprint, 2010; arXiv:1011.5862
C. Leedham-Green and S. Murray, Variants of product replacement. Computational
and statistical group theory (Las Vegas, NV/Hoboken, NJ, 2001), 97–104, Contemp.
Math., 298, Amer. Math. Soc., Providence, RI, 2002
L. Louder, Nielsen equivalence of generating sets for closed surface groups, preprint,
http://arxiv.org/abs/1009.0454
A. Lubotzky and I. Pak, The product replacement algorithm and Kazhdan’s property
(T). J. Amer. Math. Soc. 14 (2001), no. 2, 347–363
M. Lustig and Y. Moriah, Nielsen equivalence and simple homotopy type, Proc. London
Math. Soc. 62 (1991), 537–562
M. Lustig and Y. Moriah, Nielsen equivalence in Fuchsian groups and Seifert fibered
spaces, Topology 30 (1991) 191–204
M. Lustig and Y. Moriah, Generating systems of groups and Reidemeister-Whitehead
torsion, J. Algebra 157 (1993), 170–198
R. Lyndon and P. Schupp, Cominatorial Group Theory, Springer-Verlag, 1977
G.A. Noskov, Primitive elements in a free group, Math. Zametki 30 (1981), no. 4,
497–500
J. MacKay Conformal dimension and random groups, Geom. Funct. Anal. (GAFA),
22 (2012), 213–239
J. Nielsen Die Isomorphismen der allgemeinen, unendlichen Gruppe mit zwei Erzeugenden, Math. Ann. 78 (1917), no. 1, 385–397.
J. Nielsen Über die Isomorphismen unendlicher Gruppen ohne Relation, Math. Ann.
79 (1918), no. 3, 269–272.
A. Yu. Ol’shanskii, Geometry of defining relations in groups. Translated from the 1989
Russian original by Yu. A. Bakhturin. Mathematics and its Applications (Soviet Series),
70. Kluwer Academic Publishers Group, Dordrecht, 1991
I. Pak, What do we know about the product replacement algorithm?, Groups and computation, III (Columbus, OH, 1999), 301–347, Ohio State Univ. Math. Res. Inst. Publ.,
8, de Gruyter, Berlin, 2001
N. Peczynski, G. Rosenberger and H. Zieschang Über Erzeugende ebener diskontinuierlicher Gruppen Invent. Math. 29 (975), 161–180.
E. Rips, Subgroups of small cancellation groups, Bull. Lond. Math. Soc. 14 (1982),
45–47
J. R. Stallings Topology of Finite Graphs Invent. Math. 71 (1983), 551–565
Ralph Strebel, Small cancellation groups, in: “Sur les groupes hyperboliques d’apres
Mikhael Gromov.” Papers from the Swiss Seminar on Hyperbolic Groups held in Bern,
1988. Edited by E. Ghys and P. de la Harpe. Appendix, 227–273. Progress in Mathematics, 83. Birkhauser Boston, 1990
R. Weidmann, Generating tuples of free products, Bull. London Math. Soc., 39 (2007),
393–403.
34
[Z]
ILYA KAPOVICH AND RICHARD WEIDMANN
H. Zieschang, Über der Nielsensche Kürzungsmethode in freien Produkten mit Amalgam, Inventiones Math. 10 (1970), 4–37
Department of Mathematics, University of Illinois at Urbana-Champaign, 1409 West Green
Street, Urbana, IL 61801, USA
http://www.math.uiuc.edu/~kapovich/
E-mail address: [email protected]
Mathematisches Seminar, Christian-Albrechts-Universität zu Kiel, Ludewig-Meyn Str. 4,
24098 Kiel, Germany
E-mail address: [email protected]
| 4math.GR
|
Abstracting Event-Driven Systems with Lifestate Rules
Shawn Meier
Aleksandar Chakarov
Maxwell Russek
Sergio Mover
Bor-Yuh Evan Chang
arXiv:1701.00161v1 [cs.PL] 31 Dec 2016
University of Colorado Boulder
{shawn.meier, aleksandar.chakarov, maxwell.russek, sergio.mover, evan.chang}@colorado.edu
Abstract
We present lifestate rules—an approach for abstracting
event-driven object protocols. Developing applications
against event-driven software frameworks is notoriously
difficult. One reason why is that to create functioning
applications, developers must know about and understand the complex protocols that abstract the internal
behavior of the framework. Such protocols intertwine the
proper registering of callbacks to receive control from the
framework with appropriate application programming
interface (API) calls to delegate back to it. Lifestate rules
unify lifecycle and typestate constraints in one common
specification language. Our primary contribution is a
model of event-driven systems λlife from which lifestate
rules can be derived. We then apply specification mining
techniques to learn lifestate specifications for Android
framework types. In the end, our implementation is able
to find several rules that characterize actual behavior of
the Android framework.
1.
l.onClick(b)
b.onDraw()
b.enable()
b.enable()
b.disable()
b.disable()
l.onClick(b)
(a) Multi-object lifecycle
(b) Mixed lifecycle-typestate
b.disable()
b.enable()
b.disable()
b.onDraw()
b.enable()
b.disable()
b.onDraw()
b.enable()
b.disable()
b.enable()
l.onClick(b)
(c) Product of (a) and (b)
Figure 1. Event-driven object protocols are complex.
Introduction
complete. Because lifecycle constraints are so central
to implementing apps, they are typically discussed in
the framework documentation, but they are incomplete
enough that developers spend considerable manual effort
to derive more complete specifications [30].
To get a sense for the complexity of event-driven
object protocols, consider the automaton-based specifications shown in Figure 1.The meaning of such a
specification is that execution traces (projected on to
the actions of interest) must be words accepted by the
automaton. In Figure 1a, we describe a portion of a lifecycle specification for a protocol between a button object
b and its click-listener l. This automaton states that a
b.onDraw() callback notification happens before any
l.onClick(b) callback invocations. This specification
is a multi-object lifecycle constraint because it orders
callbacks on objects b and l.
In Figure 1b, we show another important property of
buttons and their click-listeners: a button can be enabled
or disabled via API calls b.enable() and b.disable(),
respectively. When button b is enabled (the upper state),
We consider the problem of specifying and mining the object protocols used by event-driven software frameworks.
Programming against event-driven frameworks is hard.
In such frameworks, programmers develop client applications (apps) against the framework by implementing callback interfaces that enable the application to be notified
when an event managed by the framework occurs (e.g.,
a user-interface (UI) button is pressed). The app may
then delegate back to the framework through method
calls to the application programming interface (API)
(e.g., to direct a change in the UI display). To develop
working apps, the application programmer must understand the complex object protocols implemented by the
event-driven framework. For example, the framework
may guarantee particular ordering constraints on callback invocations (known as lifecycle constraints), and
the application programmer may have to respect particular orderings of API calls (i.e., typestate constraints).
Unfortunately, such protocol specifications are complex to describe and maintain—and almost always in1
• We formalize a concrete semantics for event-driven
the click-listener l may receive the l.onClick(b) callback notification. Through an API call to b.disable(),
button b becomes disabled. When button b is disabled
(the lower state), the click-listener l can no longer receive
l.onClick(b) notifications. This specification mixes a
lifecycle (or callback-ordering) constraint with a typestate (or API call-ordering) constraint.
We call API calls into the framework, callins, to
clearly contrast them with callbacks that are calls back
from the framework. The specification in Figure 1b
mixes two kinds of invocations: (1) events labeled by
the callbacks they invoke to notify the app (written
slanted) and (2) callins that the app invokes to change
the state of the framework (written upright).
From just the examples in Figure 1, we see that the
specification space is rich and complex with multiple
interacting objects over callbacks and callins. Worse, the
composition of the onDraw-before-onClick specification
from (a) and the enable-disable-onClick specification
from (b) is the product automaton shown in (c). In this
product automaton, the left states are where button b
is enabled versus the right states where b is disabled,
and the upper states are where button b has yet to
be drawn versus the lower states are where b has been
drawn. While automata are descriptive and natural, the
key observation of this paper is that automata are more
precise than necessary or desired.
Thus, we take a departure from traditional automaton specifications and define a more abstract specification language consisting of lifestate rules that captures
the critical lifecycle and typestate constraints without
representing them as automata. We shall see that this
specification language is more compact than automata
and exposes lifecycle and typestate constraints as duals. At a high level, lifestate rules focus on changes in
the internal, hidden state of the event-driven framework.
For example, the enable-disable-onClick specification
from Figure 1b can be summarized as the following rules:
b.enable()
b.disable()
systems using an abstract machine model λlife that
captures an appropriate model of the internal, hidden
state of the framework (section 3). We introduce the
notion “allowedness” for API callins, which we shall
see is parallel to “enabledness” for events.
• We define a language for lifestate rules, which incorpo-
rates mixed lifecycle-typestate constraints (section 4).
• We introduce callback-driven trace slicing, a form
of trace slicing [14] adapted to events and callbacks
(section 5). The λlife model and this trace slicing
algorithm is the basis for our dynamic analysis tool,
DroidLife, that records events, callbacks, and callins
in running Android applications.
• We apply specification mining techniques to derive
lifestate specifications (section 6). In particular, we
apply unsupervised automata-based learning techniques based on probabilistic finite state automata
and hidden Markov models. We also develop a direct lifestate mining technique based on propositional
model counting (]SAT). The most significant challenge with lifestate mining is that the enables, disables,
allows, and disallows that explain callback and callin
invocations are not observable in concrete executions.
• We empirically evaluate our mining algorithms on 133
traces from a corpus of 1,909 Android apps (section 7).
Our results show that we can in fact learn several rules
corresponding to actual Android behavior and one of
which that was not in the Android documentation.
2.
Overview
In this section, we motivate the challenges in specifying
and mining event-driven object protocols and then
exhibit how our approach models event-driven systems
to mine lifestate rules about the Android framework.
We will show with our running example in Figure 2 that reasoning about the correctness of apps
requires a comprehensive understanding of the lifestate rules of various Android framework components.
In particular, the call marked with B can throw an
IllegalStateException. This code is correct, but understanding why requires a specification of subtle relationships between the various callins and the events
that trigger the callbacks. We will also see that a slight
change of this code makes it buggy.
enables l.onClick(b) for some l
disables l.onClick(b) for some l
The first rule says that whenever the invocation
b.enable() fires, the state of the framework changes
to enable the event that invokes the l.onClick(b) callback, and the second rule says when b.disable() fires,
the state changes to disable the event that invokes
l.onClick(b). These rules do not enumerate the explicit
invocation orders as in the automaton from Figure 1b.
Having defined lifestate rules for event-driven object
protocols, we investigate specification mining techniques
from dynamic traces of apps using a framework. To
address the technical challenges to this end, we make
the following contributions:
Lifestate Specification. Android is a prominent example of an event-driven framework that dispatches
a multitude of events (e.g., from user interaction or
sensors). An app gets notified about events by implementing callback interfaces that may then delegate back
to the framework through the framework’s callin interface. We consider an event-driven object protocol to be
the rules governing the use of an interface consisting of
2
callbacks and callins. Unfortunately, the protocol rules
are determined by the complex, internal behavior of the
framework, which becomes a major source of confusion
for app developers and thus of defects in apps.
To see this complexity, consider the code shown in
Figure 2. The app performs a time consuming operation (removing feeds) when the user clicks the button
b. The operation is performed asynchronously using
AsyncTask to not block the UI thread. Figure 2b shows
an execution trace of the app. Execution time flows
downwards. A box surrounds the callback and callin
invocations that happen as the result of processing of a
particular event. First event Create invokes the callback
(a:Activity).onCreate() for the object a. In this case,
object a has dynamic type MainActivity; for reasons we
will discuss below, we label the object with its most precise supertype that is a framework type. In the onCreate
callback, the code then invokes a callin for the AsyncTask
constructor (i.e., (t1:AsyncTask).<init>()) and then
the callin for setOnClickListener following the app
code at line 5. Then, the trace shows the execution of
other two events, Click, with the callback onClick, and
PostExecute, with the callback onPostExecute.
Figure 2b also depicts the back and forth flow of
control between framework and app code. The boxes
along the lines show the lifetime of activations. In
particular, we see that we can define a callback invocation
as the invocation that transfers control from framework
code to app code (arrows to the right), while a callin
is the reverse (i.e., an invocation that transfers control
from app code to framework). The difficulty in reasoning
about the code in Figure 2a is that framework behavior
that governs the possible control flow shown in Figure 2b
is obfuscated in the code.
Our running example is inspired by an actual issue in AntennaPod [1] a podcast manager for Android
with 100,000–500,000 installs from the Google Play
Store(a similar issue appears in the Facebook SDK
for Android [2]). The potential bug is that the callin
AsyncTask.execute, called via remover.execute() (line 7),
can throw an IllegalStateException.
In the example the execute call happens whenever
the user clicks button b, which in turn triggers the
event that dispatches to the OnClickListener.onClick
callback on the click-listener (set on line 5). This code
is indeed safe (i.e., does not throw the exception) because the developer enforces that the execute method
on FeedRemover is only called once. This property is
enforced by the disabling and enabling of button b
in onClick and in onPostExecute, respectively. Buttons in Android are disabled and enabled via calls to
setEnabled(false) and setEnabled(true), respectively
(corresponding to disable and enable in section 1).
3
class FeedRemover extends AsyncTask {
MainActivity a; Button b;
void doInBackground() { . . . remove feed . . . }
void onPostExecute() {
a.remover = new FeedRemover(a, b);
b.setEnabled(true);
}
1
2
}
class MainActivity extends Activity {
FeedRemover remover;
void onCreate() {
Button b = . . .;
remover = new FeedRemover(this, b);
b.setOnClickListener(new OnClickListener() {
void onClick(View v) {
b.setEnabled(false);
remover.execute();
B
}
});
}
3
4
5
6
7
}
(a) The call remover.execute() on line 7 (marked with B) can
throw an IllegalStateException if the remover task is already
running. This execute call happens whenever the user clicks button
b. This code is indeed safe because the developer enforces that
only one instance of FeedRemover is ever executing through the
disabling and enabling of the button on lines 6 and 2, respectively
(shown with dashed underline).
Framework
App
Create
(a:Activity).onCreate()
(t1:AsyncTask).<init>()
(b:Button).setOnClickListener(l:OnClickListener)
4
5
Click
(l:OnClickListener).onClick(b:Button)
(b:Button).setEnabled(false)
6
(t1:AsyncTask).execute()
7
PostExecute
(t1:AsyncTask).onPostExecute()
(t2:AsyncTask).<init>()
1
(b:Button).setEnabled(true)
2
(b) A DroidLife trace recorded dynamically. Each box is the
processing of an event. Callbacks are depicted as arrows to the
right from framework to app code; callins are arrows to the left,
which are labeled with the corresponding program location from
Figure 2a.
Figure 2. An example Android app whose safety depends on understanding the lifestate rules of two Android
framework components AsyncTask and Button.
(t:AsyncTask).execute()
small caps) and associate events with the callback(s)
they trigger (written Event onCallback()). Finally,
the last rule states that the asynchronous task constructor (t:AsyncTask).<init>() allows the t.execute()
callin. We define and discuss the formal semantics of
lifestate specifications in section 4.
We argue that specifications describing the enablingdisabling of events and the allowing-disallowing of callins
are central to event-driven object protocols. The specifications capture the relations that informally say, “When
a particular method invocation happens, the state of
the framework changes to enable or disable an event
or to allow or disallow a callin.” We unify the notions
of events and callins by introducing the term message.
To reason about lifestate specifications, we formalize an
abstraction of event-driven frameworks, called λlife , that
captures the internal state consisting of enabled and
allowed messages in section 3.
9 ci t.execute()
(b:Button).setEnabled(false)
9 evt Click (l:OnClickListener).onClick(b)
(t:AsyncTask).execute()
→ evt PostExecute t.onPostExecute()
(b:Button).setEnabled(true)
→ evt Click (l:OnClickListener).onClick(b)
(t:AsyncTask).<init>()
→ ci t.execute()
Figure 3. Some lifestate rules for AsyncTask and Button.
The first rule specifies a typestate property, while the
others are required to reason about the safety of the call
to remover.execute() in Figure 2a.
To see why this button disabling and enabling
enforces the single-executing-instance property, the
developer needs to know the interactions between
AsyncTasks, Buttons, and OnClickListeners in the Android
framework. Only with an understanding of the framework do we see that the call to b.setEnabled(false)
(line 6) disables the event that could trigger the
onClick callback (which calls the potentially problematic remover.execute()). Button b (and thus
the click event) can be enabled inside the callback
onPostExecute (line 2), but again with an understanding of the framework, we see that onPostExecute is
triggered by a post-execute event, which is in turn enabled by the call to remover.execute() (line 7). When
this happens, remover is set to a newly-allocated instance of FeedRemover (line 1), which allows for the
execute callin to be executed safely.
Without the calls to setEnabled that disables and
enables button b, (on lines 6 and 2, respectively), the app
is buggy because the user can click the button a second
time, causing a second call to remover.execute() on
the same instance of FeedRemover.
The “understanding” of the Android framework that
is necessary to reason about the code in Figure 2a
is precisely what we seek to capture with lifestate
specifications. In Figure 3, we sketch the lifestate rules
needed to reason about our running example. The first
rule, (t:AsyncTask).execute()9 ci t.execute() says
that when (t:AsyncTask).execute() is invoked for an
object t of type AsyncTask, the execute callin on t is
disallowed. We use the terms allowed and disallowed
to refer the state of callins; these terms are parallel to
enabled and disabled for the state of events. Overall, this
rule specifies that it is erroneous to call execute twice
(without something else that re-allows it).
The subsequent three rules specify disabling (written 9 evt) and enabling (written → evt) events (e.g.
setEnabled enables the onClick callback). To make
clear that the enabled-disabled state refers to events,
we give names to events, such as Click (written in
Callback-Driven Trace Slicing and Lifestate Mining. We cannot actually observe the abstraction of
the internal framework state of a running Android app.
Instead, we can only observe the sequence of method
invocations. We use the operational semantics of λlife
to derive an instrumented semantics that captures the
behavior that we can observe: messages classified into
event, callback, callin, or internal invocations. This instrumented semantics specifies the recording needed to
produce traces like the one shown in Figure 2b and is
implemented in a tool called DroidLife.
We are interested in rules about the interactions between events and callins on framework objects. To collect
traces across multiple executions, we apply the standard
type abstraction to concrete objects, resulting in an
abstract, signature message like AsyncTask.execute().
To obtain relevant lifestate rules and improve the
feasibility of mining, we slice recorded traces into subtraces that group together related method calls. Our
approach is roughly to slice a recorded trace into subtraces that collect method invocations whose arguments
share a particular concrete object, similar to Pradel
et al. [32]. However, we face the additional difficulty that
an event message is not directly a method invocation
and the relevant framework object of the event is
typically hidden in the internal state of the message.
Our insight for identifying relevant event messages
is that an event should eventually trigger a callback
where the relevant framework object is also argument
to the callback. Thus, our callback-driven trace slicing
approach determines the relevance of an event message
m based on the arguments to the callback(s) that m
eventually triggers. For example, in our trace slice
for the concrete object t1:AsyncTask, the slice includes
the PostExecute event because it eventually triggers
the callback t1.onPostExecute() and so the type4
e ∈ Expr ::= v | · · · | if v then e1 else e2
| bind v1 v2
| allow v | disallow v | invoke v1 v2
| enable v | disable v
| force m
| let x = e1 in e2
λ ::= fung x in e
g ::= app | fwk
v ∈ Val ::= x | me | · · · | λ | a | κ | h | ()
application. We also assume that there are some heap
operations of interest for manipulating a global store
ρ. Events occur non-deterministically and return to the
main event loop, so events must communicate through
the shared, global heap.
The subsequent two lines split the standard callby-value function application into multiple steps. The
bind v1 v2 expression creates a thunk κ by binding a
function value λ with an argument value v. A thunk may
be forced by direct invocation or indirect event dispatch.
Before a thunk may be forced, the allow v expression
allocates a handle h for a given thunk κ, which may
be viewed as a permission to force a particular thunk
instance or message. The permission to invoke may be
revoked by disallow v on a handle h. An invoke v1 v2 term
takes a handle-thunk pair h κ to construct a message m
where h must grant permission for forcing κ. A message
m is simply a handle-thunk pair (h, κ).
As an example, let λx be the execute method code
and let at be an address for an AsyncTask. The standard
function application (λx at ) would simply translate to
primitives
thunks
calls
events
forcing
binding
functions
packages
variables x ∈ Var
addresses a ∈ Addr
handles h ∈ Handle
thunks κ ∈ Thunk ::= λ[v]
messages m ∈ Msg ::= (h, κ)
stores
message stores
continuations
states
ρ ::= · | ρ[a 7→ v]
µ, ν ::= · | µ[(h, κ)]
k ::= • | let x = k in e | m | m k
σ ∈ State ::= he, ρ, µ, ν, ki | initial
Figure 4. The syntax and the semantic domains of
λlife , a core model of event-driven programs capturing
enabledness of events and allowedness of invocations.
abstracted sub-trace for object t1 from the recorded
trace in Figure 2b is as follows:
let k = bind
The reason to split function application into these steps
is that thunks and handles are now first-class values and
can be used in event dispatch.
The direct invocation expressions are mirrored with
expressions for event dispatch. An enable v expression
allocates a handle h for a given thunk κ and enables it
for the external event-processing system (i.e., gives the
event-processing system permission to force the message
(h, κ)), while the disable v expression disables the message named by a handle h. There is no expression directly
parallel to invoke v1 v2 . Instead, when an expression reduces to a value, control returns to the event-processing
loop to (non-deterministically) select another enabled
event. The force m expression is simply an intermediate
in evaluation that represents a message that is forceable
(i.e., has been permitted for forcing via allow or enable).
The remainder of the syntax is mostly standard, except that functions λ : fung x in e are tagged with a package g, which we discuss subsequently in subsection 3.3 regarding our instrumented semantics for recording traces.
The let expression is variable binding with the usual
semantics. Values v of this expression language are variables x, a variable used to bind the active event handle
me, whatever base values of interest · · · , closures λ, store
addresses a (which are the values of heap-allocated objects), thunks κ, handles h, and unit (). Messages m
do not need to be first class, as they are referred to
programmatically via their handles.
AsyncTask.<init>()
AsyncTask.execute()
PostExecute AsyncTask.onPostExecute()
Once we have sets of sliced traces, we apply and
evaluate specification mining techniques to learn lifestate
rules. As alluded to above, the primary challenge for
mining lifestate specifications is that lifestate rules center
around event enabling-disabling and callin allowingdisallowing, which are not observable in recorded traces.
We present our mining approach in section 6 and we
evaluate them in section 7.
3.
Modeling Event-Driven Programs
In this section, we formalize a small-step operational
semantics for event-driven programs using an abstract
machine model (λlife ) that unifies enabledness for event
lifecycles and allowedness for callin typestates. Crucially,
this semantics motivates lifestate specifications, as the
concrete states of this semantics serves as the concrete
domain from which lifestate specifications abstract (section 4). This semantics also serves as a basis for describing our dynamic analysis for recording traces of events,
callbacks, and callins (section 5).
3.1
λx at in let h = allow k in invoke h k
Thunks, Handles, and Messages
The syntax of λlife is shown at the top of Figure 4, which
is a λ-calculus in a let-normal form. The first line of
expressions e are standard, including variables and values
v, control flow, and whatever base values and operations
of interest (e.g., integers, tuples) but excluding function
3.2
Semantics: Enabling Is Not Enqueuing
In contrast to the Android implementation, the state of a
λlife program does not have a queue. Rather, the external
5
Bind
Disable
Disallow
hbind λ v, ρ, µ, ν, ki −→ hλ[v], ρ, µ, ν, ki hdisable h, ρ, µ[(h, κ)], ν, ki −→ h(), ρ, µ, ν, ki hdisallow h, ρ, µ, ν[(h, κ)], ki −→ h(), ρ, µ, ν, ki
Enable
h∈
/ dom(µ)
Allow
Invoke
h∈
/ dom(ν)
(h, κ) ∈ ν
henable κ, ρ, µ, ν, ki −→ hh, ρ, µ[(h, κ)], ν, ki hallow κ, ρ, µ, ν, ki −→ hh, ρ, µ, ν[(h, κ)], ki hinvoke h κ, ρ, µ, ν, ki −→ hforce (h, κ), ρ, µ, ν, ki
Force
(h, (fung0 x0 in e0 )[v 0 ]) = m
InvokeDisallowed
m = (h, κ)
m∈
/ν
Event
(h, κ) ∈ µ
hforce m, ρ, µ, ν, ki −→ he0 [v 0 /x0 ][h/me], ρ, µ, ν, m ki hinvoke h κ, ρ, µ, ν, ki −→ h(), ρ, µ, ν, •i hv, ρ, µ, ν, •i −→ hforce (h, κ), ρ, µ, ν, (h, κ)i
Figure 5. Semantics. Create thunks and allocate handles to disable, disallow, enable, and allow messages.
environment (e.g., user interactions) is modeled by the
non-deterministic selection of an enabled event. When
an event is processed, it begins execution in the current
state, so the effects of external events can be modeled
by effects on the global heap. This model eliminates the
complexity of a queue that is an implementation detail
for the purposes of lifestate specification.
We consider an abstract machine model with a store
and a continuation, and we enrich it with an enabled
events store µ, which is a finite map from handles to
thunks, and an allowed calls store ν, which is also a
map from handles to thunks. These stores can also
be seen as the set of enabled and allowed messages,
respectively. And thus a machine state σ : he, ρ, µ, ν, ki
consists of an expression e, a store ρ, enabled events
µ, allowed calls ν, and a continuation k (shown at
the bottom of Figure 4). The store ρ is a standard,
finite map from addresses a to values v. We explain the
special state initial while discussing the trace semantics
in subsection 3.3. A continuation k can be the top-level
continuation • or a continuation for returning to the
body of a let expression let x = k in e, which are standard.
Continuations are also used to record the active messages
via m and m k corresponding to the run-time stack of
activation records.
We define an operational semantics in terms of the
judgement form σ −→ σ 0 for a small-step transition relation. In Figure 5, we show the inference rules defining the
reduction steps related to creating, enabling-disabling,
allowing-disallowing, and finally forcing messages. The
Bind, Disable, Disallow, Enable, and Allow rules follow closely the informal semantics discussed previously
in subsection 3.1. Observe that Enable and Allow are
parallel in that they both allocate a fresh handle h, and
Disable and Disallow look up a message via its handle.
The only difference between Enable and Disable versus
Allow and Disallow is that the former pair manipulate
the enabled events µ, while the latter touches the allowed
calls ν. We write µ[(h, κ)] for the map that extends µ
with a mapping (h, κ) (i.e., is the same as µ except at
h) and similarly for other maps.
The Invoke and Event rules have similar parallels.
The Event rule says that when the expression is a value
v and the continuation is the top-level continuation •,
then a message is a non-deterministically chosen from the
enabled events µ to force. For instrumentation purposes
(see subsection 3.3), we overload continuations to record
the event message (h, κ) in the continuation. The Invoke
rule checks that the given handle-thunk pair is an allowed
message in ν before forcing.
The Force rule implements the “actual application”
that reduces to the function body e0 with the argument v 0
substituted for the formal x0 . So that the forced message
has ready access to its handle h, the h is substituted
for the me parameter. Observe that an enabled event
remains enabled after an Event reduction. A “dequeuing” event-processing semantics can be implemented by
an event disabling itself (via disable me) on execution.
To adequately model Android, it is important to be
able to model both events that are self-disabling (e.g.,
the event that invokes onCreate) and those that do
not self-disable (e.g., the event that invokes onClick).
Again for instrumentation purposes, we push the invocation message m on the continuation (via m k). The
InvokeDisallowed rule states that a disallowed message
invocation terminates the running program by dropping
the continuation. The Return and Finish rules simply
state that the recorded message m frames are popped
on return from a Force and Event, respectively. We
elide these rules and the remaining reduction rules, as
they are as expected.
3.3
Defining Callbacks and Observable Traces
We wish to record traces of “interesting” transitions from
which we mine lifestate specifications. Lifestate specifications abstract Enable, Disable, Allow, and Disallow
transitions, which are unfortunately unobservable or hidden. That is, they are transitions in λlife that are not
observable in an implementation like Android. Our challenge is to infer lifestate specifications from observable
transitions like, “An event message m was initiated.”
0
In Figure 6, we define the judgment form σ −→
τ σ ,
which instruments our small-step transition relation
6
transitions
observable kinds
initial state and a “dummy” init transition with an Init
rule. Any remaining rules defining the original transition
relation σ −→ σ 0 not discussed here are simply labeled
with the “don’t care” transition label ε.
We write JeK for the path semantics of λlife expressions
e that collects the finite (but unbounded) sequences of
alternating state-transition-state στ σ 0 triples according
0
to the instrumented transition relation σ −→
τ σ . From
paths, we derive traces $ in an expected manner by
keeping transitions but dropping intermediate states.
The design of the trace recording in DroidLife follows
0
this instrumented semantics σ −→
τ σ to obtain traces like
the one shown in Figure 2b. In particular, it maintains
a message stack corresponding to the continuation k to
emit callback cb and callin ci transitions.
τ ∈ T ⊆ Trans ::= oknd m | ε | init
oknd ::= evt | dis | cb | ci | ret
traces
$ ∈ Trans∗ ::= · | τ $
ForceCallback
(h, (funapp x0 in e0 )[v 0 ]) = m
fwk = pkg(k)
0 0
0
−→
hforce m, ρ, µ, ν, ki cb
m he [v /x ][h/me], ρ, µ, ν, m ki
ForceCallin
(h, (funfwk x0 in e0 )[v 0 ]) = m
app = pkg(k)
0 0
0
−→
hforce m, ρ, µ, ν, ki ci
m he [v /x ][h/me], ρ, µ, ν, m ki
def
msg(let x = k in e) = msg(k)
def
def
msg(m) = msg(m k) = m
def
pkg(k) = g if (h, (fung x in e)[v]) = msg(k)
Figure 6. Callbacks and callins are transitions between
framework and app code.
4.
Lifestate Specification: Abstracting
Enabledness and Allowedness
In this section, we define formally the meaning of lifestate
rules building on the λlife concrete model of event-driven
programs from section 3.
We consider the abstraction of messages to be a parameter of our approach. We call the message abstraction
ˆ and assume
of interest a message signature m
b ∈ Msg
it comes equipped with a standard concretization funcˆ → ℘(Msg) to give signatures meaning and
tion γ : Msg
ˆ
an element-wise abstraction function β : Msg → Msg
to lift a message to its signature satisfying the soundness and best-abstraction relationship. In the case of
an Android method call like (t: AsyncTask).execute()
(which corresponds to a message m : (ht , λx [at ]) for some
handle ht , function λx , and object address at ), the
message signature can be the Java method signature
AsyncTask.execute(). In DroidLife, we make a further
refinement on Java method signatures to split on values of primitive type, so Button.setEnabled(false) and
Button.setEnabled(true) are considered distinct message signatures. For the implementation, the elementwise abstraction function β is applied to recorded messages to produce message signatures for lifestate mining.
A lifestate specification R
b is a set of rules where
a rule b
r is an enable → evt, disable 9 evt, allow → ci,
or disallow 9 ci constraint. We define the meaning of
lifestate specifications by defining an abstraction of the
concrete transition relation σ −→ σ 0 for λlife .
To do so, we consider signature transitions b
τ and
a signature traces $.
b A signature transition b
τ are
simply analogous to concrete transitions τ , except over
signatures m
b instead of concrete messages m. As input
to specification mining, we are only concerned with the
event evt, callin ci, and disallowed dis transition kinds, so
we drop the callback cb and return ret kinds. A signature
trace $
b is simply a sequence of signature transitions. To
define the meaning of signature transitions and traces, we
from the previous subsection with transition labels τ .
Interesting observable transitions correspond primarily
to introduction and elimination operations on messages.
We instrument the Event rule to record that event
message m was initiated (via observable evt m). Similarly,
we instrument the InvokeDisallowed rule to record that
an invocation of a message was disallowed (via dis (h, κ)).
We use ε to instrument an uninteresting, “don’t care”
transition or an unobservable, “don’t know” transition.
The allowed invocation of a message with the Invoke
rule is an uninteresting transition, so it is simply labeled
with ε. These cases are simply adding a transition label
to the rules from Figure 5, so they are not shown here.
Recall from section 2 that we define a callback as an
invocation that transitions from framework to app code
and a callin as an invocation from app to framework
code. In λlife , this definition is captured crisply by the
context in which a message is forced. In particular,
we say that a message is a callback if the underlying
callee function is an app function (package app) and it
is called from a framework function (package fwk) as
shown in ForceCallback. The msg(·) function inspects
the continuation for the running, caller message. The
pkg(·) function gets the package of the running message.
Analogously, a message is a callin if the callee function
is in the fwk package, and the caller message is in the app
package (rule ForceCallin). There is a remaining rule
not shown here, force (ForceInternal), where there is
no switch in packages (i.e., g 0 = pkg(k) where g 0 is the
package of the callee message). It is not an interesting
transition, so it is labeled with ε.
These three cases for force replace the Force rule
from Figure 5. We also instrument the Return rule
to record returning via ret m to retain a call tree
structure in the sequence of transitions. Finally, to
simplify subsequent definitions, we introduce a “dummy”
7
νb according to the lifestate rules R.
b For example, the
Enabled rule checks if the current transition evt m
b can
happen by checking if m
b is in the enabled set µ
b. If so,
it updates the permitted state according to the rules R.
b
The helper function →evt (bı) gathers the set of messages
that is enabled by bı and similarly for disable, allow, and
disallow rules. For presentation cleanliness, we leave the
rules R
b implicit in the inference rules since it is constant.
Finally, the meaning of a lifestate specification JRK
b is
the set of signature traces that do not get stuck during
reduction (where −→∗ is the reflexive-transitive closure
R
b
of the single step relation). And this definition gives rise
to a natural description for when a trace is described by
or sound with respect to a specification:
R
b ∈ LifeSpec ::= · | R
b , br
lifestate specs
rules
br ::= bı 99K m
b
forcings bı ::= m
b | init
99K::= → evt | 9 evt | → ci | 9 ci
ˆ ::= oknd
b
τ ∈ Trans
[m
b | init
signature transitions
oknd
[ ::= evt | ci | dis
signature traces
∗
ˆ
$
b ∈Π
b ⊆ Trans
::= · | b
τ$
b
µ
b, b
ν ::= · | µ
b, m
b
ˆ ::= h$,
σ
b ∈ State
b µ
b, b
νi
signature stores
signature states
Enabled
m
b ∈b
µ
µb0 = b
µ ∪→evt (m
b)
νb0 = b
ν ∪→ci (m
b)
µb0 = b
µ ∪→evt (m
b)
νb0 = b
ν ∪→ci (m
b)
0
h(evt m
b )$
b,b
µ, b
ν i −→ h$
b , µb −9evt (m
b ), νb0 −9ci (m
b )i
Allowed
m
b ∈b
ν
0
b
Definition 1 (Trace Soundness). A signature trace $
is sound with respect to a lifestate specification R
b iff $
b
is described by R
b (i.e., $
b ∈ JRK).
b Or in other words, the
signature trace $
b can be reduced without getting stuck
(i.e., h$,
b ·, ·i −→∗ h·, µ
b, νbi for some µ
b, νb).
R
b
A specification is then sound with respect to a
λlife program if its signature traces abstract the set
of concrete traces of the program:
0
h(ci m
b )$
b,b
µ, b
ν i −→ h$
b , µb −9evt (m
b ), νb −9ci (m
b )i
Disallowed
Initialized
h(dis m
b )$
b,b
µ, b
ν i −→ h$
b,b
µ, b
νi
b
µ = →evt (init)
b
ν = →ci (init)
h(init)$
b , ·, ·i −→ h$
b,b
µ, b
νi
m
b ∈/ b
ν
def
b ∈R
b
b
m
b bı→ evt m
def
ı) = m
b bı→ ci m
b ∈R
b
→ci (b
→evt (ı) =
J·K
JR
bK
:
def
=
def
b ∈R
b
b
m
b bı9 evt m
def
ı) = m
b bı9 ci m
b ∈R
b
9ci (b
9evt (ı) =
Definition 2 (Specification Soundness). A lifestate
specification R
b is sound with respect to an expression e
iff for every trace $ of e, it is in the concretization of a
signature trace $
b of R
b (i.e., formally, JeK ⊆ γ(JRK)).
b
∗
LifeSpec → ℘(Trans )
n
o
$
b h$
b , ·, ·i −→∗ h·, b
µ, b
ν i for some b
µ, b
ν
R
b
ˆ
5.
Figure 7. Meaning of lifestate specifications. A lifestate
rule b
r is an abstraction of enabling enable v, disabling
disable v, allowing allow v, or disallowing disallow v
transitions.
Trace Slicing for Multi-Object
Event-Driven Protocols
As the specification mining problem is severely underconstrained, we wish to constrain the search space of
specifications as much as possible before applying the
learning algorithms. As a first step, we wish to slice a
trace into sub-traces that group together related messages or method invocations that (likely) correspond to
different protocols. A well-known issue is that related
method invocations may involve multiple objects. This issue is exacerbated in event-driven protocols where event
messages correspond to some internal data structures
of the framework rather than method calls per se. For
instance, in our running example, a call to the callin
(b:Button).setEnabled(. . .) that enables or disables a
button b should be associated with a Click event for b.
One reasonable and common assumption is that two
related messages should use some common values that
are readily accessible, such as arguments for method
invocations. We call this heuristic, the argument-sharing
strategy, and leave as a parameter of the strategy
a function args : Msg → ℘fin (Val) that specifies the
arguments of a message. This heuristic is similar to
the one employed by Pradel and Gross [31] for mining
multi-object API protocols. Following the argumentsharing strategy, we can slice observations so that they
lift the concretization on message signatures to signature
transitions b
τ and traces $
b in the expected way.
We can then define a signature store µ
b, νb as a set of
message signatures, which is an abstract analogue of
a message store. In a signature state σ
b : h$,
b µ
b, νbi, the
signature store µ
b is a set of enabled message signatures
(i.e., abstract messages) and νb is a set of allowed
message signatures. For convenience, we refer to the
union µ
b ∪ νb of enabled and allowed message signatures
in a state as the set of permitted signatures and all other
message signatures that are disabled or disallowed as
the prohibited signatures.
A signature state σ
b : h$,
b µ
b, νbi also includes a signature
trace $
b that controls the execution of a corresponding
abstract machine defined by the transition relation
σ
b −→R
b0 that is parametrized by a set of rules R
b and
bσ
defined in Figure 7. This transition relation is relatively
straightforward: it considers the signature transition b
τ
at the beginning of the $
b and determines whether b
τ is
permitted according to the set of permitted signatures.
If it is permitted, it updates the signature stores µ
b and
8
contain only observable transitions that share a common
argument (this strategy is essentially described as trace
slicing [14]).
However, one issue in trace slicing for event-driven
object protocols is that the messages of interest are the
event messages (in evt m) and the callin messages (in
ci m0 ). As noted above, the “argument” of interest for
the Click event message is the button b buried inside
internal implementation-specific data structures.
From the (b:Button).setEnabled(. . .) callin and
Click event example, we also see that regardless of how
b may be stored in the Click event message, the Click
event eventually invokes the callback l.onClick(b) on
some listener l where button b becomes an argument
(retrieving b in some implementation-specific manner).
Thus we extend the argument-sharing strategy to event
messages by defining the arguments of an event message
m as the arguments of any callbacks that it invokes
(instead of m’s own arguments). Once the traces have
been sliced for each object, we apply the message abstraction function β and partition the traces based on
the abstraction of the sliced object.
While our particular argument-sharing strategy focuses on associating callbacks with their initiating events,
this strategy is quite flexible in that other appropriate
context information (e.g., values reached through the
heap) could be incorporated into the trace slice.
6.
accuracy in modeling complex, often partially observable
processes. The challenge of learning a probabilistic model
of the interaction between the Android framework and
an app over the alphabet of signature messages is that
in general neither model has a direct interpretation.
For a state xi of an a HMM or PFSA A, we define the
set of permitted message signatures to be the labels of
all enabled outgoing transitions and the set of prohibited
message signatures as its complement.
For a state xi and an enabled outgoing transition
labeled m,
b we learn the lifestate enable (or resp. allow)
rules that m
b permits all message signatures prohibited
in xi but permitted in the target set of this transition.
Similarly, for the state xi and an enabled outgoing
transition labeled m,
b we learn disable (or disallow)
rules that m
b prohibits all messages permitted in xi
that become prohibited along the transition.
The final lifestate specification R
b is the union of
permit and prohibit rules of all states of the automaton.
To identify high probability rules, we associate with
each rule m
b i 99K m
bj ∈ R
b a weight that is the sum
the probabilities of all transition that define the rule
normalized by the number of permitted sets m
b i is in.
Symbolic Sampling via ]SAT. In contrast to the
two-phase probabilistic automata approaches, we propose an algorithm that directly learns a lifestate specification from the set of signature traces Π
b that is potentially
easier to interpret.
The learning algorithm proceeds by computing a
weight for each possible rule in RuleSet using the
traces in Π
b and then obtains a lifestate specification by
selecting the rules that meet a weight threshold. The
weight assigned to a rule is an average, among all the
traces in Π,
b of the fraction of the total executions that
are sound for that rule.
Intuitively, the ideal settings to compute the weight
is to observe the sequence of signature states and
transitions (a signature path). The main difficulty we
face is that the internal state of the permitted messages
(i.e., the signature state h$,
b µ
b, νbi) is not observable. To
address the problem, we propose an abstraction of the
signature paths, which we use in our learning algorithm.
We assume no prior knowledge on the internal behavior of the framework and define the most conservative
abstraction via the non-deterministic transition relation
σ
b −→n σ
b0 that, upon invoking a permitted enable or allow
signature transition b
τ , can non-deterministically change
the internal state of the system (i.e., permits or prohibits
any other message) but keeps the semantics of disallow
and disables as in Figure 7. This choice of abstraction
may produce a large number of spurious behaviors that
could affect the quality of rules we learn, but we make
the following observations: (i) the imprecision introduced
by the spurious transition may be reduced by observing
Mining Lifestate Specifications
In the previous sections, we formalized a model for eventdriven systems like Android centered around enabled
events and allowed callins. From this model, we derive
the notion of lifestate specifications. In this section, we
discuss an application of specification mining techniques
to learn lifestate specifications. For a set of signature
traces Π,
b let Σ be the set containing all the abstract
messages contained in the traces Π
b and init. We learn
specifications where the messages in the rules are elements of Σ, hence restricting our learning algorithms to
an abstraction of messages that we observed concretely.
We denote with RuleSet the set of all the possible rules
that can be learned from Σ.
The goal of the mining algorithms is to find a specification R
b ∈ ℘(RuleSet) and to this end we consider 3
different approaches. Two of these are applications of an
off-the-shelf machine learning tool Treba [23] learning
probabilistic models of the interaction between the app
and the Android framework, and a symbolic sampling
one based on model counting of propositional Boolean
formulas [20], implemented with pySMT [19] and the
solver sharpSAT [35].
Probabilistic Models. Hidden Markov Models (HMM) and Probabilistic Finite State Automata (PFSA) are
common statistical models famed for their simplicity and
9
a large number of traces; (ii) with a further knowledge
of the framework, we can reduce the non-determinism
(i.e. reducing the abstract paths), hence improving the
quality of the learned specification.
The main challenge with this approach is that enumerating all abstract paths for a signature trace is intractable. To overcome this problem, we represent the
set of possible abstract paths symbolically encoded as a
propositional logic formula, to leverage efficient model
counting solvers.
Research Questions. Given a candidate specification (i.e., a set of lifestate rules) obtained from a learning
method, we consider the following research questions.
Learn Specifications by Counting Paths. Given a signature trace $
b:b
τ1 ...b
τ n , let Paths($)
b denotes the set
of all the sequences of states and transitions (paths) under the non-deterministic semantics. For each rule kind
bı→ evt m
b (resp. 9 evt, → ci, 9 ci), we say that the rule is
“matched” in a state of the path if, in that state, the message m
b is enabled (resp. disabled, allowed, disallowed)
and the abstract transition executes the message bı.
We want to count all the paths of a trace such that
the ratio of the number of states of the path that match
the rule over the total number of times we seen bı in the
trace is greater than a weight w. For a signature trace
$,
b we call RulePaths($,b
b r) such set of paths.
Given a signature trace $
b and a rule b
r, the frequency
of b
r in $
b with weight w, is given by:
• RQ2 (Veracity): For each rule in a mined specifi-
def
Freq(w,b
r, $)
b =
• RQ1 (Sufficiency): Is a mined specification sound with
respect to previously unseen traces? We say that a
specification is sufficient for explaining a set of traces
if each trace is sound with respect to the specification
(according to Definition 1 in section 4). Foremost, we
seek specifications that are, in the limit, sufficient to
explain all actual Android behavior.
cation, does the rule correspond to actual Android
behavior? We say that a specification is veracious
if it captures when a message should be enabled or
allowed during the execution of an app and when it
should not. In the end, we want to find rules that in
fact describe the true enabling, disabling, allowing,
and disallowing behavior of Android.
Experimental Methodology. We consider the mining algorithms from section 6 based on probabilistic models (HMM and PFSA) and symbolic sampling (]SAT).
For the symbolic sampling approach, we use two weight
thresholds 0.6 and 0.8, corresponding intuitively to keeping only the rules that explain slightly more than half
the paths and those that explain most paths, respectively. This selection of parameters gives us four learning
methods to evaluate that we abbreviate as ]SAT-0.6,
]SAT-0.8, HMM, and PFSA.
Traces recorded by DroidLife are sliced for each object,
abstracted into message signatures, and grouped by the
framework type of the slicing object. Thus, we have a
corpus of sliced and abstracted traces for each framework
type of interest. To evaluate a learning method, we divide
a trace corpus into a training set and a testing set using
5-fold cross validation.
|RulePaths($,b
b r)|
.
|Paths($)|
b
The notion can naturally be lifted to a set of signature
traces Π
b by taking the geometric mean of the fractions
of the observed traces.
Let δ be a real constant. The final specification is
def
R
b = {b
r | Freq(w,b
r, Π)
b ≥ δ}. The threshold δ is needed
to select only the rules that, according to the frequency
value, more likely capture the framework behavior.
The number of paths in the sets Paths($)
b and
RulePaths($,b
b r) is exponential in the total number
of messages. Thus, it is not feasible to explicitly enumerate them. Our solution is to encode all the paths in
Paths($)
b as a propositional logic formula. The intuition
is that each (complete) model of this formula represents
a path in Paths($).
b Hence, we cast the problem of
counting the cardinality of Paths($)
b to the problem of
counting the number of models of the Boolean formula,
for which several efficient tools exist (e.g. [35]). The encoding is an adaptation of Bounded Model Checking [10],
a technique that encodes all the paths of length k of a
transition system.
7.
Corpus: Generating Sliced and Abstracted Traces
We begin with 133 traces of running Android applications recorded by DroidLife. These traces were generated
from a corpus of 1909 apps retrieved from Github.
We used two different methods for generating traces:
manual exercising of the app and automatic by means
of the UI Exerciser Monkey. We chose to generate some
traces using manual exercising because apps require user
logins, API keys, and other inputs that present problems
for automatic exploration tools (e.g., [5, 8, 27]). The
manual method resulted in 65 traces and the automatic
in 68. The process for manual trace generation was to
exercise the app in a normal manner for 10 minutes under
instrumentation. For the automatic Monkey, a script
loaded the application on an Android emulator, waited
3 minutes for any application initialization, and then
pulsed the UI exercising 40 times. Each time inputting
50 random events and delaying 30 seconds in between.
Empirical Evaluation: Mining
Here, we evaluate empirically whether our process finds
meaningful lifestate specifications, as well as compare
the quality of the specifications obtained from different
mining algorithms.
10
Table 1. The data set is partitioned by framework type
(f-type) on the sliced object. For each trace set, we
give the number of traces, the average length of the
traces (len), the total number of distinct event message
signatures (evts), callin signatures (cis), and the size of
the specification space (|LifeSpec|).
Table 2. Mined candidate specifications. For each trace
set for a framework type, we apply each method to learn
on a training set (using 5-fold cross validation). This
table shows the number of lifestate rules of each kind
on the first training set.
lifestate rules
traces
len
evts
cis
|LifeSpec|
(num)
(mean)
(num)
(num)
(num)
Fragment
FragmentV4
Button
AsyncTask
52
124
654
63
9.7
9.0
4.6
3.8
6
10
4
1
19
34
28
4
1275
3916
2080
55
summary
893
6.8
21
85
7326
f-type
→ evt
From the corpus of 133 recorded traces, we apply
callback-driven trace slicing from section 5 to get 6134
sliced traces over 184 framework types. Each framework
type corresponds to a partition of the 6134 sliced traces.
We selected four of these framework types, which had
trace sets of sufficient length and diversity of events
and callins, for evaluating our learning methods. In
Table 1, we list some statistics about the data set.
The specification space (|LifeSpec|) is the number of
possible lifestate rules as determined by the number of
observed events (evts) and callins (cis). The summary
line is the total number of traces, the mean of mean
length of traces, the total number of event and callin
signatures, the total specification space.
f-type
method
Fragment
Fragment
Fragment
Fragment
FragmentV4
FragmentV4
FragmentV4
FragmentV4
Button
Button
Button
Button
AsyncTask
AsyncTask
AsyncTask
AsyncTask
]SAT-0.6
]SAT-0.8
HMM
PFSA
]SAT-0.6
]SAT-0.8
HMM
PFSA
]SAT-0.6
]SAT-0.8
HMM
PFSA
]SAT-0.6
]SAT-0.8
HMM
PFSA
total
(num)
9 evt
(num)
→ ci
(num)
9 ci
(num)
2
0
7
12
2
0
39
27
7
2
29
20
1
0
2
1
0
0
12
14
0
0
24
45
1
1
22
25
0
0
4
2
17
3
48
39
24
2
235
145
30
7
87
87
2
1
6
5
5
0
32
41
2
0
209
213
7
4
89
90
0
0
8
5
151
150
738
705
Sound Traces (frac/on)
Sufficiency
Candidate Specifications. In Table 2, we show the
number of each rule kind learned using each method
on the first training set. Our first observation is that
the probabilistic model-based methods generate many
more candidate rules than the symbolic sampling-based
method. This observation is not unexpected, as the symbolic sampling-based method takes a more conservative
approach for generating rules: it generates rules based
on the frequency of explainable signature traces. Our
second observation is that event-disables are the least
frequently derived rules. This observation is also not
unexpected, as event-disables are the least constrained
rule kind. The total line is the total number of rules
of each kind. The magnitude is not meaningful, but it
shows the relative frequency of the rule kinds.
1
0.8
0.6
0.4
0.2
0
Fragment
Fragment v4
Bu;on
AsyncTask
Average
SAT-0.6
SAT-0.8
PFSA
HMM
Learning Method
Figure 8. Sufficiency measures the fraction of sound
traces of a given testing set for a given specification
(higher is better). For each training set for a framework
type, we measure sufficiency for each learned specification on the corresponding testing set. The y-axis shows
the average sufficiency measure over the testing sets
(from 5-fold cross validation) for each framework type
grouped by learning method. The rightmost bar in each
group shows the average over the four framework types.
RQ1: Measuring Sufficiency. Given a testing set
and a candidate specification, we measure sufficiency as
the ratio of traces that are sound with respect to the
specification to the number of traces in the testing set.
Figure 8 shows the sufficiency measure for each
learned specification, grouped by method. No one
method dominates the others by this measure. For
example, the PFSA method is the only one that gen-
erates rules that have soundness fraction over 0.5 for
Fragment but is worse than all other methods for AsyncTask.
The sufficiency measure for ]SAT-0.8 is generally high,
but recall from Table 2 that ]SAT-0.8 learns far fewer
rules than the other methods. It is quite conservative in
the rules it learns, but for the rules that it does learn,
11
each column. Since no single method dominated most
of the time, this row shows that the combination of
methods can learn a more complete specification than
any single method alone. In total we learned 75 unique
rules which model the behavior of the framework. Of
these, 55 directly correspond to the cause and effect that
a developer would attribute to the system.
The highest level point is that a number of actual
rules are learned for each framework type by both learning methods (the actual column) out of the thousands of
possible rules shown in Table 1 (7326) by only examining
the top 20 rated rules from each learning method. And
furthermore, most actual rules that are learned correspond to the real root cause in the Android framework
(the direct column). This observation provides supporting evidence for lifestate rules capturing event-driven
object protocols at an appropriate level of abstraction.
The only exception to the high fraction of actual rules
is the AsyncTask which can be fully specified by 7 rules as
we describe further below. This biases the fraction to be
low since we took the top 20 rated rules, and there were
only 7 rules to learn. And every specification needed for
AsyncTask was learned. Furthermore, most of the direct
rules are found even in the top 10.
One remaining question is how many rules are required to specify the protocol. Is it the case that the
number of rules for each framework type should be small
compared to the number of possible rules. To address
this question, we manually specified the full set of rules
for AsyncTask for the observed events and callins (since
this class is comparatively small with only 55 possible
rules compared to the thousands possible for the other
framework types, see Table 1).
Manual specification of AsyncTask resulted in 7 direct
rules. The result is that we are learning 3 out of the
7 direct rules with the ]SAT-0.6 method, 4 out of 7
with the PFSA method, and 6 out of 7 with the HMM
method. There were no rules in the manually created
set which were not learned by one of our algorithms.
Table 3. Veracity is a measure of precision in finding
actual rules by manually triaging the rules found by each
learning method. The actual column shows the number
of learned rules that correspond understood Android
behavior. The direct shows the number of actual rules
that also correspond to the root enable, disable, allow,
or disallow in Android.
rules
actual
direct
num
num
num (frac)
Fragment
Fragment
Fragment
Fragment
method
]SAT-0.6
PFSA
HMM
total unique
20
20
20
42
10
14
5
22
7
11
4
18
(0.7)
(0.8)
(0.8)
(0.8)
FragmentV4
FragmentV4
FragmentV4
FragmentV4
]SAT-0.6
PFSA
HMM
total unique
20
20
19
58
8
12
11
24
7
10
9
19
(0.9)
(0.8)
(0.8)
(0.8)
Button
Button
Button
Button
]SAT-0.6
PFSA
HMM
total unique
20
20
20
49
6
7
3
14
3
7
3
11
(0.5)
(1 )
(1 )
(0.8)
AsyncTask
AsyncTask
AsyncTask
AsyncTask
]SAT-0.6
PFSA
HMM
total unique
3
13
19
22
3
10
12
15
3
4
6
7
(1 )
(0.4)
(0.5)
(0.5)
all objects
total unique
171
75
f-type
55 (0.7)
they tend to generalize well. The ]SAT-0.6 and PFSA
methods appear to balance producing more candidate
rules and ones that generalize reasonably well. We thus
consider manual triage of the rules produced by these
two methods for evaluating veracity.
RQ2: Evaluating Veracity. To evaluate veracity,
we take a candidate lifestate specification and manually
triage the rules and categorize them with respect to
the correctness in capturing the true behavior of the
Android framework. We consider the following categories
of correctness: (actual) the rule abstracts understood
Android behavior from reading Android documentation
or the framework code; (direct) the rule is actual and
corresponds to the root cause in Android (e.g., a learned
enable rule corresponds to direct enabling in Android);
(false or unknown) the rule appears to contradict
understood Android rules. If we cannot be sure that
a learned rule corresponds to a actual rule, then we
conservatively classify it as false.
In Table 3, we show veracity results for the ]SAT-0.6,
PFSA, and HMM methods for each framework type. We
manually triaged the twenty rules with highest weight
(using the parsimony assumption that the number of
rules for each framework type should be small). We
then add a row for the total number of unique rules in
Threats to Validity. The rules that can be learned
are limited by the observations available in the traces.
For example, there was no opportunity to learn anything
about the AsyncTask.cancel() callin since our traces did
not include any observations of such an invocation. In
the end, we are limited by the extent by which framework
behavior can be explored through dynamic execution.
To try to mitigate this affect, we made traces with
as many applications as possible minimizing duplicate
traces. This exposes us to more possible ways developers
may use a given framework object. There is a potential
for bias in selecting the framework types on which to
mine specifications. To minimize this effect, we selected
by looking just at the number and length of traces and
diversity of events and callins (from Table 1).
12
Mined Specifications. From the veracity summary,
the ]SAT-0.6, PFSA, and HMM methods appear comparable, so we look more closely into the rules that
are learned. From our running example in section 2,
the most important rule that we would expect is that
(t:AsyncTask).execute() callin can only be called once.
It is indeed learned by all three methods. Another important rule is that the PostExecute event is enabled
by the (t:x AsyncTask).execute() callin. In this case,
the ]SAT-0.6 and HMM methods learned this rule but
the PFSA method did not.
The biggest source of false rules learned that we observed are rules that appear to come from common patterns in how developers use the framework. For example,
one rule learned by PFSA is that (b:Button).getWidth()
disallows itself, which perhaps corresponds to few traces
with repeated calls to getWidth.
A source of actual but not direct rules relates to
callbacks and super calls. We found that a common
pattern in Android is to require a call to the super
method. For example to implement onActivityCreated,
super.onActivityCreated would need to be called.
Thus in functioning applications, this super call is seen as
a callin that always happens with the event the triggers
the onActivityCreated callback. Thus, the left-hand–
side of a actual but not necessarily direct rule is where
the left-hand–side is either the event or the super callin.
One could reasonably consider removing required super
calls from our definition of callins.
In the end, we found a rule that captures an
undocumented behavior of Android in the official
Android documentation. The rule states that the
(f:Fragment).getResources() is allowed when the event
ActivityCreated that triggers the onActivityCreated
callback. The getResources callin will throw an exception if called before this ActivityCreated event.
8.
for lifecycle events in addition to their interaction with
typestate APIs for frameworks.
Other work also simplifies automata specifications
by considering simpler rule-based approaches (e.g., [18,
38]). Subsequently, Lo et al. [25] extend this kind
of specification to QBEC (quantified binary temporal
rules with equality constraints) specifications. However,
these rule-based approaches differ from lifestate rules
in that they express necessary causation (e.g for every
lock there must be a subsequent unlock). In contrast,
lifestate rules do not express such conditions, and these
conditions are actually undesirable, as most events in
event-driven systems enable other events but do not
cause them to occur.
Concerning static mining techniques, Alur et al. [4]
propose a technique to infer a typestate interface for
a given Java class. This technique could conceivably
be used to statically infer typestate specifications for
event-driven frameworks such as Android. However, their
technique does not account for the unique relationship
between typestate and lifecycle events in event-based
systems. Additionally, Shoham et al. [34] develop a technique for inferring the typestate of a certain API by analyzing presumably correct client programs. The order of
method invocations, determined statically, can produce
a multi-object typestate specification. Ramanathan et al.
[33] propose a similar technique, but one that can also
derive invariant-based preconditions. However, to our
knowledge these techniques also cannot be successfully
extended to event-based systems, where typestate and
lifecycle properties may interact.
Analysis of Event-Driven Systems. Recently,
more and more program analyses have focused on eventdriven systems. Both dynamic analyses, such as for race
detection [22, 29], and static analyses for information
flow [7] and safety properties [11] expect some specification of framework behavior with respect to callbacks.
There has also been a number of recent static analyses
targeting information flow concerns (e.g., [16, 21, 37])
where a significant concern is synthesizing models for
taint flow through the framework [9].
There are also some tools concerned with exposing
the implicit control flow from the framework to callbacks
by linking callback registration methods with their
callbacks [13, 28], modeling GUI components [41], and
explicating reflection [12]. This information is useful for
creating an over-approximation of the possible callbacks
but does not capture precise events that may enable or
disable a given event that then invokes the callback.
Related Work
Specification Mining. Our technique is broadly similar to mining typestate specifications for API interfaces
from traces (e.g., [3, 6, 15, 17, 31, 36, 39, 40]). Our contribution can be seen as an extension that, in addition
to inferring typestate properties, also infers the ordering constraints of lifecycle events that interleave and
interact with typestate properties. Ammons et al. [6]
presents one of the earliest attempts to infer temporal
specifications of programs by mining observed traces.
Notably, Yang et al. [40] extend the technique to scale to
larger data-sets with imperfect traces, and Pradel and
Gross [31] improve the technique to mining specifications
for collaborations involving multiple, interacting objects.
There is also work on combining automata mining with
value-based invariant mining [24, 26]. However, we are
unaware of any previous work that mines specifications
9.
Conclusion
We have presented lifestate rules, a language for specifying event-driven object protocols. The key idea underlying lifestate rules is a model of event-driven programs
13
in terms of enabled events and allowed callins unified
as suspended messages. As a result, lifestate rules are
able to capture mixed lifecycle and typestate constraints
at a higher level abstraction than automata. We then
instantiated this model in a dynamic analysis tool called
DroidLife that prepares Android traces for specification
mining using a notion of trace slicing adapted to events
and callbacks. Finally, we applied specification mining
techniques to infer lifestate specifications based on unsupervised automata-based learning techniques that have
been previously applied to typestate mining, as well as a
direct lifestate mining technique based on propositional
model counting. In the end, we were able to learn several actual lifestate rules that accurately capture the
behavior of Android in a compact way.
[11] S. Blackshear, B. E. Chang, and M. Sridharan. Selective
control-flow abstraction via jumping. In Object-Oriented
Programming Systems, Languages, and Applications
(OOPSLA), 2015.
[12] S. Blackshear, A. Gendreau, and B. E. Chang. Droidel:
A general approach to Android framework modeling. In
State of the Art in Program Analysis (SOAP), 2015.
[13] Y. Cao, Y. Fratantonio, A. Bianchi, M. Egele, C. Kruegel,
G. Vigna, and Y. Chen. Edgeminer: Automatically
detecting implicit control flow transitions through the
android framework. In Network and Distributed System
Security (NDSS), 2015.
[14] F. Chen and G. Rosu. Parametric trace slicing and monitoring. In Tools and Algorithms for the Construction
and Analysis of Systems (TACAS), 2009.
[15] V. Dallmeier, C. Lindig, A. Wasylkowski, and A. Zeller.
Mining object behavior with ADABU. In Dynamic
Systems Analysis (WODA), 2006.
References
[1] AntennaPod - FeedRemover bug report.
https:
//github.com/AntennaPod/AntennaPod/issues/1304.
Accessed: 2016-03-13.
[16] Y. Feng, S. Anand, I. Dillig, and A. Aiken. Apposcopy:
Semantics-based detection of Android malware through
static analysis. In Foundations of Software Engineering
(FSE), 2014.
[2] Facebook
SDK
for
Android
AsyncTask
bug report.
https://github.com/facebook/
facebook-android-sdk/pull/315. Accessed: 2016-0313.
[17] M. Gabel and Z. Su. Javert: Fully automatic mining
of general temporal properties from dynamic traces. In
Foundations of Software Engineering (FSE), 2008.
[3] M. Acharya, T. Xie, and J. Xu. Mining interface specifications for generating checkable robustness properties.
In Software Reliability Engineering (ISSRE), 2006.
[18] M. Gabel and Z. Su. Online inference and enforcement
of temporal properties. In International Conference on
Software Engineering (ICSE), 2010.
[4] R. Alur, P. Černý, P. Madhusudan, and W. Nam.
Synthesis of interface specifications for Java classes. In
Principles of Programming Languages (POPL), 2005.
[19] M. Gario and A. Micheli. pySMT: a solver-agnostic
library for fast prototyping of smt-based algorithms.
In Workshop on Satisfiability Modulo Theories (SMT),
2015.
[5] D. Amalfitano, A. R. Fasolino, P. Tramontana, S. D.
Carmine, and A. M. Memon. Using GUI ripping for
automated testing of Android applications. In Automated
Software Engineering (ASE), 2012.
[20] C. P. Gomes, A. Sabharwal, and B. Selman. Model
counting. In Handbook of Satisfiability. 2009.
[21] M. I. Gordon, D. Kim, J. Perkins, L. Gilham, N. Nguyen,
and M. Rinard. Information-flow analysis of Android
applications in DroidSafe. In Network and Distributed
System Security (NDSS), 2015.
[6] G. Ammons, R. Bodı́k, and J. R. Larus. Mining
specifications. In Principles of Programming Languages
(POPL), 2002.
[22] C.-H. Hsiao, J. Yu, S. Narayanasamy, Z. Kong, C. L.
Pereira, G. A. Pokam, P. M. Chen, and J. Flinn. Race
detection for event-driven mobile applications. In Programming Language Design and Implementation (PLDI),
2014.
[7] S. Arzt, S. Rasthofer, C. Fritz, E. Bodden, A. Bartel,
J. Klein, Y. L. Traon, D. Octeau, and P. McDaniel.
FlowDroid: Precise context, flow, field, object-sensitive
and lifecycle-aware taint analysis for Android apps.
In Programming Language Design and Implementation
(PLDI), 2014.
[23] M. Hulden. Treba: Efficient numerically stable EM for
PFA. In Grammatical Inference (ICGI), 2012.
[8] T. Azim and I. Neamtiu. Targeted and depth-first
exploration for systematic testing of Android apps. In
Object-Oriented Programming Systems, Languages, and
Applications (OOPSLA), 2013.
[24] D. Lo and S. Maoz. Scenario-based and value-based
specification mining: Better together. In Automated
Software Engineering (ASE), 2010.
[9] O. Bastani, S. Anand, and A. Aiken. Specification
inference using context-free language reachability. In
Principles of Programming Languages (POPL), 2015.
[25] D. Lo, G. Ramalingam, V.-P. Ranganath, and
K. Vaswani. Mining quantified temporal rules: Formalism, algorithms, and evaluation. Sci. Comput. Program.,
77(6), 2012.
[10] A. Biere, A. Cimatti, E. M. Clarke, and Y. Zhu. Symbolic
model checking without bdds. In Tools and Algorithms
for the Construction and Analysis of Systems (TACAS),
1999.
[26] D. Lorenzoli, L. Mariani, and M. PezzÃĺ. Automatic generation of software behavioral models. In International
Conference on Software Engineering (ICSE), 2008.
14
[27] A. Machiry, R. Tahiliani, and M. Naik. Dynodroid: an
input generation system for Android apps. In European
Software Engineering Conference and Foundations of
Software Engineering (ESEC/FSE), 2013.
[28] M. Madsen, F. Tip, and O. Lhoták. Static analysis of
event-driven node.js javascript applications. In ObjectOriented Programming Systems, Languages, and Applications (OOPSLA), 2015.
[29] P. Maiya, A. Kanade, and R. Majumdar. Race detection
for Android applications. In Programming Language
Design and Implementation (PLDI), 2014.
[30] S. Pomeroy.
Complete Android Fragment and
Activity lifecycle.
https://github.com/xxv/
android-lifecycle. Accessed: 2016-03-13.
[31] M. Pradel and T. R. Gross. Automatic generation of
object usage specifications from large method traces. In
Automated Software Engineering (ASE), 2009.
[32] M. Pradel, C. Jaspan, J. Aldrich, and T. R. Gross.
Statically checking API protocol conformance with
mined multi-object specifications. In International
Conference on Software Engineering (ICSE), 2012.
[33] M. K. Ramanathan, A. Grama, and S. Jagannathan.
Static specification inference using predicate mining.
In Programming Language Design and Implementation
(PLDI), 2007.
[34] S. Shoham, E. Yahav, S. Fink, and M. Pistoia. Static
specification mining using automata-based abstractions.
In Software Testing and Analysis (ISSTA), 2007.
[35] M. Thurley. sharpSAT - counting models with advanced
component caching and implicit BCP. In Theory and
Applications of Satisfiability Testing (SAT), 2006.
[36] N. Walkinshaw and K. Bogdanov. Inferring finitestate models with temporal constraints. In Automated
Software Engineering (ASE), 2008.
[37] F. Wei, S. Roy, X. Ou, and Robby. Amandroid: A
precise and general inter-component data flow analysis
framework for security vetting of Android apps. In
Computer and Communications Security (CCS), 2014.
[38] W. Weimer and G. C. Necula. Mining temporal specifications for error detection. In Tools and Algorithms
for the Construction and Analysis of Systems (TACAS),
2005.
[39] J. Whaley, M. C. Martin, and M. S. Lam. Automatic
extraction of object-oriented component interfaces. In
Software Testing and Analysis (ISSTA), 2002.
[40] J. Yang, D. Evans, D. Bhardwaj, T. Bhat, and M. Das.
Perracotta: Mining temporal API rules from imperfect
traces. In International Conference on Software Engineering (ICSE), 2006.
[41] S. Yang, D. Yan, H. Wu, Y. Wang, and A. Rountev.
Static control-flow analysis of user-driven callbacks in
Android applications. In International Conference on
Software Engineering (ICSE), 2015.
15
| 6cs.PL
|
On Structural Parameterizations of Firefighting
Bireswar Das, Murali Krishna Enduri? , Neeldhara Misra?? and I. Vinod Reddy
arXiv:1711.10227v1 [cs.DS] 28 Nov 2017
IIT Gandhinagar, India
{bireswar,endurimuralikrishna,neeldhara.m,reddy vinod}@iitgn.ac.in
Abstract. The Firefighting problem is defined as follows. At time
t = 0, a fire breaks out at a vertex of a graph. At each time step t > 0, a
firefighter permanently defends (protects) an unburned vertex, and the
fire then spread to all undefended neighbors from the vertices on fire.
This process stops when the fire cannot spread anymore. The goal is to
find a sequence of vertices for the firefighter that maximizes the number
of saved (non burned) vertices.
The Firefighting problem turns out to be NP-hard even when restricted to bipartite graphs or trees of maximum degree three. We study
the parameterized complexity of the Firefighting problem for various
structural parameterizations. All our parameters measure the distance to
a graph class (in terms of vertex deletion) on which the Firefighting
problem admits a polynomial time algorithm. Specifically, for a graph
class F and a graph G, a vertex subset S is called a modulator to F if
G \ S belongs to F. The parameters we consider are the sizes of modulators to graph classes such as threshold graphs, bounded diameter graphs,
disjoint unions of stars, and split graphs.
To begin with, we show that the problem is W[1]-hard when parameterized by the size of a modulator to diameter at most two graphs and
split graphs. In contrast to the above intractability results, we show that
Firefighting is fixed parameter tractable (FPT) when parameterized
by the size of a modulator to threshold graphs and disjoint unions of
stars, which are subclasses of diameter at most two graphs. We further
investigate the kernelization complexity of these problems to find that
Firefighting admits a polynomial kernel when parameterized by the
size of a modulator to a clique, while it is unlikely to admit a polynomial
kernel when parameterized by the size of a modulator to a disjoint union
of stars.
1
Introduction
The Firefighting problem was introduced by Hartnell [16] to model the spread
of diseases and computer viruses. It is a turn-based game between two players
(the “fire” and the “firefighter”), which is played on a graph G as follows. Initially,
at time t=0, a fire starts at a vertex s, at each following time step the following
happens. A firefighter defends one vertex which is not on fire, and the fire then
?
??
Supported by Tata Consultancy Services (TCS) research fellowship.
Supported by a DST-INSPIRE Fellowship.
spreads from each burning vertex to all its undefended neighbors. Once a vertex
is defended it remains so for all time intervals. The process stops when the
fire can no longer spread. The natural algorithmic question associated with this
game is to find a strategy that optimizes some desirable criteria, for instance,
maximizing the number of saved vertices [5], minimizing the number of rounds,
the number of firefighters per round [6], or the number of burned vertices [10,5],
and so on. These questions are well-studied in the literature, and while most
variants are NP-hard, approximation and parameterized algorithms have been
proposed for various scenarios. In this work, we will focus on the goal of finding a
sequence of defending vertices that maximizes the number of saved (not burned)
vertices and we refer to this as the Firefighting problem. We also use Saving
k-Vertices to refer to the decision version of this problem, where we are given
a demand k and the goal is to save at least k vertices.
We study the parameterized complexity of Firefighting with respect to
various structural parameters. In particular, our focus is on distance-to-triviality
parameterizations, wherein we identify classes of graphs on which the Firefighting problem is solvable in polynomial time, and understand the parameterized complexity of the problem parameterized by the distance of a graph to
these graph classes. In this paper, our notion of distance to a graph class in the
vertex deletion distance. More precisely, for a class F of graphs, we say that X is
an F-modulator of a graph G if there is a subset X ⊆ V(G) such that G \ X ∈ F. If
the size of a smallest modulator to F is k, we also say that the distance of G to
the class F is k. Throughout this paper, we will assume that a modulator is given
to us as a part of the input. This assumption is without loss of generality since
such modulators can be computed in FPT time. We are now ready to describe
our results.
Our Contributions. The Firefighting problem is FPT when parameterized by
the vertex cover and distance to a clique parameterizations. On the other hand,
it is para-NP-hard when parameterized by feedback vertex set, tree-width and
clique-width [11]. However, the parameter vertex cover is very restrictive and
significantly large for dense graphs. This motivates us to consider parameters
that are intermediate between vertex cover and clique-width. In this spirit, Ganian [15] studied the parameterized complexity of Firefighting problem for
parameter twin-cover which is a generalization of vertex cover and showed that
Firefighting is FPT with respect to twin-cover. Recently Chlebı́ková et al. [7]
showed that the problem is FPT parameterized by distance to cluster graphs
which is a generalization of twin-cover.
We study the parameterized complexity of the Firefighting problem with
respect to the distance from following graph classes: threshold graphs, disjoint
union of stars, disjoint union of graphs of diameter at most two, and split graphs.
Studying the parameterized complexity of Firefighting with respect to these
parameters improves the understanding of the boundary between tractable and
intractable parameterizations. For instance, the parameterization by distance to
cluster graphs (as studied by [7]) directly generalizes both vertex cover and distance to clique. Observe that cluster graphs are precisely the graphs whose con2
W[1]-hard
Para NP-hard
Distance to
Rank-width,
Clique-width
diameter ≤ 3 graphs*
Tree-width
Distance to
split graphs *
Feedback
vertex set
Distance to
diameter ≤ 2 graphs*
Distance
to cographs
Distance
to cluster
FPT
Distance
to stars*
Distance to
threshold graphs *
No Poly-kernel
Twin cover
Distance
to clique *
Vertex cover
Poly-kernel
Fig. 1 A schematic showing the parameterized complexity of Firefighting
with respect to various structural graph parameters. There is a line between two
parameters if the parameter below is larger than the parameter above. Results
shown in this paper are marked by an asterisk (*).
nected components have diameter one, and a natural generalization to consider
is the class of graphs whose connected components have diameter two. Here, we
show that there is a transition in complexity: the problem becomes W[1]-hard.
As a natural intermediate problem, we consider the subclass of graphs where
every connected component is a star. Here, using ideas similar to the ones that
lead to the FPT algorithm for the distance to cluster parameter, we obtain a FPT
algorithm. The case analysis here is more delicate because we have to distinguish
between the central vertex and the leaves.
1
On the other hand, the distance to threshold graphs parameter directly generalizes the distance to clique parameter, while the distance to stars parameter
is a generalization of vertex cover. For both of these parameters, we establish
that Firefighting is FPT. These being smaller parameters, our results improve
several known algorithms. Note that the next “natural” parameter to consider
after distance to stars hierarchy is the feedback vertex number, or the distance to
forests; however here the problem is already NP-hard on trees, leading to paraNP-hardness. Similarly, a natural next step from distance to threshold graphs is
the distance to split graphs, but here also we demonstrate W[1]-hardness. Finally,
a promising generalization from the distance to cluster graphs is the distance to
cographs, and here we leave the parameterized complexity of the problem open.
3
We also consider the kernelization complexity of the problem and make the
following advances: for the distance to clique parameterization, we demonstrate a
quadratic kernel, while for the distance to stars parameterization, we show that a
polynomial kernel is unlikely under standard complexity-theoretic assumptions.
The kernelization complexity of the problem relative to vertex cover, however,
remains an interesting open problem. We summarize our results below.
• We show that the problem is fixed parameter tractable (FPT) when param-
eterized by the size of a modulator to threshold graphs, cluster graphs and
disjoint unions of stars.
• We further investigate the kernelization complexity of these problems to find
that Firefighting admits a polynomial kernel when parameterized by the size
of a modulator to a clique, while it is unlikely to admit a polynomial kernel
when parameterized by the size of a modulator to a disjoint union of stars.
• Finally, in contrast to the tractability results, we show that Firefighting
is W[1]-hard when parameterized by the distance to split graphs. In fact,
the problem remains W[1]-hard with respect to the combined parameter
involving the size of the modulator and the number of vertices to be saved.
Methodology. By and large, we use a standard approach for the FPT algorithms:
we guess the behavior of the solution on the modulator and attempt to find a
solution consistent with the guessed behavior. The second part relies on exploiting the structural properties of G \ X, which is the part of the graph outside
the modulator. Usually one is able to group the vertices of G \ X based on the
structure of their neighborhoods in the modulator, and argue that all vertices
of the same “type” have a similar behavior, which leads to a controlled search
space. In the case of threshold graphs, we are able to prove that simple greedy
techniques work within a particular type. On the other hand, for disjoint unions
of stars, we have to account for several scenarios, and the classification of G \ X
is more intricate, and accordingly, we have to account for more cases in the
analysis. The hardness results follow from reductions using standard techniques,
while the kernelization algorithm uses the fact that when G \ X is a large clique,
several vertices behave in a similar fashion, and this observation allows us to
replace the large clique with a much smaller one — a careful argument is required, however, to demonstrate that the instance we constructed in this fashion
is indeed equivalent to the original.
Related Work. The Firefighting problem is known to be NP-hard even for
special classes of graphs, including bipartite graphs [19], trees of maximum degree three [11] and cubic graphs [17]. The firefighter problem can be solved in
linear time on split graphs and co-graphs [14]. From the parameterized complexity point of view, the firefighting problem is W[1]-hard when parameterized
by the number of saved vertices for bipartite graphs [1]. Cai et al. [5] propose
several FPT algorithms and polynomial kernels, and they consider the following
parameters: the number of saved vertices, the number of saved leaves, and the
number of protected vertices. Leung [18] use the random separation method to
4
give FPT algorithms on general graphs parameterized by the number of burnt
vertices and on degree bounded graphs and unicyclic graphs parameterized by
the number of protected vertices. We refer the reader to the survey [12], as well
as the references within, for more details.
2
Preliminaries
In this section, we introduce the notation and the terminology that we will need
to describe our algorithms. Most of our notation is standard. We use [k] to
denote the set {1, 2, . . . , k}. All graphs we consider in this paper are undirected,
connected, finite and simple. For a graph G = (V, E), let V(G) and E(G) denote
the vertex set and edge set of G respectively. An edge in E between vertices x
and y is denoted as xy for simplicity. For a subset X ⊆ V(G), the graph G[X]
denotes the subgraph of G induced by vertices of X. Also, we abuse notation
and use G \ X to refer to the graph obtained from G after removing the vertex
set X. For a vertex v ∈ V(G), N(v) denotes the set of vertices adjacent to v and
N[v] = N(v) ∪ {v} is the closed neighborhood of v.
Graph Classes. We now define the graph classes that we will encounter frequently.
• A graph is a split graph if its vertices can be partitioned into a clique and
an independent set. Split graphs are P5 -free [13].
• The class of P4 -free graphs are called co-graphs.
• A graph is a threshold graph if it can be constructed recursively by adding
an isolated vertex or a universal vertex.
• A cluster graph is a disjoint union of complete graphs. Cluster graphs are
also P3 -free graphs.
It is easy to see that a graph that is both split and co-graph is a threshold
graph. We denote threshold graph (or a split graph) with G = (C, I) where C
and I denote a partition of G into a clique and an independent set. For any two
vertices x, y in a threshold graph G we have either N(x) ⊆ N[y] or N(y) ⊆ N[x].
For a class of graphs F, the distance to F of a graph G is the minimum number
of vertices to be deleted from G to get a graph in F.
Parameterized Complexity. A parameterized problem is a pair Q ⊆ Σ∗ × N,
where Σ is fixed finite alphabet. For an instance (x, k) ∈ Σ∗ × N, k is called the
parameter. We say that a parameterized problem Q is fixed parameter tractable
(FPT) if there exists an algorithm and a computable function f : N → N such
that given (x, k) ∈ Σ∗ × N the algorithm correctly decides whether (x, k) ∈ Q in
f(k)|(x, k)|O(1) time. The function f is usually superpolynomial and only depends
on parameter k. The class XP contains the problems which are solvable in time
|(x, k)|f(k) , the exponent of the running time depends on the parameter k. A
problem is para-NP-hard if it is NP-hard for some fixed values of the parameter.
The complexity class of parameterized intractability is called W[1] (see [9] for
5
the definition). A kernelization algorithm is a polynomial time algorithm that
takes an instance (x, k) of a parameterized problem Q as input and outputs
an equivalent instance (x 0 , k 0 ) of P such that |x 0 | 6 h(k) for some computable
function h and k 0 6 k. If h is polynomial then we say that (x 0 , k 0 ) is a polynomial
kernel. Let P and Q be two parameterized problems. A parameterized reduction
from P to Q is a mapping g : Σ∗ → Σ∗ such that (i) for all x ∈ Σ∗ we have x
is a yes instance of P ⇔ g(x) is a yes instance of Q. (ii) g can be computed in
f(k)|x|O(1) time, where f is computable function and k is the parameter of x. (iii)
k 0 6 h(k) for some computable function h, where k and k 0 are parameters of x
and g(x) respectively. For more details on parameterized complexity the reader
is referred to [9,8].
We now briefly justify our assumption about the modulators being given as
a part of the input to our problems. Consider the following result.
Lemma 1. [4] Let F be a graph class characterized by the finite forbidden induced subgraphs H1 , · · · , Hl . Given a graph G and integer k, there is an FPT
algorithm that finds a subset X ⊆ V(G) of size at most k such that G \ X ∈ F in
O(dk nd ), where d is the size of largest forbidden subgraph.
The class of cluster graphs, threshold graphs and split graphs are P3 , P4 are
P5 free respectively. By using above lemma, Given a graph G and integer k, the
problem of deciding whether there exists a set X of vertices of size at most k
whose deletion results in a cluster graph, threshold graph and split graph is fixed
parameter tractable.
The Firefighting Problem. Finally, we formally define the problems that we consider in this paper, where F is a class of graphs or a graph property.
Firefighting[F]
Input: A graph G, a vertex s, a modulator X ⊆ V(G) such
that G \ X ∈ F.
Parameter: The size k := |X| of the modulator to F.
Question: Find a strategy that maximizes the number of saved
vertices when a fire starts at s?
Saving k-Vertices[F]
Input: A graph G, a vertex s, a modulator X ⊆ V(G) such
that G \ X ∈ F, and an integer k.
Parameter: The size ` := |X| of the modulator to F.
Question: Does there exists a strategy that saves at least k vertices when a fire starts at s?
Whenever F is clear from the context, we drop the explicit mention of it in
the name of the problem. We also abuse notation and use k differently in the
6
two definitions, to retain consistency with standard notation when we present
reductions in the context of Saving k-Vertices.
3
The Parameterized Complexity of Firefighting
Let (G, s) be an instance of Firefighting problem. The vertices that have
not been burned by the fire at the end of the process are called saved (including
defended vertices). A vertex is burned if it is on fire. A strategy for Firefighting
instance (G, s) is a sequence S of vertices {v1 , · · · , vl } where vi represents the
position of the firefighter at time step i. We say that sequence S is a valid
strategy for (G, s) if vertex vi is not burning at the start of time step i and the
process stops at time step l. A strategy S is called minimal if no subset of S
yields a strategy that saves same number of vertices as S. A strategy is optimal
if it is minimal and saves maximum number of vertices.
The following results from the literature will be useful for our algorithms.
Lemma 2. [14] Let G be a graph and s be a vertex of G. Given an ordered set S
of vertices of G, we can verify whether S is a valid strategy for the Firefighting
problem on (G, s) and count the number of vertices saved by S in O(n + m) time,
where n and m denote the number of vertices and edges in G respectively.
Proof. Let S = {v1 , · · · , vk } is strategy in this order. To verify whether the sequence S is a valid strategy, we do BFS on the graph G starting from s. Find the
distance d(s, v) from the source s to each vertex v of S on graph G[(V(G) \ S)∪ {v}].
For i = 1 to k, If S is a valid strategy then distance from source s to the vertex
vi , d(s, vi ) > i, otherwise vertex vi will be burned before time step i. Observe
that the number of vertices burned by S in G equal to the number of vertices
reachable from s in G \ S, which can be found by applying BFS on the graph
G \ S starting from s.
t
u
Corollary 1. Let S be an optimal strategy for the Firefighting problem on
(G, s) then the number of vertices burned by S in G is equal to the number of
vertices reachable from s in G \ S.
Lemma 3. [14] Let (G, s) be an instance of the Firefighting problem, and let
l be the length of a longest induced path in G starting from s. Then any optimal
strategy can defend at most l vertices.
Proof. Let S = {v1 , · · · , vt } be an optimal strategy defended in this order. Since
S is a valid strategy there is an induced path P from s to vt such that all the
vertices on P are burned except vt . Let P be a shortest such path, then P contains
at least t + 1 vertices: otherwise vt will be burned before time t. Since the length
of longest induced path in G starting from s is l, we have t + 1 6 l + 1 which
implies t 6 l.
t
u
Lemma 4. [14] The Firefighting problem can be solved in O(nl ) time on
graphs with length of longest induced path is at most l − 1.
7
Proof. Let (G, s) be an instance of firefighting problem. Since the length of the
longest induced path in G is at most l. From Lemma 3 any optimal strategy can
defend at most l vertices in G. We list out all possible subsets S of V(G) of size
at most l in O(nl ) time. For each such subset S using Lemma 2 test whether
S is valid strategy and count the number of vertices saved by S. The optimal
strategy is the one which saves maximum number of vertices.
t
u
3.1
Parameterization by Distance to Threshold Graphs
In this section we give an FPT algorithm for the Firefighting problem parameterized by the distance to threshold graphs. Without loss of generality, we assume
that threshold graph is connected, otherwise all the connected components that
do not contain the source vertex are trivially saved.
Corollary 2. Let (G, s) be an instance of Firefighting problem. Then any
optimal strategy can defend at most 2k + 2 vertices, where k is the distance to
threshold graphs.
Proof. We know that fire always spreads along an induced path in G. As the
length of the longest induced path in G is at most 2k + 2, from Lemma 3 any
optimal strategy can defend at most 2k + 2 vertices in G.
t
u
Let G be a graph and X ⊆ V(G) of size k such that G \ X = (C, I) is a threshold
graph. We partition the vertices of clique C and independent set I in G \ X
based on their neighborhoods in X. In particular, for every subset Y ⊆ X, let:
TYC := {x ∈ C | N(x) ∩ X = Y } and TYI := {x ∈ I | N(x) ∩ X = Y }.
Notice that in this way we can partition vertices of G \ X into at most 2k+1
subsets (called types), two for each Y ⊆ X. Observe that all vertices in a type
have same neighbors in X, where as they may have different neighbors inside
the threshold graph. The following result shows that when we need to choose
a defending vertex from a type the best strategy is to defend a highest degree
vertex in that type. For a strategy S, let sav(S) denote the number of vertices
saved by the strategy S.
Lemma 5. Let v1 , v2 be two vertices in a type T such that N(v2 ) ⊆ N[v1 ]. Let S
be a strategy containing v2 , which is defended at time step i. If v1 ∈
/ S and not
burning at the start of time step i then sav(S) 6 sav(S 0 ) where S 0 is obtained
from S by replacing v2 with v1 at time step i.
Proof. Using Corollary 1, the vertices burned by S in G are the vertices which
are reachable from s in G \ S. We show that every vertex u which is reachable
from s in G \ S 0 is also reachable from s in G \ S. Let P be a path between s
and u in G \ S 0 . If v2 ∈
/ P then P is a path in G \ S. If v2 ∈ P, then using the
facts that v1 is a burned vertex in strategy S and N(v2 ) ⊆ N[v1 ], we can see that
P 0 = P \ {v2 } ∪ {v1 } is a path between s and u in G \ S. Therefore the number
of vertices burned by S is at least the number of vertices burned by S 0 , which
implies sav(S) 6 sav(S 0 ).
8
Our FPT algorithm now follows by guessing, for each step in the defending
sequence, if the vertex is from the modulator or the type of the vertex from
G \ X. Then it simulates the sequence (by substituting for each guess of a type,
a greedily chosen vertex from that type) to check if it is a valid solution.
Theorem 1. The Firefighting problem can be solved in O((2k+1 +k)2k+2 (n+
m)) time when parameterized by size of modulator to threshold graphs.
Proof. Partition the vertices of G \ X into at most 2k+1 sets. Each time when we
want to defend a vertex we only choose from 2k+1 (types) +k (size of X). From
Lemma 5 we know which vertex has to be defended in a given type.
From Corollary 2 it is clear that we only need to defend at most 2k + 2 times,
therefore there are at most (2k+1 + k)2k+2 possible firefighting strategies. For
each such strategy S, using Lemma 2, test whether S is valid strategy and count
the number of vertices saved by S. The strategy which saves maximum number
of vertices is the optimal strategy. Therefore this procedure takes O((2k+1 +
k)2k+2 (n + m)) time.
t
u
3.2
Parameterization by Distance to Stars
In this section we design an FPT algorithm for the Firefighting problem parameterized by the distance to stars. Recall that this is the minimum number of
vertices to be deleted from G to get a disjoint union of stars. Let X be a k-sized
modulator to disjoint union of stars. Our first observation follows easily from
the bound on the length of the longest induced path.
Corollary 3. Let (G, s) be an instance of Firefighting problem. Any optimal
strategy can defend at most 4k + 2 vertices, where k is distance to disjoint union
of stars.
Proof. We know that fire always spreads along an induced path in G. As the
length of maximum induced path in G is at most 4k + 2, from Lemma 3 any
optimal strategy can defend at most 4k + 2 vertices in G.
t
u
We now define a notion of equivalent stars, which will lead us to partitioning of
the stars in G \ X into types as before. For a star S with center c in G \ X, we
use B(S) to denote the set of vertices in S that have a neighbor in X, and call
these the border vertices. Further, for a nonempty subset Y ⊆ X, we use BY (S) to
denote the set of vertices in S whose neighborhood in X is exactly Y .
Definition 1. Let X ⊆ V(G) such that G \ X is a disjoint union of stars. We
call two stars Si and Sj are equivalent if, (a) N(ci ) = N(cj ), where ci and cj are
the centers of stars Si and Sj respectively. (b) N(Si ) = N(Sj ), (c) |B(Si )| = |B(Sj )|
and (d) For every non-empty subset Y ⊆ X, we have that |BY (Si )| = |BY (Sj )|.
For an equivalence class T , we use bT to denote the size of the border for any
star in T . Our next result bounds the number of equivalence classes. The bound
9
follows roughly from the fact that one can associate a signature with an equivalence class based on condition (c) in Definition 1, which, in turn, can be put
in one-one correspondence with strings of length 2k over an alphabet of size `,
where ` is the maximum possible value of |B(S)| in G \ X.
Lemma 6. Let ` be defined as above and let X ⊆ V(G) be a modulator to disjoint
union of stars of size k. Then, the stars of G \ X can be partitioned into at
k
most O(22k `2 ) equivalence classes.
Proof. First, partition the stars in G \ X into at most 22k sets such that all stars
in each set satisfies conditions (a) and (b) of Definition 1. Now each set of the
partition can be further divided based on the value of |B(Si )| for each star Si
in that set. As 1 6 |B(Si )| 6 ` for all Si ∈ G \ X, each set can be partitioned
into at most ` sets. In order to satisfy condition (d) in Definition 1 each set
k
k
further partitioned into `2 −1 sets. Combining all, there are at most O(22k `2 )
equivalent star partitions of stars in G \ X.
t
u
Our FPT algorithm begins by guessing the behavior of the solution on the
modulator, and builds on the fact that there are a bounded number of equivalence
classes. We defer the details of this algorithm to a full version because of space
constraints, and note that it is similar — in spirit — to the approach used for
the cluster vertex deletion parameter in [7].
Theorem 2. The Firefighting problem is FPT when parameterized by the
distance from the class of the disjoint union of stars.
Proof. Without loss of generality, assume s ∈ X: If s ∈
/ X then G \ (X ∪ {s}) is
also a disjoint union of stars. As a first step, we guess the defended (Xd ), saved
(Xs ) (which are not in Xd ) and burned (Xb ) vertices from X in O(3k ) time. Since
the length of any optimal defending sequence is at most 4k + 2, we can guess
the set of defended vertices (Xd ) from X along with their positions in an optimal
defending sequence in O(kk (4k + 2)k ) time. The rest of this proof is divided into
two cases depending on whether Xs = ∅ or Xs 6= ∅.
Case 1: Xs = ∅. Since the fire starts at s, all the vertices in V(G)\Xd which are not
reachable from s in G[V(G) \ Xd ] are saved from fire. We begin by deleting them
from G and concentrate on the connected component Cs of G \ Xd containing
the vertex s. In particular, the instance of firefighting problem that we solve
from here is (G[Cs ], s, Xb ) where Xb is the modulator to disjoint union to stars.
Also, we note at this point that if Xb = ∅, then we are done by a straightforward
case analysis on the nature of N(s) and simulating the guessed defence sequence.
Therefore, we assume that Xb 6= ∅.
Partition the stars in Cs \ Xb into equivalence classes (with respect to Xb )
using Definition 1. We divide equivalence classes of stars in Cs \ Xb into two
categories: equivalence classes with bT 6 4k + 2 as one category and remaining
as the other. We combine all equivalence classes T , with bT > 4k + 2 to into
one new equivalence class Tnew . Therefore from Lemma 6 there are at most
10
k
O(k2k (4k + 2)2 ) equivalence classes of stars in Cs \ Xb . From Corollary 3, the
length of any optimal defending sequence S is at most 4k + 2, therefore we can
defend for at most 4k + 2 time steps. Since we have already guessed defending
vertices Xd from X, the remaining defending vertices in S has to be from Cs \ Xb .
At each time step when we want to choose a defending vertex from Cs \ Xb , we
k
choose from O(k2k (4k + 2)2 ) equivalence classes. Therefore there are at most
k
O((k2k (4k + 2)2 ))
4k+2
possible firefighting defending sequences. Let S be one
such possible defending sequence, then each position of S represents either a
vertex of Xd or an equivalence class in G \ Xb .
Now we describe the procedure to select a defending vertex from an equivalence class. Let T be an equivalence class such that bT 6 4k + 2. First, count
q, the number of times T appeared in strategy S. Since T can appear at most
4k + 2 times, q 6 4k + 2. In each star, we need to defend at least one vertex so
we can defend at most 4k + 2 stars. Order the stars in T in descending order of
size |Si \ B(Si )|. We choose defending vertices from set D = ∪4k+2
i=1 (B(Si ) ∪ {ci })
(i.e, union is over first 4k + 2 stars ordered according to |Si \ B(Si )|). The size of
D is at most (4k + 2)(4k + 3).
In equivalence class Tnew , we only need to defend center of stars in Tnew . First
count q number of times Tnew is appeared in sequence S, we can only defend
centers of q different stars. Since each star center is at different distance from
source s, so we can not greedily defend center of a star for which |Si \ B(Si )| is
maximum. The defending vertex set D is defined as follows. For each i ∈ [4k + 2],
find top i stars according to size of |Si \ B(Si )| whose center is at a distance i
from source s. Add the centers of these stars to set D. It is easy to see that
|D| 6 (4k + 2)2 .
Case 2: Xs 6= ∅. Let Xs1 , Xs2 , · · · , Xsm 0 and Xb1 , Xb2 , · · · , Xbm 00 be the connected components of G[Xs ] and G[Xb ] respectively. Recall that we are working in the connected component containing s in the graph G \ Xd . Let S be a star in G \ (Xb ∪Xs ).
We say that S is vulnerable if it has at least one vertex u that has a neighbor
in Xb and at least one vertex v that has a neighbor in Xs (potentially u = v).
For any vulnerable star S, note that any firefighting strategy that respects the
semantics of Xs , Xb and Xd must defend at least one vertex in S. Indeed, suppose
not, and let S be a firefighting strategy that violates this criteria. Then after S
is executed, S has at least one undefended vertex (say p) that is a neighbor of
some vertex in Xb (say x) and one undefended vertex (say q) that is a neighbor
of some vertex in Xd (say y). Then, note that the (x, p, q, y) is a path from Xb to
Xs in the graph after the S is executed, contradicting the semantics of Xs , Xb and
Xd . Observe that this argument works even if p = q. To discuss the consequence
of this observation further, we employ the following notation:
Bb (Si ) := {x ∈ Si | N(x) ∩ Xb 6= ∅} and Bs (Si ) := {x ∈ Si | N(x) ∩ Xs 6= ∅}.
Note now that since at least one vertex has to be defended from every vulnerable star, the total number of vulnerable stars ` is at most 4k + 2.
11
We place all the vulnerable stars in one equivalence class (call this T 0 ). Now
two cases remain: the first is that every border vertex of the star S has all its
neighbors in Xs or Bb (Si ) = ∅. We place all such stars also in one equivalence
class (call this T ? ). Observe that in this case, representative vertices for the
placement of firefighters may be chosen arbitrarily. This is because the entire
star is already safe with respect to any firefighting strategy that respects the
semantics of Xb , Xs and Xd . Now consider the stars that have all their neighbors
in Xb , and note that it is not possible to save vertices in Xs by defending vertices
in such stars. Therefore, we treat such equivalence classes just as we would in
the previous case. Overall, the guess of the template sequence S happens over
the equivalence classes determined according to Definition 1 with respect to Xb
and the two special equivalence classes that we just defined, namely T ? and
T 0 . In a valid sequence, the equivalence class T 0 must appear at least ` times.
Representatives here have to be chosen from a bounded border, so a brute force
approach suffices in this situation. Representatives from T ? are chosen arbitrarily
and for all other cases they are chosen according to the description in Case 1.
To summarize, we now have a strategy for identifying a representative vertex
from each equivalence class. For each possible defending sequence S, by using
Lemma 2, we check whether it is valid and count the number of vertices of saved
by S. We output the strategy which saves the maximum number of vertices.
3.3
Parameterization by Distance to Diameter Two Graphs
In this section, we show that Saving k-Vertices is W[1]-hard when parameterized by k and distance to diameter two graphs by giving a reduction from the
k-clique problem. The reduction is similar, in spirit, to the one used in [1].
Theorem 3. Saving k-Vertices is W[1]-hard parameterized by (k + l), where
l is the distance from the class of graphs with diameter two.
Proof. Let (G, k) be a instance of k-clique problem. We construct a graph G 0
from G as follows (see Fig. 2). For each vertex v in V(G) add vertex cv in G 0 :
this set of vertices is denoted by V . For each edge (u, v) in G add a vertex iuv
in G 0 : this set of vertices is denoted by E. Add an edge from vertex iuv in E to
both vertices cu and cv in V for each (u, v) in E(G). Add a root vertex s, and
add vertices dij (set of vertices denoted by D) for 1 6 i 6 k and 1 6 j 6 k + 1.
Add edges between dij to di 0 j 0 for all i, j, j 0 and i 0 = i + 1. Connect s to d1,j for
all j and connect dk,j to each vertex in V for all j. Add a vertex z adjacent to all
vertices of V ∪ E. The vertex set of G 0 is V(G 0 ) = {s, z} ∪ D ∪ V ∪ E. It is easy
to
see that G 0 is k(k + 1) + 1 distance to diameter two graphs. Set k 0 = k + k2 + 2
for saving vertices in G 0 .
We prove that Saving k-Vertices on (G 0 , s, k 0 ) is a yes instance if and only
if k-clique on (G, k) is a yes instance. Suppose that there is a clique K of size k in
G. Let S be a valid strategy described as follows: At each time step t = 1, · · · , k
S defends the vertices cv for all v ∈ K, at time step k + 1, S defends the vertex
z and at time step k + 2, S defends a vertex iuv , for some edge (u, v) 6∈ E(G[K]).
12
D
V
d11
d21
d31
d12
d22
d32
d13
d23
d33
E
c1
i12
3
c2
s
2
c3
i24
i34
4
c4
1
i23
5
d14
d24
i45
c5
d34
z
Fig. 2 Graph G (left) and Distance to diameter two graph G 0 (right). The vertex
z is adjacent to all vertices in V ∪ E, denoted with thick edges.
So the strategy S saves k + 1 + k2 + 1 vertices in G: k vertices in V and k2 + 1
vertices in E.
For other direction, suppose that S = {p1 , p2 , · · · , pl } is a valid strategy that
saves at least k 0 vertices in G 0 . It is easy to see that at any time step defending a
vertex dij for any i, j is not helpful, because there is at least one burning vertex
di,j 0 with same neighborhood as dij . Further defending vertices in E does not save
any vertices. Since the vertex z is universal, therefore it needs to be defended.
So the vertices in S ∩ V are responsible to save at least k2 many vertices, which
is possible only if the vertices defended in V induces k-clique in G.
t
u
We can obtain the hardness of Saving k-Vertices by a reduction from the
k-clique problem as well, in fact by making minor changes to the reduction used
to prove Theorem 3.
Corollary 4. Saving k-Vertices is W[1]-hard parameterized by (k + l), where
l is distance to split graphs.
Proof. The proof is similar to Theorem 3, except the following minor changes
in the construction of graph G 0 . Add an edge between every pair of vertices in
V and remove the universal vertex z. Add vertices dij for 1 6 i 6 k − 1 and
1 6 j 6 k. Add edges between dij to di 0 j 0 for all i, j, j 0 and i 0 = i + 1. Connects to
d1,j for all j and connect dk−1,j to each vertex in V for all j. Set k 0 = k + k2 + 1
for saving vertices in G 0 . The graph G 0 is k(k − 1) + 1 distance to a split graph.
It is easy to see that Saving k-Vertices on (G 0 , s, k 0 ) is a yes instance if and
only if k-clique on (G, k) is a yes instance.
t
u
13
4
4.1
Kernelization Complexity
Parameterization by Distance to Clique
In this section, we give a polynomial kernel for Saving k-Vertices when parameterized by distance to clique, as summarized in the following theorem.
Theorem 4. Saving k-Vertices admits a polynomial kernel of size at most
O(l2 ), where l is the distance to clique.
Let (G, s, k) be an instance of Saving k-Vertices and X ⊆ V(G) of size
l such that G \ X = C is a clique. With out loss of generality we assume that
s ∈ X, otherwise define X 0 = X ∪ {s} such that G \ X 0 is a clique with size of
modulator l + 1. Since the length of longest induced path in G is at most l + 1,
using Lemma 3 we get the following Corollary.
Corollary 5. Given an instance of Firefighting, then any optimal strategy
can defend at most l + 1 vertices, where l is size of the clique modulator X.
Let XL := {x ∈ X : |N(x) ∩ C| 6 l + 1}, XH = X \ XL . Let J := {y ∈ C | ∃x ∈
XL , y ∈ N(x)}, |J| 6 l(l + 1). Our kernelization algorithm is based on the following
observation. We show that replacing the clique G[C \ J] with another clique of
small size does not affect the solution.
Lemma 7. Let S be a valid strategy containing a vertex v ∈ C \ J, then all the
vertices of N(v) \ S are burned.
Proof. Since S is a valid strategy, there is an induced path P from s to v such
that all vertices on P are burned, except v. Let Xv := N(v) ∩ X and for every
x ∈ Xv \ S, we have |N(x) ∩ C| > l + 1: suppose there is x ∈ N(v) ∩ X such that
|N(x) ∩ C| 6 l + 1 then v ∈ J, contradiction to v ∈ C \ J.
Let u be a burned neighbor of v on P. If u ∈ C then C \ S is burned and for
every x ∈ Xv \ S, we have |N(x) ∩ C| > l + 1. From Corollary 5 we can only defend
at most l + 1 vertices, therefore all vertices of Xv \ S are burned.
If u ∈ Xv \ S then |N(u) ∩ C| > l + 1, therefore all vertices C \ S are burned.
From the first case we can see that all vertices of Xv \ S are also burned.
t
u
Reduction rules:
1. Delete vertices of C \ J from G. Add a clique K of size l + 2 and make each
vertex of K adjacent to all vertices in XH ∪ J.
2. Add another clique L of size min{l + 1, |C \ J|} and for each vertex u ∈ L, add
edges between u and J ∪ K.
Let H be the graph obtained after applying above reduction rules. It is easy
to see that X ⊆ V(H) such that H \ X is a clique. The size of the reduced instance
H is at most l2 + 4l + 3 and the reduction can be done in polynomial time.
Let C = G \ X and C 0 = H \ X. We may assume that |C \ J| > 2l + 3; otherwise,
trivially we get a kernel of size at most l2 + 4l + 3.
14
Remark 1. Let G be a graph and S be a valid strategy. If S defends a subset S1
of vertices in C \ J, then defending any subset S2 of vertices in L instead of S1 ,
with |S1 | = |S2 | is also a valid strategy S 0 for H.
Conversely let S 0 be a strategy on H. If S 0 defends a subset S1 of vertices
in K ∪ L, then defending any subset S2 of vertices in C \ J instead of S1 , with
|S1 | = |S2 | is also a valid strategy S for G.
Lemma 8. Let G and H be graphs as defined above and S, S 0 be corresponding
valid strategies as defined in the above remark. At least one vertex of C is burned
in G by strategy S iff at least one vertex of C 0 is burned in H by strategy S 0 .
Proof. Let u ∈ C 0 is burned by strategy S 0 in H. If u ∈ J, then u ∈ C will also
burned by S in G. If u ∈ K and no vertex of J burned by S 0 in H then, at least
one vertex x ∈ XH gets burned and |N(x)∩C| > l+1, therefore at least one vertex
in N(x) ∩ C gets burned by S in G.
Let u ∈ C is burned by strategy S in G. If u ∈ J then it will be burned by S 0
in H. If u ∈ C \ J and no vertex of J burned by S in G then there exists burned
vertex x ∈ XH . Since |N(x) ∩ C 0 | > l + 1, at least one vertex of K gets burned. t
u
Claim 1 If a strategy S saves at least 2l + 1 vertices in G then S saves entire
clique.
Proof. Let S be a valid strategy which saves at least 2l + 1 vertices in G. Suppose
assume that there exists a vertex v ∈ C burned by strategy S, then no vertex in
C is saved except the defended vertices. The strategy S can save at most l − 1
vertices in X and l + 1 vertices in C, total S saves at most 2l vertices which is a
contradiction to the fact that S saves at least 2l + 1 vertices.
t
u
Claim 2 Saving one undefended vertex in C is equivalent to saving entire clique.
Proof. Let v be an undefended saved vertex of C. If any vertex of C is burned
by S then v gets burned contradicting the fact that v is a saved vertex.
t
u
Lemma 9. Let G be a graph and H be the graph obtained after applying reduction rules. A strategy S saves at least k vertices in G if and only if there exists
a strategy S 0 that saves at least k 0 vertices in H.
Proof. Let S be any valid strategy on G then we can find a corresponding strategy
S 0 on H as follows. If S defends a vertex v in C \ J which is deleted in reduction
procedure, then by Remark 1 instead of v we can defend any other non-defended
vertex in L. Since S can defend at most l + 1 vertices in G and |L| = l + 1, we can
always replace vertices of C \ J in S by applying Remark 1.
Conversely, let S 0 be any valid strategy on H then we can find its corresponding strategy S on G as follows. If S 0 defends a vertex v in K ∪ L which is added
in reduction procedure, then by Remark 1 instead of v we can defend any other
non-defended vertex in C \ J. Since S 0 can defend at most l + 1 vertices in H and
|C \ J| > 2l + 3, we can always replace vertices of K ∪ J in S 0 by applying Remark 1.
The parameter k 0 is defined as follows.
15
1. If k ∈ [|C|, |C| + l − 1], then set k 0 = k − |C| + |J| + |K| + |L| 6 l2 + 4l + 3.
2. If k ∈ [2l + 1, |C| − 1], then set k 0 = 2l + 1.
3. If k ∈ [1, 2l], then set k 0 = k.
Using Lemma 8, Claims 1 and 2 we can see that the strategy S saves at least
k vertices in G if and only if the strategy S 0 saves at least k 0 vertices in H.
t
u
4.2
Parameterization by Distance to Stars
In this section, we show that Saving k-Vertices does not admit a polynomial
kernel parameterized by distance to stars unless NP ⊆ coNP/poly. First we
introduce a notion of polynomial time and parameter transformation, that under
the assumption NP * coNP/poly allows us to show non-existence of polynomial
kernels.
Definition 2. [3] Let P and Q be parameterized problems, we say that there is
a polynomial-parameter transformation from P to Q, denoted P 6ppt Q, if there
exists a polynomial time computable function f : {0, 1}∗ × N → {0, 1}∗ × N and
polynomial p : N → N, for all x ∈ {0, 1}∗ and k ∈ N, then following hold
1. (x, k) ∈ P if and only if (x 0 , k 0 ) = f(x, k) ∈ Q
2. k 0 6 p(k)
The function f is called a polynomial-parameter transformation from P to Q.
Theorem 5. [3] Let P and Q be parameterized problems, and suppose that Pc
and Qc are the derived classical problems. Suppose that Pc is NP-complete, and
Qc ∈ NP. Suppose that f is a polynomial time and parameter transformation from
P to Q. Then, if Q has a polynomial kernel, then P has a polynomial kernel.
Corollary 6. Let P and Q be parameterized problems whose unparameterized
versions Pc and Qc are NP-complete. If P 6ppt Q and P does not have a polynomial kernel, then Q does not have a polynomial kernel.
Clique [Vertex Cover]
Input: A graph G, a vertex cover X ⊆ V(G) and an integer
k.
Parameter: The size l := |X| of the vertex cover.
Question: Does there exists a clique of size k in G?
Theorem 6. [2] Clique [Vertex Cover] does not admit a polynomial kernel
unless NP ⊆ coNP/poly.
Theorem 7. Saving k-Vertices parameterized by the distance to disjoint union
of stars does not admit a polynomial kernel, unless NP ⊆ coNP/poly.
16
V
D
d11
X
d21
J
J24
2
J23
3
4
J34
2
4
1
5
s
d12
3
d22
J12
1
G
J45
d13
5
d23
H
I
Fig. 3 An instance of Clique [Vertex Cover] and the constructed graph H
represents Saving k-Vertices [Distance to stars] in the proof of Theorem 7 for
l = 2 and k = 3. In this example X = {2, 4} is a vertex cover of G and H\({s}∪D∪X)
is disjoint union of stars (graph induced by blue vertices).
Proof. We show that there is a polynomial-parameter transformation from Clique
[Vertex Cover] to Saving k-Vertices [Distance to stars]. Let (G, X, l, k) be
an instance of Clique [Vertex Cover]. We construct a graph H (see Fig. 3)
from G as follows. Initialize H as a copy of G, this set of vertices are denoted by
V = X ∪ I For each edge uv ∈ E(G) add a vertex Juv to H and make it adjacent
to both vertices u and v. we denote this set of vertices with J.
Add a root vertex s, and add vertices dij (set of vertices denoted by D) for
1 6 i 6 k − 1 and 1 6 j 6 k. Add edges between dij to di 0 j 0 for all i, j, j 0 and
i 0 = i + 1. Connect s to d1,j for all j and connect dk−1,j to each vertex of G
in H for all j. Vertex set of H is V(H) = {s} ∪ D ∪ V ∪ J. It is easy to see that
H \ ({s} ∪ D ∪ X) is a disjoint union of stars. Therefore H is k(k − 1) + l + 1 distance
to stars with X 0 = {s} ∪ D ∪ X as a modulator. Since k 6 l + 1, H is at most (l + 1)2
distance to stars. Let l 0 := |X 0 | is polynomially bounded in the input parameter
l and we can construct the graph H in polynomial time. Set k 0 = k + k2 + 1 for
saving vertices in H. Let (H, X 0 , l 0 , k 0 ) is an instance of the Saving k-Vertices
problem parameterized by a distance to stars.
We show that Saving k-Vertices on (H, X 0 , l 0 , k 0 ) is a yes instance if and
only if there is a k-clique on (G, X, l,1k) is a yes instance. Suppose that (G, X, l, k)
is a yes-instance of Clique [Vertex Cover] and C be a k-clique in G. We can
find a valid strategy S which saves at least k 0 vertices in H as follows: For first
k time steps S defends k vertices of C in H. At time step k + 1, S defends some
vertex Juv in H such that uv 6∈ E(C). The strategy S saves k vertices in C and
saves k2 vertices in H corresponding to clique edges in G and the vertex Juv .
Conversely, suppose that S is a valid strategy that saves at least k 0 vertices
in H. It is easy to see that at any time step, defending a vertex dij for any
i, j is not helpful, because there is at least one burning vertex di,j 0 with same
neighborhood as dij . Further protecting Juv vertices for some uv ∈ E(G) does
not
save any vertices. So the vertices in S ∩ V are responsible for saving at least
k
2 many vertices.
17
If |S ∩ I| = d > 2 then in the best case S can save at most k 00 = k + 1 + k−d
+
2
d(k − d) = k + 1 + (k − d)(k + d − 1)/2 vertices in H. Observe that k 00 is less than
k 0 for d > 2, contradiction to fact that the strategy S saves at least k 0 vertices
in H. Therefore |S ∩ I| 6 1. If S ∩ I = ∅ then the strategy S saves at least k 0
vertices if and only if the vertices defended in S ∩ X are forms a k-clique in G.
If |S ∩ I| = 1 then S defends one vertex v in I and k − 1 vertices in X. S saves at
least k 0 vertices if k − 1 vertices defended in X induces a (k − 1)-clique in G and v
is adjacent to k − 1 defended vertices. Combining both cases, S can save at least
k 0 vertices in H if the vertices defended in S ∩ V forms a k-clique in G.
t
u
5
Conclusion
In this paper, we studied the parameterized complexity of Firefighting problem for various structural parameters. We considered the size of a modulator
to threshold graphs, cluster graphs and disjoint unions of stars as parameters,
and showed FPT algorithms in all cases. We also established that Firefighting
admits a polynomial kernel when parameterized by the size of a modulator to a
clique, while it is unlikely to admit a polynomial kernel when parameterized by
the size of a modulator to a disjoint union of stars. In contrast to the tractability
results, we found that Firefighting is W[1]-hard when parameterized by the
distance to split graphs.
The following problems remain open. Does Saving k-Vertices admit a
polynomial kernel when parameterized by k and the size of a vertex cover? Also,
is the Firefighting problem FPT when parameterized by distance to cographs?
References
1. Bazgan, C., Chopin, M., Cygan, M., Fellows, M.R., Fomin, F.V., van Leeuwen,
E.J.: Parameterized complexity of firefighting. Journal of Computer and System
Sciences 80(7), 1285–1297 (2014)
2. Bodlaender, H.L., Jansen, B.M., Kratsch, S.: Kernelization lower bounds by crosscomposition. SIAM Journal on Discrete Mathematics 28(1), 277–305 (2014)
3. Bodlaender, H.L., Thomassé, S., Yeo, A.: Kernel bounds for disjoint cycles and
disjoint paths. Theoretical Computer Science 412(35), 4570–4578 (2011)
4. Cai, L.: Fixed-parameter tractability of graph modification problems for hereditary
properties. Information Processing Letters 58(4), 171–176 (1996)
5. Cai, L., Verbin, E., Yang, L.: Firefighting on trees:(1- 1/e)–approximation, fixed
parameter tractability and a subexponential algorithm. In: International Symposium on Algorithms and Computation. pp. 258–269. Springer (2008)
6. Chalermsook, P., Chuzhoy, J.: Resource minimization for fire containment. In: Proceedings of the twenty-first annual ACM-SIAM symposium on Discrete Algorithms.
pp. 1334–1349. Society for Industrial and Applied Mathematics (2010)
7. Chlebı́ková, J., Chopin, M.: The firefighter problem: Further steps in understanding
its complexity. Theoretical Computer Science 676, 42–51 (2017)
8. Cygan, M., Fomin, F.V., Kowalik, L., Lokshtanov, D., Marx, D., Pilipczuk, M.,
Pilipczuk, M., Saurabh, S.: Parameterized algorithms, vol. 4. Springer (2015)
18
9. Downey, R.G., Fellows, M.R.: Parameterized complexity, vol. 3. springer Heidelberg
(1999)
10. Finbow, S., Hartnell, B., Li, Q., Schmeisser, K.: On minimizing the effects of fire
or a virus on a network. Journal of Combinatorial Mathematics and Combinatorial
Computing 33, 311–322 (2000)
11. Finbow, S., King, A., MacGillivray, G., Rizzi, R.: The firefighter problem for graphs
of maximum degree three. Discrete Mathematics 307(16), 2094–2105 (2007)
12. Finbow, S., MacGillivray, G.: The firefighter problem: a survey of results, directions
and questions. The Australasian Journal of Combinatorics 43, 57–77 (2009)
13. Foldes, S., Hammer, P.L.: Split graphs. Universität Bonn. Institut für Ökonometrie
und Operations Research (1976)
14. Fomin, F.V., Heggernes, P., van Leeuwen, E.J.: The firefighter problem on graph
classes. Theoretical Computer Science 613, 38–50 (2016)
15. Ganian, R.: Improving vertex cover as a graph parameter. Discrete Mathematics
& Theoretical Computer Science 17(2), 77–100 (2015)
16. Hartnell, B.: Firefighter! an application of domination. In: 25th Manitoba Conference on Combinatorial Mathematics and Computing (1995)
17. King, A., MacGillivray, G.: The firefighter problem for cubic graphs. Discrete
Mathematics 310(3), 614–621 (2010)
18. Leung, M.L.: Fixed parameter tractable algorithm for firefighting problem. arXiv
preprint arXiv:1104.1044 (2011)
19. MacGillivray, G., Wang, P.: On the firefighter problem. Journal of Combinatorial
Mathematics and Combinatorial Computing 47, 83–96 (2003)
19
| 8cs.DS
|
1
Multi-platform Version of StarCraft: Brood War in
a Docker Container: Technical Report
arXiv:1801.02193v1 [cs.AI] 7 Jan 2018
Michal Šustr, Jan Malý and Michal Čertický
Abstract—We present a dockerized version of a real-time
strategy game StarCraft: Brood War, commonly used as a domain
for AI research, with a pre-installed collection of AI developement
tools supporting all the major types of StarCraft bots. This
provides a convenient way to deploy StarCraft AIs on numerous
hosts at once and across multiple platforms despite limited OS
support of StarCraft. In this technical report, we describe the
design of our Docker images and present a few use cases.
I. I NTRODUCTION
Games have traditionally been used as domains for Artificial Intelligence (AI) research, since they represent a welldefined challenges with various degrees of complexity. They
are also easy to understand and provide a way to compare
the performance of AI algorithms and that of human players.
After a recent success with popular games like Go [1] and
Poker [2], both of which can now be played by AIs at superhuman skill level, the attention of researchers has turned to
a more complex challenge represented by Real-Time Strategy
(RTS) games. In addition to numerous academic researchers,
commercial companies like Facebook and DeepMind have
recently expressed an interest in using the most popular RTS
game of all time, StarCraft1 , as a test environment for their
AI research [3].
Given the recent interest in StarCraft as a testbed for modern
AI techniques, the research community often needs to be
able to deploy multiple instances of the game at once and
combine it with various other tools across different platforms.
Currently, the research applicatons of StarCraft: Brood War
are limited by the fact that the game is only supported
on Microsoft Windows and Mac OS. Mass deployment of
multiple instances tends to run into licensing issues given the
proprietary nature of supported operating systems. There are
a few attempts to address some of these problems, such as
the TorchCraft ML framework [4], but we believe that simple self-contained Docker containers2 with no accompanying
requirements have a potential to provide a more versatile and
easy-to-use soltion. Our implementation is publicly available
at: https://github.com/Games-and-Simulations/sc-docker
Michal Čertický, Jan Malý and Michal Šustr are with the Games
& Simulations Research Group (http://gas.fel.cvut.cz/), Artificial Intelligence Center at Faculty of Electrical Engineering, Czech Technical University in Prague. Emails: [email protected],
[email protected], [email protected]
1 StarCraft and StarCraft: Brood War are trademarks of Blizzard Entertainment, Inc. in the U.S. and/or other countries.
2 https://www.docker.com
II. S TAR C RAFT: B ROOD WAR AS A TESTBED FOR AI
RESEARCH
In a typical RTS, a subgenre of strategy computer games,
players need to develop an economy by gathering resources,
building structures and researching technologies, train a military force and try to defeat their opponents with it. StarCraft
games, placed within a well-knon science-fiction universe,
represent a typical example of the RTS genre. Most successful
entry in the StarCraft series is StarCraft: Brood War, released
in 1998 by Blizzard Entertainment. Despite its considerable
age, the game is still immensely popular thanks to its mechanics, depth and exceptionally balanced set of rules.
RTS games like StarCraft put additional demands on modern AI techniques, compared to board game such as Go and
Chess. According to [5] most important differences are:
• In RTS, each player has a small amount of time to decide
the next action, before the game state changes.
• RTS games are simultaneous as players can issue actions
at the same time. On top of that, some actions are durative
and may not be instantaneous.
• RTS games are typically nondeterministic and partially
observable. Sometimes, there is a chance that an action
will fail or has unpredictable consequences. In most
cases, each player can only see a part of the game world.
• The complexity of RTS is estimated to be many orders
of magnitude larger than any of the board game such as
Go or Chess, both in terms of state-space size and in a
number of actions available at each decision cycle.
Additional reason for the popularity of StarCraft: Brood War
in the AI research community is the fact that the performance
of AI algorithms can be directly compared in three StarCraft
AI tournaments: SSCAIT, CIG and AIIDE. More information
about the competitions and their contestants can be found in
[6] and [7].
III. U SE CASES
In this section, we provide a few use cases for the dockerized version of StarCraft: Brood War. These should be taken as
a mere inspiration. The project has many different applications
in situations, where a user needs to programatically run many
StarCraft AI games, or when StarCraft needs to run on
unsupported operating systems.
• Testing against different bots locally: A bot developer
can easily test his own bot against many different bots
on a single machine with no need to solve their diverse
requirements. Dockerized version of StarCraft contains
pre-installed prerequisites of all the typical kinds of bots.
2
BWAPI related data. Slaves can also write to shared directories
(for logging or machine learning). Shared space is also used
as cache. For example, bots share map analysis cache files
created by BWTA2 library5 .
Fig. 1. A LAN game of four StarCraft AI bots running on one Linux OS
machine. Each bot runs in one Docker container.
•
•
User can even use provided scripts that automatically
download and start the newest version of any bot from
SSCAIT tournament using the tournament’s API.3
Testing with varying system resources: System resources
available to each Docker container are easily configurable
and can be distributed among the bots fairly. Test results
should therefore be reliable and replicable. Reliability and
replicability of bot’s results are especially important in
research. It can also help competition contestants, who
can test how their bots will perform on the hardware of
specific tournament.
Mass deployment for machine learning: Running thousands of games in parallel and as quickly as possible is
vital for training ML-based bots. For example, training a
bot using self-play helped master the game of Go [1].
IV. A RCHITECTURE
The ability to extend other docker images is what makes
Docker powerful. We used this feature for our containers as
well. First, we created Docker image with Wine4 (a compatibility layer allowing us to run MS Windows programs
under Unix-like OS). We extended this image to add StarCraft:
Brood War and a bot-coding API called BWAPI [8]. To
support Java bots, we built another container with a preinstalled Java environment. Finally, we made the last image
with a collection of useful scripts allowing users to run the
games easily.
Each docker container is connected to its host. One can
think of this setup in terms of master/slave communication
model, where docker instances are slaves and host is the
master.
This setup is depicted in Figure 2. The master contains
directories (entitled DIRS in the figure) with the data read
by the slaves. Directories are mapped through VOLUMES in
slaves. For example, each slave can read StarCraft maps and
Fig. 2. Communication between the Docker container and its host.
Our docker container supports two modes of running
StarCraft: Brood War. It is possible to run the game with
a bot in so-called headless mode, which can save some
computational resources by disabling GUI. This mode is useful
for example for self-play. The headful mode, with the GUI
enabled, is useful for debugging. In this case, the host can
connect with its VNC client to VNC server running in each
container as is shown in Figure 2.
V. C ONCLUSION
We introduced a self-contained dockerized AI research
environment consisting of StarCraft: Brood War game and a
collection of related AI-development tools. This provides a
convenient way of massive deployment of StarCraft AIs across
multiple platforms. Thanks to our scripts, it is much easier to
test, train and use custom StarCraft AIs. Researchers and AI
developers can benefit from the possibility to specify resources
for container making results more reliable and replicable. We
plan to add new features to our containers as well as support
for OpenBW6 and StarCraft: Remastered. We also consider
extending one of AI benchmark collections represented by
OpenAI Gym [9] with the StarCraft environment.
3 https://sscait.docs.apiary.io/
5 https://bitbucket.org/auriarte/bwta2
4 https://www.winehq.org
6 http://www.openbw.com
3
R EFERENCES
[1] D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. Van
Den Driessche, J. Schrittwieser, I. Antonoglou, V. Panneershelvam,
M. Lanctot et al., “Mastering the game of go with deep neural networks
and tree search,” Nature, vol. 529, no. 7587, pp. 484–489, 2016.
[2] M. Moravčı́k, M. Schmid, N. Burch, V. Lisỳ, D. Morrill, N. Bard,
T. Davis, K. Waugh, M. Johanson, and M. Bowling, “Deepstack: Expertlevel artificial intelligence in heads-up no-limit poker,” Science, vol. 356,
no. 6337, pp. 508–513, 2017.
[3] E. Gibney, “What google’s winning go algorithm will do next.” Nature,
vol. 531, no. 7594, pp. 284–285, 2016.
[4] G. Synnaeve, N. Nardelli, A. Auvolat, S. Chintala, T. Lacroix, Z. Lin,
F. Richoux, and N. Usunier, “Torchcraft: a library for machine learning
research on real-time strategy games,” arXiv preprint arXiv:1611.00625,
2016.
[5] S. Ontanón, G. Synnaeve, A. Uriarte, F. Richoux, D. Churchill, and
M. Preuss, “A survey of real-time strategy game AI research and competition in StarCraft,” IEEE Transactions on Computational Intelligence
and AI in Games, vol. 5, pp. 293–311, 2013.
[6] M. Certicky and D. Churchill, “The Current State of StarCraft AI Competitions and Bots,” in AIIDE 2017 Workshop on Artificial Intelligence
for Strategy Games, 2017.
[7] D. Churchill, M. Preuss, F. Richoux, G. Synnaeve, A. Uriarte, S. Ontanón,
and M. Certicky, “Starcraft bots and competitions,” in Encyclopedia of
Computer Graphics and Games, 2016.
[8] A. Heinermann, “Broodwar API,” https://github.com/bwapi/bwapi, 2013.
[Online]. Available: https://github.com/bwapi/bwapi
[9] G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman,
J. Tang, and W. Zaremba, “Openai gym,” 2016.
| 2cs.AI
|
1
arXiv:1703.10049v2 [cs.RO] 12 Sep 2017
Autonomous Recharging and Flight Mission Planning for
Battery-operated Autonomous Drones
CHIEN-MING TSENG, Masdar Institute
CHI-KIN CHAU, Masdar Institute
KHALED ELBASSIONI, Masdar Institute
MAJID KHONJI, Masdar Institute
Autonomous drones (also known as unmanned aerial vehicles) are increasingly popular for diverse
applications of light-weight delivery and as substitutions of manned operations in remote locations. The
computing systems for drones are becoming a new venue for research in cyber-physical systems. Autonomous
drones require integrated intelligent decision systems to control and manage their flight missions in the
absence of human operators. One of the most crucial aspects of drone mission control and management is
related to the optimization of battery lifetime. Typical drones are powered by on-board batteries, with limited
capacity. But drones are expected to carry out long missions. Thus, a fully automated management system that
can optimize the operations of battery-operated autonomous drones to extend their operation time is highly
desirable. This paper presents several contributions to automated management systems for battery-operated
drones: (1) We conduct empirical studies to model the battery performance of drones, considering various
flight scenarios. (2) We study a joint problem of flight mission planning and recharging optimization for drones
with an objective to complete a tour mission for a set of sites of interest in the shortest time. This problem
captures diverse applications of delivery and remote operations by drones. (3) We present algorithms for
solving the problem of flight mission planning and recharging optimization. We implemented our algorithms
in a drone management system, which supports real-time flight path tracking and re-computation in dynamic
environments. We evaluated the results of our algorithms using data from empirical studies. (4) To allow fully
autonomous recharging of drones, we also develop a robotic charging system prototype that can recharge
drones autonomously by our drone management system. Overall, we present a comprehensive study on flight
mission planning of battery-operated autonomous drones, considering autonomous recharging.
CCS Concepts: •Computer systems organization →Embedded and cyber-physical systems; •Computing
methodologies →Robotic planning; •Theory of computation →Dynamic graph algorithms;
Additional Key Words and Phrases: Autonomous Drones, Flight Mission Planning, Recharging Optimization,
Automated Drone Management
ACM Reference format:
Chien-Ming Tseng, Chi-Kin Chau, Khaled Elbassioni, and Majid Khonji. 2017. Autonomous Recharging and
Flight Mission Planning for Battery-operated Autonomous Drones. 1, 1, Article 1 (January 2017), 25 pages.
DOI: xx.xxxx/xxxxxxx.xxxxxxx
1
INTRODUCTION
Aerial vehicles are becoming a novel means of logistics. Often referred as drones in popular
terminology, or unmanned aerial vehicles in technical terminology, they have several advantages
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee
provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and
the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored.
Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires
prior specific permission and/or a fee. Request permissions from [email protected].
© 2017 ACM. XXXX-XXXX/2017/1-ART1 $15.00
DOI: xx.xxxx/xxxxxxx.xxxxxxx
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
1:2
Chien-Ming Tseng, Chi-Kin Chau, Khaled Elbassioni, and Majid Khonji
over ground based transportations. (1) Agility: There is little restriction in the sky, unlike on
the ground with obstacles. Drones can travel across space in straight paths. They are usually
small in size, with nimble navigating ability. (2) Swiftness: Aerial transportation is usually not
hampered by traffic congestions. The time to arrival is mostly reflected by the travelled distance.
Drones can also be rapidly launched by catapults, and drop off payloads by parachutes with short
response time. (3) Energy-efficiency: Drones are typically light-weight, which consume less energy.
They are particularly energy-efficient for transporting light-weight items in short trips, whereas
ground vehicles are useful for carrying heavier objects in long distance. (4) Safeness: There is no
on-board human operator or driver. Drones can keep a safe distance from human users. Unmanned
transportation missions are specially desirable in hazardous environments.
These advantages of drones enable diverse applications for light-weight goods transportation and
as substitutions of manned operations in remote locations. There are several notable applications of
drones. (1) Remote Surveillance: Aerial transportation can access far-away remote regions, and geographically dispersed offshore locations. Particularly, oil and gas companies and utility providers,
which rely on extensive surveillance, measurements, mapping and surveying, maintenance operations for dispersed facilities, will be major users of drones. (2) Search and Rescue: Drones can be
deployed in ad hoc manner. In emergency with damaged or unreliable infrastructure, drones can
overcome the difficulty of accessing in isolated regions, enabling fast transportation with great
convenience and flexibility. (3) Hazardous Missions: Drones are excellent solutions for unmanned
missions in risky or hazardous areas, in particular, for taking measurements in high-altitude, or
bio/chemical harmful environments. Human supervisors can remotely control the drones to carry
out dangerous operations. (4) Light-weight Items Delivery: Parcels, medical items, and mail require
speedy delivery. Drones are effective in solving the last-leg problem of the distribution chain from
depots to homes of end-users. A recent study by Amazon [13] reports that 44% of the US population
are within 20 miles from its depot facility. Hence, it is practical to employ drones for delivery.
Therefore, in the near future, fully autonomous drones are expected for extensive deployment,
giving rise to a new class of intelligent systems for logistics. Hence, the computing systems for
drones are becoming an exciting new venue for research in cyber-physical systems. Autonomous
drones require integrated intelligent decision systems to control and manage their flight missions
in the absence of human operators. Despite increasingly popular applications of drones in diverse
sectors, the operations of drones are plagued with several challenges:
• Limited Battery Lifetime: Typical drones are electric vehicles, powered by on-board batteries.
Hence, the performance of drones is critically constrained by limited battery lifetime. Many
drones are only suitable for short-range trips, which considerably limit their applicability.
To optimize the battery performance of drones, there requires an intelligent management
system to track the real-time state-of-charge, and optimize the operations accordingly.
• Dynamic Operating Environments: Drones are expected to travel in certain high altitude,
and hence are significantly susceptible by wind and weather conditions. These conditions
are highly dynamic, and should be accounted for in a real-time manner. Also, drones are
light-weight, and the impact by wind is even more substantial. Drone management system
should take explicit consideration of the dynamic uncertain operating environments.
These challenges present unique problems for battery-operated autonomous drones. Since drones
are expected to carry out long missions in dynamic environments, a fully automated management
system that can optimize the operations of battery-operated autonomous drones to extend their
operation time is highly desirable.
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
Autonomous Recharging and Flight Mission Planning for Battery-operated Autonomous Drones
1:3
1.1
Our Contributions
To improve the practical usefulness of autonomous drones, this paper presents several contributions
to automated management systems for battery-operated drones:
(1) We conduct empirical studies to model the power consumption of drones, considering
various flight scenarios. Accurate model of battery performance in different scenarios
allows further flight mission planning and recharging optimization for drones.
(2) We study a joint problem of flight mission planning and recharging optimization for drones,
using the calibrated power consumption model of a drone, with an objective to complete a
tour mission for a set of sites of interest in the shortest time. We also consider uncertainty in
dynamic environments, such as wind conditions. This problem captures diverse applications
of delivery and remote operations by drones.
(3) We present algorithms for solving the problem of flight mission planning and recharging
optimization. We implemented our algorithms in a drone management system, which
supports real-time flight path tracking and re-computation in dynamic environments. We
evaluated the results of our algorithms using data from empirical studies.
(4) To allow fully autonomous recharging of drones, we also develop a robotic charging station
prototype that can recharge drones autonomously by our drone management system.
Autonomous recharging can significantly extend the battery lifetime of drones for longdistance missions.
2
RELATED WORK
There are diverse applications for drones, including delivery (e.g., for light-weight parcels, medical
items, mail) and remote operations (e.g., wildlife surveillance, environmental surveying, search and
rescue operations). Drones had been often studied for their flight and landing control mechanisms.
For example, see the books [18, 19] for a good overview of the recent results. There are two
aspects of literature about drones: 1) low-level transient control of flight operations, for example,
controlling propellers and balance using PID controllers [3, 4], and 2) high-level planning and
management of drone missions, for example, obstacle avoidance, localization and mapping and
path planning [20]. However, the high-level studies typically focus on a single short-distance flight
paths. Long-distance flight mission planning involving multiple trips and recharging optimization
has been considered to a lesser extent, to the best of our knowledge. Moreover, there is a lack of
prior work on flight mission optimization particularly considering battery-operated drones.
Most drones are aerial electric vehicles. The prior studies of logistic optimization mainly focus
on ground electric vehicles, not on aerial electric vehicles. Nonetheless, drones exhibit different
characteristics that create some unique challenges. For example, the impact of wind is more
substantial for drone flight. In [12], graph signal sampling and recovery techniques are used to plan
routes for autonomous aerial vehicles, and a method is proposed to plan an energy-efficient flight
trajectory by considering the influences of wind. Further, there appears limited empirical studies of
battery performance of drones, although the empirical studies of ground electric vehicles have been
explored in the literature. Modelling and predicting electric vehicle power consumption has been
the subject of a number of research papers. One method is the model-based whitebox approach,
based on specific vehicle dynamics model to understand the consumption behavior of electric
vehicles [16]. The power consumption estimation can also be obtained by a blackbox approach. For
example, a general statistical approach using regression model, without vehicle dynamics model,
can estimate the power consumption of vehicles [5]. Blackbox model is more tractable and more
convenient for trip optimization. Hence, we will employ a similar blackbox model for aerial electric
vehicles, but taking into account the flight conditions.
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
1:4
Chien-Ming Tseng, Chi-Kin Chau, Khaled Elbassioni, and Majid Khonji
This work is related to the trip planning problem of electric vehicles [17]. There are recent
results for path planning of electric vehicles considering recharging operations [6]. We adopt the
solution proposed in [15] for the so-called tour gas station problem, for which efficient algorithms
are designed for obtaining a near-optimal solution under certain assumptions. A variant of the
classical algorithm [7] for the travelling salesman problem (TSP) was proposed in [15] for the
tour gas station problem. In this work, we extend those methods to solve the problem for drone
management by incorporating extensions to the settings of drone operations.
For fully autonomous drone management, drones should also be able to recharge themselves
without manual intervention. Inductive charging for drones has been proposed that can flexibly
recharge drones in an autonomous manner [2]. However, this work relies on a different solution,
with a combination of a robotic arm that can accommodate drone recharging in arbitrary positions.
To enable autonomous recharging of drones, an autonomous inductive charging system is initially
proposed in [14], which is integrated with the management system of this paper.
3
EMPIRICAL STUDIES OF BATTERY PERFORMANCE OF DRONES
In order to accurately optimize the power consumption and flight missions of drones, we first
conducted a series of empirical studies to determine the battery performance of drones, considering
various flight scenarios. In particular, we evaluate the power consumption using two commercial
drone models, 3DR Solo [1] and DJI Matrice 100 [8] (see Fig. 1 and their specifications in Table 1).
Both drones support developer kits, which allow us to extract data and program the flight paths.
After gathering sufficient measurement data, we can apply regression models to capture the power
consumption behavior of the drones.
Fig. 1. Left: 3DR Solo. Right: DJI Matrice 100.
Weight
Dimensions
Battery
Battery Weight
Motors
Max Speed
Max Altitude
Charging Duration
Software
3DR Solo
2 kg
25cm × 46cm
5200 mAh 14.8V
500 g
880 kV (×4)
20 km/h
122 ft (FAA Regulation)
90 mins
Python Developer Kit
DJI Matrice 100
2.8 kg
46cm × 46cm
5700 mAh 22.8V
600 g
350 kV (×4)
60 km/h
122 ft (FAA Regulation)
180 mins
DJI SDK & ROS
Table 1. Specifications of 3DR Solo and DJI Matrice 100 drones.
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
Autonomous Recharging and Flight Mission Planning for Battery-operated Autonomous Drones
1:5
3.1
Settings of Empirical Studies
A typical drone is equipped with a number of sensors for two main purposes: (1) for self-stabilizing
the drone in the air, and (2) for remotely tracking the drone status (e.g., the battery state-ofcharge (SoC)). The stability of a drone is controlled by three essential sensors (i.e., gyroscopes,
accelerometers and barometers), with which it can maneuver itself in the air. The SoC is measured
by the voltage and current sensors. A major part of power consumption of a drone is due to
the powering of motors to lift itself in the air. Additional power consumption is required for the
movements of the drone. The movements can be decomposed into vertical and horizontal directions.
The barometer and GPS sensors can measure the 3-dimensional movements of a drone. The speed
and position of a drone can be tracked by GPS and IMU modules, which also enable automatic
navigation. The altitude of a drone can be tracked by barometer and GPS modules.
To understand the factors that determine the power consumption of a drone, we carried out the
following experiments for obtaining empirical data in the rural areas, where the drone can fly in a
straight path without obstacles:
(1) Impact of Motion: The motions of a drone can be divided into three types: hovering,
horizontal moving and vertical moving. We study the power consumption of a test drone
in each motion type.
(2) Impact of Weight: Typical drones can carry extra payloads, such as camera equipment or
parcels. We study the impact of different weights of payloads attached to a test drone.
(3) Impact of Wind: The major environmental factor that affects the drone is wind, including
wind direction and speed. Wind may benefit the power consumption in some cases, as well
as incurring resistance to the movement in other cases. We study the power consumption
of a test drone in various wind conditions.
The experimental results are described as follows.
12
10
8
6
4
2
00
Experiment 1
Experiment 2
Altitude
Speed
50
Experiment 3
Power
480
400
320
240
160
80
4500
Power(W)
Speed (km/h)
Altitude (m)
3.1.1 Impact of Motion. To study the power consumption of motions of a drone, we conducted
three experiments. The battery power, barometer and GPS location, and speed data were collected
in each experiment to analyze the performance of test drone 3DR Solo.
100
150
200
250
Time (sec)
300
350
400
Fig. 2. Motion and battery power consumption of test drone 3DR Solo.
Fig. 2 depicts the recorded data traces of the three experiments of test drone 3DR Solo. We
discuss several observations as follows:
• Experiment 1: The test drone hovered in the air without any movement in this experiment.
Note that the drone may slightly drift around the takeoff location due to deviation error of
GPS modules. We filter the speed data that is smaller than 0.5 m/s. This experiment shows
the baseline power consumption of a flying drone. From the recorded data, we observe that
the drone can maintain a sufficiently steady flying altitude with steady power consumption.
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
1:6
Chien-Ming Tseng, Chi-Kin Chau, Khaled Elbassioni, and Majid Khonji
• Experiment 2: The test drone ascended and descended repeatedly in this experiment. The
barometer data shows the altitude of the drone. The time series data allow us to compute
the vertical acceleration and speed of the drone. We observe larger power fluctuations due
to repeatedly vertical movements. Power consumption increases slightly, when the drone
ascends steady.
• Experiment 3: The test drone moved horizontally without altering its altitude in this experiment. The GPS data comprises of speed and course angle of the drone. We also gathered
average wind speed and direction using a wind speed meter during the experiment. We
observe smaller power fluctuations due to horizontal movements. We also measure idle
power consumption of the drone between the two experiments.
25
20
15
10
5
00
Weight: 0g
Altitude
50
Weight: 250g
Speed
100
Weight: 500g
150
200
Time (sec)
500
400
300
200
100
0
Power(W)
Speed (km/h)
Altitude (m)
3.1.2 Impact of Weights. One of the practical purposes of drones is to deliver payloads, and
hence, the total weight of a drone varies as the payload it carries. We carried out several experiments
with different weights of payloads on the test drone 3DR Solo to obtain empirical data. Three
different weights were tested on the drone. The drone was set to hover in the air without any
movement to obtain the corresponding baseline power consumption.
250
300
Fig. 3. Battery power consumption of test drone 3DR Solo with different payload weights.
Fig. 3 depicts the battery power consumption of the test drone carrying three different weights.
We observe that power consumption increases almost linearly when the weight of payload increases.
The weight limit of payload depends on the thrusts that the motors can produce. Note that the
maximum payload weight is 500g for 3DR Solo.
3.1.3 Impact of Wind. Wind condition is a major environmental factor to affect the power
consumption of test drone 3DR Solo. We conducted several experiments under different wind
conditions: headwind by flying against the direction of wind, and tailwind by flying along the
direction of wind. The experiments were carried out at the same location but on different days with
different wind conditions. The wind directions and average speeds were measured using a wind
speed meter for each experiment. Once the wind direction was determined, the drone was set to fly
into a headwind or tailwind at maximum speed (18 km/h).
Fig. 4 depicts the battery power consumption of the drone under different wind conditions. We
observe smaller power consumption when flying into headwind, which is due to the increasing
thrust by translational lift, when the drone moves from hovering to forward flight. When flying
into a headwind, translational lift increases due to the relative airflow over the propellers increases,
resulting in less power consumption to hover the drone [11]. However, when the wind speed
exceeds a certain limit, the aerodynamic drag may outweigh the benefit of translational lift. In our
setting, the drone speed is relatively slow, even at maximum speed. Hence, flying into a headwind
is likely more energy-efficient.
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
Autonomous Recharging and Flight Mission Planning for Battery-operated Autonomous Drones
1:7
Power (W)
300
Headwind
Tailwind
250
200
1500
10
20
30
Time (sec)
40
Wind: 14.4 km/h
Wind: 18.7 km/h
Wind: 22.0 km/h
Wind: 28.8 km/h
50
60
Fig. 4. Battery power consumption of test drone 3DR Solo under different wind conditions.
3.2
Regression Model of Power Consumption for Drone
Since drones are aerial electric vehicles, we can apply the methodology from the literature of
general electric vehicles to model the power consumption of a drone. There are two main types of
power consumption models of a drone:
• White-box Model: A straightforward approach is to employ a white-box microscopic
behavior model for each drone that comprehensively characterizes the motor performance,
aerodynamic environment, and battery systems. However, such a white-box model requires
a large amount of data for calibration and detailed knowledge specific to a particular drone.
For example, the aerodynamic parameters such as propeller efficiencies, motor efficiencies
and drag coefficients are difficult to obtain accurately without resorting to sophisticated
experimental setups like wind tunnel.
• Blackbox Model: A blackbox approach is more desirable, because it requires minimal
knowledge of vehicle model with only a small set of measurable variables and parameters of
the drone. In the subsequent sections, a blackbox model of power consumption of a drone
will be utilized for flight mission planning and recharging optimization. The advantage of
blackbox model is that it is obtained from simple data measurements without relying on
sophisticated experimental setups.
This section describes a general multivariate blackbox model of power consumption for a drone
that has been used extensively in the literature of electric vehicles [5, 9, 10, 21, 22], which will be
verified in the later empirical studies.
Let the estimated battery power consumption of a drone be P̂, which is estimated by a number
of measurement parameters in the following linear equation:
β 1 T
P̂ = β 2
β 3
kv®xy k β 4 T
ka®xy k + β 5
kv®xy k ka®xy k β 6
kv®z k β 7 T
ka®z k + β 8
kv®z k ka®z k β 9
m
v®xy · w® xy
1
(1)
where
• v®xy and a®xy are the speed and acceleration vectors describing the horizontal movement of
the drone.
• v®z and a®z are the speed and acceleration vectors describing the vertical movement of the
drone.
• m is the weight of payload.
• w® xy is the vector of wind movement in the horizontal surface.
• β 1 , ..., β 9 are the coefficients, and kv®k denotes the magnitude of a vector.
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
1:8
Chien-Ming Tseng, Chi-Kin Chau, Khaled Elbassioni, and Majid Khonji
The coefficients β 1 , ..., β 9 can be estimated by the standard regression method, if sufficient measurement data is collected.
Assuming the uniform conditions (e.g., speed, wind) within a period of duration D, the total
energy consumption of the drone in duration D is estimated by P̂ · D.
3.3
Evaluation of Power Consumption Model
To evaluate the accuracy of the power consumption model, we conducted experiments to collect
extensive empirical data to estimate the corresponding coefficients. Two test drones (3DR Solo and
DJI Matrice 100) were used in two sets of experiments. A test drone was programmed to first fly
vertical movements, then flying into a headwind and a tailwind with different weights of payloads.
The drone maintained its altitude during the horizontal flight. We conducted experiments under
simple conditions, where the drone ascended from the source until reaching the desired altitude
and then flied directly to the destination without changing its altitude. But the experiments are
sufficiently representative of other conditions.
The following are the estimated coefficients of power consumption models for 3DR Solo and DJI
Matrice 100:
• 3DR Solo:
P̂solo
−1.526 T
= 3.934
0.968
kv®xy k 18.125 T
ka®xy k + 96.613
kv®xy k ka®xy k −1.085
kv®z k 0.220 T
ka®z k + 1.332
kv®z k ka®z k 433.9
m
v®xy · w® xy
1
(2)
• DJI Matrice 100:
P̂ dji
−2.595 T
= 0.116
0.824
kv®xy k 18.321 T
ka®xy k + 31.745
kv®xy k ka®xy k 13.282
kv®z k 0.197 T
ka®z k + 1.43
kv®z k ka®z k 251.7
m
v®xy · w® xy
1
(3)
We discussed the evaluation results of the test two drones using ground truth power consumption
data. Fig. 5-7 present the results for 3DR Solo, whereas Fig. 8-10 present the results for DJI Matrice
100. Fig. 5 and Fig. 8 depict the collected sensor data of our experiments for 3DR Solo and DJI
Matrice 100, respectively. We tested 3 different weights of payloads under similar flight paths
operations in each set of experiments. We obtain the estimated power consumption using the
respective regression model, and compare it to the ground truth power consumption data shown
in Fig. 6 and Fig. 9. We observe that the estimation is close to the actual measurement data. We
integrate power over time to obtain the power consumption of the drone in Fig. 7 and Fig. 10. The
errors of estimation of power consumption in the experiments are within 0.4%, showing relatively
good accuracy of our power consumption models for both test drones.
4
FLIGHT MISSION PLANNING AND RECHARGING OPTIMIZATION
In this section, we utilize the calibrated power consumption model of a drone from the last section
to study a joint problem of flight mission planning and recharging optimization for battery-operated
autonomous drones. The objective is to complete a flight tour mission for a set of sites of interest
in the shortest time. We consider a variable number of charging stations to allow recharging of
drones intermediately. This problem naturally captures diverse applications of delivery and remote
operations by drones. We provide efficient algorithms to determine the solutions, and implemented
our algorithms in an automated drone management system.
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
25
20
15
10
5
00
Weight: 0
Altitude
Weight: 250g
Speed
Weight: 500g
500
400
300
200
100
6000
Power(W)
Speed (km/h)
Altitude (m)
Autonomous Recharging and Flight Mission Planning for Battery-operated Autonomous Drones
1:9
100
200
300
400
500
Power (W)
Fig. 5. Sensor data for the evaluation experiment of test drone 3DR Solo.
500
400
300
200
100
00
Measurement
100
Estimation
200
300
400
500
600
Energy
consumption (Wh)
Fig. 6. Measured and estimated power consumption of test drone 3DR Solo.
40
30
20
10
00
Measurement
100
Estimation
200
300
Time (sec)
400
500
600
25
20
15
10
5
00
Weight: 0
Altitude
Weight: 300g
Speed
Weight: 600g
800
600
400
200
6000
Power(W)
Speed (km/h)
Altitude (m)
Fig. 7. Measured and estimated energy consumption of test drone 3DR Solo.
100
200
300
400
500
Power (W)
Fig. 8. Sensor data for the evaluation experiment of test drone DJI Matrice 100.
800
700
600
500
400
300
200
100
00
Measurement
100
Estimation
200
300
400
500
600
Energy
consumption (Wh)
Fig. 9. Measured and estimated power consumption of test drone DJI Matrice 100.
80
60
40
20
00
Measurement
100
Estimation
200
300
Time (sec)
400
500
600
Fig. 10. Measured and estimated energy consumption of test drone DJI Matrice 100.
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
1:10
4.1
Chien-Ming Tseng, Chi-Kin Chau, Khaled Elbassioni, and Majid Khonji
Model and Formulation
We denote a set of sites of interest by S that a drone needs to visit (e.g., drop-off locations of
parcels, or sites for measurements), and a set of charging station locations by C where a drone
can receive recharging. The base location of a drone is denoted by v 0 . Let V , S ∪ C ∪ {v 0 }. The
problem of drone flight mission planning with recharging is to find a flight mission plan (which is
a tour consisting of locations in S and C), such that the drone can visit all the sites in S, starting
and terminating at v 0 , with an objective of minimizing the total trip time, while maintaining the
state-of-charge (SoC) within the operational range. See an illustration of a flight mission plan with
recharging for a drone in Fig. 11.
Fig. 11. A flight mission plan with recharging for a drone.
Given a pair of locations (u, v), we denote the designated flight path by `(u, v), and the flight time
by τ (u, v). In this paper, we consider a simple flight path, such that the drone first ascends vertically
to a desired altitude, and then travels in a straight path, and finally descends to the destination
vertically. The model can
be generalized to consider non-straight paths.
Let E `(u, v), τ (u, v) be the required energy consumption for the drone flying along `(u, v)
within flight time τ (u, v). E(·, ·) is an increasing function that maps the combination of flight path
`(u, v) and flight time τ (u, v) to the required amount of energy. E(·, ·) can be estimated by a power
consumption model of a drone. We represent the charging strategy by a function b(·) : C 7→ R that
maps a charging station to an amount energy to be recharged. When recharging its battery at u,
let the incurred charging time be τc (b(u)). Let η c ≤ 1 and η d ≥ 1 be the charging and discharging
efficiency coefficients. If the drone flies to a charging station u ∈ C, it recharges its battery by an
amount of energy denoted by η cb(u). If the drone flies between two sites
u, v ∈ V , then it consumes
an amount of energy from the battery denoted by η d E `(u, v), τ (u, v) .
We denote a flight mission plan by F , which is a tour starting and terminating at v 0 , consisting
of a sequence of locations in S ∪ C ∪ {v 0 }. Denote k-th location by Fk . We require F1 = F| F | = v 0 .
The objective of flight mission planning is to find a flight mission plan F together with a charging
strategy b(·) that minimizes the total trip time, consisting of the flight time plus the charging time.
Let x k be the SoC when reaching the k-th location Fk in the flight mission plan. We require the
SoC to stay within feasible range [B, B]. The lower bound of SoC, B, ensures sufficient residual
energy for the drone to return to the base, in case of emergency. We set the initial SoC x 0 = B.
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
Autonomous Recharging and Flight Mission Planning for Battery-operated Autonomous Drones
1:11
With the above notations, the drone flight mission planning with recharging problem (DFP) is
mathematically formulated as follows.
(DFP)
min
F,b(·),x
|Õ
F |−1
τ (Fk , Fk +1 ) +
k =1
|F|
Õ
τc (b(Fk ))
(4)
k=1:Fk ∈ C
subject to
F1 = F| F | = v 0
(5)
S ⊆ F ⊆ S ∪ C ∪ {v 0 }
x k−1 − η d E `(Fk , Fk +1 ), τ (Fk , Fk +1 ) , if Fk ∈ S
xk =
x k−1 + η cb(Fk+1 ) − η d E `(Fk , Fk +1 ), τ (Fk , Fk +1 ) , if Fk ∈ C
B ≤ x k ≤ B, x 0 = B
(6)
(7)
(8)
The difficulty of DFP is to balance the flight decisions and charging decisions. On one hand, a
flight mission plan needs to consider the requirement of completing the mission in minimal total
trip time. On the other hand, it needs to be able to reach a charging station, in case of insufficient
battery, as well as minimizing the charging time.
The formulation of DFP can be extended to incorporate a variety of further factors for practical
flight mission plan optimization, such as restrictions of no-fly zones and attitude, and wind speed
forecast information. Users can also specify further goals, such as deadline of completion and
maximum payload weight. An efficient optimization algorithm is required to compute an optimal
flight mission plan to meet the users’ specified goals.
4.2
Case with Uniform Drone Speed and Steady Wind Condition
To provide efficient algorithms for DFP, we first consider a basic setting under some realistic
assumptions. Suppose that the horizontal speed of the drone is a uniform constant under steady
wind condition, which will be relaxed in Sec. 4.3. Then, the flight time τ (u, v) between two sites
u, v ∈ v is proportional to the length of flight path `(u, v), denoted by d(u, v). Our regression
model
of energy consumption for drone in Sec. 3 implies that the function E `(u, v), τ (u, v) is linear in
the distance d(u, v), and the charging time τc (b(u)) is linear in the amount of recharged energy
b(u). Thus, we assume the following linear objective functions:
τ (u, v) = c a d(u, v),
τc (b(u)) = cb b(u),
E `(u, v), τ (u, v) = c f (u, v) · d(u, v),
(9)
(10)
for some constants c a , cb , c f (u, v) > 0. Note that we allow c f (u, v) to be edge-dependent. This can
model non-uniform environment for each `(u, v), for instance, a path experiencing stronger wind
is expected to have a larger constant c f (u, v).
Denote the lower and upper bounds of c f by c f , min(u,v) c f (u, v) and c f , max(u,v) c f (u, v).
In this paper, we consider mostly long-distance trips (e.g., 2-3 km), for which the vertical landing
and take-off operations usually constitute a small part of the whole flight, and consume only a
small percentage of the total energy (e.g., < 1%). For clarity of presentation, we assume that the
energy consumption of landing and take-off operations is implicitly captured by c f (u, v) · d(u, v),
though our results can be easily extended to consider that explicitly.
Í
For convenience of notation, for a flight mission plan (F , b(·)), we write τ (F ) , k| F=1|−1 τ (Fk , Fk +1 )
Í|F |
Í | F |−1
and b(F ) , k =1:F ∈ C τc (b(Fk )). Also, define d(F ) , k=1 d(Fk , Fk +1 ).
k
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
1:12
Chien-Ming Tseng, Chi-Kin Chau, Khaled Elbassioni, and Majid Khonji
Under the aforementioned assumptions, the total charging time τc (b(F )), in an optimal flight
mission plan F , is proportional to the total flight time τ (F ), by the following lemma.
Lemma 4.1. In an optimal flight mission plan (F , b(·)), we have
c · d(F ) + c 0 ≤ τ (F ) + τc (b(F )) ≤ c · d(F ) + c 0
where either
1) c = c = c a and c 0 = 0, or
η
η
2) c = c a + c f cb ηdc , c = c a + c f cb ηdc , and c 0 =
cb
η c (B
− x 0 ).
Proof. See the Appendix.
Lemma 4.2. Given any feasible flight mission plan (F , b(·)), there is another feasible flight mission
plan (F , b 0(·)) such that
B − x0 c f ηd
τc (b(F )) ≤
+
· d(F )
ηc
ηc
Such a plan (F , b 0(·)) can be found in O(|V |) time.
Proof. See the Appendix.
Both Lemma 4.1 and Lemma 4.2 allow us to focus on minimizing the distance d(F ) instead of
total trip time. Hence, we simplify the problem DFP as a simplified formulation (SDFP), such that
its optimal solution is later shown to be within a constant factor with an optimal solution of DFP.
Simplified formulation (SDFP) is defined as follows.
(SDFP)
min
F,x
|Õ
F |−1
db Fk , Fk +1
(11)
k =1
subject to F1 = F| F | = v 0
S ⊆ F ⊆ S ∪ C ∪ {v 0 }
b k , Fk+1 ),
x
− η dd(F
x k = k −1
B, if Fk ∈ C
B ≤ x k ≤ B, x 0 = B
(12)
(13)
if Fk ∈ S
(14)
(15)
b ·), which is defined as follows. Recall
In SDFP, we consider a modified distance function d(·,
that V , S ∪ C ∪ {v 0 }. Consider a weighted undirected graph G 0 = (V , V2 ), whose edge lengths
b v)}u,v , which are the pairwise shortest
are defined by {c f (u, v) · d(u, v)}u,v . Then, obtain {d(u,
distances of each pair of nodes in G 0 . SDFP is related to the tour gas station problem in [15], which
optimizes a tour trip of a vehicle in minimal fuel cost, with options of refilling at given gas stations.
Note that we assume in the formulation of SDFP that the SoC is brought to its maximum at
each charging station. Once we obtain a tour under this assumption, it can be turned into a flight
mission plan with the minimal charging requirements using Lemma 4.2.
b v) be the distance to the nearest charging station from v, and
For u ∈ V , let dbu , minv ∈ C d(u,
b
su , argminv ∈ C d(u, v) be the nearest charging station from v.
B−B
Define U , ηd . Following [15], we make a mild assumption that for every u ∈ S \ {v 0 } there
is v ∈ C such that d(u, v) ≤ α U2 , where α ∈ [0, 1). This assumption can be justified (for α = 1) as
follows. For a location u ∈ S \ {v 0 }, if every v ∈ C is at distance greater than U2 , then it is infeasible
to visit u without incurring the battery level below B (as the SoC drops below B − η dU = B).
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
Autonomous Recharging and Flight Mission Planning for Battery-operated Autonomous Drones
1:13
In the following,
we present an algorithm to SDFP and then DFP. The main algorithm is
Find-plan V , d , which is a variant of Christofides Algorithm [7] for finding a tour for travelling salesman problem, based on the results in [15]. It finds a minimum spanning tree T , and then
a minimum weight perfect matching M on the odd vertices of T . The edges of T and M define an
Eulerian graph, from which an Eulerian tour F0 can be obtained in linear time. The Eulerian tour
is passed to the procedure Fix-plan for converting it to a feasible flight mission plan F , which
might use a non-optimal charging function b(·). Then, the resulting plan (F , b(·)) is further passed
to procedure Fix-charge for finding the minimal charging requirements
with respect to the flight
mission plan F . Specifically, the three procedures in Find-plan V , d are:
b u, v : This provides a lower bound for an optimal solution. Namely,
• Init-distances V , d,
e v), and the
it finds for every pair of locations u, v ∈ V , the minimum possible distance d(u,
corresponding shortest path P(u, v) to go from u to v without going out of the operational
b v) ≤ U − dbu − dbv then the drone can always go directly
range of the battery. Note that if d(u,
1
from u to v . Otherwise, at best (in an optimal solution), the drone can reach u with SoC
at most B − η ddbu , then it can visits a sequence of charging stations (only if the distance
db between two successive such stations is at most U ), then, form the last station, it has
to reach v such that the SoC at v is at least B + η ddbv (so that there is sufficient battery to
reach sv ). In particular, the distance from u to the first charging station on this path should
be at most U − dbu . Similarly, the distance from the last station on the path to v should be
at most U − dbv . This explains the definition of the graph G in line 5 of the procedure.
• Fix-plan G, F0 : starting from the flight mission plan F0 obtained using the (modified)
e this procedure reconstructs a feasible
Christofides algorithm with respect to the weights d,
flight mission plan F for problem (SDFP). It first replaces each edge (u, v) in the flight
mission plan by the corresponding path P(u, v). Since the resulting flight mission plan
maybe still infeasible, the procedure adds to every site a round trip to the closest charging
station. Finally, the added stations are dropped one by one in a greedy way as long as
feasibility ismaintained.
• Fix-charge F , b(·) : Starting from the flight mission plan (F , b(·)) constructed after calling
procedure Fix-plan G, F0 , this procedure finds a minimal amount of recharging energy,
according to Lemma 4.2.
Let OPTDFP and OPTSDFP be the optimal solutions of problems (DFP) and (SDFP), respectively.
Lemma 4.3 ([15]). The flight mission plan F returned by algorithm Find-plan V , d] has cost
b ) ≤ 3 1+α OPTSDFP .
d(F
2 1−α
The following theorem establishes that algorithm Find-plan V , d] has an asymptotic constantfactor approximation guarantee for DFP.
Theorem 4.4. The flight mission plan (F , b 0(·)) returned by algorithm Find-plan V , d] has cost
τ (F ) + τc (b(F )) = O(OPTDFP ) + O(1).
Proof. See the Appendix.
is, starting with SoC= B at su , then the drone reaches u with SoC B − η d dbu , and then it flies directly from u to v
causing the SoC to drop to B − η d (dbu + db(u, v)) = B + η d (U − dbu − db(u, v)) ≥ B + η d dbv at v. Thus, there is sufficient
battery at v to reach sv .
1 That
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
1:14
Chien-Ming Tseng, Chi-Kin Chau, Khaled Elbassioni, and Majid Khonji
Algorithm 1 Find-plan V , d
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
b v)}u,v on weighted undirected graph G 0 = (V ,
Compute pairwise shortest distances {d(u,
for each u, v ∈ V do
e v), P(u, v)) ← Init-distances V , d,
b u, v
(d(u,
end for
e where E = V
Consider the weighted undirected graph G = (V , E; d)
2
Find a minimum spanning tree T = (V , ET ) in G
V0 ← find the set of odd degree vertices in T
e
Find a minimum-weight perfect matching M = (V0 , E M ) in the graph (V0 , E; d)
F0 ← find an Eulerian tour in the graph (V , ET ∪ E M )
F ← Fix-plan[G, F0 ]
b 0(·) ← Fix-charge[F , b(·)]
return (F , b 0(·))
V
2 )
b u, v
Algorithm 2 Init-distances V , d,
1:
2:
3:
4:
5:
6:
7:
8:
9:
b v) ≤ U − dbu − dbv then
if d(u,
e v) ← d(u,
b v), P(u, v) ← {(u, v)}
d(u,
e v), P(u, v))
return (d(u,
else
Construct a weighted undirected graph G = (C ∪ {u, v}, E; w) where
b z) ≤ U − dbu Ð {v, z} : z ∈ C, d(v,
b z) ≤ U − dv0 Ð
E , {u, z} : z ∈ C, d(u,
b z 0) ≤ U
{z, z 0 } : z, z 0 ∈ C, d(z,
0
0
b z ) for all z, z 0 ∈ C ∪ {u, v}
and w(z, z ) , d(z,
P(u, v) ← shortest path between u and v in G (with a set of edge lengths {w(u, v)}u,v )
e v) ← length of P(u, v)
d(u,
e v), P(u, v))
return (d(u,
end if
Algorithm 3 Fix-plan G, F0
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
F ←∅
for each (u, v) in F0 do
Add P(u, v) to F
end for
Add to F a set of sub-tours {{(u, su ), (su , u) : u ∈ V }
for u ∈ V do
if F \ {(u, su ), (su , u) is feasible then
F ← F \ {(u, su ), (su , u)}
end if
end for
return F
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
Autonomous Recharging and Flight Mission Planning for Battery-operated Autonomous Drones
1:15
Algorithm 4 Fix-charge F , b(·)
Let Fi 1 , . . . , Fi r be the charging stations, in the order they appear on F
for j = 0, 1, . . . , r do
i j+1
Õ−1
3:
D j = ηd
c f (Fk , Fk +1 )d(Fk , Fk +1 )
1:
2:
k =i j
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
4.3
end for
for j = 1, . . . , r do
Íj
B j , η c k=1
b(Fi k )
end for
for j = r downto 1 do
j−1
r
Õ
Õ
1
Dk −
Bk )}
b 0(Fi j ) = max{0, (B − x 0 +
ηc
k =0
k =1
if b 0(Fi j ) > 0 then
exit
end if
end for
return b 0(·)
Extensions
The preceding section presents a basic setting of DFP and its efficient algorithms. In reality, an
automated drone management system requires more sophisticated options. In this section, we
present two extensions to the preceding algorithms to obtain heuristics for more practical scenarios.
4.3.1 Wind Uncertainty. Under steady wind condition, we assume in the preceding algorithms
that c f (u, v) is a constant that depends on the designated path between sites u and v. In practice,
there is sometimes uncertainty in the wind condition. Often, the wind varies as the drone flies.
This also depends on the expected wind condition on this path. Thus, it should be more precisely
represented by c f (u, v, w), where w is the wind vector whose value is in an uncertain domain
w ∈ W . For example, W is defined by the predicted speed range [|w |, |w |] and the predicted
orientation range [θ w , θ w ]. We can modify the algorithms to account for the uncertainty of
W . We proceed conservatively in our algorithm by taking the worst-case, replacing c f (u, v) by
c f (u, v) = maxw ∈W c f (u, v, w).
4.3.2 Variable Drone Speed. We consider another scenario, in which the drone can vary its
speed uniformly at all designated paths in V . In this case, we run our algorithms sequentially in
multiple rounds, with an increasing drone speed at each round, until the algorithms can not return
a feasible solution (because that higher drone speed may result in insufficient battery to reach some
sites). Then we will enumerate all the optimal solutions in all the rounds to find the best solution
with the lowest total flight time. By enumerating the possibilities of different drone speeds, the
algorithms can identify an optimal flight mission plan.
5
CASE STUDIES
We implemented the algorithms in an automated drone management system. In particular, we
evaluated the results of flight mission planning and recharging optimization for the test drones in
several case studies, based on the data from empirical studies.
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
1:16
5.1
Chien-Ming Tseng, Chi-Kin Chau, Khaled Elbassioni, and Majid Khonji
Setup
We consider a scenario with four sites of interest, and four charging stations. The drones are
programmed to begin its mission from the base. Fig. 12 depicts the geographical locations of the
sites (as black points), charging stations (as blue squares) and the base (as magenta triangle). The
choices of geographical locations and distances are based on some real locations of a suburban
community.
3.83
Case
1
2
3
2.3
8
4
5
6
3.40
Charging stations 7
8
1.7
1
Uint: km
Base
Sites
9
2.60
2.3
2.78
2.62
Fig. 12. Geographical locations of the sites, charging stations and base.
Drone Battery (Wh) kw® xy k (20 km/h)
Solo
70
South
Solo
70
North-East
Solo
140
South
Solo
140
North-East
DJI
130
South
DJI
130
North-East
DJI
260
South
DJI
260
North-East
m (g)
0
0
500
500
0
0
600
600
Fig. 13. Parameters of setup.
There are two major sets of studies conducted as follows.
(1) Study 1: We study eight sub-cases using the power consumption models of 3DR Solo and
DJI Matrice 100 under different wind and payload conditions. For each drone, we study
4 sub-cases as follows. We consider using one battery in the first two sub-cases of each
drone. Different wind conditions with average wind speed of 18 km/h are studied in the
sub-cases. Then we double the battery capacity with the same wind condition in another
two sub-cases. Since the battery capacity is doubled, extra weight is added to the drone.
The parameters of all the sub-cases are summarized in Table 13.
(2) Study 2: We consider uncertainty of wind conditions. The wind speed and orientation vary
within a certain range. The wind speed varies from 0 to 21 km/h in four discrete scales,
while the wind orientation varies from 0◦ to 360◦ in four discrete scales.
In Study 1, the cases of 3DR Solo are denoted by S1 C1 to S1 C4 , and the cases of DJI Matrice 100 are
denoted by S1 C5 to S1 C8 . Similarly, in Study 2, the cases of 3DR Solo are denoted by S2 C1 to S2 C4 ,
and the cases of DJI Matrice 100 are denoted by S2 C5 to S2 C8 .
5.2
Results and Discussion
For comparison, we also consider a benchmark algorithm, by which a drone flies to the nearest
unvisited site, or the SoC drops below a preset threshold, then the drone flies to a charging station
instead. We set the preset threshold to be the minimum SoC that can fly to a nearest charging
station from any site.
5.2.1 Study 1. The results of flight missions of Study 1 are visualized in Figs. 16-19. The numbers
indicate the path order of the drone. The colors represent the SoC of battery. The wind orientations
are displayed on the upper-left corners. We plot the trip time and energy consumption of Study 1
in Figs. 14-15. There are several interesting observations as follows:
• Our algorithm significantly outperforms the benchmark algorithm, in terms of trip time
and energy consumption. Hence, our algorithms are superior for flight mission planning.
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
Autonomous Recharging and Flight Mission Planning for Battery-operated Autonomous Drones
1:17
• In the case study, the north-east wind affects flight missions to a larger extent, which causes
a higher energy consumption than that by south wind. Besides, there is a longer trip time
due to longer charging time.
• We observe that even the trip times in S1 C7 and S1 C8 are shorter, it consumes more energy
since the drone carrying extra battery, which results in heavier loads.
• Attaching one more battery does not help to reduce the trip time for 3DR Solo, while
attaching more battery helps to reduce trip time for DJI Matrice 100. The reason is because
attaching more battery enables DJI Matrice 100 to fly longer without recharging. The total
trip time is significantly decreased due to much shorter charging time (as the blue bars of
S1 C7 and S1 C8 are shorter). But the same path is not feasible for 3DR Solo. 3DR Solo will
require to charge at the left most charging station, and hence, the trip time increases. If we
slightly increase the battery capacity to 75 Wh, the results of 3DR Solo will be the same as
that of DJI Matrice 100.
400
300
300
200
200
100
0
100
S1 C1
S1 C2
S1 C3
S1 C4
0
Flight time
Benchmark time
Benchmark energy consumption
600
600
500
500
400
400
300
300
200
200
100
100
0
S1 C5
S1 C6
S1 C7
S1 C8
0
Energy consumption (Wh)
Trip time (mins)
400
Charging time
Energy consumption
Trip time (mins)
Flight time
Benchmark time
Benchmark energy consumption
Energy consumption (Wh)
Charging time
Energy consumption
Fig. 14. Trip time and energy consumption of Study 1 Fig. 15. Trip time and energy consumption of Study 1
using 3DR Solo.
using DJI Matrice 100.
5.2.2 Study 2. We study the results of flight mission planning considering uncertainty of wind
condition. We consider two sub-cases with the shortest trip time of two different drones from Study
1 (i.e., S1 C1 and S1 C7 ) and then increase uncertainty level for each case.
The results of flight missions of Study 2 are visualized in Figs. 20-21. We represent the ranges of
wind speeds and orientations as the shaded areas on the upper-left circles. We plot the trip time
and energy consumption of Study 2 in Figs. 22-23. There are several observations as follows:
• Recall that the energy consumption of S1 C1 does not have uncertain wind condition for 3DR
Solo with 70 Wh battery. In Study 2, we investigate if a feasible solution can be obtained in
the presence of uncertainty of wind condition. We gradually increase the uncertainty from
S2 C1 to S2 C4 . For example, in S2 C1 the wind speed varies from 9 to 12 km/h and orientation
varies from -45◦ to 45◦ . In this case, our algorithms always consider the worst-case setting
in the given range of wind conditions.
• We observe that the energy consumption in general increases as the uncertainty of wind
condition increases in Fig. 22. In partcular, the energy consumption of the worst uncertainty
is in S2 C4 , in which the drone may always fly into a tailwind. Thus, S2 C4 provides the most
conservative result.
• Similarly, recall that the energy consumption of S1 C7 does not have uncertain wind condition
for DJI Matrice 100 with 260 Wh battery (two batteries). We also observe similar trends when
the wind uncertainty is increased. In general, when uncertainty increases, the difference
of trip time/energy consumption between our approach and benchmark becomes smaller.
This is because of more frequent recharging when the uncertainty becomes larger.
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
1:18
Chien-Ming Tseng, Chi-Kin Chau, Khaled Elbassioni, and Majid Khonji
1
Wind orientation
SoC Wind orientation
100
2
SoC
100
10
11
1
5
3
50
9
6
6
5
4
2
7
7
4
3
8
Base
Sites
10
9
0
Charging stations
0
(a) Case 1.
(b) Case 2.
Wind orientation
SoC Wind orientation
100
1
3
SoC
100
9
1
3
2
4
50
8
5
9
2
4
8
5
7
6
50
8
7
6
0
(c) Case 3.
50
0
(d) Case 4.
Fig. 16. Visualized flight missions for Study 1 using 3DR Solo by our algorithm.
Wind orientation
SoC Wind orientation
100
3
2
SoC
100
1
1
2
3
4
10
4
6
5
50
11
5
9
7
7
6
8
50
10
8
9
0
0
(a) Case 1 benchmark.
Wind orientation
(b) Case 2 benchmark.
SoC Wind orientation
100
3
2
10
4
5
6
7
SoC
100
3
1
50
9
4
9
5
8
1
2
6
50
8
7
0
(c) Case 3 benchmark.
0
(d) Case 4 benchmark.
Fig. 17. Visualized flight missions for Study 1 using 3DR Solo by benchmark algorithm.
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
Autonomous Recharging and Flight Mission Planning for Battery-operated Autonomous Drones
1:19
1
Wind orientation
SoC Wind orientation
100
2
SoC
100
10
11
1
5
3
9
6
50
6
5
4
2
7
7
10
50
8
4
3
8
9
0
0
(a) Case 5.
(b) Case 6.
Wind orientation
SoC Wind orientation
100
SoC
100
1
1
2
2
7
50
3
4
50
4
6
5
7
3
6
5
0
(c) Case 7.
0
(d) Case 8.
Fig. 18. Visualized flight missions for Study 1 using DJI Matrice 100 by our algorithm.
Wind orientation
SoC Wind orientation
100
3
2
SoC
100
1
1
2
3
4
10
4
5
6
50
11
5
9
7
7
6
8
50
10
8
9
0
0
(a) Case 5 benchmark.
Wind orientation
(b) Case 6 benchmark.
SoC Wind orientation
100
SoC
100
3
1
1
2
8
6
50
7
2
5
9
4
3
4
5
6
50
8
7
0
(c) Case 7 benchmark.
0
(d) Case 8 benchmark.
Fig. 19. Visualized flight missions for Study 1 using Matrice 100 by benchmark algorithm.
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
1:20
Chien-Ming Tseng, Chi-Kin Chau, Khaled Elbassioni, and Majid Khonji
Wind uncertainty
SoC Wind uncertainty
100
2
4
1
10
2
50
3
5
4
6
8
7
7
11
50
10
8
9
0
0
(a) Case 1.
(b) Case 2.
Wind uncertainty
SoC Wind uncertainty
100
3
4
2
1
2
1
12
50
5
6
11
9
3
4
50
5
8
SoC
100
12
6
7
1
3
5
9
6
SoC
100
11
7
10
8
9
10
0
0
(c) Case 3.
(d) Case 4.
Fig. 20. Visualized flight missions for Study 2 using 3DR Solo.
Wind uncertainty
SoC Wind uncertainty
100
9
1
3
3
0
(b) Case 6.
Wind uncertainty
SoC Wind uncertainty
100
2
4
7
6
0
(a) Case 5.
1
7
2
50
0
1
10
50
3
9
6
8
(c) Case 7.
4
5
9
6
SoC
100
10
3
5
8
5
7
6
50
2
4
8
5
9
1
50
2
4
SoC
100
7
8
(d) Case 8.
Fig. 21. Visualized flight missions for Study 2 using DJI Matrice 100.
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
0
Autonomous Recharging and Flight Mission Planning for Battery-operated Autonomous Drones
1:21
400
300
300
200
200
100
0
100
S1 C1
S2 C1
S2 C2
S2 C3
S2 C4
0
Flight time
Benchmark time
Benchmark energy consumption
800
800
700
700
600
600
500
500
400
400
300
300
200
200
100
0
100
S1 C7
S2 C5
S2 C6
S2 C7
S2 C8
0
Energy consumption (Wh)
Trip time (mins)
400
Charging time
Energy consumption
Trip time (mins)
Flight time
Benchmark time
Benchmark energy consumption
Energy consumption (Wh)
Charging time
Energy consumption
Fig. 22. Trip time and energy consumption of Study Fig. 23. Trip time and energy consumption of Study
2 using 3DR Solo.
2 using DJI Matrice 100.
6
AUTOMATED DRONE MANAGEMENT SYSTEM
We implemented our algorithms in an automated drone management system. The user interface is
depicted in Fig 24. The user interface allows the users to specify individual goals and to visualize
the computed flight mission plan. The system connects to a cloud computing server, which uses
the input locations from the user and computes the optimal flight mission plan. Then the drone is
programmed to follow the flight mission plan computed by the server.
Furthermore, a dynamic tracking system of drone using on-board sensors, including GPS location,
video feed, SoC, and flight status, is utilized to monitor the real-time flight status of the drone. If
abnormal measurements are detected, for example, the reported sensor measurements deviate from
estimated value by flight mission planning, then real-time re-computation will be performed to
find the minimum adjustment to the previously computed flight mission plan. The user can abort
the mission anytime by clicking the button on the system. We remark that multiple drones can
also be tracked simultaneously.
Fig. 24. Interface of automated drone management
system.
7
Fig. 25. Robotic charging system.
AUTONOMOUS ROBOTIC CHARGING SYSTEM PROTOTYPE
In this section, we present a robotic charging system prototype that can recharge drones autonomously by our drone management system. More technical details can be found in [14].
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
1:22
Chien-Ming Tseng, Chi-Kin Chau, Khaled Elbassioni, and Majid Khonji
There is a tethered rover capable of autonomous navigation, equipped with a robotic arm carrying
a charging pad that can adapt to different drone sizes and landing positions (see Fig. 25). Our
robotic rover uses 2D Light Detection and Ranging (LIDAR) sensors to navigate and detect a drone.
LIDAR sensors are more robust to environmental uncertainties and external lighting conditions
than computer vision sensors. The robotic charging system is comprised of three main components:
• Charging system: Since the goal of the charging system is to recharge drones in remote
areas, disconnected from the grid, it is vital to provide self-generating charging system, as
well as to protect the robotic rover from harsh environment conditions (i.e., rain, the wind,
etc.) when idle. So, our prototype can be powered by a solar panel. A landing zone with an
area of at least 3 × 3 m2 is designated for drone landing in front of the charging system.
• Robotic rover: To enable flexible deployment, a tethered rover is equipped with a robotic
arm that mounts a wireless inductive transmitter. Using autonomous drone detection and
navigation, the rover can cater for arbitrary landing position and orientation of a drone,
which is the case for most commercial drones that solely rely on inaccurate GPS and Inertial
Measurement Unit (IMU) for landing.
• Drones: Each drone is equipped with an inductive wireless receiver, compatible with that
of the robotic rover. In order to recharge a drone, the drone management system will send
a signal to the charging system to initiate the charging procedure.
We use Hokuyo UTM-30LX Scanning Laser Rangefinder to obtain 2D point cloud. The laser
rangefinder provides 30 meters detection range which allows the rover to detect far objects such as
drones or obstacles. The detection angle is 270◦ with 0.25◦ angular resolution. The rover odometry
is obtained based on wheel rotation using quadrature encoders.
A retractable arm is used to conduct inductive charging for the drones, which can flexibly
recharge drones in arbitrary positions. The inductive charging technology is based on resonant
inductive coupling, which can transfer energy without physical contact. It has been widely used
in wireless charging for smart phones. The use of inductive charging technology facilitates a
fully automated drone management system, without manually plugging to an external charger.
The current prototype uses six coils for inductive charging, giving a maximum charging speed of
6 × 700 = 3500 mA. We mounted six current sensors, on the robotic arm, to measure individual
coil charging rate. Our test drone is DJI Matrice 100 which is powered by two 5700 mA on-board
Li-Poly battery.
7.1
Recharging Process
We next describe the operations of recharging process. Once a drone is landed and a recharging
command is initiated, the robotic rover proceeds as follows:
(1) Finding the drone: From raw 2D LIDAR point cloud, we use a clustering technique to
separate the points of the drone from the background points.
(2) Classifying the drone and detecting its orientation: Based on supervised learning, drone
information (e.g., type, size, orientation, etc.) can be extracted from point cloud data. We
consider a single drone, and its orientation is detected using a rectangle boundary model.
The orientation of the drone is needed to instruct the rover to navigate to a position that
supports a better alignment for the charging pad with the receiver on the drone. This allows
more flexibility in the charging pad design (e.g., rectangular inductive charging pad instead
of a circular pad) to fit more inductive coils on the drone.
(3) Autonomously navigating towards the drone: the exact position of the navigation goal is
determined based on the allowable charging position.
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
Autonomous Recharging and Flight Mission Planning for Battery-operated Autonomous Drones
1:23
(4) Recharging: Position the charging pad on the drone and adjust the positioning using current
sensors readings.
Once a drone is fully charged, a termination command is initiated to the charging system. Then
the rover returns to the charging system to resume its idle state. Also, the drone management
system actively communicates with the robotic charging system to track the charging status.
8
CONCLUSION
Automated drone management system is important for practical applications of drones. This paper
provides multiple contributions to automated management systems for battery-operated drones,
including empirical studies to model the battery performance of drones considering various flight
scenarios, a study of flight mission planning and recharging optimization for drones that captures
diverse applications of delivery and remote operations by drones, and a management system
implementation with a robotic charging station to support autonomous recharging of drones.
In future work, we will incorporate a variety of further features in our automated drone management system, such as restrictions of no-fly zones and attitude, and wind speed forecast. Users may
also be able to specify further goals, such as deadline of completion and maximum payload weight.
A
PROOFS
Lemma 4.1. In an optimal flight mission plan (F , b(·)), we have
c · d(F ) + c 0 ≤ τ (F ) + τc (b(F )) ≤ c · d(F ) + c 0
where either
1) c = c = c a and c 0 = 0, or
η
η
2) c = c a + c f cb ηdc , c = c a + c f cb ηdc , and c 0 =
cb
η c (B
− x 0 ).
Proof. Consider an optimal flight plan (F , b(·)) and assume that the charging stations, in the
order they appear on F , is Fi 1 , . . . , Fi r , where without loss of generality, we assume Fi 1 , v 0 .
For completeness, let i 0 , 1 and i r +1 , |F |. For j = 0, 1, . . . , r , let
D j , ηd
i jÕ
+1 −1
c f (Fk , Fk +1 ) · d(Fk , Fk +1 ),
k =i j
and for j = 1, . . . , r , let B j , η cb(Fi j ).
Then, the feasibility of the flight mission plan F implies
x0 −
j
Õ
k =0
Dk +
j
Õ
Bk ≥ B, for j = 0, . . . , r
(16)
k=1
Let us refer to Ineq. (16) for a particular j as I (j) ≥ B. Particularly, consider I (r ) ≥ B. Suppose
that this inequality is not tight, that is, the left-hand side is strictly larger than the right-hand
side. Note that the variable b(Fi r ) = Bηrc appears only in this inequality. Since b(Fi r ) appears in the
objective function τc (b(F )) with a positive coefficient (i.e., τc (b(u)) = cb b(u)), there are two cases:
(i) b(Fi r ) = 0 at optimality, if I (r ) > B, or (ii) b(Fi r ) > 0 at optimality, if I (r ) = B. Otherwise, it
will contradict to the optimality of b(Fi r ), by reducing the value of b(Fi r ). If it is case (i), then the
inequality I (r − 1) ≥ B becomes redundant (as I (r − 1) ≥ I (r ) > 0). Removing I (r − 1) ≥ B, we obtain
that the variable b(Fi r −1 ) appears only in I (r ) ≥ B. Similarly, we can conclude that b(Fi r −1 ) = 0 and
remove the (now) redundant inequality I (r − 2) ≥ B.
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
1:24
Chien-Ming Tseng, Chi-Kin Chau, Khaled Elbassioni, and Majid Khonji
Continuing this argument, we conclude that there are two cases: (1) either all variables b(Fi j )
are set to zero in which case the value of the objective is τ (F ) = c a d(F ), or (2) we have
r
r
Õ
Õ
x0 −
Dk +
Bk = B,
k =0
k =1
In case (2), the value of the objective is
r
r
Õ
cb
cb Õ
Bk = τ (F ) + (B − x 0 +
Dk )
τ (F ) +
ηc
ηc
k =1
r
cb Õ
cb
= τ (F ) +
D k + (B − x 0 )
ηc
ηc
k =0
k =0
Therefore,
r
ηd
cb
cb Õ
cb
d(F ) + (B − x 0 ) ≤τ (F ) +
D k + (B − x 0 )
c a + c f cb
ηc
ηc
ηc
ηc
k =0
ηd
cb
≤ c a + c f cb
d(F ) + (B − x 0 ).
ηc
ηc
Lemma 4.2. Given any feasible flight mission plan (F , b(·)), there is another feasible flight mission
plan (F , b 0(·)) such that
B − x0 c f ηd
τc (b(F )) ≤
+
· d(F )
ηc
ηc
Such a plan (F , b 0(·)) can be found in O(|V |) time.
Proof. The proof follows from Lemma 4.1, which uses algorithm Fix-charge to implement the
argument used in Lemma 4.1 to find a feasible fight mission plan.
0
Theorem 4.4. The flight plan (F , b (·)) returned by algorithm Find-plan V , d] has cost
τ (F ) + τc (b(F )) = O(OPTDFP ) + O(1).
Proof. Let (F ∗ , b ∗ (·)) be an optimal flight plan for (DFP). Clearly, this plan can be trivially
turned into a feasible solution (F ∗ , x) for (SDFP) by setting x k = B for all Fk ∈ C. It follows that
b ∗ ).
OPTSDFP ≤ d(F
(17)
On the other hand, Lemma 4.3 implies that
3 1+α
b
d(F ) ≤
OPTSDFP .
2 1−α
(18)
OPTDFP ≥ c · d(F ∗ ) + c 0,
(19)
Lemma 4.1 also implies that
while Lemma 4.2 implies that
b(F ) ≤
B − x0 c f ηd
+
d(F ),
ηc
ηc
(20)
b ·) implies
and the definition of d(·,
b ) and d(F
b ∗ ) ≤ c f d(F ∗ ).
c f d(F ) ≤ d(F
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
(21)
Autonomous Recharging and Flight Mission Planning for Battery-operated Autonomous Drones
1:25
Putting together (17), (18), (19), (20), and (21), we obtain
τ (F ) + τc (b(F )) = c a d(F ) + cb b(F )
B − x0
ηd
d(F ) + cb
≤ c a + c f cb
ηc
ηc
B − x0
η d OPTDFP − c 0
3 1 + α cf
+ cb
≤
.
c a + c f cb
2 1 − α cf
ηc
c
ηc
REFERENCES
[1] 3DR Solo. 2017. https://3dr.com/solo-drone/specs/. (2017). [Online; accessed 19-Sept-2017].
How wireless power charging could recharge a flying drone.
http://www.bbc.com/news/
[2] BBC. 2016.
technology-37821459
[3] S. Bouabdallah, A. Noth, and R. Siegwart. 2004. PID vs LQ control techniques applied to an indoor micro quadrotor. In
International Conference on Intelligent Robots and Systems (IROS).
[4] Endri Bregu, Nicola Casamassima, Luca Mottola, and Kamin Whitehouse. 2016. Reactive Control of Autonomous
Drones. In ACM Mobile Systems, Applications, and Services (Mobisys).
[5] Alesaandn Cappiello, Ismail Chabini, Edward K. Nam, Alessandro Lue, and Maya Abou Zed. 2002. A Statistical Model
of Vehicle Emissions and Fuel Consumption. In IEEE Intelligent Transportation Systems Conf.
[6] Chi-Kin Chau, Khaled Elbassioni, and Chien-Ming Tseng. 2017. Drive Mode Optimization and Tour Planning for
Plug-in Hybrid Electric Vehicles. to be published in IEEE Trans. Intell. Transp. Syst. (2017).
[7] Nicos Christofides. 1976. Worst-case analysis of a new heuristic for the travelling salesman problem. Technical Report.
CMU. Report 388, Graduate School of Industrial Administration.
[8] DJI Matrice 100. 2017. https://www.dji.com/matrice100/info. (2017). [Online; accessed 19-Sept-2017].
[9] Raghu K Ganti, Nam Pham, Hossein Ahmadi, Saurabh Nangia, and Tarek F Abdelzaher. 2010. GreenGPS: a participatory
sensing fuel-efficient maps application. In ACM Mobile Systems, Applications, and Services (Mobisys).
[10] Stefan Grubwinkler and Markus Lienkamp. 2013. A modular and dynamic approach to predict the energy consumption
of electric vehicles. In Conf. Future Automotive Technology.
[11] Helicopter Flying Handbook. 2012. . US Federal Aviation Administration.
[12] Tianxi Ji, Siheng Chen, Rohan Varma, and Jelena Kovacevic. 2015. Energy-efficient route planning for autonomous aerial
vehicles based on graph signal recovery. In Annual Allerton Conference on Communication, Control, and Computing.
[13] The Wall Street Journal. 2016. Amazon’s Newest Ambition: Competing Directly With UPS and FedEx. http://www.wsj.
com/articles/amazons-newest-ambitioncompeting-directly-with-ups-and-fedex-1474994758
[14] Majid Khonji, Mohammed Alshehhi, Chien-Ming Tseng, and Chi-Kin Chau. 2017. Autonomous Inductive Charging
System for Battery-operated Electric Drones. In ACM Workshop on Electric Vehicle Systems, Data and Applications
(EV-Sys).
[15] Samir Khuller, Azarakhsh Malekian, , and Julin Mestre. 2011. To fill or not to fill: The gas station problem. ACM
Transactions on Algorithm 7 (2011), 534–545.
[16] Eugene Kim, Jinkyu Lee, and Kang G. Shin. 2013. Real-time prediction of battery power requirements for electric
vehicles. In IEEE/ACM Intl. Conf. on Cyber-Physical Systems.
[17] Mingsong Lv, Nan Guan, Ye Ma, Dong Ji, Erwin Knippel, Xue Liu, and Wang Yi. 2016. Speed Planning for Solar-Powered
Electric Vehicles. In ACM Intl. Conf. on Future Energy Systems (e-Energy).
[18] Kenzo Nonami, Farid Kendoul, Satoshi Suzuki, Wei Wang, and Daisuke Nakazawa (Eds.). 2010. Autonomous Flying
Robots: Unmanned Aerial Vehicles and Micro Aerial Vehicles. Springer Publisher.
[19] Yasmina Bestaoui Sebbane (Ed.). 2015. Smart Autonomous Aircraft: Flight Control and Planning for UAV. CRC Press.
[20] Roland Siegwart, Illah Nourbakhsh, and Davide Scaramuzza. 2015. Introduction to Autonomous Mobile Robots. MIT
Press.
[21] Chien-Ming Tseng and Chi-Kin Chau. 2017. Personalized Prediction of Vehicle Energy Consumption based on
Participatory Sensing. to be published in IEEE Trans. Intell. Transp. Syst. (2017).
[22] Chien-Ming Tseng, Chi-Kin Chau, Sohan Dsouza, and Erik Wilhelm. 2015. A Participatory Sensing Approach for
Personalized Distance-to-Empty Prediction and Green Telematics. In ACM Intl. Conf. Future Energy Systems (e-Energy).
, Vol. 1, No. 1, Article 1. Publication date: January 2017.
| 8cs.DS
|
1
An Overview of Physical Layer Security with
Finite-Alphabet Signaling
February 20, 2018
arXiv:1802.06401v1 [cs.IT] 18 Feb 2018
Sina Rezaei Aghdam*, Alireza Nooraiepour** and Tolga M. Duman*
(*) Dept. of Electrical and Electronic Engineering, Bilkent University, Ankara, Turkey
(**) Electrical and Computer Engineering Dept., Rutgers University, New Brunswick, NJ, US
Email: {aghdam, duman}@ee.bilkent.edu.tr, [email protected]
Abstract
Providing secure communications over the physical layer has been receiving growing
attention within the past decade. The vast majority of the existing studies in the area of
physical layer security focus exclusively on the scenarios where the channel inputs are Gaussian
distributed. However, in practice, the signals employed for transmission are drawn from discrete
signal constellations such as phase shift keying and quadrature amplitude modulation. Hence,
understanding the impact of the finite-alphabet input constraints and designing secure transmission schemes under this assumption is a mandatory step towards a practical implementation
of physical layer security. With this motivation, this article reviews recent developments on
physical layer security with finite-alphabet inputs. We explore transmit signal design algorithms
for single-antenna as well as multi-antenna wiretap channels under different assumptions
on the channel state information at the transmitter. Moreover, we present a review of the
recent results on secure transmission with discrete signaling for various scenarios including
multi-carrier transmission systems, broadcast channels with confidential messages, cognitive
multiple access and relay networks. Throughout the article, we stress the important behavioral
differences of discrete and Gaussian inputs in the context of the physical layer security. We
also present an overview of the practical code construction over Gaussian and fading wiretap
channels, and we discuss some open problems and directions for future research.
February 20, 2018
DRAFT
2
I. I NTRODUCTION
Wireless communications has become an indispensable part of our everyday lives. As
more and more data is being transmitted over wireless links, along with the reliability of
the transmission, ensuring secrecy of the sensitive information has become a challenging
issue. Traditionally, this issue is addressed using higher (network) layer solutions and
via computation-based mechanisms such as encryption. On the other hand, the security
provided by these methods mainly relies on the conjecture that the encryption function
is difficult to invert, which is not provable from a mathematical point of view. With
the ever-increasing computing power, encryption may no longer prevent information
leakage to sophisticated adversaries. Moreover, implementation of key-based secure
communications requires complicated protocols for key distribution and management,
which is highly challenging to implement in the cases of decentralized wireless networks.
Securing transmission at the physical layer is an alternative or a complement to the
cryptographic solutions. The basic idea is to exploit the inherent randomness of the
channel for achieving secrecy. Dissimilar to encryption based methods, in the schemes
employing physical layer security, no constraint is placed on the computational capability
of the eavesdropper. Furthermore, no key distribution/management is required in its
implementation.
During the past decade there has been an extensive growth of interest in studying the
capabilities of physical layer for securing communication. Several surveys and overview
papers are available describing the state-of-the-art on this topic. Fundamentals of physical
layer security are comprehensively explained in [1]-[4]. A review of recent results ranging
from point-to-point communications to multiuser scenarios is given in [5]. A number of
hurdles facing physical layer security are summarized in [6]. The authors in [7] present
a review of the recent research results on physical layer security under channel state
information (CSI) uncertainty in different nodes. A comprehensive overview on various
February 20, 2018
DRAFT
3
multiple antenna techniques in physical layer security is provided in [8].
The objective of this survey is to shed light on two other important (and highly related)
practical aspects, namely, secure transmission with discrete signaling and practical code
construction. We present a detailed review of the related recent results, underline the
important behavioral differences between the Gaussian inputs and discrete signaling,
and study the transmit signal design problem for securing communications in a variety
of scenarios. We also provide a review of recent results on practical code design for
secrecy, and make connections with the information theoretic results with finite inputs.
The remainder of the paper is organized as follows. In Section II, we provide a description of the fundamental concepts in physical layer security including the wiretap channel
model and different secrecy metrics. In Section III, we discuss the differences between
the secrecy performance of constellation-constrained and Gaussian inputs over Gaussian
wiretap channels. An overview of the secure multi-antenna transmission schemes with
finite-alphabet inputs is provided in Section IV. Sections V and VI discuss the scenarios
with multi-carrier transmission and multiuser channels, respectively. An overview of the
practical code design approaches for secrecy is presented in VII, followed by concluding
remarks and future research directions in Section VIII.
Notation: Vectors and matrices are denoted with lowercase and uppercase bold letters.
Non-bold letters are used to denote scalar values and calligraphic letters denote sets. The
expectation and probability mass (or density) function of a random variable W is denoted
by EW {.} and PW (.), respectively. P r(.) is used to denote the probability of an event.
An stands for a sequence of length n, i.e., An = {A(1), A(2), . . . , A(n)}. The notation
CN (m, R) denotes a circularly symmetric complex Gaussian random vector with mean
vector m and covariance matrix R.
February 20, 2018
DRAFT
4
II. F UNDAMENTALS
The wiretap channel model, originally introduced by Wyner in [10], is the most basic
scenario in the study of physical layer security. In this model, while the sender Alice
wishes to transmit a message signal to a legitimate receiver Bob; a third party, Eve,
is present with the capability of eavesdropping on Alice’s signal. The channel between
Alice and Bob, and the channel between Alice and Eve are referred to as the main
channel and the eavesdropper’s channel, respectively. In physical layer security, one is
interested in transmitting in such a way to maximize the transmission rate over the main
channel while keeping the eavesdropper ignorant about the message. This leads to the
notion of secrecy capacity, which is defined as the maximum rate at which the transmitter
can use the main link so as to deliver its message to the legitimate receiver in a way that
the eavesdropper cannot successfully obtain any information about it. Characterization
of the secrecy capacity is one of the most fundamental problems in the literature of
physical layer security in that it can provide us with vital implications on how secure
transmission techniques can be designed.
A. Different Notions of Secrecy
Consider a transmitter who wishes to transmit a message M to a legitimate receiver
while trying to keep an eavesdropper ignorant about it. M is mapped to a codeword X n
using a stochastic encoder with n denoting the number of channel uses. Then, X n is
transmitted, and Y n and Z n are received at the legitimate receiver and the eavesdropper,
respectively. The eavesdropper’s level of uncertainty is quantified using the equivocation
rate which is given by
Re =
1
H(M ∣Z n ),
n
(1)
which is nothing but a measure of how unlikely it is that the eavesdropper can infer source
information from its received signal. According to this definition, large equivocation rates
February 20, 2018
DRAFT
5
are equivalent to higher secrecy levels.
There are different metrics for measuring the secrecy guaranteed by a specific scheme
among which the information-theoretic ones are recognized as the strongest. Perfect
secrecy is achievable in a system if the message M and its corresponding encoder
output X n are statistically independent [11]. This can be expressed as I(M ; X n ) = 0,
which means that the mutual information between the message and the encoded signal
is exactly zero. This is to say, the eavesdropper is not capable of recovering the message
via observation of the encoder output in view of the fact that this encoded signal does
not provide any information about it. For this form of secrecy to be guaranteed, a secret
key of the length at least equal to the length of the message should be used. However,
weaker requirements can be adopted for evaluation of the secrecy in practical scenarios.
One example of such notions is semantic secrecy, which is achieved if it is guaranteed
that Z n does not increase the probability of guessing an arbitrary function f of M (with
an arbitrary distribution PM ) more than > 0 if Adv(M ; Z n ) < , where advantage
Adv(M ; Z n ) is defined as
Adv(M ; Z n ) = max ( ∑ PZ n (z n ) max P r(f (M ) = fi ∣z n )− max P r(f (M ) = fi )),
f,PM
zn
fi ∈supp(f )
fi ∈supp(f )
(2)
where supp(.) stands for the support of the function in its argument.
Assuming a uniform distribution for the message M , a system is called to operate
with strong secrecy if limn→∞ I(M ; Z n ) = 0, which means that if we consider codewords
of n symbols being transmitted by Alice as X n , the information leaked by observation
of the received vector Z n shall go to zero as n goes to infinity. On the other hand,
the condition limn→∞ n1 I(M ; Z n ) = 0 is referred to as weak secrecy. In particular, weak
secrecy requires the asymptotic rate of information leakage to be sublinear in n. Different
notions of secrecy, which are used to measure the secrecy guaranteed by a specific
scheme are summarized in Table I.
February 20, 2018
DRAFT
6
TABLE I: Four notions of secrecy.
Notion
Requirement
Perfect Secrecy [11]
I(M ; X n ) = 0
Semantic Secrecy [9]
Adv(M ; Z n ) <
Strong Secrecy [12]
limn→∞ I(M ; Z n ) = 0
Weak Secrecy [10]
limn→∞ n1 I(M ; Z n ) = 0
B. Secrecy Capacity of Single-Antenna Wiretap Channels
The notion of secrecy capacity was originally introduced by Wyner in [10] for degraded wiretap channels. In the scenario studied in [10], the main channel and the
eavesdropper’s channel are assumed to be discrete memoryless channels (DMCs) where
Eve’s observation of the transmitted signal is a degraded version of the signal received by
Bob. Wyner showed that, the secrecy capacity for this scenario can be expressed as the
difference of the mutual information between Alice and Bob with that of Alice and Eve,
maximized over all input distributions. Generalization of the notion of secrecy capacity
to the case of additive white Gaussian noise (AWGN) channels has been accomplished
in [13]. It is shown that, for the case of a Gaussian wiretap channel, nonzero secrecy
capacity can be attained if Bob receives the signal through a less noisy channel.
While the scenarios considered in [10] and [13] put the eavesdropper at a disadvantage,
Csiszár and Körner have studied the problem of secrecy for non-degraded channels in
[14]. In this channel model, referred to as broadcast channel with confidential messages,
the message transmitted by Alice contains a common message intended to both Bob and
Eve as well as a secret message, which is intended for Bob only, and is needed to be
kept secret from Eve. It is shown that the secrecy capacity over a discrete memoryless
February 20, 2018
DRAFT
7
wiretap channel (DMWC) is given by
Cs = max I(U ; Y ) − I(U ; Z),
(3)
PX∣U ,PU
where U is an auxiliary random variable. With a given PY Z∣X (y, z∣x), the secrecy
capacity is achieved by maximizing the difference of the mutual information terms
corresponding to the main and the eavesdropper’s channels over all joint distributions
PU X (u, x) where U satisfies the Markov chain relationship U → X → Y Z.
In order to study the physical layer security in wireless communication scenarios, we
need to extend the wiretap channel model to a model, which takes into account the fading
phenomenon. Assuming a narrowband transmission, fading is modeled as multiplicative
gains over the main and eavesdropper’s channels, denoted by HB and HE , respectively.
The first characterization of the role of fading in providing physical layer security is
given in [15] where quasi-static fading channels are considered towards the legitimate
receiver and the eavesdropper. It is shown via analyzing the secrecy outage probability
that, in presence of fading, secure transmission is possible even in the scenarios where
the eavesdropper has a better SNR. The secrecy capacity over block fading channels has
been derived in [16] under the assumption that each coherence interval is long enough
to allow for invoking proper random coding arguments. Given that the transmitter knows
either the main CSI (MCSI) or both the main and the eavesdropper CSI (ECSI) perfectly,
an achievable secrecy rate can be formulated as [16]:
+
Rs = EHB ,HE {[I(X; Y ∣HB ) − I(X; Z∣HE )] },
(4)
where [a]+ = max{a, 0}. It has been shown in [16] that the secrecy capacity that is
achieved under the full CSI assumption serves as an upper bound on the secrecy capacity
when only MCSI is known at the transmitter. This is due to the fact that with the
knowledge of instantaneous ECSI, the transmitter is capable of realizing a more efficient
transmission, which provides additional gains in terms of secrecy rates. For instance,
February 20, 2018
DRAFT
8
when perfect knowledge of both channels are available, the transmitter can optimize the
transmit power according to the instantaneous values of the channel gains HB and HE ,
providing a higher secrecy rate with respect to the case where the optimization of the
transmit power is carried out according to the main channel only.
Under the assumption that the transmitter does not know the realizations of the
channels, and it only has their statistics, assuming that each receiver knows its own
channel, the secrecy capacity can be obtained with the aid of the results for the case of
DMWC [14] as
Cs = max I(U ; Y, HB ) − I(U ; Z, HE ),
(5)
PX∣U ,PU
where U satisfies the Markov chain relationship U → X → (Y, HB ), (Z, HE ). This is
obtained by treating HB and HE as channel outputs at the legitimate receiver and at the
eavesdropper, respectively [17], [18]. Using the chain rule of mutual information and
noting the fact that HB and HE are independent of X and U , the expression in (5) can
be modified as
Cs = max EHB I(U ; Y ∣HB ) − EHE I(U ; Z∣HE ).
PX∣U ,PU
(6)
Determining the optimal joint distribution PU X (u, x) and the resulting exact secrecy
capacity is an open problem, however, by ignoring the channel prefixing, the performance
can be quantified using an achievable ergodic secrecy rate as given by
+
Rs = [EHB I(X; Y ∣HB ) − EHE I(X; Z∣HE )] .
(7)
C. Secrecy Capacity of Multi-Antenna Wiretap Channels
Multiple-input multiple-output (MIMO) wiretap channel is an extension of the wiretap
channel to a scenario where all the nodes, namely, Alice, Bob and Eve, are equipped with
multiple antennas. The secrecy capacity for this setup has been determined independently
by Khisti and Wornell [19] and Oggier and Hassibi [20]. The results in [14] are employed
February 20, 2018
DRAFT
9
by Khisti and Wornell so as to develop a genie-aided upper bound on the secrecy capacity
in the general case of non-degraded MIMO wiretap channels. While their characterization
is through a saddlepoint, Oggier and Hassibi propose an alternative approach through a
single optimization process. It has been shown in both [19] and [20] that the achievable
secrecy rate is maximized when the channel input is Gaussian, and the secrecy capacity
of a MIMO wiretap channel is given by
Cs =
max
Kx ⪰0,T r(Kx )=P
H
log det(I + HB Kx HH
B ) − log det(I + HE Kx HE ),
(8)
where Kx = E{xxH }, and HB ∈ CNE ×Nt and HE ∈ CNE ×Nt are the channel matrices
corresponding to the legitimate receiver and the eavesdropper, respectively.
An alternative characterization of the secrecy capacity of the MIMO wiretap channel
is presented by Liu and Shamai in [21] where they consider a more general matrix
constraint on the channel input. We highlight that, while [19]–[21] prove that the optimal
input is Gaussian, the optimal input covariance matrix needs to be determined using
numerical optimization approaches (see, e.g., [22]–[23]).
D. Other Secrecy Metrics
Secrecy capacity is the most widely used measure for evaluation of the secrecy
performance. However, a complete characterization of the secrecy capacity region is
prohibitive in some scenarios. In such cases, alternative metrics can be adopted as
surveyed below.
1) Secrecy Outage Probability: Secrecy outage occurs when the instantaneous secrecy
capacity falls below a target secrecy rate Rs . In delay-critical applications and for
scenarios with quasi-static fading, encoding over multiple channel states is not possible.
In these cases, secure transmission schemes can be designed such that the probability of
occurrence of an outage event is minimized. Obviously, this serves as a weaker notion
February 20, 2018
DRAFT
10
than secrecy capacity. This is due to the fact that, with this formulation, secrecy is not
guaranteed for the entire transmission duration.
2) Error Probability Based Metrics: In a number of works on physical layer security,
secure transmission schemes are designed based on constraining the bit error rate at
the legitimate receiver and the eavesdropper to pre-specified threshold values. This
leads to definition of a parameter called “security gap,” which quantifies the minimum
required difference between the received SNR values at Bob and Eve to ensure that
the bit error rate (BER) at Bob is smaller than a threshold while that at the Eve’s
receiver is larger than another threshold [24]. One may note that imposing high BER
at the eavesdropper does not satisfy any of the requirements given in Table I. However,
security gap serves as a practical measure of secrecy in a number of applications such as
explicit and implementable code design for physical layer secrecy. Some studies (e.g.,
[25]) consider the difference in the signal-to-interference-plus-noise ratio (SINR) at the
legitimate receiver and the eavesdropper as the secrecy metric. This is also a very weak
notion, which does not guarantee secrecy in an information theoretic sense.
III. G AUSSIAN W IRETAP C HANNELS WITH F INITE -A LPHABET I NPUTS
The secrecy capacity achieving input distribution over a Gaussian MIMO wiretap
channel is proved to be Gaussian [13]. However, due to the high detection complexity,
Gaussian signaling is not used in practice, and instead the transmission is carried out
with the aid of discrete inputs drawn from standard constellations such as PSK or QAM.
Hence, it is important to study the implications of discrete signaling in the context of
physical layer security.
As a first step in studying physical layer security under the finite-alphabet input
assumption, the impacts of standard constellations on the achievable secrecy rates of
Gaussian wiretap channels are studied in [26] and [27]. In [26], the authors evaluate
the constellation-constrained secrecy capacity and highlight an important behavioral
February 20, 2018
DRAFT
11
9
8
Capacity (bits/complex dim.)
7
6
BPSK
QPSK
16−PSK
16−QAM
64QAM
Gaussian
5
4
3
2
1
0
−15
−10
−5
0
5
10
15
20
25
SNR (dB)
Fig. 1: Capacity of an AWGN channel with Gaussian and constellation-constrained
inputs.
difference between finite-alphabet (e.g., PSK and QAM) and Gaussian inputs, that is, for
a fixed noise variance at Eve, the secrecy rate curves for PSK or QAM plotted against the
SNR have global maxima at finite SNR values. Investigation of the secrecy capacity of
the pulse amplitude modulation (PAM) inputs over a degraded Gaussian wiretap channel
in [27] leads to a similar conclusion. In addition, the authors in [27] demonstrate that
when finite constellations are employed, using all the available power for the information
bearing signal transmission may not be optimal. The reason behind these observations
has its roots in the mutual information expressions. That is, the capacity under an average
power constraint over an AWGN channel is given by
C = log2 (1 + SN R),
(9)
in bits per complex dimension, and it is an increasing function of the SNR. On the other
hand, the capacity of the constellation-constrained AWGN channel can be calculated as
February 20, 2018
DRAFT
12
[28]
⎧
⎫
⎪
∑x′ ∈X pY ∣X ′ (y∣x′ ) ⎪
⎪
⎪
I(X; Y ) = log2 ∣X ∣ − EX,Y ⎨ log2
⎬,
(10)
⎪
⎪
pY ∣X (y∣x)
⎪
⎪
⎩
⎭
where X is a constellation set and ∣X ∣ denotes its cardinality. Dissimilar to the capacity
expression in (9), the mutual information term in (10) converges to log2 ∣X ∣ at high SNRs
(see Fig. 1). This means that by increasing the transmit power, the mutual information
terms I(X; Y ) and I(X; Z) (calculated using (10)) will reach the saturation value, and
the secrecy rate drops to zero. This is also verified in Fig. 2, where the secrecy rates
are obtained by evaluating the difference of the mutual information terms corresponding
to the main and eavesdropper’s channels, under the assumption that the SNR at the
eavesdropper (denoted by SNRe ) is lower than the legitimate reciever’s SNR (denoted
by SNRb ) by 1.5 dB. This also clarifies that using all the available transmit power for
information transmission in the high SNR regime is not optimal [27]. Furthermore, it
can be observed from Fig. 2 that higher secrecy rates are achieved over Gaussian wiretap
channels by increasing the modulation orders.
The impact of output quantization on the secrecy rate of the binary-input Gaussian
wiretap channel is investigated in [29]. It is shown through theoretical analysis that
when the legitimate receiver has unquantized inputs where the eavesdropper has (binary)
quantized outputs higher secrecy rates are achievable with respect to the case where both
receivers have binary quantized or unquantized outputs.
To conclude this section, we note that while it may be tempting to argue that, as in
the case of transmission over an AWGN channel with no security considerations, one
can design general transmission schemes using Gaussian codebooks and simply adopt
the ideas to the case of inputs from standard constellations (e.g., PSK, QAM, etc.),
this approach does not extend in a straightforward manner. The optimal transmit signal
design strategies under Gaussian input assumption lead to considerable losses when
applied to the inputs drawn from standard constellations. For instance, it is shown in
February 20, 2018
DRAFT
13
0.45
Secrecy Rates (bits/complex dim.)
0.4
0.35
BPSK
QPSK
16−QAM
64−QAM
0.3
0.25
0.2
0.15
0.1
0.05
0
−15
−10
−5
0
5
10
SNR at Bob (SNRB)
15
20
25
Fig. 2: Secrecy rates with PSK and QAM over a degraded Gaussian wiretap channel
(SNRE = SNRB − 1.5 dB).
[27] the optimal power allocation strategy with PAM inputs is in sharp contrast to that of
Gaussian inputs. This motivates development of new algorithms and ideas for practical
constellation-constrained channels.
IV. FADING AND M ULTI -A NTENNA W IRETAP C HANNELS D RIVEN BY
F INITE -A LPHABET I NPUTS
Fading introduces new potentials for securing communications at the physical layer.
Specifically, the randomness due to the channel fluctuations can be exploited opportunistically by a transmitter to guarantee secrecy even in the scenarios where the eavesdropper
possesses a higher SNR (on average) than the legitimate receiver. Dissimilar to the
Gaussian wiretap channel where Gaussian input is secrecy capacity achieving, in the
presence of fading, discrete signaling may achieve higher secrecy rates. For instance, it
has been shown in [30] that, when Bob’s channel gain is on average worse than that of
February 20, 2018
DRAFT
14
Fig. 3: MIMO wiretap channel.
Eve’s, QAM inputs achieve higher secrecy rates at low and moderate SNR regimes. This
is because the discrete nature of the QAM inputs limits the leakage at the eavesdropper
whose channel is unusually good. At sufficiently high SNRs, however, the secrecy rate
with QAM inputs drops to zero similar to what is observed for the case of Gaussian
wiretap channels.
Taking advantage of the spatial degrees of freedom introduced by multi-antenna transmission can serve as an efficient solution to prevent secrecy rate of standard constellations
from dropping to zero at high SNRs. In the remainder of this section, we focus on a
system model where Alice, Bob and Eve are equipped with Nt , NB and NE antennas,
February 20, 2018
DRAFT
15
as depicted in Fig. 3. The received signals are given by
y = HB x + ny ,
(11)
z = HE x + nz ,
(12)
where HB and HE are the channel matrices corresponding to the legitimate receiver
and the eavesdropper. ny and nz denote circularly symmetric complex AWGN vectors,
elements of which follow CN (0, σ 2ny ) and CN (0, σ 2nz ), respectively.
Transmit precoding and artificial noise injection serve as important approaches for
enhancing secrecy [5]. Via precoding, it is possible to strengthen (or weaken) the transmitted signals in certain directions. Accordingly, it can be used as an effective tool to
increase the quality difference between the signals received at the legitimate receiver
and the eavesdropper. Beamforming is a special case of precoding where the transmitter
is restricted to using rank-one signals. It has a lower complexity with respect to the
general transmit precoding solution, achieved at the price of some performance loss.
Furthermore, injection of artificial noise [31] can effectively degrade the reception at the
eavesdropper while having no (or minimal) effect on the signal received at the legitimate
receiver. We also note that, to take advantage of the potentials offered by the channel
fading, the transmitter needs a certain level of knowledge on the CSI of both channels.
In what follows, we review the existing solutions for secure transmission over MIMO
wiretap channels with finite-alphabet inputs under different CSIT assumptions.
A. Transmit Signal Design with Full CSI
The most optimistic assumption regarding the transmitter’s CSI knowledge is availability of perfect instantaneous realizations of the channel matrices HB and HE . This
assumption is justifiable in the scenarios where Eve is an authorized user in the network,
however, she should be kept ignorant about the confidential messages transmitted from
Alice to Bob. In this case, the transmitter can take advantage of the full CSI knowledge
February 20, 2018
DRAFT
16
to design a precoder, which results in the maximal quality difference between the signals
received at the legitimate receiver and the eavesdropper.
With the aid of the perfect CSI corresponding to both channels, a generalized singular
value decomposition (GSVD) precoding solution has been proposed in [32] using which
the MIMO multi-antenna eavesdropper (MIMOME) channel is converted into a bank
of parallel channels, and a power allocation strategy is formulated to maximize the
achievable secrecy rate.
The necessary conditions for optimum transmit precoding is derived in [33], and it
is demonstrated that the GSVD precoding [32] is suboptimal. Alternatively, a gradient
descent optimization is proposed in which the precoder matrices are updated with steps
proportional to the gradient of the instantaneous secrecy rate. Furthermore, it is shown
that transmission along the null-space of the eavesdropper’s channel is the optimal
secure transmission strategy at the high-SNR regime. This is because, with a precoder
matrix along the null-space of the eavesdropper’s channel, the achievable rate at the
eavesdropper is suppressed to zero, and when the SNR is sufficiently high, the rate at
the legitimate receiver approaches the maximum value that can be achieved with the use
of specific finite-alphabet inputs, which guarantees that the secrecy rate is maximized.
However, it should be noted that, when the number of transmit antennas is less than or
equal to the number of antennas at the eavesdropper, it is not possible to suppress her
information rate by transmitting along the null-space of HE . Therefore, for maximizing
secrecy rates, only part of the available power should be allocated to data transmission.
Given Nt > NB , the authors in [33] suggest to use the excess power for transmission of
an artificial noise signal along the null-space of HB so as to prevent the secrecy rates
from dropping to zero. Fig. 4 compares the achievable secrecy rates by the schemes
proposed in [32] and [33] for a wiretap channel with Nt = 2, NB = 1 and NE = 1, and
February 20, 2018
DRAFT
17
fixed channels given by1
hB = [0.5128 − 0.3239j −0.8903 − 0.0318j ] ,
(13)
hE = [0.3880 + 1.2024j −0.9825 + 0.5914j ] .
(14)
The precoding approach proposed in [33] outperforms the GSVD-based precoding [32].
This improved performance is attained at the price of increased computational complexity. Furthermore, comparing the results given in Fig. 4 with those of Fig. 2 for QPSK
inputs over the Gaussian wiretap channel reveals the advantages offered by fading and
CSI-based precoding approaches.
An important conclusion that can be drawn from the results in [32] and [33] is that,
as in the case of non-fading channels, the structure of the optimal precoders under
the finite-alphabet input assumption is different from that of the precoders, which are
optimal for Gaussian inputs. For example, over MISO multi-antenna eavesdropper (MISOME) channels, the optimal transmission strategy is beamforming along the eigenvector
corresponding to the largest generalized eigenvalue of (Hb , He ) [34]. However, these
precoders (which correspond to signaling with rank one covariance), when applied to
the finite-alphabet inputs, undergo a considerable loss with respect to the precoders
optimized with finite-alphabet inputs (see, e.g., Fig. 7 in [33]).
As an alternative method to design a suboptimal precoder, one may also replace
the secrecy capacity with practical metrics (as opposed to those ensuring secrecy in
an information-theoretic sense) such as SINR-based metrics. For instance, the authors
in [35] design artificial noise beamformers, which serve as constructive interference to
the legitimate receiver (improving Bob’s SINR) while disrupting the reception at the
eavesdropper (degrading Eve’s SINR). In Table II, we summarize the precoder design
approaches with full CSI.
1
This example has also been considered in [33] and [34].
February 20, 2018
DRAFT
18
10
9
Secrecy capacity [34] (Gaussian inputs)
Gradient Descent Opt. [33] − QPSK inputs
GSVD−based Precoding [32] − QPSK inputs
Secrecy Rates (bits/s/Hz)
8
7
6
5
4
3
2
1
0
−10
−5
0
5
10
SNR (dB)
15
20
25
30
Fig. 4: Secrecy rates with QPSK inputs over fixed channels given in (13) and (14).
TABLE II: Precoder design with full CSI
Authors
System Model
Contributions
S. Bashar et al. [32]
MIMOME
Propose a GSVD-based precoding algorithm.
Y. Wu et al. [33]
MIMOME
Derive the necessary conditions for optimality of the precoder matrix and
develop a gradient descent based precoder optimization algorithm.
M. R. A. Khandaker et al. [35]
MISO, multiple
Design artificial noise which is constructive to the legitimate receiver and
Eves
destructive to the eavesdroppers.
B. Perfect MCSI and Statistical ECSI
While availability of the perfect ECSI at the transmitter may be practical in some
limited scenarios, in general, it is highly challenging to obtain it instantaneously. A
more practical assumption regarding CSIT is the availability of perfect MCSI along
February 20, 2018
DRAFT
19
with the statistical ECSI at the transmitter. On the other hand, the precoder design in
this scenario is not as effective as those carried out with the perfect knowledge of both
channels. In other words, in the absence of the instantaneous ECSI, precoding is not
as forceful in suppressing the reception at the eavesdropper. For instance, transmission
along the null-space of the eavesdropper’s channel is not even possible in this scenario.
An example of secure transmission scheme under perfect MCSI and statistical ECSI
is given in [37] where the authors consider a MISOME channel and define a practical
secrecy metric (instead of an information theoretic one), which quantifies the symbol
error probability of the confidential data, and show that in the absence of artificial noise,
secrecy diversity (i.e., the high-SNR slope of their defined metric) vanishes. This result
underlines the importance of artificial noise injection in these scenarios.
Artificial noise-aided transmit precoding strategies with the objective of maximizing
the secrecy rate are proposed in [38]–[40]. In [38], the authors employ naive beamforming
along with the artificial noise injection while considering single-antenna receivers. The
strategy proposed in [33], on the other hand, relies on an iterative maximization of an
approximation to the instantaneous secrecy rate. In both of these studies, it has been
shown that the optimal schemes allocate only a fraction of the total power for signal
transmission at high SNRs. Hence, the remaining power is employed for artificial noise
injection. It is shown in [39] that jointly optimizing the precoder matrix and portion of the
power allocated to the artificial noise outperforms the solutions, which rely on optimizing
the precoder only. Moreover, inspired by the idea proposed in [41], a generalized artificial
noise aided transmission is introduced in [40], which guarantees high secrecy rates for
the scenarios with Nt < NE where injection of artificial noise along the null-space of
the main channel is not possible. A summary of the existing solutions for the scenarios
with perfect MCSI and statistical ECSI at the transmitter is given in Table III.
February 20, 2018
DRAFT
20
TABLE III: Precoder design with perfect MCSI and statistical ECSI.
Authors
System Model
Contributions
T. V. Nguyen et al. [37]
MISOME
Define a weak notion of secrecy based on symbol error probability and
quantify the impact of artificial noise injection with beamforming on
diversity order.
S. Bashar et al. [38]
MISOSE
Employ a naive beamforming with the aid of MCSI only along with a
power optimization using perfect MCSI and statistical ECSI. The excess
power is used for artificial noise injection in the null-space of HB .
Y. Wu et al. [33]
MIMOME
Develops an iterative algorithm for precoder design and allocate the
excess power to inject artificial noise along the null-space of HB .
S. Rezaei Aghdam and
T. M. Duman [40]
MIMOME
Propose a joint precoder and (generalized) artificial noise optimization
algorithm.
C. Perfect MCSI Only
In the scenarios with a passive eavesdropper, a realistic assumption is that the transmitter does not know anything about the eavesdropper’s channel. Under this assumption,
[42] and [43] propose a secure transmission strategy referred to as directional modulation,
in which the amplitude and the phase of the transmit signal are adjusted by varying the
length of the reflector antennas for each symbol. This scrambles the PSK symbols in
all the directions other than that of the legitimate receiver. Other strategies for securing
communications without the knowledge of the eavesdropper’s channel are proposed in
[44] and [45]. Zhang et al. develop a Tomlinson-Harashima precoding in [44] where
the transmitter allocates part of its power in order to achieve a target mean squared
error (MSE) at the legitimate receiver, and the remaining power is used to transmit
artificial noise to degrade the eavesdropper’s reception. In [45], a secure space-time block
coding (STBC) scheme is proposed in which enables the legitimate receiver to perform
separable decoding, while requiring an exhaustive maximum likelihood (ML) detection
February 20, 2018
DRAFT
21
at the eavesdropper. Furthermore, the authors combine this scheme with artificial noise
injection to ensure a high uncoded BER at Eve. Table IV summarizes the existing
transmit signal design approaches under perfect MCSI and no ECSI at the transmitter.
TABLE IV: Transmit signal design with perfect MCSI and no ECSI.
Authors
System Model
Contributions
A. Kalantari et al. [42],
MIMOME
Steer the array beam such that the phase of the PSK modulated signal
[43]
L. Zhang et al. [44]
is scrambled in directions other than Bob.
MIMOME
Propose a nonlinear Tomlinson-Harashima precoding along with along
with artificial noise injection.
S. A. A. Fakoorian et
MIMOME Nt =
Design a STBC which results in a high BER for Eve and a low BER
al. [45]
NB = 2, NE ≥ 1
for Bob.
D. Statistical MCSI and ECSI
While most of the existing physical layer security solutions rely on the assumption
that the transmitter is capable of estimating at least the instantaneous main CSI, in
some scenarios (e.g., for fast fading channels), it may be difficult for the transmitter
to track the rapidly varying channel coefficients. For these cases, the impact of the
discrete inputs on the achievable secrecy rates is analyzed in [46]. Under the assumption
that both channels are doubly correlated, using the knowledge of transmit and receive
correlation matrices, transmit signal design algorithms are proposed in [47] to maximize
the resulting achievable secrecy rates. The results in [47] reveal that jointly optimizing
the precoder matrix and artificial noise results in increased achievable secrecy rates with
respect to precoding without artificial noise injection. The existing studies on MIMO
wiretap channels with statistical CSI are summarized in Table V.
We now compare the secrecy rates achieved by different transmit signal design approaches under different CSIT assumptions. In particular, we overview the results for the
February 20, 2018
DRAFT
22
scenarios with perfect or statistical CSI at the transmitter. When perfect CSI is available,
each element of HB and HE are modeled as independent and identically distributed
(i.i.d.) with CN (0, σB2 ) and CN (0, σE2 ), respectively. When considering the statistical
CSI corresponding to the main channel or the eavesdropper’s channel, a commonly
adopted model in the literature employs doubly correlated channels with
1/2
1/2
HB = ΨrB ĤB ΨtB ,
1/2
1/2
HE = ΨrE ĤE ΨtE ,
(15)
where the elements of ĤB and ĤE are i.i.d., and the transmitter is capable of acquiring
the transmit and receive correlation matrices, i.e., Ψtb , Ψte , Ψrb and Ψre , from the longterm statistics of the channel. The perfect instantaneous CSI is available at the receivers,
including the legitimate receiver and the eavesdropper.
Fig. 5 demonstrates the achievable secrecy rates by different transmission schemes
over a MIMOME wiretap channel with Nt = 4 and NB = NE = 2. Correlation matrices
are assumed to have exponentially decaying entries as
[Ψt (ρ)]ij = ρt ∣i−j∣ ,
[Ψr (ρ)]mk = ρr ∣m−k∣ ,
i, j = 1, 2, . . . , Nt ,
(16)
m, k = 1, 2, . . . , NB (or, NE ).
(17)
It is demonstrated that the precoder design using a gradient descent optimization
approach [33], [40] is a promising strategy for maximizing the secrecy rates. Fig. 5 also
reveals the importance of CSIT. When perfect CSI corresponding to the eavesdropper is
TABLE V: Transmit signal design with statistical MCSI and ECSI.
Authors
System Model
Contributions
M. Girnyk et al. [46]
MIMOME
Derive large system approximation for secrecy rate under arbitrary
channel inputs and quantify it for Gaussian and finite-alphabet inputs.
S. Rezaei Aghdam and
T. M. Duman [47]
February 20, 2018
MIMOME
Propose precoding and artificial noise injection approaches with the aid
of the knowledge on correlation matrices.
DRAFT
23
4
Ergodic Secrecy Rates (bits/s/Hz)
3.5
3
2.5
2
1.5
1
0.5
Perfect MCSI and ECSI [33]
Perfect MCSI and statistical ECSI [39]- ρte = 0.8, ρre=0.8
Statistical MCSI and ECSI [47]- ρtb = 0.9, ρre=0.6, ρte = 0.8, ρre=0.8
0
-10
-5
0
5
10
SNR (dB)
15
20
25
Fig. 5: Secrecy rates with different transmit signal design algorithms under different
CSIT assumptions.
available at the transmitter, it is possible to construct precoders, which are aligned with
the null-space of the eavesdropper’s channel. By this means, the achievable rate of the
eavesdropper can be suppressed to zero. Furthermore, even when perfect CSI of the main
user is available along with the statistical CSI of the eavesdropper, high secrecy rates are
still achievable with the aid of appropriately designed artificial noise injection strategies.
However, in the absence of instantaneous CSI, the precoder matrices designed with the
knowledge of correlation matrices are not capable of preventing the ergodic secrecy
rates from dropping to zero at high SNRs. In other words, since statistical CSI does
not provide sufficient degrees of freedom to take advantage of the fading for improving
secrecy, the behavior is similar to that in AWGN channels described in Section III.
February 20, 2018
DRAFT
24
E. Transmit Signal Design Based on Channel Reciprocity
A group of transmission strategies in the literature of physical layer security rely on
the assumption of channel reciprocity, that is, the transmitter and the legitimate receiver
observe the same channel, simultaneously. For the systems working in time division
duplex (TDD) mode, this property can be used to allow the transmitter to obtain the MCSI
using the pilots transmitted from the legitimate receiver. In [48], a secure transmission
scheme is proposed over a MIMO wiretap channel under the assumption that no training
signals are transmitted by Alice, and hence, no CSI is available at Bob or Eve. Alice
realizes an orthogonal space-time block coded PSK transmission and uses the CSI she
obtains from the reverse channel estimation to design a phase shifting precoder, which
compensates the phase shift by the channel. Therefore, Bob can recover the transmitted
messages with no need to CSI. On the other hand, Eve who observes an independent
channel can only recover the information with non-coherent detection, ensuring that this
strategy provides positive secrecy rates over MIMO wiretap channels.
In the scenarios with channel reciprocity, when channel estimation is carried out both
in the forward (from Alice to Bob) and the reverse (from Bob to Alice) directions, the
channel between Alice and Bob can serve as a source of common randomness, which can
further improve the secrecy. This common randomness is used in [49] for generation of
a random phase sequence, which is then used to manipulate the transmitted symbols (by
simply multiplying them by the obtained random phase terms). Since the eavesdropper
is unable to find out these random phase values, an enhanced secrecy is attained.
F. Performance-Complexity Trade-off in Transmit Signal Design
The iterative algorithms proposed in [33] and [40] for optimizing the precoder matrix
are shown to provide promising results in terms of the achievable secrecy rates. However,
their implementation complexity may be too high for some practical applications. There
February 20, 2018
DRAFT
25
are two main reasons for this high computational complexity. First, both the mutual
information and its gradient lack closed-form expressions. Therefore, Monte Carlo simulations need to be utilized to evaluate both terms, which require averaging over a
large number of noise realizations in order to maintain a high level of accuracy. More
importantly, the evaluation of the mutual information and the minimum MSE (MMSE)
term involves additions over the modulation signal space, which grows exponentially
with the number of transmit antennas, resulting in a prohibitively high computational
complexity when the number of transmit antennas is large.
Different strategies are proposed for lowering the computational complexity associated
with the transmit signal design for secrecy under the finite-alphabet input assumption.
The authors in [33] and [39] derive closed-form approximations for the secrecy rate
expressions using bounds on mutual information. While the former employs the bounds
given in [50], the latter approximates the mutual information using the cut-off rate
expression. The authors in [51]-[52] derive asymptotic secrecy rates in a regime where
the numbers of antennas at the transmitter and both receivers grow infinitely large
with a constant ratio. Maximization of this expression is shown to yield a satisfactory
performance even in the cases of small number of antennas. The authors in [53] and
[40] propose per-group precoding strategies under perfect and partial CSIT assumptions,
respectively. In these works, the channels are diagonalized and the precoder matrices are
designed by grouping the transmit antennas and designing the precoder matrix for each
group, independently.
G. Secure Spatial Modulation
Spatial modulation and space shift keying (SSK) are relatively new MIMO transmission schemes, which rely on encoding information via active antenna indices. Similar
to the amplitude or phase modulation schemes, in spatial modulation and SSK, the
channel inputs are drawn from discrete and finite sets. More specifically, in spatial
February 20, 2018
DRAFT
26
modulation [54], the data bits are mapped into two information carrying blocks: 1) an
antenna index selected from the set of transmit antennas, 2) a symbol drawn from a
standard constellation. Space shift keying is a special case of spatial modulation where
no amplitude of phase modulation is employed, i.e., the transmit antenna index serves
as the only information carrying unit [55].
Different studies explore the advantages offered by spatial modulation over state-ofthe-art MIMO schemes (see [56] for a comprehensive overview). Along with different
application areas of spatial modulation and SSK, studying these transmission schemes
in the context of physical layer security has been of recent interest. Secrecy capacity
of spatial modulation for the scenarios with two transmit antennas is studied in [57],
and the results are extended to MISOSE and MIMOME scenarios in [58] and [59],
respectively.
Different secure spatial modulation transmission schemes are also proposed with the
aid of MCSI and ECSI at the transmitter and for systems with channel reciprocity. A
precoder optimization algorithm for maximizing secrecy rates with SSK transmission
using full CSI is proposed in [60]. The authors in [61] provide a precoder design and
artificial noise injection approach for securing spatial modulation for the scenarios with
passive eavesdroppers (with no ECSI at the transmitter). The secrecy rates of spatial
modulation with artificial noise injection are characterized in [62]. While the conventional
artificial noise injection (similar to what is considered in [62]) requires multiple antennas
to be activated, a novel secure artificial noise aided transmission is introduced in [63],
which requires only two antennas to be activated at each time instance. The secure
transmit signal design approaches for spatial modulation and SSK are summarized in
Table VI.
For the scenarios where channel reciprocity holds, a transmit preprocessing technique
is proposed in [64] under the assumption that the training sequences are transmitted from
February 20, 2018
DRAFT
27
TABLE VI: Transmit signal design for secure spatial modulation and SSK.
Paper
Model
Contributions
S. Sinanovic et al. [57]
Nt = 2, NB = NE = 1,
Characterize secrecy rates for spatial modulation and SSK
No CSI at the transmitter
transmission.
MISOSE
Evaluate the secrecy rates by deriving mutual information
X. Guan et al. [58]
terms and propose a precoding with the aid of full CSI.
S. Rezaei Aghdam and
MIMOME, No CSI at the
Derive expressions for achievable secrecy rates with spatial
T. M. Duman [59]
transmitter
modulation and SSK schemes.
S. Rezaei Aghdam and
MIMOME, Full CSI at the
Develop an iterative precoder design algorithm.
T. M. Duman [60]
transmitter
F. Wu et al. [61]
MIMOME, Nt > NB ,
Propose a precoder and artificial noise design.
MCSI at the transmitter
L. Wang et al. [62]
Z. Huang et al. [63]
MIMOME, Nt > NB ,
Characterize the secrecy rates of artificial noise aided spatial
MCSI at the transmitter
modulation.
MIMOME, MCSI at the
Propose an artificial noise aided secure quadrature spatial
transmitter.
modulation which requires low-complexity hardware.
Bob to Alice, and no CSI is available at the receiver nodes. The authors in [65] propose
a dynamic antenna index assignment based on the main channel when training is carried
out at both directions, which prevents an eavesdropper from obtaining the antenna index
assignment pattern.
V. OFDM W IRETAP C HANNEL WITH D ISCRETE I NPUTS
Orthogonal frequency-division multiplexing (OFDM) provides robustness against multipath channel fading and offers flexibility in resource allocation. These features have
made OFDM a widely used technique in recent wireless standards including 4G and 5G
wireless systems. In the literature of physical layer security, OFDM can be modeled as
February 20, 2018
DRAFT
28
a set of independent parallel channels2 . It is proved in [67] that the secrecy capacity of
K independent parallel channels is achieved when each channel achieves its individual
secrecy capacity. For parallel Gaussian wiretap channels, the optimal input is Gaussian
distributed [68]. However, a more practical study of OFDM wiretap channel is possible
via assuming that the data symbols belong to finite constellations.
An initial study on parallel Gaussian wiretap channel under finite-alphabet inputs is
reported in [27] where the authors have characterized the secrecy capacity with PAM
inputs. They also provided a Mercury-waterfilling interpretation of the optimal power
allocation strategy. It is observed in [27], that dissimilar to standard parallel Gaussian
channels (with no constraints on the input distribution), it may not be optimal to use
all available power for PAM inputs. In [69], the OFDM wiretap is studied with QAM
inputs. After quantifying the loss in the secrecy rates with respect to the scenarios with
Gaussian inputs, a bit loading strategy is proposed to minimize this loss by assigning
an appropriate number of bits to each subchannel.
The sensitivity of OFDM based transmissions to frequency synchronization errors can
be exploited to provide secure transmission as well. The authors in [70] propose a secure
OFDM transmission over reciprocal channels, which relies on inducing carrier frequency
offset that is pre-compensated (using the transmitter’s knowledge on the MCSI) in
such way that it is received without inter-carrier interference at the legitimate receiver,
while the reception at the eavesdropper who receives the signal through an independent
channel is considerably degraded. In [71], an eavesdropping resilient OFDM scheme is
developed over reciprocal channels where Alice performs a subcarrier interleaving using
her knowledge on the instantaneous MCSI. The legitimate receiver who is also capable
of acquiring the MCSI can obtain the interleaving pattern and de-interleave the received
signals. In contrast, the eavesdropper cannot derive the interleaving pattern initiated by
2
This is based on the assumption that both the legitimate receiver and the eavesdropper use OFDM receivers.
February 20, 2018
DRAFT
29
TABLE VII: OFDM wiretap channel with discrete inputs
Paper
Model
Contributions
M. R. D. Rodrigues et
Parallel Gaussian Wire-
Characterize secrecy rates with PAM inputs and derive optimal
al. [27]
tap Channels
power allocation strategy.
F. Renna et al. [69]
SISOSE
Evaluate the secrecy rates for OFDM systems with QAM input
constellations and provide a bit-loading strategy.
M. Yusuf et al. [70]
H. Li et al. [71]
H. Qin et al. [72]
SISOSE with reciprocal
Suggest inducing carrier frequency offset which is pre-
main channel
compensated in the legitimate receiver’s direction.
SISOSE with reciprocal
Propose subcarrier interleaving based on the MCSI which pre-
main channel
vents eavesdropper from de-interleaving the received signals.
SISOSE
Introduce a time-domain artificial noise injection approach and
propose a joint optimization of data power and artificial noise
covariance matrix.
T. Akitaya et al. [73]
MIMOME
Study application time-domain artificial noise injection approach
in multi-antenna scenarios.
the transmitter, and as a result, fails to acquire the transmitted information.
The temporal degrees of freedom offered by the cyclic prefix can be used for artificial
noise injection so as to improve secrecy over OFDM wiretap channels. In [72], a singleantenna OFDM wiretap channel is considered and an artificial noise signal is inserted
into the time-domain signal (over the cyclic prefix samples). A framework for joint
optimization of the power allocated to each subcarrier and artificial noise covariance
matrix is also proposed. In [73], Akitaya et al. study the application of time-domain
artificial noise injection in MIMOME OFDM systems. The results in [72] and [73]
demonstrate that this approach is efficient for improving the secrecy rates in OFDM
wiretap channels.
Table VII summarizes various results on the study of OFDM wiretap channels under
the finite-alphabet input assumption.
February 20, 2018
DRAFT
30
VI. C ONSTELLATION C ONSTRAINED M ULTIUSER AND C OOPERATIVE W IRETAP
C HANNELS
The concept of physical layer security has also been extended to multi-user and
cooperative networks in recent years. In addition to the numerous studies on the limits
of secure communications in the scenarios with more than two receivers, multiple
transmitters and presence of relays/jammers/helpers (see [5] for a survey), some recent
focus has been placed on the practical constellation-constrained secure transmission for
these scenarios.
A. Multiuser and Multi-Eve Networks
1) Broadcast Channel with Confidential Messages: The achievable rate region for
Gaussian broadcast channel with confidential messages is characterized for discrete
inputs in [74]. In this scenario, the transmitter’s objective is to send a common message
to two receivers as well as to deliver a confidential message to one of them. Dissimilar
to [27] and [26] where standard constellations are considered, in [74], the symbols are
allowed to be arbitrary, and the achievable secrecy rate region is enlarged by optimizing
the symbol positions and the probability distribution of the symbols. The shrinkage of
the rate region due to employing PAM inputs is quantified with respect to the optimal
Gaussian inputs, and a number of important behavioral differences between these cases
are highlighted.
2) Multiple Access Channels: The authors in [75] focus their attention on the multiple
access wiretap channels under the assumption that the users employ the Alamouti STBC
scheme. Specifically, they consider a case where multiple transmitters communicate with
one legitimate receiver in the presence of an eavesdropper. Under the assumption that
the transmitter has the knowledge of the legitimate channel and does not know the
eavesdropper’s channel gain, an artificial noise injection is proposed along the null-
February 20, 2018
DRAFT
31
space of the main channel, which is shown to give rise to considerable improvements
in the secrecy sum-rate.
3) Cognitive Radio Networks: The linear precoder design for the cognitive multipleaccess wiretap and the cognitive multi-antenna wiretap channels with finite-alphabet
inputs are studied in [76] and [77], respectively. In [76], the authors consider a setup
in which two secondary-user transmitters are communicating with a secondary-user
receiver in presence of an eavesdropper and under interference threshold constraints at
the primary users. A two-layer precoding algorithm is proposed using statistical CSI at
the transmitters to maximize the ergodic secrecy sum rate. In [77], on the other hand, the
authors study a scenario where a multi-antenna secondary-user transmitter communicates
with a multi-antenna secondary-user receiver and the communication is wiretapped by
a multi-antenna eavesdropper. In this work, the precoder matrix is optimized using an
iterative algorithm (assuming statistical CSI of all the channels) so that the secrecy rate
of the secondary user is maximized while power leakage to the primary user sharing the
same frequency spectrum is controlled.
4) Multi-Eavesdropper Scenarios: The authors in [78] study a setup comprised of a
source, a legitimate receiver and multiple multi-antenna eavesdroppers. They first obtain
an expression for an achievable secrecy rate, and then, under the full CSI assumption,
they optimize the transmit power and the beamforming vector. The problem of power
allocation for secrecy over MIMO wiretap channels with multiple eavesdroppers is
studied in [79]. Under the assumption that the transmitter has the perfect MCSI and
statistical ECSI, the proposed power allocation strategy gives non-zero secrecy rates at
high transmit powers. This is important because in the absence of this power control
strategy, secrecy rate decreases with increasing transmit power, and it drops to zero when
the eavesdroppers’ SNRs get higher than a certain value.
February 20, 2018
DRAFT
32
B. Relay Channels and Cooperative Communications
Cooperation serves as an efficient method for enhancing secrecy in wireless networks.
In this context, along with the studies conducted on secrecy capacity in various cooperative scenarios, a number of recent studies focus on finite-alphabet inputs as the
signaling scheme. In this context, the secrecy rates achieved with decode-and-forward
(DF) relay beamforming under finite-alphabet input assumption are characterized in [80].
It is shown that by optimizing the source power and also the relay weights using global
CSI, it is possible to prevent the secrecy rates from dropping to zero. The authors in
[81] consider a cooperative jamming network in presence of multiple eavesdroppers
and develop a secure transmission scheme with finite-alphabet inputs, which relies on a
joint optimization of the artificial noise (injected in the null-space of relay-destination
links) and power allocation among the source and the relays. The study in [82] considers
secure transmission over a MIMOME setup with the aid of a multi-antenna helper, which
transmits a jamming signal to degrade the signal received by the eavesdropper. Similar
to [81], a framework is provided for joint optimization of the precoder matrix and power
allocation between the transmitter and the helper. Furthermore, low-complexity transmit
signal design schemes are developed for both low and high SNR regimes.
Table VIII summarizes the recent research results of secure transmission over finiteinput multi-user, multi-eavesdropper and cooperative channels.
VII. P RACTICAL W IRETAP C ODE C ONSTRUCTIONS
The achievability of the secrecy rates described in the previous sections is proved with
the aid of random coding arguments, i.e., from an information theoretic point of view.
However, in practice, one needs to realize secure communications with the aid of practical
codes. In this regard, explicit and implementable code design for secrecy has also been
a topic of recent interest to researchers. In particular, error-correcting codes have found
a new role since the emergence of physical layer security. Aside from providing reliable
February 20, 2018
DRAFT
33
TABLE VIII: Multi-user, multi-eavesdropper and cooperative secure transmission with
discrete inputs.
Paper
Network
Contributions
Z. Mheich et al. [74]
Broadcast channel with
Characterize secrecy rate region with PAM inputs and enhance it by
confidential message
optimizing symbol positions and the joint probability distribution.
MAC Wiretap
Develop a secure Alamouti transmission with artificial noise injec-
T. Allen et al. [75]
tion with perfect MCSI and statistical ECSI.
J. Jin et al. [76]
Cognitive MAC Wiretap
A two-layer precoding is proposed using statistical CSI to maximize secrecy sum-rate.
W. Zeng et al. [77]
K. Cao et al. [78]
Cognitive Multi-Antenna
Propose an iterative precoder optimization algorithm with statistical
Wiretap
CSI.
MISO
Multiple
Wiretap
with
Multi-Antenna
Study joint optimization of beamforming vector and transmitting
power with full CSI.
Eavesdroppers
S. Vishwakarma et al.
MIMO Wiretap with Mul-
Employ a power allocation strategy with perfect MCSI and statis-
[79]
tiple Eavesdroppers
tical ECSI.
S. Vishwakarma et al.
DF Relaying in Presence
Optimize the source power and the relay weights with the aid of
[80]
of Multiple Eavesdroppers
full CSI such that the secrecy rate is maximized.
K. Cao et al. [81]
Cooperative
Provide a joint optimization of the artificial noise and the power
Jamming
Network in Presence of
allocated between the source and the relays.
Multiple Eavesdroppers
K. Cao et al. [82]
MIMOME and a Multi-
Propose a joint optimization of precoder and power allocation
Antenna Helper
between the source and the helper
transmission, carefully designed codes are also capable of securing the transmitted
messages against an eavesdropper. On the other hand, it seems intriguing to design
these codes because they are required to meet two conflicting requirements: 1) they tend
to add more redundancy to the codewords in order to counter the randomness of the
channel and provide reliable communication. 2) in order to offer a secure transmission,
they need to maximize the randomness perceived at the eavesdropper by reducing the
February 20, 2018
DRAFT
34
redundancy.
An important step towards designing codes for the wiretap channel is choosing a
metric for measuring security. In contrast to measuring reliability where probability
of error is being used as a universally accepted metric, there are multiple metrics for
quantifying secrecy for this scenario as listed in Table I. Computation of these metrics
for explicit finite length coding schemes, however, has been limited to some specific
channels3 . One example is [83] in which the authors apply semantic secrecy to a finite
Reed-Muller code used over a binary erasure channel (BEC). For AWGN channels, a
BER based metric called security gap (as mentioned in Section II-D2) is widely used
by the research community, which is easily applicable to all practical coding schemes,
however, its relation to the information theoretic metrics is not established in a rigorous
manner.
Security gap is defined as the difference between the qualities (SNRs) of the main
and the eavesdropper’s channels. For AWGN channels without feedback, this value must
be greater than zero in order to result in a positive secrecy capacity (which equals to
the difference between the capacities of the main and the eavesdropper’s channels [13]).
Small security gaps are desirable because they make physical layer security achievable
even with a small degradation of the eavesdropper’s channel with respect to the main
one. An analytical definition of the security gap can be given as follows. Let us fix
two threshold values for the maximum desired BER over the main channel (denoted by
PBmax ≤ 21 ) and and the minimum desired BER at the eavesdropper (denoted by PEmin > 0).
Then, the security gap equals the difference between the SNR values corresponding to
these two, i.e., SN RB − SN RE measured in dB.
The major shortcoming of the above BER-based analysis is that it assumes that if
3
The authors in [98] provide a coding scheme for an AWGN channel, which satisfy strong secrecy using capacity
approaching codes of infinite length and hash functions.
February 20, 2018
DRAFT
35
min ≈ 0.5, then Eve cannot extract any information about the secret message. However,
Peve
if there exist statistical correlations between the input bit errors and the output bit errors
or there exist correlations among the output bit errors, the amount of leaked information
will not be zero. The authors in [89] employ tools from cryptography literature to eliminate such correlations. Specifically, they propose the use of the substitution permutation
networks (SPNs) without any secret key to satisfy the strict avalanche criteria (SAC) and
the bit independence criterion (BIC). The required SPNs can be implemented efficiently
[89], hence, they can be applied to the message bits before the encoding stage. Therefore,
one can design practical coding schemes, which result in BER very close to 0.5 with
no (or, small) correlation at Eve.
The general coding scheme for physical layer security is the randomized encoding
proposed by Wyner in [10] to prove that there exists a code which achieves the secrecy
capacity of the degraded wiretap channel. This method is also known as coset-coding
studied further in the subsequent literature, e.g., in [84] and [85]. In this method, allzero message is mapped to a code C with generator matrix G, and a non-zero message
s = [s1 , s2 , ..., sk ] is mapped to the coset obtained by adding the n-tuple s1 h1 + s2 h2 +
... + sk hk to all the codewords in C where hi ’s are linearly independent n-tuples outside
C. Assuming hi ’s form a matrix H of size k × n, the transmitted codeword through the
channel is
⎡ ⎤
⎢H⎥
⎢ ⎥
c = [s v] ⎢⎢ ⎥⎥
(18)
⎢G⎥
⎢ ⎥
⎣ ⎦
where v = [v1 , v2 , ..., vr ] denotes the random bit vector used for choosing a codeword in
the corresponding coset uniformly randomly. Fig. 6 illustrates this method where small
code is denoted by C and cij denotes the ith codeword in the j th coset. We note that
in conventional encoding, the leakage of information about the secret message to Eve,
H(SK ∣Zn ), is equal to the leakage of the codewords H(cn ∣Zn ) [83] since there is a oneto-one mapping between the messages and the codewords. In the randomized encoding
February 20, 2018
DRAFT
36
Fig. 6: Illustration of the coset coding scheme.
scheme, however, this is not true anymore since different codewords may represent the
same message.
Application of lattice codes to the Gaussian wiretap channel is studied in [86] where
the authors define an alternative security metric (called secrecy gain), which is related
to the theta series of lattices and shows the amount of confusion at the eavesdropper.
Without introducing a decoding method, they evaluate the performance of different
lattices based on the secrecy gain. The confusion at the eavesdropper in [86] is the result
of using a random lattice in addition to the lattice that is responsible for transmitting
the original message. As a further development, the authors in [87] prove the existence
of lattices that are semantically secure in the Gaussian wiretap channel.
The authors in [96] and [97] provide a lower bound on the mutual information between
the secret message and Eve’s observation. They utilize this lower bound to quantify the
equivocation for finite length codes for Gaussian wiretap channels. Specifically, they use
finite length LDPC codes and demonstrate the achievable points in the equivocation-rate
February 20, 2018
DRAFT
37
TABLE IX: Practical coding schemes aiming at reducing the security gap over AWGN
channel.
Paper
Secure coding scheme
Contributions
D. Klinc et al. [88]
Punctured LDPC codes
Propose puncturing information bits in LDPC codes in
order to confuse Eve.
M. Baldi et al. [90]
Non-systematic codes
Make use of scrambling technique at encoding which
amplifies the resulting errors at decoding stage.
M. Baldi et al. [91]
Y. X. Zhang et al. [93]
Coding with scrambling, con-
Investigate different techniques to reduce the resulting
catenation, and HARQ
security gaps.
Polar-LDPC
Evaluate performance of a serially concatenated coding
concatenated
codes
scheme consists of polar and LDPC codes.
A. Nooraiepour and T. M.
Randomized convolutional and
Construct coset codes based on convolutional and turbo
Duman [92]
turbo codes
codes using dual codes.
region.
Several practical coding schemes aiming directly at reducing the security gap are also
proposed in the literature. Specifically, punctured LDPC codes are exploited for physical
layer security in [88]. Furthermore, [90] demonstrates how non-systematic codes can be
effective to reduce the resulting security gap, while [91] applies different techniques
including scrambling, concatenation, and hybrid automatic repeat-request to LDPC and
BCH codes in order to further reduce the security gap. Concatenation of polar and
LDPC codes are proposed in [93]. Coset codes based on convolutional and turbo codes
are constructed in [92] and [94] for physical layer security. Finally, application of serially
concatenated low density generator matrix (SCLDGM) codes to the randomized scheme
is studied in [95]. Table IX summarizes the existing practical coding schemes in the
literature along with their main contributions. Fig. 7 demonstrates the relation between
February 20, 2018
DRAFT
38
0.5
0.45
min
BER at the Eavesdropper (Peve
)
0.4
0.35
0.3
0.25
0.2
0.15
RC, n=2056, k=1020, r=252
Punctured LDPC, n=2364, k=1576
RSCCC, n=8004, k=r=2000
RSCLDGM, n=10004, k=2250, r=250
S RSCCC, n=10004, k=r=2500
S RSCLDGM, n=10004, k=2250, r=250
0.1
0.05
0
0
0.5
1
1.5
2
2.5
3
Security gap (dB)
3.5
4
4.5
5
Fig. 7: The resulting security gaps from practical coding schemes. n, k and r denote
length of a code, number of data bits and number of random bits, respectively. S means
scrambling has been applied. Taken from [92] and [95].
the resulting security gaps from some of these practical coding schemes4 and a specific
BER at Eve. One can see that when a BER of 0.5 is desired at Eve, security gaps as
small as 1 dB can be obtained using carefully designed codes of length 104 or longer.
Authors in [98] introduce a method, which provides strong secrecy for the Gaussian
wiretap channel by means of hash functions and capacity approaching channel codes
of infinite length. Using this scheme, it is also possible to quantify the level of leakage
obtained for a fixed block length. Finally, [99] provides the maximal secrecy rate over
4
RSCCC stands for randomized serially concatenated convolutional codes scheme, RC stands for randomized
convolutional codes scheme and RSCLDGM corresponds to the randomized SCLDGM codes.
February 20, 2018
DRAFT
39
a wiretap channel subject to reliability and secrecy constraints at a given block length.
In addition to the Gaussian wiretap channels, some efforts are also made to develop
practical coding schemes for fading wiretap channels. The authors in [100] and [101]
propose polar coding schemes for achieving secrecy over block fading wiretap channels.
In [100], the secure coding scheme is designed with the aid of instantaneous MCSI and
ECSI. However, the polar coding approach proposed in [101] relies on statistical CSI of
both channels. Application of algebraic lattice codes is also proposed over block fading
wiretap channels. For instance, a sequence of non-random lattice codes is developed in
[102] that achieve strong secrecy and semantic security over ergodic fading channels.
We emphasize that the easy-to-compute property of security gap makes it a widely
used metric in majority of the aforementioned code designs, which try to avoid potential
complexities associated with computing the original information-theoretic metrics for
finite-length scenarios. Despite some limited efforts such as [97] in providing a lower
bound for information-theoretic metrics (Table I), the literature of physical layer security
still lacks a comprehensive and satisfactory finite length analysis of different metrics
especially for AWGN and fading wiretap channels.
VIII. L ESSONS L EARNED AND THE C HALLENGES A HEAD
We now provide a number of important conclusions that can be drawn from the
existing results for different wiretap channels driven by finite-alphabet inputs. We also
discuss a number of challenges ahead and present some directions for future research
in this area.
A. Concluding Remarks
An obligatory step towards practical implementation of physical layer security is understanding the impacts of employing standard constellations on the secrecy performance.
The usual approaches for designing secure transmit signals with Gaussian inputs lead
February 20, 2018
DRAFT
40
to considerable losses when applied to the signals drawn from finite constellations. This
fact motivates the need for development of secure transmission schemes under the finitealphabet input assumption.
In single-antenna transmission with constellation constrained inputs, achieving secrecy
requires adoption of a proper power allocation strategy. This is because transmission with
full power may allow the eavesdropper to successfully detect the transmitted symbols.
Moreover, the increased dimensionality in multi-antenna transmissions can be employed
to enhance secrecy under finite-alphabet inputs. Precoding serves as a promising strategy
for maximizing the quality difference between the signals received at the legitimate
receiver and the eavesdropper. Artificial noise injection serves as a promising strategy
for enhancing secrecy and it well-suits the nature of secure transmission with finitealphabet inputs since the optimal power, which is allocated to the information bearing
signal is only a fraction of the total power, and the excess power can be used for artificial
noise injection. In many cases (especially, in the absence of instantaneous ECSI), without
artificial noise injection, secrecy rates drop to zero when the transmit power increases.
Therefore, injecting artificial noise is crucial for guaranteeing a good saturation behavior
at high SNRs.
The optimal secure transmission schemes are not known in closed-form in almost all
of the scenarios studied in this survey. This is due to the non-convexity of the objective
function (i.e., secrecy rate). Furthermore, the transmit signal design algorithms, which
rely on direct maximization of secrecy rates are computationally complex mainly due to
the fact that the mutual information expression under finite-alphabet inputs lacks closedform. Therefore, different studies focus their attention on proposing lower-complexity
alternatives to direct secrecy rate maximization. There is also an important trade-off
between the secrecy rates and the computational complexity of the existing solutions
in the literature of physical layer security with finite alphabet inputs. Moreover, it is
February 20, 2018
DRAFT
41
noteworthy that the transmit signal designs based on the metrics such as SINR or BER,
though simple and may be practical in some scenarios, do not ensure secrecy in an
information theoretic sense.
The amount of CSI at the transmitter plays a central role in determining the secrecy
performance of a particular transmission scheme. Transmission algorithms employing full
CSI typically have a high capability in simultaneously increasing the information rates
at the legitimate receiver and suppressing the reception at the eavesdropper, however,
acquiring a perfect instantaneous CSI of the eavesdropper may not be practically possible
in many scenarios. Although the achievable secrecy rates undergo considerable losses
in the absence of instantaneous ECSI, the existing approaches reveal that high secrecy
rates can still be attained with (or in some cases even without) it.
Practical coding schemes exploit different tools from coding theory and cryptography
in order to ensure a certain level of security and reliability (in terms of BERs) for Eve
and Bob, respectively. By combining coset coding and scrambling approaches along with
code concatenation, security gaps as small as 1 dB can be obtained using finite-length
explicit and implementable codes. However, finding specific (and implementable) finitelength codes, which can provide physical layer security from an information-theoretic
point of view, remains as a future challenge.
B. Challenges and Open Problems
Various studies have been carried out on physical layer security with discrete signaling
with the objective of bridging the gap between information theoretic limits of secrecy to
practical and implementable secure transmission schemes. However, there still are many
important open problems and future challenges to be addressed in this area.
Physical layer security with finite-alphabet inputs has been studied only for a limited
set of CSI scenarios including the cases with perfect and statistical MCSI and ECSI at
the transmitter. Some other important scenarios regarding the CSI uncertainty at different
February 20, 2018
DRAFT
42
nodes, e.g., noisy estimation of CSI, outdated CSI or limited CSI feedback, which are
well-investigated under the Gaussian input assumption [7], have not yet been considered
for cases with finite-alphabet inputs. It would be interesting and of significant practical
value to quantify the losses in the secrecy rates under these different imperfect CSI
scenarios in the context of finite-alphabet inputs.
While the majority of artificial noise injection strategies rely on transmitting Gaussian
distributed noise either by aligning it in the null-space of the main channel or by optimizing its covariance matrix, such artificial noise signals are not necessarily optimal. For
instance, it has been shown in [103] that the worst case noise for AWGN channels with
binary inputs has a discrete distribution. This motivates seeking interference distributions
(rather than merely optimizing covariance matrices of Gaussian distributed signals),
which maximize secrecy rates.
While some steps have been taken in studying physical layer security over finite-input
multiuser networks (as surveyed in Section VI), many important scenarios have not yet
been tackled. For example, no results have been reported for interference channels with
secrecy constraints under the finite-alphabet input assumption. Secrecy issues concerning
the networks with simultaneous wireless information and power transfer (SWIPT) is also
a topic of recent interest [104]. It is demonstrated that SWIPT networks are prone to
information leakage at the energy receivers, and hence, the main focus of the existing
articles is the characterization of the trade-off between the secrecy capacity/rate and the
harvested power in these networks [105]-[107]. Studying this trade-off under practical
discrete signaling assumption is an important subject for future research. Furthermore,
physical layer security in heterogeneous networks [108] and with full duplex transmitter
or receivers (see, e.g. [109]) are other hot topics in this area, and extending the existing results for these scenarios to the case of finite-alphabet inputs constitutes another
important direction for future research.
February 20, 2018
DRAFT
43
Massive MIMO [110] and millimeter wave (mmWave) communication systems [111]
have attracted significant interest from researchers in industry and academia. Exploiting
large antenna arrays and taking advantage of a spectrum from 30 GHz to 300 GHz
away from the almost fully occupied spectral band, have proven to provide various
benefits, which make these technologies promising candidates for the next generation
wireless systems and beyond. With this motivation, investigating the potentials of massive
MIMO and mmWave communications in providing physical layer security has been of
recent interest (see e.g., [112]-[114] and [115]-[117]). Besides opportunities offered,
many challenges arise in secure transmit signal designs with finite-alphabet inputs. For
instance, most of the existing precoder design approaches (such as the ones proposed
in [33] and [40]) are intractable when the number of antennas is high. Furthermore,
communicating at the mmWave spectrum considerably changes the channel model with
respect to the microwave communication systems, and introduces some new features
(e.g., high sensitivity to blockages and considerable differences between the line-of-sight
and non-line-of-sight receptions), which motivates development of new frameworks for
designing practical secure transmission schemes.
The main problem associated with designing practical coding schemes for physical
layer security is the following: Although security gap is proposed and used as an
alternative practical metric, the most widely accepted one (i.e., equivocation) relies on
information theoretic quantities, and computing these quantities for finite-length codes
over AWGN channels is a challenging problem, and can be a promising research direction
on physical layer security. As a final suggestion for future research, we recommend the
problem of precoder design in the scenarios where practical wiretap codes are employed,
which would be an important step towards designing implementable secure transmission
schemes.
February 20, 2018
DRAFT
44
ACKNOWLEDGMENT
This work was supported by the Scientific and Technical Research Council of Turkey
(TUBITAK) under grant #113E223.
R EFERENCES
[1] Y. Liang, H. V. Poor, and S. Shamai, Information theoretic security, Found. Trends Commun. Inf. Theory, vol.
5, no. 4-5, pp. 355-580, 2008.
[2] R. Liu and W. Trappe, Securing wireless communications at the physical layer, Springer, 2010.
[3] E. A. Jorswieck, A. Wolf, and S. Gerbracht, “Trends in telecommunications technologies,” in Secrecy on the
Physical Layer in Wireless Networks. Croatia, InTech, 2010.
[4] M. Bloch and J. Barros, Physical-layer security: from information theory to security engineering, Cambridge,
2011.
[5] A. Mukherjee, S. A. A. Fakoorian, J. Huang, and A. L. Swindlehurst, “Principles of physical layer security
in multiuser wireless networks: A survey,” IEEE Commun. Surveys Tuts., vol. 16, no. 3, pp. 1550–1573, Aug.
2014.
[6] W. Trappe, “The challenges facing physical layer security,” IEEE Commun. Mag., vol. 53, no. 6, pp. 16–20,
Jun. 2015.
[7] A. Hyadi, Z. Rezki and M. S. Alouini, “An overview of physical layer security in wireless communication
systems with CSIT uncertainty,” in IEEE Access, vol. 4, pp. 6121–6132, Oct. 2016.
[8] X. Chen, D. W. K. Ng, W. H. Gerstacker and H. H. Chen, “A survey on multiple-antenna techniques for
physical layer security,” in IEEE Commun. Surveys Tuts., vol. 19, no. 2, pp. 1027–1053, Second Quarter 2017.
[9] M. Bellare, S. Tessaro, and A. Vardy, “Semantic security for the wiretap channel,” in Proc. 32nd Annu. Cryptol.
Conf., vol. 7417. 2012, pp. 294–311.
[10] A. D. Wyner, “The wire-tap channel,” Bell Sys. Tech. J., vol. 54, pp. 1355–1387, 1975.
[11] C. E. Shannon, “Communication theory of secrecy systems,” Bell Sys. Tech. J., vol. 28, pp. 656–715, 1949.
[12] U. M. Maurer and S. Wolf, “From weak to strong information-theoretic key agreement,” in IEEE Int’l. Symp.
Info. Theory (ISIT), p. 18, Sorrento, Italy, Jun. 2000.
[13] S. L. Y. Cheong and M. Hellman, “The Gaussian wire-tap channel,” IEEE Trans. Inf. Theory, vol. 24, no. 4,
pp. 451–456, July 1978.
[14] I. Csiszar and J. Korner, “Broadcast channels with confidential messages,” IEEE Trans. Inf. Theory, vol. 24,
no. 3, pp. 339–348, May 1978.
[15] J. Barros and M. R. D. Rodrigues, “Secrecy Capacity of Wireless Channels,” IEEE Int’l. Symp. Info. Theory
(ISIT), July 2006, pp. 356–360.
February 20, 2018
DRAFT
45
[16] P. Gopala, L. Lai, and H. El Gamal, “On the secrecy capacity of fading channels”, IEEE Trans. Inf. Theory,
vol. 54, no. 10, pp. 4687–4698, Oct. 2008.
[17] J. Li and A. P. Petropulu, “On ergodic secrecy rate for Gaussian MISO wiretap channels,” IEEE Trans. Wireless
Commun., vol. 10, no. 4, pp. 1176-–1187, Apr. 2011.
[18] P.-H. Lin and E. Jorswieck, “On the fast fading Gaussian wiretap channel with statistical channel state
information at the transmitter,” IEEE Trans. Inf. Forensics Security, vol. 11, no. 1, pp. 46–58, Jan. 2016.
[19] A. Khisti and G. Wornell, “Secure transmission with multiple antennas II: the MIMOME wiretap channel,”
IEEE Trans. Inf. Theory, vol. 56, no. 11, pp. 5515–5532, Nov. 2010.
[20] F. Oggier, B. Hassibi, “The secrecy capacity of the MIMO wiretap channel,” IEEE Trans. Inf. Theory, vol. 57,
no. 8, pp. 4961–4972, 2011.
[21] T. Liu and S. Shamai, “A note on the secrecy capacity of the multiple antenna wiretap channel,” IEEE Trans.
Inf. Theory, vol. 55, no. 6, pp. 2547–2553, Jun. 2009.
[22] Q. Li, M. Hong, H. T. Wai, Y. F. Liu, W. K. Ma and Z. Q. Luo, “Transmit solutions for MIMO wiretap
channels using alternating optimization,” in IEEE J. Sel. Areas Commun., vol. 31, no. 9, pp. 1714-1727, Sep.
2013.
[23] S. Loyka and C. D. Charalambous, “An algorithm for global maximization of secrecy rates in Gaussian MIMO
wiretap channels,” in IEEE Trans. Commun., vol. 63, no. 6, pp. 2288-2299, Jun. 2015.
[24] D. Klinc, J. Ha, S. W. McLaughlin, J. Barros, and B.-J. Kwak, “LDPC codes for the Gaussian wiretap channel,”
IEEE Trans. Inf. Forensics Security, vol. 6, no. 3, pp. 532–540, Sep. 2011.
[25] W.-C. Liao, T.-H. Chang, W.-K. Ma, and C.-Y. Chi, “QoS-based transmit beamforming in the presence of
eavesdroppers: An optimized artificial-noise aided approach,” IEEE Trans. Signal Processing, vol. 59, no. 3,
pp. 1202–1216, Mar. 2011.
[26] G. D. Raghava and B. S. Rajan, “Secrecy capacity of the Gaussian wiretap channel with finite complex
constellation input,” [Online]. Available: http://arxiv.org/abs/1010.1163, Oct. 2010.
[27] M. R. D. Rodrigues, A. Somekh-Baruch, and M. Bloch, “On Gaussian wiretap channels with M-PAM inputs,”
in Proc. EWConf., Apr. 2010, pp. 774–781.
[28] E. Biglieri, Coding for wireless channels, Springer-Verlag New York, 2005.
[29] C. Qi, Y. Chen, and A. J. Vinck. “On the Binary Input Gaussian Wiretap Channel with/without Output
Quantization,” Entropy, vol. 19, no. 2, Feb. 2017.
[30] Z. Li, R. Yates and W. Trappe, “Achieving secret communication for fast Rayleigh fading channels,” in IEEE
Trans. Wireless Commun., vol. 9, no. 9, pp. 2792-2799, Sep. 2010.
[31] S. Goel and R. Negi, “Guaranteeing secrecy using artificial noise,” IEEE Trans. Wireless Commun., vol. 7, no.
6, pp. 2180–2189, Jun. 2008.
[32] S. Bashar, Z. Ding, and C. Xiao, “On secrecy rate analysis of MIMO wiretap channels driven by finite-alphabet
input,” IEEE Trans. Commun., vol. 60, no. 12, pp. 3816–3825, Dec. 2012.
February 20, 2018
DRAFT
46
[33] Y. Wu, C. Xiao, Z. Ding, X. Gao, and S. Jin, “Linear precoding for finite-alphabet signaling over MIMOME
wiretap channels,” IEEE Trans. Veh. Technol., vol. 61, no. 6, pp. 2599–2612, Jul. 2012.
[34] A. Khisti and G. W. Wornell, “Secure transmission with multiple antennas—Part I: The MISOME wiretap
channel,” IEEE Trans. Inf. Theory, vol. 56, no. 7, pp. 3088–3104, Jul. 2010.
[35] M. R. A. Khandaker, C. Masouros, K. Wong, “Constructive interference based secure precoding: a new
dimension in physical layer security,” [Online]. Available: arxiv.org/abs/1612.08465, Dec. 2016.
[36] H. Reboredo, J. Xavier, and M. R. D. Rodrigues, “Filter design with secrecy constrains: The MIMO Gaussian
wiretap channel,“IEEE Trans. Signal Process., vol. 61, pp. 3799–3814, Aug. 2013.
[37] T. V. Nguyen, T. Q. S. Quek, Y. H. Kim, and H. Shin, “Secrecy diversity in MISOME wiretap channels,” in
Proc. Global Commun. Conf. (Globecom’2012), Aneheim, USA, Dec. 2012, pp. 4840–4845.
[38] S. Bashar, Z. Ding, and C. Xiao, “On the secrecy rate of multi-antenna wiretap channel under finite-alphabet
input,” IEEE Commun. Lett., vol. 15, no. 5, pp. 527-–529, May 2011.
[39] S. Rezaei Aghdam and T. M. Duman, “Low complexity precoding for MIMOME wiretap channels based on
cut-off rate,” IEEE Int. Symp. Inf. Theory (ISIT 2016), Barcelona, Jul. 2016, pp. 2988–2992.
[40] S. Rezaei Aghdam, T. M. Duman, “Joint precoder and artificial noise design for MIMO wiretap channels with
finite-alphabet inputs based on the cut-off rate,” IEEE Trans. Wireless Commun., vol. 16, no. 6, pp. 3913–3923,
Jun. 2017.
[41] P.-H. Lin S.-H. Lai, S.-C. Lin, and H.-J. Su, “On secrecy rate of the generalized artificial-noise assisted secure
beamforming for wiretap channels,” IEEE J. Sel. Areas Commun., vol. 31, no. 9, pp. 1728–1740, Sep. 2013.
[42] A. Kalantari, M. Soltanalian, S. Maleki, S. Chatzinotas, and B. Ottersten, “Secure M-PSK communication via
directional modulation,” in Proc. IEEE ICASSP, Shanghai, China, Mar. 2016, pp. 3481—3485
[43] A. Kalantari, M. Soltanalian, S. Maleki, S. Chatzinotas and B. Ottersten, “Directional modulation via symbollevel precoding: A way to enhance security,” in IEEE J. Sel. Topics Signal Process., vol. 10, no. 8, pp.
1478-1493, Dec. 2016.
[44] L. Zhang, Y. Cai, B. Champagne, and M. Zhao, “Tomlinson-Harashima precoding design in MIMO wiretap
channels based on the MMSE criterion,” in Proc. IEEE ICC, pp. 470-474, Jun. 2015.
[45] S. A. A. Fakoorian, H. Jafarkhani, and A. L. Swindlehurst, “Secure space-time block coding via artificial
noise alignment,” in Proc. Conf. Rec. 45th ASILOMAR (ASILOMAR’2011), Pacific Grove, USA, Nov. 2011,
pp. 651–655.
[46] M. A. Girnyk, M. Vehkapera, J. Yuan, and L. K. Rasmussen, “On the ergodic secrecy capacity of MIMO wiretap
channels with statistical CSI,” in Proc. Int. Symp. Inf. Theory Appl. (ISITA), Melbourne, VIC, Australia, Oct.
2014, pp. 398–402.
[47] S. Rezaei Aghdam, T. M. Duman, “Transmit signal design for MIMO wiretap channels with statistical CSI
and arbitrary inputs,” accepted in IEEE PIMRC 2017.
February 20, 2018
DRAFT
47
[48] X. Li, R. Fan, X. Ma, J. An, and T. Jiang, “Secure space-time communications over Rayleigh flat fading
channels,” IEEE Trans. Wireless Commun., vol. 15, pp. 1491–1504, Feb. 2016.
[49] H. M. Furqan, J. M. Hamamreh and H. Arslan, “Secret key generation using channel quantization with SVD
for reciprocal MIMO channels,” ISWCS, Poznan, Sep. 2016, pp. 597–602.
[50] W. Zeng, C. Xiao, M. Wang, and J. Lu, “Linear precoding for finite-alphabet inputs over MIMO fading channels
with statistical CSI,” IEEE Trans. Signal Process., vol. 60, no. 7, Jul. 2012.
[51] M. A. Girnyk, F. Gabry, M. Vehkapera, L. K. Rasmussen, and M. Skoglund, “On the transmit beamforming for
MIMO wiretap channels: Large-system analysis,” in Proc. Int. Conf. Inf. Theoretic Secur. (ICITS), Singapore,
Nov. 2013, pp. 90–102.
[52] M. A. Girnyk, F. Gabry, M. Vehkapera, L. K. Rasmussen and M. Skoglund, “MIMO wiretap channels
with randomly located eavesdroppers: Large-system analysis,” 2015 IEEE International Conference on
Communication Workshop (ICCW), London, Jun. 2015, pp. 480-484.
[53] Y. Wu, J. Wang, J. Wang, R. Schober, and C. Xiao, “Large-scale MIMO secure transmission with finite alphabet
inputs,” [Online]. Available: https://arxiv.org/pdf/1612.08328.pdf, Dec. 2016.
[54] R. Mesleh, H. Haas, S. Sinanovic, C. W. Ahn, and S. Yun, “Spatial modulation,” IEEE Trans. on Vehic.
Technol., vol. 57, no. 4, pp. 2228–2241, 2008.
[55] J. Jeganathan, A. Ghrayeb, L. Szeczecinski, A. Ceron, “Space shift keying modulation for MIMO channels,”
IEEE Trans. Wireless Commun. vol. 8, no. 7, pp. 3692-3703, Jul. 2009.
[56] M. Di Renzo, H. Haas, A. Ghrayeb, S. Suguira and L. Hanzo “Spatial modulation for generalized MIMO:
challenges, opportunities and implementation,” Proc. IEEE, vol. 102, no. 1, pp. 56–103, 2014.
[57] S. Sinanovic, N. Serafimovski, M. Di Renzo, and H. Haas, “Secrecy capacity of space keying with two
antennas,” in 2012 IEEE Veh. Technol. Conf. (VTC Fall), Sep. 2012, pp. 1–5.
[58] X. Guan, Y. Cai, and W. Yang, “On the secrecy mutual information of spatial modulation with finite alphabet,”
in Proc. IEEE Int. Conf. Wireless Commun. Signal Process. (WCSP), Oct. 2012, pp. 1–4
[59] S. Rezaei Aghdam, T. M. Duman and M. Di Renzo, “On secrecy rate analysis of spatial modulation and space
shift keying,”, IEEE BlackSeaCom 2015, pp. 63–67, May 2015.
[60] S. Rezaei Aghdam and T. M. Duman, “Physical layer security for space shift keying transmission with
precoding,” IEEE Wireless Commun. Lett., vol. 5, no. 2, pp. 180-–183, Apr. 2016.
[61] F. Wu, L. L. Yang, W. Wang and Z. Kong, “Secret precoding-aided spatial modulation,” in IEEE Commun.
Lett., vol. 19, no. 9, pp. 1544–1547, Sep. 2015.
[62] L. Wang, S. Bashar, Y. Wei, and R. Li, “Secrecy enhancement analysis against unknown eavesdropping in
spatial modulation,” IEEE Commun. Lett., vol. 19, no. 8, pp. 1351—1354, Aug. 2015.
[63] Z. Huang, Z. Gao and L. Sun, “Anti-eavesdropping scheme based on quadrature spatial modulation,” in IEEE
Commun. Lett., vol. 21, no. 3, pp. 532-535, Mar. 2017.
February 20, 2018
DRAFT
48
[64] Q.-L. Li, “Information-guided randomization for wireless physical layer secure transmission,” in Proc. IEEE
Military Commun. Conf., Orlando, FL, USA, Nov. 2012, pp. 1–6.
[65] S. Rezaei Aghdam and T. M. Duman, “Secure space shift keying transmission using dynamic antenna index
assignment,” accepted in IEEE GLOBECOM 2017, Dec. 2017.
[66] Y. Wei, L. Wang and T. Svensson, “Analysis of secrecy rate against eavesdroppers in MIMO modulation
systems,” 2015 Int. Conf. Wireless Commun. Signal Process. (WCSP), Nanjing, Oct. 2015, pp. 1-5.
[67] Z. Li, R. D. Yates, and W. Trappe, “Secrecy capacity of independent parallel channels,” in 44th Annual Allerton
CCC, Sep. 2006.
[68] F. Renna, N. Laurenti, and H. V. Poor, “Physical-layer secrecy for OFDM transmissions over fading channels,”
IEEE Trans. Inf. Forens. Security, vol. 7, no. 4, pp. 1354–1367, Aug. 2012.
[69] F. Renna, N. Laurenti, and H. V. Poor, “Achievable secrecy rates for wiretap OFDM with QAM constellations,”
in Proc. 5th Int. ICST Conf. Perform. Eval. Method. VALUETOOLS, Paris, France, 2011, pp. 679—686.
[70] M. Yusuf and H. Arslan, “Controlled inter-carrier interference for physical layer security in OFDM systems,”
2016 IEEE Veh. Technol. Conf. (VTC Fall), Montreal, QC, 2016, pp. 1-5.
[71] H. Li, X. Wang and J. Y. Chouinard, ”Eavesdropping-resilient OFDM system using sorted subcarrier
interleaving,” in IEEE Trans. Wireless Commun., vol. 14, no. 2, pp. 1155-1165, Feb. 2015.
[72] H. Qin, Y. Sun, T. Chang, X. Chen,; C. Chi, M. Zhao, J. Wang, “Power allocation and time-domain artificial
noise design for wiretap OFDM with discrete inputs,” IEEE Trans. Wireless Commun., vol. 12, no. 6, pp.
2717—2729, Jun. 2013.
[73] T. Akitaya, S. Asano, and T. Saba, “Time-domain artificial noise generation technique using time-domain and
frequency-domain processing for physical layer security in MIMO-OFDM systems,” in Proc. IEEE Int. Conf.
Commun. Workshops (ICCW), Jun. 2014, pp. 807-–812.
[74] Z. Mheich, F. Alberge, and P. Duhamel, “Achievable secrecy rates for the broadcast channel with confidential
message and finite constellation inputs,” IEEE Trans. Commun., vol. 63, no. 1, pp. 195–205, Jan. 2015.
[75] T. Allen, A. Tajer and N. Al-Dhahir, “Secure Alamouti MAC transmissions,” in IEEE Trans. Wireless Commun.,
vol. 16, no. 6, pp. 3674-3687, Jun. 2017.
[76] J. Jin, C. Xiao, M. Tao and W. Chen, “Linear precoding for cognitive multiple access wiretap channel with
finite-alphabet inputs,” 2016 IEEE ICC, Kuala Lumpur, May 2016, pp. 1-6.
[77] W. Zeng, Y. R. Zheng and C. Xiao, “Multiantenna secure cognitive radio networks with finite-alphabet inputs:
a global optimization approach for precoder design,” in IEEE Trans. Wireless Commun., vol. 15, no. 4, pp.
3044-3057, Apr. 2016.
[78] K. Cao, Y. Cai, Y. Wu, W. Yang and X. Guan, “Secure communication for MISO secrecy channel with multiple
multiantenna eavesdroppers having finite alphabet inputs,” in IEEE Access, vol. PP, no. 99, pp. 1-1.
[79] S. Vishwakarma and A. Chockalingam, “Power allocation in MIMO wiretap channel with statistical CSI and
finite-alphabet input,” 2014 SPCOM, Bangalore, Jul. 2014, pp. 1-6.
February 20, 2018
DRAFT
49
[80] S. Vishwakarma and A. Chockalingam, “Decode-and-forward relay beamforming for secrecy with finitealphabet input,” in IEEE Commun. Lett., vol. 17, no. 5, pp. 912-915, May 2013.
[81] K. Cao, Y. Cai, Y. Wu and W. Yang, “Cooperative jamming for secure communication with finite alphabet
inputs,” in IEEE Commun. Lett., vol. 21, no. 9, pp. 2025-2028, Sep. 2017.
[82] K. Cao, Y. Wu, Y. Cai and W. Yang, ”Secure transmission with aid of a helper for MIMOME network having
finite alphabet inputs,” in IEEE Access, vol. 5, pp. 3698-3708, Feb. 2017.
[83] M. R. Bloch, M. Hayashi, and A. Thangaraj, “Error-control coding for physical-layer secrecy,” Proceedings of
IEEE, vol. 103, no. 10, pp. 1725–1746, Oct. 2015.
[84] L. H. Ozarow and A. D. Wyner, “Wire-tap channel II,” Bell Lab. Tech. J., vol. 63, no. 10, pp. 2135-2157, Dec.
1984.
[85] G. Cohen and G. Zemor, “Syndrome-coding for the wiretap channel revisited,” in Proc. IEEE Information
Theory Workshop, Chengdu, China, Oct. 2006, pp. 33–36.
[86] F. Oggier, P. Sole and J. Belfiore, “Lattice codes for the wiretap Gaussian channel: construction and analysis,”
IEEE Trans. Inf. Theory, vol. 62, no. 10, pp. 5690-5707, Oct. 2016.
[87] C. Ling, L. Luzzi, J.-C. Belfiore, and D. Stehlé, “Semantically secure lattice codes for the Gaussian wiretap
channel,” Computing Research Repository, Oct. 2012, pp. 1–19.
[88] D. Klinc, J. Ha, S. M. McLaughlin, J. Barros, and B.-J. Kwak, “LDPC codes for the Gaussian wiretap channel,”
IEEE Trans. Inf. Forensics Security, vol. 6, no. 3, pp. 532-540, Sep. 2011.
[89] I. M. Kim, B. H. Kim, and J. K. Ahn, “BER-based physical layer security with finite code length: Combining
strong converse and error amplification,” IEEE Trans. Commun., vol. 64, pp. 3844-3857, Sep. 2016.
[90] M. Baldi, M. Bianchi, and F. Chiaraluce, “Non-systematic codes for physical layer security,” in Proc. IEEE
Information Theory Workshop (ITW 2010), Dublin, Ireland, Aug. 2010.
[91] M. Baldi, F. Bambozzi, and F. Chiaraluce, “Coding with scrambling, concatenation, and HARQ for the AWGN
wire-tap channel: A security gap analysis,” IEEE Trans. Inf. Forensics Security, vol. 7, no. 3, pp. 883—894,
Jun. 2012.
[92] A. Nooraiepour and T. M. Duman, “Randomized convolutional codes for the wiretap channel,” in IEEE Trans.
Commun., vol. 65, no. 8, pp. 3442-3452, Aug. 2017.
[93] Y. X. Zhang, A. Liu, “Polar-LDPC concatenated coding for the AWGN wiretap channel,” IEEE Commun.
Lett., vol. 18, no. 10, pp. 1683-1686, Oct. 2014.
[94] A. Nooraiepour and T. M. Duman “Randomized turbo codes for the wiretap channel,” in IEEE GLOBECOM,
Singapore, Dec. 2017.
[95] A. Nooraiepour and Tolga M. Duman, “Randomized serially concatenated LDGM codes for the Gaussian
wiretap channel,” Accepted for publication in IEEE Commun. Lett., Dec. 2017.
[96] C. W. Wong, T. F. Wong, and J. M. Shea, “LDPC code design for the BPSK-constrained Gaussian wiretap
channel,” in Proc. IEEE GLOBECOM 2011 Workshops, Houston, TX, Dec. 2011, pp. 898–902.
February 20, 2018
DRAFT
50
[97] M. Baldi, G. Ricciutelli, N. Maturo, and F. Chiaraluce, “Performance assessment and design of finite length
LDPC codes for the gaussian wiretap channel,” IEEE ICC 2015 Workshop on Physical-Layer Security, Jun.
2015, pp. 435-440.
[98] H. Tyagi and A. Vardy, “Explicit capacity-achieving coding scheme for the Gaussian wire- tap channel,” in
IEEE Int. Symp. Inf. Theory, Jul. 2014, pp. 956—960.
[99] W. Yang, R. F. Schaefer, and H. V. Poor, “Finite-blocklength bounds for wiretap channels,” in IEEE Int. Symp.
Inf. Theory, Barcelona, Spain, pp. 3087—3091, Jul. 2016.
[100] M. Zheng, M. Tao, and W. Chen., “Polar coding for secure transmission in MISO fading wiretap channels.”
[Online]. Available: http://arxiv.org/abs/1411.2463, Nov. 2014.
[101] H. Si, O. O. Koyluoglu and S. Vishwanath, “Hierarchical Polar Coding for Achieving Secrecy Over StateDependent Wiretap Channels Without Any Instantaneous CSI,” in IEEE Trans. Commun., vol. 64, no. 9, pp.
3609-3623, Sep. 2016.
[102] L. Luzzi, C. Ling and R. Vehkalahti, “Almost universal codes for fading wiretap channels,” IEEE Int. Symp.
Inf. Theory (ISIT), Barcelona, Jul. 2016, pp. 3082-3086.
[103] S. Shamai and S. Verdu, “Worst-case power-constrained noise for binary input channels,” IEEE Trans. Inf.
Theory, vol. 38, no. 5, pp. 1494- 1511, Sep. 1992.
[104] X. Chen, D. W. K. Ng, and H.-H. Chen, “Secrecy wireless information and power transfer: challenges and
opportunities,” IEEE Commun. Mag., May 2016.
[105] R. Zhang and C. K. Ho, “MIMO broadcasting for simultaneous wireless information and power transfer,” IEEE
Trans. Wireless Commun., vol. 12, no. 5, pp. 1989-–2001, May 2013.
[106] J. Zhang, C. Yuen, C. K. Wen, S. Jin, K. K. Wong and H. Zhu, “Large system secrecy rate analysis for SWIPT
MIMO wiretap channels,” IEEE Trans. Inf. Forensics Security, vol. 11, no. 1, pp. 74–85, Jan. 2016.
[107] K. Banawan and S. Ulukus, “MIMO wiretap channel under receiver side power constraints with applications
to wireless information transfer and cognitive radio,” IEEE Trans. on Commun., vol. 64, no. 9, pp. 3872–3885,
Sep. 2016.
[108] H. M. Wang, T. X. Zheng, J. Yuan, D. Towsley and M. H. Lee, “Physical layer security in heterogeneous
cellular networks,” in IEEE Trans. Commun., vol. 64, no. 3, pp. 1204-1219, Mar. 2016.
[109] Y. Zhou, Z. Z. Xiang, Y. Zhu, and Z. Xue, “Application of full-duplex wireless technique into secure MIMO
communication: Achievable secrecy rate based optimization,” IEEE Signal Process. Lett., vol. 21, pp. 804–808,
Jul. 2014.
[110] E. Larsson et al., “Massive MIMO for next generation wireless systems,” IEEE Commun. Mag., vol. 52, no.
2, Feb. 2014, pp. 186–95.
[111] T. Rappaport et al., “Millimeter wave mobile communications for 5G cellular: It will work!,” IEEE Access,
vol. 1, May 2013, pp. 335–49.
February 20, 2018
DRAFT
51
[112] J. Wang, J. Lee, F. Wang, and T. Q. S. Quek, “Jamming-aided secure communication in massive MIMO Rician
channels,” IEEE Trans. Commun., vol. 64, pp. 6854–6868, Dec. 2015.
[113] K. Guo, Y. Guo, and G. Ascheid, “Security-constrained power allocation in mu-massive-MIMO with distributed
antennas,” IEEE Trans. Wireless Commun., vol. 15, pp. 8139–8153, Dec. 2016.
[114] J. Zhu, R. Schober, and V. K. Bhargava, “Linear precoding of data and artificial noise in secure massive MIMO
systems,” IEEE Trans. Wireless Commun., vol. 15, pp. 2245–2261, Mar. 2016.
[115] S. Vuppala, S. Biswas, and T. Ratnarajah, “An analysis on secure communication in millimeter/micro-wave
hybrid networks,” IEEE Trans. Commun., vol. 64, pp. 3507–3519, Aug. 2016.
[116] C. Wang and H.-M. Wang, “Physical layer security in millimeter wave cellular networks,” IEEE Trans. Wireless
Commun., vol. 15, pp. 5569–5585, Aug. 2016.
[117] Y. Zhu, L. Wang, K. K. Wong and R. W. Heath, “Secure Communications in Millimeter Wave Ad Hoc
Networks,” IEEE Trans. Wireless Commun., vol. 16, no. 5, pp. 3205-3217, May 2017.
February 20, 2018
DRAFT
| 7cs.IT
|
Consistent feature attribution for tree ensembles
Scott M. Lundberg
SLUND 1@ CS . WASHINGTON . EDU
Paul G. Allen School of Computer Science, University of Washington, Seattle, WA 98105 USA
arXiv:1706.06060v6 [cs.AI] 17 Feb 2018
Su-In Lee
SUINLEE @ CS . WASHINGTON . EDU
Paul G. Allen School of Computer Science and Department of Genome Sciences, University of Washington, Seattle, WA
98105 USA
Abstract
It is critical in many applications to understand
what features are important for a model, and why
individual predictions were made. For tree ensemble methods these questions are usually answered by attributing importance values to input
features, either globally or for a single prediction.
Here we show that current feature attribution
methods are inconsistent, which means changing the model to rely more on a given feature
can actually decrease the importance assigned
to that feature. To address this problem we develop fast exact solutions for SHAP (SHapley
Additive exPlanation) values, which were recently shown to be the unique additive feature
attribution method based on conditional expectations that is both consistent and locally accurate. We integrate these improvements into the
latest version of XGBoost, demonstrate the inconsistencies of current methods, and show how
using SHAP values results in significantly improved supervised clustering performance. Feature importance values are a key part of understanding widely used models such as gradient
boosting trees and random forests. We believe
our work improves on the state-of-the-art in important ways, and so impacts any current user of
tree ensemble methods.
1. Introduction
Understanding why a model made a prediction is important for trust, actionability, accountability, debugging, and
many other common tasks. To understand predictions from
tree ensemble methods, such as gradient boosting trees or
random forests, importance values are typically attributed
to each input feature. These importance values can be computed either for a single prediction, or an entire dataset to
explain a model’s overall behavior.
Concerningly, current feature attribution methods for tree
ensembles are inconsistent, meaning they can assign higher
importance to features with a lower impact on the model’s
output. This inconsistency effects a very large number of
users, since tree ensemble methods are widely applied in
research and industry.
Here we show that by connecting tree ensemble feature attribution methods with the recently defined class of additive feature attribution methods (Lundberg & Lee, 2017)
we can motivate the use of SHapley Additive exPlanation
(SHAP) values as the only possible consistent feature attribution method with desirable properties.
SHAP values are theoretically optimal but can be challenging to compute. To address this we derive exact algorithms
for tree ensemble methods that reduce the computational
complexity of computing SHAP values from exponential
to O(T LD2 ) where T is the number of trees, L is the maximum number of leaves in any tree, and D is the maximum
depth of any tree. By integrating this new algorithm into
XGBoost, a popular tree ensemble package, we demonstrate performance that enables predictions from models
with thousands of trees, and hundreds of inputs, to be explained in a fraction of a second.
In what follows we first discuss the inconsistencies of current feature attribution methods as implemented in popular
tree enemble software packages (Section 2). We then introduce SHAP values as the only possible consistent attributions (Section 3), and present Tree SHAP as a high speed
algorithm for estimating SHAP values of tree ensembles
(Section 4). Finally, we use a supervised clustering task
to compare SHAP values with previous feature attribution
methods (Section 5).
Consistent feature attribution for tree ensembles
(A)
(B)
(Fever = Yes, Cough = Yes)
(Fever = Yes, Cough = Yes)
Fever
Cough
No
Yes
Cough
No
Cough
Yes
Fever
Fever
No
Yes
No
Yes
No
Yes
No
Yes
0
0
0
100
0
0
10
110
output = [Cough & Fever]*100
output = [Cough & Fever]*100 + [Cough]*10
Attributions
Path (output)
SHAP (output)
Split count
Path (gain)
SHAP (gain)
Fever
25
37.5
1
33
50
Cough
50
37.5
2
67
50
Attributions
Inconsistency
Inconsistency
Fever
50
37.5
2
58
37
Cough
30
47.5
1
42
63
Figure 1. Two tree models meant to demonstrate the inconsistencies of current feature attribution methods. The Cough feature has a
larger impact on tree B, but is assigned less importance by all three standard methods. The “output” attributions explain the difference
between the expected value of the model output and the current output. The “gain” represents the change in the mean squared error over
the whole dataset between when no features are used and all features are used. All calculations assume a dataset (typically a training
dataset) perfectly matching the model and evenly spread among all leaves. Section 2 describes the standard “path” methods, while
Section 3 describes the SHAP values and their interpretation.
2. Current feature attributions are
inconsistent
Tree ensemble implementations in popular packages such
as XGBoost (Chen & Guestrin, 2016), scikit-learn (Pedregosa et al., 2011), and the gbm R package (Ridgeway,
2010), allow a user compute a measure of feature importance. These values are meant to summarize a complicated
ensemble model and provide insight into what features
drive the model’s prediction. Unfortunately the standard
feature importance values provided by all of these packages are inconsistent, this means that a model can change
such that it relies more on a given feature, yet the importance assigned to that feature decreases (Figure 1).
For the above packages, when feature importance values
are calculated for an entire dataset they are by default based
on the reduction of loss (termed “gain”) contributed by
each split in each tree of the ensemble. Feature importances are then defined as the sum of the gains of all splits
for a given feature as described in Friedman et al. (Breiman
et al., 1984; Friedman et al., 2001).
Methods computing feature importance values for a single prediction are less established, and of the above packages, only the most recent version of XGBoost supports
these calculations natively. The method used by XGBoost
(Saabas) is similar to the classical dataset level feature importance calculation, but instead of measuring the reduction of loss it measures the change in the model’s output.
Both current feature attribution methods described above
only consider the effect of splits along the decision path, so
we will term them path methods. Figure 1 shows the result
of applying both these methods to two simple regression
trees. For the gain calculations we assume equal coverage
of each of the four tree leaves, and perfect regression accuracy. In other words, an equal number of dataset points fall
in each leaf, and the label of those points is exactly equal
to the prediction of the leaf. The tree in Figure 1A represents a simple AND function, while the tree in Figure 1B
represents the same AND function but with an additional
increase in predicted value when Cough is “Yes”.
The point of Figure 1 is to compare feature attributions between A and B, where it is clear that Cough has a larger impact on the model in B than the model in A. As highlighted
below each tree, we can see that current path methods (as
well as the simple split count metric) are inconsistent because they allocate less importance to Cough in B, even
though Cough has a larger impact on the output of the tree
in B. The “output” task explains the change in model output from the expected value to the current predicted value
given Fever and Cough. The “gain” explains the reduction
in mean squared error contributed by each feature (assuming a dataset as described in the previous paragraph). In
contrast to current approaches, the SHAP values (described
below) are consistent, even when the order in which features appear in the tree changes.
Consistent feature attribution for tree ensembles
3. SHAP values are the only consistent feature
attributions
It was recently noted that many current methods for interpreting machine learning model predictions fall into the
class of additive feature attribution methods (Lundberg &
Lee, 2017). This class covers all methods that explain a
model’s output as a sum of real values attributed to each
input feature.
Definition 1 Additive feature attribution methods have
an explanation model that is a linear function of binary
variables:
M
X
g(z 0 ) = φ0 +
φi zi0 ,
(1)
i=1
where z 0 ∈ {0, 1}M , M is the number of input features,
and φi ∈ R.
The zi0 variables typically represent a feature being observed (zi0 = 1) or unknown (zi0 = 0), and the φi ’s are
the feature attribution values.
As previously described in Lundberg & Lee, an important
attribute of the class of additive feature attribution methods is that there is a single unique solution in this class
with three desirable properties: local accuracy, missingness, and consistency (Lundberg & Lee, 2017). Local accuracy states that the sum of the feature attributions is equal to
the output of the function we are seeking to explain. Missingness states that features that are already missing (such
that zi0 = 0) are attributed no importance. Consistency
states that changing a model so a feature has a larger impact
on the model, will never decrease the attribution assigned
to that feature.
In order to evaluate the effect missing features have on a
model f , it is necessary to define a mapping hx that maps
between the original function input space and the binary
pattern of missing features represented by z 0 . Given such a
0
mapping we can evaluate f (h−1
x (z )) and so calculate the
effect of observing or not observing a feature (by setting
zi0 = 1 or zi0 = 0).
0
SHAP values define fx (S) = f (h−1
x (z )) = E[f (x) | xS ]
where S is the set of non-zero indexes in z 0 (Figure 2), and
then use the classic Shapley values from game theory to
attribute φi values to each feature:
φi =
X
S⊆N \{i}
|S|!(M − |S|! − 1)
[fx (S ∪ {i}) − fx (S)]
M!
(2)
where N is the set of all input features.
The SHAP values are the only possible consistent, lo-
cally accurate method that obeys the missingness property
and uses conditional dependence to measure missingness
(Lundberg & Lee, 2017). This is strong motivation to use
SHAP values for tree ensemble feature attribution, particularly since current tree ensemble feature attribution methods already obey all of these properties except consistency.
This means that SHAP values provide a strict theoretical
improvement over existing approaches by eliminating the
unintuitive consistency problems shown in Figure 1.
4. Tree SHAP: Fast SHAP value computation
for decision trees
Despite the compelling theoretical advantages of SHAP
values, their practical use is hindered by two problems:
1. The challenge of estimating E[f (x) | xS ] efficiently.
2. The exponential complexity of Equation 2.
Here we focus on tree models and propose fast SHAP value
estimation methods specific to trees and ensembles of trees.
We start by defining a straightforward, but slow, algorithm
in Section 4.1, then present the much faster and more complex Tree SHAP algorithm in Section 4.2.
4.1. Estimating SHAP values directly in O(T L2M ) time
If we ignore computational complexity then we can compute the SHAP values for a decision tree by estimating
E[f (x) | xS ] and then using Equation 2 where fx (S) =
E[f (x) | xS ]. For a tree model E[f (x) | xS ] can be estimated recursively using Algorithm 1, where v is a vector
of node values, which takes the value internal for internal nodes. The vectors a and b represent the left and right
node indexes for each internal node. The vector t contains
the thresholds for each internal node, and d is a vector of
indexes of the features used for splitting in internal nodes.
The vector r represents the cover of each node (how many
data samples fall in that subtree).
4.2. Estimating SHAP values in O(T LD2 ) time
Here we propose a novel algorithm to calculate the same
values as in Section 4.1, but in polynomial time instead
of exponential time. Specifically, we propose an algorithm that runs in O(T L log2 L) for balanced trees, and
O(T LD2 ) for unbalanced trees.
The general idea of the polynomial time algorithm is to recursively keep track of what proportion of all possible subsets flow down into each of the leaves of the tree. This
is similar to running Algorithm 1 simultaneously for all
2M subsets S in Equation 2. It may seem reasonable to
simply keep track of how many subsets (weighted by the
Consistent feature attribution for tree ensembles
Figure 2. SHAP (SHapley Additive exPlanation) values explain the output of a function as a sum of the effects φi of each feature being
introduced into a conditional expectation. Importantly, for non-linear functions the order in which features are introduced matters, so
SHAP averages over all possible orderings. Proofs from game theory show this is the only possible consistent and locally accurate
approach. In contrast, standard path methods for tree ensembles (Section 2) are similar to using a single ordering defined by a tree’s
decision path.
Algorithm 1 Estimating E[f (x) | xS ]
procedure EXPVALUE(x, S, tree = {v, a, b, t, r, d})
procedure G(j, w)
if vj 6= internal then
return w · vj
else
if dj ∈ S then
return xdj ≤ tj ? G(aj , w) : G(bj , w)
else
return G(aj , wraj /rj ) + G(aj , wraj /rj )
end if
end if
end procedure
return G(1, 1)
end procedure
cover splitting of Algorithm 1) pass down each branch of
the tree. However, this combines subsets of different sizes
and so prevents the proper weighting of these subsets, since
the weights in Equation 2 depend on |S|. To address this
we keep track of each possible subset size during the recursion. The EXTEND method in Algorithm 2 grows all
these subsets according to given fraction of ones and zeros, while the UNWIND method reverses this process. The
EXTEND method is used as we descend the tree. The UNWIND method is used to undo previous extensions when
we split on the same feature twice, and to undo each extension of the path inside a leaf to correctly compute the
weights for each feature in the path.
In Algorithm 2, m is the path of unique features we have
split on so far, and contains four attributes: d the feature
index, z the fraction of “zero” paths (where this feature is
not in the set S) that flow through this branch, o the fraction of “one” paths (where this feature is in the set S) that
flow through this branch, and w which is used to hold the
proportion of sets of a given cardinally that are present. We
use the dot notation to access these members, and for the
whole vector m.d represents a vector of all the feature indexes. (For code see https://github.com/slundberg/shap)
5. Supervised clustering experiments
One intriguing use for prediction level feature attributions
is what we term “supervised clustering”, where instead of
using an unsupervised clustering method directly on the
data features, you run clustering on the feature attributions
(Lundberg & Lee, 2016).
Supervised clustering naturally handles one of the most
challenging problems in unsupervised clustering: determining feature weightings (or equivalently, determining a
distance metric). Many times we want to cluster data using features with very different units. Features may be in
dollars, meters, unit-less scores, etc. but whenever we use
them as dimensions in a single multidimensional space it
forces any distance metric to compare the relative importance of a change in different units (such as dollars vs. meters). Even if all our inputs are in the same units, often
some features are more important than others. Supervised
clustering uses feature attributions to naturally convert all
the input features into values with the same units as the
model output. This means that a unit change in any of the
feature attributions is comparable to a unit change in any
other feature attribution. It also means that fluctuations in
the feature values only effect the clustering if those fluctuations have an impact on the outcome of interest.
Here we compare feature attribution methods by applying
supervised clustering to disease sub-typing, an area where
unsupervised clustering has contributed to important discoveries. The goal of disease sub-typing is to identify subgroups of patients that have similar mechanisms of disease (similar reasons they are sick). Here we consider
Alzheimer’s disease where the predicted outcome is the
CERAD cognitive score (Mirra et al., 1991), and the features are gene expression modules (Celik et al., 2014).
By representing the positive feature attributions as red bars
and the negative feature attributions as blue bars (as in Figure 2), we can stack them against each other to visually
represent the model output as their sum. Figure 3 does
this vertically for each participant. The explanations for
each participant are then stacked horizontally according the
Consistent feature attribution for tree ensembles
Samples (patients) ordered by explanation similarity
Model output (Alzheimer’s score)
(A) Path explanations
Samples (patients) ordered by explanation similarity
Model output (Alzheimer’s score)
(B) SHAP explanations
Figure 3. SHAP feature attributions produce better clusters than standard path attributions for supervised clustering of 518 participants
in an Alzheimer’s research study. An XGBoost model with 300 trees of max depth six was trained on 200 gene expression module
features using a shrinkage factor of η = 0.01. This model was then used to predict the CERAD cognitive score of each participant. Each
prediction was explained, and then clustered using hierarchical agglomerative clustering (imagine a dendrogram joining the samples
above each plot). Red feature attributions push the score higher, while blue feature attributions push the score lower. A) The clusters
formed with standard “path” explanations from XGBoost. B) Clusters using our Tree SHAP XGBoost implementation.
Consistent feature attribution for tree ensembles
Algorithm 2 Tree SHAP
procedure TS(x, tree = {v, a, b, t, r, d})
φ = array of len(x) zeros
procedure RECURSE(j, m, pz , po , pi )
m = EXTEND(m, pz , po , pi )
if vj 6= internal then
for i ← 2 to len(m) do
w = sum(UNWIND(m, i).w)
φmi = φmi + w(mi .o − mi .z)vj
end for
else
h, c = xdj ≤ tj ? (aj , bj ) : (bj , aj )
iz = io = 1
k = FINDFIRST(m.d, dj )
if k 6= nothing then
iz , io = (mk .z, mk .o)
m = UNWIND(m, k)
end if
RECURSE(h, m, iz rh /rj , io , dj )
RECURSE(c, m, iz rc /rj , 0, dj )
end if
end procedure
procedure EXTEND(m, pz , po , pi )
l = len(m) + 1
m = copy(m)
ml+1 .(d, z, o, w) = (pi , pz , po , l = 0 ? 1 : 0)
for i ← l − 1 to 1 do
mi+1 .w = mi+1 .w + po mi .w(i/l)
mi .w = pz mi .w[(l − i)/l]
end for
return m
end procedure
procedure UNWIND(m, i)
l = len(m)
n = ml .w
m = copy(m1...l−1 )
for j ← l − 1 to 1 do
if mi .o 6= 0 then
t = mj .w
mj .w = n · l/(j · mi .o)
n = t − mj .w · mi .z((l − j)/l)
else
mj .w = (mj .w · l)/(mi .z(l − j))
end if
end for
for j ← i to l − 1 do
mj .(d, z, o) = mj+1 .(d, z, o)
end for
return m
end procedure
RECURSE(1, [], 1, 1, 0)
return φ
end procedure
Figure 4. A quantitative performance measure of the clusterings
shown in Figure 3. If all 518 samples are placed in their own
group, and each group predicts the mean value of the group, then
the R2 value (the proportion of outcome variance explained) will
be 1. If groups are then merged one-by-one the R2 will decline
until when there is only a single group it will be 0. Hierarchical clusterings that well separate the outcome value will retain a
high R2 longer during the merging process. Here unsupervised
clustering did no better than random, supervised clustering with
the XGBoost “path” method did significantly better, and SHAP
values significantly better still.
leaf order of a hierarchical clustering. This groups participants with similar predicted outcomes and similar reasons
for that predicted outcome together. The clearer structure
in Figure 3B indicates the SHAP values are better feature
attributions, not only theoretically, but also practically.
The improvement in clustering performance seen in Figure
3 can be quantified by examining how well each clustering
explains the variance of the CERAD score outcome. Since
hierarchical clusterings encode many possible groupings,
we plot in Figure 4 the change in the R2 value as the number of groups shrinks from one group per sample (R2 = 1),
to a single group (R2 = 0).
6. Conclusion
Here we have shown that classic feature attribution methods for tree ensembles are inconsistent, meaning they can
assign less importance to a feature when the true effect
of that feature increases. In contrast, SHAP values were
shown to be the unique way to consistently attribute feature importance. By deriving fast algorithms for SHAP values and integrating them with XGBoost, we make them a
practical replacement for previous methods. Future directions include deriving fast dataset-level SHAP algorithms
for gain (as opposed to the instance-level algorithm presented here), and integrating SHAP value algorithms into
the released versions of common packages.
Consistent feature attribution for tree ensembles
Acknowledgments
We would like to thank Gabriel Erion for suggestions that
lead to a simplified algorithm, as well as Jacob Schreiber
and Naozumi Hiranuma for providing helpful input.
References
Breiman, Leo, Friedman, Jerome, Stone, Charles J, and
Olshen, Richard A. Classification and regression trees.
CRC press, 1984.
Celik, Safiye, Logsdon, Benjamin, and Lee, Su-In. Efficient dimensionality reduction for high-dimensional network estimation. In International Conference on Machine Learning, pp. 1953–1961, 2014.
Chen, Tianqi and Guestrin, Carlos. Xgboost: A scalable
tree boosting system. In Proceedings of the 22Nd ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 785–794. ACM, 2016.
Friedman, Jerome, Hastie, Trevor, and Tibshirani, Robert.
The elements of statistical learning, volume 1. Springer
series in statistics Springer, Berlin, 2001.
Lundberg, Scott and Lee, Su-In. An unexpected unity
among methods for interpreting model predictions.
arXiv preprint arXiv:1611.07478, 2016.
Lundberg, Scott and Lee, Su-In. A unified approach
to interpreting model predictions.
arXiv preprint
arXiv:1705.07874, 2017.
Mirra, Suzanne S, Heyman, A, McKeel, D, Sumi, SM,
Crain, Barbara J, Brownlee, LM, Vogel, FS, Hughes,
JP, Van Belle, G, Berg, L, et al. The consortium to
establish a registry for alzheimer’s disease (cerad) part
ii. standardization of the neuropathologic assessment of
alzheimer’s disease. Neurology, 41(4):479–479, 1991.
Pedregosa, Fabian, Varoquaux, Gaël, Gramfort, Alexandre, Michel, Vincent, Thirion, Bertrand, Grisel, Olivier,
Blondel, Mathieu, Prettenhofer, Peter, Weiss, Ron,
Dubourg, Vincent, et al. Scikit-learn: Machine learning in python. Journal of Machine Learning Research,
12(Oct):2825–2830, 2011.
Ridgeway, Greg. Generalized boosted regression models. documentation on the r package gbm, version 1.6–3,
2010.
Saabas,
Ando.
Interpreting
random
forests.
http://blog.datadive.net/
interpreting-random-forests/. Accessed:
2017-06-15.
| 2cs.AI
|
Large-Scale Evolution of Image Classifiers
Esteban Real 1 Sherry Moore 1 Andrew Selle 1 Saurabh Saxena 1
Yutaka Leon Suematsu 2 Jie Tan 1 Quoc V. Le 1 Alexey Kurakin 1
arXiv:1703.01041v2 [cs.NE] 11 Jun 2017
Abstract
Neural networks have proven effective at solving difficult problems but designing their architectures can be challenging, even for image classification problems alone. Our goal is to minimize human participation, so we employ evolutionary algorithms to discover such networks
automatically. Despite significant computational
requirements, we show that it is now possible to
evolve models with accuracies within the range
of those published in the last year. Specifically, we employ simple evolutionary techniques
at unprecedented scales to discover models for
the CIFAR-10 and CIFAR-100 datasets, starting from trivial initial conditions and reaching
accuracies of 94.6% (95.6% for ensemble) and
77.0%, respectively. To do this, we use novel and
intuitive mutation operators that navigate large
search spaces; we stress that no human participation is required once evolution starts and that the
output is a fully-trained model. Throughout this
work, we place special emphasis on the repeatability of results, the variability in the outcomes
and the computational requirements.
1. Introduction
Neural networks can successfully perform difficult tasks
where large amounts of training data are available (He
et al., 2015; Weyand et al., 2016; Silver et al., 2016; Wu
et al., 2016). Discovering neural network architectures,
however, remains a laborious task. Even within the specific problem of image classification, the state of the art
was attained through many years of focused investigation
by hundreds of researchers (Krizhevsky et al. (2012); Simonyan & Zisserman (2014); Szegedy et al. (2015); He
et al. (2016); Huang et al. (2016a), among many others).
1
Google Brain, Mountain View, California, USA 2 Google Research, Mountain View, California, USA. Correspondence to: Esteban Real <[email protected]>.
Proceedings of the 34 th International Conference on Machine
Learning, Sydney, Australia, PMLR 70, 2017. Copyright 2017
by the author(s).
It is therefore not surprising that in recent years, techniques to automatically discover these architectures have
been gaining popularity (Bergstra & Bengio, 2012; Snoek
et al., 2012; Han et al., 2015; Baker et al., 2016; Zoph
& Le, 2016). One of the earliest such “neuro-discovery”
methods was neuro-evolution (Miller et al., 1989; Stanley
& Miikkulainen, 2002; Stanley, 2007; Bayer et al., 2009;
Stanley et al., 2009; Breuel & Shafait, 2010; Pugh & Stanley, 2013; Kim & Rigazio, 2015; Zaremba, 2015; Fernando
et al., 2016; Morse & Stanley, 2016). Despite the promising results, the deep learning community generally perceives evolutionary algorithms to be incapable of matching the accuracies of hand-designed models (Verbancsics &
Harguess, 2013; Baker et al., 2016; Zoph & Le, 2016). In
this paper, we show that it is possible to evolve such competitive models today, given enough computational power.
We used slightly-modified known evolutionary algorithms
and scaled up the computation to unprecedented levels, as
far as we know. This, together with a set of novel and
intuitive mutation operators, allowed us to reach competitive accuracies on the CIFAR-10 dataset. This dataset
was chosen because it requires large networks to reach
high accuracies, thus presenting a computational challenge.
We also took a small first step toward generalization and
evolved networks on the CIFAR-100 dataset. In transitioning from CIFAR-10 to CIFAR-100, we did not modify any aspect or parameter of our algorithm. Our typical
neuro-evolution outcome on CIFAR-10 had a test accuracy
with µ = 94.1%, σ = 0.4% @ 9 × 1019 FLOPs, and our
top model (by validation accuracy) had a test accuracy of
94.6% @ 4×1020 FLOPs. Ensembling the validation-top
2 models from each population reaches a test accuracy of
95.6%, at no additional training cost. On CIFAR-100, our
single experiment resulted in a test accuracy of 77.0% @
2 × 1020 FLOPs. As far as we know, these are the most
accurate results obtained on these datasets by automated
discovery methods that start from trivial initial conditions.
Throughout this study, we placed special emphasis on the
simplicity of the algorithm. In particular, it is a “oneshot” technique, producing a fully trained neural network
requiring no post-processing. It also has few impactful
meta-parameters (i.e. parameters not optimized by the algorithm). Starting out with poor-performing models with
Large-Scale Evolution
Table 1. Comparison with single-model hand-designed architectures. The “C10+” and “C100+” columns indicate the test accuracy on
the data-augmented CIFAR-10 and CIFAR-100 datasets, respectively. The “Reachable?” column denotes whether the given handdesigned model lies within our search space. An entry of “–” indicates that no value was reported. The † indicates a result reported by
Huang et al. (2016b) instead of the original author. Much of this table was based on that presented in Huang et al. (2016a).
S TUDY
M AXOUT (G OODFELLOW ET AL ., 2013)
N ETWORK IN N ETWORK (L IN ET AL ., 2013)
A LL -CNN (S PRINGENBERG ET AL ., 2014)
D EEPLY S UPERVISED (L EE ET AL ., 2015)
H IGHWAY (S RIVASTAVA ET AL ., 2015)
R ES N ET (H E ET AL ., 2016)
E VOLUTION ( OURS )
W IDE R ES N ET 28-10 (Z AGORUYKO & KOMODAKIS , 2016)
W IDE R ES N ET 40-10+ D / O (Z AGORUYKO & KOMODAKIS , 2016)
D ENSE N ET (H UANG ET AL ., 2016 A )
no convolutions, the algorithm must evolve complex convolutional neural networks while navigating a fairly unrestricted search space: no fixed depth, arbitrary skip connections, and numerical parameters that have few restrictions on the values they can take. We also paid close attention to result reporting. Namely, we present the variability in our results in addition to the top value, we account
for researcher degrees of freedom (Simmons et al., 2011),
we study the dependence on the meta-parameters, and we
disclose the amount of computation necessary to reach the
main results. We are hopeful that our explicit discussion of
computation cost could spark more study of efficient model
search and training. Studying model performance normalized by computational investment allows consideration of
economic concepts like opportunity cost.
2. Related Work
Neuro-evolution dates back many years (Miller et al.,
1989), originally being used only to evolve the weights
of a fixed architecture. Stanley & Miikkulainen (2002)
showed that it was advantageous to simultaneously evolve
the architecture using the NEAT algorithm. NEAT has
three kinds of mutations: (i) modify a weight, (ii) add a
connection between existing nodes, or (iii) insert a node
while splitting an existing connection. It also has a mechanism for recombining two models into one and a strategy
to promote diversity known as fitness sharing (Goldberg
et al., 1987). Evolutionary algorithms represent the models
using an encoding that is convenient for their purpose—
analogous to nature’s DNA. NEAT uses a direct encoding:
every node and every connection is stored in the DNA. The
alternative paradigm, indirect encoding, has been the subject of much neuro-evolution research (Gruau, 1993; Stanley et al., 2009; Pugh & Stanley, 2013; Kim & Rigazio,
PARAMS .
C10+
C100+
R EACHABLE ?
–
–
1.3 M
–
2.3 M
1.7 M
5.4 M
40.4 M
36.5 M
50.7 M
25.6 M
90.7%
91.2%
92.8%
92.0%
92.3%
93.4%
94.6%
61.4%
–
66.3%
65.4%
67.6%
72.8%†
NO
NO
Y ES
NO
NO
Y ES
96.0%
96.2%
96.7%
77.0%
80.0%
81.7%
82.8%
N/A
Y ES
NO
NO
2015; Fernando et al., 2016). For example, the CPPN
(Stanley, 2007; Stanley et al., 2009) allows for the evolution of repeating features at different scales. Also, Kim
& Rigazio (2015) use an indirect encoding to improve the
convolution filters in an initially highly-optimized fixed architecture.
Research on weight evolution is still ongoing (Morse &
Stanley, 2016) but the broader machine learning community defaults to back-propagation for optimizing neural network weights (Rumelhart et al., 1988). Back-propagation
and evolution can be combined as in Stanley et al. (2009),
where only the structure is evolved. Their algorithm follows an alternation of architectural mutations and weight
back-propagation. Similarly, Breuel & Shafait (2010) use
this approach for hyper-parameter search. Fernando et al.
(2016) also use back-propagation, allowing the trained
weights to be inherited through the structural modifications.
The above studies create neural networks that are small in
comparison to the typical modern architectures used for image classification (He et al., 2016; Huang et al., 2016a).
Their focus is on the encoding or the efficiency of the evolutionary process, but not on the scale. When it comes to
images, some neuro-evolution results reach the computational scale required to succeed on the MNIST dataset (LeCun et al., 1998). Yet, modern classifiers are often tested
on realistic images, such as those in the CIFAR datasets
(Krizhevsky & Hinton, 2009), which are much more challenging. These datasets require large models to achieve
high accuracy.
Non-evolutionary neuro-discovery methods have been
more successful at tackling realistic image data. Snoek
et al. (2012) used Bayesian optimization to tune 9
hyper-parameters for a fixed-depth architecture, reach-
Large-Scale Evolution
Table 2. Comparison with automatically discovered architectures. The “C10+” and “C100+” contain the test accuracy on the dataaugmented CIFAR-10 and CIFAR-100 datasets, respectively. An entry of “–” indicates that the information was not reported or is not
known to us. For Zoph & Le (2016), we quote the result with the most similar search space to ours, as well as their best result. Please
refer to Table 1 for hand-designed results, including the state of the art. “Discrete params.” means that the parameters can be picked
from a handful of values only (e.g. strides ∈ {1, 2, 4}).
S TUDY
S TARTING P OINT
C ONSTRAINTS
P OST-P ROCESSING
BAYESIAN
(S NOEK
ET AL ., 2012)
3 LAYERS
FIXED ARCHITECTURE , NO
SKIPS
NONE
DISCRETE PARAMS ., MAX .
NUM . LAYERS , NO SKIPS
TUNE , RETRAIN
DISCRETE PARAMS .,
EXACTLY 20 LAYERS
SMALL GRID
SEARCH , RETRAIN
LAYERS , 2 POOL
LAYERS AT 13 AND
26, 50% SKIPS
DISCRETE PARAMS .,
EXACTLY 39 LAYERS , 2
POOL LAYERS AT 13 AND
ADD MORE FILTERS ,
SMALL GRID
SEARCH , RETRAIN
SINGLE LAYER ,
ZERO CONVS .
POWER - OF -2 STRIDES
Q-L EARNING
(BAKER
ET AL ., 2016)
–
RL (Z OPH &
L E , 2016)
20
RL (Z OPH &
L E , 2016)
39
E VOLUTION
( OURS )
LAYERS ,
SKIPS
50%
ing a new state of the art at the time.
Zoph &
Le (2016) used reinforcement learning on a deeper
fixed-length architecture.
In their approach, a neural network—the “discoverer”—constructs a convolutional
neural network—the “discovered”—one layer at a time. In
addition to tuning layer parameters, they add and remove
skip connections. This, together with some manual postprocessing, gets them very close to the (current) state of
the art. (Additionally, they surpassed the state of the art on
a sequence-to-sequence problem.) Baker et al. (2016) use
Q-learning to also discover a network one layer at a time,
but in their approach, the number of layers is decided by
the discoverer. This is a desirable feature, as it would allow
a system to construct shallow or deep solutions, as may be
the requirements of the dataset at hand. Different datasets
would not require specially tuning the algorithm. Comparisons among these methods are difficult because they explore very different search spaces and have very different
initial conditions (Table 2).
Tangentially, there has also been neuro-evolution work on
LSTM structure (Bayer et al., 2009; Zaremba, 2015), but
this is beyond the scope of this paper. Also related to this
work is that of Saxena & Verbeek (2016), who embed convolutions with different parameters into a species of “supernetwork” with many parallel paths. Their algorithm then
selects and ensembles paths in the super-network. Finally,
canonical approaches to hyper-parameter search are grid
search (used in Zagoruyko & Komodakis (2016), for example) and random search, the latter being the better of the
26
PARAMS .
N ONE
–
C10+
90.5%
C100+
–
11.2 M
93.1%
72.9%
2.5 M
94.0%
–
37.0 M
96.4%
–
5.4 M
40.4 M
94.6%
ENSEMB .
95.6%
77.0%
two (Bergstra & Bengio, 2012).
Our approach builds on previous work, with some important differences. We explore large model-architecture
search spaces starting with basic initial conditions to avoid
priming the system with information about known good
strategies for the specific dataset at hand. Our encoding
is different from the neuro-evolution methods mentioned
above: we use a simplified graph as our DNA, which is
transformed to a full neural network graph for training and
evaluation (Section 3). Some of the mutations acting on
this DNA are reminiscent of NEAT. However, instead of
single nodes, one mutation can insert whole layers—i.e.
tens to hundreds of nodes at a time. We also allow for
these layers to be removed, so that the evolutionary process
can simplify an architecture in addition to complexifying it.
Layer parameters are also mutable, but we do not prescribe
a small set of possible values to choose from, to allow for
a larger search space. We do not use fitness sharing. We
report additional results using recombination, but for the
most part, we used mutation only. On the other hand, we
do use back-propagation to optimize the weights, which
can be inherited across mutations. Together with a learning rate mutation, this allows the exploration of the space
of learning rate schedules, yielding fully trained models
at the end of the evolutionary process (Section 3). Tables 1 and 2 compare our approach with hand-designed architectures and with other neuro-discovery techniques, respectively.
Large-Scale Evolution
3. Methods
3.1. Evolutionary Algorithm
To automatically search for high-performing neural network architectures, we evolve a population of models.
Each model—or individual—is a trained architecture. The
model’s accuracy on a separate validation dataset is a measure of the individual’s quality or fitness. During each evolutionary step, a computer—a worker—chooses two individuals at random from this population and compares their
fitnesses. The worst of the pair is immediately removed
from the population—it is killed. The best of the pair is
selected to be a parent, that is, to undergo reproduction.
By this we mean that the worker creates a copy of the parent and modifies this copy by applying a mutation, as described below. We will refer to this modified copy as the
child. After the worker creates the child, it trains this child,
evaluates it on the validation set, and puts it back into the
population. The child then becomes alive—i.e. free to act
as a parent. Our scheme, therefore, uses repeated pairwise
competitions of random individuals, which makes it an example of tournament selection (Goldberg & Deb, 1991).
Using pairwise comparisons instead of whole population
operations prevents workers from idling when they finish
early. Code and more detail about the methods described
below can be found in Supplementary Section S1.
Using this strategy to search large spaces of complex image models requires considerable computation. To achieve
scale, we developed a massively-parallel, lock-free infrastructure. Many workers operate asynchronously on different computers. They do not communicate directly with
each other. Instead, they use a shared file-system, where
the population is stored. The file-system contains directories that represent the individuals. Operations on these
individuals, such as the killing of one, are represented as
atomic renames on the directory2 . Occasionally, a worker
may concurrently modify the individual another worker is
operating on. In this case, the affected worker simply gives
up and tries again. The population size is 1000 individuals,
unless otherwise stated. The number of workers is always
1
4 of the population size. To allow for long run-times with
a limited amount of space, dead individuals’ directories are
frequently garbage-collected.
3.2. Encoding and Mutations
Individual architectures are encoded as a graph that we
refer to as the DNA. In this graph, the vertices represent
rank-3 tensors or activations. As is standard for a convo2
The use of the file-name string to contain key information
about the individual was inspired by Breuel & Shafait (2010), and
it speeds up disk access enormously. In our case, the file name
contains the state of the individual (alive, dead, training, etc.).
lutional network, two of the dimensions of the tensor represent the spatial coordinates of the image and the third is
a number of channels. Activation functions are applied at
the vertices and can be either (i) batch-normalization (Ioffe
& Szegedy, 2015) with rectified linear units (ReLUs) or (ii)
plain linear units. The graph’s edges represent identity connections or convolutions and contain the mutable numerical parameters defining the convolution’s properties. When
multiple edges are incident on a vertex, their spatial scales
or numbers of channels may not coincide. However, the
vertex must have a single size and number of channels for
its activations. The inconsistent inputs must be resolved.
Resolution is done by choosing one of the incoming edges
as the primary one. We pick this primary edge to be the
one that is not a skip connection. The activations coming
from the non-primary edges are reshaped through zerothorder interpolation in the case of the size and through truncation/padding in the case of the number of channels, as in
He et al. (2016). In addition to the graph, the learning-rate
value is also stored in the DNA.
A child is similar but not identical to the parent because of
the action of a mutation. In each reproduction event, the
worker picks a mutation at random from a predetermined
set. The set contains the following mutations:
• A LTER - LEARNING - RATE (sampling details below).
• I DENTITY (effectively means “keep training”).
• R ESET- WEIGHTS (sampled as in He et al. (2015), for
example).
• I NSERT- CONVOLUTION (inserts a convolution at a random location in the “convolutional backbone”, as in Figure 1. The inserted convolution has 3 × 3 filters, strides
of 1 or 2 at random, number of channels same as input.
May apply batch-normalization and ReLU activation or
none at random).
• R EMOVE - CONVOLUTION.
• A LTER - STRIDE (only powers of 2 are allowed).
• A LTER - NUMBER - OF - CHANNELS (of random conv.).
• F ILTER - SIZE (horizontal or vertical at random, on random convolution, odd values only).
• I NSERT- ONE - TO - ONE (inserts a one-to-one/identity
connection, analogous to insert-convolution mutation).
• A DD - SKIP (identity between random layers).
• R EMOVE - SKIP (removes random skip).
These specific mutations were chosen for their similarity
to the actions that a human designer may take when improving an architecture. This may clear the way for hybrid
evolutionary–hand-design methods in the future. The probabilities for the mutations were not tuned in any way.
A mutation that acts on a numerical parameter chooses the
new value at random around the existing value. All sampling is from uniform distributions. For example, a mutation acting on a convolution with 10 output channels will
Large-Scale Evolution
result in a convolution having between 5 and 20 output
channels (that is, half to twice the original value). All values within the range are possible. As a result, the models
are not constrained to a number of filters that is known to
work well. The same is true for all other parameters, yielding a “dense” search space. In the case of the strides, this
applies to the log-base-2 of the value, to allow for activation shapes to match more easily3 . In principle, there is also
no upper limit to any of the parameters. All model depths
are attainable, for example. Up to hardware constraints, the
search space is unbounded. The dense and unbounded nature of the parameters result in the exploration of a truly
large set of possible architectures.
3.3. Initial Conditions
Every evolution experiment begins with a population of
simple individuals, all with a learning rate of 0.1. They
are all very bad performers. Each initial individual constitutes just a single-layer model with no convolutions. This
conscious choice of poor initial conditions forces evolution
to make the discoveries by itself. The experimenter contributes mostly through the choice of mutations that demarcate a search space. Altogether, the use of poor initial conditions and a large search space limits the experimenter’s
impact. In other words, it prevents the experimenter from
“rigging” the experiment to succeed.
3.5. Computation cost
To estimate computation costs, we identified the basic
TensorFlow (TF) operations used by our model training
and validation, like convolutions, generic matrix multiplications, etc. For each of these TF operations, we estimated the theoretical number of floating-point operations
(FLOPs) required. This resulted in a map from TF operation to FLOPs, which is valid for all our experiments.
For each individual within an evolution experiment, we
compute the total FLOPs incurred by the TF operations in
its architecture over one batch of examples, both during its
training (Ft FLOPs) and during its validation (Fv FLOPs).
Then we assign to the individual the cost Ft Nt + Fv Nv ,
where Nt and Nv are the number of training and validation
batches, respectively. The cost of the experiment is then
the sum of the costs of all its individuals.
We intend our FLOPs measurement as a coarse estimate
only. We do not take into account input/output, data preprocessing, TF graph building or memory-copying operations.
Some of these unaccounted operations take place once per
training run or once per step and some have a component
that is constant in the model size (such as disk-access latency or input data cropping). We therefore expect the estimate to be more useful for large architectures (for example,
those with many convolutions).
3.4. Training and Validation
3.6. Weight Inheritance
Training and validation is done on the CIFAR-10 dataset.
This dataset consists of 50,000 training examples and
10,000 test examples, all of which are 32 x 32 color images
labeled with 1 of 10 common object classes (Krizhevsky &
Hinton, 2009). 5,000 of the training examples are held out
in a validation set. The remaining 45,000 examples constitute our actual training set. The training set is augmented
as in He et al. (2016). The CIFAR-100 dataset has the same
number of dimensions, colors and examples as CIFAR-10,
but uses 100 classes, making it much more challenging.
We need architectures that are trained to completion within
an evolution experiment. If this does not happen, we are
forced to retrain the best model at the end, possibly having to explore its hyper-parameters. Such extra exploration tends to depend on the details of the model being
retrained. On the other hand, 25,600 steps are not enough
to fully train each individual. Training a large model to
completion is prohibitively slow for evolution. To resolve
this dilemma, we allow the children to inherit the parents’ weights whenever possible. Namely, if a layer has
matching shapes, the weights are preserved. Consequently,
some mutations preserve all the weights (like the identity or
learning-rate mutations), some preserve none (the weightresetting mutation), and most preserve some but not all. An
example of the latter is the filter-size mutation: only the filters of the convolution being mutated will be discarded.
Training is done with TensorFlow (Abadi et al., 2016), using SGD with a momentum of 0.9 (Sutskever et al., 2013), a
batch size of 50, and a weight decay of 0.0001. Each training runs for 25,600 steps, a value chosen to be brief enough
so that each individual could be trained in a few seconds to
a few hours, depending on model size. The loss function is
the cross-entropy. Once training is complete, a single evaluation on the validation set provides the accuracy to use as
the individual’s fitness. Ensembling was done by majority
voting during the testing evaluation. The models used in
the ensemble were selected by validation accuracy.
3
For integer DNA parameters, we actually store and mutate a
floating-point value. This allows multiple small mutations to have
a cumulative effect in spite of integer round-off.
3.7. Reporting Methodology
To avoid over-fitting, neither the evolutionary algorithm nor
the neural network training ever see the testing set. Each
time we refer to “the best model”, we mean the model with
the highest validation accuracy. However, we always report
the test accuracy. This applies not only to the choice of the
best individual within an experiment, but also to the choice
Large-Scale Evolution
of the best experiment. Moreover, we only include experiments that we managed to reproduce, unless explicitly
noted. Any statistical analysis was fully decided upon before seeing the results of the experiment reported, to avoid
tailoring our analysis to our experimental data (Simmons
et al., 2011).
4. Experiments and Results
We want to answer the following questions:
• Can a simple one-shot evolutionary process start from
trivial initial conditions and yield fully trained models
that rival hand-designed architectures?
• What are the variability in outcomes, the parallelizability, and the computation cost of the method?
• Can an algorithm designed iterating on CIFAR-10 be applied, without any changes at all, to CIFAR-100 and still
produce competitive models?
We used the algorithm in Section 3 to perform several experiments. Each experiment evolves a population in a few
days, typified by the example in Figure 1. The figure also
contains examples of the architectures discovered, which
turn out to be surprisingly simple. Evolution attempts skip
connections but frequently rejects them.
To get a sense of the variability in outcomes, we repeated
the experiment 5 times. Across all 5 experiment runs, the
best model by validation accuracy has a testing accuracy of
94.6 %. Not all experiments reach the same accuracy, but
they get close (µ = 94.1%, σ = 0.4). Fine differences in the
experiment outcome may be somewhat distinguishable by
validation accuracy (correlation coefficient = 0.894). The
total amount of computation across all 5 experiments was
4×1020 FLOPs (or 9×1019 FLOPs on average per experiment). Each experiment was distributed over 250 parallel
workers (Section 3.1). Figure 2 shows the progress of the
experiments in detail.
As a control, we disabled the selection mechanism, thereby
reproducing and killing random individuals. This is the
form of random search that is most compatible with our
infrastructure. The probability distributions for the parameters are implicitly determined by the mutations. This
control only achieves an accuracy of 87.3 % in the same
amount of run time on the same hardware (Figure 2). The
total amount of computation was 2×1017 FLOPs. The low
FLOP count is a consequence of random search generating
many small, inadequate models that train quickly but consume roughly constant amounts of setup time (not included
in the FLOP count). We attempted to minimize this overhead by avoiding unnecessary disk access operations, to no
avail: too much overhead remains spent on a combination
of neural network setup, data augmentation, and training
step initialization.
We also ran a partial control where the weight-inheritance
mechanism is disabled. This run also results in a lower
accuracy (92.2 %) in the same amount of time (Figure 2),
using 9×1019 FLOPs. This shows that weight inheritance
is important in the process.
Finally, we applied our neuro-evolution algorithm, without any changes and with the same meta-parameters, to
CIFAR-100. Our only experiment reached an accuracy
of 77.0 %, using 2 × 1020 FLOPs. We did not attempt
other datasets. Table 1 shows that both the CIFAR-10
and CIFAR-100 results are competitive with modern handdesigned networks.
5. Analysis
Meta-parameters. We observe that populations evolve
until they plateau at some local optimum (Figure 2). The
fitness (i.e. validation accuracy) value at this optimum
varies between experiments (Figure 2, inset). Since not all
experiments reach the highest possible value, some populations are getting “trapped” at inferior local optima. This
entrapment is affected by two important meta-parameters
(i.e. parameters that are not optimized by the algorithm).
These are the population size and the number of training
steps per individual. Below we discuss them and consider
their relationship to local optima.
Effect of population size. Larger populations explore the
space of models more thoroughly, and this helps reach better optima (Figure 3, left). Note, in particular, that a population of size 2 can get trapped at very low fitness values.
Some intuition about this can be gained by considering the
fate of a super-fit individual, i.e. an individual such that any
one architectural mutation reduces its fitness (even though
a sequence of many mutations may improve it). In the case
of a population of size 2, if the super-fit individual wins
once, it will win every time. After the first win, it will produce a child that is one mutation away. By definition of
super-fit, therefore, this child is inferior4 . Consequently,
in the next round of tournament selection, the super-fit individual competes against its child and wins again. This
cycle repeats forever and the population is trapped. Even if
a sequence of two mutations would allow for an “escape”
from the local optimum, such a sequence can never take
place. This is only a rough argument to heuristically suggest why a population of size 2 is easily trapped. More
generally, Figure 3 (left) empirically demonstrates a benefit from an increase in population size. Theoretical analyses of this dependence are quite complex and assume very
specific models of population dynamics; often larger populations are better at handling local optima, at least beyond
a size threshold (Weinreich & Chao (2005) and references
4
Except after identity or learning rate mutations, but these produce a child with the same architecture as the parent.
Large-Scale Evolution
94.6
91.8
85.3
Input
test accuracy (%)
C
Input
Input
C
C
C
C + BN + R
C
C + BN + R + BN + R
BN + R
C + BN + R + BN + R + BN + R + BN + R
C + BN + R + BN + R + BN + R + BN + R
C + BN + R
C + BN + R
C + BN + R
C + BN + R
C
C + BN + R
Input
C
22.6
Global Pool
C + BN + R
Output
Global Pool
Output
C
BN + R
C + BN + R
C + BN + R
C + BN + R
C + BN + R + BN + R
C
C
C
C + BN + R + BN + R
C + BN + R
Global Pool
C + BN + R
Output
Global Pool
Output
0.9
28.1
70.2
wall time (hours)
256.2
Figure 1. Progress of an evolution experiment. Each dot represents an individual in the population. Blue dots (darker, top-right) are alive.
The rest have been killed. The four diagrams show examples of discovered architectures. These correspond to the best individual (rightmost) and three of its ancestors. The best individual was selected by its validation accuracy. Evolution sometimes stacks convolutions
without any nonlinearity in between (“C”, white background), which are mathematically equivalent to a single linear operation. Unlike
typical hand-designed architectures, some convolutions are followed by more than one nonlinear function (“C+BN +R+BN +R+...”,
orange background).
therein).
Effect of number of training steps. The other metaparameter is the number T of training steps for each individual. Accuracy increases with T (Figure 3, right). Larger
T means an individual needs to undergo fewer identity mutations to reach a given level of training.
Escaping local optima. While we might increase population size or number of steps to prevent a trapped population from forming, we can also free an already trapped
population. For example, increasing the mutation rate or
resetting all the weights of a population (Figure 4) work
well but are quite costly (more details in Supplementary
Section S3).
Recombination. None of the results presented so far
used recombination. However, we explored three forms of
recombination in additional experiments. Following Tuson
& Ross (1998), we attempted to evolve the mutation probability distribution too. On top of this, we employed a recombination strategy by which a child could inherit structure from one parent and mutation probabilities from another. The goal was to allow individuals that progressed
well due to good mutation choices to quickly propagate
such choices to others. In a separate experiment, we attempted recombining the trained weights from two parents
in the hope that each parent may have learned different
concepts from the training data. In a third experiment,
we recombined structures so that the child fused the architectures of both parents side-by-side, generating wide
models fast. While none of these approaches improved our
recombination-free results, further study seems warranted.
6. Conclusion
In this paper we have shown that (i) neuro-evolution is capable of constructing large, accurate networks for two challenging and popular image classification benchmarks; (ii)
neuro-evolution can do this starting from trivial initial conditions while searching a very large space; (iii) the process, once started, needs no experimenter participation; and
(iv) the process yields fully trained models. Completely
training models required weight inheritance (Sections 3.6).
In contrast to reinforcement learning, evolution provides a
natural framework for weight inheritance: mutations can
be constructed to guarantee a large degree of similarity be-
Large-Scale Evolution
100.0
94.6
250
Evolution
94
Evolution w/o
weight inheritance
Random search
92
20.0
0
wall-clock time (hours)
250
Figure 2. Repeatability of results and controls. In this plot, the
vertical axis at wall-time t is defined as the test accuracy of the
individual with the highest validation accuracy that became alive
at or before t. The inset magnifies a portion of the main graph.
The curves show the progress of various experiments, as follows.
The top line (solid, blue) shows the mean test accuracy across 5
large-scale evolution experiments. The shaded area around this
top line has a width of ±2σ (clearer in inset). The next line down
(dashed, orange, main graph and inset) represents a single experiment in which weight-inheritance was disabled, so every individual has to train from random weights. The lowest curve (dotteddashed) is a random-search control. All experiments occupied the
same amount and type of hardware. A small amount of noise in
the generalization from the validation to the test set explains why
the lines are not monotonically increasing. Note the narrow width
of the ±2σ area (main graph and inset), which shows that the high
accuracies obtained in evolution experiments are repeatable.
50
test accuracy (%)
150
100
test accuracy (%)
test accuracy (%)
100
2
10 43
1000
population size
75
256
2560
25600
training steps
Figure 3. Dependence on meta-parameters. In both graphs, each
circle represents the result of a full evolution experiment. Both
vertical axes show the test accuracy for the individual with the
highest validation accuracy at the end of the experiment. All populations evolved for the same total wall-clock time. There are 5
data points at each horizontal axis value. LEFT: effect of population size. To economize resources, in these experiments the
number of individual training steps is only 2560. Note how the accuracy increases with population size. RIGHT: effect of number
of training steps per individual. Note how the accuracy increases
with more steps.
tween the original and mutated models—as we did. Evolution also has fewer tunable meta-parameters with a fairly
predictable effect on the variance of the results, which can
be made small.
While we did not focus on reducing computation costs,
we hope that future algorithmic and hardware improvement
will allow more economical implementation. In that case,
evolution would become an appealing approach to neurodiscovery for reasons beyond the scope of this paper. For
example, it “hits the ground running”, improving on arbitrary initial models as soon as the experiment begins. The
mutations used can implement recent advances in the field
and can be introduced without having to restart an experiment. Furthermore, recombination can merge improvements developed by different individuals, even if they come
from other populations. Moreover, it may be possible to
combine neuro-evolution with other automatic architecture
discovery methods.
Figure 4. Escaping local optima in two experiments. We used
smaller populations and fewer training steps per individual (2560)
to make it more likely for a population to get trapped and to reduce resource usage. Each dot represents an individual. The vertical axis is the accuracy. TOP: example of a population of size 100
escaping a local optimum by using a period of increased mutation
rate in the middle (Section 5). BOTTOM: example of a population
of size 50 escaping a local optimum by means of three consecutive weight resetting events (Section 5). Details in Supplementary
Section S3.
Large-Scale Evolution
Acknowledgements
We wish to thank Vincent Vanhoucke, Megan Kacholia, Rajat Monga, and especially Jeff Dean for their support and valuable input; Geoffrey Hinton, Samy Bengio, Thomas Breuel, Mark DePristo, Vishy Tirumalashetty,
Martin Abadi, Noam Shazeer, Yoram Singer, Dumitru Erhan, Pierre Sermanet, Xiaoqiang Zheng, Shan Carter and
Vijay Vasudevan for helpful discussions; Thomas Breuel,
Xin Pan and Andy Davis for coding contributions; and the
larger Google Brain team for help with TensorFlow and
training vision models.
References
Abadi, Martı́n, Agarwal, Ashish, Barham, Paul, Brevdo,
Eugene, Chen, Zhifeng, Citro, Craig, Corrado, Greg S,
Davis, Andy, Dean, Jeffrey, Devin, Matthieu, et al. Tensorflow: Large-scale machine learning on heterogeneous
distributed systems. arXiv preprint arXiv:1603.04467,
2016.
Baker, Bowen, Gupta, Otkrist, Naik, Nikhil, and
Raskar, Ramesh. Designing neural network architectures using reinforcement learning. arXiv preprint
arXiv:1611.02167, 2016.
Bayer, Justin, Wierstra, Daan, Togelius, Julian, and
Schmidhuber, Jürgen. Evolving memory cell structures
for sequence learning. In International Conference on
Artificial Neural Networks, pp. 755–764. Springer, 2009.
Goodfellow, Ian J, Warde-Farley, David, Mirza, Mehdi,
Courville, Aaron C, and Bengio, Yoshua. Maxout networks. International Conference on Machine Learning,
28:1319–1327, 2013.
Gruau, Frederic. Genetic synthesis of modular neural networks. In Proceedings of the 5th International Conference on Genetic Algorithms, pp. 318–325. Morgan Kaufmann Publishers Inc., 1993.
Han, Song, Pool, Jeff, Tran, John, and Dally, William.
Learning both weights and connections for efficient neural network. In Advances in Neural Information Processing Systems, pp. 1135–1143, 2015.
He, Kaiming, Zhang, Xiangyu, Ren, Shaoqing, and Sun,
Jian. Delving deep into rectifiers: Surpassing humanlevel performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, pp. 1026–1034, 2015.
He, Kaiming, Zhang, Xiangyu, Ren, Shaoqing, and Sun,
Jian. Deep residual learning for image recognition. In
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770–778, 2016.
Huang, Gao, Liu, Zhuang, Weinberger, Kilian Q, and
van der Maaten, Laurens. Densely connected convolutional networks. arXiv preprint arXiv:1608.06993,
2016a.
Bergstra, James and Bengio, Yoshua. Random search
for hyper-parameter optimization. Journal of Machine
Learning Research, 13(Feb):281–305, 2012.
Huang, Gao, Sun, Yu, Liu, Zhuang, Sedra, Daniel, and
Weinberger, Kilian Q. Deep networks with stochastic
depth. In European Conference on Computer Vision, pp.
646–661. Springer, 2016b.
Breuel, Thomas and Shafait, Faisal. Automlp: Simple,
effective, fully automated learning rate and size adjustment. In The Learning Workshop. Utah, 2010.
Ioffe, Sergey and Szegedy, Christian. Batch normalization:
Accelerating deep network training by reducing internal
covariate shift. arXiv preprint arXiv:1502.03167, 2015.
Fernando, Chrisantha, Banarse, Dylan, Reynolds, Malcolm, Besse, Frederic, Pfau, David, Jaderberg, Max,
Lanctot, Marc, and Wierstra, Daan. Convolution by evolution: Differentiable pattern producing networks. In
Proceedings of the 2016 on Genetic and Evolutionary
Computation Conference, pp. 109–116. ACM, 2016.
Kim, Minyoung and Rigazio, Luca. Deep clustered convolutional kernels. arXiv preprint arXiv:1503.01824, 2015.
Goldberg, David E and Deb, Kalyanmoy. A comparative
analysis of selection schemes used in genetic algorithms.
Foundations of genetic algorithms, 1:69–93, 1991.
Goldberg, David E, Richardson, Jon, et al. Genetic algorithms with sharing for multimodal function optimization. In Genetic algorithms and their applications: Proceedings of the Second International Conference on Genetic Algorithms, pp. 41–49. Hillsdale, NJ: Lawrence
Erlbaum, 1987.
Krizhevsky, Alex and Hinton, Geoffrey. Learning multiple
layers of features from tiny images. 2009.
Krizhevsky, Alex, Sutskever, Ilya, and Hinton, Geoffrey E.
Imagenet classification with deep convolutional neural
networks. In Advances in Neural Information Processing
Systems, pp. 1097–1105, 2012.
LeCun, Yann, Cortes, Corinna, and Burges, Christopher JC. The mnist database of handwritten digits, 1998.
Lee, Chen-Yu, Xie, Saining, Gallagher, Patrick W, Zhang,
Zhengyou, and Tu, Zhuowen. Deeply-supervised nets.
In AISTATS, volume 2, pp. 5, 2015.
Large-Scale Evolution
Lin, Min, Chen, Qiang, and Yan, Shuicheng. Network in
network. arXiv preprint arXiv:1312.4400, 2013.
Miller, Geoffrey F, Todd, Peter M, and Hegde, Shailesh U.
Designing neural networks using genetic algorithms. In
Proceedings of the third international conference on Genetic algorithms, pp. 379–384. Morgan Kaufmann Publishers Inc., 1989.
Morse, Gregory and Stanley, Kenneth O. Simple evolutionary optimization can rival stochastic gradient descent in neural networks. In Proceedings of the 2016 on
Genetic and Evolutionary Computation Conference, pp.
477–484. ACM, 2016.
Pugh, Justin K and Stanley, Kenneth O. Evolving multimodal controllers with hyperneat. In Proceedings of
the 15th annual conference on Genetic and evolutionary
computation, pp. 735–742. ACM, 2013.
Rumelhart, David E, Hinton, Geoffrey E, and Williams,
Ronald J. Learning representations by back-propagating
errors. Cognitive Modeling, 5(3):1, 1988.
Saxena, Shreyas and Verbeek, Jakob. Convolutional neural
fabrics. In Advances In Neural Information Processing
Systems, pp. 4053–4061, 2016.
Silver, David, Huang, Aja, Maddison, Chris J, Guez,
Arthur, Sifre, Laurent, Van Den Driessche, George,
Schrittwieser, Julian, Antonoglou, Ioannis, Panneershelvam, Veda, Lanctot, Marc, et al. Mastering the game of
go with deep neural networks and tree search. Nature,
529(7587):484–489, 2016.
Simmons, Joseph P, Nelson, Leif D, and Simonsohn, Uri.
False-positive psychology: Undisclosed flexibility in
data collection and analysis allows presenting anything
as significant. Psychological Science, 22(11):1359–
1366, 2011.
Simonyan, Karen and Zisserman, Andrew. Very deep convolutional networks for large-scale image recognition.
arXiv preprint arXiv:1409.1556, 2014.
Snoek, Jasper, Larochelle, Hugo, and Adams, Ryan P.
Practical bayesian optimization of machine learning algorithms. In Advances in neural information processing
systems, pp. 2951–2959, 2012.
Springenberg, Jost Tobias, Dosovitskiy, Alexey, Brox,
Thomas, and Riedmiller, Martin. Striving for simplicity: The all convolutional net. arXiv preprint
arXiv:1412.6806, 2014.
Srivastava, Rupesh Kumar, Greff, Klaus, and Schmidhuber, Jürgen. Highway networks. arXiv preprint
arXiv:1505.00387, 2015.
Stanley, Kenneth O. Compositional pattern producing networks: A novel abstraction of development. Genetic programming and evolvable machines, 8(2):131–162, 2007.
Stanley, Kenneth O and Miikkulainen, Risto. Evolving
neural networks through augmenting topologies. Evolutionary Computation, 10(2):99–127, 2002.
Stanley, Kenneth O, D’Ambrosio, David B, and Gauci, Jason. A hypercube-based encoding for evolving largescale neural networks. Artificial Life, 15(2):185–212,
2009.
Sutskever, Ilya, Martens, James, Dahl, George E, and Hinton, Geoffrey E. On the importance of initialization and
momentum in deep learning. ICML (3), 28:1139–1147,
2013.
Szegedy, Christian, Liu, Wei, Jia, Yangqing, Sermanet,
Pierre, Reed, Scott, Anguelov, Dragomir, Erhan, Dumitru, Vanhoucke, Vincent, and Rabinovich, Andrew.
Going deeper with convolutions. In Proceedings of
the IEEE Conference on Computer Vision and Pattern
Recognition, pp. 1–9, 2015.
Tuson, Andrew and Ross, Peter. Adapting operator settings
in genetic algorithms. Evolutionary computation, 6(2):
161–184, 1998.
Verbancsics, Phillip and Harguess, Josh. Generative
neuroevolution for deep learning.
arXiv preprint
arXiv:1312.5355, 2013.
Weinreich, Daniel M and Chao, Lin. Rapid evolutionary
escape by large populations from local fitness peaks is
likely in nature. Evolution, 59(6):1175–1182, 2005.
Weyand, Tobias, Kostrikov, Ilya, and Philbin, James.
Planet-photo geolocation with convolutional neural networks. In European Conference on Computer Vision, pp.
37–55. Springer, 2016.
Wu, Yonghui, Schuster, Mike, Chen, Zhifeng, Le, Quoc V.,
Norouzi, Mohammad, et al. Google’s neural machine
translation system: Bridging the gap between human and
machine translation. arXiv preprint arXiv:1609.08144,
2016.
Zagoruyko, Sergey and Komodakis, Nikos. Wide residual
networks. arXiv preprint arXiv:1605.07146, 2016.
Zaremba, Wojciech. An empirical exploration of recurrent
network architectures. 2015.
Zoph, Barret and Le, Quoc V.
Neural architecture
search with reinforcement learning. arXiv preprint
arXiv:1611.01578, 2016.
Large-Scale Evolution of Image Classifiers
Supplementary Material
S1. Methods Details
This section contains additional implementation details, roughly following the order in Section 3. Short code snippets
illustrate the ideas. The code is not intended to run on its own and it has been highly edited for clarity.
In our implementation, each worker runs an outer loop that is responsible for selecting a pair of random individuals from
the population. The individual with the highest fitness usually becomes a parent and the one with the lowest fitness is
usually killed (Section 3.1). Occasionally, either of these two actions is not carried out in order to keep the population size
close to a set-point:
def evolve_population(self):
# Iterate indefinitely.
while True:
# Select two random individuals from the population.
valid_individuals = []
for individual in self.load_individuals(): # Only loads the IDs and states.
if individual.state in [TRAINING, ALIVE]:
valid_individuals.append(individual)
individual_pair = random.sample(valid_individuals, 2)
for individual in individual_pair:
# Sync changes from other workers from file-system. Loads everything else.
individual.update_if_necessary()
# Ensure the individual is fully trained.
if individual.state == TRAINING:
self._train(individual)
# Select by fitness (accuracy).
individual_pair.sort(key=lambda i: i.fitness, reverse=True)
better_individual = individual_pair[0]
worse_individual = individual_pair[1]
# If the population is not too small, kill the worst of the pair.
if self._population_size() >= self._population_size_setpoint:
self._kill_individual(worse_individual)
# If the population is not too large, reproduce the best of the pair.
if self._population_size() < self._population_size_setpoint:
self._reproduce_and_train_individual(better_individual)
Much of the code is wrapped in try-except blocks to handle various kinds of errors. These have been removed from the
code snippets for clarity. For example, the method above would be wrapped like this:
def evolve_population(self):
while True:
try:
# Select two random individuals from the population.
...
except:
except exceptions.PopulationTooSmallException:
self._create_new_individual()
continue
Large-Scale Evolution
except exceptions.ConcurrencyException:
# Another worker did something that interfered with the action of this worker.
# Abandon the current task and keep going.
continue
The encoding for an individual is represented by a serializable DNA class instance containing all information except for the
trained weights (Section 3.2). For all results in this paper, this encoding is a directed, acyclic graph where edges represent
convolutions and vertices represent nonlinearities. This is a sketch of the DNA class:
class DNA(object):
def __init__(self, dna_proto):
"""Initializes the ‘DNA‘ instance from a protocol buffer.
The ‘dna_proto‘ is a protocol buffer used to restore the DNA state from disk.
Together with the corresponding ‘to_proto‘ method, they allow for a
serialization-deserialization mechanism.
"""
# Allows evolving the learning rate, i.e. exploring the space of
# learning rate schedules.
self.learning_rate = dna_proto.learning_rate
self._vertices = {} # String vertex ID to ‘Vertex‘ instance.
for vertex_id in dna_proto.vertices:
vertices[vertex_id] = Vertex(vertex_proto=dna_sproto.vertices[vertex_id])
self._edges = {} # String edge ID to ‘Edge‘ instance.
for edge_id in dna_proto.edges:
mutable_edges[edge_id] = Edge(edge_proto=dna_proto.edges[edge_id])
...
def to_proto(self):
"""Returns this instance in protocol buffer form."""
dna_proto = dna_pb2.DnaProto(learning_rate=self.learning_rate)
for vertex_id, vertex in self._vertices.iteritems():
dna_proto.vertices[vertex_id].CopyFrom(vertex.to_proto())
for edge_id, edge in self._edges.iteritems():
dna_proto.edges[edge_id].CopyFrom(edge.to_proto())
...
return dna_proto
def add_edge(self, dna, from_vertex_id, to_vertex_id, edge_type, edge_id):
"""Adds an edge to the DNA graph, ensuring internal consistency."""
# ‘EdgeProto‘ defines defaults for other attributes.
edge = Edge(EdgeProto(
from_vertex=from_vertex_id, to_vertex=to_vertex_id, type=edge_type))
self._edges[edge_id] = edge
self._vertices[from_vertex_id].edges_out.add(edge_id)
self._vertices[to_vertex].edges_in.add(edge_id)
return edge
# Other methods like ‘add_edge‘ to manipulate the graph structure.
...
The DNA holds Vertex and Edge instances. The Vertex class looks like this:
class Vertex(object):
def __init__(self, vertex_proto):
# Relationship to the rest of the graph.
Large-Scale Evolution
self.edges_in = set(vertex_proto.edges_in) # Incoming edge IDs.
self.edges_out = set(vertex_proto.edges_out) # Outgoing edge IDs.
# The type of activations.
if vertex_proto.HasField(’linear’):
self.type = LINEAR # Linear activations.
elif vertex_proto.HasField(’bn_relu’):
self.type = BN_RELU # ReLU activations with batch-normalization.
else:
raise NotImplementedError()
# Some parts of the graph can be prevented from being acted upon by mutations.
# The following boolean flags control this.
self.inputs_mutable = vertex_proto.inputs_mutable
self.outputs_mutable = vertex_proto.outputs_mutable
self.properties_mutable = vertex_proto.properties_mutable
#
#
#
#
Each vertex represents a 2ˆs x 2ˆs x d block of nodes. s and d are positive
integers computed dynamically from the in-edges. s stands for "scale" so
that 2ˆx x 2ˆs is the spatial size of the activations. d stands for "depth",
the number of channels.
def to_proto(self):
...
The Edge class looks like this:
class Edge(object):
def __init__(self, edge_proto):
# Relationship to the rest of the graph.
self.from_vertex = edge_proto.from_vertex # Source vertex ID.
self.to_vertex = edge_proto.to_vertex # Destination vertex ID.
if edge_proto.HasField(’conv’):
# In this case, the edge represents a convolution.
self.type = CONV
# Controls the depth (i.e. number of channels) in the output, relative to the
# input. For example if there is only one input edge with a depth of 16 channels
# and ‘self._depth_factor‘ is 2, then this convolution will result in an output
# depth of 32 channels. Multiple-inputs with conflicting depth must undergo
# depth resolution first.
self.depth_factor = edge_proto.conv.depth_factor
# Control the shape of the convolution filters (i.e. transfer function).
# This parameterization ensures that the filter width and height are odd
# numbers: filter_width = 2 * filter_half_width + 1.
self.filter_half_width = edge_proto.conv.filter_half_width
self.filter_half_height = edge_proto.conv.filter_half_height
# Controls the strides of the convolution. It will be 2ˆstride_scale.
# Note that conflicting input scales must undergo scale resolution. This
# controls the spatial scale of the output activations relative to the
# spatial scale of the input activations.
self.stride_scale = edge_proto.conv.stride_scale
elif edge_spec.HasField(’identity’):
self.type = IDENTITY
else:
raise NotImplementedError()
# In case depth or scale resolution is necessary due to conflicts in inputs,
# These integer parameters determine which of the inputs takes precedence in
# deciding the resolved depth or scale.
self.depth_precedence = edge_proto.depth_precedence
Large-Scale Evolution
self.scale_precedence = edge_proto.scale_precedence
def to_proto(self):
...
Mutations act on DNA instances. The set of mutations restricts the space explored somewhat (Section 3.2). The following
are some example mutations. The AlterLearningRateMutation simply randomly modifies the attribute in the DNA:
class AlterLearningRateMutation(Mutation):
"""Mutation that modifies the learning rate."""
def mutate(self, dna):
mutated_dna = copy.deepcopy(dna)
# Mutate the learning rate by a random factor between 0.5 and 2.0,
# uniformly distributed in log scale.
factor = 2**random.uniform(-1.0, 1.0)
mutated_dna.learning_rate = dna.learning_rate * factor
return mutated_dna
Many mutations modify the structure. Mutations to insert and excise vertex-edge pairs build up a main convolutional
column, while mutations to add and remove edges can handle the skip connections. For example, the AddEdgeMutation
can add a skip connection between random vertices.
class AddEdgeMutation(Mutation):
"""Adds a single edge to the graph."""
def mutate(self, dna):
# Try the candidates in random order until one has the right connectivity.
for from_vertex_id, to_vertex_id in self._vertex_pair_candidates(dna):
mutated_dna = copy.deepcopy(dna)
if (self._mutate_structure(mutated_dna, from_vertex_id, to_vertex_id)):
return mutated_dna
raise exceptions.MutationException() # Try another mutation.
def _vertex_pair_candidates(self, dna):
"""Yields connectable vertex pairs."""
from_vertex_ids = _find_allowed_vertices(dna, self._to_regex, ...)
if not from_vertex_ids:
raise exceptions.MutationException() # Try another mutation.
random.shuffle(from_vertex_ids)
to_vertex_ids = _find_allowed_vertices(dna, self._from_regex, ...)
if not to_vertex_ids:
raise exceptions.MutationException() # Try another mutation.
random.shuffle(to_vertex_ids)
for to_vertex_id in to_vertex_ids:
# Avoid back-connections.
disallowed_from_vertex_ids, _ = topology.propagated_set(to_vertex_id)
for from_vertex_id in from_vertex_ids:
if from_vertex_id in disallowed_from_vertex_ids:
continue
# This pair does not generate a cycle, so we yield it.
yield from_vertex_id, to_vertex_id
def _mutate_structure(self, dna, from_vertex_id, to_vertex_id):
"""Adds the edge to the DNA instance."""
edge_id = _random_id()
edge_type = random.choice(self._edge_types)
if dna.has_edge(from_vertex_id, to_vertex_id):
return False
else:
new_edge = dna.add_edge(from_vertex_id, to_vertex_id, edge_type, edge_id)
Large-Scale Evolution
...
return True
For clarity, we omitted the details of a vertex ID targeting mechanism based on regular expressions, which is used to
constrain where the additional edges are placed. This mechanism ensured the skip connections only joined points in the
“main convolutional backbone” of the convnet. The precedence range is used to give the main backbone precedence over
the skip connections when resolving scale and depth conflicts in the presence of multiple incoming edges to a vertex. Also
omitted are details about the attributes of the edge to add.
To evaluate an individual’s fitness, its DNA is unfolded into a TensorFlow model by the Model class. This describes how
each Vertex and Edge should be interpreted. For example:
class Model(object):
...
def _compute_vertex_nonlinearity(self, tensor, vertex):
"""Applies the necessary vertex operations depending on the vertex type."""
if vertex.type == LINEAR:
pass
elif vertex.type == BN_RELU:
tensor = slim.batch_norm(
inputs=tensor, decay=0.9, center=True, scale=True,
epsilon=self._batch_norm_epsilon,
activation_fn=None, updates_collections=None,
is_training=self.is_training, scope=’batch_norm’)
tensor = tf.maximum(tensor, vertex.leakiness * tensor, name=’relu’)
else:
raise NotImplementedError()
return tensor
def _compute_edge_connection(self, tensor, edge, init_scale):
"""Applies the necessary edge connection ops depending on the edge type."""
scale, depth = self._get_scale_and_depth(tensor)
if edge.type == CONV:
scale_out = scale
depth_out = edge.depth_out(depth)
stride = 2**edge.stride_scale
# ‘init_scale‘ is used to normalize the initial weights in the case of
# multiple incoming edges.
weights_initializer = slim.variance_scaling_initializer(
factor=2.0 * init_scale**2, uniform=False)
weights_regularizer = slim.l2_regularizer(
weight=self._dna.weight_decay_rate)
tensor = slim.conv2d(
inputs=tensor, num_outputs=depth_out,
kernel_size=[edge.filter_width(), edge.filter_height()],
stride=stride, weights_initializer=weights_initializer,
weights_regularizer=weights_regularizer, biases_initializer=None,
activation_fn=None, scope=’conv’)
elif edge.type == IDENTITY:
pass
else:
raise NotImplementedError()
return tensor
The training and evaluation (Section 3.4) is done in a fairly standard way, similar to that in the tensorflow.org tutorials for
image models. The individual’s fitness is the accuracy on a held-out validation dataset, as described in the main text.
Parents are able to pass some of their learned weights to their children (Section 3.6). When a child is constructed from a
parent, it inherits IDs for the different sets of trainable weights (convolution filters, batch norm shifts, etc.). These IDs are
embedded in the TensorFlow variable names. When the child’s weights are initialized, those that have a matching ID in
the parent are inherited, provided they have the same shape:
graph = tf.Graph()
Large-Scale Evolution
with graph.as_default():
# Build the neural network using the ‘Model‘ class and the ‘DNA‘ instance.
...
tf.Session.reset(self._master)
with tf.Session(self._master, graph=graph) as sess:
# Initialize all variables
...
# Make sure we can inherit batch-norm variables properly.
# The TF-slim batch-norm variables must be handled separately here because some
# of them are not trainable (the moving averages).
batch_norm_extras = [x for x in tf.all_variables() if (
x.name.find(’moving_var’) != -1 or
x.name.find(’moving_mean’) != -1)]
# These are the variables that we will attempt to inherit from the parent.
vars_to_restore = tf.trainable_variables() + batch_norm_extras
# Copy as many of the weights as possible.
if mutated_weights:
assignments = []
for var in vars_to_restore:
stripped_name = var.name.split(’:’)[0]
if stripped_name in mutated_weights:
shape_mutated = mutated_weights[stripped_name].shape
shape_needed = var.get_shape()
if shape_mutated == shape_needed:
assignments.append(var.assign(mutated_weights[stripped_name]))
sess.run(assignments)
S2. FLOPs estimation
This section describes how we estimate the number of floating point operations (FLOPs) required for an entire evolution
experiment. To obtain the total FLOPs, we sum the FLOPs for each individual ever constructed. An individual’s FLOPs
are the sum of its training and validation FLOPs. Namely, the individual FLOPs are given by Ft Nt + Fv Nv , where Ft is
the FLOPs in one training step, Nt is the number of training steps, Fv is the FLOPs required to evaluate one validation
batch of examples and Nv is the number of validation batches.
The number of training steps and the number of validation batches are known in advance and are constant throughout the
experiment. Ft was obtained analytically as the sum of the FLOPs required to compute each operation executed during
training (that is, each node in the TensorFlow graph). Fv was found analogously.
Below is the code snippet that computes FLOPs for the training of one individual, for example.
import tensorflow as tf
tfprof_logger = tf.contrib.tfprof.python.tools.tfprof.tfprof_logger
def compute_flops():
"""Compute flops for one iteration of training."""
graph = tf.Graph()
with graph.as_default():
# Build model
...
# Run one iteration of training and collect run metadata.
# This metadata will be used to determine the nodes which were
# actually executed as well as their argument shapes.
run_meta = tf.RunMetadata()
with tf.Session(graph=graph) as sess:
feed_dict = {...}
_ = sess.run(
Large-Scale Evolution
[train_op],
feed_dict=feed_dict,
run_metadata=run_meta,
options=tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE))
# Compute analytical FLOPs for all nodes in the graph.
logged_ops = tfprof_logger._get_logged_ops(graph, run_meta=run_metadata)
# Determine which nodes were executed during one training step
# by looking at elapsed execution time of each node.
elapsed_us_for_ops = {}
for dev_stat in run_metadata.step_stats.dev_stats:
for node_stat in dev_stat.node_stats:
name = node_stat.node_name
elapsed_us = node_stat.op_end_rel_micros - node_stat.op_start_rel_micros
elapsed_us_for_ops[name] = elapsed_us
# Compute FLOPs of executed nodes.
total_flops = 0
for op in graph.get_operations():
name = op.name
if elapsed_us_for_ops.get(name, 0) > 0 and name in logged_ops:
total_flops += logged_ops[name].float_ops
return total_flops
Note that we also need to declare how to compute FLOPs for each operation type present (that is, for each node type in the
TensorFlow graph). We did this for the following operation types (and their gradients, where applicable):
• unary math operations: square, squre root, log, negation, element-wise inverse, softmax, L2 norm;
• binary element-wise operations: addition, subtraction, multiplication, division, minimum, maximum, power, squared
difference, comparison operations;
• reduction operations: mean, sum, argmax, argmin;
• convolution, average pooling, max pooling;
• matrix multiplication.
For example, for the element-wise addition operation type:
from tensorflow.python.framework import graph_util
from tensorflow.python.framework import ops
@ops.RegisterStatistics("Add", "flops")
def _add_flops(graph, node):
"""Compute flops for the Add operation."""
out_shape = graph_util.tensor_shape_from_node_def_name(graph, node.name)
out_shape.assert_is_fully_defined()
return ops.OpStats("flops", out_shape.num_elements())
S3. Escaping Local Optima Details
S3.1. Local optima and mutation rate
Entrapment at a local optimum may mean a general lack of exploration in our search algorithm. To encourage more
exploration, we increased the mutation rate (Section 5). In more detail, we carried out experiments in which we first
waited until the populations converged. Some reached higher fitnesses and others got trapped at poor local optima. At this
point, we modified the algorithm slightly: instead of performing 1 mutation at each reproduction event, we performed 5
mutations. We evolved with this increased mutation rate for a while and finally we switched back to the original singlemutation version. During the 5-mutation stage, some populations escape the local optimum, as in Figure 4 (top), and none
Large-Scale Evolution
get worse. Across populations, however, the escape was not frequent enough (8 out of 10) and took too long for us to
propose this as an efficient technique to escape optima. An interesting direction for future work would be to study more
elegant methods to manage the exploration vs. exploitation trade-off in large-scale neuro-evolution.
S3.2. Local optima and weight resetting
The identity mutation offers a mechanism for populations to get trapped in local optima. Some individuals may get
trained more than their peers just because they happen to have undergone more identity mutations. It may, therefore,
occur that a poor architecture may become more accurate than potentially better architectures that still need more training.
In the extreme case, the well-trained poor architecture may become a super-fit individual and take over the population.
Suspecting this scenario, we performed experiments in which we simultaneously reset all the weights in a population that
had plateaued (Section 5). The simultaneous reset should put all the individuals on the same footing, so individuals that
had accidentally trained more no longer have the unfair advantage. Indeed, the results matched our expectation. The
populations suffer a temporary degradation in fitness immediately after the reset, as the individuals need to retrain. Later,
however, the populations end up reaching higher optima (for example, Figure 4, bottom). Across 10 experiments, we find
that three successive resets tend to cause improvement (p < 0.001). We mention this effect merely as evidence of this
particular drawback of weight inheritance. In our main results, we circumvented the problem by using longer training
times and larger populations. Future work may explore more efficient solutions.
| 9cs.NE
|
1
Dispersion and Line Formation in Artificial Swarm
Intelligence
DONGHWA JEONG and KIJU LEE, Case Western Reserve University
1.
INTRODUCTION
One of the major motifs in collective or swarm intelligence is that, even though individuals follow
simple rules, the resulting global behavior can be complex and intelligent. In artificial swarm systems,
such as swarm robots, the goal is to use systems that are as simple and cheap as possible, deploy
many of them, and coordinate them to conduct complex tasks that each individual cannot accomplish
[Cao 1995; Simmons 2000; Sahin 2005]. The system may involve a group of artificially intelligent
agents with limited sensing, communication, and computing capabilities and can achieve a higher
level task by the agents locally or globally collaborating with each other. Local collaboration can be
established among the neighbors within a certain communication range or physical/social distance.
Global collaboration can be realized by centralized control possibly with a designated leader.
Global shape formation is one of the challenging problems in artificial swarm intelligence. In nature,
it is performed for various purposes, such as engergy saving and path optimization. For instance, a
flock of large birds fly together while forming a V shape in order to reduce the air resistance and
fatigue for physically weak birds [Weimerskirch 2001]. Ants form a line to optimize the path to a food
source by laying a trail of pheromone and regulate foraging activities without any central control or
spatial information [Prabhakar 2012; Gordon 2013]. This trail enables the ant colony find the shortest
path to the food source [Bonabeau 2000]. Shape formation in artificial intelligence systems is usually
required for specific task-oriented performance, including 1) forming sensing grids [Spears 2004], 2)
exploring and mapping in space, underwater, or hazardous environments [Nawaz 2010; Wang 2011],
and 3) forming a barricade for surveillance or protecting an area/person [Cheng 2005]. This paper
presents a dynamic model of an artificial swarm system based on a virtual spring damper model and
algorithms for dispersion without a leader and line formation with an interim leader using only the
distance estimation among the neighbors.
2.
DYNAMIC MODEL OF COLLECTIVE SYSTEMS
Our dynamic model of swarm systems involving multiple agents is based on a virtual spring damper
model. To realize attractive and repulsive forces in a stable manner, virtual damper is added to the
spring system to depress the oscillatory motion of the agents. Modeling dynamic multi-agent systems
frequently requires global optimization by designing objective functions. As the number of agents increases, the computational cost for optimization increases exponentially. To address this problem, we
used geometric conditions that guarantee global shape formation from geometric primitives based on
the attractive and repulsive forces between the agents as shown in Fig. 1. Each node and edge represent an agent and a communication/sensing/social connection respectively where the spring constant
and damping ratio may be differently defined depending on each application. The mass of each node
may indicate the weight or relative importance of that agent within the group.
Let A1 , A2 , · · · , An be the agents in two-dimensional space. For the ith agent, the two-dimensional
th
th
Cartesian coordinate is denoted by ~xi = [xi , yi ]T . Likewise,
p the vector from the i agent to the j
T
2
2
agent is given by ~xij = [xi − xj , yi − yj ] where ||~xij || = (xi − xj ) + (yi − yj ) Assuming that the
Collective Intelligence 2014.
1:2
•
D. Jeong and K. Lee
Fig. 1. Attractive and repulsive forces between the nodes are modeled with a spring and a damper.
ith agent has a mass, mi . Then it follows the Newton’s second law: f~im = mi ~x¨i where f~im is the force
at the ith node and ~xi is its position. The force exerted by the spring and damper is described by
f~ik = −kij d~ij and f~ib = −bij ~x˙ ij where kij is the spring constant, bij is the damping coefficient between
agent i and neighboring agent j, and d~ij is a displacement vector between the two. In this model, each
agent knows the distances to its neighbors only. However, d~ij is expressed as a vector since the sum
of multiple interactions to an agent will force the agent to a certain direction in the global coordinate
system. The net force acting on the ith agent in the global coordinate can be derived as
X
f~im = f~ik + f~ib = −
(kij d~ij + bij ~x˙ ij )
j∈Mi
th
where M
Pi is the ~number˙ of neighboring agents of the i agent. The acceleration is calculated as
¨
~xi = − j∈Mi (kij dij + bij ~xij )/mi . In practical control with a microprocessor, the above equation can
be rewritten in discrete time representation given by
X
~x˙ i [t + 1] − ~x˙ i [t] = −
(kij d~ij + bij ~x˙ ij [t])/mi .
j∈Mi
In case of mi = 1 for all i = 1, · · · , n, the anticipated velocity and position at t + 1 and t + 2 are given by
X
X
~x˙ i [t + 1] = ~x˙ [t] −
(kij d~ij + bij ~x˙ ij [t]); ~xi [t + 2] = 2~xi [t + 1] − ~xi [t] −
(kij d~ij + bij ~x˙ ij [t]).
j∈Mi
3.
j∈Mi
DISPERSION AND LINE FORMATION
Algorithms
Two algorithms are introduced here: 1) dispersion based on trigonal plannar elements without a leader
and 2) line formation using paired line elements using an interim leader.
Fig. 2. Elementary geometric shapes: (a) trigonal planar and (b) paired line.
Collective Intelligence 2014.
Dispersion and Line Formation in Artificial Swarm Intelligence
•
1:3
Dispersion using trigonal planar elements. Let N be a set of dynamic agents. For every O ∈ N , label
the closest three neighbor agents as A, B, and C. By using the proposed attractive and repulsive forces,
let LOA = LOB = LOC = Ld where Ld is the desired distance between the agents.
Lining using paired line elements. Let N be a set of dynamic agents. For every O ∈ N , label the
closest agent as A. Let LOA = and form a pair, (O, A). Find the nearest pair and label as (B, C). Let
LOB = LOC = LAB = LAC . Note that LOB = LOC for LOA = LBC = as → 0. When all rectangulars
are connected each other, two ends are designated as interim leaders of the connected chains in order
to stretch the line to opposite directions.
Simulation Results
MATLAB simulation is performed with 50 agents randomly spread in two-dimensional space. The
simple trigonal elements guarantee dispersion of the agents and paired line elements guarantees line
shape while maintaining connectivity as shown in Fig. 3 and 4. If each agent has a limited sensing/communication range, the system may result in multiple disc formation or outliers. Fig. 5 shows
the simulation results for two different communication ranges.
Fig. 3. Dispersion based on trigonal planar elements
Fig. 4. Line formation using paired line elements
Fig. 5. Separated discs and connected disc with different sensing range: (a) 40 inch, (b) 60 inch with 100 agents
Collective Intelligence 2014.
1:4
•
D. Jeong and K. Lee
REFERENCES
Y U. Cao, A S. Fukunaga, A B. Kahng, and, F. Meng. 1995. Cooperative mobile robotics: antecedents and directions. IEEE/RSJ
International Conference on Intelligent Robots and Systems 1, 226–234.
R. Simmons, D. Apfelbaum, W. Burgard, D. Fox, M. Moors, S. Thrun, and, H. Younes. 2000. Coordination for Multi-Robot
Exploration and Mapping. AAAI/IAAI .
E. Sahin. 2005. Swarm Robotics: From Sources of Inspiration to Domains of Application. Swarm Robotics, Lecture Notes in
Computer Science ,3342, 10–20.
H. Weimerskirch, J. Martin, Y. Clerquin, P. Alexandre, and, S. Jiraskova. 2001. Energy saving in flight formation. Nature ,413,
697–698.
S. Shen, N. Michael, Y. Mulgaonkar, and, V. Kumar. 2013. Vision-based State Estimation for Autonomous Rotorcraft MAVs in
Complex Environments. IEEE International conference on Robotics and Automation .
R. W. Beard, J. Lawton, and, F. Y Hadaegh. 2001. A coordination architecture for spacecraft formation control. IEEE Transactions on control Systems Technology ,9, 777–790.
E. Bonabeau, and, G. Theraulaz. 2000. Swarm Smarts. Scientific American ,282(3).
B. Prabhakar, N. Dektar. Katherine, and, D. Y. Gordon. 2012. The Regulation of Ant Colony Foraging Activity without Spatial
Information. PLoS Comput Biol ,8(8).
D. Y. Gordon. 2013. The rewards of restraint in the collective regulation of foraging by harvester ant colonies. Nature ,498.91–93
X. Li, and, X. Liang. 2011. A 3D Local Interaction Strategy for Swarm Robot. Journal of computers ,6(10).2237–2242.
C. M Breder. 1954. Equations descriptive of fish schools and other animal aggregations. Ecology ,35.361–370.
S. Sakai. 1973. A model for group structure and its behavior. Biophysics Japan ,13. 82–90.
H. S Niwa. 1994. Self-organizing dynamic model of fish schooling. J. Theor. Biol. ,171.123–136.
J. A Beecham, and, K. D Farnsworth. 1999. Animal group forces resulting from predator avoidance and competition minimization. J. Theor. Biol. ,198.533–548.
K. Warburton, and, J. Lazarus. 1991. Tendency-distance models of social cohesion in animal groups. J. Theor. Biol. ,150.473–488.
W. M Spears, D. F Spears, J. C Hamann, and, R. Heil. 2004. Distributed, Physics-Based Control of Swarms of Vehicles.
Autonomous Robots ,17.137–162.
S. Nawaz, M. Hussain, S. Watson, N. Trigoni, and, P. N Green. 2010. An Underwater Robotic Network for Monitoring Nuclear
Waste Storage Pools. Sensor Systems and Software ,24.236–255.
Y. Wang, A. Liang, and, H. Guan. 2011. Frontier-based multi-robot map exploration using Particle Swarm Optimization. Swarm
Intelligence , 1–6.
J. Cheng, W. Cheng, and, R. Nagpal. 2005. Robust and self-repairing formation control for swarms of mobile agents. AAAI ,
59–64.
Collective Intelligence 2014.
| 9cs.NE
|
arXiv:1508.07547v2 [cs.PL] 22 Mar 2016
Protocol Programming: A Connection
of the Digital World
Yanping Chen,1∗ Qinghua Zheng,1 Ping Chen2
1
2
Xi’an Jiaotong University, China
University of Massachusetts Boston, USA
E-mail: [email protected], [email protected], [email protected]
The current computer programmings encapsulate attributes and behaviours
into objects, but miss the mechanism to support the connection among objects.
A programming paradigm is presented to connect all objects. The connection
supports communications. Protocols are defined to coordinate the behaviours
between objects, which enable the interaction of objects across different platforms. The connection also provides an efficient mechanism to support the
concurrency, parallelism, distribution, pipeline and adaptability, etc. They
can be governed transparently, autonomously, even adaptively. In this paper,
an implementation is also discussed to show the effectiveness of protocol programming.
Connection is the norm
The connection between entities is an important aspect of the world. Large objects as celestial
bodies or small objects as biomolecules are connected, by gravity or conjugation, and organized
into a galaxy or an organism. The connection should not be treated as an attribute of an object.
It has a powerful influence on the connected objects. For example, gravity controls trajectories
of celestial bodies, chemical reactions determine the transformation between substances. And
relations, e.g., food chain, heredity, etc. judge the fittest and influence the evolution of species.
The connection condenses substances into the cosmos, synthesizes amino acid and brings the
life. All things are influenced or developed by connecting with others.
1
In this article, we define “connection” as the channel for interactions between objects. Usually interactions are not random, they show a clear pattern. The patterned interaction is referred
to as “communication” in this paper. The connection supports communications between objects. It is the core idea of our method, we give an example for easier understanding. Let’s
consider the ocean, the land and the sky as a platform for the the biological world. Animals
wander around the platform. It should be emphasized that the ability for an animal to ramble
is important, referred to as “autonomy”. It enables animals to change to fit its surroundings.
Strolling animals are connected by inhabiting or encountering with each others. In the wild,
passing each other does nothing about the communication. The communication between them
is established by giving birth to, feeding, preying, fighting, etc.
In the digital world, to imitate or encode a system under a programming paradigm, the modularity is an important issue, which supports structured programming. Modules (e.g., classes
or functions (1, 2)) are used to encapsulate objects and partition a system into different parts.
However, in practice, it is common that a complex problem can not be perfectly partitioned. For
example, to design a building, some parts of a building (e.g., the water supply or power system)
are difficult to be encapsulated. This is also known as the crosscutting issue (3). It makes the
code tangled and increases the cost of communication among developers when designing and
developing a system (4). Furthermore, in object-oriented programming, an object is merely
nested in another object as a member variable. The inner object is restricted by the holder that
difficult to be shared with others. It also sacrifices the autonomy of an object, making it difficult
to update independently.
Without a mechanism to support the connection appropriately, a hard partition hurts the
communications between objects. In the traditional programming paradigms, various approaches
are applied to support communications between objects. For example, objects can be used as
function parameters or member variables of other objects. Inheritance is used to represent relations between classes. It does nothing about the communication unless static members are
defined in the declaration (5). Then instantiated objects can share the same variables for communications. Others such as message passing, semaphore, shared memory, socket, etc. are
widely adopted to support communication. Because the lack of the connection between objects,
these approaches are used to support the communication weakened by separating a system into
discrete parts. Traditionally, the connection as a mechanism to support communications has received less attention. The communications between objects are temporary and irregular, mainly
used to exchange data or control resource competition if required. It is rarely formalized as a
2
connection mechanism to support communications.
Compared to the traditional paradigms, connection in protocol programming is developed
into a new territory. Under the protocol programming, programming is divided into two steps:
protocol implementation and domain implementation. Protocol implementation builds an infrastructure to support the communication between objects. Protocols are defined to control
the communication between them. By inheritance from a protocol implementation all instantiated objects are automatically connected into a network. Domain implementation is developed
by traditional programming paradigms, encapsulating attributes and behaviours into objects.
While supported by protocol implementation, partitioning a system into discrete models and
reorganizing them become easier and more effective.
Connection creates possible
Building a system is usually not a trivial matter. It is often partitioned into discrete modules
(e.g., procedures, functions, objects, agents, aspect, etc.), then a team is set up to implement
each module separately. It is generally accepted that a modular design is the key to successful
programming. All programming paradigms have mechanisms to support structured programming, where the modularity is guaranteed. Systems developed by structured programming have
good stability, efficiency, maintainability and expansibility. The main obstacle for structured
programming is that some problems are difficult to be partitioned. In some projects, design
decision is scattered throughout a system. A hard partition breaks the connections between
modules and hurts the communications between them.
In the digital world, to support the partition and connection between modules, many program paradigms are developed, e.g., agent-oriented programming (6), aspect-oriented programming (7). Based on message passing, mechanisms are proposed for communicating between
threads or processes (e.g., Message Passing Interface (8), Erlang (9), Hoare et al. (10)). The
communication across different systems are also designed (e.g., MapReduce (11), serviceoriented computing (12), CORBA (13)). In a program, mechanisms (e.g., semaphore, shared
memory, socket, message passing) are used to support the communication between partitioned
modules. Because implementation of these mechanisms is not transparent to programmers,
it may lead to many issues such as mutual exclusion, interrupts, data conflict, resource competition, etc. When building a system, these mechanisms increase the burden of design and
implementation, and make the code difficult to understand and maintain. Furthermore, when a
3
system is complex enough, partitioning it, defining the interfaces and integrating every part are
very challenging. The cost to design, develop and maintain such a system and the communication between developers are increased considerably (4).
To better illustrate our ideas, we analyse two systems (the Internet and the biosphere), which
show the motivation of protocol programming.
Since the invention of packet switching theory (14), for half a century, the Internet has expanded into a huge system with billions of access points and numerous applications (15). At
the dawn of computer communication (16), it is difficult to foresee that human society can build
such a complex system. Despite various techniques are developed to accelerate the growth of the
Internet, the most important technique is the Internet protocol deigned for the open-architecture
network environment. Under the internet protocols, every equipment is independently connected into the Internet and the information between them are routed automatically. The Internet protocol hides the underneath implementation of connected modules. By following the
Internet protocol, each part of the Internet can conveniently interact with each other across heterogeneous platforms or systems. Moreover, because the Internet protocol is a connectionless
service, each part can be updated independently.
The biosphere is a largely self-regulated system. The earth can be seen as a platform for life
to compete. Individual organisms of the earth encounter on the land, in the air or the water. The
connection (encounter) is an important aspect for life on the earth. After the connection was
constructed, the interaction between them is not random. They follow inherent patterns (e.g.,
foster, prey, competition, hereditary and variation) in the same way as the protocol. Life in the
biosphere is ruled by these patterns, referred to as laws in biology. Hereditary and variation are
two important laws determining the survival of a species and the evolution of the living nature.
Many factors can disturb the natural stability of the biosphere, e.g., the geographical or climatic
factor, the species diversity and the food chain, etc. They are the motive force for the species
variation. The variation enables that a species can adapt to the environmental change. “survival
of the fittest” ensures that the suitable change is inherited. Accumulated changes in individuals
or species give rise to the evolution of the species.
Designing or constructing a system like the Internet or the biosphere in one step is far beyond
human ability. They are not built, but evolved. Five characteristics can be induced from such
systems. Each system should consist of modules or individuals that are mutually independent,
which enables the separableness. Individuals can interact with each other and the connection
is guaranteed. The interactions between individuals are not random, they follow predefined
4
or evolved patterns. Then the communication between them is established. For a system that
supports evolution, individuals have the ability to change, which enables the fittest is reserved,
and the adaptability is possible.
A system is a set of interdependent and interacting modules (components or individuals),
which forms an integrated whole. If partitioned modules are unrelated, the system is static and
inactive, shows no interaction. In a dynamic, autonomous or adaptive system, the interconnectivity between modules is required. The separableness ensures that a system can be partitioned
into modules, so each module can be designed or developed independently. It is the premise
for concurrency, parallelism and distribution. Modules of a system can be connected by structural, functional, behavioural relationships. The connection bands partitioned modules together.
Then the interaction between modules would occur frequently. Many interactions are not random. They will show some kind of patterns (e.g., protocols and laws in the Internet and the
biosphere). These patterns are the manufactured products of the communication, which makes
them coordinate, compete, develop and struggle for existence. The connection supports communications. In traditional paradigms of the digital world, the communication is supported by
limited approaches, e.g., semaphore, shared memory. They are mainly designed for exchanging
data between partitioned modules in a system. No valid channel is provided for every object
that they can communicate without any constraints.
Motivated by issues discussed above, a protocol programming paradigm is presented for
designing and implementing a system. Protocols are predefined communicating patterns. They
are used to connect and coordinate each module in a system. In this paradigm, programming is
divided into two aspects: protocol implementation and domain implementation. Protocol implementation provides a platform for connecting or coordinating instantiated objects. Domain
implementation can use the traditional methods to encapsulate objects. But supported by a protocol implementation, when designing and encoding a system, the development process become
simple. Because the communication can be designed to be connectionless, modules can have
the ability to update itself without stopping the system.
In order to show the methodology of protocol programming, next, we give a simple discussion about the implementation developed in our work.
5
Protocol Implementation
Our protocol implementation has two layers: transmission layer and service layer. The transmission layer supports the connection, providing a channel for communication. In this layer, a
protocol is designed to establish a thread network, referred to as the connecting network. The
service layer supports the connected objects to register, require and execute a service according
to a defined protocol. The framework about the programming paradigm is shown in Figure 1.
Object
2
Working Thread
2
2
2
Domain Layer
Domain Imp.
2
Acting Thread
Service Layer
Connecting Thread
Configuring Thread
Connecting Network
Transmission Layer
Protocol Imp.
Executing Thread
Figure 1: Protocol Programming Architecture
The protocol implementation is divided into different layers. Each layer is implemented by
a class. If a class implements a non-bottom layer, this class should inherit from a parent class
that implements a lower layer. It is better that each layer is designed to be transparent, only the
interface is known by the upper layer the same as the OSI model (17). Every object instantiated from the transmission layer is connected into the connecting network, where all messages
between objects are transmitted automatically. In the service layer, the service configuration,
registration, request and execution are performed. In the domain layer, each object implements
specific work. Next, the paradigm is discussed.
In the protocol implementation, the connection and communication are done by transmitting
messages from one thread to another. Unlike transmitting an IP package through the Internet,
in a local environment, all messages share the same memory space. Therefore, only the address
of message is posted between threads. The structure of a message is given as follows, where
6
PMessage is a class. Every message is an instantiation of this class or its sub-class.
PMessage{
LAYER m Layer;
ACTION m Action;
ADD m Intend;
ADD m Creator;
bool m NeedEchoF lag;
size t m T imeT oLive;
TIME m CreateT ime;
vectorhADDi m Route v;
vectorhstringi m P aram v; }
The LAYER and ACTION are enumerated data types. The field m Layer is used to indicate
which layer a message should be delivered for further processing. Three types are defined, representing each layer respectively. Each layer supports a certain numbers of actions suggested
in the field m Action. The ADD represents the address (or identification) of an object. The
fields m Intend and m Creator are the destination and creator of a message. In the field
m NeedEchoF lag, a true value indicates that a response is required by m Creator. The
m T imeT oLive is a hop count used to limit a message’s lifetime. The m CreateT ime can
be used to line up arrived messages for avoiding problems such as data access violation when
accessing a critical resource. When a message is transmitted through the network, the field
m Route v can be used to record the route information. The field m P aram v stores parameters of message.
Transmission Layer
The main purpose of the transmission layer is to build an infrastructure for message transmission. We design a class named as PTrans to do this. Three threads are created in each
instantiated object of PTrans: configuring thread, connecting thread and acting thread, used for
configuration, connection and action respectively. They run concurrently.Five kinds of actions
are defined in the transmission layer: CONFIG, HELLO, ECHO, TICK and EXIT.
All messages are posted through the connecting thread. A connecting thread retrieves messages from its message queue. If a message’s destination matches itself, the message is taken
off. Otherwise, it is retransmitted according to a routing protocol. Because other messages in
its message queue may be waiting for transmitting, it is not reasonable to spend too much time
for processing a message in the connecting thread. The connecting thread pushes its message
7
into a transmitting message list and return immediately. All messages in the list are processed
by the acting thread. If the message list is empty, the acting thread is blocked. If a message is
added, the connecting thread sets a semaphore. Then the acting thread is waken up to process
the arrived message.
In protocol programming, the most important problem is how to make every object connected. The topology of connecting network determines the strategy to establish it. In our
implementation, we design a tree structure. The object in the structure also is referred to as
node. The Master is the root of the structure and a leaf node is an object without any child. In
our implementation, every object is identified by the connecting thread handle. The child object
(the child) is used as a member variable of a parent object (the parent). The child has its parent’s
connecting handle as a parameter for initialization. Then, after all objects are initialized, except
the root object, every object knows its parent’s connecting handle.
After a child is initialized, using parent’s connecting handle, the child reports its connecting
handle to its parent by posting a CONFIG message, then waits for a response. The parent
collects the child’s identification from the message and posts a message in return. After this
process, each pair of parent and child know the connecting threads between each other. All
configuring threads exist except that in the Master, which will be running until the system
terminates.
The configuring thread in Master broadcasts a HELLO message to all its children to start a
connecting process. The Master will wait until all responses of its children are returned. After
a HELLO is received from its parent, every object also broadcasts a HELLO to its children.
It responds an ECHO to its parent in two conditions: all feedbacks from children are returned
or it has not child at all. In a concurrent environment, some threads may be blocked or suspended, if a time-out error occurs,, the Master can restart a new connecting process. When
children’s ECHO arrive at the Master, it means that every object is instantiated successfully.
A connecting tree structure is established.
All messages are transmitted in the connecting network. If there is no message to process, a
connecting thread is blocked by calling a GetMessage(&msg, ...) function. A blocked connecting thread is aroused by arrived messages. Because all threads are running concurrently that
some threads may be jammed or blocked, the TICK message is used to check the state of an
object. The EXIT is used for system to exit. Every node retransmits this message to its children.
All nodes from the leaves to the root will release its resources and send a response to its parent.
It the arrived message does not belong to the transmission layer, the acting thread calls a
8
pure virtual function Trans Outlet to process it.
Interface
The connecting thread and the Trans Outlet are the interface between the transmission layer
and the service layer. In the service layer, a requirement is submitted by using the handle as
parameter to post messages, and the Trans Outlet is overridden to provide services.
Service Layer
In this paper, a domain service (or a service for short) is defined as a function that can be
executed to support a request in an application. It is provided in the domain layer. The service
layer doesn’t provide any domain service. The major function of the service layer is to supervise
the services provided in the domain layer.
The service layer is implemented by a PService class inherited from the PTrans. The pure
virtual function Trans Outlet in transmission layer is overridden in the PService class. Because the PTrans is inherited, any object instantiated from the PService class is automatically
connected into the connecting network. If a matched message cannot be processed in the transmission layer, the function Trans Outlet will be called. Three actions are defined in the service
layer: REG, REQ and ACK.
Because the implementation of service layer is transparent and independent from the domain
service. Therefore, the service layer should be given the information about a domain service.
A PServicedepict class is defined to contain this information, named as the service description. Each service must have a service description. The content of the service description is
introduced next.
The service register is started by the Master through broadcasting a REG message to its
children. The REG message is spread from the root node to all leaf nodes. Every node collects
service descriptions from its children and reports all services it knows to its parent. When the
service description transmits through the connecting network, routing information is collected.
With this protocol, every object contains service descriptions about it and its children. The
Master knows the all services in a system.
If a service is requested by an object, it is done by packaging a REQ message into the
connecting network, where the name of the service and possible parameters are filled. In our
implementation, a string matching strategy is used to match a service. When requesting a
9
service, the synchronization and asynchronization should be considered. In asynchronously
calling, the object keeps going after a requirement was posted. Otherwise, it is blocked to wait
an ACK message. The message is packaged by calling a requesting function defined in the
PService class, then transmitted from the requesting object towards the Master. If an object
has a service description matching this requirement, the message is retransmitted to the provider
directly by routing information in the service description.
If a REQ message arrives at its provider, the message is appended into a local message
list of the PService class, named as the service message list. An executing thread defined in
the class is always waiting for processing each message respectively. Two approaches can be
used to implement a service: a thread or a function. Their addresses are stored in the service
description. In the service description, two fields are declared to indicate which approach is
supported and whether or not the service is reenterable.
Many mechanisms can be designed to support the execution of a service. For example, for
a specific service, if the service is executed by a thread, then the service message can be posted
into the thread directly. If a service is declared in a class in the domain layer, then for each
arrived message, a new object of the class can be instantiated to handle this service. If the codes
are reenterable, it is possible to create a new thread to process each message separately and
concurrently. Furthermore, the dynamic scheduling can be designed to support the load balancing. These mechanisms are automatically supervised in the service layer. Only the service
descriptions are shared between the service layer and the domain layer.
After a REQ message is executed, if a response is required in the synchronously calling, it
will return an ACK message.
Interface
One interface of the service layer is the requesting function, where the REQ message is generated and posted into the connecting network. Another interface is a service description list. It is
a member variable of the PService class used by the domain layer to report its services.
Domain Implementation
In domain implementation, the traditional programming methods can be used to provide the
domain services. However, by inheriting a protocol implementation, all instantiated objects are
connected into the connecting network. Every connected object can communicate with each
10
other by predefined protocols. Therefore, the difficulty to partition a problem into discrete tasks
is reduced. Following the predefined protocol, every module can be developed independently.
Because the concurrency, distribution, or load balancing, etc. can be supported transparently by
the protocol implementation, it also reduces the burden to develop the domain implementation.
Evaluation
Three classes, inherited from the PService, are declared in the domain layer. Each provides
a toy service, named as Service1, Service2 and Service3. They separately execute 1 × 106 ,
2 × 106 and 1 × 106 Floating Multiply Instructions (FMI). Three objects are instantiated from
each class respectively. They are connected into a connecting network with a Master object
instantiated from the PService class.
In the first experiment, the monotone model runs every service in a single function without
transmitting message. In the sequence model, the Master orderly executes each service by
synchronously posting the REQ message to the three objects. In the pipeline model, the objects
are organized in a line and asynchronously called from one to another. They are executed
simultaneously. Because the computation complexity of the Service2 is doubled, it becomes
the bottleneck in the pipeline model. In the super pipeline, two objects are instantiated to
provide the Service2 service. Running time1 of the four models is shown in Figure 2(a).
5
16
14
105
104
103
4
10
Speedup
Time (sec.)
12
106
Monotone
Sequence
Pipeline
Super pipeline
8
6
3
2
4
1
2
100 200 300 400 500 600 700 800 900 1000
100 200 300 400 500 600 700 800 900 1000
Rounds
Rounds
(a) Time Consuming
(b) Speedup of Super Pipeline
Figure 2: Evaluation of Protocol Programming
In Figure 2(a), the performance of the monotone model can be seen as a benchmark. In the
1
The evaluation uses a Window 7 OS with Intel i5-2400 3.10GHz 4 Core CPU and a 8 GB memory.
11
sequence model, because synchronously calling is used, the time consuming is increased due to
message transmitting. In the pipeline model, the expense caused by transmitting messages can
be offset by allowing the services to be processed simultaneously. The super pipeline has a load
balancing mechanism, therefore, the highest performance is received.
In Figure 2(b), the speedup2 of the super pipeline model is compared using different numbers of FMI. When the number of FMI is decreased exponentially in each service, the speedup
is decreased accordingly. When the number of FMI is decreased to 103 , the cost of transmitting
message negated the superiority of the super pipeline. Therefore, the speedup is less than 1.
The implementation has a very high speedup, because, instead of a single thread, multi
threads are used to implement the services. The main advantage of protocol programming is
that the concurrency is managed automatically by the protocol implementation.
Discussion
In protocol programming, three types of objects exist: the Master object, the Server object
and the Router object. The Master is designed for governing a system. It is also helpful to
use multi-master objects in a system. Each performs a specific assignment. The Server refers
to the object who provides the service. A Router is a node, which only transmits messages and
no service or management is assigned. It can can be used to organize a system or partition the
message space. In a system, mixed objects can be used.
After a system is initialized, every object inherited from the PService class will have at
least three threads. All communications between them are performed by message transmitting.
Therefore, the concurrency is naturally supported. When a system is running, objects can be
added or erased from the connecting network dynamically, which can be used to support the
dynamic scheduling and the load balancing.
Because the connecting network provides a connectionless service to all objects, the service
provided by the program can be updated without stopping the system. Other methods developed in the Internet protocol can be introduced for the protocol implementation, e.g., domain
name service, error control, structured naming strategy, etc. In addition to keeping the program
and the service description locally, Both can be registered and stored in a remote server, then
accessed though the Internet. If a service is supported in a remote system, the message can be
transmitted across the Internet and executed in the remote system. If possible, the program can
2
It is computed by dividing the time consuming in the monotone model and the super pipeline model.
12
also be downloaded directly and run locally.
In summary, the advantages of the protocol programming include: (a) a novel paradigm for
programming is presented, which divides the process into two aspects: protocol implementation
and domain implementation. (b) the connection as a mechanism to support communication for
programming is emphasized and developed. (c) protocols are designed to control the communication between objects, which enables the communications across heterogeneous programs,
systems and the Internet. (d) Because one protocol implementation can be shared by different
applications, it ensures a broad range of code reuse. (e) mechanism such as concurrency, parallelism, distribution, pipeline and adaptability, etc. can be governed transparently, autonomously,
even adaptively. (g) it supports the autonomy of objects, which enables an object to update itself
when the system is running.
References
1. P. Hudak, ACM Computing Surveys 21, 359 (1989).
2. D. J. Armstrong, Communications of the ACM 49, 123 (2006).
3. T. Elrad, M. Aksit, G. Kiczales, K. J. Lieberherr, H. Ossher, Communications of the ACM
44, 33 (2001).
4. F. P. Brooks, The mythical man-month, vol. 1995 (Addison-Wesley Reading, MA, 1975).
5. R. J. Wirfs-Brock, R. E. Johnson, Communications of the ACM 33, 104 (1990).
6. Y. Shoham, Artificial intelligence 60, 51 (1993).
7. G. Kiczales, et al., Aspect-oriented programming (Springer, 1997).
8. W. Gropp, E. Lusk, N. Doss, A. Skjellum, Parallel computing 22, 789 (1996).
9. J. Armstrong, R. Virding, C. Wikström, M. Williams (1993).
10. C. A. R. Hoare, Communications of the ACM 21, 666 (1978).
11. J. Dean, S. Ghemawat, Communications of the ACM 51, 107 (2008).
12. M. P. Papazoglou, Proceedings of the WISE ’03 (IEEE, 2003), pp. 3–12.
13
13. J. Siegel, Communications of the ACM 41, 37 (1998).
14. L. Kleinrock, RLE Quarterly Progress Report 1 (1961).
15. B. M. Leiner, et al., ACM SIGCOMM Computer Communication Review 39, 22 (2009).
16. T. Marill, L. G. Roberts, Proceedings of the AFIPS ’66 (ACM, 1966), pp. 425–431.
17. H. Zimmermann, IEEE Transactions on Communications 28, 425 (1980).
14
| 6cs.PL
|
arXiv:1711.06504v1 [cs.CV] 17 Nov 2017
Detecting hip fractures with radiologist-level
performance using deep neural networks
William Gale∗, Gustavo Carneiro
School of Computer Science
The University of Adelaide
Adelaide, SA 5000
[email protected]
[email protected]
Luke Oakden-Rayner∗, Lyle J. Palmer
School of Public Health
The University of Adelaide
Adelaide, SA 5000
{luke.oakden-rayner,lyle.palmer}
@adelaide.edu.au
Andrew P. Bradley
Faculty of Science and Engineering
Queensland University of Technology
Brisbane, QLD 4001
[email protected]
Abstract
We developed an automated deep learning system to detect hip fractures from
frontal pelvic x-rays, an important and common radiological task. Our system
was trained on a decade of clinical x-rays (≈53,000 studies) and can be applied to
clinical data, automatically excluding inappropriate and technically unsatisfactory
studies. We demonstrate diagnostic performance equivalent to a human radiologist
and an area under the ROC curve of 0.994. Translated to clinical practice, such
a system has the potential to increase the efficiency of diagnosis, reduce the need
for expensive additional testing, expand access to “expert level” medical image
interpretation, and improve overall patient outcomes.
1 Introduction
Hip fractures represent a significant clinical and public health problem worldwide. They are among
the most common causes of hospitalisation, morbidity, and mortality1 in the elderly, with a lifetime
risk of 17.5% for women and 6% for men2 . The all-cause mortality rate is over 20% within one year,
and less than 50% of patients regain the ability to live independently3.
Diagnosis of a fracture is usually made with pelvic x-ray imaging, and such imaging accounts for 6%
of all imaging referrals from the emergency department at our institution, a tertiary public hospital.
To limit misdiagnosis, 5-10% of at-risk patients undergo further imaging, including additional xrays, nuclear medicine bone scans, computed tomography (CT), or magnetic resonance imaging
(MRI), of which only a third demonstrate a fracture4 . Not only does this increase diagnostic costs
and resource utilisation, but without access to these advanced imaging modalities (for example in
remote and under-serviced regions) delayed or missed diagnosis is likely to result in worse patient
outcomes including increased mortality rate5 , length of hospitalisation6 , and cost of care7 .
Recent advances in medical image analysis using deep learning8 have produced automated systems
that can perform as well as human experts in some medical tasks9,10 . Deep learning is a computer
science method that can be used to teach computers to recognise patterns that are useful in discriminating between groups of images, such as images with or without a certain disease8 .
* These authors contributed equally to the work
Highly sensitive and specific automation of hip fracture assessment using x-ray studies would lead
to earlier and more accurate diagnosis and hence improve patient outcomes. Such automation would
also reduce the need for expensive CT and MRI studies, which could improve service efficiency and
increase access to highly accurate detection of hip fractures in under-serviced regions. Automation
could also improve reproducibility, given the reported variation in diagnostic certainty among human
experts of different experience levels11 . Here we investigate the application of deep learning using
convolutional neural networks (CNNs) for the task of fracture detection, and present the first large
scale study where a deep learning system achieves human-level performance on a common and
important radiological task.
2 Dataset
2.1 Developing ground truth labels for hip fractures
Hip fractures are a promising target for machine learning approaches because of the availability of
near-perfect ground truth labels. Clinically, patients with hip fractures do not remain undetected.
Because of the weight bearing nature of the region, clinically ’silent’ fractures rapidly progress to
severe pain and immobility. As such, all patients with hip fractures that have imaging in a hospital
should be identified in the radiology reports, the orthopaedic operative records, or the mortality
records. While a small subset of patients may be lost to the records (due to hospital transfers, for
example), our clinical experience suggests that the ground-truth label accuracy will be >99%.
2.2 Obtaining and efficiently labelling our dataset
Data for this study were obtained from the clinical radiology archive at the The Royal Adelaide
Hospital (RAH), a large tertiary teaching hospital. Ethics approval was granted by the RAH human
research ethics committee. All pelvis x-rays between 2005 and 2015 were included in the study,
obtained with a wide variety of equipment used across the decade during normal clinical practice.
Initial fracture labels were obtained by combining the orthopaedic surgical unit records, and findings
from the radiology report archive (using regular expressions). This combination of sources resulted
in labelling accuracy of around 95%, evaluated on the hold out test set (described below) which was
labelled manually by a radiologist using all of the available sources of information. To improve this
further while avoiding the need for manual review of the entire dataset, a deep learning model was
trained on the original labels and the false positive cases that this model identified were reviewed
by a radiologist. As the "default" label for a case was negative (i.e., the case was not present in
the orthopaedic database, nor had matching keywords in the radiology report), the majority of label
errors were unrecognised fractures (which a well-trained model identifies as false positives). This
process was repeated several times, and finally a single review of the false negatives was performed.
This process improved the label accuracy on the hold-out test set from around 95% to >99.9%, while
only requiring 3371 cases to be labeled by an expert (7.4% of the dataset). All manual review of
cases was performed by a consultant radiologist (Dr Oakden-Rayner).
Each case provided two images, one from each hip, resulting in a total dataset of 53,278 images.
These were randomly divided into a training set (45,492 images), a validation set (4,432 images) for
model selection, and a held-out test set (3,354 images). There was no overlap of patients between
the sets. The test set included only images referred from the emergency department (ED), which was
considered the most clinically challenging setting, where lateral films and cross-sectional imaging
are often not immediately available and management is often required prior to a formal radiology
report. The prevalence of fractures among the patients in the test set was 19%, which is the clinical
prevalence in the ED at our centre. The prevalence of fractures among patients in the training and
validation sets was lower (≈12%) due to the presence of outpatient and inpatient cases.
To ensure an unbiased set of labels for model evaluation, the entire test set was reviewed manually.
2.3 Managing medical data heterogeneity
To deal with the variation inherent in clinical studies, the dataset was processed with a series of
artificial neural networks. Firstly, a small CNN (CNN-frontal) was trained to identify frontal pelvis
x-rays within each “case”, which often include other images like lateral hip films, chest, and spinal
x-rays, requiring the network to learn to discriminate between gross anatomical features. Secondly,
2
Model
Training set
Validation set
Precison
Recall
Accuracy
Parameters
CNN-frontal
CNN-bounding
CNN-metal
581 cases
300 cases
5330 cases
300 cases
440 cases
300 cases
>0.99
>0.99
>0.99
1.0
>0.99
0.97
>0.99
2,408
746,624
11,792
Table 1: Performance of pre-processing deep learning models. CNN-frontal: identifies frontal pelvic
x-rays, and excludes all other images. CNN-bounding: localises the neck of femur region for extraction. CNN-metal: excludes cases with metal in the region of interest. The CNN-metal training set
was identified using regular expressions to find appropriate keywords from the radiology reports.
a regression-based CNN (CNN-bounding) was trained to localise the neck of femur, which is the
only location where a relevant fracture could occur. This reduced the input size of the x-rays from
over 3000 x 3000 pixels to a much more manageable 1024 x 1024 pixels, while maintaining image resolution. This task is more challenging than that of CNN-frontal, as it requires the system
to localise fine-grained anatomical landmarks. Finally, a third CNN (CNN-metal) was trained to
exclude cases with implanted metal from hip fractures and other similar operations (which represent
a separate diagnostic challenge). Each network was trained on a small volume of data, requiring less
than one hour of annotation effort by a radiologist. In Table 1 we present the performance of these
models on unseen validation data. The accuracy of CNN-bounding is estimated by manual review
of a held out test set, where adequacy was defined as coverage including the femoral head, and the
greater and lesser trochanters.
3 Methods
3.1 Model selection
To analyse the pelvic x-rays, we applied a type of CNN known as a DenseNet12 . This architecture
utilises extensive feed-forward connections between the layers, which is thought to improve feature
propagation in these networks.
The validation set was used to determine the following hyper-parameters (using a grid search strategy): the layer width (number of units per layer), the choice of activation function and leak rate, the
use of a secondary loss function, the types and extent of data augmentation, the level of regularisation, and the learning rate.
The final network was 172 layers deep, with 12 features/units per layer (a total of 1,434,176 parameters). We used leaky relu13 non-linear activations with a leak rate of 0.5 and pre-activation batch
normalisation14. The model optimised two loss functions; a primary loss related to the presence
or absence of fractures, and a secondary loss to learn more specific fracture location information
(intra-capsular, extra-capsular, and no fracture). This was motivated by similar approaches which
improved performance in previous work9,10.
We applied extensive data augmentation, consisting of small translations, rotations, and shears, as
well as histogram matching to account for the residual variation in pixel values even after processing (a common issue with medical images). Each augmentation technique resulted in an absolute
improvment of around 0.01 AUC on the validation set.
The network was regularised with a dropout15 rate of 0.2 and a weight decay rate of 1e-5. The
network was trained via stochastic gradient descent using the Adam optimiser16 , with a learning rate
of 0.0001. The network was trained for 25 epochs, with a batch size of 14. The system was trained
using PyTorch17 on a on a workstation comprised of a hexa-core Intel i7-6850k processor, 64GB of
DDR4 RAM and two 12GB NVidia Titan X Pascal graphics cards, resulting in a wall-clock training
time of around 22 hours for the final model.
We also compared the performance of our model against a ten layer fully-convolutional CNN with
4,722,944 parameters, and pre-trained deep neural networks that required downsampling of the input
images. The DenseNet had significantly higher performance on the validation data (absolute AUC
improvement = 0.035 in both cases).
3
3.2 Evaluation
The algorithm was evaluated using a hold-out test set containing 3,354 images, with 348 fractures.
We compared the performance of the algorithm against recently published work on automated hip
fracture detection18, as well as against the original radiology reports to achieve an estimate of human
expert performance.
3.3 Estimating human performance
Our task of identifying hip fractures using only a frontal pelvic x-ray is a common clinical one, as
often the frontal film is the only test that is available to make this diagnosis. However, assessing human performance from radiology reports is confounded by the fact that other sources of information
can be available at the time of reporting, and the role that this information plays in the diagnosis is
usually unstated in the radiology report. Sources of such information could include discussions with
clinicians (including physical examination and surgical findings), or other imaging such as lateral xray images or follow-up films (such as repeat x-ray, CT or MRI). At our centre the x-ray reports are
often only finalised several hours after imaging, which allows radiologists to “peek” at any follow
up imaging.
To provide the strongest possible baseline we generate an estimate of the “upper bound” of human
performance. That is, we assume that no unstated information was used to report the films. A report
is only considered a human error if it provides the wrong diagnosis (later proven on follow-up), if
it clearly states that additional information was required to makle the diagnosis (i.e., a lateral film),
or it is unequivocal that further imaging will be necessary to make a diagnosis. When reports had
vague wording or it was unclear if the radiologist was recommending further imaging, we treated
the stated diagnosis as the result of the frontal x-ray only. For example, "there is a subtle fracture,
consider CT to confirm" would be treated as a finding of a fracture, but "equivocal appearance, CT
recommended" or "the fracture was only demonstrated on the lateral film" would both be considered
a failure to make the diagnosis with the frontal film alone.
4 Results
Figure 1 shows the receiver operator characteristic curve for our model on the hold-out test data.
The human upper bound performance is also presented as a point in ROC space.
Figure 1: ROC curve showing the performance of the model with AUC 0.994, with a point reflecting
the optimistic upper bound of human performance.
4
Model
Accuracy
Precision
Recall
F1
Our model
Kazi et al. (STN)
Kazi et al. (LBM)
Kazi et al. (UBM)
0.97
0.84 (-13)
0.81 (-16)
0.88 (-9)
0.99
0.74 (-27)
0.76 (-25)
0.91 (-8)
0.95
0.93 (-2)
0.84 (-11)
0.85 (-10)
0.97
0.82 (-15)
0.80 (-17)
0.88 (-9)
Table 2: Comparison against recent research. STN: spatial transformer network. LBM: lower bound
model. UBM: upper bound model. The upper bound model presented by Kazi et al. required human
localisation of the neck of femur region, our system performs this localisation automatically.
Model
Acc (CI95% )
Prec (CI95% )
Rec (CI95% )
F1 (CI95% )
Radiologist (estimate)
Our results (high prec)
Our results (high sens)
0.99 (99-100)
0.99 (99-100)
0.99 (99-100)
0.93 (90-95)
0.97 (95-99)
0.92 (89-94)
0.97 (95-99)
0.92 (89-95)
0.95 (92-97)
0.95 (93-97)
0.95 (93-97)
0.94 (92-97)
Table 3: Comparison against estimated human baseline. The upper bound of human performance is
estimated from the original (clinical) radiology reports for the test set cases. This reflects the upper
limit of human performance given the data as we assume the radiologists used no additional information other than the frontal pelvic x-rays for their diagnoses (which is unlikely). As demonstrated,
there is no significant difference between the human upper bound estimate and our model at either
operating point on this dataset.
In Table 2 we compare our model to recently published results on automated hip fracture detection18 .
Kazi et al. developed several models using a dataset of 669 frontal pelvic x-rays. They split these
cases into two separate hip images, resulting in 900 images for training and 270 for testing. Their
test data had a prevalence of 50% and they presented results at a single operating point for each
model. For consistency, we present results here with a similar data distribution (prevalence = 50%)
using all 348 fractures from our test set and 348 randomly selected non-fracture test cases. This
subset was selected so the performance metrics were comparable (given the use of precision in Kazi
et al., which varies with prevalence).
In Table 3 we present results for the entire larger test set (348 fractures and 2997 non-fractures),
comparing our results against the performance on the original radiology reports. Note that in these
tests the radiologists often had access to more information than our models, including any lateral
x-rays, clinical information and follow-up studies. We present the performance of our system at two
operating points, with high precision and high recall. Confidence intervals are calculated using exact
Clopper-Pearson intervals20 .
5 Discussion
We present a deep learning model that achieves human-level performance at the important radiology
task of hip fracture detection in a large-scale study, and demonstrate state of the art performance
compared to recently published work in this area by a large margin. The ROC AUC value of 0.994
is, to the best of our knowledge, the highest level ever reported for automated diagnosis in any largescale medical task, not just in radiology. Furthermore, unlike in previous work, our fully automated
pipeline can take in any frontal pelvis x-ray and exclude ineligible cases, localise the neck of femur,
and identify the presence of proximal femoral fractures automatically. Our research shows that
despite the challenges specific to radiological image data, the development of large, clean datasets
is sufficient to achieve high-level automated performance with deep-learning systems.
Our method of using small CNNs trained on small labeled datasets (300 cases for each model,
taking less than one hour for an expert to annotate each dataset) made processing clinical image
data practical. Many of these ’simple’ tasks such as anatomy localisation can be achieved with
very high accuracy with only a modest expert time commitment. Likewise, our process of iterative
labelling using partially trained neural networks (where an expert reviews the false positives and
false negatives rather than the entire dataset) allowed us to improve label accuracy from around 95%
to over 99.9% while only hand-labelling 7.4% of the dataset.
5
We also show several further new results; that high performance (AUC = 0.994) can be achieved
without pre-training of models (for example, on ImageNet data as demonstrated in Gulshan et al. and
Esteva et al.9,10 ), and that very deep networks are useful in this setting. While these results should
be expected given the results of CNNs in image analysis more generally, we have anecdotally seen
many discussions around these points and we hope our findings can help to answer these concerns.
Baseline human performance on this task is difficult to estimate due to the presence of unreported
external sources of information which can contribute to diagnostic decisions, and the use of inexact
language by radiologists. While we are confident that our upper bound estimate of human performance is optimistic, and our team is currently working to test our system against human experts
directly in a controlled environment.
References
[1] Brauer, C. A., Coca-Perraillon, M., Cutler, D. M. & Rosen, A. B. Incidence and mortality of hip fractures
in the United States. Jama 302, 1573-1579 (2009).
[2] Kannus, P. et al. Epidemiology of hip fractures. Bone 18, S57-S63 (1996).
[3] Morrison, R. S., Chassin, M. R. & Siu, A. L. The medical consultant’s role in caring for patients with hip
fracture. Annals of internal medicine 128, 1010-1020 (1998).
[4] Cannon, J., Silvestri, S. & Munro, M. Imaging choices in occult hip fracture. The Journal of emergency
medicine 37, 144-152 (2009).
[5] Shiga, T., Wajima, Z. i. & Ohe, Y. Is operative delay associated with increased mortality of hip fracture
patients? Systematic review, meta-analysis, and meta-regression. Canadian Journal of Anesthesia 55, 146
(2008).
[6] Simunovic, N., Devereaux, P. & Bhandari, M. Surgery for hip fractures: Does surgical delay affect outcomes? Indian journal of orthopaedics 45, 27 (2011).
[7] Shabat, S. et al. Economic consequences of operative delay for hip fractures in a non-profit institution.
Orthopedics 26, 1197-1199 (2003).
[8] LeCun, Y., Bengio, Y. & Hinton, G. Deep learning. Nature 521, 436-444 (2015).
[9] Gulshan, V. et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA 316, 2402-2410 (2016).
[10] Esteva, A. et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature 542,
115-118 (2017).
[11] Collin, D., Dunker, D., Göthlin, J. H. & Geijer, M. Observer variation for radiography, computed tomography, and magnetic resonance imaging of occult hip fractures. Acta Radiologica 52, 871-874 (2011).
[12] Huang, G., Liu, Z., Weinberger, K. Q. & van der Maaten, L. Densely connected convolutional networks.
arXiv preprint arXiv:1608.06993 (2016).
[13] Maas, A. L., Hannun, A. Y., and Ng, A. Y. (2013). Rectifier nonlinearities improve neural network acoustic
models. In Proc. ICML, volume 30.
[14] Ioffe, S. & Szegedy, C. in International Conference on Machine Learning. 448-456.
[15] Srivastava, N., Hinton, G. E., Krizhevsky, A., Sutskever, I. & Salakhutdinov, R. Dropout: a simple way to
prevent neural networks from overfitting. Journal of Machine Learning Research 15, 1929-1958 (2014).
[16] Kingma, D. & Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014).
[17] Paszke, A., Gross, S. & Chintala, S. (2017).
[18] Kazi, A. et al. in International Workshop on Machine Learning in Medical Imaging. 70-78 (Springer).
6
| 1cs.CV
|
CLASSIFICATION OF PLETHORIES IN
CHARACTERISTIC ZERO
arXiv:1701.01314v1 [math.AC] 5 Jan 2017
MAGNUS CARLSON
Abstract. We classify plethories over fields of characteristic zero,
thus answering a question of Borger-Wieland and Bergman. All
plethories over characteristic zero fields are linear, in the sense that
they are free plethories on a bialgebra. For the proof we need some
facts from the theory of ring schemes where we extend previously
known results. We also classify plethories with trivial Verschiebung
over a perfect field of non-zero characteristic and indicate future
work.
Contents
1. Introduction
1
Notation and conventions
3
Acknowledgements
4
2. Ring schemes
4
3. Plethories and k − k-birings.
8
4. Classification of plethories over a field of characteristic zero. 11
5. Some classification results in characteristic p > 0.
13
References
17
1. Introduction
Plethories, first introduced by Tall-Wraith [13], and then studied by
Borger-Wieland [3], are precisely the objects which act on k-algebras,
for k a commutative ring. There are many fundamental questions regarding plethories which remain unanswered. One such question is,
given a ring k, whether one can classify plethories over k, in this paper
we will take a first step towards a classification.
For some motivation, let us start by looking at the category of modules Modk over a commutative ring k. If we consider the category of
representable functors Modk → Modk , there is a monoidal structure
given by composition of functors. Then one defines a k-algebra R as
a k-module R together with a comonad structure on the representable
endofunctor Modk (R, −) ∶ Modk → Modk with respect to composition
Date: January 6, 2017.
1
2
MAGNUS CARLSON
of functors. Heuristically, this says that a k-algebra is precisely the
kind of object which knows how to act on k-modules. This can be extended to a non-linear setting, so that instead of looking at k-modules
we look at k-algebras Algk and consider representable endofunctors
Algk → Algk . A comonoid with respect to composition of functors is
then called a plethory and analogously, a plethory is what knows how
to act on k-algebras. One particular important example of a plethory
is the Z-algebra Λ which consist of the ring of symmetric functions in
infinitely many variables with a certain biring structure. The functor
Algk (Λ, −) ∶ Algk → Algk represents the functor taking a ring R to its
ring of Witt vectors. Using plethories one gets a very conceptual view
of Witt vectors and in [2] James Borger develops the geometry of Witt
vectors using the plethystic perspective.
Let now k be a field. If we let P k denote the category of plethories over
k, there is a forgetful functor
F ∶ P k → Bialgk
into the category of cocommutative counital bialgebras over k. This
functor has a left adjoint S(−) ∶ Bialgk → P and we say that a plethory
P is linear if P ≅ S(Q) for some cocommutative, counital bialgebra Q.
Heuristically, a plethory P is linear if every action of P on an algebra
A comes from an action of a bialgebra on A. The main theorem of this
paper is:
Theorem 1.1. Let k be a field of characteristic zero. Then any kplethory is linear.
This answers a question of Bergman-Hausknecht [1, p.336] and BorgerWieland [3] in the positive. The theorem is proved by studying the
category of affine ring schemes. We have the following results, extending those of Greenberg [8] to arbitrary fields and not necesarily reduced
schemes:
Theorem 1.2. Let k be a field. Then any connected ring scheme of
finite type is unipotent.
Theorem 1.3. Let P be a connected ring scheme of finite type over
k. Then P is affine.
For the case of characteristic p > 0 our classification results on plethories
are not as complete and further work is needed to have a complete
classification. To explain our classification results here we need some
definitions. Let Fk be the Frobenius homomorphism of k and k⟨F ⟩
be the non-commutative ring which as underlying set is k[F ] and has
multiplication given by
F i F j = F i+j
CLASSIFICATION OF PLETHORIES IN CHARACTERISTIC ZERO
3
and
F a = Fk (a)F.
We define
to be the category of cocommutative, counital bialgebras over k which also are modules over k⟨F ⟩. Once again, for a
plethory over a perfect field k of char k > 0 there is a forgetful functor
Bialgkp
P k → Bialgkp
which has a left adjoint S[p] . Call a plethory P p-linear if P ≅ S[p] (Q)
for some Q ∈ Bialgkp . We have then the following classification result:
Theorem 1.4. Let k be a perfect field of characteristic p > 0. Assume
that P is a plethory over k such that the Verschiebung VP = 0. Then
P is p-linear.
The structure of this paper is as follows. In section 2 we study ring
schemes and prove some results which we will need for our classification theorem. The main theorems of this section that are needed for
later purposes are Theorem 2.6 and Theorem 2.7. In section 3 we introduce plethories and k − k-birings and provide some examples. This
section contains no new results and gives just a brief introduction to
the relevant objects as defined in Borger-Wieland [3]. In section 4 we
prove that all plethories over a field k of characteristic zero is linear
using the results from section 2. We also show that any k − k-biring
is connected. In section 5 we prove some initial classification results
regarding plethories in characteristic p > 0.
Notation and conventions
Ring
BRk,k
Pk
Bialgk
Bialgkp
⊙
Algk
+
∆A , ∆×A
ǫ+A , ǫ×A
βA
+
∆2 , ∆×2
category of commutative and unital rings.
category of k − k-birings.
category of k-plethories.
category of cocommutative k-bialgebras.
category of cocommutative k-p-bialgebras.
composition product of k − k-birings, Def. 3.2.
category of commutative algebras over the ring k.
coaddition resp. comultiplication map for a biring A.
counit for coaddition resp. comultiplication for a biring A.
co-k-algebra strucutre on a k − k-biring A.
abbreviation for the composite (1 ⊗ ∆+ ) ○ ∆+ resp. (1 ⊗ ∆× ) ○
∆× .
P primitive elements functor
OX structure sheaf of a scheme X.
Schk category of k-schemes for k a commutative ring.
Ga the affine line viewed as a group scheme, see Ex. 3.1
Gm the multiplicative group scheme, after Def. 2.5.
4
MAGNUS CARLSON
µp the p-th root of unity group scheme, Ex. 5.2
αp see Ex. 3.3
π0 (G) the group scheme of connected components of a group scheme
G over the field k, Def. 4.2
S free plethory functor on a cocommutative bialgebra. Def. 4.1
[p]
S
free plethory functor on a cocommutative p-bialgebra, after
Def. 5.1.
○
G the identity component of a group scheme G.
FG , VG the Frobenius resp. Verschiebung morphism of a group scheme
G over a perfect field of characteristic p > 0.
k⟨F ⟩ the twisted polynomial algebra.
For us, all rings are commutative and unital. We will use Sweedler
(1)
(2)
notation for coaddition ∆+ and ∆× , so that ∆+ (x) = ∑i xi ⊗ xi and
[1]
[2]
∆× (x) = ∑i xi ⊗ xi if x ∈ A where A is a biring. For concepts from
the theory of group schemes not introduced properly here, we refer to
[11] or [6] .
Acknowledgements. I am very grateful to James Borger who supplied me with the conjecture for characteristic zero and the idea of
”weakly linear” plethories. He has been more than generous with his
knowledge and many of the ideas in this paper come from conversations
with him. I also thank David Rydh and Lars Hesselholt for the many
useful comments they gave which helped improve the quality of this
article. I would also like to thank my advisor Tilman Bauer for his
support and for his many thoughtful suggestions on this article.
2. Ring schemes
Let k be a commutative ring. Recall that R is a ring scheme over k if
R is a scheme together with a lift of the functor
Schk (−, R) ∶ Schk → Set
to Ring. We say that a ring scheme is a k-algebra scheme if the lift
factors through the category of k-algebras. We will mostly be concerned
with affine ring schemes. Ring schemes were studied by Greenberg in [8]
and he showed that for connected, reduced ring schemes of finite type
over an algebraically closed field k, the underlying scheme is always
affine. Further, he shows that the underlying group variety is always
unipotent. We improve on these results by showing that any connected
ring schemes of finite type over an arbitrary field is affine, and that the
underlying group scheme is always unipotent. From now on, in this
section, k is always a field.
Definition 2.1. A k-scheme X is anti-affine if OX (X) = k. We say
that a group scheme is anti-affine if its underlying scheme is anti-affine
CLASSIFICATION OF PLETHORIES IN CHARACTERISTIC ZERO
5
For example, abelian varieties are all anti-affine group schemes. An
anti-affine group scheme has the property that any morphism from
it into an affine group scheme is trivial. Anti-affine groups are very
important for the structure of group schemes as the following theorem
shows:
Theorem 2.2 ([4, Theorem 1]). If G is a group scheme of finite type
over a field k there is an exact sequence of group schemes
0 → Gant → G → G/Gant → 0
such that Gant is anti-affine and G/Gant is affine.
We will now want to show that all connected finite type ring schemes
are affine, i.e that in the above exact sequence Gant = Spec k. For this,
we will need the following lemma.
Lemma 2.1. Let X, Y, Z be k-schemes with X quasi-compact and antiaffine and Y locally noetherian and irreducible.Suppose that f ∶ X ×Y →
Z is a morphism such that there exist k-rational points x0 ∈ X, y0 ∈ Y
such that f (x, y0 ) = f (x0 , y0 ) for all x. Then f (x, y) = f (x0 , y) for all
x, y.
Proof. see [4, Lemma 3.3.3.].
Theorem 2.3. Let R be a connected ring scheme of finite type over
k. Then R is affine.
Proof. We know by Theorem 2.2 that R sits in the middle of an extension of an affine group scheme by an anti-affine group. Let
0 → Rant → R → Raf f → 0
be the corresponding extension where Rant is anti-affine and Raf f the
affine quotient. Since R is of finite type, Raf f is a ring scheme. This
follows from the fact that if X, Y are quasi-compact k-schemes, then
the obvious map OX (X)⊗k OY (Y ) → OX×k Y (X ×k Y ) is an isomorphism
(see [4, Lemma 2.3.3.]). This implies that Rant defines an ideal scheme
in R, i.e for all rings S over k, Rant (S) is an ideal of R(S). Now, we will
apply the above lemma with Y = R (note that R is irreducible) and
X = Z = Rant . Taking x0 = eRant and y0 = eR to be the rational points
corresponding to the additive identity of Rant (k) and R(k) respectively,
we have that m(x, y0 ) = m(x0 , y0 ) is identically equal to zero. Thus, we
have that m(x, y) = m(x0 , y) is identically zero. But, letting 1R be the
rational point corresponding to the multiplicative identity of R(k) we
have that m(x, 1R ) is zero. But multiplication by 1 is always injective,
and thus, Rant is trivial and R is affine.
We don’t know if the condition for R to be of finite type is necessary in
2.3. Let us recall the following definition from the theory of algebraic
groups.
6
MAGNUS CARLSON
Definition 2.4. Let G be a commutative group scheme over k. We say
that G is unipotent if it is affine and if every non-zero closed subgroup
H of G admits a non-zero homomorphism H → Ga .
The data of a homomorphism G → Ga is the same as specifying an
element x ∈ AG in the underlying Hopf algebra of G that satisfies
∆G (x) = x⊗1 +1 ⊗x, i.e specifying a primitive element. If G = Spec AG
is an affine group scheme and AG the Hopf algebra associated to G,
then saying that G is unipotent is the same as saying that it is coconnected (or conilpotent). The following definition will be useful for the
proof of Theorem 2.6.
Definition 2.5. Let G be a commutative affine group scheme over a
field. We say that G is multiplicative if every homomorphism G → Ga
is zero.
An example of a multiplicative group is Gm = Spec k[x, x−1 ]. There
can in general be no homomorphism from a multiplicative group into
a unipotent group and no morphisms from a unipotent group to a
multiplicative group (for a proof, see [11, Corollary 15.19-15.20]).
The following theorem was shown for reduced ring varieties over an
algebraically closed fields by Greenberg, but the results carry over for
perfect fields without any modification. We improve on this by carrying
through the proof when R is not necesarily reduced and over any field
k. Further,the theorem can be extended to ring schemes not necesarily
of finite type if the ring scheme is already known to be affine.
Theorem 2.6. Over a field k, all connected ring schemes R of finite
type are unipotent.
Proof. By Theorem 2.3, a connected ring scheme is affine. We know
that R contains a greatest multiplicative subgroup Rm that has the
property that for all endomorphisms α of RS , (where RS is the base
change of R to S) for S a k-algebra, that α((Rm )S ) ⊂ (Rm )S ([11,
Theorem 17.16]). Thus, since any x ∈ R(k) defines an endomorphism
of R (as a group scheme) through multiplication by x, we have that Rm
is an ideal of R. It is known that any action of a connected algebraic
group on a multiplicative group must be trivial, i.e for G connected and
H multiplicative, a map G → Aut(H, H) must have image the identity.
We will need the following, which says that any map G → End(H, H)
where G is any connected group scheme and H is multiplicative is
trivial. This is basically just deduced, mutatis mutandis, from the
proof of [11, Theorem 14.28]. So, we see that 0 and 1 defines the same
endomorphisms on the ideal scheme Rm . But this is only the case if
Rm = 0.
To extend this to all connected ring schemes, we need the following:
Theorem 2.7. Let k be a field and R be an affine ring scheme over k.
Then R is a filtered limit of ring schemes of finite type.
CLASSIFICATION OF PLETHORIES IN CHARACTERISTIC ZERO
7
Proof. The following proof is inspired by the analogue theorem for Hopf
algebras over a field, as occurs in for example Milne [11, Prop. 11.32].
Write R = Spec AR . We know that AR is a bialgebra and we see that we
can reduce to proving that any a ∈ AR is contained in a sub-bialgebra of
finite type. Let ∆+ ∶ AR → AR ⊗ AR be the coaddition giving the additive group structure on R and ∆× ∶ AR → AR ⊗AR the comultiplication
defining the multiplication on R. Consider
∆+2 (a) = ∑ ci ⊗ xij ⊗ dj
i,j
with ci and dj linearly independent. Now, by the fundamental theorem
of coalgebras, we know that if we take X to be the subspace of AR
generated by {xij }, then this is a subcoalgebra, i.e that ∆+ (xij ) ⊂
X ⊗ X. Now, for each xij in this system, consider
∆×2 (xij ) = ∑ ei ⊗ ykl ⊗ fl
k,l
with ei and fl linearly independent. With the same arguments, one sees
that for the subspace Y generated by {ykl } we have ∆× (ykl ) ⊂ Y ⊗ Y.
Let now Z be subalgebra generated by the finite-dimensional subspace
spanned by {xij , ykl }. We claim that Z actually is closed under both
the operation ∆+ and ∆× . It is clear that
∆× (xij ) ⊂ Z ⊗ Z
and the same holds for coaddition. It is also easy to verify that
∆× (ykl ) ⊂ Z ⊗ Z. We will now prove that ∆+ (ykl ) ⊂ Z ⊗ Z and for
this, consider the following diagram which is easily verified if we reverse all arrows and think of it in terms of rings.
∆×
AR
AR ⊗ AR
∆+
AR ⊗ AR
AR ⊗ AR
∆× ⊗∆×
AR ⊗ AR ⊗ AR ⊗ AR
AR ⊗ AR
1⊗T ⊗1
AR ⊗ AR ⊗ AR ⊗ AR
AR ⊗ AR
∆× ⊗1⊗1⊗∆×
AR ⊗ (AR ⊗ AR ) ⊗ (AR ⊗ AR ) ⊗ AR
1⊗T ⊗T ⊗1
(AR ⊗ AR ) ⊗ AR ⊗ AR ⊗ (AR ⊗ AR )
M ⊗1⊗1⊗M
AR ⊗ AR ⊗ AR ⊗ AR
AR ⊗ AR
∆× ⊗1
(AR ⊗ AR ) ⊗ AR
1⊗∆+ ⊗1
AR ⊗ AR ⊗ AR ⊗ A
8
MAGNUS CARLSON
Here M is the multiplication map and T swithces the factors. What
the diagram is saying, is just relating different ways of forming
abd + acd
for a, b, c, d in a ring. So this says, that
(1 ⊗ ∆+ ⊗ 1)(∆×2 (xij ) = ∑ ek ⊗ ∆+ (ykl ) ⊗ fl ∈ Z ⊗ Z ⊗ Z ⊗ Z.
k,l
Now, since ek are independent, this means that
∑ ∆+ (ykl ) ⊗ fl ∈ Z ⊗ Z ⊗ Z
l
and by linear independence of each fl this means that
∆+ (ykl ) ∈ Z ⊗ Z.
Now, let W be the sub-algebra generated by Z ∪ S(Z) where S ∶ AR →
AR is the antipode. It is easily verified that
∆+ ○ S = (S ⊗ S) ○ ∆+
and that
∆× (S(Z)) ⊂ W
follows from the identity
∆× ○ S = (1 ⊗ S) ○ ∆× .
We thus see that W is a bialgebra and we are done.
Corollary 2.2. Any affine connected ring scheme over a field is unipotent.
Proof. Indeed, we know that we can write P = lim Pi where Pi ranges
←Ð
over ring schemes of finite type. Now, unipotence is stable under inverse
limits and this immediately gives that P is unipotent.
3. Plethories and k − k-birings.
Let k be an arbitrary commutative ring. In this section we will recall
the definition of a plethory as defined in [3].
Definition 3.1. A k-biring A is a coring object in the category of
k-algebras. Explicitly, A is a k-algebra together with maps
∆+ ∶ A → A ⊗k A,
∆× ∶ A → A ⊗k A,
S ∶ A → A,
ǫ+ ∶ A → k
and ǫ× ∶ A → k such that:
CLASSIFICATION OF PLETHORIES IN CHARACTERISTIC ZERO
9
● The triple (∆+ , ǫ+ , S) defines a cocommutative Hopf algebra
structure on A with S the antipode and ǫ+ the counit.
● ∆× is cocommutative coassociative and codistributes over ∆+
and ǫ× ∶ A → k is a counit for ∆× .
We say that A is a k − k-biring if, in addition to the above data, it has
a map
β ∶ k → Ringk (A, k)
of rings, where we endow Ringk (A, k) with the ring structure induced
from the coring structure on A.
Equivalently, a k − k-biring A is just an affine scheme together with a
lift of the functor Ringk (A, −) to the category of k-algebras, i.e it is an
affine k-algebra scheme.
Example 3.1. Let define the k-algebra scheme which we will call Ga .
Ga will represent the identity functor Ringk → Ringk . The underlying
scheme of Ga is Spec k[e]. The coaddition and comultiplication is given
by ∆+ (e) = e ⊗ 1 + e ⊗ 1, ∆× (e) = e ⊗ e, the additive resp. multiplicative
counit by ǫ+ (e) = 0, ǫ× (e) = 1 the antipode by S(e) = −e and the co-klinear structure by β(c)(e) = c for all c ∈ k.
Example 3.2. Consider Z[e, x]. On e, we define all the structure maps
as in the previous example. We then define
∆+ (x) = x ⊗ 1 + 1 ⊗ x,
∆× (x) = x ⊗ e + e ⊗ x
and ǫ× (x) = ǫ+ (x) = 0, S(x) = −x. This Z-ring scheme represents the
functor taking a ring R to R[ǫ]/(ǫ2 ) , the ring of dual numbers over
that ring.
Example 3.3. Let k = Fq be a finite field of characteristic p and consider
αp = Spec k[e]/(ep )
as a group scheme where the group structure is induced from Spec k[e].
Define a multiplication
αp × αp → αp
by saying that xy = 0 for any x, y ∈ αp (R) for R a k-algebra. Consider
now the constant group scheme
Fq = ∐ k.
x∈k
Then we can define a structure of a ring scheme on Fq × αp by defining
the multiplication to be (x, y)(z, w) = (xz, xw + yz) for (x, y), (z, w) ∈
(Fq × αp )(R). This is a non-reduced ring scheme.
10
MAGNUS CARLSON
A famous example is also that the functor taking a ring R to W (R),
its ring of big Witt vectors, is also representable by a ring scheme.
Let us note that we can form the category of k − k-birings, with morphisms between objects those morphism of k-algebras respecting the
biring structure. We let BRk,k be the category of k − k-birings. Let us
recall the following definition from [3].
Definition 3.2. Let A be a k − k-biring. Then the functor
Ringk (A, −) ∶ Algk → Algk
has a left adjoint,
A ⊙k − ∶ Algk → Algk .
Explicitly, for a k-algebra R, A ⊙ R is the k-algebra generated by all
symbols a ⊙ r subject to the conditions that:
●
●
●
●
●
●
∀a, a′ ∈ A, r ∈ R, aa′ ⊙ r = (a ⊙ r)(a′ ⊙ r).
∀a, a′ ∈ A, r ∈ R, (a + a′ ) ⊙ r = (a ⊙ r) + (a′ ⊙ r).
∀c ∈ k, c ⊙ r = c.
(1)
(2)
∀a ∈ A, r, r ′ ∈ R, a ⊙ (r + r ′ ) = ∑i (ai ⊙ r)(ai ⊙ r ′ ).
[1]
[2]
∀a ∈ A, r, r ′ ∈ R, a ⊙ rr ′ = ∑i (ai ⊙ r)(ai ⊙ r ′ ).
∀a ∈ A, c ∈ k, a ⊙ c = β(c)(a).
It is easy to see that (⊗i Ai ) ⊙ R ≅ ⊗i (Ai ⊙ R) and that A ⊙ (⊗i Ri ) ≅
⊗i (A⊙Ri ). If further, R is a k−k-bialgebra, we note that A⊙R is a k−kbialgebra. Indeed, we have that Ringk (A⊙R, S) ≅ Ringk (R, Ringk (A, S))
and since the latter set has a ring structure, so does the former. One
then verifies that ⊙k gives a monoidal structure to BRk,k . The unit
of this monoidal structure is k[e]. BRk,k is a monoidal category, but
it is not symmetric. Now, the Yoneda embedding sets up an equivalence of categories between the category of representable endofunctors
Algk → Algk and BRk,k and under this equivalence, ⊙ corresponds to
○, composition of representable endofunctors as given in the introduction. Denote the category of representable endofunctors Algk k → Algk
by Algkend .
Definition 3.3. A k-plethory is a comonoid in Algkend where the monoidal
structure is composition of endofunctors. Explicitly, on the level of representing objects, a k-plethory P is a monoid in BRk,k . This means that
P is a biring together with an associative map of birings P ⊙ P → P
and a unit k[e] → P.
Remark 3.4. For a plethory P one can define an action of P on a
k-ring R to be a map ○ ∶ P ⊙ R → R such that (p1 ⊙ p2 ) ○ r = p1 ⊙ (p2 ○ r)
and e ○ r = r, ∀p1 , p2 ∈ P, r ∈ R. A ring R together with an action of P
on R is called a P -ring.
Example 3.5. If k is a finite ring, then k k , the set of functions k → k
is a plethory where ○ is given by composition of functions.
CLASSIFICATION OF PLETHORIES IN CHARACTERISTIC ZERO
11
4. Classification of plethories over a field of
characteristic zero.
In this section we will prove that all plethories over a field of characteristic zero are linear. This question was asked by Bergman-Hausknecht
[1] and Borger-Wieland [3]. To understand what it means for a plethory
to be linear, we will introduce some terminology.
Definition 4.1. Let A be a cocommutative bialgebra (not necesarily
commutative) over k with comultiplication ∆. Then there is a free kplethory on A over k. The underlying algebra structure is S(A), the
symmetric algebra on A, and the coaddition
∆+ ∶ S(A) → S(A) ⊗ S(A)
is induced from the map
A → S(A) ⊗ S(A)
sending a to a ⊗ 1 + 1 ⊗ a. The comultiplication ∆× is similarily induced
from ∆. The plethysm
○ ∶ S(A) ⊙ S(A) → S(A)
is given by
S(m)
S(A) ⊙ S(A) ≅ S(A ⊗ A) ÐÐ→ S(A)
where m is the multiplication on A. Among the pairs consisting of a
plethory P and a morphism of bialgebras f ∶ A → P the pair S(A) and
j ∶ A → S(A) is initial with this property.
Call a plethory P linear if P ≅ S(A) for some bialgebra A. The reason
for calling it linear is that if P ≅ S(A) for some bialgebra A then
Ringk (S(A), −) = Modk (A, −).
Let us note now that by Theorem 2.6 any connected ring scheme of
finite type is unipotent. Over Q (or more generally any field of characteristic zero) all group schemes are reduced by a theorem of Cartier.
We say that a group scheme G is étale if G is a finite scheme and geometrically reduced. This is equivalent to asking for the underlying
Hopf algebra AG to be an étale algebra. Let us recall the following
definition from the theory of group schemes (see for example [6, II, §5,
Proposition 1.8] or [11, Definition 9.4])
Definition 4.2. Let G be a group scheme of finite type over k. Let
AG be the underlying Hopf algebra of G and consider the largest étale
k-subalgebra π0 (AG ) of AG . π0 (AG ) then has a Hopf algebra structure
induced from the one on AG and we let π0 (G) = Spec π0 (AG ) be the
group scheme associated to this Hopf algebra.
Note that there is a canonical map G → π0 (G). It is easy to see that
if π0 (G) = Spec k, then G is geometrically connected since in that case
AG has no nontrivial idempotents.
12
MAGNUS CARLSON
Lemma 4.1. Any k-algebra scheme R of finite type over any infinite
field k is geometrically connected.
Proof. Consider the connected-étale exact sequence
0 → R○ → R → π0 (R) → 0
of group schemes where R○ is the identity component of R. We will first
show that π0 (R) has a natural k-algebra scheme structure. Indeed, for
this it is enough to show that R○ is a k-ideal scheme in R. Let us start
by proving that m(R○ × R) ⊂ R○ . We know that the multiplication
m ∶ R○ × R → R
takes the additive identity e ∈ R(k) to itself, i.e m(e, x) = e for any
x ∈ R(k). Further, the k-algebra structure on R○ is induced from the kalgebra structure on R. This clearly implies that R○ is a k-ideal scheme.
Thus, the quotient R/R○ ≅ π0 (R) is a k-algebra scheme. Let us see that
π0 (R) is isomorphic to Spec k. One knows that the underlying algebra
of π0 (R) is a product of finite separable k-extensions. We consider
Schk (π0 (R), π0 (R)), this is a k-algebra (since π0 (R) is a ring scheme).
Because the underlying algebra of π0 (R) is a finite product of finite
separable field extensions, Schk (π0 (R), π0 (R)) is a finite set. However,
for a finite set to have a k-algebra structure it must just contain one
element, i.e it has to be the zero ring. This implies that π0 (R) = Spec k
so R is geometrically connected.
Now, let us consider a Hopf algebra H and denote the primitive elements of H by P(H). We say that a Hopf algebra is primitively generated if P(H) generates H as an algebra. Over characteristic zero
all unipotent affine group schemes are primitively generated. We then
have the classical Milnor-Moore theorem [12].
Theorem 4.3. For any commutative connected affine unipotent group
scheme H over a field of characteristic zero, the canonical map
Spec H → Spec S(P(H))
is an isomorphism of group schemes. In particular, the underlying
scheme is affine space.
Remark 4.2. Let us note that we can view P(H) as a Lie algebra
with trivial commutator. Then the construction S(P(H)) is the same
as the universal enveloping Lie algebra of P(H).
In [3] it is shown that if Q is a plethory over a field k, then P(Q). the
+
primitives with respect to δQ
, is a cocommutative k-bialgebra. Briefly,
the multiplication in P(Q) is given by the plethysm ○ and the maps
∆× ∶ Q → Q ⊗ Q,
ǫ× ∶ Q → k induces a comultiplication respectively a counit on P(Q)
making it a cocommutative counital bialgebra.
CLASSIFICATION OF PLETHORIES IN CHARACTERISTIC ZERO
13
Theorem 4.4. Let Q be a plethory over a field of characteristic zero
k. Then Q is linear, i.e
Q ≅ S(P(Q))
where S(P(Q)) has the plethory structure as given in Definition 3.1.
Proof. Suppose that Q is a plethory over k. P(Q) naturally has a
bialgebra structure as explained above. Given this, we can form the
free plethory on P(Q), S(P(Q)). We always have a natural map
v ∶ S(P(Q)) → Q
of Hopf algebras, and this is bijective by Milnor-Moore. Thus to show
that any plethory is linear, it suffices to show that this is actually a
morphism of plethories. But this is clear: the pair S(P(Q)) and
j ∶ P(Q) → S(P(Q))
is initial in the category of pairs consisting of a plethory P and a morphism f ∶ P(Q) → P of bialgebras. It is immediate that the canonical
map v is induced by this universal property, when we note that there
clearly is a map P(Q) → Q of bialgebras. We will of course need to
show that v is an isomorphism in the category of plethories. This follows easily from the fact that v is an isomorphism of affine schemes and
thus has an inverse in the category of affine schemes. What remains to
be checked is that this inverse is a morphism of plethories, but this is
immediate since v is.
Corollary 4.3. Let k be a field of characteristic zero. Then the category of plethories over k is equivalent to the category of cocommutative
k-bialgebras.
Proof. By the above theorem, the counit map is an isomorphism and
it is immediate to see that the unit map is an isomorphism as well.
Remark 4.4. If k is a field of characteristic zero and Q a plethory over
k, this shows that the category of Q-rings is equivalent to the category
of rings with an action of the bialgebra P (Q).
5. Some classification results in characteristic p > 0.
In this section we will start a classification for plethories over a perfect field k of characteristic p. Our classification results here only apply to a certain class of plethories. We state future research directions, as well as give some ”pathological” examples which a complete
classification must take into account. For any scheme X over k with
structure map f ∶ X → Spec k we let Gp be the pullback of f along
F ∶ Spec k → Spec k, the Frobenius.
14
MAGNUS CARLSON
Let us briefly recall that for perfect fields k, group schemes over k
have two especially important maps, the (relative) Frobenius
FG ∶ G → Gp ≅ G
and the Verschiebung
VG ∶ G ≅ Gp → G.
These satisfy the property that FG VG = VG FG = p. A ring scheme R is
called elementary unipotent if VR = 0, i.e the Verschiebung is zero. Call
a plethory Q weakly linear if there is a map of plethories f ∶ P → Q
where P is a linear plethory (as defined in the previous section) such
that f when viewed as a map of algebras is surjective. This will, in
particular, imply that Q is primitively generated and is a quotient of
P by a P − P -ideal as defined in [3]. Not all plethories over a perfect
field k are primitively generated, as the following example shows (built
on an example from [10], Remark 1.6.2).
Example 5.1. Let G be the group scheme
Ga ×f αp
which as a scheme, is just Ga × αp . We let the the group structure be
given by
(g1 , h1 )(g2 , h2 ) = (g1 g2 , h1 + h2 + f (g1 , g2 ))
for
g1 , g2 , h1 , h2 ∈ Ga (R) × αp (R)
where f (x, y) = ((x+y)p −xp −y p )/p. This is a p-torsion group scheme but
is not elementary unipotent. One can define a non-unital ring scheme
structure on G be definining the multiplication to be trivial and then,
when k is finte, i.e k ≅ Fq ”unitalize” this by taking the direct product
with
Fq = ∐ Fq
a∈Fq
to get a ring scheme, as we did in Example 3.3. The underlying group
scheme of this ring scheme is clearly not elementary unipotent, since the
Verschiebung acts on each factor separately. Taking the free plethory
on a biring (see [3] 2.1 ) will then give us a plethory with its underlying
group scheme not elementary unipotent.
Another feature which differs from the case over a field of characteristic
zero is that there are plethories which have a non-trivial multiplicative
subgroups. This stems from the fact that there are ring schemes with
non-trivial multiplicative subgroups.
Example 5.2. Consider µp = Spec k[x, x−1 ]/(x − 1)p with comultiplication ∆ ∶ x → x ⊗ x and counit ǫ(x) = 1. This is an example of a multiplicative group scheme which is p-torsion and we can as before define
a trivial multiplication on µp , making it a non-unital ring scheme. We
CLASSIFICATION OF PLETHORIES IN CHARACTERISTIC ZERO
15
can then as previously stated, for finite fields, unitalize it to get a ring
scheme by taking the direct product with
Fq
and after that we can form the free plethory to get a plethory Q with
a non-trivial multiplicative subgroup.The fact that it has a non-trivial
multiplicative subgroup comes from , for example, the fact that there
is a non-zero homomorphism of group schemes µp → Q.
These two examples are rather artificial, but they show that plethories
behave wildly different in characteristic p > 0 than in characteristic 0.
We know that for any group scheme G over a perfect field k of characteristic p > 0, the group P(G) of primitive elements has a natural action
of the Frobenius, taking x ∈ P(G) to xp . In fact, P(G) becomes a module over a certain ring. As we previously stated, P(G) = Hom(G, Ga ).
We thus have that P(G) is naturally a module over the endomorphism
ring End(Ga , Ga ).
Definition 5.1. Let k⟨F ⟩ be the non-commutative polynomial ring
over k in one variable F with multplication given by, for a ∈ k aF =
Fk (a)a where Fk is the Frobenius endomorphism of k.
It is a quick calculation to show that End(Ga , Ga ) ≅ k⟨F ⟩. We now see
that P(G) is a module over k⟨F ⟩. Let us denote the category of modules
over k⟨F ⟩ by Modk⟨F ⟩ . Given a k⟨F ⟩-module M one can construct an
elementary unipotent group scheme S[p] (M) as follows (for details we
refer the reader to [11]) . Form S(M), the symmetric algebra on M,
with its obvious Hopf algebra structure and consider the map j ∶ M →
S(M). We then quotient out by the ideal generated by the elements
j(F x) − j(x)p
to get S[p] (M). One notes that for any commutative algebraic group G
one always has a map G → S[p] (P(G)). We have the following classical
theorem (see [6, IV,§3, Proposition 6.6])
Theorem 5.2. Let G be an affine group scheme. The following are
equivalent:
(i) The Verschiebung VG is zero.
(ii) G is a closed subgroup of Gra for some r.
(iii) The canonical homomorphism G → S[p] (P(G)) is an isomorphism.
Remark 5.3. What we call S[p] (P(Q)) is the same as the enveloping
p-algebra (also called the restricted universal enveloping algebra) on
the p-Lie algebra P(Q) where P(Q) has trivial commutator.
Lemma 5.4. When Q is a plethory, then S[p] (P(Q)) has the structure
of a plethory.
16
MAGNUS CARLSON
Proof. We know that P(Q) has a k⟨F ⟩ module structure where the
action of F is just taking the pth power. Further, S[p] (P(Q)) is the
quotient of S(P(Q)), which we know is a plethory, by the ideal J
generated by j(x)p − j(xp ), where j ∶ P(Q) → S(P(Q)) is the inclusion
in degree 1. It now suffices to show that this is a Q−Q-ideal (see [3] 6.1)
for U [p] (P(Q)) to be a plethory. This is equivalent to showing that for
a generating set S of J that
∆+Q (S) ⊂ Q ⊗ J + J ⊗ Q,
∆×Q (X) ⊂ Q ⊗ J + J ⊗ Q,
and
βQ (c)(S) = 0
∀c ∈ k and that
P(Q) ⊙ X ⊙ Q ⊂ J.
The first is immediate, since taking S to be the set of all j(x)p − j(xp ),
we have
∆+ (j(x)p ) − ∆+ (j(xp )) = ∆+ (j(x))p − (j(xp ) ⊗ 1 + 1 ⊗ j(xp ))
which is equal to
j(x)p ⊗ 1 + 1 ⊗ j(x)p − j(xp ) ⊗ 1 − 1 ⊗ j(xp ) ⊂ J ⊗ Q + Q ⊗ J.
Further,
[1]
[2]
[1]
[2]
∆× (j(x)p )−∆× (j(xp )) = ∑ j(xi )p ⊗j(xi )p − ∑ j((xi )p )⊗j((xi )p )
i
i
and this is equal to
[1]
[2]
[2]
[1]
[1]
[2]
∑ j(xi )p ⊗(j(xi )p −j((xi )p ))+∑((j((xi )p )−j(xi )p )⊗j((xi )p ))
i
but this is in J ⊗ P + P ⊗ J. We also need to show that βQ (c)(S) = 0,
this is clear. The last containment is similarily easy to verify.
Theorem 5.3. When Q is a plethory over a perfect field such that
VQ = 0, then S[p] (P(Q)) ≅ Q. We then say that Q is a p-linear plethory.
Proof. All one has to verify is that the canonical map f ∶ Q → S[p] (P(Q))
is a map of plethories. But this is obvious since this map is just the
composition of the two plethory maps Q → S(P(Q)) and S(P(Q)) →
S[p] (P(Q)).
Remark 5.5. We have seen that plethories need not be elementary
unipotent and not purely unipotent either (i.e it can have a non-trivial
multiplicative subgroup) Let us note that there can be no non-trivial
finite plethories over an infinite perfect field k. Indeed, from what we
have seen all plethories Q are connected over an infinite field. By
classical Dieudonné theory we can then decompose Q as Q = Qloc,red ×
Qloc,loc . This would imply that the Frobenius is nilpotent, but this can
never happen: the Frobenius is always a map of ring schemes.
CLASSIFICATION OF PLETHORIES IN CHARACTERISTIC ZERO
17
It seems to us that to classify plethories over a perfect field one should
establish an extension of ordinary Dieudonné theory to account for ring
schemes, which has been done to some extent by Hedayatzadeh in [9]
and for Hopf rings by Goerss [7] and Buchstaber-Lazarev [5]. Note that
Hedayatzadeh work with finite / profinite group schemes and with local
group schemes, which limits their applications to ring schemes since we
have seen that there are no non-trivial finite connected ring schemes
over a perfect field.
References
[1] George M. Bergman and Adam O. Hausknecht. Co-groups and co-rings in categories of associative rings, volume 45 of Mathematical Surveys and Monographs.
American Mathematical Society, Providence, RI, 1996.
[2] James Borger. The basic geometry of Witt vectors, I: The affine case. Algebra
Number Theory, 5(2):231–285, 2011.
[3] James Borger and Ben Wieland. Plethystic algebra. Adv. Math., 194(2):246–
283, 2005.
[4] M. Brion. Some structure theorems for algebraic groups. 2015.
[5] Victor Buchstaber and Andrey Lazarev. Dieudonné modules and p-divisible
groups associated with Morava K-theory of Eilenberg-Mac Lane spaces. Algebr.
Geom. Topol., 7:529–564, 2007.
[6] Michel Demazure and Pierre Gabriel. Groupes algébriques. Tome I: Géométrie
algébrique, généralités, groupes commutatifs. Masson & Cie, Éditeur, Paris;
North-Holland Publishing Co., Amsterdam, 1970. Avec un appendice ıt Corps
de classes local par Michiel Hazewinkel.
[7] Paul G. Goerss. Hopf rings, Dieudonné modules, and E∗ Ω2 S 3 . In Homotopy
invariant algebraic structures (Baltimore, MD, 1998), volume 239 of Contemp.
Math., pages 115–174. Amer. Math. Soc., Providence, RI, 1999.
[8] Marvin J. Greenberg. Algebraic rings. Trans. Amer. Math. Soc., 111:472–481,
1964.
[9] Mohammad Hadi Hedayatzadeh. Exterior powers of Barsotti-Tate groups. PhD
thesis, ETH Zürich, 2010.
[10] Tatsuji Kambayashi, Masayoshi Miyanishi, and Mitsuhiro Takeuchi. Unipotent algebraic groups. Lecture Notes in Mathematics, Vol. 414. Springer-Verlag,
Berlin-New York, 1974.
[11] J. S Milne. Algebraic groups (algebraic group schemes over fields) v1.20. 2015.
[12] John W. Milnor and John C. Moore. On the structure of Hopf algebras. Ann.
of Math. (2), 81:211–264, 1965.
[13] D. O. Tall and G. C. Wraith. Representable functors and operations on rings.
Proc. London Math. Soc. (3), 20:619–643, 1970.
E-mail address: [email protected]
| 0math.AC
|
Canonical Representations for Circular-Arc Graphs
Using Flip Sets
Maurice Chandoo
Leibniz Universität Hannover, Theoretical Computer Science,
Appelstr. 4, 30167 Hannover, Germany
[email protected]
arXiv:1702.05763v2 [cs.DS] 20 Dec 2017
Abstract
We show that computing canonical representations for circular-arc (CA) graphs reduces to computing certain subsets of vertices called flip sets. For a broad class of CA graphs, which we call
uniform, it suffices to compute a CA representation to find such flip sets. As a consequence
canonical representations for uniform CA graphs can be obtained in polynomial-time. We then
investigate what kind of CA graphs pose a challenge to this approach. This leads us to introduce
the notion of restricted CA matrices and show that the canonical representation problem for CA
graphs is logspace-reducible to that of restricted CA matrices. As a byproduct, we obtain the
result that CA graphs without induced 4-cycles can be canonized in logspace.
1998 ACM Subject Classification G.2.2 Graph Theory
Keywords and phrases canonization, circular-arc graphs, isomorphism
Digital Object Identifier 10.4230/LIPIcs...
1
Introduction
We consider an arc to be a connected set of points on the unit circle including the endpoints.
A CA graph is a graph whose vertices can be assigned arcs such that two vertices are adjacent
iff their corresponding arcs intersect. More formally, given a graph G we call it a CA graph if
there exists a function ρ which maps every vertex u of G to an arc ρ(u) such that u and v are
adjacent iff their arcs ρ(u) and ρ(v) have non-empty intersection. We call such a mapping ρ
a CA representation of G. CA graphs are a form of geometrical intersection graphs. Let X
be a family of sets over some ground set. Any subset Y of X defines a graph GY which has
Y as its vertex set and two vertices are adjacent if they have non-empty intersection. The
graph GY is called intersection graph of Y . We say a (finite) graph G is an intersection graph
of X if it is isomorphic to the intersection graph of Y for some Y ⊆ X . In this language CA
graphs are intersection graphs of arcs. The intersection graphs of intervals on the real line
are called interval graphs. In this sense any set of geometrical objects defines a (geometrical
intersection) graph class. CA graphs are a generalization of interval graphs since every set of
intervals on the real line can be ‘bent’ into arcs while preserving the intersection relation.
Therefore every interval graph is a CA graph.
Being a generalization of interval graphs—the archetype of geometrical intersection
graphs—CA graphs are quite prominent as well and have been known for decades. Since
then structural properties and algorithmic problems for this class have been thoroughly
investigated with [5] and [16] being two of the earliest works in this regard. In particular,
finding characterizations of CA graphs and constructing a CA representation for a given
CA graph have received a great deal of attention. Remarkably, finding a forbidden induced
subgraph characterization of CA graphs is still an open problem. See [12] for a survey on this
line of research and [1] for one of the most recent results in that direction. It should also be
© Maurice Chandoo;
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
2
Canonical Representations for Circular-Arc Graphs Using Flip Sets
mentioned that CA graphs are of practical relevance with applications arising in disciplines
such as genetics and operations research. An explanation of the connection between genetics
and interval graphs in layman’s terms can be found in [17]. For a specialized account on this
connection emphasizing circularity see [15]. An example of how CA graphs can be used to
model the problem of phasing traffic lights is given in [6].
In this work we consider the canonical representation problem for CA graphs. The
representation problem for CA graphs is as follows. Given a CA graph G as input
we want to output a CA representation ρG of G. The canonical variant of this problem imposes the additional requirement that for every pair of isomorphic CA graphs G
and H their representations ρG and ρH should have identical underlying sets of arcs,
i.e. {ρG (v) | v ∈ V (G)} = {ρH (v) | v ∈ V (H)}. Notice that solving the representation problem for CA graphs implies solving the recognition problem for CA graphs, i.e. the question
given a graph G is it a CA graph. Likewise, solving the canonical representation problem for
CA graphs implies solving the isomorphism problem for CA graphs, i.e. deciding whether
two given CA graphs are isomorphic.
Consider the following generalization of interval graphs: 2-interval graphs are intersection
graphs of two intervals on the real line. It is easy to see that this class contains CA graphs
because given a set of arcs one can cut the circle at some point and straighten the arcs.
The arcs which are cut can be modeled as two intervals. It is interesting to note that the
isomorphism problem for interval graphs is logspace-complete [9] while the one for 2-interval
graphs is already GI-complete and CA graphs lie inbetween these two classes. The GIcompleteness for 2-interval graphs follows from the fact that every line graph is a 2-interval
graph and line graphs are already GI-complete. To see why this inclusion holds consider a
graph G and its line graph L(G). Assign every vertex v of G an interval Iv on the real line
such that no two intervals Iu and Iv intersect for every pair of distinct vertices u and v of
G. The 2-interval model for L(G) is obtained by mapping every edge {u, v} of G to the two
intervals Iu and Iv .
While a polynomial-time algorithm for deciding isomorphism of interval graphs is known
since 1976 due to Booth and Lueker this question still remains open for CA graphs. There
have been two claimed polynomial-time algorithms for deciding isomorphism of CA graphs
in [18] and [7] which were shown to be incorrect in [4] and [3] respectively. For interval
graphs even a linear-time algorithm for isomorphism is known [13]. A more recent result
is that canonical interval representations for interval graphs can be computed in logspace
and that this is optimal in the sense that recognition and deciding isomorphism for interval
graphs is logspace-complete [9]. These two hardness results also carry over to the class of
CA graphs. Furthermore, the isomorphism problem for proper CA graphs [11] and Helly
CA graphs [10] have been shown to be decidable in logspace. It is also shown how to obtain
canonical representations for these subclasses in logspace.
In this article we explain how the method used in [10] to obtain canonical representation
for Helly CA graphs can be adapted to CA graphs in general. Following this approach,
canonical representations for CA graphs can be found by computing certain subsets of vertices
called flip sets in an isomorphism-invariant manner. We introduce the class of uniform CA
graphs for which this method yields canonical representations in polynomial-time. We then
aim to isolate the instances of CA graphs which are difficult to handle with this method.
We try to capture these hard instances by what we call restricted CA matrices and show
that the canonical representation problem for CA graphs is logspace-reducible to that of
restricted CA matrices. During this isolation process we find a subset of uniform CA graphs,
namely ∆-uniform CA graphs, for which canonical representations can be computed in
M. Chandoo
logspace. The ∆-uniform CA graphs contain Helly CA graphs and every CA graph without
an induced 4-cycle. This generalizes the canonization result for Helly CA graphs given in
[10]. A preliminary version of this work appeared in [2].
The paper is organized as follows. In the third section we formalize the idea of computing
invariant flip sets in order to obtain canonical representations for CA graphs. This leads to
the definition of invariant flip set functions. In the fourth section we investigate for what
CA graphs a particular invariant flip set function is easy to compute. This leads to the
class of uniform CA graphs. We also provide an alternative characterization of uniform
CA graphs in terms of whether certain triangles in a CA graph have an unambiguous
representation. The main result of this section is that the representation problem for uniform
CA graphs, the canonical representation problem for uniform CA graphs and the non-Helly
triangle representability problem (introduced in section 4) for uniform CA graphs are all
logspace-equivalent. In the fifth section we consider the structure of non-uniform CA graphs,
introduce restricted CA matrices and show how the canonical representation problem for
CA graphs can be reduced to that of restricted CA matrices. In the process of proving this
reduction the class of ∆-uniform CA graphs is defined and it is shown that this class can be
canonized in logspace.
2
Preliminaries
For a number n ∈ N we write [n] for {1, . . . , n}. Given two sets A, B we say A and B intersect
if A ∩ B 6= ∅. We say A and B overlap, in symbols A G B, if A ∩ B, A \ B and B \ A are
non-empty.
We consider graphs without self-loops which sometimes have colored vertices and colored
edges. They can be seen as relational structures with the vertex set as universe and vertex
colors encoded as unary relations and edge colors as binary relations. The standard notion
of isomorphism for relational structures applies. We describe a graph with vertex colors as
tuple (G, c) where c is a function that maps the vertices of G to the colors. We talk about
a graph with edge colors as a square matrix whose entries represent the edge colors and
identify the indices of the matrix and the vertices of the graph. Consequently, we identify a
square matrix with the graph that it represents and talk about it in graph-theoretical terms.
By a class of (relational) structures we mean a set of such structures which is closed under
isomorphism.
We call a bijective function τ which maps the vertices of a graph G to some set V 0
a relabeling of G and τ (G) denotes the graph obtained after relabeling the vertices of G
according to τ . Let G and H be two graphs and let X ⊆ V (G) and Y ⊆ V (H). We say X
and Y are in the same orbit, in symbols X ∼orb Y , if there exists an isomorphism π from G
to H such that π(X) = Y . Let f be a function which maps a graph along with a subset of its
vertex set to a binary string, i.e. f (G, X) ∈ {0, 1}∗ and X ⊆ V (G). We call f an invariant
for a graph class C if f (G, X) = f (H, Y ) whenever X ∼orb Y and G, H ∈ C. Let us call a
function f which maps a graph G to a family of subsets of its vertex set, i.e. f (G) ⊆ P(V (G)),
a vertex set selector. For example, the function that maps a graph to the set of its cliques is
a vertex set selector. The characteristic function χf of a vertex set selector f is defined as
χf (G, X) = 1 ⇔ X ∈ f (G). We say a vertex set selector f is invariant for a graph class C if
its characteristic function χf is an invariant for C. We call f globally invariant if χf is an
invariant for all graphs. Intuitively, a vertex set selector f is invariant for C if a graph G ∈ C
can be arbitrarily relabeled and f still returns the ‘same’ vertex sets as before w.r.t. ∼orb .
The following definitions are with respect to a graph G. Throughout the paper it will be
3
4
Canonical Representations for Circular-Arc Graphs Using Flip Sets
always clear from context with respect to what graph these expressions are to be interpreted.
For a vertex v we define its open neighborhood N (v) as the set of vertices which are adjacent
to v and its closed neighborhood N [v] = N (v) ∪ {v}. A vertex v is called universal if
N [v] = V (G). For two vertices u, v we say that u and v are twins if N [u] = N [v]. A graph G
is twin-free if for every pair of distinct vertices u =
6 v it holds that N [u] 6= N [v]. A twin class
is an inclusion-maximal set of vertices X such that for all u, v ∈ X it holds that u and v are
twins. For two subsets of vertices S, S 0 with S 0 ⊆ S we define the exclusive neighborhood
NS (S 0 ) as all vertices v ∈ V (G) \ S such that v is connected to all vertices in S 0 and to none
in S \ S 0 . Let A be a square matrix with entries from a set E. For a vertex u of A and x ∈ E
we define N x (u) = {v ∈ V | Au,v = x}.
Logspace Transducers and Reductions. We assume deterministic Turing machines as
default model of computation. A logspace transducer is a deterministic Turing machine M
with a read-only input tape, a work tape and a write-only output tape. The work tape is
only allowed to use at most O(log n) cells where n denotes the input length. To write onto
the output tape M has a designated state called output state with the following semantic. If
M enters the output state then the symbol in the current cell of the work tape is written to
the current cell of the output tape and the head on the output tape is moved one cell to the
right. Other than that, M cannot write or move the head on the output tape. This means
as soon as something is written to the output tape it cannot be modified afterwards. Let
Σ and Γ be the input and work alphabet of M respectively. Then M computes a function
fM : Σ∗ → Γ∗ . We say a (partial) function f is computed by a logspace transducer M if
f (x) = fM (x) whenever f (x) is defined. We call f logspace-computable if there exists a
logspace transducer M which computes f . The class of logspace-computable functions is
closed under composition. Let f be a function which maps words over some alphabet to words
over some other alphabet. We say that the length of f is polynomially bounded if |f (x)| is
polynomially bounded by |x|. Only functions whose length is polynomially bounded can be
logspace-computable since the runtime of a logspace transducer is polynomially bounded. A
language L ⊆ Σ∗ is in logspace if its characteristic function is logspace-computable.
Given two functions f and g we say f is logspace-reducible to g if there exists l ∈ N and
logspace-computable functions r1 , . . . , rl such that f can be expressed as composition of g
and r1 , . . . , rl . Intuitively, this means that an oracle which can compute g can be queried a
constant number of times when constructing a logspace transducer for f . For two functions f
and g we say that they are logspace-equivalent if f is logspace-reducible to g and vice versa.
Analogously, given three functions f, g1 , g2 we say f is logspace-reducible to g1 and g2 if
there exists l ∈ N and logspace-computable functions r1 , . . . , rl such that f can be expressed
as composition of g1 , g2 and r1 , . . . , rl .
Circular-Arc Graphs and Representations. A CA model is a set of arcs A = {A1 , . . . , An }
on the circle. Let p =
6 p0 be two points on the circle. Then the arc A specified by [p, p0 ]
is given by the part of the circle that is traversed when starting from p going in clockwise
direction until p0 is reached. We say that p is the left and p0 the right endpoint of A and
write l(·), r(·) to denote the left and right endpoint of an arc in general. If A = [p, p0 ] then
the arc obtained by swapping the endpoints A = [p0 , p] covers exactly the opposite part of
the circle plus the endpoints. We say A is obtained by flipping A. In our context, we are only
interested in the intersection structure of a CA model and thus only the relative position of
the endpoints to each other matter. All endpoints can w.l.o.g. be assumed to be pairwise
different and no arc covers the full circle. Under these assumptions, a CA model A with n
M. Chandoo
5
A1
2
ρ(i) = Ai
1
A2
ρ(G)
G
3
A5
5
4
A3
A4
Figure 1 A CA graph and a representation of it
arcs can be described as a unique string as follows. Pick an arbitrary arc A ∈ A and relabel
the arcs with 1, . . . , n in order of appearance of their left endpoints when traversing the circle
clockwise starting from the left endpoint of A. Then write down the endpoints in order of
appearance when traversing the circle clockwise starting from the left endpoint of A. Do
this for every arc and pick the lexicographically smallest resulting string as representation
for A. For example, the smallest such string for the CA model in Figure 1 would result
from choosing A1 : (l(1), r(1), l(2), r(5), l(3), r(2), . . . ). Let str(A) denote this smallest string
representation. For a CA model A let Ar be the CA model obtained after reversing the
order of its endpoints. Observe that reversing the endpoints does not affect the intersection
structure of a CA model. Therefore we consider two CA models A and B to be equal if
str(A) = str(B) or str(Ar ) = str(B).
Let G be a graph and ρ = (A, f ) consists of a CA model A and a bijective mapping f
from the vertices of G to the arcs in A. Then ρ is called a CA representation of G if for all
u=
6 v ∈ V (G) it holds that {u, v} ∈ E(G) ⇔ f (u) ∩ f (v) 6= ∅. We write ρ(x) to mean the arc
f (x) corresponding to the vertex x, ρ(G) for the CA model A and for a subset V 0 ⊆ V (G)
let ρ[V 0 ] = {ρ(v) | v ∈ V 0 }. A graph is a CA graph if it has a CA representation.
We say a CA model A has a hole if there exists a point on the circle which isn’t contained
by any arc in A. Every such CA model can be understood as interval model (a set of intervals
on the real line) by straightening the arcs. Conversely, every interval model can be seen as
CA model by bending the intervals. Therefore a graph is an interval graph iff it admits a
CA representation with a hole.
A family of sets F over some ground set is called Helly if for all subsets F 0 of F such that
all elements in F 0 intersect pairwise it holds that ∩A∈F 0 A is non-empty. A CA graph G is
called Helly (HCA graph) if it has a CA representation ρ with a Helly CA model ρ(G). This
is the case iff for all inclusion-maximal cliques C in G it holds that the overall intersection of
T
C in ρ is non-empty, i.e. v∈C ρ(v) 6= ∅. Every interval model has the Helly property and
therefore every interval graph is a Helly CA graph.
The intersection type of two circular arcs A and B can be one of the following five types:
di: A and B are disjoint — A ∩ B = ∅
cs: A contains B — B ⊂ A
cd: A is contained by B — A ⊂ B
cc: A and B jointly cover the circle (circle cover) — A G B and A ∪ B = whole circle
ov: A and B overlap — A G B and A ∪ B 6= whole circle
Using these types we can associate a matrix with every CA model. An intersection
matrix is a square matrix with entries {cc, cd, cs, di, ov}. Given a CA model A we define its
intersection matrix µA such that (µA )A,B ∈ {cc, cd, cs, di, ov} reflects the intersection type
of the arcs A =
6 B ∈ A. An intersection matrix µ is called a CA (interval) matrix if it is the
intersection matrix of some CA model (with a hole). See Figure 2 for an example of a CA
model and the CA matrix which it induces. Given an intersection matrix µ and two distinct
6
Canonical Representations for Circular-Arc Graphs Using Flip Sets
vertices u, v of µ we sometimes write u α v instead of µu,v = α if µ is clear from the context.
Also, we sometimes talk about an intersection matrix µ as if it were a graph. In that case we
consider two vertices u, v of µ to be adjacent if they do not have a di-entry in µ.
When trying to construct a CA representation for a CA graph G it is clear that whenever
two vertices are non-adjacent their corresponding arcs must be disjoint in every CA representation of G. For two adjacent vertices the intersection type of their corresponding arcs might
depend on the particular CA representation of G that one considers. Hsu has shown that
this ambiguity can be removed as follows [7].
We adopt the notation of [10].
I Definition 2.1. For a graph G we define its neighborhood matrix λG which is an intersection
matrix as
di , if {u, v} ∈
/ E(G)
cd , if N [u] ( N [v]
cs , if N [v] ( N [u]
(λG )u,v = cc , if N [u] G N [v] and N [u] ∪ N [v] = V (G)
and ∀w ∈ N [u] \ N [v] : N [w] ⊂ N [u]
and ∀w ∈ N [v] \ N [u] : N [w] ⊂ N [v]
ov , otherwise
for all u 6= v ∈ V (G).
Let µ be an intersection matrix with vertex set V and let ρ = (A, f ) where A is a CA
model and f is a bijective mapping from V to A. We say ρ is a CA representation of µ if f
is an isomorphism from µ to the intersection matrix µA of A. We denote the set of such CA
representations for µ with N (µ). The representation problem for CA matrices is to compute
a CA representation for a given CA matrix µ. The canonical representation problem for
CA matrices is defined analogously to the canonical representation problem for CA graphs.
We say ρ is a normalized CA representation for a graph G if ρ is a CA representation
for the neighborhood matrix λG of G. An example of a normalized representation can be
seen in Figure 3. Let us denote the set of all normalized CA representations for G with
N (G) = N (λG ).
I Lemma 2.2 (Corollary 2.3. [7]). Every twin-free CA graph G without a universal vertex
has a normalized CA representation, that is N (G) 6= ∅.
I Lemma 2.3. The canonical representation problem for CA graphs is logspace-reducible
to the canonical representation problem for vertex-colored twin-free CA graphs without a
universal vertex.
Proof. For a graph G let G0 denote the induced subgraph of G that is obtained by removing
all universal vertices from G and only taking one vertex from each twin-class and deleting
b
a
A
d
c
µA
a
b
c
d
a
cd
ov
cc
Figure 2 A CA model A and its intersection matrix µA
b
cs
di
di
c
ov
di
ov
d
cc
di
ov
-
M. Chandoo
7
the rest. Let c0 be a coloring of G0 which assigns each vertex the cardinality of its twin class
in G. It holds that (G0 , c0 ) and the number of universal vertices in G suffice to reconstruct
G. Let G be a CA graph. Compute the graph (G0 , c0 ). Since (G0 , c0 ) is twin-free and
without universal vertices we can compute a canonical representation ρ0 for it. For a vertex
v of G let v0 denote the twin of v that occurs in G0 . A canonical representation of G is
given by v 7→ ρ0 (v0 ) for every non-universal vertex v of G and every universal vertex of G is
represented by an arc which intersects with all other arcs.
J
Therefore for our purposes it suffices to consider only twin-free graphs without universal
vertices and a vertex-coloring.
I Proviso. From this point on we assume every graph to be twin-free and without a universal
vertex unless explicitly stated otherwise. As a consequence we view CA graphs as a set of
CA matrices in the sense that the neighborhood matrix of every CA graph is a CA matrix.
Flips in Intersection Matrices. McConnell [14] observed that the operation of flipping arcs
in CA models has a counterpart in intersection matrices. He called this counterpart operation
algebraic flips. Note that for two arcs A, B with intersection type α ∈ {cc, cd, cs, di, ov} the
intersection type of A and B is solely determined by α. More precisely, the intersection type
of A and B is Z10 (α) where Z10 is the function defined in Table 1. Similarly, the intersection
type of A and B is given by Z01 (α). Using the functions Zij we can define the operation of
flipping a set of vertices in an intersection matrix.
I Definition 2.4. Let µ be an intersection matrix with vertex set V and X ⊆ V . We define
the intersection matrix µ(X) obtained after flipping the vertices X in µ as
µ(X)
u,v = Zij (µu,v ) with i = 1 iff u ∈ X and j = 1 iff v ∈ X
for all u 6= v in V .
Since flipping the same set of arcs twice is an involution it follows that (µ(X) )(X) = µ.
I Definition 2.5. Let V be a set of vertices, let A be a set of arcs and let ρ be a function
that maps V to A. Then ρ(X) : V → A for X ⊆ V is defined as follows:
(
ρ(v) , if v ∈ X
(X)
ρ (v) =
ρ(v) , if v ∈
/X
Notice that flipping vertices in an intersection matrix is equivalent to flipping arcs in a
CA representation in the following sense. Given an intersection matrix λ and a subset of
its vertices X it holds that ρ ∈ N (λ) ⇔ ρ(X) ∈ N (λ(X) ). Also, it is not difficult to observe
that flipping is an isomorphism-invariant operation in the sense that flipping sets of vertices
which are in the same orbit lead to isomorphic intersection matrices.
4
4
5
cd/cs
1
cc
2
cd/cs
6
1
5
6
2
3
3
Figure 3 A CA graph and a normalized representation thereof. Every non-labeled edge corresponds
to an ov-entry in the neighborhood matrix.
8
Canonical Representations for Circular-Arc Graphs Using Flip Sets
3
Flip Trick
In this section we generalize the idea used by Köbler, Kuhnert and Verbitsky in [10] to
compute canonical representations for Helly CA graphs. They showed that finding canonical
representations for Helly CA graphs can be reduced to finding canonical representations for
vertex-colored interval matrices. We show that the idea behind this reduction also works for
CA matrices in general. Recall that CA graphs can be seen as special case of CA matrices
since the neighborhood matrix of every CA graph is a CA matrix. The converse does not
hold, i.e. there exist CA matrices which are not expressible as the neighborhood matrix of a
CA graph (for instance any CA matrix with only two vertices that are not disjoint). The key
result here, which is used in the subsequent sections, is that finding canonical representations
for CA matrices is logspace-reducible to the task of computing what we call an invariant flip
set function.
McConnell showed in [14] that CA representations for CA graphs can be computed as
follows. Given a CA graph G with neighborhood matrix λ one can compute a set of vertices
X of G such that λ(X) is an interval matrix. We call such a set X a flip set. Then by
computing an interval representation ρ for λ(X) and flipping back the arcs X in ρ one obtains
a CA representation for λ and therefore for G as well [14]. We essentially use the same
argument to obtain canonical CA representations.
I Definition 3.1. Let λ be a CA matrix. A subset of vertices X of λ is called a flip set if
there exists a representation ρ ∈ N (λ) and a point x on the circle such that v ∈ X iff ρ(v)
contains the point x.
The concept of flip sets has already been implicitly defined and used in both [14] and
[10]. They observed that λ(X) is an interval matrix whenever X is a flip set of a CA matrix
λ. In fact, the other direction holds as well leading to the following characterization.
I Lemma 3.2. Let λ be a CA matrix and X is a subset of vertices of λ. It holds that X is
a flip set iff λ(X) is an interval matrix.
Proof. “⇒”: Let X be a flip set of λ. Let ρ ∈ N (λ) be a witnessing representation of the
fact that X is a flip set, i.e. there exists a point x on the circle such that every arc ρ(v)
with v ∈ X contains x and every arc ρ(v) with v ∈
/ X does not contain x. Consider the
representation ρ(X) ∈ N (λ(X) ). It holds that no arc ρ(X) (v) with v ∈ V (λ) contains the
point x which implies that there is a hole in ρ(X) and thus λ(X) is an interval matrix.
“⇐”: Let X be a subset of vertices of λ such that λ(X) is an interval matrix. We argue
that X must be a flip set. Let ρ ∈ N (λ(X) ) be a CA representation of λ(X) containing a hole
at point x on the circle. Such a representation must exist since λ(X) is an interval matrix.
This means the arc ρ(v) does not contain the point x for every vertex v ∈ V (λ). Consider the
Table 1 Algebraic flip functions Zxy : {cc, cd, cs, di, ov} → {cc, cd, cs, di, ov}
Zxy (α)
Z00
Z01
Z10
Z11
cc
cc
cs
cd
di
cd
cd
di
cc
cs
cs
cs
cc
di
cd
di
di
cd
cs
cc
ov
ov
ov
ov
ov
M. Chandoo
9
(X)
representation ρ(X) ∈ N ((λ(X) ) ) = N (λ). Then it can be checked that ρ(X) (v) contains
the point x iff v is in X and therefore X is a flip set with respect to λ.
J
We already mentioned that the canonical representation problem for vertex-colored
interval matrices can be solved in logspace due to [10]. However, since the theorem that
we reference just states this result for uncolored interval matrices we shortly explain how
to modify the proof to incorporate the coloring, which is a straightforward task for anyone
familiar with the proof.
I Theorem 3.3 ([10, Thm. 5.5]). The canonical representation problem for vertex-colored
interval matrices can be solved in logspace.
Proof. In Theorem 5.5 of [10] it is stated that a canonical interval representation for an
interval matrix can be found in logspace. To prove this they convert the input interval
matrix λ into a colored tree T(λ) called ∆ tree which is a complete invariant for interval
matrices. The leafs of this tree correspond to the vertices of λ. By appending the color of a
vertex from our vertex-colored interval matrix λ to the existing color of its corresponding
leave node in the colored ∆ tree T(λ) one obtains a complete invariant for vertex-colored
interval matrices. Then by applying the same argument given in the proof of Theorem 5.5
one can also compute a canonical representation for a vertex-colored interval matrix using
this slightly modified colored ∆ tree.
J
A consequence of Lemma 3.2 and Theorem 3.3 is that flip sets can be recognized in
logspace. Given an intersection matrix λ and a subset of vertices X of λ it suffices to check
whether λ(X) is an interval matrix by trying to compute an interval representation.
I Definition 3.4. Let C be a class of CA matrices and f is a vertex set selector. The function
f is called an invariant flip set function for C if the following conditions hold:
1. For every λ ∈ C there exists an X ∈ f (λ) such that X is a flip set of λ
2. f is invariant for C
Recall that f is globally invariant if f is invariant for all intersection matrices.
I Theorem 3.5. Let C be a class of CA matrices. The canonical representation problem
for vertex-colored C is logspace-reducible to the problem of computing an invariant flip set
function for C.
Proof. Let f be an invariant flip set function for C. Given a vertex-colored CA matrix (λ, c)
with λ ∈ C a canonical representation can be computed as follows. For every flip set X ∈ f (λ)
we associate it with the colored interval matrix IX = (λ(X) , cX ) where cX (v) = (c(v), red) if
v is in X and (c(v), blue) if v is not in X for all v ∈ V (λ). For a colored interval matrix I let
ρ̂I denote a canonical representation of I. Such a canonical representation can be computed
in logspace due to Theorem 3.3. Let X̂ denote a flip set in f (λ) such that the interval model
of ρ̂IX̂ is lexicographically minimal, i.e. for all flip sets X in f (λ) it holds that the model
of ρ̂IX is not smaller than the model of ρ̂IX̂ . Let ρ̂ denote the CA representation that is
obtained after flipping the red arcs in ρ̂IX̂ . Since these are the arcs that were flipped to
convert λ into IX it holds that ρ̂ is a representation of λ. To see that this leads to a canonical
representation consider two isomorphic vertex-colored CA matrices (λ, c) and (µ, d) with
λ, µ ∈ C and V (λ) and V (µ) are disjoint. Let Iλ be the set of colored interval matrices IX
for all flip sets X ∈ f (λ), and the set Iµ is defined analogously. Let Mλ be the set of interval
models M such that there exists an I ∈ Iλ and M is the model underlying the canonical
representation ρ̂I of I. The set Mµ is defined analogously. Since f is invariant it follows that
10
Canonical Representations for Circular-Arc Graphs Using Flip Sets
for every I ∈ Iλ there exists an I 0 ∈ Iµ such that I and I 0 are isomorphic, and vice versa.
Since the models in Mλ and Mµ only depend on the isomorphism type of the matrices
in Iλ and Iµ it follows that Mλ = Mµ . The CA models which underlie the canonical
representations of λ and µ are both derived from the smallest element in Mλ = Mµ and
thus are identical.
J
Suppose that there is a partition of the set of CA graphs into two classes C and D such
that you can efficiently compute invariant flip set functions for both classes. One might be
misled into thinking that this implies canonical representations for all CA graphs can be
found efficiently. However, this is not the case unless the class C (or D) can be efficiently
recognized, or one of the two invariant flip set functions is globally invariant.
I Lemma 3.6. Let C and D be classes of CA matrices. The canonical representation problem
for C ∪ D is logspace-reducible to the canonical representation problem for C and the problem
of computing a globally invariant flip set function for D.
Proof. Let f be a globally invariant flip set function for D. Let D0 be the set of CA matrices
λ such that f (λ) contains a flip set. Clearly, D is a subset of D0 . It holds that f (λ) contains
a flip set iff λ ∈ D0 . Since f is globally invariant it follows that f is an invariant flip set
function for D0 . To obtain a canonical representation for a matrix λ ∈ C ∪ D first compute
f (λ). If f (λ) contains a flip set it holds that λ ∈ D0 and therefore the output of f can be
used to find a canonical representation for λ. If f (λ) contains no flip set it must be the case
that λ ∈ C and therefore the canonization algorithm for C can be applied.
J
We conclude this section by restating the invariant flip set function that was used in [10]
to compute canonical representations for Helly CA graphs and explain why it is correct:
fHCA (G) = N [u] ∩ N [v] | u, v ∈ V (G)
In a Helly CA graph G every inclusion-maximal clique C of G is a flip set. To see why
this holds let ρ be a representation of G with the Helly property. Since C is a clique this
means every pair of arcs ρ(u) and ρ(v) with u, v ∈ C intersects. By the Helly property it
T
follows that the overall intersection v∈C ρ(v) is non-empty. This means there exists a point
x on the circle such that every arc ρ(v) with v ∈ C contains x. Assume there exists a vertex
w ∈ V (G) \ C such that ρ(w) contains x. This means w must be adjacent to every vertex in
C, which contradicts that C is inclusion-maximal. Hence C is a flip set.
In [10, Thm. 3.2] it is shown that every Helly CA graph contains at least one inclusionmaximal clique which can be expressed as the common neighborhood of two vertices. Therefore
fHCA (G) returns at least one flip set for every Helly CA graph G. Also, it is trivial to see
that fHCA is globally invariant.
4
Uniform Circular-Arc Graphs
We define the class of uniform CA graphs for which computing a particular invariant flip set
function reduces to computing a representation. As a consequence, canonical representations
for this class of CA graphs can be computed in polynomial-time. This is an interesting
class for two reasons. First, it seems to capture the instances where it is easy to apply
the flip trick. Secondly, its complement (within the CA graphs) is a rather exotic class
of CA graphs with a quite particular structure. While the initial definition of uniformity
makes it apparent why it suffices to find an arbitrary representation in order to obtain a
canonical one, it is rather impractical when trying to understand what constitutes a uniform
M. Chandoo
11
CA graph. We provide a more pleasant characterization of uniform CA graphs in terms of
how certain triangles in a CA graph can be represented. This alternative characterization
also reveals that every Helly CA graph is uniform. Additionally, we show that the canonical
representation problem for uniform CA graphs is logspace-equivalent to what we call the
non-Helly triangle representability problem. This problem is: given a CA graph G and a set
T of three pairwise overlapping vertices as input, does there exist a representation ρ of G
such that T covers the whole circle in ρ?
The following kind of flip set will lead us to uniform CA graphs when trying to compute
canonical representations. Given a CA matrix λ recall that X is a flip set of λ if there exists
a representation ρ ∈ N (λ) and a point x on the circle such that x ∈ ρ(v) iff v ∈ X for all
vertices v of λ. We impose the additional restriction that x is not allowed to be an arbitrary
point on the circle but instead has to be one of the endpoints in ρ.
I Definition 4.1. Let λ be a CA matrix and u ∈ V (λ). A flip set X of λ is a u-flip set if
there exists a representation ρ ∈ N (λ) and an endpoint x of ρ(u) such that v ∈ X iff ρ(v)
contains the point x.
Clearly, every CA graph has a u-flip set for every vertex u. On the other hand, there
are CA graphs that have flip sets which are not u-flip sets for any vertex u. For example,
consider the cycle graph with n ≥ 4 vertices. Every flip set that consists of exactly one vertex
is not a u-flip set for any vertex u of the cycle graph.
Consider the following task: given a CA graph G and a vertex u, find a u-flip set of
G. Clearly, no vertex v which is disjoint from u or contained by u belongs to X since in
every representation the arc of v does not contain any of the two endpoints of the arc of u.
Similarly, if a vertex v contains u or forms a circle cover with u then in every representation
the arc of v contains both endpoints of u and therefore must be included in X. See Figure 4
for a schematic overview.
It remains to decide for the set of vertices N ov (u) that overlap with u whether they should
be included in X. A vertex v which overlaps with u contains exactly one of the endpoints
of u in any representation. Let x, y be two vertices that overlap with u. We say x and y
overlap from the same side with u in ρ if ρ(x) and ρ(y) contain the same endpoint of ρ(u).
Evidently, this is an equivalence relation with respect to v and ρ which partitions N ov (u)
into two parts, namely the part which contains the left endpoint and the one which contains
the right endpoint. If X is a u-flip set then X ∩ N ov (u) must be an equivalence class of the
‘overlap from the same side with u in ρ’-relation for some ρ ∈ N (G).
I Definition 4.2. For a CA matrix λ and a vertex u of λ we say a partition Y of N ov (u)
into two parts is a u-ov-partition if there exists a representation ρ ∈ N (λ) such that two
vertices x, y ∈ N ov (u) are in the same part of Y iff ρ(x) and ρ(y) overlap from the same side
with ρ(u). We say ov-partition to mean an u-ov-partition for an arbitrary u ∈ V (λ).
u
X1
X2
u
a
b
Figure 4 Exemplary u-flip sets X1 and X2
d
Pu = {{a, b}, {c, d}}
c
Figure 5 Example of a u-overlap partition Pu
12
Canonical Representations for Circular-Arc Graphs Using Flip Sets
In general, for a vertex u of a CA graph G there can be multiple u-ov-partitions. In
fact, there are instances with exponentially many u-ov-partitions with respect to |N ov (u)|.
A trivial way of obtaining at least one u-ov-partition for every vertex u of a CA graph G
is to compute an arbitrary representation ρ ∈ N (G). But the ov-partitions obtained by
this method are not invariant and thus do not yield canonical representations. However, if
one considers CA graphs where there is only one u-ov-partition for every vertex u then an
arbitrary representation suffices.
I Definition 4.3 (Uniform CA Graphs). A CA graph G is uniform if for every vertex u in G
there exists exactly one u-ov-partition. This partition is denoted by Pu = {Pu,1 , Pu,2 }.
I Lemma 4.4. The following mapping is an invariant flip set function for uniform CA
graphs. Let G be a uniform CA graph.
[
Funiform (G) =
{u} ∪ N cd (u) ∪ N cc (u) ∪ Pu,i
u∈V (G)
i∈{1,2}
Proof. Let G be a uniform CA graph and X is in Funiform (G) with X = {u} ∪ N cd (u) ∪
N cc (u) ∪ Pu,i for some u ∈ V (G) and i ∈ {1, 2}. It follows from Figure 4 and the definition
of ov-partitions that X is a u-flip set. The invariance of Funiform (G) follows from the fact
that the intersection type of two vertices as well as the property of being an ov-partition is
independent of the vertex labels.
J
We remark that the function Funiform is undefined for non-uniform CA graphs since the
sets Pu,1 and Pu,2 are not well-defined in that context.
I Theorem 4.5. Canonical representations for uniform CA graphs can be computed in
polynomial-time.
Proof. Let G be a uniform CA graph. Compute a normalized representation ρ of G
and extract the u-ov-partition for each vertex u from ρ. Then compute Funiform (G) from
Lemma 4.4 to obtain a canonical CA representation for G. Since CA representations can be
computed in polynomial-time (see for instance [14]) it follows that this procedure also works
in polynomial-time.
J
Considering that our definition of uniform CA graphs arose from the desire to compute
invariant u-flip sets one might expect that these graphs are only a small special case of CA
graphs. Surprisingly, quite the opposite is the case as we will see. We give an alternative
definition of uniform CA graphs which gives a better intuition as to why many CA graphs
are uniform.
I Definition 4.6. Let λ be a CA matrix. An ov-triangle T of λ is a set of three vertices that
overlap pairwise, i.e. for all u 6= v in T it holds that u ov v. An ov-triangle T is representable
as non-Helly triangle (interval triangle) if there exists a representation ρ ∈ N (λ) such that
the set of arcs {ρ(x) | x ∈ T } does (not) cover the whole circle. Let TNHT (λ) and TIT (λ)
denote the sets of ov-triangles representable as non-Helly triangles and interval triangles
respectively.
This definition also applies to CA graphs via their neighborhood matrix, i.e. TIT (G) =
TIT (λ) and TNHT (G) = TNHT (λ) where λ is the neighborhood matrix of G. See Figure 6 for
M. Chandoo
13
an example where the vertices u, x, z are represented as non-Helly triangle on the left and
interval triangle on the right.
Recall that a set of arcs which intersect pairwise but have overall empty intersection
is called non-Helly. Since three pairwise overlapping arcs that cover the whole circle have
overall empty intersection we call such a set a non-Helly triangle. In fact, one can verify that
this is the only non-Helly arrangement of three arcs. A complete list of inclusion-minimal
non-Helly CA models can be found in [8, Corrollary 3.1].
I Theorem 4.7. A CA graph G is uniform iff TIT (G) ∩ TNHT (G) = ∅.
Proof. “⇒”: Assume there exists a uniform CA graph G with TIT (G) ∩ TNHT (G) 6= ∅.
Let T be an ov-triangle in TIT (G) ∩ TNHT (G) and T = {x, y, z}. This means there exist
two representations ρI , ρN ∈ N (G) such that T is represented as interval triangle in ρI
and as non-Helly triangle in ρN . We assume w.l.o.g. that ρI (y) ⊂ ρI (x) ∪ ρI (z), i.e. y is
placed in-between x and z in ρI . This means y and z must be in the same part of the
unique x-ov-partition Px . However, y and z do not contain the same endpoint of x in the
representation ρN , which contradicts that G is uniform.
“⇐”: Assume there exists a CA graph G with TIT (G) ∩ TNHT (G) = ∅ that is not uniform.
This means there exist a vertex u, two vertices x, y ∈ N ov (u) and two representations
ρ, ρ0 ∈ N (G) such that x and y overlap from the same side with u in ρ but not in ρ0 . This
implies that x and y must overlap and therefore T = {u, x, y} is an ov-triangle. Notice
that T must be represented as interval triangle in ρ because x and y both contain the same
endpoint of u. It holds that T is represented as interval triangle in ρ0 as well since otherwise
T ∈ TIT (G) ∩ TNHT (G). Also, we assume w.l.o.g. that ρ(y) ⊂ ρ(x) ∪ ρ(u). Since u and y
overlap it holds that N [u] \ N [y] 6= ∅. Due to ρ0 it follows that N [u] \ N [y] ⊆ N [u] ∩ N [x].
For a vertex z ∈ N [u] \ N [y] to intersect with both u and x it is necessary that z overlaps
with u and x due to the representation ρ. It follows that {u, x, z} is represented as non-Helly
triangle in ρ. On the other hand, {u, x, z} must be represented as interval triangle in ρ0 and
therefore TIT (G) ∩ TNHT (G) 6= ∅, contradiction. See Figure 6 for a schematic overview of ρ
and ρ0 .
J
Observe that if an ov-triangle T of G is representable as non-Helly triangle then this
implies that T must have certain structural properties in G. For example, every vertex of
G must be adjacent to at least one of the vertices in T since T covers the whole circle in
some representation. Similarly, if T is representable as interval triangle this also implies
some structural properties. For instance, there must be an x ∈ T such that every vertex that
is adjacent to x must also be adjacent to at least one other vertex in T . If an ov-triangle
is representable as both non-Helly triangle and interval triangle then it must satisfy all of
these structural properties at once. As a consequence such an ov-triangle must have a very
particular structure which extends to the whole graph as we will see in the next section.
A CA graph is Helly if it has a Helly CA representation. In [8, Theorem 4.1] it is shown
that every ‘stable’ representation of a Helly CA graph is Helly. Since every normalized
y
u
x
ρ
z
x
u
y
ρ0
z
Figure 6 “⇐”-direction in the proof of Theorem 4.7
14
Canonical Representations for Circular-Arc Graphs Using Flip Sets
representation has the ‘stable’ property it follows that a CA graph is Helly iff every normalized
representation of it is Helly. If a CA graph G is Helly this implies that TNHT (G) is empty,
and therefore every Helly CA graph is uniform.
A natural question to consider is the computational complexity of deciding whether an
ov-triangle is representable as non-Helly triangle or interval triangle. Given a CA graph G
and an ov-triangle T of G let us call the problem of deciding whether T is in TNHT (G) the
non-Helly triangle representability problem. Analogously, deciding whether T is in TIT (G) is
called the interval triangle representability problem. In the case of uniform CA graphs these
two problems are complementary, i.e. an ov-triangle T is in TNHT (G) iff T is not in TIT (G).
In the following, we show that solving either of these two problems for uniform CA graphs is
logspace-equivalent to computing a canonical representation for uniform CA graphs.
I Definition 4.8. Let G be a CA graph and T = {u, v, w} is an ov-triangle of G. We say v
is amidst u and w if one of the following conditions holds:
1. NT (u) and NT (w) are non-empty
2. there exists a z ∈ NT (u, w) such that {u, w, z} ∈ TNHT (G)
I Lemma 4.9. Let G be a uniform CA graph and T = {u, v, w} is an ov-triangle of G with
T ∈
/ TNHT (G). Then the following statements are equivalent:
1. v is amidst u and w
2. ∃ρ ∈ N (G) : ρ(v) ⊂ ρ(u) ∪ ρ(w)
3. ∀ρ ∈ N (G) : ρ(v) ⊂ ρ(u) ∪ ρ(w)
Proof. “2 ⇒ 1”: Let ρ be in N (G) such that ρ(v) ⊂ ρ(u) ∪ ρ(w) and assume that v is not
amidst u, w. Since v overlaps with u and w it holds that N [u] \ N [v] and N [w] \ N [v] are
non-empty. Because NT (u) = NT (w) = ∅ it must hold that NT (u, w) 6= ∅. Let z ∈ NT (u, w).
For z to intersect with u and w in ρ it must hold that {u, w, z} is represented as non-Helly
triangle in ρ. This contradicts the assumption that v is not amidst u, w.
“1 ⇒ 3”: Let v be amidst u and w and assume that there exists a ρ ∈ N (G) such that
ρ(v) 6⊂ ρ(u) ∪ ρ(w). Since T ∈
/ TNHT (G) and G is uniform it follows by Theorem 4.7 that
T must be represented as interval triangle in every representation, which includes ρ. We
assume w.l.o.g. that ρ(w) ⊂ ρ(u) ∪ ρ(v). From that it follows that NT (w) is empty and
therefore there must be a z ∈ NT (u, w) such that {u, w, z} is a non-Helly triangle in ρ, which
is impossible.
“3 ⇒ 2”: clear.
J
I Definition 4.10. Let G be a CA graph and u ∈ V (G). Let the binary relation ∼u on
N ov (u) be defined such that x ∼u y holds if one of the following holds:
1. x = y
2. x cd y or x cs y
3. x ov y, {u, x, y} ∈
/ TNHT (G) and u is not amidst x and y
I Lemma 4.11. For every uniform CA graph G and u ∈ V (G) it holds that the partition
induced by ∼u equals the unique u-ov-partition Pu . Stated differently, x ∼u y iff x and y are
in the same part of Pu .
Proof. “⇒”: Let x ∼u y and assume for the sake of contradiction that x and y are not in
the same part of the u-ov-partition. This means there exists a representation ρ ∈ N (G)
such that ρ(x) and ρ(y) contain different endpoints of ρ(u). This is only possible if x and
y overlap. Since {u, x, y} ∈
/ TNHT (G) this means {u, x, y} must be represented as interval
triangle in ρ. In order for ρ(x) and ρ(y) to contain different endpoints of ρ(u) it must
M. Chandoo
hold that ρ(u) ⊂ ρ(x) ∪ ρ(y), which implies that u is amidst x and y by Lemma 4.9. This
contradicts x ∼u y.
“⇐”: Let x and y be in the same part of the u-ov-partition and assume that x ∼u y does
not hold. This implies that x and y must overlap and therefore {u, x, y} form an ov-triangle.
For x ∼u y to not hold it must be either the case that {u, x, y} is only representable as
non-Helly triangle or u is amidst x and y. In both cases this contradicts x and y being in
the same part of the u-ov-partition.
J
I Theorem 4.12. The representation, canonical representation, non-Helly triangle representability and interval triangle representability problem for uniform CA graphs are logspaceequivalent.
Proof. The non-Helly triangle representability and interval triangle representability problem
for uniform CA graphs are logspace-equivalent because they are complementary in the sense
that an ov-triangle is representable as non-Helly triangle iff it is not representable as interval
triangle. This follows from the fact that an ov-triangle can only be either represented as
non-Helly triangle or interval triangle and these two possibilities are mutually exclusive in
the case of uniform CA graphs. As a consequence these two problems are trivially reducible
to the representation problem for uniform CA graphs. Given a uniform CA graph G, an
ov-triangle T of G and a representation ρ ∈ N (G) it holds that T ∈ TNHT (G) iff T ∈
/ TIT (G)
iff T is represented as non-Helly triangle in ρ.
The representation problem is obviously reducible to the canonical representation problem.
Therefore it remains to show that the canonical representation problem for uniform CA
graphs is reducible to the non-Helly triangle representability problem. To obtain a canonical
representation for a uniform CA graph we can use the invariant flip set function given in
Lemma 4.4. To compute this function we need to figure out the unique ov-partitions for each
vertex. By Lemma 4.11 this can be done by computing the equivalence relation ∼u for each
vertex u. It can be verified that this relation is computable in logspace using queries of the
form T ∈ TNHT (G).
J
The isomorphism problem for CA graphs can be reduced to the one for non-uniform CA
graphs in polynomial-time due to Theorem 4.5. However, a reduction from the canonical
representation problem for CA graphs to the one for non-uniform CA graphs does not
immediately follow from Theorem 4.5 unless uniform CA graphs can be recognized in
polynomial-time. An alternative approach to construct such a reduction is to solve the
non-Helly triangle representability problem for uniform CA graphs with an additional
requirement.
I Definition 4.13. The globally invariant non-Helly triangle representability problem for
uniform CA graphs is defined as follows. Let A be an algorithm that correctly decides the
non-Helly triangle representability problem for uniform CA graphs. Let fA be the function
computed by A, i.e. for a graph G and an ov-triangle T of G it holds that fA (G, T ) = 1 iff
A accepts (G, T ). We say A decides the globally invariant non-Helly triangle representability
problem for uniform CA graphs if fA is an invariant for all graphs. Stated differently, the
output of A must be independent of the vertex labels.
I Lemma 4.14. The canonical representation problem for CA graphs is logspace-reducible
to the globally invariant non-Helly triangle representability problem for uniform CA graphs
and the canonical representation problem for vertex-colored non-uniform CA graphs.
15
16
Canonical Representations for Circular-Arc Graphs Using Flip Sets
Proof. Suppose we are given an algorithm A which solves the globally invariant non-Helly
triangle representability problem for uniform CA graphs. We argue that A can be used to
compute a globally invariant flip set function for uniform CA graphs. From Lemma 3.6 it
then follows that the canonical representation problem for CA graphs reduces to that for
vertex-colored non-uniform CA graphs.
Given a CA graph G let ∆(G, A) be the set of ov-triangles T of G such that A accepts
(G, T ). If G is a uniform CA graph then ∆(G, A) = TNHT (G). Consider Definition 4.8 and
4.10 and suppose that each occurrence of TNHT (G) is replaced by ∆(G, A). Let us call the
A
new relation ∼A
u . Clearly, in the case of uniform CA graphs ∼u and ∼u coincide. Next,
consider the following variant of Funiform :
[
A
Funiform
(G) =
{u} ∪ N cd (u) ∪ N cc (u) ∪ X
u∈V (G)
X∈(N ov (u)/∼A
u)
A
A
where (N ov (u)/ ∼A
u ) denotes the equivalence classes of ∼u . If ∼u is not an equivalence
relation let (N ov (u)/ ∼A
u ) = ∅. If G is a uniform CA graph then it follows from Lemma 4.11
A
A
that Funiform (G) = Funiform (G). Therefore Funiform
is an invariant flip set function for uniform
A
CA graphs. Additionally, it can be verified that Funiform
is globally invariant due to the fact
A
that the answer of A is independent of the vertex labels. Also, the function Funiform
can be
computed in logspace using queries of the form T ∈ ∆(G, A). Observe that ∆(G, A) only
provides n3 bits of information with n = |V (G)| and therefore can be computed ‘in a single
query’ by a functional oracle which outputs the n3 bits of information.
J
5
Non-Uniform CA Graphs and Restricted CA Matrices
In the first part of this section we examine the structure of non-uniform CA graphs. Every
such graph must have two ov-triangles which have exactly one vertex in common and both are
representable as interval triangle and as non-Helly triangle. This pair of ov-triangles enforces
a particular structure in non-uniform CA graphs. In the second part we introduce restricted
CA matrices, which try to partly capture this structure. Roughly speaking, restricted CA
matrices can be seen as a generalization of the neighborhood matrices of non-uniform CA
graphs. We pay the price of considering this more general class of structures in order to
provide a logspace reduction from the canonical representation problem for CA graphs to
that of restricted CA matrices.
I Definition 5.1. Given a CA graph G, an induced 4-cycle C = (u, w, w0 , u0 ) of G and
v ∈ V (G) \ C. We say (C, v) is a non-uniformity witness of G if {u, v, w}, {u0 , v, w0 } ∈
TIT (G) ∩ TNHT (G). We also simply call (C, v) a witness of G.
I Theorem 5.2. A CA graph G is non-uniform iff G has a non-uniformity witness.
Proof. “⇒”: Let G be a non-uniform CA graph. Due to Theorem 4.7 there exists an
ov-triangle T of G with T ∈ TIT (G) ∩ TNHT (G). Let T = {u, v, w} and ρI ∈ N (G) such
that v is in-between u and w, i.e. ρI (v) ⊂ ρI (u) ∪ ρI (w). First, we show that there exists an
induced 4-cycle C = (u, w, w0 , u0 ) in G.
From the non-Helly triangle representation of T it follows that N [u]∪N [v]∪N [w] = V (G).
Since v is in-between u and w this means N [u] ∪ N [w] = V (G). It holds that u and w overlap.
Therefore one of the conditions in the definition of the neighborhood matrix for u and w to
form a circle cover must be violated. Let us assume w.l.o.g. that the violated condition is that
there exists a u0 ∈ N [u] \ N [w] such that N [u0 ] 6⊆ N [u]. This means u0 must overlap with u
M. Chandoo
u
17
w
u
w
v
v
3K2
x
X0
X1
X2
v0
u0
w0
X3
u0
w0
X4
Figure 7 Examples of non-uniform CA graphs and one uniform CA graph X4
and there exists a w0 ∈ N [u0 ] \ N [u]. Since w0 ∈
/ N [u] it follows from N [u] ∪ N [w] = V (G)
that w0 ∈ N [w] and because w is disjoint from u0 , and because w0 intersects with both u0
and w it follows that w0 overlaps with u0 and w. Therefore C = (u, w, w0 , u0 ) is an induced
4-cycle in G.
It remains to show that {u0 , v, w0 } is an ov-triangle and that it is in both TIT (G) and
TNHT (G). Consider the representation ρI from before. Assume for the sake of contradiction
that v does not overlap with u0 . Then due to ρI it must be the case that u0 is disjoint from v
and thus u0 ∈ NT (u). However, due to fact that T is representable as non-Helly triangle this
would imply that u0 is contained by u, which is not the case. Therefore u0 overlaps with v as
the other intersections types are out of question. For the same reason w0 overlaps with v and
hence T 0 = {u0 , v, w0 } is an ov-triangle. Now, it can be verified that in every representation
of G where T is a non-Helly triangle it follows that T 0 must be an interval triangle and vice
versa. This concludes that T 0 is in TIT (G) ∩ TNHT (G).
“⇐”: Follows directly from Theorem 4.7.
J
In Figure 7 five non-uniform CA graphs and one uniform CA graph (X4 ) are given by
their CA models. We explain how to verify this claim. First, we have to check that every
CA model is normalized. This means the graphs which are induced by these models must
be twin-free and without a universal vertex. Additionally, the intersection types of the arcs
must match the intersection types in the induced graph (or more precisely its neighborhood
matrix). A quick way to determine whether two overlapping arcs also overlap in the graph is
to check if they jointly occur in an induced n-cycle for some n ≥ 4.
To see that the first five CA graphs are non-uniform we have to find an ov-triangle that is
representable as both interval and non-Helly triangle. In the case of 3K2 this ov-triangle can
be {u, v, w}. In the given representation {u, v, w} is represented as interval triangle. Observe
that v and v 0 are in the same orbit and therefore the labels v and v 0 can be swapped in the
representation. After swapping v and v 0 the ov-triangle {u, v, w} is represented as non-Helly
triangle. For the graph X0 we can also choose the ov-triangle {u, v, w}. In this case there
is an automorphism which swaps u with u0 and w with w0 and has the other vertices as
fix-points. After changing the labels in the representation according to this automorphism it
18
Canonical Representations for Circular-Arc Graphs Using Flip Sets
1
u
w
w0
u0
cs
di
di
cs
ov
di
di
cs
2
cs
di
di
ov
ov
di
di
ov
di
cs
cs
di
di
ov
cs
di
3
di
cs
ov
di
di
ov
ov
di
ov
di
ov
ov
4
ov
di
ov
cs
ov
ov
di
ov
5
cs
ov
di
ov
ov
ov
ov
di
6
ov
cs
ov
di
di
ov
ov
ov
di
ov
cs
ov
7
ov
ov
ov
ov
Table 2 Intersection types of restricted CA matrices with witness cycle (u, w, w0 , u)
holds that {u, v, w} is represented as non-Helly triangle. We remark that 3K2 and X0 are
minimal in the sense that no induced subgraph of them is a non-uniform CA graph. Next, let
us consider the graphs X1 to X3 . Observe that the black arcs in each of these graphs form
an induced 3K2 subgraph. We assume that the black arcs are labeled with u, u0 , v, v 0 , w, w0
in the same way that the representation of 3K2 is labeled. It holds that v and v 0 are in
the same orbit in all of these four graphs because they have the same open neighborhood.
Therefore {u, v, w} is representable as both interval and non-Helly triangle due to the same
argument that we made for 3K2 .
To show that X4 is uniform we argue that it has a unique normalized representation,
i.e. |N (X4 )| = 1. Observe that this graph has a unique CA model. Additionally, it has no
non-trivial automorphism (it is rigid). Therefore X4 has a unique CA representation.
I Fact 5.3. Every non-uniform CA graph contains 3K2 or X0 as induced subgraph.
Proof. Let G be a non-uniform CA graph. Due to Theorem 5.2 there exists a witness
(C, v) of G with C = (u, w, w0 , u0 ). Since G does not contain a universal vertex it holds
that V (G) \ N [v] is non-empty. Due to the fact that {u, v, w} and {u0 , v, w0 } can be
represented as interval triangles it follows that NC (C \ {x}) ⊆ N [v] for all x ∈ C. Therefore
V (G) \ N [v] ⊆ NC (C) ∪ NC (u, u0 ) ∪ NC (w, w0 ). Suppose there is a v 0 ∈ NC (C) \ N [v].
Then the vertices of C along with v and v 0 form an induced 3K2 -subgraph of G. Assume
that this is not the case, i.e. NC (C) ⊆ N [v]. Since u and v overlap it must hold that
N [u] \ N [v] 6= ∅. The only vertices that can be adjacent to N [u] but not to N [v] must be
in NC (u, u0 ) since NC (C) ⊆ N [v]. Therefore there exists a vertex x ∈ NC (u, u0 ) that is not
adjacent to v. For the same reason there must be a vertex y ∈ NC (w, w0 ) not adjacent
to v because N [w] \ N [v] 6= ∅. The vertices of C along with v, x and y form an induced
X0 -subgraph.
J
I Definition 5.4 (Restricted CA Matrix). Let λ be a CA matrix. We say λ is a restricted CA
matrix if it contains an induced 4-cycle C = (u, w, w0 , u0 ) called witness cycle such that:
1. NC (u, w), NC (u0 , w0 ) and NC (x) are empty for every x ∈ C
2. For all x ∈ NC (C) it holds that x overlaps with all vertices in C
Observe that the intersection matrix of every CA model that is shown in Figure 7 is a
restricted CA matrix.
I Fact 5.5. Given an intersection matrix λ, vertices x, y1 , . . . , yk of λ and intersection types
α1 , . . . , αk , we say x is an (α1 , . . . , αk )-neighbor of (y1 , . . . , yk ) if λx,yi = αi for all i ∈ [k].
A CA matrix λ is restricted iff λ contains an induced 4-cycle C = (u, w, w0 , u0 ) such that for
all vertices x ∈ V (λ) \ C there exists a column α in Table 2 such that x is a α-neighbor of C.
Proof. We use the numbers in the table headline to refer to the different columns. For example, 2.3 refers to the third column from left in the second part of the table: (di, cs, ov, di).
M. Chandoo
“⇒”: Let λ be a restricted CA matrix with witness cycle C = (u, w, w0 , u0 ). We need to
show for every x ∈ V (λ) \ C there exists a column α in Table 2 such that x is a α-neighbor
of C. Due to the definition of restricted CA matrices it must hold that x is in (exactly) one
of the following seven sets: NC (C), NC (u, u0 ), NC (w, w0 ) or NC (C \ {z}) for a z ∈ C. If x is
in NC (C) then x overlaps with every vertex of C by definition. This corresponds to the last
column 7.1 of the table. If x ∈ NC (u, u0 ) then x is disjoint from w and w0 . In that case x is
an α-neighbor of C where α must be one of the four columns in part one of the table. For
the same reason if x ∈ NC (w, w0 ) then it is an α-neighbor of C where α corresponds to one
of the two columns in the second part of the table. If x is in NC (C \ {w}) then x is disjoint
from w and x overlaps with both u and w0 . The intersection type between x and u0 can be
one of the following: x overlaps with u or x is contained by u or x contains u. The first two
cases are covered by the third part of the table. However, if x contains u then there exists no
corresponding column in the table since it does not have any cd-entries. This can be resolved
by using the following observation: if x is in NC (C \ {w}) and contains u0 then (u, w, w0 , x)
is a witness cycle of λ as well. As a consequence we can assume without loss of generality
that a witness cycle C of λ can be chosen such that there exists no x ∈ NC (C \ {w}) which
contains u0 . The same argument applies to the remaining three cases x ∈ NC (C \ {z}) with
z ∈ {u, u0 , w0 }.
“⇐”: clear.
J
In the remainder of this section we prove that the canonical representation problem for
CA graphs is logspace-reducible to the canonical representation problem for vertex-colored
restricted CA matrices. The proof outline looks as follows. First, we define a subset of
uniform CA graphs, namely ∆-uniform CA graphs, for which the globally invariant nonHelly triangle representability problem can be solved in logspace. Therefore the canonical
representation problem for CA graphs is logspace-reducible to that of CA graphs which are
not ∆-uniform. This reduction follows from a slightly modified version of Lemma 4.14. Then
we show that the neighborhood matrix of a non-∆-uniform CA graph can be converted into
a vertex-colored restricted CA matrix by flipping ‘long’ arcs. By coloring the flipped arcs
the isomorphism type is preserved.
I Definition 5.6. For a graph G we define ∆G as the following set of ov-triangles (see
Definition 4.6). An ov-triangle T of G is in ∆G if there exist three pairwise different vertices
u, v, w in T such that the following holds:
1. N [u] ∪ N [v] ∪ N [w] = V (G)
2. For all z ∈ T it holds that if a vertex x ∈ NT (z) then x cd z
3. If there exist u0 , w0 such that (u, w, w0 , u0 ) is an induced 4-cycle and v overlaps with u0
and w0 then N [v] ⊆ N [u0 ] ∪ N [w0 ]
I Definition 5.7. A CA graph G is ∆-uniform if ∆G ∩ TIT (G) = ∅.
Let us explain the intuition behind these two definitions. The set ∆G approximates
TNHT (G). More precisely, whenever an ov-triangle T = {u, v, w} is in TNHT (G) this implies
that T satisfies certain constraints such as for example N [u] ∪ N [v] ∪ N [w] = V (G). The
set ∆G consists of three such constraints. Therefore if an ov-triangle is representable as
non-Helly triangle it must also be in ∆G , i.e. TNHT (G) ⊆ ∆G . The ∆-uniform CA graphs
can be alternatively seen as the subset of uniform CA graphs where the constraints of ∆G
suffice to characterize TNHT (G), i.e. ∆G = TNHT (G).
I Lemma 5.8. For every graph G it holds that TNHT (G) ⊆ ∆G . If G is a ∆-uniform CA
graph then TNHT (G) = ∆G .
19
20
Canonical Representations for Circular-Arc Graphs Using Flip Sets
Proof. For the first claim consider a graph G. If G is not a CA graph then TNHT (G) = ∅.
Therefore we can assume that G is a CA graph. Given an ov-triangle T ∈ TNHT (G) we
show that it must be in ∆G . Let ρ ∈ N (G) be a representation such that T = {u, v, w} is
represented as non-Helly triangle in it. Since ρ(u) ∪ ρ(v) ∪ ρ(w) covers the whole circle it
follows that N [u] ∪ N [v] ∪ N [w] = V (G), which is the first condition of Definition 5.6. To see
that the second condition holds we consider a vertex x ∈ NT (u) without loss of generality.
Since x is not adjacent to v and w it holds that ρ(x) ⊆ C \ (ρ(v) ∪ ρ(w)) where C denotes
the whole circle. Since C \ (ρ(v) ∪ ρ(w)) ⊂ ρ(u) it follows that ρ(x) ⊂ ρ(u). Due to the fact
that ρ is a normalized representation this implies that x is contained by u. To see that the
third condition of ∆G holds let u0 , w0 be vertices such that (u, w, w0 , u0 ) is an induced 4-cycle
of G. Since T is represented as non-Helly triangle in ρ it must hold that {u0 , v, w0 } is an
interval triangle in ρ with ρ(v) ⊂ ρ(u0 ) ∪ ρ(w0 ) and therefore N [v] ⊆ N [u0 ] ∪ N [w0 ].
For the second claim let G be a ∆-uniform CA graph. From the previous claim we know
that TNHT (G) ⊆ ∆G . Since every ov-triangle must be in TNHT (G) ∪ TIT (G) it follows that
∆G ⊆ TNHT (G) ∪ TIT (G). The definition of ∆-uniform requires ∆G ∩ TIT (G) = ∅ and thus
∆G ⊆ TNHT (G).
J
I Fact 5.9. ∆-uniform CA graphs are a strict subset of uniform CA graphs.
Proof. Assume there exists a ∆-uniform CA graph G which is not uniform. This means
there exists an ov-triangle T ∈ TNHT (G) ∩ TIT (G). Due to the previous lemma it holds that
TNHT (G) ⊆ ∆G . This implies that T ∈ ∆G ∩ TIT (G) which contradicts that G is ∆-uniform.
Therefore every ∆-uniform CA graph is uniform.
An example of a uniform CA graph that is not ∆-uniform is the graph X4 in Figure 7. In
the third paragraph after Theorem 5.2 we argued that X4 is a uniform CA graph because it
has a unique normalized representation. Assume that the black arcs of X4 are labeled with
u, u0 , v, v 0 , w, w0 in the same way that the representation of 3K2 is labeled in Figure 7. To
see that X4 is not ∆-uniform it suffices to check that the ov-triangle {u, v, w} is in ∆X4 and
represented as interval triangle.
J
I Corollary 5.10. The globally invariant non-Helly triangle representability problem for
∆-uniform CA graphs can be solved in logspace.
Proof. Given a CA graph G and an ov-triangle T output yes iff T ∈ ∆G . This is correct
because in the case of a ∆-uniform CA graph G it holds that ∆G = TNHT (G) (Lemma 5.8).
Clearly, ∆G is computable in logspace and an invariant.
J
I Lemma 5.11. Let G be a CA graph that is not ∆-uniform. Then there exists an induced
4-cycle C = (u, w, w0 , u0 ) such that N [u] ∪ N [w] = N [u0 ] ∪ N [w0 ] = V (G) and a vertex v that
overlaps with every vertex in C.
Proof. The argument is essentially the same as the one made for the “⇒”-direction in
the proof of Theorem 5.2. The difference is that instead of the stronger assumption that
T ∈ TNHT (G) we only require that T ∈ ∆G .
Since G is not ∆-uniform there exists an ov-triangle T = {u, v, w} of G such that T ∈ ∆G
and there is a representation ρ ∈ N (G) such that T is represented as interval triangle in ρ.
Furthermore, let us assume w.l.o.g. that ρ(v) ⊂ ρ(u) ∪ ρ(w). Since T ∈ ∆G it holds that
N [u] ∪ N [v] ∪ N [w] = V (G). Due to the interval representation of T in ρ it follows that
N [u] ∪ N [w] = V (G). Since u and w do not form a circle cover it must hold that there
exists a vertex u0 ∈ N [u] \ N [w] such that N [u0 ] \ N [u] is non-empty. If u0 is disjoint from
v it follows that u0 must be contained by u from the second condition in Definition 5.6 of
M. Chandoo
∆G . This cannot be the case and therefore u0 ∈ NT (u, v). For u0 to have a neighbor which
is not adjacent to u it must hold that ρ(u0 ) 6⊆ ρ(u). Therefore u0 overlaps with u and v.
Let w0 ∈ N [u0 ] \ N [u]. If w0 ∈ NT (w) then w0 would be contained by w due to the second
condition of ∆G . Again, this cannot be the case and therefore w0 ∈ NT (v, w). From the
representation ρ it follows that w must overlap with u0 , v and w. Then C = (u, w, w0 , u0 ) is
an induced 4-cycle of G such that v overlaps with every vertex of C. It remains to show that
N [u0 ] ∪ N [w0 ] = V (G). Due to the third condition of ∆G it holds that N [v] ⊆ N [u0 ] ∪ N [w0 ].
Additionally, it holds that ρ(u) \ ρ(v) ⊂ ρ(u0 ) and ρ(w) \ ρ(v) ⊂ ρ(w0 ). As a consequence
N [u0 ] ∪ N [w0 ] = V (G).
J
I Corollary 5.12. Canonical representations for CA graphs without induced 4-cycle can be
computed in logspace.
Proof. By Lemma 5.11 the class of CA graphs without induced 4-cycle is a subset of ∆uniform CA graphs and due to Corollary 5.10 and Theorem 4.12 a canonical representation
for such graphs can be computed in logspace.
J
I Corollary 5.13. Helly CA graphs are a strict subset of ∆-uniform CA graphs.
Proof. Assume G is a Helly CA graph which is not ∆-uniform. Then due to Lemma 5.11
there exists an induced 4-cycle C and a vertex v not in C which overlaps with every vertex
in C. In any normalized representation of G it must hold that v forms a non-Helly triangle
with two vertices from C. This contradicts that G is Helly. The graph
is a ∆-uniform
CA graph which is not Helly.
J
I Theorem 5.14. The canonical representation problem for CA graphs is logspace-reducible
to the canonical representation problem for vertex-colored restricted CA matrices.
Proof. For brevity let Z denote the set of all CA graphs which are not ∆-uniform. Since the
globally invariant non-Helly triangle representability problem for ∆-uniform CA graphs can
be solved in logspace (see Corollary 5.10) it follows from a modified version of Lemma 4.14
that the canonical representation problem for CA graphs is logspace-reducible to the canonical
representation problem for vertex-colored Z. To see this replace ‘uniform’ with ‘∆-uniform’
and ‘non-uniform’ with ‘non-∆-uniform’ in the statement (and proof) of Lemma 4.14.
(X)
For a CA graph G let us say a subset of vertices X of G is an R-flip set if λG is a
restricted CA matrix. To find canonical representations for Z we construct an invariant
vertex set selector f such that f (G) contains at least one R-flip set for every G ∈ Z. Then
to obtain a canonical representation for G ∈ Z let X̂ denote the R-flip set in f (G) such that
(X̂)
canon(λG , cX̂ ) is lexicographically minimal with cX being the coloring which assigns every
vertex v ∈ X the color red and the other vertices are blue. Let ρ be a canonical normalized
(X̂)
representation for (λG , cX̂ ). Then ρ(X̂) is a canonical representation for G. Notice, that
ρ(X̂) can be computed in logspace by computing canonical representations for vertex-colored
restricted CA matrices. The correctness of this approach follows from the same argument
made in the proof of Theorem 3.5 in the flip trick section. The analogy is straightforward.
The R-flip sets in this context correspond to flip sets and the invariant vertex set selector f
takes the place of the invariant flip set function. Given a CA graph G and X ⊆ V (G) it can
(X)
be easily checked in logspace whether λG is a restricted CA matrix.
For a CA graph G let C(G) denote the set of all ordered induced 4-cycles in G. Now, we
claim that the following logspace-computable function f is an invariant vertex set selector
with the desired property:
21
22
Canonical Representations for Circular-Arc Graphs Using Flip Sets
[
f (G) =
{x ∈ V (G) \ C | ∃y ∈ C : x cs y}
C∈C(G)
It is not difficult to check that f is invariant. It remains to argue why f (G) contains at
least one R-flip set for every G ∈ Z. Let G ∈ Z and C = (u, w, w0 , u0 ) is an induced 4-cycle
in G such that N [u] ∪ N [w] = N [u0 ] ∪ N [w0 ] = V (G). The existence of such an induced
4-cycle is guaranteed by Lemma 5.11. Observe that if there exists a u1 ∈ NC (u, w, u0 ) with
u1 cs u then C1 = (u1 , w, w0 , u0 ) also satisfies the previous condition N [u1 ] ∪ N [w] = V (G).
Therefore we can assume that there exists no z ∈ C and z1 ∈ NC (N [z] ∩ C) such that z1 cs z.
From N [u] ∪ N [w] = N [u0 ] ∪ N [w0 ] = V (G) it immediately follows that NC (u, w), NC (u0 , w0 )
and NC (x) are empty for every x ∈ C.
We prove that λ(X) is a restricted CA matrix with witness cycle C where λ is the
neighborhood matrix of G and X = {x ∈ V (G) \ C | ∃y ∈ C : x cs y}. Note that X ∈ f (G)
via C. To reference the neighborhoods of G (which are the same as the ones of λ) or λ(X)
(X)
(X)
we write N G and N λ
to distinguish between them. First, we show that NCλ (u, w) = ∅.
(X)
Assume the opposite, i.e. there exists x ∈ NCλ (u, w). If x was not flipped, i.e. x ∈
/ X,
then it also holds that x ∈ NCG (u, w), which contradicts that NCG (u, w) is empty. If x
was flipped, i.e. x ∈ X, then it must be the case that x contains u0 and w0 in λ. This
means NG [u0 ] ∪ NG [w0 ] ⊆ NG [x] which implies that x is a universal vertex in G since
(X)
NG [u0 ] ∪ NG [w0 ] = V (G), contradiction. For the same reason it holds that NCλ (u0 , w0 )
(X)
(X)
and NCλ (z) are empty for all z ∈ C. It remains to show that for all x ∈ NCλ (C) it
(X)
holds that x overlaps with all vertices of C in λ(X) . Notice that λx,z ∈ {ov, cs, cc} for
every z ∈ C. Otherwise x would not be in NC (C). We consider the following two cases: in
the first one we assume that x contains one vertex of C in λ(X) and in the second one we
assume that x forms a circle cover with one vertex of C in λ(X) . We prove that neither of
these cases can occur and therefore x must overlap with all vertices of C in λ(X) . For the
first case assume that w.l.o.g. x contains u in λ(X) and intersects with the other vertices
of C in λ(X) . If x ∈ X then it was flipped. It follows that x was disjoint from u in λ and
therefore x ∈ NCG (w, w0 , u0 ). Since x ∈ X it also must hold that x contains at least one of the
vertices w, w0 , u0 in G. It follows that x contains w0 since it cannot contain the other two in λ.
However, this contradicts our choice of C which says that there exists no w10 ∈ NCG (w, w0 , u0 )
such that w10 contains w0 in λ. If x ∈
/ X then it must hold that x already contained u in λ.
But then x should be in X, contradiction. For the second case assume x forms a circle cover
with u in λ(X) . If x forms a circle cover with u then this implies that x contains w0 in λ(X)
and therefore this reduces to the first case. We conclude that both conditions of Definition
5.4 are satisfied and hence λ(X) is a restricted CA matrix.
J
6
Further Research
Finding a polynomial-time isomorphism test for CA graphs remains an open problem. We
have shown that it suffices to consider only non-uniform CA graphs for this problem. This
particular class of CA graphs offers quite a lot of structure, which is caused by what we
named non-uniformity witnesses. It seems plausible that such witnesses can be exploited
to devise an isomorphism test. Additionally, we proved that the canonical representation
problem for CA graphs is logspace-reducible to that of restricted CA matrices. The central
question with regard to the flip trick is how invariant flip sets for restricted CA matrices or
non-uniform CA graphs can be computed. Also, we remark that CA representations for CA
graphs can be computed in logspace if flip sets for restricted CA matrices can be found in
M. Chandoo
logspace. Another interesting problem is to extend Definition 5.6 of ∆G such that it captures
TNHT (G) on uniform CA graphs, i.e. ∆G = TNHT (G) for all uniform CA graphs G. If this
can be done in such a way that ∆G remains an invariant and computable in logspace then
everything that is said about ∆-uniform CA graphs in section 5 also applies to uniform CA
graphs.
Acknowledgements. We thank the anonymous reviewers for their insightful comments and
suggestions that helped us to improve the quality of this work.
References
1
2
3
4
5
6
7
8
9
10
Yixin Cao, Luciano N. Grippo, and Martín D. Safe.
Forbidden induced subgraphs of normal Helly circular-arc graphs: Characterization and detection. Discrete Applied Mathematics, 216, Part 1:67 – 83, 2017.
Special Graph Classes
and Algorithms — in Honor of Professor Andreas Brandstädt on the Occasion of
His 65th Birthday.
URL: http://www.sciencedirect.com/science/article/pii/
S0166218X15004308, doi:http://dx.doi.org/10.1016/j.dam.2015.08.023.
Maurice Chandoo. Deciding Circular-Arc Graph Isomorphism in Parameterized Logspace.
In Nicolas Ollinger and Heribert Vollmer, editors, 33rd Symposium on Theoretical Aspects of Computer Science (STACS 2016), volume 47 of Leibniz International Proceedings
in Informatics (LIPIcs), pages 26:1–26:13, Dagstuhl, Germany, 2016. Schloss Dagstuhl–
Leibniz-Zentrum fuer Informatik. URL: http://drops.dagstuhl.de/opus/volltexte/
2016/5727, doi:http://dx.doi.org/10.4230/LIPIcs.STACS.2016.26.
Andrew Curtis, Min Lin, Ross McConnell, Yahav Nussbaum, Francisco Soulignac, Jeremy
Spinrad, and Jayme Szwarcfiter. Isomorphism of graph classes related to the circular-ones
property. Discrete Mathematics and Theoretical Computer Science, 15(1), 2013. URL:
http://www.dmtcs.org/dmtcs-ojs/index.php/dmtcs/article/view/2298.
Elaine Marie Eschen. Circular-Arc Graph Recognition and Related Problems. PhD thesis,
Vanderbilt University, Nashville, TN, USA, 1998. UMI Order No. GAX98-03921.
F. Gavril. Algorithms on circular-arc graphs. Networks, 4(4):357–369, 1974. URL: http:
//dx.doi.org/10.1002/net.3230040407, doi:10.1002/net.3230040407.
Martin Charles Golumbic. Algorithmic Graph Theory and Perfect Graphs (Annals of Discrete Mathematics, Vol 57). North-Holland Publishing Co., Amsterdam, The Netherlands,
The Netherlands, 2004.
Wen-Lian Hsu. O(M · N ) algorithms for the recognition and isomorphism problems on
circular-arc graphs. SIAM J. Comput., 24(3):411–439, June 1995. URL: http://dx.doi.
org/10.1137/S0097539793260726, doi:10.1137/S0097539793260726.
Benson L. Joeris, Min Chih Lin, Ross M. McConnell, Jeremy P. Spinrad, and Jayme L.
Szwarcfiter. Linear-time recognition of Helly circular-arc models and graphs. Algorithmica,
59(2):215–239, 2011. URL: http://dx.doi.org/10.1007/s00453-009-9304-5, doi:10.
1007/s00453-009-9304-5.
Johannes Köbler, Sebastian Kuhnert, Bastian Laubner, and Oleg Verbitsky. Interval graphs:
Canonical representations in logspace. SIAM Journal on Computing, 40(5):1292–1315, 2011.
doi:10.1137/10080395X.
Johannes Köbler, Sebastian Kuhnert, and Oleg Verbitsky. Helly circular-arc graph isomorphism is in logspace. In Mathematical Foundations of Computer Science 2013, volume 8087 of Lecture Notes in Computer Science, pages 631–642. Springer Berlin Heidelberg, 2013. URL: http://dx.doi.org/10.1007/978-3-642-40313-2_56, doi:10.1007/
978-3-642-40313-2_56.
23
24
Canonical Representations for Circular-Arc Graphs Using Flip Sets
11
12
13
14
15
16
17
18
Johannes Köbler, Sebastian Kuhnert, and Oleg Verbitsky.
Solving the canonical
representation and star system problems for proper circular-arc graphs in logspace.
Journal of Discrete Algorithms, 38-41(Supplement C):38 – 49, 2016. URL: http:
//www.sciencedirect.com/science/article/pii/S1570866716300016, doi:https://
doi.org/10.1016/j.jda.2016.03.001.
Min Chih Lin and Jayme L. Szwarcfiter. Characterizations and recognition of circulararc graphs and subclasses: A survey. Discrete Mathematics, 309(18):5618 – 5635,
2009. Combinatorics 2006, A Meeting in Celebration of Pavol Hell’s 60th Birthday (May 1–5, 2006). URL: http://www.sciencedirect.com/science/article/pii/
S0012365X08002161, doi:http://dx.doi.org/10.1016/j.disc.2008.04.003.
George S. Lueker and Kellogg S. Booth. A linear time algorithm for deciding interval graph
isomorphism. J. ACM, 26(2):183–195, April 1979. URL: http://doi.acm.org/10.1145/
322123.322125, doi:10.1145/322123.322125.
Ross M. McConnell. Linear-time recognition of circular-arc graphs. Algorithmica,
37(2):93–147, 2003. URL: http://dx.doi.org/10.1007/s00453-003-1032-7, doi:10.
1007/s00453-003-1032-7.
Franklin W. Stahl. Circular genetic maps. Journal of Cellular Physiology, 70(S1):1–12, 1967.
URL: http://dx.doi.org/10.1002/jcp.1040700403, doi:10.1002/jcp.1040700403.
Alan Tucker. Characterizing circular-arc graphs. Bull. Amer. Math. Soc., 76(6):1257–1260,
11 1970. URL: http://projecteuclid.org/euclid.bams/1183532398.
Michael S. Waterman and Jerrold R. Griggs.
Interval graphs and maps of
dna. Bulletin of Mathematical Biology, 48(2):189 – 195, 1986. URL: http://www.
sciencedirect.com/science/article/pii/S0092824086800064, doi:http://dx.doi.
org/10.1016/S0092-8240(86)80006-4.
Tsong-Ho Wu. An O(n3 ) Isomorphism Test for Circular-Arc Graphs. PhD thesis, SUNY
Stony Brook, New York, NY, USA, 1983.
| 8cs.DS
|
Online Identification of Time-Varying Systems: a Bayesian approach
arXiv:1609.07393v1 [cs.SY] 23 Sep 2016
G. Prando, D. Romeres and A. Chiuso
Abstract— We extend the recently introduced regularization/Bayesian System Identification procedures to the estimation
of time-varying systems. Specifically, we consider an online
setting, in which new data become available at given time steps.
The real-time estimation requirements imposed by this setting
are met by estimating the hyper-parameters through just one
gradient step in the marginal likelihood maximization and by
exploiting the closed-form availability of the impulse response
estimate (when Gaussian prior and Gaussian measurement
noise are postulated). By relying on the use of a forgetting factor,
we propose two methods to tackle the tracking of time-varying
systems. In one of them, the forgetting factor is estimated
by treating it as a hyper-parameter of the Bayesian inference
procedure.
I. I NTRODUCTION
The identification of time-varying systems plays a key role in
different applications, such as adaptive and model predictive
control, where a good real-time tracking of the system to
be controlled is necessary. In addition, the detection of
changes or drifts in plant parameters is crucial in terms
of process monitoring and fault detection. Online System
Identification (SysId) and the estimation of time-varying
systems are typically strictly connected problems: one would
like to exploit the new data that become available in order
to track possible changes in the system dynamics.
Recursive Prediction Error Method (RPEM), a variant of
the classical PEM [1], [2], represents nowadays a wellestablished technique, through which the current estimate can
be efficiently updated, as soon as new data are provided.
RPEM are parametric approaches, relying on Recursive
Least-Squares (RLS) routines, which compute the parameter
estimate by minimizing a functional of the prediction errors
[1]. An extension of these approaches to the identification of
time-varying systems involves the adoption of a forgetting
factor, through which old data become less relevant in the
estimation criterion. Convergence and stability properties of
Forgetting Factor RPEM have been well-studied within the
SysId community [3], [4].
Alternative approaches model the coefficients trajectories as
stochastic processes [5], thus exploiting Kalman filtering [6]
or Bayesian inference [7] for parameter estimation. Combinations of bases sequences (e.g. wavelet basis [8]) have also
been considered to model the parameters time evolution.
The above-mentioned parametric procedures share the criticality of the model selection complexity: this step is especially crucial when the model complexity has to be modified
This work has been partially supported by the FIRB project “Learning
meets time” (RBFR12M3AC) funded by MIUR.
† Dept. of Information Engineering, University of Padova (e-mail:
{romeresd,prandogi,chiuso}@dei.unipd.it)
†
in response to changes in the true system dynamics. In
addition, classical complexity selection rules (e.g. crossvalidation or information criteria) may not be applicable in
online settings, due to the excessive computational effort
they require. Model complexity issues have been partially
addressed in the SysId community through the recent introduction of non-parametric methods, relying on Gaussian
processes and Bayesian inference [9], [10]. In this framework
model complexity is tuned in a continuous manner by
estimating the hyper-parameters which describe the prior
distribution chosen by the user [11]. This property makes
these new techniques appealing for the online identification
of time-varying systems: indeed, model complexity can be
continuously adapted whenever new data become available.
In a previous work [12] we started exploring this research
direction by adapting the newly introduced Bayesian procedures to an online identification setting. The methodologies
proposed in [12] are extended in this new paper by dealing
with time-varying systems. Two approaches, relying on the
use of a forgetting factor, are proposed; in particular, following the approach in [13], we investigate the online estimation
of the forgetting factor by treating it as a hyper-parameter
of the Bayesian inference procedure. These techniques are
experimentally compared with the classical parametric counterparts: the results appear favourable and promising for the
methods we propose.
The paper is organized as follows. Sec. II presents the online
identification framework and the challenges we will try to
address. Sec. III provides a brief review of parametric realtime identification techniques, while Sec. IV illustrates the
Bayesian approach to linear SysId, both in the batch and
online scenarios. In particular, Sec. IV-C focuses on the
estimation of time-varying systems. Experimental results are
reported in Sec. V, while conclusions are drawn in Sec. VI.
II. P ROBLEM F ORMULATION
Consider a dynamical system described through an outputerror model, i.e.:
y(t) = [h ∗ u] (t) + e(t),
y(t), u(t) ∈ R
(1)
where h(t) denotes the model impulse response and e(t) is
assumed to be a zero-mean Gaussian noise with variance σ 2 .
SysId techniques aim at estimating the impulse response h of
N
the system, once a set D := {y(t), u(t)}t=1
of measurements
of its input and output signals is provided.
In this work we consider an online setting, in which a new
set of input-output measurements becomes available every T
time steps. Specifically, let us define the variable i := k/T by
assuming w.l.o.g. that k is a multiple of T , and the ith −dataset
iT
as Di = {u(t), y(t)}t=(i−1)T
+1 .
We suppose that at time k an impulse response estimate ĥ(i)
has been computed using the data coming from a collection
S
iT
of previous datasets il=1 Dl = {u(t), y(t)}t=1
; at time k + T
new data Di+1 become available and we would like to update
the previous estimate ĥ(i) by exploiting them. In addition
we assume that the underlying system undergoes certain
variations that we would like to track: this situation could
often arise in practice, due to e.g. variations of the internal
temperature, of the masses (e.g. after grasping an object).
Furthermore, online applications typically require that the
new estimate is available before the new dataset Di+2 is
provided, thus limiting the computational complexity and the
memory storage of the adopted estimation methods.
In this paper, the recently proposed Bayesian approach to
SysId [10] is adapted in order to cope with the outlined
online setting. Its performances are compared with the ones
achieved using classical parametric approaches.
Remark 1: We stress that in the remainder of the paper
we will use the indexes k and iT interchangeably.
III. PARAMETRIC A PPROACH
Standard parametric approaches to SysId rely on the a-priori
choice of a model class M (e.g. ARX, ARMAX, OE, etc.),
which is completely characterized by a parameter θ ∈ Rm .
A. Batch Approach
N
In the batch setting, when a dataset D = {y(t), u(t)}t=1
is
provided, the identification procedure reduces to estimate θ
by minimizing the sum of squared prediction errors:
1 N
θ̂ = arg minm VN (θ , D) = arg minm ∑ (y(t) − ŷ(t|θ ))2 (2)
θ ∈R
θ ∈R 2 t=1
where ŷ(t|θ ) denotes the one-step ahead predictor [1].
B. Online Approach
The extension of these procedures to an online setting relies
on RLS (or pseudo LS) methods.
For ease of notation, let us assume T = 1 in this section.
Suppose that at time k + 1 a new input-output data pair Di+1
is provided; then θ̂ (i) is updated as:
−1
θ̂ (i+1) = θ̂ (i) + µ (i+1) Q(i+1) ∇θ Vk+1 (θ̂ (i) ,
l=1 Dl )
Si+1
(3)
where
Dl ) denotes the gradient of the
loss function computed in the previous estimate and in the
new data; µ (i+1) ∈ R and Q(i+1) ∈ Rm×m are appropriate
scalings which assume different shapes according to the
specific algorithm which is adopted (see [2] and [1], Ch.
11, for further details). Notice that (3) is simply a scaled
S
gradient step w.r.t. the loss function Vk+1 (θ , i+1
l=1 Dl ).
S
∇θ Vk+1 (θ̂ (i) , i+1
l=1
C. Dealing with time-varying systems
In order to cope with time-varying systems, a possible
strategy involves the inclusion of a forgetting factor γ̄ in
the loss function Vk (θ , D):
Vk (θ , D) =
γ
1 k k−t
∑ γ̄ (y(t) − ŷ(t|θ ))2 ,
2 t=1
γ̄ ∈ (0, 1]
(4)
In this way old measurements become less relevant for
the computation of the estimate. A recursive update of the
estimate θ̂ (i) (as the one in (3)) can be derived ([1], Ch. 11).
As an alternative, a sliding window approach can be adopted:
at each time step only the last Nw data are used for computing
the current estimate (with Nw being the window length).
However, since this approach does not admit an update rule
as the one in (3), the computational complexity of the new
estimate will depend on the window length.
A crucial role in the application of parametric SysId techniques is played by the model order selection step: once a
model class M is fixed, its complexity has to be chosen
using the available data. This is typically accomplished
by estimating models with different complexities and by
applying tools such as cross-validation or information criteria
to select the most appropriate one. However, the estimation of
multiple models may be computationally expensive, making
this procedure not suited for the online identification of timevarying systems. Indeed, in this framework, it should ideally
be applied every time new data become available.
The recently proposed approach to SysId, relying on regularization/Bayesian techniques, overpasses the above-described
issue by jointly performing estimation and order selection. Next section will illustrate how the batch regularization/Bayesian method can be tailored to the online identification of time-varying systems.
IV. R EGULARIZATION /BAYESIAN A PPROACH
A. Batch Approach
We discuss how the regularization/Bayesian technique
works in the standard batch setting, i.e. when data D =
N
{y(t), u(t)}t=1
are given. For future use, let us define the
vector YN = [y(1) ... y(N)]> ∈ RN .
According to the Bayesian estimation, the impulse response
h is considered as a realization of a stochastic process with
a prior distribution pη (h), depending on some parameters
η ∈ Ω. The prior pη (h) is designed in order to account for
some desired properties of the estimated impulse response,
such as smoothness and stability [9], [10]. In the Bayesian
framework, the parameters η are known as hyper-parameters
and they need to be estimated from the data, e.g. by optimizing the so-called marginal likelihood (i.e. the likelihood
once the latent variable h has been integrated out) [11]:
Z
η̂ = arg max pη (YN ) = arg max
η∈Ω
η∈Ω
p(YN |h)pη (h)dh
(5)
Once the hyper-parameters η have been estimated, the
minimum variance estimate of h needs to be computed; it
coincides with the posterior mean given the observed data:
Z
ĥ := Eη̂ [h|YN ] =
h
p(YN |h)pη̂ (h)
dh
pη̂ (YN )
(6)
In the SysId context, h is typically modelled as a zeromean Gaussian process (independent of the noise e(t)) with
covariance E [h(t), h(s)] = K̄η (t, s) (aka kernel in the Machine
Learning literature) [9], [14]. Thanks to this asssumption, the
marginal likelihood pη (YN ) is Gaussian and the estimate (6)
is available in closed form.
Furthermore, for simplicity the IIR model in (1) can be accurately approximated by a FIR model of order n, whenever n is
chosen large enough to catch the relevant components of the
system dynamics. By collecting in h := [h(1) · · · h(n)]> ∈
Rn the first n impulse response coefficients, the following
Gaussian prior can be defined:
pη (h) ∼ N (0, Kη ),
η ∈ Ω ⊂ Rd , Kη ∈ Rn×n
(7)
as 1st or 2nd order optimization algorithms [15] or the
Expectation-Maximization (EM) algorithm [16], [17]. Since
these methods may require a large number of iterations
before reaching convergence, they may be unsuited for
online applications. We should recall that, when adopted
for marginal likelihood optimization, each iteration of these
algorithms has a computational complexity of O(n3 ), due to
the objective function evaluation. Specifically, f(i+1)T (η) can
be robustly evaluated as [18]
The hyper-parameters η can then be estimated by solving
η̂ = arg min − ln pη (YN ) = arg min fN (η)
η∈Ω
u(−1)
..
.
u(N) u(N − 1)
(8)
···
..
.
···
u(−n + 1)
..
.
u(N − n + 1)
(9)
(10)
(11)
In the batch setting we are considering the quantities
u(−n + 1), ..., u(0) can be either estimated or set to zero.
Here, we follow the latter option. Once η̂ has been computed,
the corresponding minimum variance estimate is given by
b : = Eη̂ [h|YN ] = arg min (YN − ΦN h)> (YN − ΦN h) + σ 2 h> K −1 h
h
η̂
∈ n
h R
2 −1 −1 >
= (Φ>
N ΦN + σ Kη̂ ) ΦN YN
+ σ −2 ( Y
η∈Ω
fN (η) = YN> Σ(η)−1YN + ln det Σ(η)
2
Σ(η) = ΦN Kη Φ>
N + σ IN
where ΦN ∈ RN×n :
u(0)
ΦN := ...
f(i+1)T (η) =((i + 1)T − n) ln σ 2 + 2 ln |S|
(i+1)
− kS−1 L>Ye (i+1) k22 )
(16)
where L and S are Cholesky factors: Kη =: LL> and σ 2 In +
L> R(i+1) L =: SS> (whose computation is O(n3 )).
To tackle the real-time constraints, the approach proposed
in [12] is adopted: η̂ (i+1) is computed by running just one
iteration of a Scaled Gradient Projection (SGP) algorithm
(a 1st order optimization method) applied to solve problem
(8) [15]. Algorithm 1 summarizes its implementation. Notice
that it is initialized
with the previous estimate η̂ (i) (obtained
Si
using the data l=1 Dl ) which is likely to be close to a local
optimum of the objective function fiT (η) ≡ fk (η). If the
number of new data T << k, it is reasonable to suppose
that arg minη∈Ω fiT (η) ≈ arg minη∈Ω f(i+1)T (η). Therefore,
by just performing one SGP iteration, η̂ (i+1) will be sufficiently close to a local optimum of f(i+1)T (η).
(12)
b in (12) can be computed once
Remark 2: The estimate h
a noise variance estimate σ̂ 2 is available. For this purpose,
σ 2 can be treated as a hyper-parameter and estimated by
solving (8) or it can be computed from a LS estimate of h.
In this work the latter option is adopted.
B. Online Approach
We now adapt the batch technique described in Sec. IVA to the online setting outlined in Sec. II. At time k + T ,
(i+1)T
when data Di+1 = {u(t), y(t)}t=iT +1 are provided, the current
b(i) is updated through formula
impulse response estimate h
(12), once the data matrices are enlarged with the new data
and a new hyper-parameter estimate η̂ (i+1) is computed. The
data matrices are updated through the following recursions
(i+1)T > (i+1)T
(i)
ΦiT +1 (13)
R(i+1) := Φ>
(i+1)T Φ(i+1)T = R + ΦiT +1
>
(i+1)T
(i+1)T
e (i)
Ye (i+1) := Φ>
YiT +1
(14)
(i+1)T Y(i+1)T = Y + ΦiT +1
(i+1)
(i)
(i+1)T > (i+1)T
>
Y
:= Y(i+1)T
Y(i+1)T =Y + YiT +1
YiT +1
(15)
(i+1)T
where Y(i+1)T = [y(1) · · · y(iT + T )]> ∈ R(i+1)T , YiT +1 =
[y(iT + 1) · · · y(iT + T )]; Φi is defined as in (11) with N
(i+1)T
replaced by (i + 1)T , while ΦiT +1 has the same structure
of matrix (11) but it contains the data from iT − n + 1 to
(i + 1)T . The computational cost of (13)-(15) is, O(n2 T ),
O(nT ) and O(T 2 ), respectively.
The minimization of f(i+1)T (η) in (9), needed to determine
η̂ (i+1) , is typically performed through iterative routines, such
Algorithm 1 1-step Scaled Gradient Projection (SGP)
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
Inputs: previous estimates {η̂ (i) , η̂ (i−1) }, ∇ fiT (η̂ (i−1) ),
(i+1)
2
R(i+1) , Ye (i+1) , Y
, σ̂ (i+1)
Initialize: c = 10−4 , δ = 0.4
Compute ∇ f(i+1)T (η̂ (i) )
r(i−1) ← η̂ (i) − η̂ (i−1)
w(i−1) ← ∇ f(i+1)T (η̂ (i) ) − ∇ fiT (η̂ (i−1) )
Approximate the inverse Hessian of f(i+1)T (η̂ (i) ) as
B(i) = α (i) D(i) (using the procedure outlined in [15])
Project onto the feasible set:
z ← ΠΩ,D(i) ( η̂ (i) − B(i) ∇ f(i+1)T (η̂ (i) ) )
∆η̂ (i) ← z − η̂ (i)
ν ←1
if
f(i+1)T (η̂ (i) + ν∆η̂ (i) ) ≤ f(i+1)T (η̂ (i) ) +
cν∇ f(i+1)T (η̂ (i) )> ∆η̂ (i) then
Go to step 12
else
ν ← δν
η̂ (i+1) ← η̂ (i) + ν∆η̂ (i)
Output: η̂ (i+1)
The key step in Algorithm 1 is 4, where the inverse Hessian
is approximated as the product between the positive scalar
α (i) ∈ R+ and the diagonal matrix D(i) ∈ Rd×d . α (i) is chosen
by alternating the so-called Barzilai-Borwein (BB) rules [19]:
>
(i)
α1 :=
r(i−1) r(i−1)
,
>
r(i−1) w(i−1)
>
(i)
α2 :=
r(i−1) w(i−1)
>
w(i−1) w(i−1)
(17)
with r(i−1) and w(i−1) specified at steps 2 and 3 of Algorithm
1. The definition of D(i) depends on the constraints set and
on the objective function. The authors in [15] exploit the
following decomposition of ∇η f(i+1)T (η) (defined in (9)):
∇η f(i+1)T (η) = V (η) −U(η),
V (η) > 0, U(η) ≥ 0 (18)
to specify D(i) . We refer the interested reader to [15] for
further details.
The projection operator adopted at step 5 of Algorithm 1 is
ΠΩ,D(i) (z) = arg minx∈Ω
−1
(x − z)> D(i) (x − z)
In this section we deal with the identification of timevarying systems: specifically, estimators have to be equipped
with tools through which past data become less relevant for
the current estimation. In the following we propose two
routines which combine the “online Bayesian estimation”
above sketched with the ability to “forget” past data.
1) Fixed Forgetting Factor: Following a classical practice
in parametric SysId (see Sec. III), we introduce a forgetting
factor γ̄ ∈ (0, 1] into the data we are provided in order to base
the estimation mainly on the more recent data. Specifically,
we assume that the first k data are generated according to
the following linear model:
ḠkYk = Ḡk Φk h + E, E = [e(1)...e(k)]> ∼ N (0, σ 2 Ik ) (20)
where Ḡk Ḡk =: Γ̄k and Γ̄k := diag γ̄ k−1 , γ̄ k−2 , ..., γ̄ 0 . Therefore, when adopting the regularized regression criterion (12),
the estimate at time k is computed as:
(21)
= arg minn (Yk − Φk h)> Γ̄k (Yk − Φk h) + σ 2 h> Kη̂−1 h
h∈R
>
= (Φk Γ̄k Φk + σ 2 Kη̂−1 )−1 Φ>
k Γ̄kYk
(22)
Correspondingly, the hyper-parameters are estimated solving:
η̂ = arg min Yk> Ḡk Σγ̄ (η)−1 ḠkYk + ln det Σγ̄ (η)
(23)
η∈Ω
2
Σγ̄ (η) = Ḡk Φk Kη Φ>
k Ḡk + σ Ik
Algorithm 2
identification
In particular,
and η̂ (i) are
Inputs: forgetting factor γ̄, previous estimates
(i) (i) (i)
{η̂ (i) , η̂ (i−1) }, previous data matrices {Rγ̄ , Yeγ̄ ,Yγ̄ },
(i+1)T
C. Dealing with time-varying systems
h∈R t=1
Algorithm 2 Online Bayesian SysId: Fixed Forgetting Factor
(19)
Remark 3: Besides SGP, in [12] other inverse Hessian
approximations are investigated (e.g. the BFGS formula).
In this work we only consider the SGP approximation,
since it appears preferable to the others, according to the
experiments we performed (both in the time-invariant and
-variant domain). [12] also considers the EM algorithm as
an alternative to 1st order optimization methods to solve
problem (8). Even if the results reported for EM in [12]
are comparable to the ones achieved through SGP, the latter
approach appears superior to EM in the time-varying setting
we are considering.
k
bγ̄ := arg min ∑ γ̄ k−t y(t) − Φtt h 2 + σ 2 h> K −1 h
h
η̂
n
solving, respectively, (21) and (23); these estimates are
then online updated after the new data Di+1 are provided. Once γ̄ is chosen by the user, it is inserted in
(i+1)
e
the data matrices Rγ̄
:= Φ>
(i+1)T Γ̄(i+1)T Φ(i+1)T , Yγ̄ (i+1) :=
>
>
Φ(i+1)T Γ̄(i+1)T Y(i+1)T , Yγ̄ (i+1) := Y(i+1)T Γ̄(i+1)T Y(i+1)T , updated at steps 1-3 of the algorithm.
(24)
illustrates the online implementation of the
procedure based on equations (22) and (23).
b(i)
it assumes that at time k the estimates h
available and they have been computed by
new data Di+1 = {u(t), y(t)}t=iT +1
(i+1)
(i)
(i+1)T >
(i+1)T
1: Rγ̄
← γ̄ T Rγ̄ + ΦiT +1
Γ̄T ΦiT +1
>
e (i+1) ← γ T Yeγ(i) + Φ(i+1)T Γ̄T Y (i+1)T
2: Y
γ̄
iT +1
iT +1
>
(i+1)
(i)
(i+1)T
(i+1)T
← γ̄ TYγ̄ + YiT +1
Γ̄T YiT +1
3: Y γ̄
−1
4:
b(i+1) ← R(i+1) Ye (i+1)
h
γ̄
γ̄
LS
2
5: σ̂ (i+1) ←
1
(i+1)T −n
(i+1)
(i+1)> b(i+1) b(i+1)> (i+1) b(i+1)
Ȳγ̄
− 2Yeγ̄
hLS + hLS
Rγ̄
hLS
η̂ (i+1) ← arg minη∈Ω f(i+1)T (η) (use Algorithm 1)
−1
2
(i+1)
b(i+1) ← R(i+1) + σ̂ (i+1) K −1
7: h
Yeγ̄
(i+1)
γ̄
γ̄
η̂
b(i+1) , η̂ (i+1)
Output: h
6:
2) Treating the Forgetting Factor as a Hyper-parameter:
The Bayesian framework provides the user with the possibility to treat the forgetting factor as a hyper-parameter and
to estimate it by solving:
η̂, γ̂ = arg minη∈Ω,γ∈(0,1] fk (η, γ)
fk (η, γ) = Yk> Gk Σ(η, γ)−1 GkYk + ln det Σ(η, γ)
2
Σ(η, γ) = Gk Φk Kη Φ>
k Gk + σ Ik
(25)
(26)
(27)
where Gk Gk =: Γk and Γk := diag
.
Remark 4: Notice that the model (20) is equivalent to
>
Yk = Φk h + Eγ̄ , Eγ̄ = eγ̄ (1), ..., eγ̄ (k) ∼ N (0, σ 2 Γ̄−1
k )
γ k−1 , γ k−2 , ..., γ 0
Therefore, treating the forgetting factor as a hyper-parameter
is equivalent to modeling the noise with a non-constant
variance and to give to the diagonal entries of the covariance
matrix an exponential decaying structure.
The online implementation of this approach is detailed in
Algorithm 3, where
>
(i)
(i−1)
b(i) ΦiT
Rγ̂γ := γ̂ (i) Rγ̂γ
+ ΦiT
Γ
(28)
T
(i−1)T +1
(i−1)T +1
b(i) = diag((γ̂ (i) )T −1 , .., (γ̂ (i) )0 ). Ye (i) and Yγ̂γ(i) are analowith Γ
T
γ̂γ
gously defined.
We should stress that the objective function in (26) does not
admit the decomposition (18); we have
∂ fk (η, γ)
= V (η, γ) +U(η, γ),
∂γ
V (η, γ) > 0, U(η, γ) ≥ 0
Thus, when γ is treated as an hyper-parameter, Algorithm
1 is run setting D(i) = Id at step 4; α (i) is still determined
alternating the BB rules (17).
Algorithm 3 Online Bayesian SysId: Forgetting Factor as a
hyper-parameter
Inputs: previous estimates {η̂ (i) , η̂ (i−1) , γ̂ (i) , γ̂ (i−1) }, pre(i) (i) (i)
vious data matrices {Rγ̂γ , Yeγ̂γ ,Yγ̂γ }, new data Di+1 =
(i+1)T
{u(t), y(t)}t=iT +1
(i)
(i+1)T >
(i+1)T
← γ T Rγ̂γ + ΦiT +1
ΓT ΦiT +1
>
eγ(i+1) ← γ T Ye (i) + Φ(i+1)T ΓT Y (i+1)T
2: Y
iT +1
iT +1
γ̂γ
(i+1)
(i)
(i+1)T >
(i+1)T
T
← γ Yγ̂γ + YiT +1
ΓT YiT +1
3: Y γ
(i+1)
(i)
(i)
−1
b
4: h
← (R ) Ye
1:
(i+1)
Rγ
LS
5: σ̂ 2
(i+1)
γ̂γ
←
1
(i+1)T −n
(i)γ̂γ
(i)
b(i+1)
b(i+1) + (h
b(i+1) )> R(i) h
Yγ̂γ − 2(Yeγ̂γ )> h
LS
LS
γ̂γ LS
η̂ (i+1) , γ̂ (i+1) ← arg minη∈Ω,γ∈(0,1] f(i+1)T (η, γ)
(use Algorithm
1)
−1
(i+1)
b(i+1) ← R(i+1) + σ̂ 2(i+1) (γ̂ (i+1) ) K −1
7: h
Yeγ̂γ
(i+1)
γ̂γ
η̂
b(i+1) , η̂ (i+1)
Output: h
6:
V. E XPERIMENTAL R ESULTS
In this section we test the online algorithms for parametric
and Bayesian SysId described in Sec. III and IV. Their
performance are compared through a Monte-Carlo study over
200 time-varying systems.
A. Data
200 datasets consisting of 3000 input-output measurement
pairs are generated. Each of them is created as follows:
the first 1000 data are produced by a system contained in
the data-bank D4 (used in [20]), while the remaining 2000
data are generated by perturbing the D4-system with two
additional poles and zeros. These are chosen such that the
order of the D4-system changes, thus creating a switch on
the data generating system at time k = 1001.
The data-bank D4 consists of 30th order random SISO
dicrete-time systems having all the poles inside a circle of
radius 0.95. These systems are simulated with a unit variance
band-limited Gaussian signal with normalized band [0, 0.8].
A zero mean white Gaussian noise, with variance adjusted
so that the Signal to Noise Ration (SNR) is always equal to
1, is then added to the output data.
B. Estimators
The parametric estimators are computed with the roe Matlab routine, using the BIC criterion for the model complexity
selection. In the following this estimator will be denoted as
PEM BIC. Furthermore, as a benchmark we introduce the
parametric oracle estimator, called PEM OR, which selects
the model complexity by choosing the model that gives the
best fit to the impulse response of the true system. The order
selection is performed every time a new dataset becomes
available: multiple models with orders ranging from 1 to 20
are estimated and the order selection is performed according
to the two criteria above-described.
For what regards the methods relying on Bayesian inference
we adopt a zero-mean Gaussian prior with a covariance
matrix (kernel) given by the so-called TC-kernel:
KηTC (k, j) = λ min(β k , β j ),
η = [λ , β ]
(29)
with Ω = {(λ , β ) : λ ≥ 0, 0 ≤ β ≤ 1} [14]. The length n
of the estimated impulse responses is set to 100. In the
following, we will use the acronym TC to denote these
methods. Furthermore, the notation OPT will refer to the
standard Bayesian procedure, in which the SGP algorithm
adopted to optimize the marginal likelihood fk (η) is run until
the relative change in fk (η) is less than 10−9 . From here on,
the online counterpart (illustrated in Sec. IV) will be referred
to as the 1-step ML. We will also use the acronyms TC FF
when a fixed forgetting factor is adopted, TC est FF when
the forgetting factor is estimated as a hyper-parameter.
For each Monte Carlo run, the identification algorithms are
300
initialized using the first batch of data Dinit = {u(t), y(t)}t=1
.
After this initial step, the estimators are updated every T = 10
(i+1)T
time steps, when new data Di+1 = {u(t), y(t)}t=iT are
provided. The forgetting factor in the TC FF and PEM
methods is set to 0.998, while its estimation in TC est FF
method is initialized with 0.995.
C. Performance
The purpose of the experiments is twofold. First, we will
compare the two routines we have proposed in Sec. IV-C
to explicitly deal with time-varying systems. Second, we
will compare the parametric and the Bayesian identification
approaches while dealing with time-varying systems.
As a first comparison, we evaluate the adherence of the
b to the true one h, measured as
estimated impulse response h
b
b = 100 · 1 − kh − hk2
(30)
F (h)
khk2
b at four time instants.
Figure 1 reports the values of F (h)
It is interesting to note that immediately before the change
in the data generating system (k = 1000) the TC methods
slightly outperform the ideal parametric estimator PEM OR.
After the switch (occurring at k = 1001), among regularization/Bayesian routines TC est FF recovers the fit performance a bit faster than TC FF; even at regime it outperforms
the latter because it can choose forgetting factor values that
retain a larger amount of data.
We also observe how the 1-step ML procedures and the
corresponding OPT routines provide analogous performance
at each time step k, validating the method we propose for
online estimation and confirming the results in [12].
The unrealistic PEM OR represents the reference on the
achievable performance of the PEM estimators; it outperforms TC methods in the transient after the switch, while it
has comparable performance at regime. Instead, the recursive
PEM BIC estimator performs very poorly.
As a second comparison, Table I reports the computational
cumulative time of the proposed algorithms in terms of mean
and standard deviation after the estimators are fed with all
3000
the data D = {u(t), y(t)}t=1
. The 1-step ML methods are
one order of magnitude faster than the corresponding OPT
k = 1000
100
update of the Bayesian estimate, resembling the one which
is available for parametric techniques.
b
F (h)
80
R EFERENCES
60
40
TC OPT FF
TC FF
TC est FF
PEM OR
PEM BIC
PEM OR
PEM BIC
PEM OR
PEM BIC
PEM OR
PEM BIC
k = 1300
100
b
F (h)
80
60
40
TC OPT FF
TC FF
TC est FF
k = 1800
100
b
F (h)
80
60
40
TC OPT FF
TC FF
b
F (h)
TC est FF
k = 3000
100
80
60
40
TC OPT FF
TC FF
TC est FF
b achieved at four time instants k (correFig. 1: Fit F (h)
sponding to the number of data available for the estimation).
mean
std
OPT FF
6.70
1.28
TC
FF
0.44
0.03
est FF
5.45
0.67
OR
18.44
0.69
PEM
BIC
18.44
0.69
TABLE I: Computational cumulative time after data D =
3000
{u(t), y(t)}t=1
are used: mean and std over 200 datasets.
ones. The TC est FF estimator is slower than TC FF: this
should be a consequence of having set D(i) = Id in Algorithm
1. On the other hand the RPEM estimators are three times
slower than the OPT ones, thus appearing not particularly
appealing for online applications.
VI. C ONCLUSION AND F UTURE W ORK
We have adopted recently developed SysId techniques relying on the Gaussian processes and Bayesian inference to
the identification of time-varying systems. Specifically, we
have focused on an online setting by assuming that new data
become available at predefined time instants. To tackle the
real-time constraints we have modified the standard Bayesian
procedure: hyper-parameters are estimated by performing
only one gradient step in the corresponding marginal likelihood optimization problem. In order to cope with the timevarying nature of the systems to be identified, we propose
two approaches, based on the use of a forgetting factor.
One of them treats the forgetting factor as a constant, while
the other estimates it as a hyper-parameter of the Bayesian
inference procedure.
We believe that the preliminary investigation performed in
this work may pave the way for further research in this
topic. A future research direction could consider the recursive
[1] L. Ljung, System Identification - Theory for the User, 2nd ed. Upper
Saddle River, N.J.: Prentice-Hall, 1999.
[2] L. Ljung and T. Söderström, Theory and Practice of Recursive
Identificationn, ser. Signal Processing, Optimization, and Control. The
MIT Press, 1983.
[3] S. Bittanti, P. Bolzern, and M. Campi, “Convergence and exponential
convergence of identification algorithms with directional forgetting
factor,” Automatica, vol. 26, no. 5, pp. 929–932, 1990.
[4] L. Guo, L. Ljung, and P. Priouret, “Performance analysis of the
forgetting factor rls algorithm,” International journal of adaptive
control and signal processing, vol. 7, no. 6, pp. 525–538, 1993.
[5] G. C. Chow, “Random and changing coefficient models,” in Handbook
of Econometrics, 1st ed. Elsevier, 1984, vol. 2, ch. 21, pp. 1213–1245.
[6] L. Guo, “Estimating time-varying parameters by the kalman filter
based algorithm: stability and convergence,” Automatic Control, IEEE
Transactions on, vol. 35, no. 2, pp. 141–147, 1990.
[7] A. H. Sarris, A Bayesian approach to estimation of time-varying
regression coefficients. National Bureau of Economic Research, 1973.
[8] M. K. Tsatsanis and G. B. Giannakis, “Time-varying system identification and model validation using wavelets,” Signal Processing, IEEE
Transactions on, vol. 41, no. 12, pp. 3512–3523, 1993.
[9] G. Pillonetto, A. Chiuso, and G. De Nicolao, “Prediction error
identification of linear systems: a nonparametric Gaussian regression
approach,” Automatica, no. 47, pp. 291–305, 2011.
[10] G. Pillonetto, F. Dinuzzo, T. Chen, G. De Nicolao, and L. Ljung, “Kernel methods in system identification, machine learning and function
estimation: a survey,” Automatica, vol. 50, no. 3, pp. 657–682, 2014.
[11] G. Pillonetto and A. Chiuso, “Tuning complexity in kernel-based
linear system identification: the robustness of the marginal likelihood
estimator,” Automatica, vol. 58, pp. 106–117, 2015.
[12] D. Romeres, G. Prando, G. Pillonetto, and A. Chiuso, “On-line
bayesian system identification,” in Proc. of ECC 2016, 2016.
[13] G. Pillonetto and A. Aravkin, “A new kernel-based approach for identification of time-varying linear systems,” in 2014 IEEE International
Workshop on Machine Learning for Signal Processing (MLSP), Sept
2014, pp. 1–6.
[14] T. Chen, H. Ohlsson, and L. Ljung, “On the estimation of transfer
functions, regularizations and Gaussian processes - revisited,” Automatica, vol. 48, no. 8, pp. 1525–1535, 2012.
[15] S. Bonettini, A. Chiuso, and M. Prato, “A scaled gradient projection
methods for Bayesian learning in dynamical systems,” SIAM Journal
on Scientific Computing, p. in press, 2015.
[16] G. Bottegal, A. Y. Aravkin, H. Hjalmarsson, and G. Pillonetto, “Robust
em kernel-based methods for linear system identification,” Automatica,
vol. 67, pp. 114–126, 2016.
[17] G. Bottegal, R. S. Risuleo, and H. Hjalmarsson, “Blind system identification using kernel-based methods,” IFAC-PapersOnLine, vol. 48,
no. 28, pp. 466 – 471, 2015.
[18] T. Chen and L. Ljung, “Implementation of algorithms for tuning
parameters in regularized least squares problems in system identification,” Automatica, vol. 49, no. 7, pp. 2213–2220, 2013.
[19] J. Barzilai and J. M. Borwein, “Two-point step size gradient methods,”
IMA Journal of Numerical Analysis, vol. 8, no. 1, pp. 141–148, 1988.
[20] T. Chen, M. Andersen, L. Ljung, A. Chiuso, and G. Pillonetto, “System
identification via sparse multiple kernel-based regularization using
sequential convex optimization techniques,” IEEE Transactions on
Automatic Control, 2014.
| 3cs.SY
|
A linear programming based heuristic framework for min-max
regret combinatorial optimization problems with interval costs
Lucas Assunçãoa,∗, Thiago F. Noronhab, Andréa Cynthia Santosc , Rafael Andraded
arXiv:1609.09179v1 [cs.DS] 29 Sep 2016
a
Departamento de Engenharia de Produção, Universidade Federal de Minas Gerais
Avenida Antônio Carlos 6627, CEP 31270-901, Belo Horizonte, MG, Brazil
b Departamento de Ciência da Computação, Universidade Federal de Minas Gerais
Avenida Antônio Carlos 6627, CEP 31270-901, Belo Horizonte, MG, Brazil
c ICD-LOSI, UMR CNRS 6281, Université de Technologie de Troyes
12, rue Marie Curie, CS 42060, 10004, Troyes CEDEX, France
d
Departamento de Estatı́stica e Matemática Aplicada, Universidade Federal do Ceará
Campus do Pici - Bloco 910, 60455-900, Fortaleza, CE, Brazil
Abstract
This work deals with a class of problems under interval data uncertainty, namely interval robusthard problems, composed of interval data min-max regret generalizations of classical NP-hard
combinatorial problems modeled as 0-1 integer linear programming problems. These problems
are more challenging than other interval data min-max regret problems, as solely computing the
cost of any feasible solution requires solving an instance of an NP-hard problem. The state-ofthe-art exact algorithms in the literature are based on the generation of a possibly exponential
number of cuts. As each cut separation involves the resolution of an NP-hard classical optimization problem, the size of the instances that can be solved efficiently is relatively small. To smooth
this issue, we present a modeling technique for interval robust-hard problems in the context of a
heuristic framework. The heuristic obtains feasible solutions by exploring dual information of a
linearly relaxed model associated with the classical optimization problem counterpart. Computational experiments for interval data min-max regret versions of the restricted shortest path problem and the set covering problem show that our heuristic is able to find optimal or near-optimal
solutions and also improves the primal bounds obtained by a state-of-the-art exact algorithm and
a 2-approximation procedure for interval data min-max regret problems.
Keywords: Robust optimization, Matheuristics, Benders’ decomposition
∗ Corresponding
author.
Email addresses: [email protected] (Lucas Assunção), [email protected] (Thiago F. Noronha),
[email protected] (Andréa Cynthia Santos), [email protected] (Rafael Andrade)
Preprint submitted to arXiv.org
September 30, 2016
1. Introduction
Robust optimization [1] is an alternative to stochastic programming [2] in which the variability of the data is represented by deterministic values in the context of scenarios. A scenario
corresponds to a parameters assignment, i.e., a value is fixed for each parameter subject to uncertainty. The two main approaches adopted to model robust optimization problems are the discrete
scenarios model and the interval data model. In the former, a discrete set of possible scenarios
is considered. In the latter, the uncertainty referred to a parameter is represented by a continuous interval of possible values. Differently from the discrete scenarios model, the infinite many
possible scenarios that arise in the interval data model are not explicitly given. Nevertheless, in
both models, a classical (i.e., parameters known in advance) optimization problem takes place
whenever a scenario is established.
With respect to robust optimization criteria, the min-max regret (also known as robust deviation criterion) is one of the most used in the literature (see, e.g., [3, 4, 5, 6]). The regret of a
solution in a given scenario is defined as the cost difference between such solution and an optimal one in this scenario. In turn, the robustness cost of a solution is defined as its maximum
regret over all scenarios. Min-max regret problems aim at finding a solution with the minimum
robustness cost, which is referred to as a robust solution. We refer to [1] for details on other
robust optimization criteria.
Robust optimization versions of several combinatorial optimization problems have been studied in the literature, addressing, for example, uncertainties over costs. Handling uncertain costs
brings an extra level of difficulty, such that even polynomially solvable problems become NPhard in their corresponding robust versions [4, 6, 7, 8, 9]. A recent trend in this field is to investigate robust optimization problems whose classical counterparts are already NP-hard. We refer
to these problems as robust-hard problems. The interval data min-max regret restricted shortest
path problem, introduced in this study, belongs to this class of problems, along with interval data
min-max regret versions of the traveling salesman problem [5], the 0-1 knapsack problem [10]
and the set covering problem [11].
This work addresses interval min-max regret problems, which consist of min-max regret versions of combinatorial problems with interval costs. Precisely, we consider interval data min-max
regret generalizations of classical combinatorial problems modeled by means of 0-1 Integer Linear Programming (ILP). Notice that a large variety of combinatorial problems are modeled as 0-1
2
ILP problems, including (i) polynomially solvable problems, such as the shortest path problem,
the minimum spanning tree problem and the assignment problem, and (ii) NP-hard problems,
such as the 0-1 knapsack problem, the set covering problem, the traveling salesman problem and
the restricted shortest path problem [12, 13]. In this study, we are particularly interested in a
subclass of interval min-max regret problems, namely interval robust-hard problems, composed
of interval data min-max regret versions of classical NP-hard combinatorial problems as those
aforementioned in (ii).
Aissi at al. [14] (also see [15]) showed that, for any interval min-max regret problem (including
interval robust-hard problems), the robustness cost of a solution can be computed by solving a
single instance of the classical optimization problem counterpart in a particular scenario. Therefore, one does not have to consider all the infinite many possible scenarios during the search for
a robust solution, but only a subset of them, one for each feasible solution. Notice, however, that,
in the case of interval robust-hard problems, computing the cost of a solution still involves solving an NP-hard problem. Thus, these problems are more challenging than other interval min-max
regret problems.
The state-of-the-art exact algorithms used to solve these problems are based on the generation
of a possibly exponential number of cuts (see, e.g., [5, 11]). Since each cut separation implies
solving an instance of an NP-hard problem (which corresponds to the classical counterpart of the
robust optimization problem considered), the size of the instances efficiently solvable is considerably smaller than that of the instances solved for the corresponding classical counterparts. In
fact, to the best of our knowledge, no work in the literature provides a compact ILP formulation
(with a polynomial number of variables and constraints) for any interval robust-hard problem.
Although we do not close the aforementioned issues in this work, we smooth them by presenting an alternative way to handle interval robust-hard problem. We propose a Linear Programming
(LP) based heuristic framework inspired by the modeling technique introduced in [7, 16] and formalized by Kasperski [9]. The heuristic framework is suitable to tackle interval min-max regret
problems in general and consists in solving a Mixed Integer Linear Programming (MILP) model
based on the dual of a linearly relaxed formulation for the classical optimization problem counterpart. An optimal solution for the model solved by our heuristic is not necessarily optimal for
the original robust optimization problem. However, computational experiments for interval data
min-max regret versions of the restricted shortest path problem and the set covering problem [11]
3
showed that the heuristic is able to find optimal or near-optimal solutions, and that it improves
the upper bounds obtained by both a state-of-the-art exact algorithm and a 2-approximation procedure for interval min-max regret problems.
The theoretical quality of the solutions obtained by the heuristic in general is beyond the scope
of this work, as it may rely on specific problem dependent factors, such as the strength of the
formulation adopted to model the classical counterpart problem within the heuristic framework.
In fact, as we will discuss in the sequel, the study of approximative procedures for interval minmax regret problems is a relatively unexplored field, and not much is known.
The remainder of this work is organized as follows. Related works on robust-hard problems are
discussed in Section 2. In Section 3, we present a standard modeling technique for interval minmax regret problems and a state-of-the-art logic-based Benders’ framework for solving problems
in this class. In this study, the solutions obtained by this exact approach are used to evaluate the
quality of the solutions found by the proposed LP based heuristic framework, which is presented
in Section 4. In Section 5, we formally describe the interval robust-hard problems used as case
studies for the heuristic framework. In this same section, we introduce the interval data minmax regret restricted shortest path problem, along with a brief motivation and related works.
Computational experiments are showed in Section 6, while concluding remarks and future work
directions are discussed in the last section.
2. Related works
As far as we know, Montemanni et al. [5] (also see [17]) were the first to address an interval robust-hard problem. The authors introduced the interval data min-max regret traveling
salesman problem, along with a mathematical formulation containing an exponential number
of constraints. Moreover, three exact algorithms were presented and computationally compared
at solving the proposed formulation: a branch-and-bound, a branch-and-cut and a logic-based
Benders’ decomposition [18] algorithm. The latter algorithm has been widely used to solve robust optimization problems whose classical counterparts are polynomially solvable (see, e.g.,
[4, 6, 8]). Computational experiments showed that the logic-based Benders’ algorithm outperforms the other exact algorithms for the interval data min-max regret traveling salesman problem.
Later on, Pereira and Averbakh [11] introduced the interval data min-max regret set covering
problem. The authors proposed a mathematical formulation for the problem that is similar to
4
the one proposed by [5] for the interval data min-max regret traveling salesman problem. As in
[5], the formulation has an exponential number of constraints. They also adapted the logic-based
Benders’ algorithm of [4, 5, 8] to this problem and presented an extension of the method that aims
at generating multiple Benders’ cuts per iteration of the algorithm. Moreover, the work presents
an exact approach that uses Benders’ cuts in the context of a branch-and-cut framework. Computational experiments showed that such approach, as well as the extended logic-based Benders’
algorithm, outperforms the standard logic-based Benders’ algorithm at solving the instances considered. This robust version of the set covering problem was also addressed in [19], where the
authors propose scenario-based heuristics with path-relinking.
A few works also deal with robust optimization versions of the 0-1 knapsack problem. For
instance, the studies [1] and [20] address a version of the problem where the uncertainty over
each item profit is represented by a discrete set of possible values. In these works, the absolute
robustness criterion is considered. In [20], the author proved that this version of the problem is
strongly NP-hard when the number of possible scenarios is unbounded and pseudo-polynomially
solvable for a bounded number of scenarios. Kouvelis et al. [1] also studied a min-max regret
version of the problem that considers a discrete set of scenarios of item profits. They provided a
pseudo-polynomial algorithm for solving the problem when the number of possible scenarios is
bounded. When the number os scenarios is unbounded, the problem becomes strongly NP-hard
and there is no approximation scheme for it [21].
More recently, Feizollahi and Averbakh [22] introduced the min-max regret quadratic assignment problem with interval flows, which is a generalization of the classical quadratic assignment
problem in which material flows between facilities are uncertain and vary in given intervals.
Although quadratic, the problem presents a structure that is very similar to the robust versions
of combinatorial problems we address in this work. The authors proposed two mathematical
formulations and adapted the logic-based Benders’ algorithm of [4, 5, 8] to solve them through
the linearization of the corresponding master problems. They also developed a hybrid approach
which combines Benders’ decomposition with heuristics.
Regarding heuristics for interval robust-hard problems, a simple and efficient scenario-based
procedure to tackle interval min-max regret problems in general was proposed in [23] and successfully applied in several works (see, e.g., [17, 23, 24]). The procedure, called Algorithm Mean
Upper (AMU), consists in solving the corresponding classical optimization problem in two spe5
cific scenarios: the so called worst-case scenario, where the cost referred to each binary variable
is set to its upper bound, and the mid-point scenario, where the cost of the binary variables are
set to the mean values referred to the bounds of the respective cost intervals. With this heuristic,
one can obtain a feasible solution for any interval min-max regret optimization problem (including interval robust-hard problems) with the same worst-case asymptotic complexity of solving
an instance of the classical optimization problem counterpart. Moreover, it is proven in [24] that
this algorithm is 2-approximative for any interval min-max regret optimization problem. Notice
that AMU does not run in polynomial time for interval robust-hard problems, unless P = NP.
The study of approximative procedures for interval min-max regret problems in general is a
relatively unexplored field, and not much is known. Conde [25] proved that AMU gives a 2approximation for an even broader class of min-max regret problems. Precisely, while the works
of Kasperski et al. [23, 24] only address combinatorial (finite and discrete) min-max regret optimization models, Conde [25] extends the 2-approximation result for models with compact constraint sets in general, including continuous ones. Recently, some research has been conducted to
refine the constant factor of the aforementioned approximation. For example, in [26], the author
attempts to tighten this factor of 2 through the resolution of a robust optimization problem over
a reduced uncertainty cost set. Moreover, in [27], the authors introduced a new bound that gives
an instance dependent performance guarantee of the mid-point scenario solution for interval data
min-max regret versions of combinatorial problems. They show that the new performance ratio is at most 2, and the bound is successfully applied to solve the interval data min-max regret
shortest path problem [1] within a branch-and-bound framework.
Kasperski and Zieliński [28] also developed a Fully Polynomial Time Approximation Scheme
(FPTAS) for interval min-max regret problems. However, this FPTAS relies on two very restrictive conditions: (i) the problem tackled must present a pseudopolynomial algorithm, and (ii) the
corresponding classical counterpart has to be polynomially solvable. Condition (ii) comes from
the fact that the aforementioned FPTAS uses AMU within its framework. Notice that, from (ii),
this FPTAS does not naturally hold for interval robust-hard problems.
The existence of efficient approximative algorithms is closely related to another almost unexplored aspect of interval robust-hard problems, which is their computational complexity. Although NP-hard by definition, whether these problems are necessarily strongly NP-hard or not is
still an open issue.
6
3. Modeling and solving interval min-max regret problems
In this section, we discuss a standard modeling technique in the literature of interval min-max
regret problems, as well as a state-of-the-art exact algorithm to solve problems in this class. To
this end, consider G, a generic 0-1 ILP combinatorial problem defined as follows.
(G)
min
s.t.
cy
(1)
Ay ≥ b
(2)
y ∈ {0, 1}n .
(3)
The binary variables are represented by an n-dimensional column vector y, whereas their corresponding cost values are given by an n-dimensional row vector c. Moreover, b is an mdimensional column vector, and A is an m × n matrix. The feasible region of G is given by
Ω = {y : Ay ≥ b, y ∈ {0, 1}n }.
Let R be an interval data min-max regret robust optimization version of G, where a continuous
cost interval [li , ui ], with li , ui ∈ Z+ and li ≤ ui , is associated with each binary variable yi ,
i = 1, . . . , n. The following definitions describe R formally.
Definition 1. A scenario s is an assignment of costs to the binary variables, i.e., a cost cis ∈ [li , ui ]
is fixed for all yi , i = 1, . . . , n.
Let S be the set of all possible cost scenarios, which consists of the cartesian product of the
continuous intervals [li , ui ], i = 1, . . . , n. The cost of a solution y ∈ Ω in a scenario s ∈ S is given
n
P
by c s y = cis yi .
i=1
Definition 2. A solution opt(s) ∈ Ω is said to be optimal for a scenario s ∈ S if it has the
smallest cost in s among all the solutions in Ω, i.e., opt(s) = arg min c s y.
y∈Ω
Definition 3. The regret (robust deviation) of a solution y ∈ Ω in a scenario s ∈ S, denoted by
rys , is the difference between the cost of y in s and the cost of opt(s) in s, i.e., rys = c s y − c s opt(s).
Definition 4. The robustness cost of a solution y ∈ Ω, denoted by Ry , is the maximum regret of y
among all possible scenarios, i.e., Ry = max rys .
s∈S
∗
Definition 5. A solution y ∈ Ω is said to be robust if it has the smallest robustness cost among
all the solutions in Ω, i.e., y∗ = arg min Ry .
y∈Ω
7
Definition 6. The interval data min-max regret problem R consists in finding a robust solution.
For each scenario s ∈ S, let G(s) denote the corresponding 0-1 ILP problem G under the cost
vector c s ∈ Rn+ referred to s. Also consider an n-dimensional column vector x of binary variables.
Then, R can be generically modeled as follows.
G(s)
(R)
z }| {
max (c y − min c s x)
(4)
y ∈ Ω.
(5)
s
min
x∈Ω
s∈S
s.t.
Theorem 1 (Aissi et al. [14]). The robust deviation (regret) of any feasible solution ȳ ∈ Ω is
maximum in the scenario s(ȳ) induced by ȳ, defined as follows:
for all i ∈ {1, . . . , n},
cis(ȳ)
ui , if ȳi = 1,
=
li , if ȳi = 0.
Theorem 1, which is also stated in [15], reduces the number of scenarios to be considered
during the search for a robust solution. Accordingly, R can be rewritten taking into account only
the scenario induced by the solution that the y variables define. This scenario, referred to as s(y),
is represented by the cost vector c s(y) = l1 + (u1 − l1 )y1 , . . . , ln + (un − ln )yn . Then, R can be
rewritten as
G(s(y))
(R̃)
z }| {
c s(y) y − min c s(y) x
min
x∈Ω
y ∈ Ω.
s.t.
(6)
(7)
In order to obtain an MILP formulation for R, we reformulate R̃ according to [14]. Precisely,
we add a free variable ρ and linear constraints that explicitly bound ρ with respect to all the
feasible solutions that x can represent. The resulting MILP formulation is provided from (8) to
(11).
(F )
min
(
n
X
ui yi − ρ)
(8)
i=1
s.t.
ρ≤
n
X
(li + (ui − li )yi ) x̄i
∀ x̄ ∈ Ω,
(9)
i=1
y ∈ Ω,
(10)
ρ free.
(11)
8
Constraints (9) ensure that ρ does not exceed the value related to the inner minimization in (6).
Note that, in (9), x̄ is a constant vector, one for each solution in Ω. These constraints are tight
whenever x̄ is optimal for the classical counterpart problem G in the scenario s(y). Constraints
(10) define the feasible region referred to the y variables, and constraint (11) gives the domain of
the variable ρ.
The number of constraints (9) corresponds to the number of feasible solutions in Ω. As the
size of this region may grow exponentially with the number of binary variables, this fomulation is
particularly suitable to be handled by decomposition methods, such as the logic-based Benders’
decomposition [18] algorithm detailed below.
3.1. A logic-based Benders’ decomposition algorithm
Benders’ decomposition method was originally proposed in [29] (also see [30]) to tackle MILP
problems by exploring duality theory properties. Several methodologies were later studied to
improve the convergence speed of the method (see, e.g., [31, 32, 33]). More recently, Hooker and
Ottosson [18] introduced the idea of a logic-based Benders’ decomposition, which is a Benderslike decomposition approach suitable for a broader class of problems. This specific method
is intended to address any optimization problem by devising valid cuts solely based on logic
inference.
The logic-based Benders’ algorithm presented below is a state-of-the-art exact method for
interval min-max regret problems [4, 5, 6, 8, 11]. The algorithm solves formulation F , given
by (8)-(11), by assuming that, since several of constraints (9) might be inactive at optimality,
they can be generated on demand whenever they are violated. The procedure is described in
Algorithm 1. Let Ωψ ⊆ Ω be the set of solutions x̄ ∈ Ω (Benders’ cuts) available at an iteration
ψ. Also let F ψ be a relaxed version of F in which constraints (9) are replaced by
ρ≤
n
X
(li + (ui − li )yi ) x̄i
∀ x̄ ∈ Ωψ .
(12)
i=1
Thus, the relaxed problem F ψ , called master problem, is defined by (8), (10), (11) and (12).
Let ubψ keep the best upper bound found (until an iteration ψ) on the solution of F . Notice
that, at the beginning of Algorithm 1, Ω1 contains the initial Benders’ cuts available, whereas
ub1 keeps the initial upper bound on the solution of F . In this case, Ω1 = ∅ and ub1 := +∞. At
each iteration ψ, the algorithm obtains a solution by solving a corresponding master problem F ψ
9
and seeks a constraint (9) that is most violated by this solution. Initially, no constraint (12) is
considered, since Ω1 = ∅. An initialization step is then necessary to add at least one solution to
Ω1 , thus avoiding unbounded solutions during the first resolution of the master problem. To this
end, it is computed an optimal solution for the worst-case scenario su , in which c su = u (Step I,
Algorithm 1).
Algorithm 1: Logic-based Benders’ algorithm.
Input: Cost intervals [li , ui ] referred to yi , i = 1, . . . , n.
Output: A robust solution for F , and its corresponding robustness cost.
ψ ← 1; ub1 ← +∞; Ω1 ← ∅;
Step I. (Initialization)
Find an optimal solution x̄1 = opt(su ) for the worst-case scenario su ;
Ω1 ← Ω1 ∪ { x̄1 };
Step II. (Master problem)
Solve the relaxed problem F ψ , obtaining a solution (ȳψ , ρ̄ψ );
Step III. (Slave problem)
Find an optimal solution x̄ψ = opt(s(ȳψ )) for the scenario s(ȳψ ) induced by ȳψ and use it to
compute Rȳψ , the robustness cost of ȳψ ;
Step IV. (Stopping condition)
lbψ ← uȳψ − ρ̄ψ ;
if lbψ ≥ Rȳψ then
Return (ȳ∗ , R∗ );
end
else
ubψ+1 ← ubψ ← min{ubψ , Rȳψ };
Ωψ+1 ← Ωψ ∪ { x̄ψ };
ψ ← ψ + 1;
Go to Step II;
end
After the initialization step, at each iteration ψ, the corresponding relaxed problem F ψ is
solved (Step II, Algorithm 1), obtaining a solution (ȳψ , ρ̄ψ ). Then, the algorithm checks if (ȳψ , ρ̄ψ )
violates any constraint (9) of the original problem F . For this purpose, it is solved a slave
10
problem that computes Rȳψ (the actual robustness cost of ȳψ ) by finding an optimal solution
x̄ψ = opt(s(ȳψ )) for the scenario s(ȳψ ) induced by ȳψ (see Step III, Algorithm 1). Notice that
each slave problem involves solving the classical optimization problem G, given by (1)-(3), in
the scenario s(ȳψ ).
Let lbψ = uȳψ − ρ̄ψ be the value of the objective function in (8) related to the solution (ȳψ , ρ̄ψ ) of
the current master problem F ψ . According to [34], lbψ and Rȳψ give, respectively, a lower (dual)
and an upper (primal) bounds on the solution of F . If lbψ does not reach Rȳψ , ubψ and ubψ+1
are both set to the best upper bound found by the algorithm until the iteration ψ. In addition,
a new constraint (12) is generated from x̄ψ and added to F ψ+1 by setting Ωψ+1 ← Ωψ ∪ { x̄ψ }.
Otherwise, if lbψ equals Rȳψ , the algorithm stops (see Step IV of Algorithm 1). As proved in
[34], this stopping condition is always satisfied in a finite number of iterations.
The algorithm detailed above is applied to obtain optimal solutions for the interval robust-hard
problems considered in this work, and their corresponding bounds are used to evaluate the quality
of the solutions produced by the heuristic proposed in the next section. We highlight that other
exact algorithms can be adapted and used for this purpose, such as the branch-and-cut algorithm
proposed by Pereira and Averbakh [11] for the interval data min-max regret set covering problem.
In this work, we chose the logic-based Benders’ algorithm, as it has already been successfully
applied to solve a wide range of interval min-max regret problems [4, 5, 6, 8, 11].
4. An LP based heuristic framework for interval min-max regret problems
In this section, we present an LP based heuristic framework for interval robust-hard problems
which is applicable to interval min-max regret problems in general. Consider an interval minmax regret problem R̃, as defined by (6) and (7) in Section 3. For a given ȳ ∈ Ω, the inner
minimization in (6) is a 0-1 ILP problem, namely G(s(ȳ)). Particularly, it corresponds to problem
G, defined by (1)-(3), under the cost vector c s(ȳ) , where s(ȳ) is the scenario induced by ȳ.
Relaxing the integrality on the x variables of G(s(ȳ)), we obtain the LP problem
ϑ(ȳ) = min
s.t.
11
c s(ȳ) x
(13)
Ax ≥ b,
(14)
I x ≤ 1,
(15)
x ≥ 0,
(16)
whose corresponding dual problem is given by
(D(ȳ)) ϑ(ȳ) = max
s.t.
(bT λ + 1T µ)
(17)
AT λ + I T µ ≤ (c s(ȳ) )T ,
(18)
λ ≥ 0,
(19)
µ ≤ 0.
(20)
Here, I is the identidy matrix, and the dual variables λ and µ are associated, respectively, with
constraints (14) and (15) of the primal problem. Notice that, in both LP problems, ϑ(ȳ) corresponds to an optimal solution cost and gives a lower (dual) bound on the optimal value of G(s(ȳ)).
Precisely,
G(s(ȳ))
z }| {
ϑ(ȳ) ≤ min c s(ȳ) x
x∈Ω
∀ ȳ ∈ Ω.
(21)
Replacing G(s(y)) by D(y) in (6), we obtain
min
s.t.
c s(y) y − max (bT λ + 1T µ)
(22)
AT λ + I T µ ≤ (c s(y) )T .
(23)
Constraints (7), (19) and (20),
Recall that c s(y) = l1 + (u1 − l1 )y1 , . . . , ln + (un − ln )yn is the cost vector referred to the scenario
s(y) induced by the y variables. Notice that the nested maximization operator can be omitted,
giving the following formulation.
(R̂)
min
s.t.
c s(y) y − bT λ − 1T µ
(24)
Constraints (7), (19), (20) and (23).
Proposition 1. The cost value referred to an optimal solution for R̂ gives an upper bound on
the optimal solution value of R̃, and this bound is tight (optimal) if (i) the restriction matrix A is
totally unimodular, and (ii) the column vector b is integral.
Proof. Let (ỹ, x̃) and (ŷ, λ̂, µ̂) be optimal solutions for R̃ and R̂, respectively. The value of the
objective function in (6) referred to (ỹ, x̃) is given by (c s(ỹ) ỹ − c s(ỹ) x̃), whereas that of (24) referred
to (ŷ, λ̂, µ̂) is (c s(ŷ) ŷ − (bT λ̂ + 1T µ̂)). Since (ỹ, x̃) is optimal for R̃, we have that (c s(ỹ) ỹ − c s(ỹ) x̃) ≤
(c s(y) y − min c s(y) x) for all y ∈ Ω. In particular, as ŷ ∈ Ω, it holds that
x∈Ω
c s(ỹ) ỹ − c s(ỹ) x̃ ≤ c s(ŷ) ŷ − min c s(ŷ) x.
x∈Ω
12
(25)
As (ŷ, λ̂, µ̂) is optimal for R̂, it follows, from (13)-(21), that
c s(ŷ) ŷ − min c s(ŷ) x ≤ c s(ŷ) ŷ − ϑ(ŷ) = c s(ŷ) ŷ − (bT λ̂ + 1T µ̂).
x∈Ω
(26)
Then, from (25) and (26), we obtain that
c s(ỹ) ỹ − c s(ỹ) x̃ ≤ c s(ŷ) ŷ − (bT λ̂ + 1T µ̂).
(27)
Now, also suppose that (i) the matrix A is totally unimodular, and (ii) the column vector b is
integral. As (ŷ, λ̂, µ̂) is optimal for R̂, it follows, from (13)-(20), that
c s(ŷ) ŷ − (bT λ̂ + 1T µ̂) = c s(ŷ) ŷ − ϑ(ŷ) ≤ c s(y) y − ϑ(y) ∀y ∈ Ω.
(28)
Particularly, as ỹ ∈ Ω,
c s(ŷ) ŷ − (bT λ̂ + 1T µ̂) ≤ c s(ỹ) ỹ − ϑ(ỹ).
(29)
Additionally, from assumption (i), it follows that the restriction matrix (A, I)T referred to (17)(20) is also totally unimodular. Thus, from assumption (ii), inequation (21) is tight, i.e.,
ϑ(ȳ) = min c s(ȳ) x
x∈Ω
∀ȳ ∈ Ω.
(30)
As (ỹ, x̃) is optimal for R̃, we obtain, from (29) and (30),
c s(ŷ) ŷ − (bT λ̂ + 1T µ̂) ≤ c s(ỹ) ỹ − ϑ(ỹ) = c s(ỹ) ỹ − min c s(ỹ) x = c s(ỹ) ỹ − c s(ỹ) x̃,
x∈Ω
(31)
which, along with (27), implies (c s(ŷ) ŷ − (bT λ̂ + 1T µ̂)) = (c s(ỹ) ỹ − c s(ỹ) x̃).
For a given problem R, defined by (4) and (5), the heuristic framework consists in (I) solving
the corresponding formulation R̂, obtaining a solution (ŷ, λ̂, µ̂), and (II) computing the robustness
cost (maximum regret) referred to ŷ, considering Theorem 1. One may note that the heuristic is
applicable not only to interval robust-hard problems, but to any interval min-max regret problem of the general form of R. From Proposition 1, whenever the classical optimization problem
counterpart can be modeled as a 0-1 ILP of the form of G, with a totally unimodular restriction
matrix and b integral, there is a guarantee of optimality at solving R̂. In fact, applying the framework detailed above to the interval data min-max regret versions of the polynomially solvable
shortest path and assignment problems presented, respectively, in [16] and [23], leads to the same
compact MILP formulations proposed and computationally tested in these works.
13
Notice that, whenever G is compact, the resulting formulation R̂ is also compact. We also
highlight that, although the heuristic framework was detailed by the assumption of G being a
minimization problem, the results also hold for maximization problems, with minor modifications. In addition, we conjecture that the heuristic framework also provides valid bounds for the
wider class of interval data min-max regret problems with compact constraint sets addressed in
[25, 26]. We believe that Theorem 1 can be extended to this class of problems through the definition of a specific scenario similar to the one induced by a solution. However, a more careful
study needs to be conducted to close this issue.
In this work, we do not present any theoretical guarantee of quality for the solutions obtained
by the heuristic, as it may rely on specific problem dependent factors, such as the strength of
the formulation adopted to model the classical counterpart problem G. Nevertheless, in the next
section, we present two very successful applications of the heuristic in solving interval robusthard problems.
5. Case studies in solving interval robust-hard problems
In this section, we define the interval robust-hard problems used as case studies for the proposed heuristic framework. For each problem, we give a mathematical formulation according to
the modeling technique presented in Section 3 and devise the corresponding MILP formulation
tackled by the heuristic. For the interval data min-max regret restricted shortest path problem,
which is introduced in this work, we give a more detailed description, along with a brief motivation and some related works. For simplicity, let robust stand for interval data min-max regret in
the designation of each problem.
5.1. The Restricted Robust Shortest Path problem (R-RSP)
The Restricted Robust Shortest Path problem (R-RSP) is an interval data min-max regret version of the Restricted Shortest Path problem (R-SP), an extensively studied NP-hard problem
[12, 13, 35, 36, 37, 38]. Consider a digraph G = (V, A), where V is the set of vertices, and A
is the set of arcs. With each arc (i, j) ∈ A, we associate a resource consumption di j ∈ Z+ and
a continuous cost interval [li j ,ui j ], where li j ∈ Z+ is the lower bound, and ui j ∈ Z+ is the upper
bound on this interval of cost, with li j ≤ ui j . An origin vertex o ∈ V and a destination one t ∈ V
are also given, as well as a value β ∈ Z, parameter used to limit the resource consumed along a
14
path from o to t in G, as discussed in the sequel. An example of an R-RSP instance is given in
Figure 1.
0
}
{2
8]
,
[5
1
}
{1
2]
,
[2
[3,5]{1}
[1,7]{3}
[3,6]{1}
2
[2,2]{1}
3
Figure 1: Example of an R-RSP instance, with origin o = 0 and destination t = 3. The notation
[li j , ui j ]{di j } means, respectively, the cost interval [li j , ui j ] and the resource consumption {di j }
associated with each arc (i, j) ∈ A.
Here, a scenario s is an assignment of arc costs, where a cost cisj ∈ [li j , ui j ] is fixed for all
(i, j) ∈ A. Let P be the set of all paths from o to t and A[p] be the set of the arcs that compose
a path p ∈ P. Also let S be the set of all possible scenarios of G. The cost of a path p ∈ P in
P
a scenario s ∈ S is given by C ps =
cisj . Similarly, the resource consumption referred to a
(i,
j)∈A[p]
P
path p ∈ P is given by D p =
di j . Also consider P(β) = {p ∈ P | D p ≤ β}, the subset of
(i, j)∈A[p]
paths in P whose resource consumptions are smaller than or equal to β.
Definition 7. A path p∗ (s, β) ∈ P(β) is said to be a β-restricted shortest path in a scenario s ∈ S
if it has the smallest cost in s among all paths in P(β), i.e., p∗ (s, β) = arg min C ps .
p∈P(β)
Definition 8. The regret (robust deviation) of a path p ∈ P(β) in a scenario s ∈ S, denoted by
β)
r(s,
, is the difference between the cost C ps of p in s and the cost of a β-restricted shortest path
p
β)
p∗ (s, β) ∈ P(β) in s, i.e., r(s,
= C ps − C ps ∗ (s, β) .
p
Definition 9. The β-restricted robustness cost of a path p ∈ P(β), denoted by Rβp , is the maximum
(s, β)
β
regret of p among all possible scenarios, i.e., R p = max r p
s∈S
.
Definition 10. A path p∗ ∈ P(β) is said to be a β-restricted robust path if it has the smallest
β-restricted robustness cost among all paths in P(β), i.e., p∗ = arg min Rβp .
p∈P(β)
Definition 11. R-RSP consists in finding a β-restricted robust path p∗ ∈ P(β).
15
R-RSP has applications in determining paths in urban areas, where travel distances are known
in advance, but travel times may vary according to unpredictable traffic jams, bad weather conditions, etc. Here, uncertainties are represented by values in continuous intervals, which estimate
the minimum and the maximum traveling times to traverse each pathway. For instance, R-RSP
can model situations involving electrical vehicles with a limited battery (energy) autonomy, when
one wants to find the fastest robust path with a length (distance) constraint. A similar application
arises in telecommunications networks, when one wants to determine a path to efficiently send
a data packet from an origin node to a destination one in a network. Due to the varying traffic
load, transmission links are subject to uncertain delays. Moreover, each link is associated with a
packet loss rate. In order to guarantee Quality of Service (QoS) [38, 39], a limit is imposed on
the total packet loss rates of the path used.
The classical R-SP is a particular case of R-RSP in which li j = ui j ∀ (i, j) ∈ A. As R-SP is
known to be NP-hard [12, 13] even for acyclic graphs [38], the same holds for R-RSP. The main
exact algorithms to solve R-SP can be subdivided into two groups: Lagrangian relaxation and
dynamic programming procedures. The former procedures use Lagrangian relaxation to handle
ILP formulations for the problem (see, e.g., [13, 36]). In addition, preprocessing techniques have
been presented in [35] and refined in [36]. These techniques identify arcs and vertices that cannot
compose an optimal solution for R-SP through the analysis of the reduced costs related to the
resolution of dual Lagrangian relaxations. More recently, [40] proposed a path ranking approach
that linearly combines the arc costs and the resource consumption values to generate a descent
direction of search. In turn, dynamic programming procedures for R-SP consist of label-setting
and label-correcting algorithms, such as the one proposed in [41] and further improved in [42] by
the addition of preprocessing strategies. Recently, Zhu and Wilhelm [43] developed a three-stage
label-setting algorithm that runs in pseudo-polynomial time and figures among the state-of-theart methods to solve R-SP, along with the algorithms presented in [40] and [42]. We refer to [44]
for a survey on exact methods to solve R-SP.
Although NP-hard, R-SP can be solved efficiently by some of the aforementioned procedures
(particularly, the ones proposed in [40, 42, 43]). Moreover, optimization softwares, as CPLEX1 ,
are also competitive in handling reasonably-sized instances (with up to 3000 vertices) of the
problem [43].
1 http://www-01.ibm.com/software/commerce/optimization/cplex-optimizer/
16
R-RSP is also a generalization of the interval data min-max regret Robust Shortest Path problem (RSP) [1, 3, 45, 46]. RSP consists in finding a robust path (from the origin vertex to the
destination one) considering the min-max regret criterion, with no additional resource consumption restriction on the solution path. Therefore, RSP can be reduced to R-RSP by considering
β = 0 and di j = 0 ∀ (i, j) ∈ A.
Preprocessing techniques able to identify arcs that cannot compose an optimal solution for
RSP have been proposed in [16] and later improved in [47]. A compact MILP formulation for
the problem, based on the dual of an LP formulation for the classical shortest path problem, was
presented in [16]. Exact algorithms have been proposed for RSP, such as the branch-and-bound
algorithm of [48] and the logic-based Benders’ algorithm of [8], which is able to solve instances
with up to 4000 vertices. Moreover, the FPTAS of [28] for interval min-max regret problems
is applicable to RSP when the problem is considered in series-parallel graphs. However, as
pointed out by the end of Section 2, this FPTAS does not naturally extend to interval robust-hard
problems, such as R-RSP.
5.1.1. Mathematical formulation
The mathematical formulation here presented makes use of Theorem 1, which can be stated
for the specific case of R-RSP as follows.
Proposition 2. Given a value β ∈ Z and a path p ∈ P(β), the regret of p is maximum in the
scenario s(p) induced by p, where the costs of all the arcs of p are in their corresponding upper
bounds and the costs of all the other arcs are in their lower bounds.
2{
1}
5{1}
1
1{3}
3{1}
2
2{1}
5{
2}
0
3
Figure 2: Scenario s p̃ induced by the path p̃ = {0, 1, 2, 3} in the digraph presented in Figure 1.
s( p̃)
s( p̃)
For each arc (i, j) ∈ A, the notation ci j {di j } means, respectively, the arc cost ci j in the scenario
s( p̃) and its resource consumption {di j }.
17
Consider the digraph presented in Figure 1 and the scenario s( p̃) induced by the path p̃ =
{0, 1, 2, 3} (as showed in Figure 2). Also let the resource limit be β = 3. According to Proposition
p̃), β)
2, the β-restricted robustness cost of p̃ is given by Rβp̃ = r(s(
= C p̃s( p̃) − C ps(∗p̃)
p̃
(s( p̃), β) = (2 +
2 + 5) − (3 + 5) = 1. Proposition 2 reduces R-RSP to finding a path p∗ ∈ P(β) such that
(s(p), β)
p∗ = arg min r p
p∈P
s(p)
, i.e., p∗ = arg min {C p
p∈P(β)
s(p)
− C p∗ (s(p), β) }. Nevertheless, computing the β-
restricted robustness cost of any feasible solution p for R-RSP still implies finding a β-restricted
shortest path p∗ (s(p), β) in the scenario s(p) induced by p, and this problem is NP-hard.
Now, let us consider the following result, which helps describing an optimal solution path for
R-RSP.
Property 1. Given two arbitrary sets Z1 and Z2 , it holds that Z1 = (Z1 ∩ Z2 ) ∪ (Z1 \Z2 ).
Theorem 2. Given a value β ∈ Z and a non-elementary path p ∈ P(β), for any elementary path
p̃),β)
p̃ ∈ P(β) such that A[ p̃] ⊂ A[p], it holds that r(s(
≤ r(s(p),β)
.
p
p̃
Proof. Consider a value β ∈ Z and a non-elementary path p ∈ P(β). By definition, p contains at
least one cycle. Let G[p] be the subgraph of G induced by the arcs in A[p] and p̃ be an elementary
path from o to t in G[p]. Clearly, A[ p̃] ⊂ A[p] and, therefore, p̃ ∈ P(β). By definition,
(s( p̃),β)
r p̃
s( p̃)
= C p̃
s( p̃)
− C p∗ (s( p̃),β) .
(32)
Consider the set Ā = A[p]\A[ p̃] of the arcs in p which do not belong to p̃. As p is supposed to
contain at least one cycle, then Ā , ∅. Since A[ p̃] ⊂ A[p], the difference between scenarios s(p)
and s( p̃) consists of the cost values assumed by the arcs in Ā. More precisely,
cis(j p̃) = cis(p)
j
cis(p)
j = ui j
s( p̃)
ci j = li j
∀(i, j) ∈ A\Ā,
(33)
∀(i, j) ∈ Ā,
(34)
∀(i, j) ∈ Ā.
(35)
It follows that
C p̃s( p̃) =
X
(i, j)∈A[ p̃]
ui j =
X
(i, j)∈A[p]
X
ui j −
(i, j)∈A[p]\A[ p̃]
18
ui j = C ps(p) −
X
(i, j)∈Ā
ui j .
(36)
Applying Property 1 to the sets A[p∗ (s( p̃), β)] and A[p]:
(a1 )
z
}|
{
A[p (s( p̃), β)] = (A[p∗ (s( p̃), β)] ∩ A[p]) ∪(A[p∗ (s( p̃), β)]\A[p]).
∗
(37)
Since A[ p̃] ⊂ A[p] and Ā = A[p]\A[ p̃], it follows that
A[p] = A[ p̃] ∪ (A[p]\A[ p̃]) = A[ p̃] ∪ Ā.
(38)
Therefore, expression (a1 ) of (37) can be rewritten as
(A[p∗ (s( p̃), β)] ∩ A[p]) = (A[p∗ (s( p̃), β)] ∩ (A[ p̃] ∪ Ā)) =
(A[p∗ (s( p̃), β)] ∩ A[ p̃]) ∪ (A[p∗ (s( p̃), β)] ∩ Ā).
(39)
s
p
Applying (37) and (39) to C ps(∗p̃)
(s( p̃),β) and C p∗ (s( p̃),β) , we obtain:
X
C ps(∗p̃)
(s( p̃),β) =
(i, j)∈A[p∗ (s( p̃),β)]∩A[ p̃]
cis(j p̃)
(i, j)∈A[p∗ (s( p̃),β)]∩Ā
X
+
X
cis(j p̃) +
cis(j p̃) ,
(40)
(i, j)∈A[p∗ (s( p̃),β)]\A[p]
X
C ps(p)
∗ (s( p̃),β) =
(i, j)∈A[p∗ (s( p̃),β)]∩A[ p̃]
X
+
X
cis(p)
j +
cis(p)
j
(i, j)∈A[p∗ (s( p̃),β)]∩Ā
s(p)
ci j .
(41)
(i, j)∈A[p∗ (s( p̃),β)]\A[p]
Considering (33)-(35), (40) and (41), we deduce that the difference between the cost of path
p∗ (s( p̃), β) in s(p) and its cost in s( p̃) is given by the arcs which are simultaneously in Ā and in
A[p∗ (s( p̃), β)]. Thus, expressions (40) and (41) can be reformulated as
X
C ps(∗p̃)
(s( p̃),β) =
(i, j)∈A[p∗ (s( p̃),β)]∩A[ p̃]
+
X
ui j +
X
(i, j)∈A[p∗ (s( p̃),β)]\A[p]
19
li j
(i, j)∈A[p∗ (s( p̃),β)]∩Ā
li j ,
(42)
X
C ps(p)
∗ (s( p̃),β) =
(i, j)∈A[p∗ (s( p̃),β)]∩A[ p̃]
ui j
(i, j)∈A[p∗ (s( p̃),β)]∩Ā
X
+
X
ui j +
li j .
(43)
(i, j)∈A[p∗ (s( p̃),β)]\A[p]
Subtracting (43) from (42),
X
s(p)
C ps(∗p̃)
(s( p̃),β) − C p∗ (s( p̃),β) =
X
li j −
(i, j)∈A[p∗ (s( p̃),β)]∩Ā
ui j .
(44)
(i, j)∈A[p∗ (s( p̃),β)]∩Ā
Therefore,
X
s(p)
C ps(∗p̃)
(s( p̃),β) = C p∗ (s( p̃),β) −
(ui j − li j ).
(45)
(i, j)∈A[p∗ (s( p̃),β)]∩Ā
Applying (36) and (45) in (32):
p̃),β)
r(s(
= C ps(p) −
p̃
X
(i, j)∈Ā
ui j − C ps(p)
∗ (s( p̃),β) −
X
C ps(p) − C ps(p)
∗ (s( p̃),β) +
X
(i, j)∈A[p∗ (s( p̃),β)]∩Ā
(ui j − li j ) −
(i, j)∈A[p∗ (s( p̃),β)]∩Ā
(ui j − li j ) =
X
ui j .
(46)
(i, j)∈Ā
One may note that
X
(ui j − li j ) −
(i, j)∈A[p∗ (s( p̃),β)]∩Ā
X
(i, j)∈Ā
ui j ≤
X
(ui j − li j ) −
X
(i, j)∈Ā
(i, j)∈Ā
P
Since Ā ⊂ A, and li j ≥ 0 for all (i, j) ∈ A, it follows that
ui j ≤
X
(−li j ).
(47)
(i, j)∈Ā
(−li j ) ≤ 0 and, thus,
(i, j)∈Ā
X
(ui j − li j ) −
(i, j)∈A[p∗ (s( p̃),β)]∩Ā
X
ui j ≤ 0.
(48)
(i, j)∈Ā
From (46) and (48),
p̃),β)
r(s(
≤ C ps(p) − C ps(p)
∗ (s( p̃),β) .
p̃
20
(49)
As p∗ (s(p), β) is a path with the smallest cost in s(p) among all the paths in P(β), including
s(p)
p∗ (s( p̃), β), it holds that C ps(p)
∗ (s( p̃),β) ≥ C p∗ (s(p),β) . Thus,
p̃),β)
s(p)
r(s(
≤ C ps(p) − C ps(p)
− C ps(p)
∗ (s( p̃),β) ≤ C p
∗ (s(p),β) .
p̃
(50)
(s( p̃),β)
By definition, r(s(p),β)
= C ps(p) − C ps(p)
≤ r(s(p),β)
.
∗ (s(p),β) . Therefore, r p̃
p
p
Now, we can devise a mathematical formulation for R-RSP from the generic model R̃, defined
by (6) and (7). Consider the decision variables y on the choice of arcs belonging or not to a βrestricted robust path: yi j = 1 if the arc (i, j) ∈ A belongs to the solution path; yi j = 0, otherwise.
Likewise, let the binary variables x identify a β-restricted shortest path in the scenario induced
by the path defined by y, such that xi j = 1 if the arc (i, j) ∈ A belongs to this β-restricted shortest
path, and xi j = 0, otherwise. A nonlinear compact formulation for R-RSP is given by
!
X
X
(li j + (ui j − li j )yi j )xi j .
min
ui j yi j − min
y∈P(β)
x∈P(β)
(i, j)∈A
(51)
(i, j)∈A
As discussed in Section 3, we can derive an MILP formulation from (51) by adding a free
variable ρ and linear constraints that explicitly bound ρ with respect to all the feasible paths that
x can represent. The resulting formulation is provided from (52) to (59).
min
X
u i j yi j − ρ
(52)
(i, j)∈A
s.t.
X
y jo −
X
y ji −
X
y jt −
j:( j,o)∈A
yok = −1,
(53)
k:(o,k)∈A
j:( j,i)∈A
X
yik = 0
X
ytk = 1,
∀ i ∈ V\{o, t},
(54)
k:(i,k)∈A
j:( j,t)∈A
X
X
(55)
k:(t,k)∈A
di j yi j ≤ β,
(56)
(i, j)∈A
ρ≤
X
(li j + (ui j − li j )yi j ) x̄i j
∀ x̄ ∈ P(β),
(57)
(i, j)∈A
yi j ∈ {0, 1} ∀ (i, j) ∈ A,
(58)
ρ free.
(59)
The flow conservation constraints (53)-(55), along with the domain constraints (58), ensure
that the y variables define a path from the origin to the destination vertices. In fact, as pointed out
21
in [16], these constraints do not prevent the existence of additional cycles of cost zero disjoint
from the solution path. Notice, however, that every arc (i, j) of these cycles necessarily has
li j = ui j = 0 and, thus, they do not modify the optimal solution value. Hence, these cycles are
not taken into account hereafter.
Constraint (56) limits the resource consumption of the path defined by y to be at most β,
whereas constraints (57) guarantee that ρ does not exceed the value related to the inner minimization in (51). Note that, in (57), x̄ is a constant vector, one for each path in P(β). Moreover,
these constraints are tight whenever x̄ identifies a β-restricted shortest path in the scenario induced by the path that y defines. Constraint (59) gives the domain of the variable ρ.
Notice that, in the definition of R-RSP, we do not impose that a vertex in the solution path must
be traversed at most once. However, if that is the case, Theorem 2 indicates that the formulation
presented above can also be used to determine an elementary solution path for R-RSP by simply
discarding some edges from the cycles that may appear in the solution. In fact, Proposition 2 and
Theorem 2 imply that, for any β ∈ Z, if P(β) , ∅, then there is an elementary path p ∈ P(β)
which is a β-restricted robust path.
In this study, we use the logic-based Benders’ algorithm discussed in the Section 3.1 to solve
the formulation detailed above.
5.1.2. An LP based heuristic for R-RSP
In this section, we apply to R-RSP the heuristic framework presented in Section 4. To this
end, consider the following R-SP ILP formulation used to compute a β-restricted shortest path
p∗ (s, β) ∈ P(β) in a scenario s ∈ S. The binary variables x define p∗ (s, β), such that xi j = 1 if the
arc (i, j) ∈ A belongs to A[p∗ (s, β)], and xi j = 0, otherwise.
(I1 )
min
X
cisj xi j
(60)
(i, j)∈A
s.t.
X
x jo −
X
x ji −
X
x jt −
j:( j,o)∈A
X
xok = −1,
(61)
k:(o,k)∈A
j:( j,i)∈A
j:( j,t)∈A
X
X
xik = 0
X
xtk = 1,
∀ i ∈ V\{o, t},
(62)
k:(i,k)∈A
(63)
k:(t,k)∈A
di j xi j ≤ β,
(64)
(i, j)∈A
xi j ∈ {0, 1} ∀ (i, j) ∈ A.
22
(65)
The objective function in (60) represents the cost, in the scenario s, of the path defined by x,
while constraints (61)-(65) ensure that x identifies a path in P(β). Relaxing the integrality on x,
we obtain the following LP formulation:
X
(L1 ) θ(s, β) = min
cisj xi j
(66)
(i, j)∈A
s.t.
Constraints (61)-(64),
xi j ≥ 0
∀ (i, j) ∈ A.
(67)
The domain constraints xi j ≤ 1 for all (i, j) ∈ A were omitted from L1 , since they are redundant. Let θ(s, β) be the optimal value for problem L1 in a scenario s. Observe that θ(s, β) provides
a lower bound on the solution of I1 in the scenario s. For the sake of clarity, let us define a new
metric to evaluate the quality of a path in P(β).
Definition 12. The β-heuristic robustness cost of a path p ∈ P(β), denoted by H βp , is the difference
s(p)
between the cost C p
of p in the scenario s(p) induced by p and the relaxed cost θ(s(p), β) in s(p),
i.e., H βp = C ps(p) − θ(s(p), β) .
Definition 13. A path p̃∗ ∈ P(β) is said to be a β-heuristic robust path if it has the smallest
β
β-heuristic robustness cost among all the paths in P(β), i.e., p̃∗ = arg min H p .
p∈P(β)
In the case of R-RSP, the proposed heuristic aims at finding a β-heuristic robust path and relies
on the hypothesis that such a path is a near-optimal solution for R-RSP. The problem of finding
a β-heuristic robust path can be modeled by adapting formulation (51). To this end, the binary
variables y now represent a β-heuristic robust path in P(β). Furthermore, considering the scenario
s(y) induced by the path defined by y, the nested minimization in (51) is replaced by θ(s(y), β) . We
obtain:
min
y∈P(β)
X
(i, j)∈A
!
ui j yi j − θ(s(y), β) .
(68)
Given a scenario s ∈ S, the optimal value assumed by θ(s, β) can be represented by the dual of
L1 , as follows:
(L̃1 )
θ(s,β) = max
s.t.
(λt − λo − βµ)
λ j ≤ λi + cisj + di j µ
(69)
∀ (i, j) ∈ A,
(70)
µ ≥ 0,
(71)
λk free ∀ k ∈ V.
23
(72)
The dual variables {λk : k ∈ V} and µ are associated, respectively, with constraints (61)-(63)
and with constraint (64) in the primal problem L1 . Since L̃1 is a maximization problem, its
objective function, along with (70)-(72), can be used to replace the relaxed cost θ(s(y),β) in (68),
thus deriving the following formulation:
X
min
y∈P(β)
s.t.
(i, j)∈A
From (69) !
z
}|
{
ui j yi j − (λt − λo − βµ)
λ j ≤ λi + li j + (ui j − li j )yi j + di j µ
(73)
∀ (i, j) ∈ A,
(74)
µ ≥ 0,
(75)
λk free ∀ k ∈ V.
(76)
Notice that constraints (74) consider the cost of each arc (i, j) ∈ A in the scenario s(y) induced
by the path identified by the y variables, i.e., the cost of each arc (i, j) ∈ A is given by (li j + (ui j −
li j )yi j ). The domain constraints (75) and (76) related to L̃1 remain the same. Now, we give an
MILP formulation for the problem of finding a β-heuristic robust path.
!
X
(H1 ) min
ui j yi j − λt + λo + βµ
(77)
(i, j)∈A
s.t.
X
y jo −
X
y ji −
X
y jt −
j:( j,o)∈A
X
yok = −1,
(78)
k:(o,k)∈A
j:( j,i)∈A
j:( j,t)∈A
X
X
yik = 0 ∀ i ∈ V\{o, t},
(79)
X
ytk = 1,
(80)
k:(i,k)∈A
k:(t,k)∈A
di j yi j ≤ β,
(81)
(i, j)∈A
Constraints (74)-(76),
yi j ∈ {0, 1} ∀ (i, j) ∈ A.
(82)
The objective function in (77) represents the β-heuristic robustness cost of the path defined
by the y variables. Constraints (78)-(81) and (82) ensure that y belongs to P(β). Constraints
(74)-(76) are the remaining restrictions related to L̃1 .
The heuristic consists in solving the corresponding problem H1 in order to find a β-heuristic
robust path p̃∗ ∈ P(β). Note that p̃∗ is also a feasible solution path for R-RSP, and, according to
Proposition 1, its β-heuristic robustness cost provides an upper bound on the solution of R-RSP.
Such bound can be improved by the evaluation of the actual β-restricted robustness cost of p̃∗ .
24
5.2. The Robust Set Covering problem (RSC)
The Robust Set Covering problem (RSC), introduced by Pereira and Averbakh [6], is an interval data min-max regret generalization of the Set Covering problem (SC). The classical SC is
known to be strongly NP-hard [12], and, thus, the same holds for RSC.
Let O = (oi j ) be an i × j binary matrix, such that I = {1, . . . , i} and J = {1, . . . , j} are its
corresponding rows and columns sets, respectively. We say that a column j ∈ J covers a row
i ∈ I if oi j = 1. In this sense, a covering is a subset K ⊆ J of columns such that every row in
I is covered by at least one column from K. Hereafter, we denote by Λ the set of all possible
coverings.
In the case of RSC, we associate with each column j ∈ J a continuous cost interval [l j ,u j ],
with l j , u j ∈ Z+ and l j ≤ u j . Accordingly, a scenario s is an assignment of column costs, where
a cost c sj ∈ [l j , u j ] is fixed for all j ∈ J. The set of all these possible cost scenarios is denoted by
P s
S, and the cost of a covering K ∈ Λ in a scenario s ∈ S is given by C Ks =
c j.
j∈K
∗
Definition 14. A covering K (s) ∈ Λ is said to be optimal in a scenario s ∈ S if it has the
smallest cost in s among all coverings in Λ, i.e., K ∗ (s) = arg min C Ks .
K∈Λ
Definition 15. The regret (robust deviation) of a covering K ∈ Λ in a scenario s ∈ S, denoted by
rKs , is the difference between the cost C Ks of K in s and the cost of an optimal covering K ∗ (s) ∈ Λ
in s, i.e., rKs = C Ks − C Ks ∗ (s) .
Definition 16. The robustness cost of a covering K ∈ Λ, denoted by RK , is the maximum regret
of K among all possible scenarios, i.e., RK = max rKs .
s∈S
∗
Definition 17. A covering K ∈ Λ is said to be a robust covering if it has the smallest robustness
cost among all coverings in Λ, i.e., K ∗ = arg min RK .
K∈Λ
Definition 18. RSC consists in finding a robust covering K ∗ ∈ Λ.
As far as we are aware, RSC was only addressed in [6] and [19]. On the other hand, the classical SC has been widely studied in the literature, especially because it can be used as a basic model
for applications in several fields, including production planning [49] and crew management [50].
We refer to [51] for an annotated bibliography on the main applications and the state-of-the-art of
SC. We highlight that, as well as for the classical counterpart of R-RSP, directly solving compact
ILP formulations for RSC via CPLEX’s branch-and-bound is competitive with the best exact
algorithms for the problem [52].
25
5.2.1. Mathematical formulation
The mathematical formulation presented below was proposed by Pereira and Averbakh [6]. As
the one for R-RSP presented in Section 5.1.1, this formulation is based on the modeling technique
for interval min-max regret problems in general discussed in Section 3. For the specific case of
RSC, Theorem 1 can be stated as follows.
Proposition 3. Given a covering X ∈ Λ, the regret of K is maximum in the scenario induced by
K, where the costs of all the columns of K are in their corresponding upper bounds and the costs
of all the other columns are in their lower bounds.
Now, consider the decision variables y on the choice of columns belonging or not to a robust
covering, such that y j = 1 if the column j ∈ J belongs to the solution; y j = 0, otherwise. Also
let ρ be a continuous variable. The MILP formulation of [6] is provided from (83) to (87).
min
X
u jy j − ρ
X
oi j y j ≥ 1
ρ≤
X
(83)
j∈J
s.t.
∀ i ∈ I,
(84)
j∈J
(l j + (u j − l j )y j ) x̄ j
∀ x̄ ∈ Λ,
(85)
j∈J
y j ∈ {0, 1} ∀ j ∈ J,
(86)
ρ free.
(87)
The objective function in (83) considers Proposition 3 and gives the robustness cost of a robust
covering. Constraints (84) and (86) ensure that the y variables represent a covering, whereas
constraints (85) guarantee that ρ does not exceed the cost of an optimal covering in the scenario
induced by the covering defined by y. In (85), x̄ is a constant vector, one for each possible
covering. Constraint (87) gives the domain of the variable ρ.
As for R-RSP, we use the logic-based Benders’ algorithm discussed in the Section 3.1 to solve
the formulation described above.
5.2.2. An LP based heuristic for RSC
Consider the following ILP formulation used to model the classical SC in a given cost scenario
s ∈ S. Here, the binary variables x define an optimal covering K ∗ (s) ∈ Λ in s, such that x j = 1 if
26
the column j ∈ J belongs to K ∗ (s), and x j = 0, otherwise.
(I2 )
min
X
c sj x j
(88)
X
oi j x j ≥ 1 ∀ i ∈ I,
(89)
j∈J
s.t.
j∈J
x j ∈ {0, 1} ∀ j ∈ J.
(90)
The objective function in (88) gives the cost, in the scenario s, of the covering defined by x,
while constraints (89) and (90) ensure that x identifies a covering in Λ. Relaxing the integrality
on x, we obtain:
(L2 )
min
X
c sj x j
(91)
j∈J
s.t.
Constraint (89),
xj ≥ 0
∀ j ∈ J.
(92)
Notice that we omitted the domain constraints x j ≤ 1 for all j ∈ J, since they are redundant
in this case. Considering formulations I2 , L2 and the generic model R̂, given by (7), (19), (20),
(23) and (24), we can devise a heuristic formulation for RSC. Now, the y variables represent a
heuristic solution covering, and the variables {λi : i ∈ I} are the ones of the dual problem related
to L2 .
(H2 )
min
X
j∈J
s.t.
X !
u jy j −
λi
(93)
i∈I
X
oi j y j ≥ 1 ∀ i ∈ I,
X
oi j λi ≤ l j + (u j − l j )y j
(94)
j∈J
∀ j ∈ J,
(95)
i∈I
λi ≥ 0 ∀ i ∈ I,
(96)
y j ∈ {0, 1} ∀ j ∈ J.
(97)
According to Proposition 1, the objective function in (93) gives an upper bound on the robustness cost of a robust solution for RSC. Constraints (94) and (97) ensure that y defines a covering
in Λ, while constraints (95) are the ones related to the dual of L2 . Notice that constraints (95)
consider the cost of each column j ∈ J in the scenario induced by the covering identified by the
y variables. Restrictions (96) give the domain of the dual variables λ.
27
In the case of RSC, the LP based heuristic consists in solving formulation H2 and, then,
computing the robustness cost of the solution obtained.
6. Computational experiments
In this section, we evaluate, out of computational experiments, the effectiveness and the time
efficiency of the proposed heuristic at solving the two interval robust-hard problems considered
in this study. For short, the Logic-Based Benders’ decomposition algorithm is referred to as
LB-Benders’, whereas the LP based Heuristic is referred to as LPH. LB-Benders’, LPH and
the 2-approximation heuristic for interval min-max regret problems, namely AMU [23], were
implemented in C++, along with the optimization solver ILOG CPLEX 12.5. The computational
experiments were performed on a 64 bits Intel R Xeon R E5405 machine with 2.0 GHz and 7.0
GB of RAM, under Linux operating system. LB-Benders’ was set to run for up to 3600 seconds
of wall-clock time.
6.1. The Restricted Robust Shortest Path problem (R-RSP)
In all of the algorithms implemented, whenever a classical R-SP instance had to be solved,
we used CPLEX to handle the ILP formulation I1 , defined by (60)-(65), directly. We also used
CPLEX to solve each master problem in LB-Benders’ and the heuristic formulation H1 , defined
by (77)-(82)
6.1.1. Benchmarks description
Due to the lack of R-RSP instances in the literature, we generated two benchmarks of instances
inspired by the applications described in Section 5.1. These benchmarks were adapted from
two sets of RSP instances: Karaşan [16] and Coco [53] instances, which model, respectively,
telecommunications and urban transportation networks.
Karaşan instances have been largely used in experiments concerning RSP [8, 16, 48, 53, 54].
They consist of layered [55] and acyclic [56] digraphs. In these digraphs, each of the κ layers
has the same number ω of vertices. There is an arc from every vertex in a layer b ∈ {1, . . . , κ − 1}
to every vertex in the adjacent layer b + 1. Moreover, there is an arc from the origin o to every
vertex in the first layer, and an arc from every vertex in the layer κ to the destination vertex t.
These instances are named K-v-Φmax -δ-ω, where v is the number of vertices (aside from o and
28
t), Φmax is an integer constant, and 0 < δ < 1 is a continuous value. The arc cost intervals were
generated as follows. For each arc (i, j) ∈ A, a random integer value Φi j was uniformly chosen
in the range [1, Φmax ]. Afterwards, random integer values li j and ui j were uniformly selected,
respectively, in the ranges [(1 − δ) · Φi j , (1 + δ) · Φi j ] and [li j , (1 + δ) · Φi j ]. Note that Φ plays
the role of a base-case scenario, and δ determines the degree of uncertainty. Figure 3 shows an
example of an acyclic digraph with 3 layers of width 2.
1
3
5
0
7
2
4
6
Figure 3: An acyclic digraph with 3 layers of width 2. Here, o = 0 and t = 7.
Coco instances consist of grid digraphs based on n × m matrices, where n is the number of
rows and m is the number of columns. Each matrix cell corresponds to a vertex in the digraph,
and there are two bidirectional arcs between each pair of vertices whose respective matrix cells
are adjacent. The origin o is defined as the upper left vertex, and the destination t is defined as
the lower right vertex. These instances are named G-n×m-Φmax -δ, with 0 < δ < 1, where Φmax
is an integer value. Given Φmax and δ values, the cost intervals are generated as in the Karaşan
instances. Figure 4 gives an example of a grid digraph.
0
1
2
3
4
5
6
7
Figure 4: A 2 × 4 grid digraph, with o = 0 and t = 7.
For all instances, the resource consumption associated with each arc is given by a random
integer value uniformly selected in the interval (0, 10]. The small interval amplitude allows
the generation of instances in which most of the arcs are candidate to appear in an optimal
solution, increasing the number of feasible solutions. The symmetry with respect to arc resource
consumptions was preserved, i.e., we considered di j = d ji for any pair of adjacent vertices i and
j such that (i, j) ∈ A and ( j, i) ∈ A.
The resource consumption limit β of a given instance was computed as follows. Consider
29
the set P of all the paths from o to t, and let p̄ ∈ P be a shortest path in terms of resource
consumption, i.e., p̄ = arg min D p . We set β = 1.1·D p̄ , which means that is given a 10% tolerance
p∈P
with respect to the minimum resource consumption D p̄ . This way, the resource consumption limit
is tighter.
We generated Karaşan and Coco instances of 1000 and 2000 vertices, with Φmax ∈ {20, 200},
δ ∈ {0.5, 0.9} and ω ∈ {5, 10, 25}. Considering these values, a group of 10 instances was generated for each possible parameters configuration. In summary, 480 instances were used in the
experiments.
6.1.2. Results
Computational experiments were carried out in order to evaluate if the proposed heuristic
efficiently finds optimal or near-optimal solutions for the two benchmarks of instances described
above. Results for Karaşan and Coco instances are reported in Tables 1 and 2, respectively. The
first column displays the name of each group of 10 instances. The second and third columns
show, respectively, the number of instances solved at optimality by LB-Benders’ within 3600
seconds, and the average wall-clock processing time (in seconds) spent in solving these instances.
If no instance in the group was solved at optimality, this entry is filled with a dash. The fourth and
fifth columns show, respectively, the average and the standard deviation (over the 10 instances)
of the relative optimality gaps given by 100 ·
UBb −LBb
UBb .
Here, LBb and U Bb are, respectively, the
best lower and upper bounds obtained by LB-Benders’ for a given instance. The sixth column
displays the average wall-clock processing time (in seconds) of AMU. The seventh column shows
the average (over the 10 instances) of the relative gaps given by 100 ·
UBamu −LBb
UBamu ,
where U Bamu
is the best upper bound obtained by AMU for a given instance. The standard deviation of these
gaps is given in the eighth column. Likewise, the ninth column shows the average wall-clock
processing time of LPH, and the last two columns give the average and the standard deviation
(over the 10 instances) of the gaps given by 100 ·
UBlph −LBb
UBlph .
Here, U Blph is the β-restricted
robustness cost of the solution obtained by LPH for a given instance.
Notice that the gaps referred to the solutions obtained by AMU and LPH consider the best
lower bounds obtained by LB-Benders’ (within 3600 seconds of execution), which might not
correspond to the cost of optimal solutions. Thus, the aforementioned gaps may overestimate the
actual gaps between the cost of the solutions obtained and the cost of optimal ones.
30
Table 1: Computational results for the layered and acyclic digraph instances.
LB-Benders’
Test set
#opt
Time (s)
AvgGAP (%)
AMU
StDev (%)
Time (s)
AvgGAP (%)
LPH
StDev (%)
Time (s)
AvgGAP (%)
StDev (%)
31
K-1000-20-0.5-5
8
1515.25
0.11
0.24
3.13
4.21
3.59
10.33
0.11
0.24
K-1000-20-0.9-5
2
2108.61
7.47
5.29
3.12
10.06
4.21
28.84
5.79
3.75
K-1000-200-0.5-5
10
1165.08
0.00
0.00
3.13
3.68
2.18
8.52
0.16
0.28
K-1000-200-0.9-5
1
1919.16
4.85
3.30
3.07
9.77
1.83
21.36
3.80
2.57
K-1000-20-0.5-10
10
83.30
0.00
0.00
4.15
3.06
2.70
11.23
0.07
0.22
K-1000-20-0.9-10
10
218.69
0.00
0.00
4.17
4.52
3.20
17.75
0.00
0.00
K-1000-200-0.5-10
10
42.54
0.00
0.00
3.96
1.57
1.84
10.45
0.31
0.97
K-1000-200-0.9-10
10
413.43
0.00
0.00
3.94
4.50
2.98
24.44
0.05
0.15
K-1000-20-0.5-25
10
17.61
0.00
0.00
7.35
3.12
4.52
21.55
0.00
0.00
K-1000-20-0.9-25
10
32.22
0.00
0.00
7.51
2.35
3.69
26.97
0.24
0.76
K-1000-200-0.5-25
10
18.17
0.00
0.00
7.25
1.27
2.71
23.67
0.00
0.00
K-1000-200-0.9-25
10
41.04
0.00
0.00
7.33
1.13
1.89
30.92
0.00
0.00
K-2000-20-0.5-5
0
0.00
14.47
5.89
10.61
14.30
4.06
64.61
10.43
3.87
K-2000-20-0.9-5
0
0.00
25.01
2.69
10.68
21.35
2.17
241.13
17.94
2.38
K-2000-200-0.5-5
0
0.00
14.45
3.10
10.38
14.15
2.87
69.62
10.70
2.21
K-2000-200-0.9-5
0
0.00
25.77
3.39
10.43
22.13
3.17
454.79
18.15
2.79
K-2000-20-0.5-10
8
1297.89
0.91
2.23
13.01
4.37
3.69
92.04
0.89
1.99
K-2000-20-0.9-10
0
0.00
7.34
4.06
12.75
9.46
4.30
252.91
5.91
3.45
K-2000-200-0.5-10
4
846.44
1.37
2.30
12.34
4.12
2.94
112.94
1.04
1.66
K-2000-200-0.9-10
0
0.00
5.99
2.44
12.39
8.78
3.54
211.80
4.77
2.31
K-2000-20-0.5-25
10
155.30
0.00
0.00
21.72
5.98
7.21
122.22
0.00
0.00
K-2000-20-0.9-25
10
408.79
0.00
0.00
21.60
1.46
1.54
222.00
0.07
0.22
K-2000-200-0.5-25
10
138.88
0.00
0.00
21.88
1.53
2.11
121.82
0.04
0.13
K-2000-200-0.9-25
10
572.74
0.00
0.00
21.18
2.70
2.76
215.21
0.02
0.05
4.49
1.46
6.65
3.15
3.35
1.25
Average
Table 2: Computational results for the grid digraph instances.
LB-Benders’
AvgGAP (%)
StDev (%)
Time (s)
AvgGAP (%)
LPH
32
Test set
#opt
G-32x32-20-0.5
10
7.31
0.00
0.00
3.21
4.16
9.20
3.80
0.22
0.70
G-32x32-20-0.9
10
8.63
0.00
0.00
3.43
5.38
7.49
5.01
0.69
1.67
G-32x32-200-0.5
10
6.83
0.00
0.00
3.35
1.56
2.72
3.88
3.11
4.72
G-32x32-200-0.9
10
9.79
0.00
0.00
3.22
3.90
5.89
5.33
0.00
0.00
G-20x50-20-0.5
10
6.66
0.00
0.00
2.86
2.71
5.94
3.96
0.42
1.34
G-20x50-20-0.9
10
9.14
0.00
0.00
3.00
1.80
2.93
4.76
0.76
1.10
G-20x50-200-0.5
10
6.23
0.00
0.00
3.24
2.54
5.09
4.47
0.78
2.45
G-20x50-200-0.9
10
13.43
0.00
0.00
3.15
2.33
4.43
5.43
0.58
0.74
G-5x200-20-0.5
10
397.06
0.00
0.00
2.92
5.29
4.89
10.97
0.10
0.31
G-5x200-20-0.9
9
670.84
0.28
0.89
2.91
4.64
1.69
19.83
0.36
0.87
G-5x200-200-0.5
10
153.69
0.00
0.00
2.97
4.82
3.57
8.75
0.29
0.66
G-5x200-200-0.9
8
1059.09
0.20
0.50
3.01
8.00
3.73
22.34
0.32
0.57
G-44x44-20-0.5
10
23.52
0.00
0.00
10.36
0.97
1.92
13.15
0.81
1.42
G-44x44-20-0.9
10
29.43
0.00
0.00
10.32
3.92
8.06
15.50
0.55
1.04
G-44x44-200-0.5
10
23.00
0.00
0.00
9.77
5.05
4.02
14.58
0.00
0.00
G-44x44-200-0.9
10
38.13
0.00
0.00
9.53
2.70
3.23
19.98
0.13
0.22
G-20x100-20-0.5
10
34.77
0.00
0.00
10.90
2.81
5.26
20.38
0.00
0.00
G-20x100-20-0.9
10
90.28
0.00
0.00
10.76
5.41
4.98
29.39
0.61
1.13
G-20x100-200-0.5
10
47.27
0.00
0.00
11.39
1.64
1.97
20.17
0.25
0.66
G-20x100-200-0.9
10
59.16
0.00
0.00
10.53
3.37
3.27
26.32
0.21
0.37
G-5x400-20-0.5
1
3444.71
2.57
1.98
10.26
6.82
3.90
62.50
2.14
1.62
G-5x400-20-0.9
0
0.00
7.76
3.95
10.20
10.83
3.16
150.45
5.67
2.89
G-5x400-200-0.5
1
2719.17
5.24
4.32
10.27
8.67
3.59
79.37
3.94
3.17
G-5x400-200-0.9
0
0.00
11.50
3.45
9.93
13.27
3.75
249.81
8.19
2.24
1.15
0.63
4.69
4.36
1.26
1.24
Average
Time (s)
AMU
StDev (%)
Time (s)
AvgGAP (%)
StDev (%)
With respect to Karaşan instances (Table 1), the average gaps referred to the solutions provided by LB-Benders’, AMU and LPH are up to, respectively, 7.47%, 10.06% and 5.79% for
the instances with 1000 vertices (see K-1000-20-0.9-5). For the instances with 2000 vertices,
the average gaps referred to the solutions provided by LB-Benders’, AMU and LPH are up to,
respectively, 25.77%, 22.13% and 18.15% (see K-2000-200-0.9-5). Notice that the average gaps
of AMU over all Karaşan instances is 6.65%, while that of LPH is only 3.35%. In fact, the
average gaps of the solutions provided by LPH are smaller than those of AMU for all the sets
of instances considered. It can also be observed that the smaller the value of ω is, the larger the
average gaps achieved by the three algorithms are. Furthermore, for the hardest instances (with
ω = 5), the average gaps of the solutions provided by LPH are smaller than or equal to those of
LB-Benders’ (except for K-1000-200-0.5-5).
Regarding Coco instances (Table 2), it can be seen that the average optimality gaps referred
to the solutions provided by LB-Benders’, AMU and LPH are at most, respectively, 11.50%,
13.27% and 8.19% (see G-5x400-200-0.9). Also in this case, the average gaps of the solutions
provided by LPH are smaller than those of AMU for all the sets of instances (except for G32x32-200-0.5). The average optimality gap of LB-Benders’ over all Coco instances is very
small (1.15%), while that of LPH is very close to this value (1.26%). It can also be observed
that the instances based on 5x400 grids are much harder to solve than the other grid instances.
Moreover, the average relative gaps referred to the solutions provided by LPH is always smaller
than those of LB-Benders’ for the hardest grid instances.
For both benchmarks, LPH clearly outperforms AMU in terms of the quality of the solutions
obtained. In fact, the heuristic is even able to achieve better upper bounds than LB-Benders’ at
solving some instance sets, especially the hardest ones. Furthermore, LPH required significantly
smaller computational effort than LB-Benders’. Therefore, LPH arises as an effective and time
efficient heuristic to solve Karaşan and Coco instances. One may note that, although the bounds
provided by AMU are not as good as those given by LPH, the 2-approximation procedure performs within tiny average computational times when compared with LB-Benders’, especially
for the hardest instances. Thus, the results suggest that AMU should be considered in the cases
where time efficiency is preferable to the quality of the solutions.
Our computational results also suggest that, for both benchmarks, instances generated with a
higher degree of uncertainty (in particular, δ = 0.9) tend to become more difficult to be handled
33
by LB-Benders’ and LPH. As pointed out in related works [11, 16], high δ values can increase
the occurrence of overlapping cost intervals and, thus, decrease the number of dominated arcs.
As a consequence, the task of finding a robust path becomes more difficult. Such behaviour does
not apply to AMU, as it simply solves R-SP instances in specific scenarios that do not rely on
the amplitude of the cost intervals.
In Tables 3a and 3b, we detail the improvement of LPH over AMU in terms of the bounds
achieved for Karaşan and Coco instances, respectively. The first column displays the name of
each instance set, and the second one shows the number of times (out of 10) that LPH gives
better bounds than AMU. The third and fourth columns show, respectively, the minimum and the
maximum percentage improvements given by 100 ·
UBamu −UBlph
,
UBamu
while the last two columns give
the average and the standard deviation (over the 10 instances in each set) of these values. Notice
that the percentage improvement assumes a negative value whenever AMU gives a better bound
than LPH.
Regarding Karaşan instances (Table 3a), the percentage improvement is up to 21.47% (see
K-2000-20-0.5-25). In fact, AMU only overcomes LPH on two instances from the whole benchmark, one from K-1000-20-0.9-25 and other from K-2000-200-0.9-25. For the grid digraph
instances (Table 3b), the percentage improvement is up to 29.31% (see G-32x32-20-0.5). Moreover, AMU gives better bounds than LPH only on 13 out of the 240 instances tested. We could
not identify a precise pattern on these instances that might indicate why AMU performs better.
However, most of them belong to instance sets based on square-like grids (such as 32x32 and
44x44). Thus, we conjecture that the balance between the number of lines and columns in the
corresponding grids plays a role in such behaviour.
34
Table 3: Details on the improvement of LPH over AMU for the two benchmarks of R-RSP instances.
(b) Grid digraph instances.
(a) Layered and acyclic digraph instances.
LPH Improvement
Test set
#Better
Min (%)
Max (%)
LPH Improvement
Avg (%)
StDev (%)
Test set
#Better
Min (%)
Max (%)
Avg (%)
StDev (%)
9.17
35
K-1000-20-0.5-5
10
0.00
9.79
4.10
3.59
G-32x32-20-0.5
10
0.00
29.31
3.94
K-1000-20-0.9-5
10
0.60
9.62
4.51
3.22
G-32x32-20-0.9
9
-5.49
20.17
4.69
7.95
K-1000-200-0.5-5
10
0.68
8.15
3.52
2.26
G-32x32-200-0.5
7
-11.21
0.00
-1.73
3.71
K-1000-200-0.9-5
10
2.06
12.49
6.14
3.28
G-32x32-200-0.9
10
0.00
18.21
3.90
5.89
K-1000-20-0.5-10
10
0.00
7.79
2.99
2.56
G-20x50-20-0.5
10
0.00
18.89
2.29
5.97
K-1000-20-0.9-10
10
0.00
12.10
4.52
3.20
G-20x50-20-0.9
9
-2.40
8.89
1.03
3.14
K-1000-200-0.5-10
10
0.00
4.18
1.27
1.51
G-20x50-200-0.5
9
-8.41
15.57
1.70
6.14
K-1000-200-0.9-10
10
1.33
12.16
4.45
2.97
G-20x50-200-0.9
9
-1.04
12.78
1.77
4.13
K-1000-20-0.5-25
10
0.00
11.90
3.12
4.52
G-5x200-20-0.5
10
0.00
14.34
5.20
4.89
K-1000-20-0.9-25
9
-2.47
10.39
2.10
3.94
G-5x200-20-0.9
10
2.58
8.01
4.29
1.75
K-1000-200-0.5-25
10
0.00
7.27
1.27
2.71
G-5x200-200-0.5
10
0.00
11.14
4.54
3.51
K-1000-200-0.9-25
10
0.00
5.68
1.13
1.89
G-5x200-200-0.9
10
1.22
12.15
7.70
3.82
K-2000-20-0.5-5
10
0.41
9.88
4.29
3.20
G-44x44-20-0.5
8
-3.95
5.77
0.15
2.40
K-2000-20-0.9-5
10
1.34
8.01
4.13
2.09
G-44x44-20-0.9
9
-3.29
23.26
3.40
7.95
K-2000-200-0.5-5
10
0.39
7.15
3.86
2.21
G-44x44-200-0.5
10
0.00
11.05
5.05
4.02
K-2000-200-0.9-5
10
1.61
8.25
4.86
2.22
G-44x44-200-0.9
9
-0.47
7.45
2.56
3.32
K-2000-20-0.5-10
10
0.00
9.80
3.52
2.81
G-20x100-20-0.5
10
0.00
14.09
2.81
5.26
K-2000-20-0.9-10
10
1.48
6.72
3.80
1.79
G-20x100-20-0.9
10
0.00
11.06
4.85
4.25
K-2000-200-0.5-10
10
0.00
5.66
3.12
1.95
G-20x100-200-0.5
10
0.00
5.80
1.40
1.80
K-2000-200-0.9-10
10
1.04
8.19
4.23
2.31
G-20x100-200-0.9
8
-0.66
7.65
3.17
3.38
K-2000-20-0.5-25
10
0.00
21.57
5.98
7.21
G-5x400-20-0.5
10
1.27
10.58
4.80
2.87
K-2000-20-0.9-25
10
0.00
3.91
1.39
1.60
G-5x400-20-0.9
10
1.59
13.90
5.42
3.65
K-2000-200-0.5-25
10
0.00
6.46
1.49
2.14
G-5x400-200-0.5
10
0.16
10.92
4.88
3.63
K-2000-200-0.9-25
9
-0.14
7.20
2.68
2.78
G-5x400-200-0.9
10
1.72
11.72
5.53
3.26
6.2. The Robust Set Covering problem (RSC)
In all of the algorithms regarding RSC, we used CPLEX to handle the ILP formulation I2 ,
defined by (88)-(90), whenever a classical SC instance had to be solved. We also used CPLEX
to solve each master problem in LB-Benders’ and the heuristic formulation H2 , defined by (93)(97).
6.2.1. Benchmarks description
In our experiments, we considered three benchmarks of instances from the literature of RSC,
namely Beasley, Montemanni and Kasperski-Zieliński benchmarks [6]. The three of them are
based on classical SC instances from the OR-Library [57]. However, the way the column cost
intervals are generated differs from benchmark to benchmark.
Regarding Beasley instances, let Φ j represent the cost of a column j ∈ J in the original SC
instance, and let 0 < δ < 1 be a continuous value used to control the level of uncertainty referred
to an RSC instance. For each j ∈ J , the corresponding cost interval [l j , u j ] is generated by
uniformly selecting random integer values l j and u j in the ranges [(1 − δ) · Φ j , Φ j ] and [Φ j , (1 +
δ) · Φ j ], respectively. These instances are named B.<SCinst>-δ, where <SCinst> stands for the
name of the original SC instance set considered. For each original instance of the classical SC,
we considered three RSC instances, one for each value δ ∈ {0.1, 0.3, 0.5}. In total, 75 instances
from Beasley benchmark were used in our experiments.
In Montemanni instances, the column costs of the original SC instances are discarded, and,
for each column j ∈ J , the corresponding cost interval [l j , u j ] is generated as follows. First, a
random integer value u j is uniformly chosen in the range [0, 1000], and, then, a random integer
value l j is uniformly selected in the range [0, u j]. These instances are named M.<SCinst>-1000,
where <SCinst> is the name of the original SC instance set used. Each original SC instance
considered gives the backbone to generate three RSC instances, and a total of 75 instances from
Montemanni benchmark were used in our experiments.
Kasperski-Zieliński benchmark also consider classical SC instances without the original column costs. In these instances, the cost interval [l j , u j ] of each column j ∈ J is generated as
follows. First, a random integer value l j is uniformly chosen in the range [0, 1000], and, then,
a random integer value u j is uniformly selected in the range [l j , l j + 1000]. These instances are
named KZ.<SCinst>-1000, where <SCinst> is the name of the original SC instance set used.
Each of the SC instances considered gives the backbone to generate three RSC instances, and
36
a total of 75 instances from Kasperski-Zieliński benchmark were used in our experiments. In
summary, 225 RSC instances were considered in the experiments.
6.2.2. Results
Table 4 shows the computational results regarding the three benchmarks of RSC instances
considered. The first column displays the name of each instance set, while the second one gives
(i) the number of instances solved at optimality by LB-Benders’ within 3600 seconds, over (ii)
the cardinality of the corresponding instance set. The average wall-clock processing time (in
seconds) spent in solving these instances at optimality are given in the third column. This entry
is filled with a dash if no instance in the group was solved at optimality. The fourth and fifth
columns show, respectively, the average and the standard deviation (over all the instances in
each set) of the relative optimality gaps given by 100 ·
UBb −LBb
UBb .
Recall that LBb and U Bb are,
respectively, the best lower and upper bounds obtained by LB-Benders’ for a given instance.
The sixth column displays the average wall-clock processing time (in seconds) of AMU. The
seventh column shows the average (over all the instances in each set) of the relative gaps given
by 100 ·
UBamu −LBb
UBamu ,
where U Bamu is the best upper bound obtained by AMU for a given instance.
The standard deviation of these gaps is given in the eighth column. Likewise, the ninth column
shows the average wall-clock processing time of LPH, and the last two columns give the average
and the standard deviation (once again, over all the instances in each set) of the gaps given by
100 ·
UBlph −LBb
UBlph .
Accordingly, U Blph is the robustness cost of the solution obtained by LPH for a
given instance.
As for R-RSP, the gaps referred to the solutions obtained by AMU and LPH consider the (not
necessarily optimal) lower bounds obtained by LB-Benders’ within 3600 seconds of execution.
Thus, these gaps may overestimate the actual gaps between the cost of the solutions obtained and
the cost of optimal ones.
From the results, it can be seen that the average gaps of the solutions provided by LPH are
smaller than those referred to AMU for all the sets of instances tested. Furthermore, for the
hardest instances (Kasperski-Zieliński benchmark), the average gaps of the solutions provided
by LPH are even smaller than those of LB-Benders’. As discussed in [6], Kasperski-Zieliński
instances are especially challenging, mostly because of the high probability of overlap between
arbitrary cost intervals in these instances.
37
Table 4: Computational results for the three benchmarks of RSC instances.
LB-Benders’
Time (s)
AvgGAP (%)
AMU
StDev (%)
Time (s)
AvgGAP (%)
LPH
Beasley set
#opt
StDev (%)
Time (s)
AvgGAP (%)
StDev (%)
B.scp4-0.1
10/10
1.03
0.00
0.00
0.23
4.24
4.92
0.21
0.59
1.86
B.scp5-0.1
10/10
3.32
0.00
0.00
0.34
9.01
8.27
0.42
2.60
3.62
38
B.scp6-0.1
5/5
2.23
0.00
0.00
0.88
8.02
7.41
0.64
1.43
3.19
B.scp4-0.3
10/10
9.82
0.00
0.00
0.24
5.24
3.46
0.44
0.98
1.12
B.scp5-0.3
10/10
28.23
0.00
0.00
0.34
6.21
2.61
0.92
1.78
2.62
B.scp6-0.3
5/5
12.18
0.00
0.00
1.08
2.44
2.70
1.05
1.63
1.51
B.scp4-0.5
10/10
93.98
0.00
0.00
0.29
3.56
2.12
0.64
1.20
1.49
B.scp5-0.5
10/10
94.20
0.00
0.00
0.34
5.38
4.33
1.21
0.87
1.00
B.scp6-0.5
5/5
22.57
0.00
0.00
1.00
4.08
4.21
1.16
1.44
2.23
0.00
0.00
5.35
4.45
1.39
2.07
Average
LB-Benders’
Montemanni set
StDev (%)
Time (s)
Time (s)
M.scp4-1000
27/30
598.94
0.04
0.14
0.23
1.13
0.78
0.61
0.05
0.16
M.scp5-1000
30/30
481.98
0.00
0.00
0.21
1.00
0.90
0.70
0.01
0.04
M.scp6-1000
15/15
13.90
0.00
0.00
0.42
0.78
0.94
0.80
0.06
0.14
0.01
0.05
0.97
0.87
0.04
0.11
LB-Benders’
Kasperski-Zieliński set
#opt
K.scp4-1000
0/30
K.scp5-1000
0/30
K.scp6-1000
3/15
Average
Time (s)
AvgGAP (%)
LPH
#opt
Average
AvgGAP (%)
AMU
StDev (%)
Time (s)
AMU
AvgGAP (%)
StDev (%)
0.00
14.25
2.93
0.00
8.63
3.45
1881.73
2.95
3.13
8.61
3.17
Time (s)
AvgGAP (%)
StDev (%)
LPH
AvgGAP (%)
StDev (%)
Time (s)
AvgGAP (%)
StDev (%)
3.08
14.27
2.73
124.86
12.90
2.62
2.90
8.63
3.45
44.93
7.99
3.01
11.96
4.04
2.90
115.31
2.77
2.82
8.98
3.03
7.89
2.82
Notice that LPH clearly outperforms AMU in terms of the quality of the solutions obtained
and requires significantly smaller computational effort than LB-Benders’. In fact, for most of the
instance sets considered, the average execution times of LPH are comparable to those of AMU.
Therefore, LPH arises as an effective and time efficient heuristic in solving the three benchmarks
of RSC instances tested. Nevertheless, we highlight that AMU should still be considered in the
cases where time efficiency is preferable to the quality of the solutions, especially for the more
challenging instances.
Table 5: Details on the improvement of LPH over AMU for the three benchmarks of RSC instances.
LPH Improvement
Test set
#Better
Min (%)
Max (%)
Avg (%)
StDev (%)
B.scp4-0.1
10/10
0.00
12.50
3.66
5.05
B.scp5-0.1
10/10
0.00
20.83
6.53
8.33
B.scp6-0.1
5/5
0.00
15.00
6.69
6.79
B.scp4-0.3
10/10
0.00
11.76
4.29
3.57
2.98
B.scp5-0.3
10/10
0.00
8.06
4.47
B.scp6-0.3
4/5
-2.50
6.52
0.80
3.37
B.scp4-0.5
10/10
0.52
5.36
2.39
1.42
B.scp5-0.5
9/10
-1.52
13.64
4.52
5.06
B.scp6-0.5
5/5
0.00
6.58
2.70
2.94
M.scp4-1000
30/30
0.17
3.30
1.08
0.79
M.scp5-1000
30/30
0.00
3.51
0.99
0.88
M.scp6-1000
15/15
0.00
2.87
0.72
0.94
KZ.scp4-1000
29/30
-0.04
4.59
1.56
1.33
KZ.scp5-1000
29/30
-0.08
3.13
0.71
0.80
KZ.scp6-1000
14/15
-0.13
4.25
1.29
1.61
In Table 5, we detail the improvement of LPH over AMU in terms of the bounds achieved
for the three benchmarks of RSC instances. The first column displays the name of each instance
set, and the second one shows (i) the number of times LPH gives better bounds than AMU,
over (ii) the cardinality of the corresponding instance set. The third and fourth columns show,
respectively, the minimum and the maximum percentage improvements given by 100·
UBamu −UBlph
,
UBamu
while the last two columns give the average and the standard deviation (over all the instances in
each set) of these values. These percentage improvements assume negative values whenever
39
AMU gives better bounds than LPH.
Regarding Beasley instances, the percentage improvement is up to 20.83% (see B.scp5-0.1).
In fact, AMU only overcomes LPH on two instances from the whole benchmark, one from
B.scp6-0.3 and other from B.scp5-0.5. With respect to Montemanni instances, the percentage improvement is up to 3.51% (see M.scp5-1000), and LPH always gives better bounds than
AMU. Regarding Kasperski-Zieliński instances, the percentage improvement is at most 4.59%
(see KZ.scp4-1000), and AMU overcomes LPH on three instances, one from each instance set.
We could not identify any specific pattern on the instances for which AMU outperforms LPH.
However, such behaviour is only verified in 5 out of the 225 instances tested.
7. Concluding remarks
In this study, we proposed a novel heuristic approach to address the class of interval robusthard problems, which is composed of robust optimization versions of classical NP-hard combinatorial problems. We applied this new technique to two interval robust-hard problems, namely the
Restricted Robust Shortest Path problem (R-RSP) and the Robust Set Covering problem (RSC).
To our knowledge, the former problem was also introduced in this study.
In order to evaluate the quality of the solutions obtained by our heuristic (namely LPH), we
adapted to R-RSP and to RSC a logic-based Benders’ decomposition algorithm (referred to as
LB-Benders’), which is a state-of-the-art exact method to solve interval data min-max regret
problems in general. We also compared the results obtained by the proposed heuristic with a
widely used 2-approximation procedure, namely AMU.
Regarding R-RSP, the computational experiments showed the heuristic’s effectiveness in solving the two benchmarks of instances considered. For the layered and acyclic digraph instances,
the average relative gaps of LB-Benders’ and AMU were, respectively, 4.49% and 6.65%, while
that of LPH was only 3.35%. For the grid digraph instances, the average relative gaps of LBBenders’, AMU and LPH were, respectively, 1.15%, 4.69% and 1.26%. Moreover, the average
processing times of LPH were much smaller than those of LB-Benders’ for both benchmarks
considered and remain competitive with those referred to AMU for most of the instances. In addition, LPH was able to provide better bounds than AMU for 465 out of the 480 R-RSP instances
tested.
With respect to RSC, the results show that the average gaps of the solutions provided by LPH
40
were smaller than those referred to AMU for all the sets of instances from the three benchmarks
used in the experiments. Moreover, LPH achieved better primal bounds than LB-Benders’ for the
hardest instances (Kasperski-Zieliński benchmark) and was able to provide better bounds than
AMU for 220 out of the 225 RSC instances tested.
The results point out to the fact that the proposed heuristic framework may also be efficient in
finding optimal or near-optimal solutions for other interval robust-hard problems, which can be
explored in future studies. Other directions of future work remain open. For instance, research
can be conducted to close the conjecture that the heuristic framework can be applied to the
wider class of interval data min-max regret problems with compact constraint sets addressed in
[25, 26]. Future works can also add to our framework local search strategies, such as Local
Branching [58], to further improve the quality of the solutions obtained by the heuristic.
Acknowledgments
This work was partially supported by the Brazilian National Council for Scientific and Technological Development (CNPq), the Foundation for Support of Research of the State of Minas
Gerais, Brazil (FAPEMIG), and Coordination for the Improvement of Higher Education Personnel, Brazil (CAPES).
References
[1] P. Kouvelis, G. Yu, Robust Discrete Optimization and Its Applications, Kluwer Academic Publishers, Boston, 1997.
[2] J. C. Spall, Introduction to Stochastic Search and Optimization: Estimation, Simulation and Control, Wiley, New
York, 2003.
[3] I. Averbakh, Computing and minimizing the relative regret in combinatorial optimization with interval data, Discrete Optimization 2 (4) (2005) 273–287.
[4] R. Montemanni, A Benders decomposition approach for the robust spanning tree problem with interval data, European Journal of Operational Research 174 (3) (2006) 1479–1490.
[5] R. Montemanni, J. Barta, L. M. Gambardella, The robust traveling salesman problem with interval data, Transportation Science 41 (3) (2007) 366–381.
[6] J. Pereira, I. Averbakh, Exact and heuristic algorithms for the interval data robust assignment problem, Computers
& Operations Research 38 (8) (2011) 1153–1163.
[7] H. Yaman, O. E. Karaşan, M. Ç. Pinar, The robust spanning tree problem with interval data, Operations Research
Letters 29 (1) (2001) 31–40.
[8] R. Montemanni, L. M. Gambardella, The robust shortest path problem with interval data via Benders decomposition, 4OR 3 (4) (2005) 315–328.
41
[9] A. Kasperski, Discrete Optimization with Interval Data: Minmax Regret and Fuzzy Approach (Studies in Fuzziness
and Soft Computing), Springer Berlin, 2008.
[10] V. G. Deineko, G. J. Woeginger, Pinpointing the complexity of the interval min-max regret knapsack problem,
Discrete Optimization 7 (4) (2010) 191–196.
[11] J. Pereira, I. Averbakh, The robust set covering problem with interval data, Annals of Operations Research 207 (1)
(2013) 217–235.
[12] M. R. Garey, D. S. Johnson, Computers and Intractability: A Guide to the Theory of NP-Completeness, W. H.
Freeman & Co., New York, NY, USA, 1979.
[13] G. Handler, I. Zang, A dual algorithm for the constrained shortest path problem, Networks 10 (4) (1980) 293–309.
[14] H. Aissi, C. Bazgan, D. Vanderpooten, Minmax and minmax regret versions of combinatorial optimization problems: A survey, European Journal of Operational Research 197 (2) (2009) 427–438.
[15] I. Averbakh, On the complexity of a class of combinatorial optimization problems with uncertainty, Mathematical
Programming 90 (2) (2001) 263–272.
[16] O. E. Karaşan, M. Ç. Pinar, H. Yaman, The robust shortest path problem with interval data, Tech. rep., Bilkent
University, Ankara, Turkey (2001).
[17] R. Montemanni, J. Barta, L. M. Gambardella, Heuristic and preprocessing techniques for the robust traveling
salesman problem with interval data, Tech. rep., Dalle Molle Institute for Artificial Intelligence (2006).
[18] J. Hooker, G. Ottosson, Logic-based Benders decomposition, Mathematical Programming 96 (1) (2003) 33–60.
[19] A. A. Coco, A. C. Santos, T. F. Noronha, Scenario-based heuristics with path-relinking for the robust set covering
problem, in: Proceedings of the XI Metaheuristics International Conference (Agadir, Marocco, 2015), 2015, pp.
1–8.
[20] G. Yu, On the max-min 0-1 knapsack problem with robust optimization applications, Operations Research 44 (2)
(1996) 407–415.
[21] H. Aissi, C. Bazgan, D. Vanderpooten, Approximation of min-max and min-max regret versions of some combinatorial optimization problems, European Journal of Operational Research 179 (2) (2007) 281–290.
[22] M. J. Feizollahi, I. Averbakh, The robust (minmax regret) quadratic assignment problem with interval flows, INFORMS Journal on Computing 26 (2) (2014) 321–335.
[23] A. Kasperski, P. Kobylański, M. Kulej, P. Zieliński, Minimizing maximal regret in discrete optimization problems
with interval data, in: Issues in Soft Computing. Decisions and Operations Research, O. Hryniewicz, J. Kacprzyk,
D. Kuchta (eds.), Akademicka Oficyna Wydawnicza EXIT, Warszawa, 2005, pp. 193–208.
[24] A. Kasperski, P. Zieliński, An approximation algorithm for interval data minmax regret combinatorial optimization
problems, Information Processing Letters 97 (5) (2006) 177–180.
[25] E. Conde, A 2-approximation for minmax regret problems via a mid-point scenario optimal solution, Operations
Research Letters 38 (4) (2010) 326–327.
[26] E. Conde, On a constant factor approximation for minmax regret problems using a symmetry point scenario, European Journal of Operational Research 219 (2) (2012) 452 – 457.
[27] A. B. Chassein, M. Goerigk, A new bound for the midpoint solution in minmax regret optimization with an application to the robust shortest path problem, European Journal of Operational Research 244 (3) (2015) 739 –
747.
42
[28] A. Kasperski, P. Zieliński, On the existence of an FPTAS for minmax regret combinatorial optimization problems
with interval data, Operations Research Letters 35 (4) (2007) 525–532.
[29] J. F. Benders, Partitioning procedures for solving mixed-variables programming problems, Numerische Mathematik
4 (1) (1962) 238–252.
[30] A. M. Geoffrion, Generalized Benders decomposition, Journal of Optimization Theory and Applications 10 (4)
(1972) 237–260.
[31] D. McDaniel, M. Devine, A modified Benders’ partitioning algorithm for mixed integer programming, Management Science 24 (3) (1977) 312–319.
[32] T. L. Magnanti, R. T. Wong, Accelerating Benders decomposition: Algorithmic enhancement and model selection
criteria, Operations Research 29 (3) (1981) 464–484.
[33] M. Fischetti, D. Salvagnin, A. Zanette, A note on the selection of Benders’ cuts, Mathematical Programming
124 (1-2) (2010) 175–182.
[34] L. Assunção, A. C. Santos, T. F. Noronha, R. Andrade, On the finite optimal convergence of logic-based Benders’
decomposition in solving 0-1 min-max regret optimization problems with interval costs, Lecture Notes in Computer
Science 9849 (1) (2016) 1–12.
[35] Y. P. Aneja, V. Aggarwal, K. P. K. Nair, Shortest chain subject to side constraints, Networks 13 (2) (1983) 295–302.
[36] J. Beasley, N. Christofides, An algorithm for the resource constrained shortest path problem, Networks 19 (4)
(1989) 379–394.
[37] R. Hassin, Approximation schemes for the restricted shortest path problem, Mathematics of Operations Research
17 (1) (1992) 36–42.
[38] Z. Wang, J. Crowcroft, Quality-of-Service routing for supporting multimedia applications, IEEE on Selected Areas
in Communications 14 (7) (1996) 1228–1234.
[39] G. Apostolopoulos, R. Guérin, S. Kamat, S. K. Tripathi, Quality of Service based routing: A performance perspective, SIGCOMM Computer Communication Review 28 (4) (1998) 17–28.
[40] L. Santos, J. Coutinho-Rodrigues, J. R. Current, An improved solution algorithm for the constrained shortest path
problem, Transportation Research Part B: Methodological 41 (7) (2007) 756–771.
[41] H. Joksch, The shortest route problem with constraints, Journal of Mathematical Analysis and Applications 14 (2)
(1966) 191–197.
[42] I. Dumitrescu, N. Boland, Improved preprocessing, labeling and scaling algorithms for the weight-constrained
shortest path problem, Networks 42 (3) (2003) 135–153.
[43] X. Zhu, W. E. Wilhelm, A three-stage approach for the resource-constrained shortest path as a sub-problem in
column generation, Computers & Operations Research 39 (2) (2012) 164–178.
[44] L. D. P. Pugliese, F. Guerriero, A survey of resource constrained shortest path problems: Exact solution approaches,
Networks 62 (3) (2013) 183–200.
[45] M. J. Rosenhead, M. Elton, S. K. Gupta, Robustness and optimality as criteria for strategic decisions, Operational
Research Quarterly 23 (4) (1972) 413–431.
[46] P. Zieliński, The computational complexity of the relative robust shortest path problem with interval data, European
Journal of Operational Research 158 (3) (2004) 570–576.
[47] D. Catanzaro, M. Labbé, M. Salazar-Neumann, Reduction approaches for robust shortest path problems, Computers
43
& Operations Research 38 (11) (2011) 1610–1619.
[48] R. Montemanni, L. M. Gambardella, A. V. Donati, A branch and bound algorithm for the robust shortest path
problem with interval data, Operations Research Letters 32 (3) (2004) 225–232.
[49] J. K. Cochran, A. M. Uribe, A set covering formulation for agile capacity planning within supply chains, International Journal of Production Economics 95 (2) (2005) 139–149.
[50] A. Caprara, M. Fischetti, P. Toth, D. Vigo, P. L. Guida, Algorithms for railway crew management, Mathematical
Programming 79 (1) (1997) 125–141.
[51] S. Ceria, P. Nobili, A. Sassano, Set covering problem, in: Annotated Bibliographies in Combinatorial Optimization,
M. DellAmico, F. Maffioli and S. Martello (eds.), Wiley, 1997, pp. 415–428.
[52] A. Caprara, P. Toth, M. Fischetti, Algorithms for the set covering problem, Annals of Operations Research 98 (1)
(2000) 353–371.
[53] A. A. Coco, J. C. A. Júnior, T. F. Noronha, A. C. Santos, An integer linear programming formulation and heuristics
for the minmax relative regret robust shortest path problem, Journal of Global Optimization 60 (2) (2014) 265–287.
[54] R. Montemanni, L. M. Gambardella, An exact algorithm for the robust shortest path problem with interval data,
Computers & Operations Research 31 (10) (2004) 1667–1680.
[55] K. Sugiyama, S. Tagawa, M. Toda, Methods for visual understanding of hierarchical system structures, IEEE
Transactions on Systems, Man & Cybernetics 11 (2) (1981) 109–125.
[56] J. A. Bondy, U. S. R. Murty, Graph Theory with Applications, Elsevier, New York, 1976.
[57] J. E. Beasley, OR-Library: Distributing test problems by electronic mail, The Journal of the Operational Research
Society 41 (11) (1990) 1069–1072.
[58] M. Fischetti, A. Lodi, Local branching, Mathematical Programming 98 (1-3) (2003) 23–47.
44
| 8cs.DS
|
Requirements-driven Dynamic Adaptation to Mitigate
Runtime Uncertainties for Self-adaptive Systems
Zhuoqun Yang
Academy of Mathematics and Systems Science
Chinese Academy of Sciences
Beijing, China
[email protected]
Abstract: Self-adaptive systems are capable of adjusting their
behavior to cope with the changes in environment and itself. These
changes may cause runtime uncertainty, which refers to the system
state of failing to achieve appropriate reconfigurations. However, it is
often infeasible to exhaustively anticipate all the changes. Thus,
providing dynamic adaptation mechanisms for mitigating runtime
uncertainty becomes a big challenge. This paper suggests solving this
challenge at requirements phase by presenting REDAPT, short for
REquirement-Driven adAPTation. We propose an adaptive goal
model (AGM) by introducing adaptive elements, specify dynamic
properties of AGM by providing logic based grammar, derive
adaptation mechanisms with AGM specifications and achieve
adaptation by monitoring variables, diagnosing requirements
violations, determining reconfigurations and execution. Our approach
is demonstrated with an example from the Intelligent Transportation
System domain and evaluated through a series of simulation
experiments.
Keywords: Self-adaptive systems, requirements modeling, runtime
uncertainty, specification, dynamic adaptation
I. INTRODUCTION
Self-adaptive systems are capable of adjusting their
behavior to cope with the changes in environment and itself.
These changes may cause runtime uncertainty, which refers to
the system state of failing to achieve appropriate
reconfigurations under requirements violation [1]. We divide
the sources of runtime uncertainty into context uncertainty and
components uncertainty. Generally, context uncertainty means
the changes in execution environment, e.g. changes of
bandwidth for a service system or changes of temperature for
an air conditioner, while components uncertainty represents
changes in the system itself, e.g. sensors break for componentsbased systems. For keeping continuous satisfaction of
requirements at runtime, self-adaptive systems need
appropriate adaptation mechanisms. This paper suggests
achieving adaptation mechanisms at requirements phase.
Requirements phase is considered as the first stage during
the life cycle of a system. Definitely different from traditional
requirements engineering (RE), for self-adaptive systems, it is
need to capture not only functional requirements (FR) and nonfunctional requirements (NFR) but also adaptive requirements
(AR), which refer to the requirements that can be hold through
adaptation. Additionally, adaptation mechanisms also should
be taken into consideration at this phase for making clear when
to adapt and how to adapt [2]. Therefore, requirements
Wei Zhang, Haiyan Zhao, Zhi Jin
Institute of Software
School of EECS, Peking University
Beijing, China
{zhangw, zhhy}@sei.pku.edu.cn, [email protected]
engineering for self-adaptive systems should provide two kinds
of support: methods for modeling requirements and adaptation
mechanisms for mitigating runtime uncertainty. Recently, the
RE community made great strides in introducing methods and
techniques for providing the two supports. Wittle et al. [4]
introduced RELAX, a formal requirements specification
language to specify the uncertain requirements of self-adaptive
systems. Cheng et al. extended RELAX with goal modeling to
specify uncertainty in the goal model [5]. Requirements-aware
approaches [11-14] considered requirements as runtime entities
and monitoring as meta-requirements about whether the
requirements hold during runtime. Baresi et al. [16] introduced
adaptive goal into KAOS model and provided adaptation
mechanism by adopting fuzzy membership function. Other
works include monitoring and diagnose [8-10], self-repair [17]
[18] and architecture-based adaptation [20-23].
However, on account of growing complexity of system
structure, inherent volatility of environment and increasing
diversity of seamless interaction between the system and
environment, it becomes infeasible to predict and anticipate all
the runtime changes at requirements phase. Moreover, in more
and more situations, changes cannot be handled off-line, but
require the system to adapt its behavior dynamically without
human intervention. Therefore, providing requirements models
at runtime for resolving unpredictable uncertainty and modeldriven adaptation mechanisms become research urgencies and
challenges [24] [25].
This paper tackles these challenges by seeking answers to
the following questions: how to identify the boundary of
uncertainty sources (Q1); how to provide runtime requirements
model for modeling adaptive requirements and uncertainty
sources in a generic process (Q2); how to represent dynamic
properties of the runtime requirements model (Q3); how to
derive adaptation mechanisms and achieve adaptation (Q4).
To achieve these ends, we endeavor to present REDAPT,
short for REquirements-Driven adAPTation. For answering Q1,
this paper introduces how to exploring problem world for
identifying the boundary of uncertainty sources. To answer Q2,
we propose an adaptive goal model (AGM) by introducing
context uncertainty, components uncertainty, kinds of adaptive
elements and MAPE loop [26] into traditional goal model. For
answering Q3, we provide first-order linear-time temporal logic
based formal grammar for specifying AGM elements so as to
bridge the gap between requirements model and adaptation
mechanisms. To answer Q4, we derive adaptation mechanism
algorithms by integrating AGM with its specifications. The
adaptation can be achieved by monitoring variables, diagnosing
requirements violations, determining reconfigurations and
execution. We demonstrate REDAPT by applying it to a
Highway-Rail Control System (HRCS) from Intelligent
Transportation System domain (ITS) and evaluate it through a
series of simulation experiments. The results illustrate
REDAPT’s ability of modeling requirements as runtime
entities and mitigating runtime uncertainty by parametric
adaptation and structural adaptation.
The rest of the paper is structured as follows. Section II
introduces the motivating example. Section III overviews our
approach. Section IV presents the processes of modeling AGM
and deriving its specifications. Section V provides the
adaptation mechanism algorithm. We illustrate and evaluate the
proposed adaptation mechanism in Section VI. The related
work is discussed in Section VII. Finally, Section VIII
concludes the paper and identifies avenues for future work.
II. MOTIVATING EXAMPLE
AR
Description
AR1
When illuminance is above 20lx, gates open/close time
interval is set to 4s for pass efficiency.
AR2
For safety efficiency, when illuminance is under 20lx, gates’
closing time should be set within (1s, 4s], while opening time
should be set within [4s, 7s) to achieve safety efficiency.
AR3
Dispatch time interval should ensure at least 50% of the
vehicles can pass through the whole highway within 400s.
AR4
Dispatch time interval should ensure the amount of vehicles
on the highway is always under 350 at any time.
AR5
Components can be replaced once failure occurs.
III. APPROACH OVERVIEW
This section provides the framework of REDAPT and the
processes of deriving adaptive goal model, AGM specifications
and achieving adaptation.
A. REDAPT Framework
Figure 1 depicts the REDAPT framework. It is composed
of three basic layers: Problem Layer, Model Layer and
Adaptation Layer. Model Layer models the scenarios from
Problem Layer, while Adaptation Layer achieves the
adaptation mechanisms derived from Model Layer. The three
layers compose a feedback loop that controls the adaptation of
self-adaptive systems at runtime.
PROBLEM LAYER
MODEL LAYER
Shared
System
Environment
System
Evnironmentphenomena
M1
AGM
M2
A1
A2
Monitor
A6
Knowledge
Base
A3
AGM’
Execute
Analyze
A4
A5
A7
Plan
ADAPTATION LAYER
Adaptation Transitions
A1: monitoring context uncertainty
A2: monitoring components uncertainty
A3: provide monitored data
A4: requirements are violated
A5: provide adaptation
A6: execute adaptation on system
A7: trigger evolution when fails planning
M4
Requirements Model
Intelligent transport systems are advanced applications
which, without embodying intelligence as such, aim to provide
innovative services relating to different modes of transport and
traffic management and enable various users to be better
informed and make safer. Highway-rail crossing control is a
fundamental concern [29]. Basic scenarios can be described as:
Rails are built across a highway for trains dispatched from
east and west. At the crossing, gates are built at both sides
of the highway for block the vehicle flow temporarily from
south and north when train is coming.
Train Dispatch End is in charge of determining appropriate
dispatch time interval according to vehicle flow.
Gate Control End is in charge of closing/opening gates on
both side of the highway, when the train is monitored
approaching/leaving the crossing.
However, the environment and system itself are ever
changed. Thus, the changing scenarios can be described as:
When illuminance is above 20lx, the closing/opening time
interval is set to 4s for pass efficiency; otherwise the
closing time interval should be set within (1s, 4s] and
opening time interval should be set within [4s, 7s) for safety
efficiency.
Vehicle flow on highway is changing. For preventing traffic
jam and accident, we add some constraints that the dispatch
time interval should ensure at least 50% of the vehicles can
pass through the whole highway within 400s and the
amount of vehicles on the highway is always under 350.
Sometimes sensors may fail to monitor context variables or
the monitored value may suffer from noises.
The scenarios above imply that the dispatch time interval is
not fixed but varies according to the changed vehicle flow.
Thus, the requirement of determining dispatch time interval
should adapt to vehicle flow changes. We call this kind of
requirement Adaptive Requirements (AR). Adaptive
requirements of HRCS are described in Table I.
TABLE I. ADAPTIVE REQUIREMENTS OF HRCS
M3
AGM
Specification
Model Transitions
M1: capturing system requirements, context
uncertainty and components uncertainty
M2: evolution of AGM when unpredicted
uncertainty occurs
M3: describing dynamic properties of AGM
M4: achieving adaptation mechanism from
AGM specification
Fig. 1. Overview of the REDAPT framework
Problem layer aims at illustrating the relation between the
system and the environment. By analyzing the shared
phenomena and system, we can capture system’s requirements,
context uncertainty and components uncertainty (M1).
Model layer consists of Adaptive Goal Model (AGM) and
AGM Specifications. AGM is the requirements model derived
by extending traditional goal model with some new elements,
e.g. context uncertainty, adaptive goal and adaptive task. We
consider AGM as runtime entity, which can evolve at runtime
(M2). For bridging the gap between AGM and adaptation
mechanisms, dynamic properties of AGM are represented with
AGM specifications (M3). Thus we can achieve adaptation
mechanism algorithms from the specifications (M4).
Adaptation layer presents adaptation mechanisms based on
AGM specification. We adopt the MAPE feedback loop from
autonomic computing [26] for monitoring context uncertainty
(A1) and components uncertainty (A2), diagnosing
requirements violations (A4), deciding resolutions (A5) and
executing the decisions (A6). When the system fails planning,
unpredicted uncertainty should be modeled in AGM (A7).
B. From Model Layer to Adaptation Layer
1) Deriving Adaptive Goal Model and Specifications
Figure 2 presents a generic process of deriving AGM and
its specifications in Model Layer. The input is scenarios and
the output is AGM specifications. Each sub-process connects
with each other by artifacts coming from the former process.
Directed lines depict the order of processes. We give a brief
introduction to these sub-processes next.
P1: Adaptive requirements are elicited from the scenarios
and we can derive the initial requirements model in P1. The
adaptive requirements are described in adaptive tasks.
P2: Context uncertainty and components uncertainty are
identified from the adaptation scenarios.
P3: Attach context uncertainty to the adaptive tasks in
existing requirements model and then switch the adaptive tasks
into adaptive goals.
P4: Refine adaptive goals with MAPE loop. Thus, each
adaptive goal is operationalized through four tasks: Monitor,
Analyze, Plan and Execute.
P5: Attach components uncertainty to the tasks that are
accomplished by components and switch the tasks into
adaptive goals again.
P6: Refine new adaptive goals with MAPE tasks again.
P7: Specify elements of AGM with proposed grammar.
AGM can be derived from P1 to P6. P7 is involved in the
process because it has close relation with AGM. The details of
specifications are elaborated in Section IV.
Adaptation
Scenarios
Problem Layer
P2
Identify relevant sources
of runtime uncertainty
P1
Elicit and model
adaptive requirement
Context
uncertainty
(ConU)
Goal model
with adaptive
tasks (At)
P3
Add context uncertainty to
goal model and switch tasks
into adaptive goals (Ag)
AGM under
context uncertainty
Adaptation Layer
Goal model
with M/A/P/E
tasks
AGM
specifications
Components
uncertainty
(ComU)
P7
Specify elements of
AGM
P5
Add components uncertainty to
MAPE tasks and switch tasks
to adaptive goals (Ag)
AGM and
classification
of violation
AGM under
components unerrtainty
P4
P6
Refine adaptive goal with
MAPE process operating by
related components
Refine adaptive goal
with MAPE process
AGM
specification
grammar
Fig. 2. A generic process of deriving adaptive goal model
Layer
Process
Artifact
C. Process of Achieving Requirements-driven Adaptation
Figure 3 depicts a generic process of achieving adaptation
in Adaptation Layer. The input is AGM specifications and the
output is new configurations.
Monitoring: Monitor algorithms are derived from
specifications of Monitor tasks. The monitored variables are
specified in attributes of monitor tasks. By monitoring, we get
runtime data of both context and components.
Analyzing: Specifications of Analyze tasks describe how to
diagnose requirements violation at runtime. We suggest
adopting quantitative verification into diagnosing process. For
choosing appropriate algorithms, we should match the type of
violation according to the classification derived from both
requirements perspective and uncertainty perspective.
Planning: Planning algorithms should be chosen according
to the type of violation. What should be noticed is that the
planning process is intertwined with analyzing process, because
the modification of parameters or structures is always followed
by iterative verification.
Executing: Realize parametric or structural adaptation
according to the results from planning process.
After adaptation, new configuration is achieved and
runtime uncertainty is mitigated. The achieving details are
provided in Section V.
AGM
Specification
Model Layer
Context/Components
uncertianty
Specification
M task
Specification
Monitoring
Monitor the changes
of variables at runtime
Classification
of violation
A task
Specification
Analyzing
Diagnose violations
and match the
classification of
violation
Runtime
Data
Diagnose
Result
Problem Layer
P task
Specification
Planning
Choose planning
algorithm and achieve
appropriate solutions
Violation
information
E task
Specification
System in new
configurations
Executing
Reconfigure system
parameters and components
Parametric/Structural
adaptation
Layer
Process
Artifact
Fig. 3. A generic process of achieving requirements-driven adaptation
IV. ADAPTIVE GOAL MODEL AND FORMAL SPECIFICATION
REDAPT provides the processes of deriving AGM and its
specifications. First, we should make clear the boundary of
uncertainty and then carry out the processes in Section IV-B.
A. Identifying Uncertainty
1) Capturing Uncertainty
Uncertainty should be identified according to concrete
scenarios. Problem Layer in Section III consists of two
intertwined parts: system and environment. The uncertainty
existing in environment that can be observed by system
belongs to context uncertainty. While, the uncertainty existing
in system itself belongs to components uncertainty. Both
context uncertainty and components uncertainty are
unpredicted at runtime.
In HRCS, context uncertainty consists of:
Illuminance: Outdoor illuminance is a continuous variable
changing over time. 20lx is a boundary of whether driver
can drive safely without auxiliary illumination instruments
g0
decomposition, father goal can be achieved by achieving all the
sub-goals or sub-tasks, while for OR decomposition it can be
achieved by achieving only one of the alternative goals or tasks.
For example, g2 can be achieved by achieving both t1 and t2,
while g4 can be achieved by achieving either t4 or t5. Tasks are
leaves in goal model and can be operated by certain agents.
Tasks can contribute to softgoals through Help-contribution or
Hurt-contribution. For instance, t4 has positive effects on sg1,
while t3 has negative effects on sg1.
g0
g1
g2
g3
t3
g4
t4
t5
Help
Hurt
sg1
Meaning
ei
Value of illuminance gauged by I_sensori
fi
g3gauged by
g4I_camera t1
Value of vehicle flow
i
E
Mean value of illuminance
to tdispatch Hurt
Utility of closing time for passing vehicles, proportional to tclose
Uopen
Appropriateto topen
Utility of opening time for passingAppropriate
vehicles, inversely proportional
tclose and topen
tclose and topen
Utility of illuminance
Ag2: Set
At2: Set
At2: Set
Appropriate
Upass
Utility of safety efficiency P3
at crossing
NFR Violation
tclose and topen
Utility of pass efficiency at crossingConU2:
p
Percentage of vehicles whose driving time is under 300
n
Number of vehicles on highway at
each moment
Appropriate
tclose and topen
P3
NFR Violation
ConU2:
Illuminance
At1: Determine
tdispatch to make
P>50% and N<350
Hurt
sg3: Pass Efficiency
t1: Dispatch
(Upass)
according to
tdispatch
Help
g2: Gates
Control
Help
sg4: Safety Efficiency
t2: Monitor
(Usafety)
Train Location
At2: Set
Appropriate
tclose and topen
Fig. 5. Goal model with adaptive tasks
Figure 6 provides the processes of refining At1. Vehicle
Ag1: Determine
Flow, represented as ConU1,
belongs to context
uncertainty
At1: Determine
tdispatch
to make
Dispatch
tdispatch of
to make
and it affects the achievement
At1. Thus,P>50%
for and
P3,
link
ConU1 g0:Trains
and
N<350
P>50% and N<350
Control Gates
to At1At1:
with
Affect relation, represented as two-arrow directed
Determine
tdispatch to make
P3 label of FR violation.
P3
FR Violation
line, attached
with the
For refiningg1:At1,
Dispatch
FR Violation
g2: G
P>50% and N<350
ConU1:
Train According
Con
switch adaptive task into adaptive
goal Flow
(Ag), ConU1:
i.e. Ag1. For
to P4,
t
ConU1: Vehicle
Vehicle
Flow
Flow
Ag1 is decomposed into fourVehicle
tasks
(M1,
A1,
P1
and
E1),
which
FR Violation
t1: Dispatch
P4
Determine
are adopt from MAPE loop. ComU2:
Thus, the Ag1:
original
adaptation
t2: Mo
according to
t
to make
Train L
ComU1:
ComU2:
Sensor
t
P>50%
and N<350 processes
mechanisms
are
presented
as
M/A/P/E
tasks.
Similar
Noise
Sensor
Sensor
Ag1: Determine
E1:
Failure
ComU1:
NFR
tdispatch to make
can be carried
out forNoise
modeling
At2
(Figure 7).
Refoncigure
M2: Gau
Sensor
Violation
P>50% and N<350
FR
NFR
Illumi
t
Failure
After modeling
all the context
uncertainty,
we take E1:into
Sen
Violation Violation
Ag3: Gauge F
FR
P1: Decide
M1:
Gauge
F
by
Precisely
Refoncigure
Violation
sg1: Precision of the components uncertainty (P5). M1 can be
consideration
t
to hold
Infrared Sensors
P5
tdispatch
Efficiency
A2:
P>50% and N<350
achieved by Infrared
Sensors, which may be affected
by Sensor
U
M1: Gauge F M3: Gauge F by
E3:
A1: Verify R5 Reconfigure
Help
Infrared
Sensors
by
Infrared
P1:
Decide
tmay
dispatch
Failure (ComU1) and Sensor
Noise
(ComU2).
ComU1
A1:
Verify R5 and
and
R6
at
components
and Record Data
Sensors
R6 at
runtime
to hold P>50%
and
P2.1: Ch
result in FR violation, while ComU2
mayruntime
cause
NFR violation.
P6
N<350
P3: Choose
betwee
A3: Detect
Available Sensor
Similar to At1, we switch M1Empty/Noise
into
Ag3
and
refine
it.
Thus
we
Ag3: Gauge F for replacing
Help
Data
Precisely
derive the original
adaptation mechanisms
for achieving
Ag3.
E3:
M3: Gauge F by
P4
dispatch
dispatch
M2: Gauge E by
E3: Reconfigure
B. Deriving Adaptive
Goal Model
Illumination
tclose and topen
Sensors
P3: Choose tclose and
Goal-oriented methods are wildly accepted in RE [27].
topen by Trade-off between
A2: Verify
Goal models are used Uto
model Uand
analyze stakeholder
safety
safety and Upass
objectives. Systems’ FR are represented as hard goals, while
P3.2: Choose topen
P3.1: Choose tclose
NFR are represented as softgoals.
Figure 1 presents
an example
between [4s, 7s)
between (1s, 4s]
of goal model. Goal can be refined through AND/OR
Hurt
Help
Help
decompositions into sub-goals
andHurtsub-tasks.
For
AND
sg3: Pass Efficiency
(Upass)
Help
dispatch
Illuminance
Ag2: Set
F
Viol
P3.2: Choose topen
between [4s, 7s)
Train According
t2
Uclose
Usafty
ComU1:
Sensor
Failure
Control Gates
Time interval of gate opening after detecting train’s leaving
UE
At1: Determine
tdispatch to mak
P>50% and N<3
sg1: Precision
Efficiency
topen
Trade-off between
g0:by
Dispatch
Usafetyand
and Upass
Trains
A2: Verify
P3.1: Choose tclose
g1: Dispatchbetween (1s, 4s]
topen
tdispatch
P>50% and N<350
P3: Choose tclose and
g2
tclose
t
At1: Determine
tdispatch to make
Ag2: Set
At2: Set
Sensors
goal
t4at each highway
t5
Mean value oft3
vehicle flow
entrance
task
softgoal
Time
Help
Hurt
And
Dispatching time interval
Or
sg1
Help/
contribution
Hurt
Time interval of gate close after detecting train’s coming
F
goal
task
softgoal
And
Or
contribution
Appropriate
Figure 5 presents theAppropriate
goal model with
adaptive tasks (At)
tclose and topen
tclose and topen
At2:
Set
derived through P1 in Section IV-B. Adaptive tasks model and
Appropriate
P3
P3
Violation
NFR
Violation
represent
thetopenadaptive
requirements.
In NFR
general,
At1 models
tclose and
ConU2:
ConU2:
AR1 and AR2, while At2
models
AR3
and
AR4.
P2 is the
Illuminance
Illuminance
process of identifying context uncertainty (ConU) and
Ag2: Set
components uncertaintyAppropriate
(ComU), which is accomplished in
t
and topen
close
Section M2:
V-A.
Next,
we
conduct
the processes
from P3 to P6 for
Gauge E by
E3: Reconfigure
both At1 Illumination
and At2. First we take At1 for illustration.
tclose and topen
Usafety
g1
Help/
Hurt
t
Fig. 4. An example of goal model
TABLE II. SYMBOL ASSUMPTIONS OF VARIABLES
Symbol
g1
Tra
t2
t1
P4
or aids. Thus, when illuminance is under 20lx, the system
should adjust it behavior for safety efficiency requirement.
Vehicle flow: Vehicle flow can be viewed as discrete
variable because the monitored results are always integer.
Components uncertainty consists of:
Sensor Failure: Sensors fails to monitor the changes in
context. There is no return value from sensors. Failed
sensors should be replaced by available sensors.
Sensor Noise: Sensors succeed in monitoring variables, but
the monitored value is unstable. Sensor noise can also
tackled by replacement of available sensors.
2) Symbol Assumptions
For the convenience of deriving AGM and specifications,
Table II introduces some symbols to represent variables and
constants in the adaptive scenarios. Utility is used to depict the
satisfaction degree of requirements, varying from 0 to 1. 0
refers to completely dissatisfaction while 1 refers to completely
satisfaction. The value between (0, 1) refers to partial
satisfaction. For instance, Uclose represents the drivers’
satisfaction degree of closing gates when train is coming. The
small tclose is, the small Uclose will be, because drivers want to
pass the crossing as soon as possible. On the contrary, the small
topen is, the large Uopen will be, because drivers don’t want to
wait at the crossing for too long. Utility is used to get better
tradeoff between NFR and mitigate the runtime uncertainty of
NFR. Among the symbols, Uclose is the function of tclose, while
Uopen is the function of topen. Usafety is the function of UE, Uclose
and Uopen, while Upass is the function of Uclose and Uopen. We will
give the concrete function in Section VI.
sg4: Safety Efficiency
(Usafety)
dispatch
dispatch
s
Infrared Sensors
and Record Data
A3: Detect
Empty/Noise
Data
sg1: Precision
Reconfigure
Efficiency
components
P3: Choose
Available Sensor
for replacing
sg2: Pas
(U
dispatch
At1: Determine
t1: Dispatch
according to
tdispatch to make
p>50% and n<350
l
At2: Set
Appropriate
t2: Monitor
Train Location
tdispatch
tclose and topen
ution
TABLE III. CLASSIFICATION OF VIOLATIONS
Ag2: Set
ppropriate
Ag1: Determine
tdispatch to make
p>50% and n<350
p>50% and n<350
At1: Determine
e and topen
At1: Determine
tdispatch to make
P3
tdispatch to make
P3
p>50% and n<350
ConU1:
Vehicle Flow
ConU2:
uminance
Functional
Requirements
FR Violation
FR Violation
R Violation
ConU1:
Vehicle Flow
Non-functional
Requirements
P4
ComU1:
Sensor
Failure
ComU2:
Sensor
Noise
FR
Violation
econfigure
and topen
Ag1: Determine
tdispatch to make
g0
P>50% and N<350
NFR
Violation
M1: Gauge fi by
Infrared Sensors
P5
g1
g2
M1: Gauge fi
by Infrared
Sensors
g3
tdispatch
A1: Verify p>50%
and n<350
at runtime
g4
P1: Decide tdispatch
to hold p>50% and
n<350
t2
t1
elp
t3
goal
task
Ag3: Gauge fi
Precisely
t4
M3: Gauge fi by
Infrared Sensors
Hurt
and Record Data
iency
t5
Help
A3: Detect
Empty/Noise
Data
sg1
E3:
Reconfigure
softgoal
components
And
P3: Choose
sg1: Precision
Available Sensor
Help
Or
Efficiency
for replacing Help/
contribution
Hurt
Fig. 6. Refinement of At1
g0: Dispatch
Trains and
Control Gates
tdispatch
ConU2:
P3
Illuminance
NFR Violation
tclose and topen
t1: Dispatch
according to
ConU2:
Ag2: SetIlluminance
Appropriate
t2: Monitor
Train Location
tdispatch
P1: Decide tdispatch
to hold P>50% and
N<350
A1: Verify P>50%
and N<350
at runtime
ConU2:
Illuminance
P4
Ag2: Set
Appropriate
M2: Gauge E
by Illumination
Sensors
NFR Violation
NFR Violation
tclose and topen
tdispatch
E1:
Refoncigure
tclose and topen
tclose and topen
At2: Set g2: Gates
Appropriate Control P3
g1: Dispatch Train
According to
Ag2: Set
Appropriate
At2: Set
Appropriate
E2: Reconfigure
tclose and topen
tclose and topen
P2: Choose tclose and topen
by Trade-off between Usafety
M2: Gauge E byA2: Verify
Illumination U
safety
Sensors
A2: Verify
Usafety
and U
P2: Choose pass
tclose and
tclose and topen
topen by Trade-off between
P2.2: Choose
P2.1: Choose
Usafety and Upass
topen between
between
[4s, 7s)
(1s, 4s]
tclose
HurtPass Efficiency
sg2:
Hurt
Help
P2.2: Choose topen
Help
between [4s, 7s)
sg3:
Safety Efficiency
Help
Help
(Upass)
(Usafety)
sg3: Pass Efficiency
sg4: Safety Efficiency
(Upass)
(Usafety)
Components Uncertainty
Context uncertainty
caused FR violation
(violation of Ag1)
Context uncertainty
caused NFR violation
(violation of Ag2)
Components uncertainty
caused FR violation
(violation of Ag3)
Components uncertainty
caused NFR violation
(violation of Ag3)
to tdispatch
At1:
Determine
t1: Dispatch
At2: Set
//
ELEMENTS
t2: Monitor
according to
Appropriate
tdispatch to make
Train Location
entity
:=
goal
|
softgoal
|
task
|
uncertainty
p>50% and n<350
tdispatch
tclose and topen
goal := goal-type mode name [attributes] [initialization]
[invariant] [variant] [fulfillment]
softgoal := Softgoal name [attribute] [tradeoff-softgoal]
[invariant] [variant] [fulfillment]
task := task-type name From goal-type
name input output Ag1: Determine
At1: Determine
[attributes] [initialization]
[fulfillment]
tdispatch to make
tdispatch
to make
and
n<350
uncertainty
:= uncertainty-typep>50%
name
[attribute]
[violation] p>50% and n<350
At1: Determine
P3
P3
tdispatch to :=
make
goal-type
Ordinary
Goal | Adaptive Goal
p>50% and n<350
FR Violation
FR Violation
mode := achieve | maintain
ConU1:
ConU1:
tradeoff-softgoal := Tradeoff With
Softgoal name
Vehicle Flow
Vehicle Flow
task-type := Ordinary Task| Monitor | Analyze | Plan | Execute
P4
input := Input name
ComU1:
ComU2:
output
:=
Output
name
Ag1: Determine
Sensor
Sensor
Failure
Noise
uncertainty-type
:=
Context Uncertainty | Components
tdispatch toUncertainty
make
P>50%
and N<350
violationFR:= Affected
violation-type
NFR Adaptive Goal name formula
E1:
Violation
violation-type :=Violation
FR Violation| NFR
Violation
M1:
Gauge fi by
Refoncigure
P5
Infrared Sensors
tdispatch
// ATTRIBUTES
M1: Gauge fi attribute+
attribute := Attribute
A1: Verify p>50%
P1: Decide tdispatch
by Infrared
and n<350
attribute := attribute-type
: name
to hold p>50% and
Sensors
at runtime
n<350
attribute-type := Numeric | Boolean | Class
// INVARIANT, VARIANT, INITIALIZATION, FULFILLMENT
+
initialization := Initialization
initial-property
Ag3: Gauge
fi
Precisely
initial-property := property-type
conditional-type formula
E3:
M3: Gauge fi by
Reconfigure
invariant
:= Invariant invar-property+
Infrared Sensors
components
and Record Data
invar-property
:=A3:
Constraint
formula
Detect
P3: Choose
sg1: Precision
Empty/Noise
Available
Sensor
+
Help
variant := Variant var-property
Efficiency
Data
for replacing
var-property := Possibility formula
fulfillment := Fulfillment fulfill-property+
fulfill-property := property-type conditional-type formula
g0:
Dispatch
condition-type
:= PreCondition | TriggerCondition | PostCondition
P6
P2.1: Choose tclose
Hurt
Hurt
between (1s, 4s]
sg1: Precision Efficiency
E2: Reconfigure
Context Uncertainty
C. Adaptive Goal Model Specificationsl
g0: Dispatch
For representing the dynamic
Trains and properties of AGM and
Control Gates
deriving adaptation mechanisms
later, we provide AGM
specification’s
grammar in Figure 8, inspired by Formal Tropos
g1: Dispatch
g2: Gates
Train According
[27] and KAOS
specifications [28].
Control
P6
e topen
s, 7s)
E1:
Refoncigure
Uncertainty Type
Requirements
Type
Fig. 7. Refinement of At2
After deriving the AGM, we should classify the violations
for the convenience of choosing appropriate adaptation
mechanisms during runtime. According to the Affect relation in
Trains and
Control Gates
Section V-B, we suggest classifying violations into four kinds
Fig. 8. AGM specification grammar
from both uncertainty perspective
ConU1:and requirements perspective.
Vehicle Flow
g2: Gates
g1: Dispatch
Train
Table III presents the classification
of requirements
violations
ConU2:of entities, i.e.
AGM specifications
consist of specifications
Control
According to
Illuminance
of HRCS. Based on the fourFRkinds
of
violation,
we
can
design
Violation
goal, softgoal, task and uncertainty. Attribute presents some
tdispatch
respective adaptation
mechanism algorithms in Section VI.
NFR Violation
ComU2:
properties related to the entity. Attribute-type
can be Numeric,
Sensorof violations, Ag1:
Determine
Among the four kinds
components
uncertainty
t1:
Dispatch
Boolean
or
Class.
Numeric
attributes
depicts
the
variables and
Ag2:
Set
Appropriate
t2:
Monitor
Noise
tdispatch
to make
according to
Train Location
caused violations
can be solved through
structural
adaptation.
tclose to
and tachieving
ComU1:
open
constants
that
are
needed
the
entity.
Boolean
P>50% and N<350
Sensor uncertainty
While context
caused violations can be mitigatingtdispatch attributes always function as the output of verification
NFR
activity.
Failure
E2:
Reconfigure
M2: Gauge E
Violation
E1:
by parametric adaptation.
tduring
by Illumination
close and tachieving
open
Class
attributes
refer
to
other
entities
involved
FR
Refoncigure
Ag3: Gauge F
Sensors
Violation
Precisely
attribute
M3 (Gauge F by
tdispatchthe specified entity. For example,
P2: Choose
tclose andof
topen
M3: Gauge F by
Infrared Sensors
and Record Data
A3: Detect
Empty/Noise
Data
A2: Verify
E3:
Reconfigure
components
P1: Decide tdispatch
to hold P>50% and
N<350
by Trade-off between Usafety
Usafety
and Upass
P2.1: Choose
P3: Choose
Available Sensor
for replacing
A1: Verify P>50%
and N<350
at runtime
tclose
between
(1s, 4s]
Hurt
Hurt
P2.2: Choose
topen between
[4s, 7s)
Help
Infrared Sensors) contains numeric attribute F (Vehicle Flow)
and Class attribute I_sensor (Infrared Sensor). Attribute of A1
contains a Boolean to record the result of verifying R3 and R4.
Besides, father goals’ attributes contain sub-goals’ attributes
and sub-tasks’ attributes.
For goal, softgoal and task, initialization and fulfillment
refer to the activating process and the terminating process of an
entity respectively. Condition-type consists of PreCondition,
TriggerCondition and PostCondition. PreCondition means the
condition before initialization or fulfillment; TriggerCondition
means the trigger condition; PostCondition means the result
after initialization or fulfillment. It is known that father entities
can be satisfied by achieving child entities. That is to say, child
entities’ initialization conditions are always triggered by the
activation of their father entities, while father entities’
fulfillment conditions are always triggered by fulfillment of
their child entities. For instance, Ag1’s initialization is triggered
by activation of g1, while its fulfillment is triggered by
fulfillment of M1, A1, P1 and E1.
Invariant refers to the constraints that the entities should
hold all the time, while variant refers to the possibilities that
may occur. For example, R3 and R4 are constraints to Ag1,
while the priority of sg2 and sg3 is changeable. Violation is
unique in uncertainty, describing the affected adaptive goals
and violation types.
The formulas in the grammar are specified by first-order
linear-time temporal logic. The syntax is given by:
t :: x | c | f(t1 ,..., tn )
:: t | P(t1 ,..., tn ) |
| | | |
X | F | G | U
x | x
A term t is a variable x, a constant c, or a function f of a
number of terms. A formula Φ is either a term, a predicate of a
number of terms, a Boolean operation, a timed operation or a
quantifiers operation. XΦ refers to Φ should hold in the next
state reached by the system. FΦ refers to Φ should eventually
hold in some future state. GΦ refers to Φ should hold in all state
of the system. Φ1 U Φ2 refers to Φ1 should hold until Φ2 holds.
The grammar provides a template for representing dynamic
properties of model entities. Several examples are provided.
Context Uncertainty Vehicle Flow
Attribute
Numeric f i, F
Class I_sensor
Affected Adaptive Goal Determine tdispatch to make p>50% and n<350
F ( p (tdispatch , F ) 50% n(tdispatch , F ) 350)
FR Violation
Components Uncertainty Sensor Failure
Attribute
Numeric f i
Class I_sensor
Affected Adaptive Goal Gauge f i Precisely
I _ sensori I _ sensori .value '' ''
FR Violation
Fig. 9. Specification of ConU1 and ComU1
Adaptive Goal Achieve Determine tdispatch to make p>50% and n<350
Attribute
Numeric tdispatch , f i, F, p, n
Class I_sensor
Initialization PreCondition
Fulfill ( Dispatch Train According to tdispatch )
Initialization TriggerCondition
Activate( Dispatch Train According to tdispatch )
Initialization PostCondition
tdispatch " "
Invariant
G ( p 50%0 n 350)
Fulfillment PreCondition
tdispatch F ( p(tdispatch , F ) 50% n(tdispatch , F ) 350)
Fulfillment TriggerCondition
new
new
new
tdispatch
( p (tdispatch
, F ) 50% n(tdispatch
, F ) 350)
Fulfillment PostCondition
new
tdispatch tdispatch
Fig. 10. Specification of Ag1
Softgoal Maintain Safety Efficiency
Attribute
Numeric tclose, topen, E, Usafety, Upass
Tradeoff With Softgoal Pass Efficiency
Invariant Constraint
{tclose , topen , E} G(U safety (tclose , topen , E) 0)
Variant Possibility
t (Prior( safety efficiency, pass efficiency , t )
Prior( pass efficiency, safety efficiency , t )
equalPrior(pass efficiency, safety efficiency , t ))
Fulfillment PreCondition
desired
{tclose , topen } U safety (tclose , topen , E) U safety
Fulfillment TriggerCondition
new
new
new
new
desired
{tclose
, topen
} (U safety (tclose
, topen
, E) U safety
)
MAX _ Tradeoff (U safety , U pass )
Fulfillment PostCondition
new
new
tclose tclose
topen topen
Fig. 11. Specification of sg2
Monitor Gauge fi by Infrared Sensors
From Adaptive Goal Determine tdispatch to make p>50% and n<350
Attribute
Numeric f i
Class I_sensor
Input None
Output f 1…f 10
Initialization PreCondition
Fulfill ( Determine tdispatch to make p 50% and n 350)
Initialization TriggerCondition
activate( Determine tdispatch to make p 50% and n 350)
Initialization PostCondition
I _ sensor1 ,...I _ sensor10 I _ sensori .select TRUE
Fulfillment PreCondition
I _ sensori , i {1,...,10} I _ sensori .value " "
Fulfillment TriggerCondition
I _ sensori , i {1,...,10} I _ sensori .gauge TURE
Fulfillment PostCondition
I _ sensori .value guagedValue output ( I _ sensori .value)
Fig. 12. Specification of M1
Analyze Verify p>50% and n<350 at runtime
From Adaptive Goal Determine tdispatch to
make p>50% and n<350
Attribute
Numeric f i, F, tdispatch, p, n
Boolean sat
Input f 1…f 10, tdispatch
Output sat
Initialization PreCondition
Fulfill ( Determine tdispatch to make p 50% and n 350)
Initialization TriggerCondition
I _ sensori .value " "
Initialization PostCondition
input output ( I _ sensori .value)
Fulfillment PreCondition
sat " "
Fulfillment TriggerCondition
Verification( p (tdispatch , F ) 50% n(tdispatch , F ) 350)
Fulfillment PostCondition
sat Return value of Verification output (sat )
Fig. 13. Specification of A1
Plan Decide tdispatch to hold p>50% and n<350
From Adaptive Goal Determine tdispatch to
make p>50% and n<350
Attribute
new
Numeric F, tdispatch, tdispatch
Boolean sat
Input sat, F, tdispatch
new
Output tdispatch
Initialization PreCondition
Fulfill ( Determine tdispatch to make p 50% and n 350)
Initialization TriggerCondition
input 0
Initialization
PostCondition
new
tdispatch " "
Fulfillment PreCondition
sat " "
Fulfillment TriggerCondition
while( sat 1)
new
{tdispatch
tdispatch 1;
new
new
Verification( p (tdispatch
, F ) 50% n(tdispatch
, F ) 350)
new
tdispatch tdispatch
}
return tdispatch
Fulfillment PostCondition
new
new
tdispatch
tdispatch output (tdispatch
)
Fig. 14. Specification of P1
V. REQUIREMENTS-DRIVEN ADAPTATION MECHANISM
This section presents how to get adaptation mechanisms
from AGM specifications. Adaptation mechanisms are
composed of four parts: monitoring, analyzing, planning and
executing, which are also known as adaptive tasks. We give the
generic algorithm based on the classification of violations in
Section IV-B.
Monitoring algorithm should provide the ability of
monitoring context uncertainty and components uncertainty.
The variables and sensors are specified in specifications
(Figure 12).
Algorithm 1 Monitor Context and Components Uncertainty
monitorUncertainty (AGM specifications){
for each Monitor task M{
select M.Specification.Attribute.Numeric.variables
as monitored variables
select M.Specification.Attribute.Class.instance as monitor
gauge values V
return V
}}
Analyzing algorithm is used for diagnosing whether the FR
or NFR are violated. The inputs of analyzing algorithm are
specifications of Analyze task (figure 13) and the returned
value of variables from monitoring algorithm.
Algorithm 2 Diagnose Requirements Violations
diagnoseViolations (AGM specifications, monitored value){
for each Analyze task A{
violationType VT
if A.InitializationPostCondition=TRUE
Match violation type
if (father Ag is affected by context uncertainty and Ag is
related to FR)
{ computing and verifying Ag.Invariant
if (Ag.Invariant=FALSE) VT=ConU_FR
else VT=None
return VT}
end if
if (father Ag is affected by context uncertainty and Ag is to
with sg)
{ computing sg.Attribute.Numeric.Utility
if(Utility<desired value) VT=ConU_NFR
else VT=None
return VT}
end if
if (father Ag is affected by components uncertainty and Ag
is to related with FR)
{ if (monitored value=” ”) VT=ComU_FR
else VT=None
return VT}
end if
if (father Ag is affected by components uncertainty and Ag
is related to NFR)
{ if (detect noise=TURE) VT=ComU_NFR
else VT=None
return VT }
end if
end if
}}
Planning algorithm is designed for determining how to
reconfigure the system for mitigating these violations. The
inputs are specifications of Plan task (Figure 14) and the
violation type.
Algorithm 3 Determine reconfiguration
decisionMaking (AGM specifications, violation_type){
for each Plan task P{
if P.InitializationPostCondition=TRUE
Match violation type
if (violation_type = None) return currentconfiguration
if (violation_type = ConU_FR or ConU_NFR)
{while (violation_type≠None)
do {adjust P.Attribte.Numeric.parameterValue
diagnoseViolations }
end do
return new parameterValue }
end if
if (violation_type = ComU_FR or ComU_NFR)
{while (violation_type≠None)
do { replace failed components with
M.Specification.Attribute.Class.instance
diagnoseViolations }
end do
return new structureconfiguration }
end if
end if
}}
Executing algorithm is built for carrying out the
reconfigurations derived in planning algorithm. The
reconfiguration can either be parametric or structural.
Algorithm 4 Execute Reconfiguration
executeDecision (AGM specifications, reconfiguration){
for each Execute task E{
if E.InitializationPostCondition=TRUE
if (reconfiguration= new parameterValue)
P.Attribte.Numeric.parameterValue=
new parameterValue
end if
if ( reconfiguration= new structureconfiguration)
M.Specification.Attribute.Class.instance=
new structureconfiguration
end if
end if
}}
Thus, we derive the generic adaptation mechanism
algorithms according to AGM specifications. In next Section,
we conduct two simulation experiments to demonstrate the
adaptation mechanism algorithms’ ability.
VI. EXPERIMENTAL EVALUATION
To evaluate the effectiveness of our approach, we conduct
two experiments based on the scenarios of HRCS on a
computer with Intel Core 3110M CPU and 2G memory. HRCS
is modeled by using AnyLogic. Road Traffic Library and
Railway Library are used for runtime simulation. The first
experiment focuses on mitigating violation of R2. We present
how to trade-off between softgoal sg2 and sg3 by parametric
adaptation. In the second experiment, we illustrate how to
reconfigure the system for mitigating violation of functional
requirements R3 and R4.
A. Mitigating Violation of Non-functional Requirements
For diagnosing violation of R2, we suggest leveraging
utility-based quantitative verification. Utility can be used to
represent the satisfaction degree of NFR, just as what we state
in Section IV-A. First, we give the function relation of these
utility and variables.
E 20
1
UE
0
E 20
UE 0
4
topen
[4, 7) U E 0
(1)
Uopen
UE 0
4
tclose
(1, 4] U E 0
Uclose
7 topen
3
tclose 1
3
1
Usafety =U E + (1-sgnU E ) Uopen +Uclose -2
2
Uopen +Uclose
U pass
2
(2)
(3)
(4)
(5)
Equation 1 depicts utility function of illuminance is 0-1
function. We assume the optimal close/open time interval are
both 4s when E>20. When E<20, the time interval can be
adjusted according to Eq. 2 and Eq. 3. In Eq. 4, sgnU E is the
symbolic function of UE. When E>20, Usafety equals to UE. That
is to say under better illumination, Usafety is not affected by tclose
and topen. When UE decreases to 0, Usafety can be tuned by
adjusting tclose and topen. Upass means that the more time vehicles
can have to pass the crossing, the larger value it will be.
Integrate Eq. 4 and Eq. 5, we can derive
U wait U safety 1
MRSU safetyU wait
U pass
U safety
(6)
1
(7)
MRS refers to marginal rate of substitution. Equation 7
implies that time resource distributed into safety efficiency and
pass efficiency result in Pareto Optimality. That means our
assumption is reasonable. Under Pareto Optimality, we can
better trade-off between the two “commodities”. When E<20,
we assume the desired Usafety is above 0.7. The dynamic
adaptation process is shown in Figure 15.
X-axis refers to time. Illuminance is monitored under 20lx
and Usafety is diagnosed to be 0 at 9, 13 and 18 time point.
System first tries to tune tclose and topen once violation occurs.
However, with iterative verification, R4 is still dissatisfied.
Then system tune tclose and topen for the second time at 10, 14
and 19 time point. We can see Usafety of the former two cases
are above 0.7, which means R2 hold at 10 and 14 time point.
Thus the adaptation is achieved. For the third case, further
tuning is needed.
Utility
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
(a)
0
2
4
6
illuminance
8
close
10
open
12
Time
14
16
18
safety efficiency
20
22
24
pass efficiency
Fig. 15. Adaptation of NFR violation
B. Mitigating Violation of Functional Requirements
This experiment mainly presents the adaptation of R4 and
R5. We design two comparison experiments. The context
uncertainty is simulated through changes of vehicle flow.
(b)
Fig. 16. Histogram of driving time (a) and vehicle amount curve (b) in
experiment 1 when tdispatch=5s and R3, R4 hold
TABLE IV. CONFIGURATION OF PARAMETERS
Comparison
Experiment
1
2
Configuration of Parameters
Vehicle Flow
Vehicle Flow
tclose
from North
from South
15Vehicles/m
18 Vehicles/m
4s
20 Vehicles/m
18Vehicle/m
4s
topen
4s
4s
(a)
In experiment 1, tdispatch is set to 5min. The x-axis of Figure
16(a) refers to driving time, while y-axis refers to the
percentage of vehicles. The x-axis of Figure 16(b) refers to
virtual time, while y-axis refers to the number of vehicles on
highway. In Figure 16(a), the yellow histogram refers to the
vehicles from south to north, while the blue histogram refers to
those from north to south. The green vertical line refers to the
mean of driving time from south to north while the red vertical
line refers to the mean of driving time from the other direction.
Based on the statistical result of 165118 samples, we can
calculate pblue=66.84% and pyellow=80.04%, so R3 hold. Besides,
the cureve in Figure 16(b) depicts R4 holds.
In experiment 2, the vehicle flow at both sides of highway
is monitored increasing by Infrared sensors, while tdispatch is also
5min. We can derive the result in the same way. In Figure 17,
pblue=38.77%, pyellow=48.51% and n>350. Thus, both R3 and R4
are diagnosed violated.
According to the adaptation mechanism algorithm, we
should tune the parameter tdispatch at runtime by increasing its
new
value. Do tdispatch
tdispatch 1 6s and verify R3 and R4
(b)
Fig. 17. Histogram of driving time (a) and vehicle amount curve (b) in
experiment 2 when tdispatch=5s and R3, R4 are violated
(a)
again. The result is presented in Figure 18 with pblue=90.57%,
pyellow=87.45% and n<350. Thus, R3 and R4 hold again and
adaptation is accomplished.
(b)
Fig. 18. The result after parametric adaptation
VII. RELATED WORK
Over the past decade, researches and practitioners have
developed a variety of methodologies, frameworks, and
technologies intended to support building adaptation.
Requirements model. Wittle et al. [4] introduced RELAX,
a formal requirements specification language that leverage
Fuzzy Branching Temporal Logic to specify the uncertain
requirements of self-adaptive systems. Our specification work
differs from theirs in that they just specify the static properties
of requirement, e.g. involved environment and relations with
other entities. However, AGM specification can describe the
dynamic properties of requirements, e.g. the initialization and
fulfillment of goals or tasks, input and output of tasks, the
variant of goals. In a subsequent work [5], Cheng et al.
extended RELAX with goal modeling to specify uncertainty in
the goals, while our AGM can not only describes uncertainty,
but also describes how to mitigate uncertainty by refining
adaptive goal with MAPE loop. There are some other works on
addressing uncertainty. Goldsby and Cheng [6] presented
behavior models to deal with uncertainty in environment
through processes in model-driven engineering. Our REDAPT
is also a model-driven method, but we focus both on
environment and system itself. To managing requirements time
uncertainty, Salay et al. [7] proposed partial models to
represent uncertainty in requirement and illustrate uncertainty
reduction by reason through the traceability relations.
Monitoring and Diagnosing. Requirement monitoring
aims to track systems’ runtime behavior for detecting
requirements violations. Diagnosing always comes along with
monitoring. Fickas and Feather proposed goal-oriented
methods for runtime requirements monitoring [8] and
mechanisms for repairing deviations caused by unsatisfied
domain assumptions [9]. However, our work also focuses on
how to monitor the deviation of system itself. Besides,
REDAPT gives an answer to the question about how to provide
monitoring specifications [8] [10]. Another representative work
on monitoring is requirements awareness. Requirements-aware
approaches [11-14] considered requirements as runtime entities
and monitoring as meta-requirements about whether the
requirements hold during runtime. Our work not only treats
requirements as runtime entities, but also treat requirements
model as runtime entities, i.e. requirements models at runtime.
Besides, Ramirez and Cheng [15] proposed to use utility
changes to depict requirements violation, while we just
leverage utility for represent satisfaction degree of NFR.
Achieving adaptation. Notable work of Baresi et al. [16]
introduced adaptive goal into KAOS model and provided
adaptation mechanism by adopting fuzzy membership function.
Our work differs from theirs in that adaptive goals in FLAGS
refers to the goals that can be fuzzy or modified, while adaptive
goals in REDAPT represent the adaptation needs of both NR
and NFR. In addition, adaptive goals in FLAGS couldn’t adjust
themselves online, because it has no online adaptation
mechanisms. On the contrary, REDAPT provides plenty and
generic adaptation mechanisms for supporting adjustment at
runtime. Wang et al. [17] [18] presented goal-oriented methods
for diagnosing and repairing requirements violations by
selecting the optimal structural configuration that contributes
most positively to system’s NFR. Our work not only focuses on
structural adaptation but also parametric adaptation. Esfahani et
al. [19] provided POISED for tackling the challenge posed by
internal uncertainty by introducing possibility theory to
quantify uncertainty. While REDAPT introduces utility
functions to quantify satisfaction of NFR.
Architecture-based methods. Researches also endeavor to
develop architectural methods and techniques. Cheng and
Garlan [20] described three specific sources of uncertainty and
address the way of mitigating them in Rainbow framework [21].
In notable work of Oreizy et al. [22], A broad framework for
studying and describing evolution is introduced by addressing
several issues on architectural, that serves to unify the wide
range of work in the field of dynamic software adaptation.
Vogel and Giese [23] proposed a model-driven approach to
provide multiple architectural runtime models at different
levels of abstraction as a basis for adaptation.
VIII. CONCLUSION
This paper presents REDAPT, short for RequirementsDriven adaptation, a new method of mitigating runtime
uncertainty for self-adaptive systems. The contributions of
REDAPT are as follows. First, by developing adaptive goal
model (AGM), the approach provides generically
comprehensive processes for reflecting Problem Layer into
Model Layer through modeling adaptive goals, context
uncertainty, components uncertainty and refining adaptive
goals with MAPE loops. Second, by specifying elements of
AGM, REDAPT presents the dynamic properties of AGM that
bridge the gap between Model Layer and Adaptation Layer.
Third, by proposing adaptation algorithm, we can achieve
dynamic adaptation for mitigating runtime uncertainty through
monitoring, analyzing, planning and executing. Thus, the
adaptation problem returns to Problem Layer again. The three
layers compose a feedback loop that controls the adaptation of
self-adaptive systems.
In our future work, we will concentrate on investigating
appropriate behavior models at runtime for representing selfadaptive systems and transitions from our AGM to behavior
model. In parallel, we will also intend to provide novel
adaptation mechanisms, especially runtime verification and
decision making.
REFERENCES
[1] A. J. Ramirez, A. C. Jensen, and B. H. C. Cheng, "A taxonomy
of uncertainty for dynamically adaptive systems," in
Proceedings of the 7th ICSE Workshop on Software
Engineering for Adaptive and Self-Managing Systems
(SEAMS’12), East Lansing, MI, USA, 2012, pp. 99-108.
[2] B. H. Cheng, R. Lemos, H. Giese, P. Inverardi, J. Magee, J.
Andersson, et al., "Software Engineering for Self-Adaptive
Systems: A Research Roadmap," Software Engineering for SelfAdaptive Systems, vol. 5525, pp. 1-26, 2009.
[3] Y. Brun, G. M. Serugendo, C. Gacek, H. Giese, H. Kienle, M.
Litoiu, H. Müller, M. Pezzè and M. Shaw "Engineering SelfAdaptive Systems through Feedback Loops," in Software
Engineering for Self-Adaptive Systems, 2009, pp. 48-70.
[4] J. Whittle, P. Sawyer, N. Bencomo, B. H. C. Cheng, and J. M.
Bruel, "RELAX: Incorporating Uncertainty into the
Specification of Self-Adaptive Systems," in Proceedings of the
17th IEEE International Conference on Requirements
Engineering (RE’09), Lancaster, UK, 2009, pp. 79-88.
[5] B. H. Cheng, P. Sawyer, N. Bencomo, and J. Whittle, "A GoalBased Modeling Approach to Develop Requirements of an
Adaptive System with Environmental Uncertainty," in
Proceedings of the 12th International Conference on Model
Driven Engineering Languages and Systems (MODELS’09),
Denver, CO, 2009, pp. 468 - 483 .
[6] H. J. Goldsby and B. H. Cheng, "Automatically Generating
Behavioral Models of Adaptive Systems to Address
Uncertainty," in Proceedings of the 11th International
Conference on Model Driven Engineering Languages and
Systems (MODELS’08), Toulouse, France, 2008, pp. 568-583.
[7] R. Salay, M. Chechik, and J. Horkoff, "Managing requirements
uncertainty with partial models," in Proceedings of the 20th
IEEE International Conference on Requirements Engineering
(RE’12), 2012, pp. 1-10.
[8] S. Fickas and M. S. Feather, "Requirements monitoring in
dynamic environments," in Proceedings of the 2nd IEEE
International Symposium on Requirements Engineering (RE’95),
Eugene, OR, USA, 1995, pp. 140-147.
[9] M. S. Feather, S. Fickas, A. van Lamsweerde, and C. Ponsard,
"Reconciling system requirements and runtime behavior," in
Proceedings of the 9th International Workshop on Software
Specification and Design, 1998, pp. 50-59.
[10] K. Welsh and P. Sawyer, "When to Adapt? Identification of
Problem Domains for Adaptive Systems," in Proceedings of the
14th international conference on Requirements Engineering:
Foundation for Software Quality (REFSQ’08), Montpellier,
France, 2008, pp. 198-203.
[11] N. Bencomo, J. Whittle, P. Sawyer, A. Finkelstein, and E. Letier,
"Requirements reflection: requirements as runtime entities," in
Proceedings of 32nd ACM/IEEE International Conference on
Software Engineering (ICSE), Lancaster, UK, 2010, pp. 199-202.
[12] P. Sawyer, N. Bencomo, J. Whittle, E. Letier, and A. Finkelstein,
"Requirements-Aware Systems: A Research Agenda for RE for
Self-adaptive Systems," in Proceedings of 18th IEEE
International Conference on Requirements Engineering (RE’10),
Lancaster, UK, 2010, pp. 95-103.
[13] K. Welsh, P. Sawyer, and N. Bencomo, "Towards requirements
aware systems: Run-time resolution of design-time
assumptions," in Proceedings of the 26th IEEE/ACM
International Conference on Automated Software Engineering
(ASE’11), Lawrence, KS, USA, 2011, pp. 560-563.
[14] V. E. S. Souza, A. Lapouchnian, W. N. Robinson, and J.
Mylopoulos, "Awareness Requirements for Adaptive Systems,"
in Proceedings of the 6th International Symposium on Software
Engineering for Adaptive and Self-Managing Systems
(SEAMS’11), Waikiki, Honolulu, HI, USA, 2011, pp. 60-69.
[15] A. J. Ramirez and B. H. C. Cheng, "Automatic derivation of
utility functions for monitoring software requirements," in
Proceedings of the 14th international conference on Model
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
driven engineering languages and systems (MODELS’11),
Wellington, New Zealand, 2011, pp. 501-516.
L. Baresi, L. Pasquale, and P. Spoletini, "Fuzzy Goals for
Requirements-Driven Adaptation," in Proceedings of the 18th
IEEE International Conference on Requirements Engineering
(RE’10), Milan, Italy, 2010, pp. 125-134.
Y. Wang, S. A. Mcilraith, Y. Yu, and J. Mylopoulos,
"Monitoring and diagnosing software requirements," Automated
Software Engineering, vol. 16, pp. 3-35, March 2009.
W. Yiqiao and J. Mylopoulos, "Self-Repair through
Reconfiguration: A Requirements Engineering Approach," in
Proceedings of the 24th IEEE/ACM International Conference on
Automated Software Engineering (ASE’09), Washington, DC,
USA, 2009, pp. 257-268.
N. Esfahani, E. Kouroshfar, and S. Malek, "Taming uncertainty
in self-adaptive software," in Proceedings of the 19th ACM
SIGSOFT symposium and the 13th European conference on
Foundations of software engineering (ESEC/FSE’11), Szeged,
Hungary, 2011, pp. 234-244.
S.W. Cheng and D. Garlan, "Handling Uncertainty in
Autonomic Systems," In Proceedings of International Workshop
on Living with Uncertainties (IWLU'07), co-located with the
22nd International Conference on Automated Software
Engineering (ASE'07), Atlanta, Georgia, USA, 2007.
D. Garlan, S.W. Cheng, A.C. Huang, B. Schmerl, and P.
Steenkiste, "Rainbow: Architecture-Based Self-Adaptation with
Reusable Infrastructure," Computer, vol. 37, pp. 46-54, October
2004.
P. Oreizy, N. Medvidovic, and R. N. Taylor, "Runtime software
adaptation: framework, approaches, and styles," in Proceedings
of the 29th International Conference on Software engineering
(ICSE’08), Leipzig, Germany, 2008, pp. 899-909
T. Vogel and H. Giese, "Adaptation and abstract runtime
models," in Proceedings of the 5th ICSE Workshop on Software
Engineering for Adaptive and Self-Managing Systems
(SEAMS’10), Cape Town, South Africa, 2010, pp. 39-48.
G. Blair, N. Bencomo, and R. B. France, "Models@ run.time,"
Computer, vol. 42, pp. 22-26, 2009.
L. Baresi and C. Ghezzi, "The disappearing boundary between
development-time and run-time," in Proceedings of the
FSE/SDP workshop on Future of software engineering research,
Santa Fe, New Mexico, USA, 2010, pp. 17-22.
J. O. Kephart and D. M. Chess, "The Vision of Autonomic
Computing," Computer, vol. 36, pp. 41-50, January 2003.
A. Fuxman, L. Liu, J. Mylopoulos, M. Pistore, M. Roveri, and P.
Traverso, "Specifying and analyzing early requirements in
Tropos," Requir. Eng., vol. 9, pp. 132-150, May 2004.
A. Dardenne, A. v. Lamsweerde, and S. Fickas, "Goal-directed
requirements acquisition," Sci. Comput. Program., vol. 20, pp.
3-50, April 1993.
C.A. Bell, K.M. Hunter, “Low Volume Highway-rail Grade
Crossing Treatments for the Oregon High Speed Rail Corridor”,
technical report, Transportation Research Institute, Oregon State
University, 1997.
| 3cs.SY
|
Analysing the Degree of Meshing in Medium Voltage Target Grids - An Automated
Technical and Economical Impact Assessment
Leon Thurnera,d , Alexander Scheidlerb,d , Alexander Probstc , Martin Brauna,b
b Fraunhofer
a University of Kassel, Germany
Institute for Wind Energy and Energy System Technology, Kassel, Germany
c Netze BW GmbH, Stuttgart, Germany
d contributed equally
arXiv:1802.01492v1 [cs.CE] 5 Feb 2018
Abstract
There are different medium voltage (MV) grid concepts with regard to mode of operation and protection system layout.
The increasing installation of distributed generation (DG) raises the question if the currently used concepts are still
optimal for future power systems. We present a methodology that allows the automated calculation and comparison of
target grids within different concepts. Specifically, we consider radial grids, closed ring grids and grids with switching
stations. A target grid structure is optimized for each of those grid concepts based on geographical information. To
model a realistic planning process, compliance with technical constraints for normal operation, contingency behaviour
and reliability figures are ensured in all grid concepts. We present a multiphase approach to solve the optimization
problem based on an iterated local search meta-heuristic. We then economically compare the grids with regards to
investment and operational cost for primary and secondary equipment, to analyse which concept leads to the most
overall cost-efficient target grids. Since the methodology allows an automated evaluation of a large number of grids,
it can be used to draw general conclusions about the cost-efficiency of specific concepts. The methodology is applied
to 44 real MV grids spanning about 4800 km of lines, for which the results show that a radial grid structure is overall
cost effective compared to grid topologies with switching stations or closed rings even in grid areas with a large DG
penetration. The contribution of this paper is threefold: first, a comprehensive methodology to compile automated target
grid plans under realistic premises is presented. Second, the practical applicability of the approach is demonstrated by
its application in a large scale case study with a high degree of automation. And third, the results of the case study
allow to draw conclusions about the techno-economical differences of different MV grid concepts.
Keywords: grid planning, distribution systems, expansion planning, reliability, meshing, switching stations, single
contingency policy, radial grid, mode of operation, distributed generation, optimal planning
1. Introduction
The increasing installation of distributed generation
(DG) in electric distribution systems leads to a gradual
transformation from a centralized to a decentralized power
generation. This paradigm shift presents a significant challenge for distribution system operators (DSO), especially
in medium voltage (MV) and low voltage (LV) grids. For
example, the German DSO Netze BW GmbH expects necessary investment costs between 1.5 and 1.9 billion Euros
to integrate additional DG units between the years 2012
and 2030 [1].
1.1. DG Integration and Increased Degree of Meshing
Several improvements in grid operation have been considered to reduce grid integration costs, such as controllable MV/LV substations, innovative HV/MV transformer
control or intelligent curtailment of DG feed-in [1]. It has
also been suggested that an increased degree of meshing
in the MV grid can facilitate the integration of DG by
improving the operational behaviour of the grid [2, 3, 4],
which could decrease the need for additional grid extension. But closed ring systems also require a more sophisticated protection system layout than open ring systems
[5, 6], which leads to additional costs. So while meshing
can be a technical solution for DG integration, the question arises if it is also economically feasible. It is therefore
necessary to study not only the qualitative effect of additional integration of DG by closed ring structures, but
to quantify the benefits in terms of saved grid extension
and compare them to the additional costs for a more sophisticated protection system layout. Because of the high
diversity of distribution grids, it is difficult to draw substantiated conclusions from exemplary case studies. For
statistically relevant results, a grid study has to be based
on the evaluation of a large number of grids. This can only
be achieved with a high degree of automation [7].
Preprint submitted to International Journal of Electrical Power & Energy Systems
February 6, 2018
1.2. Automated Comparison of Grid Concepts
In this paper we introduce a methodology that allows
the automated comparison of different grid concepts for
future power systems. We consider the current grid structure as well as the expected changes in future power systems, such as increasing installation of DG and increasing
share of underground cables. We then derive three different possible future grid structures following the three
different grid concepts: radial grid without switching station, radial grid with switching station and closed ring
grid. To ensure that all grids have equivalent technical
capabilities, all grids have to comply with the same technical constraints. These include constraints on voltage and
line loading in normal operation and in contingency operation as well constraints that limit the outage times in the
grid. The different grid plans are then compared with regard to their cost for primary and secondary equipment.
This methodology is applied to 44 real MV grid areas that
span about 4800 km of lines to reach a sound conclusion
about the technical and economic feasibility of each grid
concept. The presented approach takes the existing grid
infrastructure as well as a prognosis for the future development of the grid into account. The results allow to draw
conclusions about strategic decisions in grid planning and
adaptation of planning principles.
Figure 1: Basic structural MV grid elements: a. open ring, b. closed
ring, c. switching station ring.
Radial grid. A radial grid is built exclusively out of open
ring structures as shown in Figure 1 a. The radial structure ensures that short-circuit currents are always unidirectional, which makes fault detection and location easier than in grid structures with closed rings. We therefore consider a simple protection system layout with overcurrent protection systems at HV/MV substations (primary substations) for fault cut-off and non-directed shortcircuit indicators at the MV/LV substations (secondary
substations) for fault location in this concept.
Closed Ring grid. A closed ring grid has the same structure as a radial grid, only that some of the rings are operated as closed rings as shown in Figure 1 b. Since this
allows bi-directional short-circuit currents, the protection
system needs to be more sophisticated than in the open
ring case. We therefore assume an impedance protection
system in the primary substations and directional shortcircuit indicators in the secondary substations.
1.3. Overview
Switching Station grid. The switching station concept is a
radial grid concept with additional MV substations placed
in load or generation centres to stabilize the grid. Switching stations are MV substations with over-current protection devices that allow parallel operation of cables (see
Figure 1 c.). Apart from the additional protection system
in the MV substations, the protection system layout is
the same as in a radial grid. Using switching stations can
be seen as a way to allow closed rings in the grid without
changing the overall protection system layout by installing
additional circuit breakers in the MV grid.
The paper is organized as follows: Section 2 defines the
three grid concepts that are considered in this study and
Section 3 gives an overview of the overall methodology that
is used to replace switching stations. Section 4 introduces
the grid data used in the case study and the scenario for
future power systems which is applied. Section 5 describes
the optimization problem and its solution with heuristic
random search algorithms in general. Section 6 outlines
the specific multi-phase target grid optimization approach
in detail. Finally, the results of applying the methodology
to all grids are presented and interpreted in Section 7.
Section 8 gives a summary and an outlook of how the
presented methodology can be used in future studies.
3. Methodology for Replacing Switching Stations
We use the existing switching stations as starting point
for the analysis and optimize the respective area for the
three concepts. We assume that the switching station will
have to be renewed within the planning horizon, so that a
replacement investment will be necessary if the switching
station topology is maintained. The question then arises,
if it is more economical to avoid this replacement investment by changing the grid concept to a strictly radial or
a closed ring grid structure. Founding new switching stations in grid areas that are currently radial is out of the
scope of this paper. An example for the methodology
can be seen in Figure 2, where a current grid structure
with switching station is shown on the left. The grid area
around the switching station is now dismantled with respect to the planning horizon of the year 2030 (Figure 2,
2. Grid Concepts
Grid layout and protection system layout have to be
inter-coordinated to guarantee a safe grid operation [8].
For the grid layout, we consider three basic structural elements: open line rings, closed line rings and ring structures
with switching stations (see Figure 1). A grid concept is
then defined as the combination of these elements with an
appropriate protection system layout.
Every DSO defines its grid concept depending on individual requirements and boundary conditions. For this
study, we consider three different MV grid concepts:
2
Current Network Configuration
(with Switching Station)
Radial Target Network 2030
(with Switching Station)
Dismantled Network 2030
(with Switching Station)
Radial Target Network 2030
(without Switching Station)
Dismantled Network 2030
(without Switching Station
& Supply Lines)
Closed Ring Target Network 2030
(without Switching Station)
Primary Substation
Secondary Substation
Long Line
Switching Station
Automated Station
Supply Line
Closed Ring
Figure 2: Overview of target grid optimization methodology: the current grid structure (left) is dismantled with and without switching station
with assumptions for the grid state in 2030 (middle) and then target grids for switching station, open ring and closed ring grid concept is
optimized (right)
middle). Since asset data of the lines was not available,
the dismantling is based on the importance of the line for
the grid topology. Short lines which connect two secondary
substations will be necessary in any grid to ensure the supply of all stations. Longer supply lines however might be
replaced by different cable routes depending on the chosen concept. We therefore remove all long lines, which are
defined as lines with a length of over 2 km for this case
study. The methodology works the same with other criteria for which lines should be removed, for example by
installation date or standard type if this data is available.
To develop an alternative radial grid structure, we create
a second dismantled grid where the switching station with
all of its supply lines is removed in addition (Figure 2,
middle). From these two topologies, we then generate target grids for open ring, closed ring and switching stations
concepts as shown in Figure 2 on the right. The optimization process that is used to create these grids is outlined
in detail in Section 6.
level through 51 HV/MV transformers. The grids cover
a total of about 4800 km of lines and service loads with
about 770 MVA and DG with about 750 MVA installed
power (see Table 1). The grids are all located in southern
Germany and operated by the DSO Netze BW GmbH. All
MV grids are currently operated as an open ring grid with
switching stations in load or generation centres. There are
a total of 49 switching stations in the grid data. The grid
area around each switching station is taken as a starting
point for the analysis as outlined in Section 3. We model
the situation for the planning horizon of 2030 by applying a load prognosis as well as installing additional DG
based on a prognosis for the expected expansion of phoGrid
Group
1
2
3
4
5
Total
4. Grid Data and Scenario
In this paper we study 5 MV grid groups at the 20 kV
voltage level which are connected to the 110 kV voltage
HV/MV
Transf.
15
8
8
12
8
51
Switching
Stations
16
8
8
7
10
49
Load
[MVA]
249
87
151
169
117
773
DG
[MVA]
183
275
231
28
34
751
Table 1: Parameters of considered grid groups
3
Lines
[km]
1363
907
1329
655
525
4779
Grid Component
Cable NA2XS2Y 3x1x300
MV switching station
Communication link
Directed short-circuit indicator (update)
Impedance protection (update)
tovoltaic systems (PV) and wind power plants [1]. Table
1 shows that the load remains predominant in some grids
while in others DG clearly dominates. The set of grid
groups was chosen in order to form a good representation of the expected situation in 2030. The DSOs internal
planning principles dictate that new line trails are always
built as underground cables. We therefore assume that all
new lines are built as cables and overhead lines will be replaced by underground cables until the planning horizon
in 2030. The grid is modelled and analysed in the open
source power system analysis tool pandapower [9].
Annual Cost
7,000 e/km
35,900 e
1,200 e
30 e/Station
400 e/Feeder
Table 2: Annual costs of elements considered in this study
5.3. Cost Assumptions
Different measures, such as cables or communication
links, differ in operational costs, investment costs and life
expectancy. To make the costs of these measures comparable, we chose the Net Equivalent Uniform Annual Cost
[10] as a measure of cost for all grid elements. The annual
costs Cpa are equal to the sum of annual operational costs
Cop,pa [EU R/a] and annual investment costs Cinv,pa [EU R/a]:
5. Heuristic Grid Optimization
The goal of this paper is to economically compare different grid concepts taking into account the currently existing grid structure. Figure 2 shows an example of such a
topology optimization. This section outlines, how the optimization problem is formulated and solved with a heuristic
random search algorithm.
Cpa = Cop,pa + Cinv,pa
(1)
While operational costs are usually already given as yearly
costs, investment costs are typically given as total costs
Cinv,total [EU R]. The annual investment costs can be calculated by discounting the total investment costs with the
calculation interest rate i [%] over the life expectancy m [a]
of the component:
5.1. Technical Constraints
An economic comparison is only valid if the compared
grids are equivalent with respect to their technical capabilities. To ensure all grids meet the technical requirements, we introduce constraints regarding grid topology,
operational behaviour in normal operation, operational behaviour in contingency operation and outage times. The
goal of the optimization is then to find a grid that complies with all these constraints at the minimal possible
costs. Which constraints are applied in the different optimization phases is explained in detail in Section 6.
Cinv,pa = Cinv,total ·
(1 + i/100)m · i/100
(1 + i/100)m − 1
(2)
The cost assumptions used for all relevant components in
this study are depicted in Table 2. The cable costs are
assumed to be constant per kilometre for both adding new
lines an replacing existing lines. This is because the main
cost factor is the excavator work, which has to be carried out when replacing an existing line as well as when
laying a new cable. The full switching station costs are
assumed even if the switching station already exists in the
current grid, in accordance with the strategic viewpoint of
expected renewal outlined in Section 3.
5.2. Measures
To generate realistic grid structures, all options that a
grid planner has in the planning process have to be considered as degrees of freedom in the optimization. In this
study, we consider the following measures:
• Replacing existing overhead lines with cables
• Replacing existing cables with cables of higher diameter
5.4. Heuristic Solution
• Optimizing the position of normally open switches
(sectioning points)
We formulate the DEP as a combinatorial optimization
problem with non-linear constraints. Given a fixed number of possible measures, the goal is to find the cheapest
subset of these measures that fulfils all constraints. It
is not considered feasible to formulate this optimization
problem in a closed form and solve it analytically. Instead,
heuristic algorithms such as genetic algorithms [11, 12, 13],
evolutionary algorithms [14], tabu search [15, 16], particle
swarm optimization [17, 18] or artificial immune systems
[19] are popular methods for its solution. Most of these
studies assume the radial structure of the grid as a constraint [20, 11, 15, 21, 16, 13, 12, 18] and are therefore not
designed to optimize and compare grids in other grid concepts. The approach described in [7] is however flexible
enough to handle the heuristic optimization with varying
• Adding cables in parallel to existing line trails
• Adding cables in new line trails
• Equipping secondary substations with communication
links to improve reliability
• Removing or renewing switching stations
Each of the possible measures can either be applied or not
applied, which leads to the formulation of the Distribution system Expansion Problem (DEP) as a combinatorial
optimization problem [7]. The details of the different measure and at what point in the optimization process they
are deployed is explained in detail in Section 6.
4
Current Grid
Preparation and
Dismantling
With
Switching
Station
Without
Switching
Station
Secondary Substation
1a
1...m
Topology
Optimization
1b
Topology
Optimization
2a
Reconfiguration
Reinforcement
2b
Reconfiguration
Reinforcement
3a
Automation for
Reliability
Switching
Station
Network
5
1...m
3b
Line
Primary Substation
Supply Line
Switching Station
New Line
Figure 4: Considered new line trails in the example grid area
6.1.1. Topological Constraints
Topological constraints specify requirements for the
topological structure of the grid without taking electric
parameters into account. They can be analysed with
graph searches using the pandapower topology package [9].
Which topological constraints are relevant depend on the
grid concept and protection system layout. Here, we assume the following constraints:
4
Automation for
Reliability
Meshing
Radial
Network
Closed Ring
Network
Cost Comparison
Supply connection. All secondary substations have to be
connected to an HV/MV substation to be supplied. This
means there has to be at least one path from each station
to an external grid connection.
Figure 3: Methodology for the calculation and comparison of the
three grid structures
measures and constraints, so that a study of different grid
concepts is possible. The DEP is solved using an Iterated Local Search (ILS) algorithm with Hill Climbing as
local search. This paper concentrates on the application
of the optimization framework to the specific questions of
structural optimization in different grid concepts. Detailed
information on the formulation of the optimization problem, codification of measures, definition of the step-wise
cost function and neighbourhood function can be found in
[7] and [22].
Contingency supply. All secondary substations must have
a second connection for resupply in case of a single contingency, which is provided by ring structures in the grid.
Secondary substations which are located on stubs and
therefore do not have a contingency connection in the current configuration are exempted from this rule. Note that
this constraint only guarantees the topological possibility
of a resupply, operational constraints in the contingency
state are addressed in section 6.2.2.
Radiality. The grid has to be radial, in that no feeders
can be galvanically connected through closed rings in the
MV grid. Rings are only permitted if they go through
a switching station. Closed ring structures as shown in
Figure 1 b. are note permitted, as they will be investigated
separately in phase 4 (see Section 6.4).
6. Multiphase Target Grid Optimization
Because of the complexity of the combinatorial optimization problem, it is not feasible to optimize sectioning
points, new line trails, line replacement, parallel lines and
secondary substation automation all in one optimization.
We therefore split the problem into several sub-problems
that are solved subsequently. An overview of the multiphase optimization process that is used to find grid structures for the three concepts starting from the current grid
configuration is shown in Figure 3. Each phase is explained
in this section with the help of the example grid area shown
in Figure 2.
6.1.2. Line Trail Optimization
To construct grid structures that comply with the constraints outlined above from a dismantled grid, new lines
are added between two secondary substations which are
not currently connected. The length of a new line is calculated as the air-line distance between the stations multiplied with a factor of 1.5. The costs per kilometre line
are given in Table 2. We find possible new line trails to
complete the grid with a Delaunay-triangulation [23] between all stations from which a line was removed. The
considered line trails in the example grid are shown in
the example in Figure 4 for the layout with and without
switching station. Since there are a lot of possible combinations and the topologies cannot be easily compared, we
6.1. Phase 1: Topology Optimization
A valid grid has to comply with several constraints regarding its structural layout. In the first step, we therefore
optimize the topology of the grid.
5
Scaling Load
Scaling DG (PV)
Scaling DG (Wind)
Transformer Setpoint
Peak Load
1.0
0.0
0.0
1.0 pu
Peak Generation
0.3
0.8
1.0
1.05 pu
Figure 5: Resupply switching sequence for a line fault in an open
ring: a. normal operation, b. line fault with protection trip, c. fault
isolation, d. resupply
Table 3: Definition of worst-case scenarios
consider several different solutions. Specifically, we create
50 different topologies for each grid area. Figure 6 shows
three of the 50 topologies for the example grid area. All
steps that are explained in the following are conducted 50
times per studied grid area.
scenario, since the whole ring has to be supplied through
only one feeder. A typical switching sequence for a line
outage in a radial grid is shown in Figure 5. When a line
fails, the circuit breaker in the primary substation opens to
disrupt the short-circuit current, and thereby cuts the affected half ring from power supply (Figure 5 b.). After the
fault is located, it is isolated by manually opening the loadbreak switches at the secondary substations connected to
the faulted line segment (Figure 5 c.). Finally, the circuit
breaker in the primary substation as well as the sectioning point are closed to allow a resupply of all stations on
the ring (Figure 5 d.). This switching sequence is automatically carried out for all feeders using the pandapower
topology package to find a resupplied state. Compliance
with contingency constraints is checked with a power flow
in the resupplied state. The constraints are checked for the
peak load case as defined in Table 3. The peak generation
scenario is not checked in contingency situations because
DG are not considered to be subject to the SCP.
6.2. Phase 2: Reconfiguration and Reinforcement
After the first optimization step, the grids comply with
topological constraints, so that all stations are supplied
and the grid is radial. In the second optimization step, it
is ensured that the electric parameters of the grids comply
with the grid planning principles. This is achieved by grid
reconfiguration and reinforcement.
6.2.1. Normal Operation Constraints
Constraints for equipment loading and bus voltages have
to be complied with in every possible load or generation
scenario. We therefore define two worst-case scenarios for
the peak load and peak generation situation. The simultaneous factors used to generate the worst case scenarios
can be seen in Table 3. The scaling factors for the loads
refer to the drag pointer measurements at the secondary
substations and to the installed power of DG. Since we
assume V(P) transformer control, the voltage set point for
the transformer voltage controller depends on the active
power flow and is therefore different in the two worst-case
scenarios. Adherence with the power flow constraints in
normal operation is checked with one power flow for each
worst-case scenario using pandapower [9]. In both cases,
all bus voltages have to be within the voltage band defined
by the DSO and the line loadings have to be below 100 %.
6.2.3. Optimization of Operational Behaviour
We consider switching and cable measures to improve
the electric parameters in a grid structure and ensure compliance with constraints for normal and contingency operation.
Sectioning Point Optimization. In a radial grid structure,
sectioning points are necessary to fulfill the topological
constraints, specifically to avoid direct galvanic connection of feeders or transformers. Sectioning points can be
relocated to improve the operational behaviour of the grid.
The heuristic optimization aims to find a switching state
which minimizes the voltage and line loading violations
while still complying with all topological constraints. Details about the used composite objective function are described in [22].
6.2.2. Constraints in Contingency Operation
The Single Contingency Policy (SCP) dictates, that the
power system must continue functioning after the outage
of one power system element. The topological contingency
constraint ensures that a resupply is possible after the failure of a line through a backup connection (see Section
6.1.1). However, it also has to be ensured that the grid
operates safely in the resupplied state. Specifically, the
bus voltages have to be within the contingency voltage
band (which is usually broader than the voltage band for
normal operation) and the line loadings have still to be
below 100 %.
To check this constraint, we calculate a resupply scenario for each feeder line directly connected to transformer
or switching station. A fault of such a line is the worst case
Cable Measures. Existing line segments can be replaced by
a new cable with higher ampacity, to mitigate line overloading, or with a lower impedance, to mitigate voltage
problems. If line replacement is not enough to mitigate
all constraint violations, additional lines can be added to
the grid in parallel to existing line routes. Parallel lines
can only be connected to secondary substations and the
radiality constraints have to be respected.
Constraint violations are first mitigated through sectioning point optimization, since opening and closing
switches is not associated with costs. If not all constraint
6
failure rates of the components are taken from statistical
analysis where it is given as a function of the line type
and, in case of underground cables, the insulation material. The effect of a line failure can be calculated from
the resupply switching sequence as shown in Figure 5. By
assuming time constants for fault location, on-site switching and remote switching, we calculate the outage time for
each of the m stations in the case of a failure of each of
the n lines in the grid. By weighting the calculated outage time with the yearly failure rate, we get the expected
yearly outage time tout,i [h/a] for a station i:
tout,i =
n
X
hk · tout,ki
(3)
k=0
where hk [1/a] is the yearly failure rate of line k and
tout,ik [h] is the outage time for station i after a failure
of line k. The yearly outage energy Eout,i [kW h/a] for a
station i can then be calculated with the installed power
Pk as:
Eout,i = Pk · tout,i
Secondary Substation
Primary Substation
(4)
The ASIDI of the whole grid is then defined as the sum of
all the outage energy in all m stations in relation to the
total installed power:
Pm
i=0 Eout,i
(5)
ASIDI = P
m
i=0 Pi
Figure 6: Necessary reinforcement for the three example grids with
and without switching station
violations can be mitigated through reconfiguration, the
grid has to be reinforced with additional cable measures
until all violations are mitigated.
The necessary reinforcement for the three example grids
can be seen in Figure 6. It can be seen that there is overall
less need for reinforcement in the switching station grid,
since the switching station serves as a junction that stabilizes the grid.
To prevent any decline in service reliability, we define
the following reliability constraints:
System wide Criterion. To assure that the target grid
topologies are at least as reliable as the current grid, the
ASIDI of the grid must not increase compared to the current grid.
After the second optimization step, all grids comply with
topological constraints as well as operational constraints
for normal and contingency operation. To ensure that the
grids are also sufficiently reliable, the grid is equipped with
communication links to speed up resupply and reduce outage times if necessary.
Station Criterion. To assure that no individual station
suffers a great setback in service reliability compared to
the current grid, the expected outage energy for each station is limited to an allowed maximum of Eout,max . For
stations that already violate this constraint in current grid
configuration, no further increase of the outage energy is
allowed. In this case study, we use a maximum outage
energy of Eout,max = 150 kWh/a.
6.3.1. Reliability Constraints
The SCP demands that the grid complies with the operational constraints in the resupplied state. It does however
not draw any conclusion about the frequency of fault occurrences in the grid or about which secondary substations
will experience outages for how long. In this optimization
step, it is ensured that the grid structure is sufficiently
reliable with respect to outage frequency and restoration
times. The service reliability can be measured in reliability
figures, the most prominent of which is the Expected Average System Interruption Duration Index (ASIDI), which
can be computed with a Failure mode and effects analysis
(FMEA) [24]. The FMEA takes into account the failure
rates of components as well as the effect of the failure. The
6.3.2. Automating Resupply
Secondary substations can be equipped with a communication link to allow remote controllable switching of the
sectioning point. This measure accelerates the fault isolation and resupply process, so that it can be used to improve the service reliability. If there are violations of the
reliability constraints in a feeder, the secondary substation
in the load centre of the feeder is chosen for automation.
This results in a partitioning, which confines the spread
of the fault area and speeds up the resupply process. secondary substations are repeatedly selected until all constraints are met. The resulting grids now comply with all
planning constraints and are used as the reference grid for
the radial and switching station grid respectively.
6.3. Phase 3: Automation for Reliability
7
122.1
101.3
Closed Ring
Cost [Thousand EUR / Year]
100
Switching Station
Network Concept
Radial Network
120
Cabling
Switching
Station
Secondary
Equipment
0
50
100
Cost [Thousand EUR /Year]
150
200
90.7
80
60
40
20
0
Radial
Network
Closed
Rings
Switching
Station
Figure 7: Cost comparison for example grid area: costs for all 50 separately optimized grids (left) and cost minimal reference grid (right) for
each concept
solution, since the additional cost for secondary equipment
is greater than the cost savings in grid reinforcement. We
therefore conclude, that the best concept for the example
grid area is a radial grid without a switching station.
6.4. Phase 4: Meshing
In a last step, we construct a closed ring structure from
the radial grid. Closed rings are only allowed with a more
sophisticated protection system as detailed in Section 2.
The costs for updating the protection system are considered as per Table 2. In return, the meshing constraint
is now relaxed so that the meshing of two feeders is allowed even without a switching station. The Iterated Local Search meta-heuristic is used to create a closed ring
solution from each of the fifty valid radial topologies. As
explained in Section 2, closing sectioning points leads to a
decrease in service reliability, so that compliance with the
reliability constraints cannot be guaranteed after sectioning points are closed. We therefore demand that for every
sectioning point that is closed, the associated secondary
substations has to be automated. Since the open sectioning point is replaced with an automated sectioning point,
the grid is considered at least as reliable as the radial grid.
The optimization creates fifty closed ring grids for each
grid area and the cost minimal grid is chosen as the reference grid for the closed ring concept.
7. Results
The methodology outlined in Section 6 for one grid area
is applied to all switching stations in the case study grid.
Of the 49 switching stations, 5 where dismissed due to
faulty or inconsistent data. The methodology described in
Section 6 is applied to all remaining 44 switching station
grid areas. As 50 topologies are calculated for all the three
network concepts in each of the 44 grid areas, a total of
3 · 50 · 44 = 6600 independent network plans are compiled
for this study. To provide enough computational power,
calculations were carried out on a computation cluster and
over 840 million power flows were conducted to obtain the
results. The results are now interpreted to draw a conclusion about the overall cost efficiency of the different MV
grid concepts.
6.5. Phase 5: Cost Comparison
7.1. Switching Stations
We have constructed fifty grids for each of the three
meshing concepts which all fulfil the planning constraints,
but differ in their structure and therefore in their total
costs. Figure 7, shows the costs of the fifty topologies for
each grid concept on the left. We select the cost minimal
grid for each concept as the reference grid. These three
grids are then compared to draw a conclusion about the
differences of the three concepts. This comparison is shown
for the example area in Figure 7 on the right. We can
see, that the best open ring grid is cheaper than the best
switching station solution for the example grid by around
31,400 e per year. We can also see, that the best closed
ring solution is more expensive than the best open ring
We compare radial and switching station grids to evaluate the profitability of switching stations. The switching
station improves reliability because of the included protection systems as well as the operational behaviour by allowing closed line rings. On the other hand, the costs for
the switching station are equivalent to over 5 km of cable
and it needs additional supply lines to create a junction
in the grid. The comparison in Figure 8 shows, that the
radial grid is more cost efficient in all 44 analysed cases.
This means that none of the switching stations are able to
save enough grid reinforcement for it to be cost-efficient.
Reliability only seems to be a minor impact factor, since
compliance with the reliability constraints can be achieved
8
are outweighed by the costs for the switching station itself and the necessary supply lines. We therefore conclude
that it is more cost-efficient to develop the grid towards
a radial structure in the long run, avoiding replacement
investments for switching stations that have reached the
end of their life cycle. The closed ring concept is only cost
effective in few grid areas, where the expected savings are
small compared to the additional cost in other grid areas.
We therefore conclude, that the radial grid in combination with selective automation of secondary substations is
the most cost efficient mode of operation for the analysed
grids. The contributions of this paper can be summarized
as follows:
1. Automated target grid planning: the methodology
presented in this paper allows to automatically compile target grid plans based on geographical information and a wide range of technical constraints. This
includes constraints for grid topology, normal operation power flow parameters, contingency operation
power flow parameters and service reliability.
Figure 8: Cost difference between switching station and radial grid
for all 44 grid areas (GA)
with 2.4 automated stations on average, which is the cost
equivalent of 0.4 km of cable.
We therefore conclude, that a radial grid with automated secondary substations is preferable to a structure
with switching stations. The radial grid structure with secondary substation automation is equivalent to the switching station grid in relation to grid operation and service
reliability, but comes with significantly lower investment
and maintenance costs.
2. Proof of concept: the successful application of the approach to 44 grid areas with about 4800 km of lines
demonstrates the practical applicability of the approach. The high degree of automation makes the
approach feasible for large-scale grid studies as well
as for grid planning assistance systems.
3. Comparison of grid concepts: The results of the case
study allow to draw conclusions about the technoeconomical differences of different MV grid concepts.
A higher degree of meshing for the integration of DG
was not found to be cost-efficient, as the additional
costs for secondary equipment outweighed the cost
savings in network reinforcement.
7.2. Closed Rings
We now compare the closed ring grids with the open
ring grids to evaluate the profitability of a closed ring concept. The comparison for all 44 cases is shown in Figure
9. In five grid areas, the closed ring solution is economical (green). In ten grids, the costs for the upgrade of
the protection and fault location systems exceed the savings in grid reinforcement (orange). In the remaining 29
grids, closed rings do not lead to any savings in grid reinforcement and are therefore not cost-efficient (red). In
conclusion, closed rings are only cost-efficient in very few
grids, and even in those grids the saving potential with less
than 4,000 e per year in average is not very high. Since
the DSO usually aims to have the same grid concept in all
MV grids to allow harmonized procedures in grid planning
and operation, the radial structure is more cost effective
than the closed ring structure for the overall grid.
It should be noted that the conclusion about the grid concepts are only valid for the considered grids and under the
considered boundary conditions. Even though the grids
where chosen with the aim to include structurally different
grids, a generalization for other grid groups is not easily
possible. The boundary conditions with regard to protection system layout, spatial distribution of loads and DG
or ratio of underground cables and overhead lines might
differ significantly in other grids, especially of a different
DSO. While meshed operation for the integration of DG
could still be cost-efficient in other grids, our study does
not confirm the assumption that meshing is always better
for integration of DG. Instead, a comprehensive consideration of expected grid extension cost and protection system
layout is necessary to determine the best mode of operation for the future power system. The approach could be
further extended in the future to include transformation
paths from the current power system to a future target
grid or probabilistic scenarios for DG installation.
8. Conclusion
In this study we economically compare target grids for
radial, closed ring and switching station grids to come to
a conclusion about the optimal grid concept in 44 real
MV grid areas for a planning horizon in 2030. The results
show, that switching stations are economically inefficient
under the assumed constraints, since the technical benefits
9
GA 01
GA 02
GA 03
GA 04
GA 05
GA 06
GA 08
GA 07
GA 09
GA 10
GA 11
GA 12
GA 13
GA 14
GA 15
GA 16
GA 17
GA 18
GA 19
GA 20
GA 21
GA 22
GA 23
GA 24
GA 25
GA 26
GA 27
GA 28
GA 29
GA 30
GA 31
GA 32
GA 33
GA 34
GA 35
GA 36
GA 37
GA 38
GA 39
GA 40
GA 41
GA 42
GA 43
GA 44
Figure 9: Cost difference between closed ring and radial grid for all 44 grid areas (GA)
Acknowledgement
[7] A. Scheidler, L. Thurner, M. Braun, Automated distribution
system planning for large-scale network integration studies, IET
Renewable Power Generation (submitted)Preprint.
URL https://arxiv.org/abs/1711.03331
[8] J. Schlabbach, K.-H. Rofalski, Power System Engineering:
Planning, Design, and Operation of Power Systems and Equipment, Wiley-VCH Verlag GmbH & Co. KGaA, 2008.
[9] L. Thurner, A. Scheidler, F. Schäfer, J.-H. Menke, J. Dollichon,
F. Meier, S. Meinecke, M. Braun, pandapower - an Open Source
Python Tool for Convenient Modeling, Analysis and Optimization of Electric Power Systems, preprint (2017).
URL https://arxiv.org/abs/1709.06743
[10] H. Seifi, M. S. Sepasian, Electric Power System Planning - Issues, Algorithms and Solutions, Springer, 2011.
[11] V. Camargo, M. Lavorato, R. Romero, Specialized genetic algorithm to solve the electrical distribution system expansion
planning, in: 2013 IEEE Power Energy Society General Meeting, 2013, pp. 1–5. doi:10.1109/PESMG.2013.6672615.
[12] V. Miranda, J. V. Ranito, L. M. Proenca, Genetic algorithms in
optimal multistage distribution network planning, IEEE Transactions on Power Systems 9 (4) (1994) 1927–1933. doi:10.1109/
59.331452.
[13] H. Falaghi, C. Singh, M.-R. Haghifam, M. Ramezani, Dg integrated multistage distribution system expansion planning,
International Journal of Electrical Power & Energy Systems
33 (8) (2011) 1489 – 1497. doi:http://dx.doi.org/10.1016/
j.ijepes.2011.06.031.
[14] Z. Zmijarević, M. Skok, H. Keko, D. Škrlec, S. Krajcar, N. LangKosić, G. S. čki, A comprehensive methodology for long-term
planning of distribution networks with intrinsic contingency
support, in: 18th International Conference and Exhibition on
Electricity Distribution, CIRED, Turin, Italy, 2005, pp. 1–5.
doi:10.1049/cp:20051353.
[15] A. M. Cossi, L. G. W. D. Silva, R. A. R. Lazaro, J. R. S.
Mantovani, Primary power distribution systems planning taking into account reliability, operation and expansion costs, IET
Generation, Transmission Distribution 6 (3) (2012) 274–284.
doi:10.1049/iet-gtd.2010.0666.
This work was supported by the German Federal Ministry for Economic Affairs and Energy and the Projektträger Jülich GmbH (PTJ) within the framework of the
project Smart Grid Models (FKZ: 0325616).
References
References
[1] C. Rehtanz, M. Greve, B. Gwisdorf, A. El-Hadidy, J. Kays,
M. Kch, V. Liebenau, J. Teuwsen, EnBW-Verteilnetzstudie
(2014).
[2] G. Celli, F. Pilo, G. Pisano, V. Allegranza, R. Cicoria, A. Iaria,
Meshed vs. radial mv distribution network in presence of large
amount of dg, in: Power Systems Conference and Exposition,
2004. IEEE PES, 2004, pp. 709–714 vol.2. doi:10.1109/PSCE.
2004.1397664.
[3] S. Repo, A. Nikander, H. Laaksonen, P. Jrventausta, A method
to increase the integration capacity of distributed generation on
weak distribution networks, in: 17th International Conference
on Electricity Distribution, 2003.
[4] A. Nikander, R. E. P. O. Sami, P. Järventausta, Utilizing the
ring operation mode of medium voltage distribution feeders,
2003.
[5] F. Viawan, D. Karlsson, A. Sannino, J. Daalder, Protection
scheme for meshed distribution systems with high penetration
of distributed generation, in: Power Systems Conference: Advanced Metering, Protection, Control, Communication, and
Distributed Resources, 2006. PS ’06, 2006, pp. 99–104. doi:
10.1109/PSAMP.2006.285378.
[6] S. Salman, S. Tan, Comparative study of protection requirements of active distribution networks using radial and ring operations, in: Power Tech, 2007 IEEE Lausanne, 2007, pp. 1182–
1186. doi:10.1109/PCT.2007.4538483.
10
[16] B. R. P. Junior, A. M. Cossi, J. Contreras, J. R. S. Mantovani,
Multiobjective multistage distribution system planning using
tabu search, IET Generation, Transmission Distribution 8 (1)
(2014) 35–45. doi:10.1049/iet-gtd.2013.0115.
[17] S. Ganguly, N. Sahoo, D. Das, A novel multi-objective pso
for electrical distribution system planning incorporating distributed generation, Energy Systems 1 (3) (2010) 291–337.
[18] N. Sahoo, S. Ganguly, D. Das, Multi-objective planning of electrical distribution systems incorporating sectionalizing switches
and tie-lines using particle swarm optimization, Swarm and
Evolutionary Computation 3 (2012) 15 – 32. doi:http://dx.
doi.org/10.1016/j.swevo.2011.11.002.
[19] H. Keko, M. Skok, D. Skrlec, Solving the distribution network
routing problem with artificial immune systems, in: Proceedings
of the 12th IEEE Mediterranean Electrotechnical Conference
(MELECON), Vol. 3, Dubrovnik, Croatia, 2004, pp. 959–962
Vol.3. doi:10.1109/MELCON.2004.1348212.
[20] N. G. Boulaxis, M. P. Papadopoulos, Optimal feeder routing in
distribution system planning using dynamic programming technique and gis facilities, IEEE Transactions on Power Delivery
17 (1) (2002) 242–247. doi:10.1109/61.974213.
[21] E. Diaz-Dorado, J. Cidras, E. Miguez, Application of evolutionary algorithms for the planning of urban distribution networks
of medium voltage, IEEE Transactions on Power Systems 17 (3)
(2002) 879–884. doi:10.1109/TPWRS.2002.800975.
[22] L. Thurner, A. Scheidler, A. Probst, M. Braun, IET Generation, Transmission & Distribution 11 (2017) 4264–4273(9).
[link].
URL
http://digital-library.theiet.org/content/
journals/10.1049/iet-gtd.2017.0729
[23] M. de Berg, O. Cheong, M. van Krefeld, M. Overmars, Computational Geometry: Algorithms and Applications, 2008.
[24] R. E. Brown, Electric power distribution reliability, 2nd Edition,
Vol. 31 of Power engineering, CRC, Boca Raton, 2009.
11
| 5cs.CE
|
1
Diffusion Based Cooperative Molecular Communication in Nano-Networks
arXiv:1710.01882v2 [cs.IT] 15 Oct 2017
Neeraj Varshney, Student Member, IEEE, Adarsh Patel, Student Member, IEEE, and Aditya K.
Jagannatham, Member, IEEE
Abstract—This work presents a novel diffusion based dualphase molecular communication system where the source leverages multiple cooperating nanomachines to improve the end-toend reliability of communication. The Neyman-Pearson Likelihood Ratio Tests are derived for each of the cooperative as
well as the destination nanomachines in the presence of multiuser interference. Further, to characterize the performance of
the aforementioned system, closed form expressions are derived
for the probabilities of detection, false alarm at the individual
cooperative, destination nanomachines, as well as the overall
end-to-end probability of error. Simulation results demonstrate
a significant improvement in the end-to-end performance of the
proposed cooperative framework in comparison to multiple-input
single-output and single-input single-output molecular communication scenarios in the existing literature.
Index Terms—Cooperation, diffusion, Likelihood Ratio Test
(LRT), molecular communication.
I. I NTRODUCTION
Nanoscale molecular communication has gained significant
prominence in recent times due to its ability to provide
novel solutions for various problems arising in biomedical,
industrial, and surveillance scenarios. Efficient drug delivery,
as described in [1], is one such example of its envisaged
potential. Several research efforts [2]–[6] have been devoted
to developing models as well as analyzing the performance
of diffusion based molecular systems. It has been shown in
[7] that the distance between two nanomachines significantly
affects the performance of communication, since the molecular
concentration decays inversely as the cube of this distance.
To overcome this impediment, the work in [8] proposed a
novel multiple-input multiple-output (MIMO) molecular communication framework, wherein a noticeable performance improvement has been demonstrated through suitable allocation
of molecules among the transmitting nodes. However, the
work therein focuses exclusively on applications of transmit/
receive diversity and spatial multiplexing in molecular communication, while not exploring cooperative molecular communication that can significantly enhance the communication
range with improvement in the reliability through cooperative
diversity.
Some works in the existing literature [9], [10] and the
references therein have analyzed the performance of relayassisted molecular communication systems in terms of their
error rate, capacity, etc. However, to the best of our knowledge,
none of the works in the existing literature consider dual-phase
multiple half-duplex nanomachine assisted cooperative molecular communication in the presence of multi-user interference
(MUI) and exploit cooperative diversity, which is the key focus
of this work.
Neeraj Varshney, Adarsh Patel, and Aditya K. Jagannatham are with the
Department of Electrical Engineering, Indian Institute of Technology Kanpur,
Kanpur UP 208016, India (e-mail:{neerajv,adarsh,adityaj}@iitk.ac.in).
Towards this objective, the optimal Likelihood Ratio Test
(LRT) based decision rule at each cooperating nanomachine
is initially derived, followed by a characterization of its
probabilities of detection and false alarm. Subsequently, the
optimal fusion rule is derived for the destination nanomachine
incorporating also the probabilities of false alarm and detection
at each of the cooperating nanomachines. Thus, the framework
developed is practically applicable, unlike the selective decode
and forward protocol considered in works such as [11], [12]
that assume retransmission only when the symbol is accurately
decoded at the relay. Finally, closed form expressions are
derived for the end-to-end probabilities of detection, false
alarm, and probability of error for the cooperative scenario
under consideration. Simulation results demonstrate the improved performance of the proposed cooperative nanosystem
in comparison to the existing multiple-input single-output
(MISO) and single-input multiple-output (SIMO) systems [8].
Further, while this work employs multiple molecule types
for the source and cooperating nanomachines, it results in a
simplistic linear combining based decision rule and also does
not require a large capacity backhaul that is necessitated in [8].
Due to space limitations, detailed derivations of some results
are given in a supplemental technical report [13].
II. D IFFUSION BASED C OOPERATIVE M OLECULAR
C OMMUNICATION S YSTEM M ODEL
Consider a diffusion based molecular communication system with N cooperating nanomachines R1 , R2 , · · · , RN as
shown schematically in Fig. 1. Further, it can be noted
that the diffusion takes place in 3-dimensional and infinite
space. The N intermediate nanomachines cooperate with the
source nanomachine to relay its information to the destination nanomachine. The transmitter is modeled as a point
source while the the destination node is a perfectly absorbing
surface. Further, the relay nodes during reception act as
perfect absorbing surfaces while during transmission act as
point sources. End-to-end communication between the source
and destination nanomachines occurs in two phases. In first
phase, the source nanomachine emits either Q0 molecules in
the propagation medium for information symbol 1 generated
with a prior probability β or remains silent for information
symbol 0. In the second phase, the intermediate nanomachines
Ri , i = 1, 2, · · · , N that employ decode-and-forwarding protocol, decode the symbol using the concentration received
from the source followed by retransmission to the destination nanomachine using different types of molecules1 . As
described in [14], nanomachines such as eukaryotic cells can
1 The use of different types of molecules at each transmitting nanomachine
allows the cooperative nanomachines to operate either in half or full duplex
modes [10].
where ηei ∼ N (e
µi , σ
ei2 ) represents the MUI at the destination
nanomachine and Qi is the number of molecules emitted
by node Ri corresponding to information symbol 1. Similar
to Q0 hp (di ) in (2), the peak concentration Qi hp (dei ) at the
destination nanomachine corresponding to the emission by Ri
is obtained as
3/2
3
−3
e
e
,
(4)
Qi hp (di ) = Qi (di )
2πe
Fig. 1. Schematic diagram of a cooperative molecular nano-network
with di and dei denoting distances from source nanomachine to Ri ,
and Ri to destination nanomachine respectively.
be genetically modified to emit different types of molecules.
The molecules released by the transmitter nanomachine are
assumed to diffuse freely through the propagation medium
with Brownian motion and are absorbed by the receiver once
they reach it. Furthermore, similar to [8], this work also
assumes that the inter symbol interference (ISI) and receiver
noise are suppressed and negligible in comparison to the MUI.
The effect of ISI can be ignored by considering the symbol
duration Ts to be significantly larger than tp , where tp is
the time instant when the channel impulse response has a
maximum as can be seen in [15, Fig.3]. Further, since such
nanonetworks are expected to operate in a distributed and
uncoordinated manner, this work focuses on MUI and assumes
ISI, noise effects to be negligible in comparison.
The peak concentration sensed at the ith cooperating
nanomachine Ri corresponding to transmission by the source
nanomachine in the first phase is given as [8]
ci = x0 Q0 hp (di ) + ηi ,
(1)
where x0 ∈ {0, 1} denotes the binary information symbol.
The term ηi represents the MUI at the cooperative node Ri .
Similar to [8], assuming the MUI to arise from a large number
of interfering sources, it can be modeled as a Gaussian random
variable with mean µi and variance σi2 , denoted by N (µi , σi2 ),
as per the central limit theorem (CLT) [16]. The molecular
concentration is given as
d2i
1
−
,
Q0 h(t) = Q0
3 exp
4Dt
(4πDt) 2
where di denotes the distance between the source 2and the ith
d
node Ri . The peak concentration is given at t = 6Di , obtained
by solving dh(t)
dt = 0. The peak concentration is obtained as
3/2
3
−3
.
(2)
Q0 hp (di ) = Qdi
2πe
Each cooperative nanomachine employs the minimum probability of error criterion for decoding the transmitted symbol
followed by retransmission of the decoded symbol x
bi ∈ {0, 1}
to the destination nanomachine. Therefore, the peak concentration sensed at the destination corresponding to the ith
cooperating nanomachine Ri is
ci = x
e
bi Qi hp (dei ) + ηei ,
(3)
where dei is distance of the destination from Ri . The end-to-end
error rate of the cooperative molecular communication system
depends on the detection performance of the individual cooperative nanomachines. The next section begins by determining
the probabilities of detection and false alarm at Ri .
III. D ETECTION AND E RROR R ATE A NALYSIS
WITH
MUI
The problem of sensing the peak concentration at the
cooperative nanomachine Ri in (1) can be formulated as the
binary hypothesis testing problem
H0 : ci = ηi
(5)
H1 : ci = Q0 hp (di ) + ηi ,
where the null and alternative hypotheses H0 , H1 correspond
to the binary symbols 0, 1 respectively. The peak concentration
ci at the ith cooperative nanomachine corresponding to the
individual hypotheses is distributed as
H0 : ci ∼ N (µi , σi2 )
(6)
H1 : ci ∼ N (Q0 hp (di ) + µi , σi2 ).
The optimal detection statistic Λ(ci ) for the cooperating
nanomachine Ri can be obtained employing the NeymanPearson LRT and is given as
p(ci |H1 ) H1
1−β
Λ(ci ) = ln
.
(7)
≷ ln
p(ci |H0 ) H0
β
Substituting PDFs p(ci |H0 ) and p(ci |H1 ) from (6), the logarithm of the LRT above can be evaluated as
Λ(ci )=[−(ci − Q0 hp (di ) − µi )2 + (ci − µi )2 ]/2σi2 .
(8)
On further simplification, the test above reduces to
H1
T (ci ) = α
bi ci ≷ γ,
(9)
H0
Q0 hp (di )
and
σi2
(Q0 hp (di )) +2Q0 hp (di )µi
+
2σi2
where the parameter α
bi is defined as α
bi =
2
the threshold γ is obtained as γ =
. The result below determines the detection perforln 1−β
β
mance of the individual cooperative nanomachines.
(i)
Lemma 1: The probabilities of detection (PD ) and false
(i)
alarm (PF A ) at the ith cooperative nanomachine corresponding to the test statistic T (ci ) in (9) are given as
′
γ − Q0 hp (di ) − µi
(i)
PD = Q
,
(10)
σ
′
i
γ − µi
(i)
PF A = Q
,
(11)
σi
where γ ′ is defined as γ ′ = γ/b
αi and the function Q(·) is the
tail probability of the standard normal random variable.
Proof: Given in the technical report [13, Appendix A].
The cooperative nanomachines employ different types of
molecules. Therefore, the joint concentration vector e
c =
[e
c1 , e
c2 , · · · , e
cN ]T ∈ RN ×1 at the destination2 corresponding
to the N cooperating nanomachines can be written as
e
e+n
e,
c=x
(12)
N ×1
e ∈ R
where the concatenated vector x
is defined as
e = [b
x
x1 Q1 hp (de1 ), x
b2 Q2 hp (de2 ), · · · , x
bN QN hp (deN )]T . Sime ∈ RN ×1 is given as
ilarly, the concatenated noise vector n
T
e = [e
n
η1 , ηe2 , · · · , ηeN ] . The sensing problem at the destination
nanomachine can be formulated as the binary hypothesis
testing problem
e
H0 : e
c=n
(13)
e+n
e,
H1 : e
c=x
e = [Q1 hp (de1 ), Q2 hp (de2 ), · · · , QN hp (deN )]T . Considwhere x
ering the logarithm of the LRT at the destination nanomachine,
the decision statistic corresponding to the N cooperative
nanomachine transmissions is obtained as
"N
#
Y p(e
ci |H1 )
p(e
c|H1 )
= ln
,
(14)
Λ(e
c) = ln
p(e
c|H0 )
p(e
ci |H0 )
i=1
where (14) follows from the fact that the peak concentrations
at the destination nanomachine are conditionally independent
i.e., corresponding
to the alternative hypothesis H1 the PDF
Q
p(e
c|H1 ) = N
ci |H1 ). The individual PDFs p(e
ci |H0 ) and
i=1 p(e
p(e
ci |H1 ) in (14) corresponding to the two hypotheses H0 and
H1 are Gaussian with
p(e
ci |H0 ) ∼ N (e
µi , σ
ei2 )
(15)
p(e
ci |H1 ) ∼ N (Qi hp (dei ) + µ
ei , σ
e2 ).
i
Result in Lemma 2 below derives the simplified optimal
Neyman-Pearson LRT at the destination nanomachine for the
cooperative nano-network for moderate to high concentration
scenarios with molecules (> 109 ).
Lemma 2: The optimal test statistic T (e
c) and the resulting
test at the destination nanomachine for the binary hypothesis
testing problem in (13), for a scenario in which a moderate to
high number of molecules is emitted, is obtained as
N
X
H1
αi e
ci ≷ γ ′′ ,
(16)
T (e
c) =
i=1
H0
PN
where γ ′′ = ln 1−β
+ i=1 θi . The quantities αi and θi
β
are defined as
h
i
(i)
(i)
(17)
αi = Qi hp (dei )/e
σi2 (PD −PF A ),
#
"
(Qi hp (dei ))2 +2Qi hp (dei )e
µi
(i)
(i)
(18)
(PD −PF A ),
θi =
2e
σi2
with the expressions for the individual probabilities of de(i)
(i)
tection PD and false alarm PF A for the ith cooperative
nanomachine as obtained in (10) and (11) respectively.
Proof: Given in the technical report [13, Appendix B].
2 The destination nanomachine responds independently to each of the
cooperative nanomachines as the various cooperating nanomachines employ
different types of molecules. This is similar to the Multi-Molecule-Type MultiHop Network (MMT-MH) considered in works such as [17] that also employs
multiple molecule types.
Employing the distributions of e
ci for the null and the alternative hypotheses obtained in (15), together with the expression
for the test statistic T (e
c) in (16), its distributions under the
various hypotheses can be determined as
X
XN
N
2 2
α σ
H0 : N
e
αi µ
ei ,
i=1 i i
i=1
(19)
X
XN
N
α2i σ
ei2 .
H1 : N
ei ,
αi Qi hp (dei ) + µ
i=1
i=1
The end-to-end probability of error for the cooperative nanonetwork follows as described in the result below.
Theorem 1: The end-to-end probability of error (Pe ) at
the destination nanomachine in the diffusion based cooperative molecular nano-network comprising of N cooperating
nanomachines,
is given
as
PN
′′
e
γ
−
α
(Q
h
(
d
)
+
µ
e
)
i p i
i
i=1 i
qP
Pe =β 1 − Q
N
2
2
ei
i=1 αi σ
P
N
γ ′′ −
αi µ
ei
,
(20)
+ (1 − β)Q qP i=1
N
2
2σ
e
α
i=1 i i
P
where γ ′′ = ln 1−β
+ N
i=1 θi . The parameters αi and θi
β
are defined in (17) and (18) respectively.
Proof: Given in the technical report [13, Appendix C].
IV. S IMULATION R ESULTS
For simulation purposes, a cooperative nanonetwork is considered with diffusion coefficient D = 10−6 cm2 /s. The MUI
at each receiving node is modeled as a Gaussian distributed
RV with mean µi = µ
ei = µI = 4 × 1016 molecules/cm3
that is equivalent to five interfering transmissions, each with
Q = 3 × 109 molecules at a distance of 30 µm. Similar
to [8], the coefficient of variation of the medium is set as
0.3 with σi = σ
ei = 0.3µI . Fig. 2(a) shows the probability
of error Pe versus the number of molecules Q employed at
each node. For simulation, the distances between the source
and cooperative nanomachines, cooperative and destination
nanomachines, and the source and destination nanomachines
are set as di = dsr = 10 µm, dei = drd = 20 µm, i ∈ {1, 2, 3},
and d = 25 µm respectively. It can be observed from Fig.
2(a) that the analytical values obtained using (20) coincide
with those obtained from simulations, thus validating the
derived analytical results. One can also observe that the endto-end performance of the system is significantly enhanced by
cooperation in comparison to the direct source-destination only
communication scenario. Moreover, the end-to-end probability
of error decreases progressively as the number of cooperative
nanomachines N increases.
Figs. 2(b) and 2(c) compare the error rate of the cooperative molecular system with that of the existing multi-input
single-output (MISO) and single-input multi-output (SIMO)
systems proposed in [8] for a fixed number of molecules
and a fixed distance respectively. The number of cooperating
nanomachines is set as N = 2, with two nano-transmission
nodes at the information source for the MISO system, and
two reception nodes at the information sink for the SIMO
system. Further, the distance between each transmitting and
0
0
10
−2
10
N =1
−3
10
d = 25µm
−4
10
−5
10
−6
N =3
N =2
10
(c)
−1
10
di = 2d/3, dei = d/3
−2
10
di = d/3, dei = 2d/3
−3
10
di = dei = d/2
−4
10
Proposed
MISO, 2 Tx
SIMO, 2 Rx, MRC
−5
10
−7
10
0.2
0.4
0.6
0.8
1
1.2
1.4
Number of Molecules (Q)
1.6
1.8
2
10
x 10
Probability of Error (Pe )
10
10
(b)
Direct Communication
Proposed (Analytical)
Proposed (Simulated)
Probability of Error (Pe )
Probability of Error (Pe )
0
10
(a)
−1
0.9
1
1.1
1.2
1.3
1.4
Distance (d)
1.5
1.6
di = 20µm,
dei = 10µm
−1
10
−2
10
di = 10µm,
dei = 20µm
−3
10
−4
10
−5
Proposed
MISO, 2 Tx
SIMO, 2 Rx, MRC
di = dei = 15µm
10
0.4
1.7
−3
x 10
0.6
0.8
1
1.2
1.4
Number of Molecules (Q)
1.6
1.8
2
10
x 10
Error rate of the diffusion based cooperative molecular communication system for various scenarios with (a) direct communication
(b) MISO and SIMO systems with fixed number of molecules (c) MISO and SIMO systems with fixed distance.
Fig. 2.
receiving node in the SIMO and MISO systems is considered
to be d and 30 µm in Figs. 2(b) and 2(c) respectively. For a fair
comparison, the total molecules Q are uniformly distributed
among the transmitting nodes, i.e., Q0 = Q1 = Q2 = Q
3 in
the cooperative system, with Q0 = Q for the SIMO system,
and Q0,1 = Q0,2 = Q
2 for the MISO system. One can
observe from Fig. 2(b) that the performance of the molecular
communication systems with Q = 1×109 molecules degrades
as the distance increases. Moreover, it can be seen in Figs.
2(b)-(c) that the cooperative molecular system outperforms
the MISO and the SIMO systems for the scenario when the
distance between the source and cooperative nanomachines is
less than the distance between the cooperative and destination
nanomachines. This is owing to the fact that when di < dei ,
the cooperative nanomachines are able to decode the source
symbol correctly with high probability. Therefore, the endto-end performance under such a scenario is dominated by
the communication distance between the cooperative and the
destination nanomachines. On the other hand, for the scenario
with di > dei , the resulting probability of error at the cooperative nanomachines is relatively higher, leading to erroneous transmissions to the destination nanomachine. Hence,
the performance of the cooperative system is dominated by
the source-cooperative nanomachine distances. In contrast to
the SIMO system with Q0 = Q molecules, the cooperative
scenario uses Q0 = Q
3 molecules at the source. Therefore,
the cooperative system has a higher probability of error in
comparison to the SIMO system. However, the cooperative
system outperforms the MISO system for this scenario as well.
Moreover, a significant improvement in the error performance
is attained for the scenario when the cooperative nanomachines
are at equal distances from both the source and destination
nanomachines. Due to space limitations, additional simulation
results for detection performance at the destination nanomachine are presented in the technical report [13].
V. C ONCLUSION
This work presented a performance analysis for diffusion
based cooperative molecular communication systems wherein
multiple nanomachines cooperate with the source to enhance
the end-to-end reliability in the presence of MUI. Analytical
expressions were derived for the optimal test statistics at
the cooperative, destination nanomachines, together with the
resulting probabilities of detection, false alarm as well as the
end-to-end error rate. Moreover, the proposed system was seen
to yield a performance improvement in comparison to the
existing SIMO and MISO systems.
R EFERENCES
[1] A. Singhal, R. K. Mallik, and B. Lall, “Performance analysis of amplitude modulation schemes for diffusion-based molecular communication,” IEEE Trans. Wireless Commun., vol. 14, no. 10, pp. 5681–5691,
2015.
[2] T. Nakano, Y. Okaie, and J.-Q. Liu, “Channel model and capacity
analysis of molecular communication with Brownian motion,” IEEE
Commun. Lett., vol. 16, no. 6, pp. 797–800, 2012.
[3] M. Pierobon and I. F. Akyildiz, “A physical end-to-end model for molecular communication in nanonetworks,” IEEE J. Sel. Areas Commun.,
vol. 28, no. 4, pp. 602–611, 2010.
[4] T. Nakano, T. Suda, Y. Okaie, M. J. Moore, and A. V. Vasilakos,
“Molecular communication among biological nanomachines: A layered
architecture and research issues,” IEEE Trans. Nanobiosci., vol. 13,
no. 3, pp. 169–197, 2014.
[5] I. Llatser, E. Alarcón, and M. Pierobony, “Diffusion-based channel characterization in molecular nanonetworks,” in proc. 2011 IEEE Conference
on Computer Communications Workshops, 2011, pp. 467–472.
[6] I. Llatser, A. Cabellos-Aparicio, and E. Alarcon, “Networking challenges
and principles in diffusion-based molecular communication,” IEEE
Wireless Commun., vol. 19, no. 5, pp. 36–41, 2012.
[7] I. Llatser, A. Cabellos-Aparicio, M. Pierobon, and E. Alarcón, “Detection techniques for diffusion-based molecular communication,” IEEE J.
Sel. Areas Commun., vol. 31, no. 12, pp. 726–734, 2013.
[8] L.-S. Meng, P.-C. Yeh, K.-C. Chen, and I. F. Akyildiz, “MIMO communications based on molecular diffusion,” in proc. 2012 IEEE Global
Communications Conference, 2012, pp. 5380–5385.
[9] X. Wang, M. D. Higgins, and M. S. Leeson, “Relay analysis in molecular
communications with time-dependent concentration,” IEEE Commun.
Lett., vol. 19, no. 11, pp. 1977–1980, 2015.
[10] N. Tavakkoli, P. Azmi, and N. Mokari, “Performance evaluation and optimal detection of relay-assisted diffusion-based molecular communication with drift,” IEEE Trans. Nanobiosci., vol. 16, no. 1, pp. 34–42,
2017.
[11] A. K. Sadek, W. Su, and K. R. Liu, “Multinode cooperative communications in wireless networks,” IEEE Trans. Signal Process., vol. 55,
no. 1, pp. 341–355, 2007.
[12] K. R. Liu, Cooperative communications and networking. Cambridge
University Press, 2009.
[13] N. Varshney, A. Patel, and A. K. Jagannatham, “Technical
Report: Diffusion based cooperative molecular communication
in nano-networks,” IIT Kanpur, Tech. Rep., 2017, [Online]
http://www.iitk.ac.in/mwn/documents/MWNLab_TR_DCoop_2017.pdf
[14] Y. Moritani, S. Hiyama, and T. Suda, “Molecular communication
among nanomachines using vesicles,” in Proc. NSTI nanotechnology
conference, 2006, pp. 705–708.
[15] L.-S. Meng, P.-C. Yeh, K.-C. Chen, and I. F. Akyildiz, “On receiver
design for diffusion-based molecular communication,” IEEE Trans.
Signal Process., vol. 62, no. 22, pp. 6032–6044, 2014.
[16] A. Papoulis and S. U. Pillai, Probability, random variables, and stochastic processes. Tata McGraw-Hill Education, 2002.
[17] A. Ahmadzadeh, A. Noel, and R. Schober, “Analysis and design of
multi-hop diffusion-based molecular communication networks,” IEEE
Trans. Mol. Biol. Multi-Scale Commun., vol. 1, no. 2, pp. 144–157,
2015.
| 7cs.IT
|
Improvised Comedy as a Turing Test
arXiv:1711.08819v2 [cs.AI] 2 Dec 2017
Kory Mathewson ∗
Department of Computing Science
University of Alberta
Edmonton, Alberta, Canada
[email protected]
Piotr Mirowski ∗
HumanMachine
London, UK
[email protected]
Abstract
The best improvisational theatre actors can make any scene partner, of any skill level
or ability, appear talented and proficient in the art form, and thus "make them shine".
To challenge this improvisational paradigm, we built an artificial intelligence (AI)
trained to perform live shows alongside human actors for human audiences. Over
the course of 30 performances to a combined audience of almost 3000 people, we
have refined theatrical games which involve combinations of human and (at times,
adversarial) AI actors. We have developed specific scene structures to include
audience participants in interesting ways. Finally, we developed a complete show
structure that submitted the audience to a Turing test and observed their suspension
of disbelief, which we believe is key for human/non-human theatre co-creation.
1
Background
Theatrical improvisation is a form of live theatre where artists perform "real-time dynamic problem
solving" through semi-structured spontaneous storytelling [1]. Improvised comedy involves both
performers and audience members in interactive formats. We present explorations in a theatrical
Turing Test as part of an improvised comedy show. We have developed an artificial intelligence-based
improvisational theatre actor—a chatbot with speech recognition and speech synthesis, with a physical
humanoid robot embodiment [2, 3] and performed alongside it in improv showsA at performing
arts festivals, including ImproFest UK and the Brighton, Camden, and Edinburgh Fringe Festivals
[4]. Over these first 30 shows, one or two humans performed improvised scenes with the AI. The
performers strove to endow the AI with human qualities of character/personality, relationship, status,
emotion, perspective, and intelligence, according to common rules of improvisation [5, 6]. Relying
on custom state-of-the-art neural network software for language understanding and text generation,
we were able to produce context-dependent replies for the AI actor.
The system we developed aims to maintain the illusion of intelligent dialogue. Improvised scenes
developed emotional connections between imaginary characters played by humans and AI improvisors.
The human-like characterization elicited attachment for the AI from audience members. Through
various configurations (e.g. human-human, human-AI, and AI-AI) and different AI embodiments (e.g.
voice alone, visual avatar, or robot), we challenged the audience to discriminate between human- and
AI-led improvisation. In one particular game setup, through a Wizard-of-Oz illusion, we performed a
Turing test inspired structure. We deceived the audience into believing that an AI was performing,
then we asked them to compare that performance with a performance by an actual AI. Feedback
from the audience, and from performers who have experimented with our system, provide insight for
future development of improv games. Below we present details on how we debuted this technology
to audiences, and provide strictl anecdotal observations collected over multiple performances.
∗
A
Both authors contributed equally.
Show listings and recordings are available at https://humanmachine.live
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, California, USA.
Workshop on Machine Learning for Creativity and Design.
2
Methods
We named our AI improviser A.L.Ex, the Artificial Language Experiment, an homage to Alex the
Parrot, trained to communicate using a vocabulary of 150 words [7]. The core of A.L.Ex consists of
a text-based chatbot implemented as a word-level sequence-to-sequence recurrent neural network
(4-layer LSTM encoder, similar decoder, and topic model inputs) with an output vocabulary of 50k
words. The network was trained on cleaned and filtered subtitles from about 100k filmsA . Dialogue
turn-taking, candidate sentence selection, and sentiment analysis [8] on the input sentences are based
on heuristics. The chatbot communicates with performers through out-of-the-box speech recognition
and text-to-speech software. The chatbot runs on a local web server for modularity and allows for
integration with physical embodiments (e.g. parallel control of a humanoid robotB . The server also
enables remote connection which can override the chatbot and give dialog control to a human operator.
Further technological implementation details are provided by Mathewson and Mirowski [4].
An improvisational scene starts by soliciting suggestion for context from the audience (e.g., “nongeographical location” or “advice a grandparent might give”). The human performer then says several
lines of dialogue to prime the AI with dense context. The scene continues through alternating lines of
dialog between the human improviser(s) and the AI. Often through human justification, performers
aim to maintain scene reality and ground narrative in believable storytelling. A typical scene lasts
between 3-6 minutes, and is interrupted by the human performer when it reaches a natural ending
(e.g. narrative conclusion or comical high point).
The first versions of the improvising artificial stage companions had their stage presence reduced to
projected video and amplified sound. We evolved to physical embodiments (i.e. the humanoid robot)
to project the attention of the performer(s) and audience on a material avatar. Our robotic performers
are distinctly non-human in size, shape, material, actuation and lighting. We chose humanoid robotics
because the more realistic an embodiment is the more comfortable humans often are with it; though
comfort sharply drops when creatures have human-like qualities but are distinctly non-human [9].
The performances at the Camden and Edinburgh Fringe festivals involved a Turing test inspired scene
conducted with the willing audience. We performed the scene by first deceiving the audience into
believing that an AI was performing (whereas the chatbot and the robot were controlled by a human);
then we performed a second scene with an actual AI. In game (1), we explained the Turing test first,
then performed the two scenes consecutively and finally asked the audience to discriminate, through
a vote, which scene was AI-led. In a different game (2), we performed the Wizard-of-Oz scene and
then immediately asked, in character and as part of the performance, if the audience suspected that a
human was in control of the chatbot.
3
Preliminary Observations and Conclusions
We summarize here anecdotal observations from our performance. In game (1), nearly everyone
identified the AI from the human. However, we noted that in game (2) approximately half the
audience members believed that an AI was performing flawlessly alongside human improvisor(s).
When not forewarned about the Turing test, the audience (of various ages and genders) was convinced
that the dialog system understood the details of the scene and responded immediately and contextually.
The propensity of this delusion is likely driven by several factors: the context within which they are
viewing the deception, the lack of personal awareness of the current state-of-the-art AI abilities, and
emotional connections with the scene. Post-show discussions with audience members confirmed
that when a performer tells the audience that an AI is controlling the robot’s dialogue, the audience
members will trust this information. Being at an improvisational show, they expect to suspend
disbelief and use their imagination. Most of them were also unaware of capabilities and limitations of
state-of-the-art AI systems, which highlights the responsibility of the AI community to communicate
progress in AI effectively and to effectively invite public understanding of AI ability. Finally, we
observed that the introduction of a humanoid robot, with a human-like voice, increased the audiences’
propensity to immerse themselves in the imaginative narrative presented to them.
A
B
Subtitles from 100k movies were collected from https://opensubtitles.org
The robot was manufactured by https://www.ez-robot.com
2
We plan to conduct an experimental study of the audience beliefs in shared AI and human
creativity[10]. We hope to better understand the way that audiences enjoy art when co-created
by humans and AIs, to create better tools and mediums for human expression.
References
[1] Brian Magerko, Waleed Manzoul, Mark Riedl, Allan Baumer, Daniel Fuller, Kurt Luther, and
Celia Pearce. An empirical study of cognition and theatrical improvisation. In Proceedings of
the seventh ACM conference on Creativity and cognition, pages 117–126. ACM, 2009.
[2] Ken Perlin and Athomas Goldberg. Improv: A system for scripting interactive actors in virtual
worlds. In Proceedings of the 23rd annual conference on Computer graphics and interactive
techniques, pages 205–216. ACM, 1996.
[3] Barbara Hayes-Roth and Robert Van Gent. Improvisational puppets, actors, and avatars. In
Computer Game Developers Conference, 1996.
[4] Kory Mathewson and Piotr Mirowski. Improvised theatre alongside artificial intelligences. In
AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, 2017.
[5] Keith Johnstone. Impro: Improvisation and the theatre. Routledge, 2012.
[6] Mick Napier. Improvise: Scene from the inside out. Heinemann Drama, 2004.
[7] Irene M Pepperberg and Irene M Pepperberg. The Alex studies: cognitive and communicative
abilities of grey parrots. Harvard University Press, 2009.
[8] Clayton J Hutto and Eric Gilbert. Vader: A parsimonious rule-based model for sentiment
analysis of social media text. In Eighth international AAAI conference on weblogs and social
media, 2014.
[9] Masahiro Mori. The uncanny valley. Energy, 7(4):33–35, 1970.
[10] Jürgen Schmidhuber. Formal theory of creativity, fun, and intrinsic motivation (1990–2010).
IEEE Transactions on Autonomous Mental Development, 2(3):230–247, 2010.
Supplementary Material
Acknowledgments
We would like to acknowledge the fantastic collaborators and supporters throughout the project, in
particular Adam Meggido, Colin Mochrie, Matt Schuurman, Paul Blinov, Lana Cuthbertson, Patrick
Pilarski, as well as Alessia Pannese, Stephen Davidson, Stuart Moses, Roisin Rae, John Agapiou,
Arfie Mansfield, Luba Elliott, Shama Rahman, Holly Bartolo, Benoist Brucker, Charles Sabourdin,
Katy Schutte and Steve Roe. Thank you to Rapid Fire Theatre for a space for new ideas to flourish.
Illustrations
Figure 1: System diagram of the Artificial Language Experiment (A.L.Ex).
3
Figure 2: Visual and physical embodiments of the AI improviser.
Figure 3: Two human performers and an audience volunteer improvising with a robot.
4
| 2cs.AI
|
Berezin quantization of noncommutative projective varieties
Andreas Andersson
arXiv:1506.01454v8 [math.OA] 4 Feb 2018
Email: [email protected]
Max Planck Institute for Mathematics in the Sciences. Inselstrasse 22, D-04103 Leipzig, Germany
Wollongong University, School of Mathematics and Applied Statistics, 2522 Wollongong, Australia
Mathematics Subject Classification 2010 Primary: 58B32; Secondary: 58B34, 46L89
Keywords: Berezin quantization, Toeplitz operators, Cuntz–Pimsner algebras, compact quantum groups, generalized
inductive limits, noncommutative geometry, subproduct systems, noncommutative random walks, Arveson’s conjecture,
balanced metrics.
February 6, 2018
Abstract
We use operator algebras and operator theory to obtain new result concerning Berezin quantization of compact Kähler manifolds. Our main tool is the notion of subproduct systems of finitedimensional Hilbert spaces, which enables all involved objects, such as the Toeplitz operators, to be
very conveniently expressed in terms of shift operators compressed to a subspace of full Fock space.
This subspace is not required to be contained in the symmetric Fock space, so from finite-dimensional
matrix algebras we can construct noncommutative manifolds with extra structure generalizing that
of a projective variety endowed with a positive Hermitian line bundle and a Kähler metric in the class
of the line bundle. Even in the commutative setting these constructions are very fruitful. Firstly, we
show that the algebra of smooth functions on any smooth projective variety can be quantized in a
strong sense of inductive limits, as was previously only accomplished for homogeneous manifolds. In
this way the Kähler manifold is recovered exactly from quantization and not just approximately. Secondly, we obtain a strict quantization also for singular varieties. Thirdly, we show that the Arveson
conjecture is true in full generality for shift operators compressed to the subspace of symmetric Fock
space associated with any homogeneous ideal. For noncommutative examples we consider homogeneous spaces for compact matrix quantum groups which generalize q-deformed projective spaces,
and we show that these can be obtained as the cores of Cuntz–Pimsner algebras constructed solely
from the representation theory of the quantum group. We also discuss interesting connections with
noncommutative random walks.
Contents
1
Introduction
1.1 Kähler geometry . . . . . . . . . .
1.2 Algebraic geometry . . . . . . . .
1.3 Operator algebra . . . . . . . . . .
1.4 Operator theory . . . . . . . . . .
1.5 Quantum homogeneous manifolds
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3
3
5
6
7
8
2
3
4
5
6
7
Subproduct systems
2.1 Basic properties . . . . .
2.2 Toeplitz algebras . . . . .
2.2.1 The Toeplitz core
2.2.2 The right shifts .
2.2.3 Normal ordering
2.3 Cuntz–Pimsner algebras .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
9
9
10
11
11
12
12
Review of quantization of projective varieties
3.1 Berezin quantization with prequantum condition
3.2 Projectively induced quantization . . . . . . . .
3.3 The circle bundle . . . . . . . . . . . . . . . . .
3.4 Singular varieties . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
13
13
15
17
18
Inductive limits
4.1 Relaxed definition of inductive limits . . . . .
4.2 Inductive limits from subproduct systems . . .
4.3 Cuntz–Pimsner algebras from inductive limits
4.4 Formulas for covariant symbols . . . . . . . . .
4.5 Commutative case and the Arveson conjecture
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
19
19
21
23
24
25
Projective limits
5.1 Relaxing the notion of projective limit . .
5.2 Changing the inner products . . . . . . .
5.3 The isometries . . . . . . . . . . . . . . .
5.4 Projective system for subproduct systems
5.5 The state on OH . . . . . . . . . . . . . .
5.6 Contravariant symbols . . . . . . . . . .
5.7 The asymptotic multiplication . . . . . .
5.8 The adjoint of the total Toeplitz map . .
(0)
5.9 OH as a projective limit . . . . . . . . .
5.10 Strict quantization . . . . . . . . . . . . .
5.11 OH assembled from projective limits . . .
5.12 Commutative case . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
26
27
27
28
30
32
33
35
35
37
38
39
39
Application to compact matrix quantum groups
6.1 Compact matrix quantum groups . . . . . . . . . . . . .
6.1.1 Representations and actions . . . . . . . . . . . .
6.1.2 Universal quantum groups . . . . . . . . . . . . .
6.1.3 The dual discrete quantum group . . . . . . . . .
6.2 First-row and first-column algebras . . . . . . . . . . . .
6.3 Subproduct systems of G-representations . . . . . . . . .
6.4 Berezin quantization of C(G/K) . . . . . . . . . . . . . .
6.4.1 Covariant symbols as first-row matrix coefficients
6.4.2 Intertwining the actions . . . . . . . . . . . . . .
6.4.3 C(G/K) as an inductive limit . . . . . . . . . . .
6.4.4 C(SG ) as an inductive limit . . . . . . . . . . . .
6.5 Comparison with Poisson and Martin boundaries . . . .
6.5.1 Poisson integral versus total Toeplitz map . . . .
6.5.2 Markov operator . . . . . . . . . . . . . . . . . .
6.5.3 Martin boundaries . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
40
40
40
42
42
44
45
46
46
46
47
49
49
49
50
51
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Concluding remarks
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
52
2
8
References
52
1 Introduction
With motivations from physics, Berezin introduced a way of approximating projective compact Kähler
manifolds M by finite-dimensional matrix algebras B(Hm ) parameterized by m ∈ N0 [Bere1, Bere2,
CGR1, Lan2, Schl1]. When M = G/K is a homogeneous space of some Lie group G, each Hilbert
space Hm carries an irreducible representation of G, and it is obtained by the Borel–Weil construction:
Hm is the space of holomorphic sections of a suitable line bundle L⊗m over M .
In order to apply Berezin quantization to quantum physics, the parameters (time or temperature
or energy etc.) should be chosen such that the limit m → ∞ simplifies the description of the system
at hand. This is a powerful method; for instance it gives the Hartree–Fock approximation as a special
case [Raj1]. The limit behavior is captured by a classical (compact Kähler) manifold M .
Nowadays it is however becoming more and more important to have a versatile theory of open
quantum systems. Recently we observed how to obtain a simplifying infinite-m limit in the standard
framework of quantum channels as driving the evolution of open quantum systems [An4, An5]. It so
happens though, that the infinite-m system is (in general) not given by a classical manifold but by
a noncommutative manifold, i.e. there is a noncommutative C ∗ -algebra C(M) which is supposed to
encode the properties of the dynamics and which is a surprisingly good analogue of the commutative
C ∗ -algebra C(M ) of continuous functions on a projective variety M . Here we use the symbol M
for a nonexisting object appearing only in the notation for the algebra C(M), while M denotes an
honest manifold. Such a C ∗ -algebra C(M) of continuous functions on a “noncommutative projective
variety” will be obtained as an inductive limit of matrix algebras B(Hm ). These in turn arise from
a sequence H• = (Hm )m∈N0 of finite-dimensional Hilbert spaces such that Hm+l ⊆ Hm ⊗ Hl for
all m, l ∈ M0 . Such a sequence has been referred to as a “subproduct system” [ShSo1] and it
generalizes the structure needed to perform ordinary Berezin quantization. In [An4] we discussed the
physical meaning of the special case when M = G/K is a “projective” quantum homogeneous space
(a generalization of compact coadjoint orbits).
The purpose of the present paper is to introduce and study the C ∗ -algebras C(M) defining these
noncommutative manifolds. They will be constructed in such a way that they possess a lot of extra
structure generalizing that complex-analytic structure, a positive line bundle and a Kähler metric.
The construction is new also in the commutative setting and we will solve some very interesting
problems in operator theory which provide new insight in the geometry of compact Kähler manifolds.
In this way we set the stage for a new interplay between operator theory and complex differential
geometry. We begin by outlining the results from several different viewpoints.
1.1 Kähler geometry
Let (M, L) be a polarized manifold, i.e. M is a compact connected complex-analytic manifold
admitting a Kähler metric and L is a holomorphic line bundle over M with the property that any
choice of basis for the finite-dimensional vector space H 0 (M ; L) of holomorphic sections of L gives
a holomorphic embedding of M into the projective space CPn−1 , where n := dim H 0 (M ; L). For
m ∈ N, let Lm be the mth tensor power of the line bundle L and let L0 = OM be the trivial
holomorphic line bundle. The idea of Berezin quantization is that if one chooses an inner product
on H 0 (M ; Lm ) for each m, so that we obtain a sequence H• = (Hm )m∈N0 of Hilbert spaces, then
the finite-dimensional matrix algebras B(Hm ) should give an increasingly good approximation of the
C ∗ -algebra C(M ) of continuous functions on M as m goes to infinity. This works for any polarized
manifold (M, L) [BMS] if the inner product on each Hm is given by
hφ|ψih,ω := md
ˆ
hm (φ(x), ψ(x))
M
3
ω(x)d
,
d!
∀ψ, φ ∈ H 0 (M, Lm )
for some fixed Kähler metric ω on M and a Hermitian metric h on L related to the Kähler metric
via the “prequantum condition”
√
¯ log h = ω.
(1.1)
−1∂∂
The approximation of C(M ) by the matrix algebras B(Hm ) is, more precisely, a “strict quantization”
of C(M ) realized by explicit surjective positive maps (“Toeplitz maps”) (see §3 for more details)
ς˘(m) : C(M ) → B(Hm )
and explicit injective positive maps (covariant symbol maps)
ς (m) : C(M ) → B(Hm ).
such that ς (m) is the adjoint of ς˘(m) with respect to the inner product on B(Hm ) and C(M ) defined
by the trace and by ω respectively, and such that the “Berezin transforms” ς (m) ◦ ς˘(m) converge to
the identity map on C(M ) as the “quantum number” m goes to infinity.
One could imagine however, that it should be possible to approximate C(M ) in some stronger
sense. It turns out that is not true for any polarized manifold (M, L), if we want the prequantum condition (1.1). We are now interested in the problem of relating properties of (M, L) to the
well-behavedness of the quantization H• . As has been well known every since the beginnings of quantization theory, and made very clear in [CGR1, CGR2], if (M, L) is a homogeneous manifold then
it can be quantized in a stronger sense than that provided by a strict quantization. Hawkins made
this very explicit in a C ∗ -algebraic setting [Hawk1] by realizing C(M ) as an inductive limit of the
matrix algebras B(Hm ). We shall show that these inductive limits are arising very intrinsically from
a choice of embedding M ֒→ CPn−1 and that the construction works for any projective variety M if
we drop the prequantum condition (1.1) and instead choose h and ω to be related in a different way.
In fact we need not refer to Hermitian metrics or Kähler metrics for the construction of the inductive
limit, and the covariant symbol ς (m) and the Toeplitz maps ς˘(m) arise naturally from the very maps
ιm,l : B(Hm ) → B(Hl ) which defines the inductive system. Nevertheless it is easy to recover the
metrics h and ω such that ς (m) and ς˘(m) attains the same geometric meaning as before.
Thus we drop the prequantum condition (1.1) and instead look at the Hilbert spaces Hm obtained
from a sequence (ω, hm ) where the metric on Lm varies with m ∈ N0 , or equivalently a sequence
(ωm , h) where the Kähler metric varies with m; the important datum is the sequence H• = (Hm )m∈N0
of Hilbert spaces. We call H• a “projectively induced quantization” of (M, L) (again see §3 for details)
if H• is a subproduct system in the sense of [ShSo1], i.e.
Hl ⊂ Hm ⊗ Hl−m ,
∀m ≤ l ∈ N.
This is the condition which allows the construction of an inductive system of unital completely
positive maps
ιm,l : B(Hm ) → B(Hl ),
∀m ≤ l ∈ N0
such that C(M ) is recovered as a C ∗ -algebra as the limit of the matrix algebras B(Hm ) as m goes to
infinity.
An inductive system which recovers C(M ) could possibly be constructed also from a prequantum
quantization, but the inductive system will have less nice properties, depending on the geometry of
M . It would be interesting to study further if one could characterize (“stability”) properties of M
as existence of an inductive system of matrix algebras with certain properties (see Conjecture 1.2
below).
What is special to the case when M is a coadjoint orbit is that then the maps ιm,l intertwine
the tracial states. In that case Hawkins showed that one obtains C(M ) as a projective limit is well
[Hawk2]. He also showed that a similar construction works for general polarized manifolds if one uses
a prequantum quantization [Hawk2]. Our contribution is the description of these constructions on a
single Hilbert space, namely the Hilbert-space direct sum
M
Hm ,
HN :=
m∈N0
4
which (as will be discussed in much more detail in future work) can be realized as a Hilbert space of
analytic functions on a subvariety of the unit ball in Cn (cf. [DRS1, DRS2]). The space HN sits as a
subspace of the symmetric Fock space H∨N over H1 , and the point is that H∨N can be identified with
the so-called Drury–Arveson space which has a very central role in multivariable operator theory.
Our ongoing work uses the results of the present paper to study the geometry of (M, L), in particular
the stability of vector bundles over M , using operator theory in Drury–Arveson space.
The subproduct property of the sequence H• says precisely that HN is invariant under the backward
shifts on H∨N . The inductive and projective limits mentioned above encode the data of a “strict
quantization”, as needed to generalize the classical setting, but they are much more convenient than
just knowing that there is a strict quantization because they are constructed using the shift operators
on HN . The notion of subproduct systems and the associated operator algebras of shift operators
provide a machinery for explicit calculations that has previously been available mainly in the case of
CPn−1 , where Berezin quantization and fuzzy geometry has been successfully described in terms of
creation and annihilation operators (the unnormalized shift operators) (see e.g. [BDLMC]).
1.2 Algebraic geometry
There is an analogue of Berezin quantization in projective algebraic geometry saying that the vector
spaces H 0 (M ; Lm ) of algebraic sections of higher tensor powers of a very ample line bundle L over
a projective complex variety M ⊂ CPn−1 determine all coherent sheaves on M . Namely, take the
Abelian category gr(A) of finitely generated graded modules over the graded algebra
M
H 0 (M ; Lm ).
A=
m∈N0
Elements of A are not functions on the variety M . In order to pass to objects defined on M we have
to work modulo the modules which are finite-dimensional as vector spaces over C, by replacing the
homomorphism spaces Homgr(A) (E, F ) by
Homqgr(A) (E, F ) := lim Homgr(A) (E≥m , F ),
m→∞
L
where E≥m := l≥m El is a tail of a graded A-module E = k∈Z Ek . It was shown in [Serre2] that in
this way we obtain an Abelian category qgr(A) which is equivalent to the Abelian category coh(M ) of
coherent sheaves on M . This result is also the starting point for noncommutative projective
L algebraic
geometry [ArZh1], namely one starts with a finitely generated graded algebra A =
m∈N0 Am
satisfying less stringent assumptions than commutativity and one studies the category qgr(A) guided
by the geometric intuition from analogy with commutative algebraic geometry.
Our observation in §4 is that if we endow the subspaces Am ⊂ A by inner products such that
the resulting Hilbert spaces Hm form a subproduct system then the above inductive system can be
explicitly described by unital completely positive maps and is compatible with the inner products in
a sense (namely we obtain a generalized inductive
limit of C ∗ -algebras as described below). Instead
L
of A one considers a N0 -graded ring R = m∈N0 Rm which can be concretely described as a (non∗-closed) algebra of operators on the Fock space HN . The limit
L
0
B∞ := lim Endgr(R) (R≥m )
m→∞
exists for the same reason as in the case of A. We show that 0 B∞ coincides with the normally ordered
part of the algebraic “Cuntz–Pimsner core” of the subproduct system (for the definitions see §2). In
the commutative case it is the algebra of real-algebraic C-valued functions on the variety M ⊂ CPn−1
associated with H• . Moreover, there is a natural norm on 0 B∞ such that its completion B∞ is (in
the commutative case) the C ∗ -algebra C(M ) of continous functions on M .
The quotient category qgr(R) is (as will be discussed in another publication) in the commutative
case equivalent to the category of torsionfree modules over the sheaf of real-algebraic functions on the
variety M . From there it is easy to go to modules over the sheaf of continuous functions. Thus we are
5
discussing a continuous version of Serre’s theorem and thereby a continuous analogue of Artin–Zhang
approach to noncommutative projective geometry.
The ring A is contained as a graded subring of R,
Am ⊂ Rm ,
∀m ∈ N0 ,
and one can define a “holomorphic structure” on (the continuous extensions of) objects in qgr(R)
as an operator on modules over B∞ possessing certain properties. Then we are lead to the following
research problem:
Problem 1.1 (Noncommutative GAGA). Find examples of noncommutative N0 -graded algebras A
such that qgr(A) is equivalent to the holomorphic subcategory of qgr(R).
The works [DRS1, DRS2] can be regarded as an operator-theoretic approach to noncommutative
projective geometry using shift operators on Fock space. What we are doing here suggest that this
gives the analytic version of Artin–Zhang’s algebraic approach.
1.3 Operator algebra
The C ∗ -algebra B∞ is a special case of a “generalized inductive limit” of C ∗ -algebras in the sense of
[BlKi1]. What is special with the inductive system coming from a subproduct system is not only that
it is “NF”, i.e. defined by completely positive maps, but also that the underlying algebraic inductive
limit of sets is already an algebra.
A tracial state ω on a C ∗ -algebra such as B∞ can have finite-dimensional approximation properties of various strengths, e.g. “quasidiagonality”, “uniform quasidiagonality”, “amenability” and
“uniform amenability”; see [Brow1, §3]. These notions involve tracial states φm : B(Hm ) → C on
finite-dimensional matrix algebras such that ω is in some sense the limit of the φm ’s as m → ∞,
in that there are completely positive maps ς˘(m) : B∞ → B(Hm ) intertwining ω and φm up to some
error. In this case we call (˘
ς (m) )m∈N0 an explicit realization of the approximation property of ω.
Consider the case of a commutative smooth projective variety M , where as mentioned we shall
show that the inductive limit B∞ coincides with C(M ). Berezin quantization can be regarded as
an explicit realization of the quasidiagonality of states on C(M ), namely the states associated with
volume forms of Kähler metric on M . The point is not to show that some traces are quasidiagonal,
but to obtain an explicit realization where the approximating finite-dimensional tracial states φm :
B(Hm ) → C are defined on matrix algebras whose dimension nm = χ(OM (m)) depend on the
complex-analytic structure of M.
With a prequantum quantization H• of (M, L) one will not obtain a subproduct system unless
the quantization is “regular” in the sense of [CGR2]. However, Blackadar–Kirchberg’s notion of
generalized inductive limit is very general and one can probably still obtain C(M ) as such an inductive
limit, albeit in a weaker sense than as the Cuntz–Pimsner algebra of a subproduct system. If one
could characterize precisely the properties of such an inductive limit needed for the existence of a
constant scalar curvature Kähler metric in the class c1 (L), one could hope to find the correct stability
condition on the manifold (M, L) which characterizes the existence of such a metric.
By analogue with the Donaldson–Tian–Yau conjecture about the equivalence of the exstence of a
constant scalar curvature Kähler metric and some stability property (“K-stability”) [Don9, Don12,
Tian1, Tian2] we formulate the following speculative but suggesting conjecture:
Conjecture 1.2. There exists an approximation property (say “QK-stability”) of traces on C ∗ algebras (perhaps quasidiagonality or amenability) such that the following holds: If ω is a Kähler
form on a smooth projective variety M ⊂ CPn−1 then ω has constant scalar curvature if and only if
there is a prequantum quantization H• of ω which explictly realizes ω as a QK-stable state on C(M ).
The inductive and projective limits constructed in the present papers may serve as guidance for
less well-behaved structures associated with quantizations with prequantum condition. Our ongoing
work is more focused on showing that such a characterization is possible for the analogous problem
of stability of vector bundles (where the analogue of the Donaldson–Tian–Yau conjecture is known
to be true [LuTe1]).
6
1.4 Operator theory
L
Starting from any subproduct system H• = (Hm )m∈N0 one can form its Fock space HN := m∈N0 Hm ,
L
which is a subspace of the full Fock space H⊗N := m∈N0 H⊗m which is invariant under the backward
shift operators. The shift operators S1 , . . . , SL
n on HN commute pairwisely if and only if HN is
contained in the symmetric Fock space H∨N = m∈N0 H∨m , where H∨m denotes the mth symmetric
tensor power of the Hilbert space H = H1 = Cn . The symmetric Fock space H∨N can be identified
with a Hilbert space of analytic functions on the unit ball Bn ⊂ Cn , called the Drury–Arveson
space and usually denoted by Hn2 , in such a way that the shifts on H∨N identify with the operators
M1 , . . . , Mn on Hn2 acting by multiplication by the coordinate functions z1 , . . . , zn on Bn .
Subspaces of Hn2 which are invariant under the adjoint shifts M1∗ , . . . , Mn∗ (briefly, “quotient
modules” of Hn2 ) have a simple description generalizing that of the Beurling representation of quotient
modules of the Hardy space H12 = H 2 (S1 ) of the unit circle [McTr1]. The most basic open question
about these quotient modules is whether the following is true.
Conjecture 1.3 (Arveson’s conjecture). Let HN be a graded quotient module of the Drury–Arveson
space H∨N . Then the shift operators S1 , . . . , Sn on HN form an essentially normal n-tuple, i.e.
[Sj∗ , Sk ] ∈ K,
∀j, k = 1, . . . , n.
In the present paper we shall prove Conjecture 1.3 by operator-algebraic methods. Moreover, the
geometric interpretations discussed here lead to an alternative more geometric proof which we will
present in a separate paper.
The conjecture has been proven previously in several special cases by different means (see e.g.
[Arv8, Arv9, Doug4, DoWa2, DoWa4, EnEs1, Esch1, Kenn1, KeSh1, KeSh2, Sha3] and references
therein), and the proofs have provided great new insights in the structure of submodules and quotient
modules of Drury–Arveson space, Hardy space and Bergman space. It is probably not possible to
have essential normality for any quotient module, as counterexamples have been found e.g. in the
case of the Hardy space of the polydisk [Doug4], and one has to restrict attention to some special
subclass of quotient modules. In relation to projective geometry one only needs graded quotient
modules, which is why Conjecture 1.3 is very important there.
Let TH be the C ∗ -algebra generated by S1 , . . . , Sn and let OH be the quotient TH /K by the ideal
of compact operators. Then Conjecture 1.3 can be reformulated by saying that the C ∗ -algebra OH
is commutative. The algebras TH and OH both have a natural Z-grading due to the grading of HN .
Our proof of the essential normality of the tuple (S1 , . . . , Sn ) is based on the observation that the
(0)
degree-0 part OH of OH can be constructed as an inductive system of sets. The inductive system
is given by explicit unital completely positive maps ιm,l : B(Hm ) → B(Hl ) for m ≤ l ∈ N0 . To show
that the limit is a unital C ∗ -algebra it is therefore enough to show that the ιm,l ’s are “asymptotically
multiplicative”. To see this we forget about the Hilbert-space structure and regard the ιm,l ’s as maps
between matrix rings. Then we observe that these matrix rings all sit inside a bigger N0 -graded ring
R and that the inductive system is of a certain type which is known to have a ring structure on the
set-theoretic limit.
The importance of the Arveson conjecture comes from the fact that, if the C ∗ -algebra TH generated
by S1 , . . . , Sn on HN is commutative modulo compacts, we have a short exact sequence of C ∗ -algebras
0 → K → TH → C(S) → 0,
where S is a subset of the unit sphere S2n−1 in Cn . Taking the U(1)-invariant part of this sequence
one obtains
(0)
0 → Γ0 → TH → C(M ) → 0,
and M is a compact Kähler manifold endowed with a positive line bundle L with associated circle
(0)
bundle equal to S. The algebra TH acts on HN which, just as the ambient space Hn2 , has the
property that every invariant subspace under S1 , . . . , Sn has a “Beurling decomposition” [McTr1].
Such submodules give rise to coherent sheaves on M and sometimes to vector bundles over M . We
expect that vector bundles arising like this have certain stability properties not possessed by an
arbitrary vector bundle over M .
7
1.5 Quantum homogeneous manifolds
Smooth projective varieties include in particular all coadjoint orbits G/K and, going noncommutative, we observe that every compact matrix quantum group G defines a “noncommutative manifold”
G/K with properties resembling very much those of a coadjoint orbit. Our main aim is then to show
that the C ∗ -algebra C(G/K) can be recovered from a suitably chosen subproduct system H• via the
noncommutative version of strict quantization sketched above.
In fact the definition of G/K is very simple. Let G be a compact matrix quantum group with
defining unitary representation u ∈ Mn (C) ⊗ C(G). We let zj := u1,j for j = 1, . . . , n denote the
elements of the first row of u. Then C(G/K) is defined as the C ∗ -algebra generated by elements
of the form zj1 · · · zjm zk∗m · · · zk∗1 for all multi-indices (j1 , . . . , jm ) and (k1 , . . . , km ) of equal length.
We also denote by C(SG ) the C ∗ -algebra generated by z1 , . . . , zn (and refer to it as the “first-row
algebra”).
Suppose that the compact matrix quantum group G is such that every element of C(G/K) can be
“normally ordered”, in the sense that all zj∗ ’s are to the right of the zk ’s (this is a crucial assumption).
Assume also that the Haar state is faithful on C(G/K). As recalled in §6.1, every compact quantum
group G has a dual discrete quantum group Ĝ, and both G and Ĝ “act” on the C ∗ -algebra C(G).
Hence they also act on the subalgebra C(SG ).
Theorem 1.4. For each m ∈ N0 , let Hm be the Hilbert space spanned by products zj1 · · · zjm with
inner product coming from the Haar measure on G (the sequence of Hm ’s is the “G-subproduct system”
H• ). Then there are explicit G-Ĝ-biequivariant injections (“Berezin covariant symbol maps”)
(m)
: B(Hm ) → C(G/K)
ςG
and explicit surjections (“Toeplitz quantization maps”)
(m)
: C(G/K) → B(Hm )
ς˘G
(m)
(m)
such that ςG ◦ ς˘G
isomorphism
converges point-norm to the identity on C(G/K) as m → ∞. This effects an
C(SG ) ∼
= OH
of the first-row algebra C(SG ) with the Cuntz–Pimsner algebra OH of the subproduct system H• , and
this isomorphism is equivariant for the natural ergodic actions of G and its discrete dual group Ĝ.
(0)
Denoting by OH the U(1)-invariant part of OH under the gauge action, we also have
(0)
C(G/K) ∼
= OH ,
(0)
(1.2)
(0)
and it is the C ∗ -algebra OH which will occupy most of the paper. We will first realize OH as a
“generalized inductive limit” in the sense of [BlKi1], [Hawk1] (as in the case of a general subproduct
system as described before), and then as a generalized projective limit in the spirit of [Hawk1]. Then
we do the same thing for C(G/K) to obtain the desired isomorphism (1.2).
The quantization of coadjoint orbits is studied in detail in [Hawk1] and [Rie2]. Berezin quantization for quantum homogeneous spaces of compact quantum groups with tracial Haar state was
discussed in [Sain].
The idea of looking at the G-subproduct system was partially motivated by Woronowicz’ reconstruction of a compact matrix quantum group G from its irreducible representations [Wor3]. Since
H• only contains a subset of all irreducible representations (in general), we recover not C(G) but
C(SG ).
For a classical manifold M , with quantization defined by a line bundle L over M , elements
(k)
(m)
of OH are continuous sections of the line bundle L⊗k . The subspace Hm ⊂ OH consists of the
holomorphic sections of the line bundle. For M = G/K a homogeneous space, the Hm ’s are irreducible
(k)
(0)
representations of G. In general, the subspaces OH for k ∈ Z \ {0} are Hilbert modules over OH ,
8
and for a quantum homogeneous space M = G/K each Hm is an irreducible representation of G. In
this sense, OH is a kind of “Borel–Weil algebra” for G (cf. [Seg1, Thm. 14.1]).
We shall also compare our results with “noncommutative random walks” on duals of compact
(0)
quantum groups [Iz1, INT1, Iz4]. The Toeplitz core TH plays the role of Martin compactification
(0)
of a walk restricted to the “dual” of G/K while OH is the boundary of the walk. Recalling that
the simplest Cuntz–Pimsner algebras C(S2n−1 ) of functions on spheres behave like boundaries of the
corresponding Toeplitz algebras, these observations are not too surprising.
This shows that in the presence of time-reversal symmetry, one obtains a noncommutative random walk, not on the dual of a compact quantum group, but on the dual of a compact quantum
homogeneous space. In this way one can use results from the theory of noncommutative random
walks to study physically relevant quantum walks, even though these two notions of “walks” are a
priori completely unrelated. Most significant is the possibility to describe the infinite-time limit in
a mathematically rigorous way, as provided by the Martin boundary of a noncommutative random
walk (or rather, since we really end up with random walks on homogeneous spaces, one has the
dequantization manifold replacing the Martin boundary as the infinite-time limit).
Acknowledgment. The author thanks Adam Rennie for great comments, support and inspiration.
Many thanks also to Orr Shalit and Guy Salomon for finding an important error in an earlier version
of this paper. We thank Orr Shalit also for other remarks of great value.
2 Subproduct systems
2.1 Basic properties
In this paper we write N0 := N ∪ {0} = {0, 1, 2, . . . }.
Definition 2.1. [[ShSo1, Def. 6.2]] A subproduct system is a sequence H• = (Hm )m∈N0 of finitedimensional Hilbert space Hm such that H0 = C and
Hm+l ⊆ Hm ⊗ Hl ,
∀m, l ∈ N0 .
(2.1)
We shall always denote H1 by H and, throughout this paper, n ∈ N will always be the dimension
of H,
H∼
= Cn .
Example 2.2. Given a finite-dimensional Hilbert space H, we set Hm := H⊗m for each m and refer to
it as the “product system” associated with H. Another example of a subproduct system is obtained
by taking Hm := H∨m to be the mth symmetric power of H; this is the “symmetric subproduct
system”.
Definition 2.3. A subproduct system is commutative if Hm ⊆ H∨m for all m ∈ N0 .
Lemma 2.4. [[ShSo1, Lemma 6.1]] Let H• be a subproduct system. Then the projections pm :
H⊗m → Hm satisfy
pl (pm ⊗ pl−m )pl = pl = pl (pl−m ⊗ pm )pl
(2.2)
whenver m ≤ l.
Proof. Replacing l by l − m for m ≤ l, the condition (2.1) reads
Hl ⊆ Hm ⊗ Hl−m ,
∀m ≤ l ∈ N0 .
Writing (2.3) in terms of projections,
pl ≤ pm ⊗ pl−m ,
the result is clear.
9
(2.3)
Definition 2.5. The Fock space associated with a subproduct system H• is the Hilbert space
M
HN :=
Hm .
m∈N0
P
L
We regard HN as a subspace of full Fock space H⊗N := m∈N0 H⊗m and denote by pN = m∈N0 pm
the projection from H⊗N onto HN . Thus pN is the identity in B(HN ), just as pm is the identity in
B(Hm ).
Example 2.6. If H• = H⊗• is the product system over a fixed Hilbert space H then HN is the full
(or “Boltzmannian”) Fock space H⊗N over H.
Example 2.7. If H• = H∨• is theLfull commutative subproduct system then HN is the symmetric
(or “Bosonic”) Fock space H∨N := m∈N0 H∨m over H.
Notation 2.8. We denote by Chzi = Chz1 , . . . , zn i the algebra of polynomials in n freely commuting
variables. We denote by C[z] = C[z1 , . . . , zn ] the algebra of polynomials in n commuting variables.
P
For any polynomial f (z1 , . . . , zn ) = j1 ,...,jn fj1 ,...,jn z1j1 · · · znjn in Chz1 , . . . , zn i, evaluation on the
basis e1 , . . . , en for H = H1 defines an element in Fock space,
X
fj1 ,...,jn ej11 ⊗ · · · ⊗ ejnn ∈ HN ,
f (e1 , . . . , en ) :=
j1 ,...,jn
and f is homogeneous iff
f (e1 , . . . , en ) ∈ H⊗m ,
for some m ∈ N0 .
Lemma 2.9. Every subproduct system H• = (Hm )m∈N0 defines a homogeneous ideal in Chzi, where
n = dim(H). Conversely, every homogeneous ideal in Chzi corresponds to a subproduct system
[ShSo1, Prop. 7.2]. For commutative subproduct systems the same is true with C[z] instead [DRS1,
§2.3].
Sketch of proof. Given H• , define a homogeneous ideal I in Chzi by
I := {f ∈ Chzi|f (e1 , . . . , en ) ∈ H⊗m ⊖ Hm for some m ∈ N0 }.
Conversely, given a homogeneous ideal I, we associate the Hilbert spaces
Hm := H⊗m ⊖ {f (e1 , . . . , en )|f ∈ I (m) },
where I (m) is the degree-m component of I.
2.2 Toeplitz algebras
Having fixed a subproduct system H• , we shall always denote by S1 , . . . , Sn the operators on Fock
space HN defined by
Sk φ := pm+1 (ek ⊗ φ),
∀φ ∈ Hm
for all m ∈ N0 , where pm+1 : H⊗(m+1) → Hm+1 is the orthogonal projection. They are the compressions to HN of the left shifts ψ → ek ⊗ ψ on full Fock space H⊗N . For more about compressed
n-tuples of shift operators, see [Pop1], [Pop2], [ShSo1], [DRS1], [DRS2].
Definition 2.10. The Toeplitz algebra of a subproduct system H• is the unital C ∗ -algebra TH of
operators on HN generated by the shifts S1 , . . . , Sn .
The backward shifts on H⊗N preserves the subspace HN , so the adjoints S1∗ , . . . , Sn∗ of S1 , . . . , Sn
are just the restrictions to HN of the backward shifts on H⊗N .
10
Definition 2.11. The vacuum state on the Toeplitz algebra TH is the restriction ε̂ : TH → C of
the vector state on B(HN ) defined by the unit vector Ω ∈ H0 = C. That is,
ε̂(X) := hΩ|XΩi,
∀X ∈ B(HN ).
If p0 denotes the unit in B(H0 ) then
Xp0 = ε̂(X)p0 = p0 X,
∀X ∈ B(HN ).
(2.4)
Notation 2.12. Let F+
n be the free unital semigroup generated by n elements 1, . . . , n (the empty
+
word ∅ is the identity in F+
n ). We write a word k ∈ Fn as k = k1 · · · km and refer to |k| := m as the
length of k. For the shifts S1 , . . . , Sn and the basis vectors e1 , . . . , en we then write
Sk∗ := (Sk )∗ = Sk∗m · · · Sk∗1 ,
Sk := Sk1 · · · Skm ,
ek := ek1 ⊗ · · · ⊗ ekm ,
and similarly for other n-tuples of elements defined below. Finally, jk := j1 · · · jl k1 · · · km for j, k ∈ F+
n
with |j| = l and |k| = m.
2.2.1
The Toeplitz core
L
Let N = m m pm be the number operator on HN . It generates a unitary group on HN which
implements an action γ• of the circle group U(1) on TH ,
γt (Sk ) := eit Sk ,
∀eit ∈ U(1), k ∈ {1, . . . , n},
(2.5)
referred to as the gauge action on TH .
The gauge action (2.5) splits TH into the C ∗ -direct sum of the subspaces
(k)
TH
:= {T ∈ TH |γt (T ) = eikt T for all eit ∈ U(1)},
k ∈ Z.
(0)
The fixed-point subalgebra TH (the Toeplitz core) will be of great importance to us. It is generated
by polynomials in the shifts Sj and Sk∗ which are “homogeneous of degree zero” in the sense that
each term contains equally many forward shifts Sj as backward shifts Sk∗ .
2.2.2
The right shifts
In addition to the “sinister” shift Sk by the basis vector ek ∈ H, we shall need the “rectus” shift
Rk ψ := pm+1 (ψ ⊗ ek ),
∀ψ ∈ Hm , m ∈ N0
(2.6)
acting on the same Fock space HN . Note that Rk commutes with each Sj , and that Rk∗ commutes
with each Sj∗ , but [Rk , Sj∗ ] 6= 0.
The following formulas will be used extensively.
Lemma 2.13. For all m, l ∈ N with m ≤ l we have
X
X
Rr Rr∗
Sr Sr∗ H = pl =
l
Hence
X
|r|=m
and, in particular (m = 1),
Hl
.
|r|=m
|r|=m
n
X
k=1
Sr Sr∗ =
X
pl
l≥m
Sk Sk∗ = 1 − |ΩihΩ|.
11
(2.7)
∗
Proof. For r, s ∈ F+
n with |r| = m = |s| we have Sr Ss |Hl = pl (|er ihes |⊗pl−m )pl = pl (|er ihes |⊗pl−m )pl ,
so
X
X
|er iher | ⊗ pl−m pl
Sr Sr∗ H = pl
l
|r|=m
|r|=m
= pl (pm ⊗ pl−m )pl = pl ,
and similarly for the right shifts.
From (2.7) we see that the vacuum projection p0 = |ΩihΩ| belongs to the Toeplitz algebra.
Considering multiplying p0 from both sides with different shift operators, it is a simple matter to
deduce the following.
Corollary 2.14. The Toeplitz algebra TH contains the C ∗ -algebra K of all compact operators on Fock
space HN as a norm-closed two-sided ideal.
2.2.3
Normal ordering
In §5.9 we will obtain the following important result about the Toeplitz algebra, which we state here
for emphasis.
Lemma 2.15 (Normal ordering). Let H• be a subproduct system. Let AH denote the norm-closed
(non-∗) algebra generated by the shifts S1 , . . . , Sn and the identity in B(HN ). Then
span(AH A∗H ) = TH .
In particular, span(AH A∗H ) is an algebra.
2.3 Cuntz–Pimsner algebras
Definition 2.16 ([Vis2, Cor. 3.2]). The Cuntz–Pimsner algebra of a subproduct system H• is
the quotient of the Toeplitz algebra TH by the ideal K of all compact operators on HN ,
OH := TH /K.
We denote by Z1 , . . . , Zn the generators of OH , i.e. the images of the shifts S1 , . . . , Sn in the
quotient. They satisfy the sphere relation
n
X
Zk Zk∗ = 1,
k=1
Pn
which suggests viewing OH as the “boundary” of TH ; in the latter holds k=1 Sk Sk∗ ≤ 1, as we saw
in (2.7).
The formula (2.5), but with Zk replacing Sk , defines the gauge action on OH , which gives a
splitting
M (k) k·k
OH
OH =
k∈Z
of OH into spectral subspaces for this U(1)-action.
Remark 2.17 (Known examples). The most straightforward example of a subproduct Cuntz–
Pimsner algebra is the Cuntz algebra On , obtained from H• = H⊗• . As a commutative example, OH
for the symmetric subproduct system H∨• was shown in [Arv6c] to be isomorphic to the C ∗ -algebra
C(S2n−1 ) of continuous functions on the unit sphere S2n−1 ⊂ Cn . Cuntz–Pimsner algebras coming
from monomial ideals were described in [KaSh1].
We can reformulate Conjecture 1.3 as a statement about Cuntz–Pimsner algebras:
12
Conjecture 2.18 (Arveson’s conjecture). For every commutative subproduct system H• ⊂ H•∨ , the
Cuntz–Pimsner algebra OH is commutative.
(k)
For any subproduct system H• , the spectral subspaces OH for the gauge action on OH are Hilbert
(0)
∗
C -bimodules over the fixed-point subalgebra OH , with left and right inner products
hξ|ηiright := ξ ∗ η,
hξ|ηileft := ξη ∗
(2.8)
(k)
for ξ, η ∈ OH .
3 Review of quantization of projective varieties
Let us formulate Berezin quantization of complex submanifolds of projective n-space CPn−1 in terms
of subproduct systems, just to make it clear how the results of the subsequent sections relate to the
classical ones.
3.1 Berezin quantization with prequantum condition
Recall that Chow’s theorem says that a submanifold of projective space P[Cn ] = CPn−1 is a nonsingular (i.e. smooth) projective variety, i.e. the zero-set of some finitely generated homogeneous ideal
in C[z1 , . . . , zn ]. These manifolds can be characterized without even referring to CPn−1 (see Lemma
3.3 below), but for this we need to recall some complex geometry. A Kähler manifold is a pair
(M, ω) consisting of a complex manifold M and a closed nondegenerate 2-form ω on M which equals
the imaginary part of a Hermitian metric on M .
Remark 3.1 (Poisson bracket). The Kähler form ω is in particular a symplectic (i.e. closed and
nondegenerate) form, making M a symplectic manifold. The nondegeneracy of ω allows us to use
the inverse ω −1 to define a Poisson bracket on C ∞ (M ) by
{f, g} := ω j,k
∂f ∂g
,
∂xj ∂xk
(3.1)
if we denote by ω j,k the coefficients of ω −1 in local Darboux coordinates xj .
Recall that for a holomorphic line bundle L with a fixed choice of Hermitian metric h, there is
a unique connection, the “Chern connection”, which is compatible with both the metric and the
holomorphic structure in a suitable sense [Huy, Prop. 4.2.14]. If we locally represent h by a matrix¯ log h.
valued function, the curvature of this connection is given by ∂∂
Definition 3.2 ([BeSl1, §2.1]). A compact Kähler manifold (M, ω) is quantizable if there is a
¯ log h of the Chern
holomorphic Hermitian line bundle (L, h) over M such that the curvature ∂∂
connection satisfies the prequantum condition
√
¯ log h = ω.
(3.2)
−1∂∂
Then (L, h) gives a quantization of the Kähler manifold (M, ω).
Condition (3.2) is there to ensure ensures the following.
Lemma 3.3 ([Schl2], [BeSl1, §2.1]). Every quantization (L, h) of a compact Kähler manifold (M, ω)
gives an embedding of M as a submanifold of CPn−1 for some n ∈ N, and hence M can be regarded
as a projective algebraic variety (by Chow’s theorem). Conversely, every smooth projective algebraic
variety is a quantizable compact Kähler manifold.
Example 3.4 (The hyperplane bundle). The “tautological line bundle” over CPn−1 is the holomorphic line bundle, usually denoted by O(−1), whose fiber over a point in CPn−1 is the corresponding
line in Cn . The dual O(1) := O(−1)∗ of this line bundle is the hyperplane line bundle over CPn−1 .
The triple (CPn−1 , O(1), ωFS ), where ωFS is the Fubini–Study form and O(1) is equipped with the
Fubini–Study metric, is the prototypical example of a quantizable Kähler manifold (M, L, ω) [Schl1].
13
The embedding into CPn−1 mentioned in Lemma 3.3 requires a sufficiently positive line bundle,
and (3.2) says only that L is positive (“ample” ). It may therefore be necessary to use some tensor
power L⊗m of the line bundle L. However, by replacing L by L⊗m and rescaling the Kähler form ω
to mω we can, and shall, assume that L is itself sufficiently positive (“very ample”).
The important requirement in Lemma 3.3 is that M admits a Kähler metric, but we do not
need to choose one in order to an embedding of M into CPn−1 as a complex-analytic submanifold
(similarly we do not have the choose a Hermitian metric on L). Also, if we choose a Kähler metric ω
on M and a Hermitian metric h on L, so that we can embed M as a Kähler submanifold of CPn−1 ,
there is no need to require the prequantum relation (3.2) between h and ω; it is just the existence of
metrics satisfying the prequantum condition which is needed, to ensure that there is an ample line
bundle on M . Therefore, we will often speak of a polarized manifold, i.e. a pair (M, L) where M
is a compact Kähler manifold (with no choice of Kähler metric) and L is a positive line bundle on M
(which we shall assume very ample for convenience, with no choice of Hermitian metric specified).
Let (M, ω) be a compact Kähler manifold and let (L, h) be a Hermitian line bundle over M .
The space H 0 (M ; L) of global holomorphic sections of L is finite-dimensional and hence made into
a Hilbert space after fixing any inner product on H 0 (M ; L). In Berezin quantization one looks at
the limit of large m for the spaces H 0 (M ; L⊗m ) of sections of the tensor powers of L, and therefore
the inner products on these spaces should be comparable in some way. A consequence of the fact
that the Kähler form ω is symplectic is that ω d /d! (where d := dimC M ) is a volume form on M (the
“Liouville form”) and we can take the inner product
hφ|ψih,ω := md
ˆ
hm (φ(x), ψ(x))
M
ω(x)d
,
d!
∀ψ, φ ∈ H 0 (M, Lm ),
(3.3)
where hm is the Hermitian metric on L⊗m induced by h. Note that (3.3) is determined for all m by
the choice of inner product on H 0 (M ; L). If (L, h) is a quantization of (M, ω) then it is reasonable
to leave out either h or ω from the notation in h·|·ih,ω .
Example 3.5 (Sections of the hyperplane bundle). Recall the hyperplane line bundle O(1) over
P[H∗ ] introduced in Example 3.4. Fixing a basis for H∗ ∼
= Cn , the global holomorphic sections of
the mth tensor power O(m) of O(1) are identified with the degree-m homogeneous polynomials on
Cn . In particular, the space of holomorphic sections of O(1) is just the space H of continuous linear
functionals on H∗ . If we define Hm to be the Hilbert space of holomorphic sections of O(m) with
inner product (3.3) then for L = O(1) we simply have Hm = H∨m (symmetrized tensor product).
Thus, the Hm ’s form the full symmetric subproduct system (see Example 2.2).
Let (L, h) be a quantization of (M, ω) and endow H 0 (M ; Lm ) with the inner product (3.3). The
Hilbert space L2 (M ; L⊗m ) of all square-integrable sections of the line bundle L⊗m is much larger
than the holomorphic subspace H 0 (M ; Lm ). If Πm denotes the projection from L2 (M ; L⊗m ) to
H 0 (M ; Lm ) then every function f ∈ C(M ) defines an operator ς˘(m) (f ) on H 0 (M ; Lm ) by
ς˘(m) (f )φ := Πm (f φ),
∀φ ∈ H 0 (M ; Lm ),
(3.4)
i.e. acting as multiplication by f (recall that a section of a line bundle can be multiplied by continuous
functions to yield a new section) followed by projection back to H 0 (M ; Lm ) (the latter step is needed
since f is not holomorphic unless it is constant). In the case (L, h) is a quantization of (M, ω), we
have the following.
Proposition 3.6 ([BMS, Thms. 4.1,4.2, §5]). Let (L, h) be a quantization of a compact Kähler
manifold (M, ω) and define a structure of Hilbert space on H 0 (M ; Lm ) by (3.3). Then the collection
of endomorphism algebras EndC H 0 (M ; Lm ) and maps C ∞ (M ) ∋ f → ς˘(m) (f ) ∈ EndC H 0 (M ; Lm )
defined by (3.4) gives a strict quantization of the algebra C ∞ (M ) in the sense of [Lan1, Def. 1.1.1],
i.e. for all f, g ∈ C ∞ (M ) we have
(i) limm→∞ k˘
ς (m) (f )k = kf k (Rieffel condition),
(ii) limm→∞ k˘
ς (m) (f g) − ς˘(m) (f )˘
ς (m) (g)k = 0 (von Neumann condition),
14
(iii) limm→∞ km−1 [˘
ς (m) (f ), ς˘(m) (g)] − {f, g}k = 0 (Dirac condition),
and every operator in EndC H 0 (M ; Lm ) is of the form ς˘(m) (f ) for some f ∈ C ∞ (M ) [BMS, Prop.
4.2]. Here {·, ·} is the Poisson bracket (3.1).
For any volume form ω d /d! of (M, L), we have the associated Lebesgue space L2 (M, ω). The
algebra C(M ) identifies with a subspace of L2 (M, ω). Let ς (m) : B(Hm ) → C(M ) denote the adjoint
of ς˘(m) : C(M ) → B(Hm ) with respect to the L2 -inner product on L2 (M, ω) and the normalized
Hilbert-Schmidt inner product on B(Hm ).
For A ∈ B(Hm ), the function ς (m) (A) is also called the Berezin covariant symbol of A, and if
A = ς˘(m) (f ) then f is a (non-unique) contravariant symbol of A. The map
ς (m) ◦ ς˘(m) : C(M ) → C(M )
is the Berezin transform at level m. Similar to the famous expansion of the integral kernel for the
Bergman projections Πm [Zeld1], the Berezin transform ς (m) ◦ ς˘(m) has an asymptotic expansion at
large m [KaSc1].
Since the Toeplitz maps ς˘(m) are surjective by Proposition 3.6, their adjoints ς (m) are injective.
Hence we may regard the B(Hm )’s as embedded in C ∞ (M ) as vector subspaces. One may then ask if
it is possible to use the covariant symbols to approximate the whole C ∗ -algebraic structure of C(M )
by that of finite-dimensional matrix algebras B(Hm ).
We shall see that by changing the definition of the Toeplitz maps ς˘(m) we can in fact obtain C(M )
as an inductive limit for any polarized manifold (M, L).
3.2 Projectively induced quantization
The vector spaces H 0 (M ; Lm ) equipped with the inner products (3.3) do not always form a subproduct system. For that one has to choose ω and h appropriately, and for most polarized manifolds
(M, L) one cannot choose them to satisfy the prequantum condition (3.2) at the same time.
If we do not require that h and ω are related as in (3.2) then any two of (i) an inner product on
H 0 (M ; L), (ii) a Hermitian metric on L with positive curvature and (iii) a volume form ω d /d! on M
determines the third via (3.3). By the Calabi–Yau theorem [Yau1], any volume form on M can be
obtained as ω d /d! for some ω in the cohomology class c1 (L) (for any choice of polarization L, after
normalization of the volume form). If we do not require h to have positive curvature then there are
infinitely many Hermitian metrics h giving rise to the same inner product via the same volume form
(see [LMS1, Eq. (3.25)]).
For us, the choice of inner product h·|·i on H 0 (M ; L) will be the important input, and it will not
matter which Hermitian metric on L and volume form on M was used to define it.
Given a polarized manifold (M, L), a choice of basis for the n-dimensional vector space H 0 (M ; L)
allows use to embed M into the projectivization P[H 0 (M ; L)∗ ] of the vector space dual to H 0 (M ; L).
The elements of the basis for H 0 (M ; L) become the restrictions of the homogeneous coordinate functions z1 , . . . , zn on P[H 0 (M ; L)∗ ] to the embedded M . Choosing an inner product h·|·i on H 0 (M ; L)
we obtain an n-dimensional Hilbert space H which after a choice of orthonormal basis identifies with
Cn , and so M embeds into P[H∗ ] = CPn−1 . Whatever inner product on H 0 (M ; L) we used to define
the Hilbert space H, it will produce the symmetric subproduct system H∨• of holomorphic sections of
the hyperplane bundle on P[H∗ ] as in Example 3.5. What Lemma 3.3 says is that the ideal determined
by the algebraic relations among the zj ’s, appearing when we restrict them to the submanifold M ,
is homogeneous. The subspaces H 0 (M ; Lm ) ⊂ H 0 (M ; L)∨m of holomorphic sections of the tensor
powers of L endowed with the inner product as a subspace of H∨m will be denoted by
Hm = (H 0 (M ; L), h·|·i).
Here h·|·i is thus the inner product (3.3) in the special case when ω and h are the restrictions to M of
the Fubini–Study metrics on P[H∗ ], depending only on the inner product on H 0 (M ; L) which defines
the one-particle Hilbert space H. We set H0 := C.
15
We therefore have a description of polarized manifolds (M, L) with the extra datum of an inner
product on H 0 (M ; L) as a collection of Hilbert spaces Hm satisfying (2.1) with Hm ⊆ H∨m , where ∨
is the symmetrized tensor product (recall Lemma 2.9). The subproduct system H• is obtained from
H∨• by quotiening out by the ideal in C[z1 , . . . , zn ] which defines the embedded M .
Corollary 3.7. Every commutative subproduct system H• = (Hm )m∈N0 (see Definition 2.3) determines (via its associated homogeneous ideal in C[z1 , . . . , zn ]) a polarized manifold (M, L) with a fixed
structure of Hilbert space on H 0 (M ; L). Conversely, every such datum (M, L, h·|·i) determines a
commutative subproduct system.
We stress that for obtaining the subproduct system H• , the inner product on H 0 (M ; L) is arbitrary;√we do not require h and ω in (3.3) to satisfy the pre-quantum condition. Even if we did require
¯ log h, the inner product on Hm ⊂ H∨m for m ≥ 2 would in general differ from the inner
ω = −1∂∂
product h·|·ih defined by the initial ω and h.
Lemma 3.8. Let (M, L) be a polarized manifold with d := dimC M . Then for any inner product
h·|·i on H 0 (M ; L), there exists a unique volume form on M , which we can express as ω d for a
Kähler metric in the class c1 (L), such that h·|·i is ω d -balanced in the sense of [Don3, §2.2], i.e. if
Z1 , . . . , Zn are the homogeneous coordinates
on M ⊂ P[H∗ ] associated with any orthonormal basis for
Pn
0
H = (H (M ; L), h·|·i), normalized to k=1 Zk Zk∗ = 1 ∈ C ∞ (M ), then
1
vol(M, L)
where vol(M, L) :=
´
M
ˆ
M
Zj∗ Zk
δj,k
ωd
=
,
d!
n
∀j, k = 1, . . . , n,
ω d /d!.
Proof. The statement follows from the Calabi–Yau theorem [Yau1] and the fact that every polarized
manifold (M, L) admits a unique ω d -balanced metric for every volume form ω d on M [BLY1], [Don3,
§2.2] (equivalently, every line bundle is stable and from this it follows that every very ample line
bundle is “balanced as a line bundle” in the sense of [Wa1]).
To say that (M, L) is balanced as a polarized manifold [Don1] means precisely that there exists a
Hilbert space structure H on H 0 (M ; L) such that the volume form ω d in Lemma 3.8 is the restriction
to M of the Fubini–Study volume form on P[H∗ ]. Not every polarized manifold is balanced, and so in
general one cannot form a subproduct system from a quantization (L, h) of a compact Kähler manifold
(M, ω) in the sense of the last section. In fact that would require (M, Lm ) to be balanced for each
m ∈ N, in which case one says that the quantization is regular [CGR2]. The only polarized manifolds
known to admit a regular quantization are coadjoint orbits; cf. [ArLo1]. Since the subproduct
condition clearly will lead to a stronger kind of quantization (as will be shown in this paper), to be
able to use it for any polarized manifold we therefore consider a new kind of quantization.
We shall see that the normalized traces on the B(Hm )’s converge to a faithful state ω on C(M ),
which we call the limit state of the subproduct system H• .
Definition 3.9. A projectively induced quantization of a polarized manifold (M, L) is the datum
of a subproduct system H• ⊆ H∨• associated with some choice of inner product h·|·i on H 0 (M ; L)
together with the covariant and contravariant symbol maps specified by H• . That is, we define
ς (m) : B(Hm ) → C(M ) as the Berezin covariant symbol map (see e.g. [?]) and we take the Toeplitz
map ς˘(m) : C(M ) → B(Hm ) to be the adjoint of ς (m) with respect to the limit state ω and the
normalized trace on B(Hm ).
Remark 3.10. The terminology in Definition 3.9 is slightly nonstandard unless (M, L) is balanced;
indeed, (M, L) is a balanced polarized manifold (in the sense of [Don1]) if and only if there exists a
¯ log h) is projectively induced for
Hermitian metric h on L such that the quantization (L, h) of (M, ∂∂
0
some choice of inner product on H (M ; L). In that sense the term “projectively induced” appeared in
[CGR1] (it says precisely that the epsilon function discussed there is constant at each level, i.e. that
we have a “regular” quantization in the sense of [CGR2]). In the less restricted sense of Definition 3.9,
which works for any polarized manifold (M, L) since we do not require the prequantization condition,
16
the families covariant symbols maps associated with projectively induced quantizations were referred
to as “Berezin–Bergman quantizations” in [LMS1, §5] (this is the only work we know of where it has
been discussed for not necessarily balanced manifolds). In order to choose Toeplitz operators one
needs
An explicit formula for the covariant symbol map ς (m) is easy to write down; see Theorem 4.15.
In order to define Toeplitz operators one also needs to choose a state ω : C(M ) → C. The choice of
state ω is very important if one wants the Toeplitz maps ς˘(m) to give a strict quantization. We will
show that there is a canonical choice of ω, appearing as the limit of the normalized traces on the
B(Hm )’s, which produces a strict quantization of (M, L) with covariant symbols determned by the
subproduct system H• . We will show that a projectively induced quantization gives
(i) C(M ) as the “generalized inductive limit” of the C ∗ -algebras B(Hm ), and
(ii) a strict quantization of (M, L).
In this paper we will mainly discuss (ii) in the case of (quantum) homogeneous manifolds, but due
to our results concerning point (i) in the next sections we can fill in the details to obtain (ii) for
any polarized manifold. The latter will be very important in future work where more details will be
given.
Both (i) and (ii) could be satisfied without the subproduct condition. However, the inductive
system in (i) has very nice properties in the subproduct case which strictly relies on this assumption.
Concerning (ii) it is not necessarily the case that the subproduct condition gives something extra.
Rather, the important fact is that one can have a strict quantization at the same time as an inductive
system.
3.3 The circle bundle
Recall the Toeplitz quantization maps ς˘(m) : C(M ) → B(Hm ) (for definiteness and later relevance we
will focus on the case of a projectively induced quantization). We can assemble them into a single
map
Y
B(Hm ),
(3.5)
ς˘ : C(M ) →
m∈N0
∗
where the C -algebra on the right-hand side is the C ∗ -direct product.
Let L∗ be the dual line bundle of L. Under the embedding of M into CPn−1 , when L becomes the
restriction of the hyperplane line bundle, L∗ becomes the restriction of the tautological line bundle.
Denote by
SM := {ζ ∈ L∗ | kζk = 1}
(3.6)
the total space of the associated principal U(1)-bundle. The U(1)-action on SM induces a Z-grading
C(SM ) =
M
k·k
C(SM )(k)
k∈Z
of the C ∗ -algebra of continuous functions on SM , with C(SM )(0) = C(M ) and more generally
C(SM )(k) = Γ(M ; Lk )
as vector spaces, where Γ(M ; Lm ) is the space of all global continuous sections of Lk . The strongly
Z-graded algebraic structure on C(SM ) comes from the C(M )-module structure on each Γ(M, Lm )
and the tensor operation
Γ(M ; Lj ) ⊗C(SM ) Γ(M ; Lk ) = Γ(M ; Lj+k ).
The ∗-operation on C(SM ) is obtained by endowing each module Γ(M, Lk ) with the structure of
a Hilbert module given by hk , where h is the Fubini–Study Hermitian metric on L. That is, the
17
C ∗ -algebra C(SM ) is generated by the homogeneous coordinate functions Z1 , . . . , Zn on M ⊂ CPn−1
satisfying
n
n
X
X
Zk Zk∗ = 1 =
Zk∗ Zk .
k=1
k=1
These “sphere” relations are a manifestation of the fact that SM is nothing but the preimage of M
under the map S2n−1 → CPn−1 which defines projective space. Thus we can make an identification
SM ⊂ S2n−1 .
Now let ω : C(M ) → C be any faithful state. It extends canonically to a faithful state on C(SM )
by setting
ω(Zj Zk∗ ) := 0,
∀j, k ∈ F+
n , |j| 6= |k|.
L
k·k
As just mentioned, we have C(SM ) = k∈Z Γ(M ; Lk ) as a C ∗ -algebra if we endow each Γ(M, Lk )
with the metric hk . The GNS space of ω is the L2 -space L2 (S, ω) of ω-square-integrable functions
on S, and the closed subspace H 2 (S, ω) spanned by the coordinate functions Z1 , . . . , Zn on S is a
Hardy-type space.
2
Fock space HN are completions of the homogeneous coordinate ring A =
L Both H (S, ω) and the
n−1
A
of
M
⊂
CP
in such a way that elements of Am are orthogonal to elements of Al
m
m∈N0
whenever m 6= l. If M = G/K is a coadjoint orbit and H• is a regular quantization of G/K then one
simply has (see the proof of Lemma 6.30)
hψ|ϕiL2 =
1
hψ|ϕi.
dim Hm
So in this case Fock space can be embedded into H 2 (SM ),
h·|·i
M
Hm
⊂ H 2 (SM ) ⊂ L2 (SM ).
HN =
m∈N0
In general the relation between HN and H 2 (SM , ω) is more involved. Hence H 2 (SM , ω) is far from
sitting inside H 2 (S2n−1 , ωFS ) as a coinvariant subspace. For this reason, the Fock space HN is better
suited for studying how well the geometric properties of M are compatible with an embedding
M ֒→ CPn−1 and the associated pullback quantities.
Suppose now that ω is the state on C(SM ) associated with a Kähler form on M (denoted by the
same symbol ω), and let Π : L2 (SM , ω) → HN be the orthogonal projection. Then, with ς˘ as in (3.5),
we have
ς˘(f )ψ = Π(f ψ),
∀ψ ∈ HN
if we identify f ∈ C(M ) with the multiplication operator it defines on L2 (SM , ω). Then f is just the
contravariant symbol of the operator ς˘(f ) in the general sense of [Bere3]. The interior of SM (the
disk bundle) is a bounded symmetric domain and Berezin quantization on spaces such as SM has
been studied even more than in the setting of compact Kähler manifolds, see e.g. [UnUp1], [Bere2].
The circle bundle comes with the structure of a Cauchy–Riemann manifold, the details of which
can be found in [Schl1, §2], [Zeld1, §2].
We can define the Berezin transform and the covariant symbol of an operator on HN , just as we
did on the components Hm , using the fact that HN is a reproducing kernel Hilbert space. Again the
covariant symbol map ς is the adjoint of ς˘. In the noncommutative setting we will just calculate the
adjoint of ς˘ and take that as the definition of the covariant symbol map.
3.4 Singular varieties
We have seen that Berezin quantization of quantizable Kähler manifolds is really the quantization
of smooth projective varieties. It was suggested in [Schl2] that it may be possible to quantize also
singular (non-smooth) projective varieties in the same fashion. We shall see that this is in fact so:
it will be covered by the constructions in the next two sections, as the case when the subproduct
system H• is commutative (Corollary 5.35).
18
4 Inductive limits
Recall [ArMa1, Def. 1.1.11] that a sequence (Bm )m∈N0 of ∗-algebras Bm forms an “inductive system”
if there are ∗-homomorphisms ιm,l : Bm
S → Bm+l for each m ≤ l, and that the algebraic inductive
limit of such a sequence is the algebra m∈N0 Bm obtained as the quotient of the algebra of eventually
constant sequences of elements in the Bm ’s,
n
o
Y
(bj )j∈N0 ∈
Bj ∃m ∈ N0 such that bj = bm for all j ≥ m ,
j∈N0
S
by its ideal of sequences (bj )j∈N0 which are eventually 0. If each Bm is a C ∗ -algebra, m∈N0 Bm can
be completed in a canonical C ∗ -norm to obtain a C ∗ -algebra which, if the ιm,l ’s are injective, can be
S
k·k
identified with the non-disjoint union m∈N0 Bm [ArMa1, §1.2].
It was observed in [Hawk1] that the sequence of algebras Bm := B(Hm ) arising in quantization
has a structure resembling that of an inductive system, although the map from Bm to Bm+1 is
not a homomorphism in the category of C ∗ -algebras. If we want to obtain a C ∗ -algebra C(M ) of
continuous functions on a manifold as an inductive limit of finite-dimensional matrix algebras, then
requiring Bm ⊂ Bm+1 says by definition that C(M ) is an AF algebra. This forces M to be totally
disconnected. Hence we must relax the notion of inductive limit.
4.1 Relaxed definition of inductive limits
Blackadar and Kirchberg introduced a more general inductive-limit-type construction [BlKi1]. Although never pointed out in the literature, the system of finite-dimensional C ∗ -algebras B(Hm )
obtained from a projective quantization (M, ω, L) fits perfectly into their framework. This is most
apparent in [Hawk1] where similar notions were introduced independently. We will follow the notation
of [Hawk1] as closely as possible.
Notation 4.1. If B• = (Bm )m∈N0 is a sequence of C ∗ -algebras, we write
Y
Bm
Γb (B• ) =
m∈N0
for the full C ∗ -direct product of the Bm ’s, i.e. the set of sequences X• = (Xm )m∈N0 of elements
Xm ∈ Bm with finite supremum norm
kX• k := sup kXm kBm < ∞.
m∈N0
The multiplication and ∗-operation in Γb (Bm ) is pointwise. We also write
Γ0 (B• ) =
M
m∈N0
Bm
for the C ∗ -direct sum, the closed two-sided ideal in Γb (B• ) consisting of the sequences converging
to zero in norm. We simply write Γb := Γb (B• ) etc. if it is clear which sequence B• it concerns. We
let
π : Γb → Γb /Γ0
be the quotient map
Since we will only deal with a special kind of the “generalized inductive systems” defined in [BlKi1]
(namely the “NF” ones), we will simply refer to them as “inductive systems”. See also [BrOz1, §11].
Definition 4.2. An inductive system is a sequence (B• , ι• ) of full matrix algebras Bm = Mk(m) (C)
and unital completely positive maps ιm,l : Bm → Bm+l for l ≥ m (with ιm,m := id) satisfying
ιm,l = ιr,l ◦ ιm,r ,
19
if m ≤ r ≤ l
(4.1)
and which are asymptotically multiplicative in the sense that for all A, B ∈ B(Hm ), ε > 0, there
are r ≤ l such that
kιr,l (ιm,r (A)ιm,r (B)) − ιm,l (A)ιm,l (B)k < ε.
(4.2)
The inductive limit of an inductive system (B• , ι• ) is the C ∗ -algebra
B∞ ⊂ Γb (B• )/Γ0 (B• )
generated by the elements
ς (m) (A) := π((ιl (A))l≥m ),
A ∈ B(Hm )
(4.3)
for all m ∈ N0 , where π : Γb → Γb /Γ0 is the quotient map.
Remark 4.3 (Norm). A norm on the quotient C ∗ -algebra Γb /Γ0 is given by
kπ(X• )k = lim sup kXm k,
∀X• ∈ Γb ,
m→∞
and this norm satisfies the C ∗ -identity, hence it is the unique C ∗ -norm on Γb /Γ0 . Moreover, since
the ιm,l ’s are norm-decreasing,
lim sup kς (m) (A)k = lim kιm,l (A)k,
l→∞
m→∞
∀A ∈ B(Hm ),
so the norm on B∞ is just the “norm-at-infinity” of π −1 (B∞ ).
It follows that the maps
ς (m) : Bm → B∞
(4.4)
(m)
are completely positive, and we refer to ς
as the covariant Berezin symbol map at level m.
The motivation for this terminology will become clear below. Due to (4.1), the covariant symbol
maps satisfy
ς (l) ◦ ιm,l = ς (m) ,
∀m ≤ l ∈ N0 .
(4.5)
We may say that a sequence A• ∈ Γb is eventually constant under ι•,• if there is a large enough
m ∈ N0 such that Al = ιr,l (Ar ) for all l ≥ r ≥ m. Then B∞ is the image under π of the norm closure
of the algebra of eventually constant sequences.
Remark 4.4 (Asymptotic multiplicativity). The condition (4.2) is chosen precisely to ensure that
ς (m) (A)ς (m) (B) belongs to the C ∗ -algebra B∞ for all A, B ∈ B(Hm ), without requiring it to be close
to ς (m) (AB). Conversely, if ς (m) (A)ς (m) (B) belongs to B∞ for all A and B then each ς (m) (A)ς (m) (B)
is an eventually constant sequence, so (4.2) must hold.
Remark 4.5 (Continuous fields). Suppose that B(•) is a continuous field of matrix algebras over
N0 ∪ {∞}. For each A ∈ B(∞), thereQis a continuous section x → A(x) of B(•) with A(∞) = A,
and this section
L defines an element of m∈N0 B(m). Two sections evaluating to A at ∞ differ by an
element in m∈N0 B(m). Hence [BlKi1, Prop. 2.2.3]
Y
. M
B(∞) ⊂
B(m)
B(m) .
m∈N0
m∈N0
In fact, a C ∗ -algebra is an inductive limit (in the sense of Definition 4.2) if and only if it is a nuclear
separable C ∗ -algebra which is of the form B(∞) for some continuous field of matrix algebras over
N0 ∪ {∞} [BlKi1, Thm. 5.2.2].
Remark 4.6 (Quasi-diagonality). If a C ∗ -algebra B∞ is an inductive limit in the sense of Definition
4.2, there exists a short-exact sequence
0 −→ K −→ D −→ B∞
which is an “essential quasi-diagonal extension” of B∞ , meaning that D is a C ∗ -algebra of quasidiagonal operators containing the C ∗ -algebra K of compact operators as an essential ideal. In fact,
such an extension of a C ∗ -algebra B∞ exists if and only if B∞ can be embedded into Γb (B• )/Γ0 (B• )
for some sequence of full matrix algebras Bm [BlKi1].
20
Remark 4.7 (Nuclearity). Since B∞ is nuclear, the Choi–Effros lifting theorem [Blac1, IV.2.3.4]
says that the identity mapping id : B∞ → B∞ can be lifted to a unital completely positive map
ς˘ : B∞ → Γb such that if
ς˘(m) (f ) := ς˘(f )pm ∈ B(Hm ),
∀f ∈ B∞ ,
then the sequence ς (m) ◦ ς˘ converges in the point-norm topology to id : B∞ → B∞ . We shall calculate
ς˘ and its inverse ς explicitly in §5.8.
4.2 Inductive limits from subproduct systems
Lemma 4.8. Let H• be a subproduct system and define completely positive maps ιm,l : Bm → Bl by
ιm,l (A) := pl (A ⊗ 1H⊗(l−m) )pl ,
∀A ∈ B(Hm ),
where pl : H⊗l → Hl is the projection. Then for all A ∈ B(Hm ) we have the formulae
X
Rk ARk∗ H
ιm,l (A) =
l
(4.6)
(4.7)
|k|=l−m
=
X
Aj,k Sj Sk∗
Hl
(4.8)
|j|=m=|k|
where Rk is the right shift by the vector ek as in (2.6) and Aj,k := hej |Aek i.
Proof. Formula (4.8) is immediate from
Sj Sk∗ |Hl = pl (|pm ej ihpm ek | ⊗ 1Hl−m )pl .
Similarly, the expression (4.7) is deduced from straightforward calculations.
Theorem 4.9. Every subproduct system H• defines a generalized inductive system (B• , ι•,• ) by setting
Bm := B(Hm ) and letting ιm,l : Bm → Bl be as in (4.6).
Proof. It is clear that each ιm,l : Bm → Bl is unital and completely positive. For m ≤ r ≤ l and
A ∈ B(Hm ) we have
ιr,l ◦ ιm,r (A) = ιm,l (pr (A ⊗ 1H⊗(r−m) )pr ) = pl (pr (A ⊗ 1H⊗(r−m) )pr ⊗ 1H⊗(l−r) )pl
= pl (A ⊗ 1H⊗(l−m) )pl ,
where the last equality is due to (2.2). That is, the coherence condition (4.1) holds.
It remains to show that ι•,• is asymptotically multiplicative, i.e. that it satisfies (4.2). We have
ιr,l (ιm,r (A)ιm,r (B)) = ιr,l pr (A ⊗ 1)pr (B ⊗ 1)pr
= pl pr (A ⊗ 1)pr (B ⊗ 1)pr ⊗ 1 pl
= pl (A ⊗ 1)pr (B ⊗ 1) ⊗ 1 pl ,
so it is the failure of A ⊗ 1 to commute with the projection pr which spoils multiplicativity. It seems
hard to show directly from norm estimates that the maps (4.6) satisfy the asymptotic multiplicativity
condition (4.2). We shall instead obtain that by showing that the set of elements (4.3) forms an
algebra.
Consider the positively graded algebraic part R of the Toeplitz algebra TH , i.e. the N0 -graded
ring
M
M (m)
R=
Rm :=
TH .
m∈N0
21
m∈N0
Denote by Gr(R) the Abelian category of Z-graded right R-modules, with morphisms
L the gradingpreserving morphisms in the category of R-modules. Write R≥m for the R-module l≥m Rl . Then
it is straightforward to see that
EndGr(R) (R≥m ) = B(Hm )
as rings. Indeed, we have the left R-action on R≥m and the grading-preseving elements can all be
obtained by taking linear combinations of the elements Sj Sk∗ with |j| = m = |k|. Namely, the operator
Sj Sk∗ on HN is the direct sum T ⊕ 0 of an operator T ∈∈ B(H≥m ) and the zero operator 0 ∈ B(H<m ).
So Sj Sk∗ can be viewed as an operator of the R-module R≥m , and Sj Sk∗ is right R-linear because it
acts by multiplication from the left. Now the family (Sj Sk∗ )|j|=m=|k| forms an overcomplete set of
matrix units in B(Hm ), so their C-linear span identifies with B(Hm ). We have an inductive system
EndGr(R) (R≥m ) ∋ X|H≥m → Y |R≥l ∈ EndGr(R) (R≥l ),
∀m ≤ l ∈ N0
obtained by restriction to shorter tails R≥l ⊂ R≥m . From Equation (4.8) we see that this is precisely the algebraic inductive system underlying ιm,l : B(Hm ) → B(Hl ), under the identification
EndGr(R) (R≥m ) = B(Hm ).
By [Sten1, §IX.1] (see also [ArZh1, Example 5.4]), the algebraic inductive limit
0
B∞ = lim EndGr(R) (R≥m )
m→∞
is a ring (and an algebra over C since the maps ιm,l are C-linear). Therefore the norm closure
B∞ of 0 B∞ is an algebra as well. In particular, the asymptotic multiplicativity condition (4.2) is
satisfied.
Thus, subproduct systems give rise to generalized inductive limits of C ∗ -algebras with the special
property that the algebraic direct limit 0 B∞ is already an algebra (no need for norm closure). Still,
the asymptotic multiplicativity condition (4.2) cannot be formulated in a weaker fashion even for
subproduct systems, since the set of eventually constant sequences under ι•,• is only an algebra after
taking norm closure.
(0)
Our aim is to identify the inductive limit B∞ with the Cuntz–Pimsner core OH . For that we
need a lemma.
Q
Lemma 4.10. Let π −1 (B∞ ) be the norm closure of the subset of m B(Hm ) consisting of sequences
which are eventually constant under the coherent system ι•,• defined by (4.6). Then π −1 (B∞ ) coin(0)
cides with the normally ordered part of the Toeplitz core TH . Hence the normally ordered part of
(0)
(0)
TH is an algebra, and must coincide with all of TH , so
(0)
π −1 (B∞ ) = TH .
In this way we have proven Lemma 2.15.
(0)
Proof. It is clear that every normally ordered element of TH defines a sequence A• = (Am )m∈N0 of
operators Am ∈ B(Hm ) with ιm,l (Am ) = Al for sufficiently large m ≤ l. For example,
Y
Sj Sk∗ = (Sj Sk∗ |Hm )m∈N0 ∈
B(Hm ).
m∈N0
Suppose now that A = (Am )m∈N0 is any element of
(0)
TH
Q
m∈N0
B(Hm ) which is eventually constant.
Since B(Hm ) is contained in
for each m, we may for simplicity just as well look at the case
where ιr,l (Ar ) = Al for all r ≤ l for some r ∈ N0 while Am = 0 for m ≤ r. Then (4.8) in Lemma 4.8
shows that A is a combination of shift operators.
From the fact that B∞ is an algebra we have that π −1 (B∞ ) is an algebra, whence the last
statement.
22
Remark 4.11. Let B(•) be the continuous field of C ∗ -algebras over N0 ∪{∞} such that the fiber over
m ∈ N is B(m) = B(Hm ) and the fiber over ∞ is B(∞) = B∞ , the inductive limit (cf. Remark 4.5).
(0)
Then Lemma 4.10 says that TH is the algebra of continuous sections of this field. For commutative
case see also [Hawk2, Thm. 3.3].
(0)
Theorem 4.12. Let H• be a subproduct system and let OH denote the U(1)-invariant part of the
Cuntz–Pimsner algebra of H• . Then we have
(0)
OH ∼
= B∞ ,
where the right-hand side is the inductive limit defined by the inductive system ι•,• in (4.6).
(0)
Proof. The Toeplitz core TH is the norm closure of linear combinations of elements of the form
Sj Sk∗ with |j| = |k| as well as their products with the vacuum projection p0 = |ΩihΩ|. The elements
(0)
which are products with |ΩihΩ| belong to Γ0 = K ∩ TH . Hence,
(0)
(0)
(0)
(0)
B∞ = TH /Γ0 = TH /(K ∩ TH ) = OH ,
as asserted.
Remark 4.13. The quasi-diagonal extension of B∞ mentioned in Remark 4.6 can now be taken as
(0)
D = TH + K (cf. [Blac1, V.4.2.16]).
4.3 Cuntz–Pimsner algebras from inductive limits
For m > 0, let H−m := Hm denote the conjugate Hilbert space of Hm . For all k ∈ Z we can consider
the B(Hm )-module
(k)
Em
:= B(Hm , Hm+k ),
(k)
(k)
(k)
and the maps ιm,l : Em → El
defined by
(k)
ιm,l (X) :=
X
Rr XRr∗
Hl
,
|r|=l−m
(k)
(k)
∀X ∈ Em
.
(k)
We define the C ∗ -algebras Γb (E• ) and Γ0 (E• ) in the same way as in Notation 4.1 and we denote
(k)
(k)
(k)
by π (k) : Γb (E• ) → Γb (E• )/Γ0 (E• ) the quotient map.
0 (k)
Define E
to be the vector space consisting of all elements of the form
(k)
ς (m,k) (X) := π (k) ιm,l (X) l≥m ,
(k)
X ∈ Em
(0)
for all m ∈ N0 . In particular, 0 E (0) ≡ 0 B∞ is the algebraic part of B∞ ≡ E (0) ∼
= OH . Each 0 E (k) is
0
a module over B∞ . The linear span of
B∞ 0 E (k) := {f ψ|f ∈ B∞ , ψ ∈ 0 E (k) }
is a module over B∞ , which we denote by E (k) .
Theorem 4.14. The Cuntz–Pimsner algebra OH is isomorphic to the C ∗ -algebra generated by E (1)
and B∞ . It allows the decomposition
k·k
M
E (k)
OH ∼
=
k∈Z
and E (k) ∼
=
(k)
OH
is the spectral subspace for the gauge action corresponding to k ∈ Z.
23
Proof. The vector space B(Hm , Hm+k ) has an overcomplete basis given by the operators Sk |Hm for
all k ∈ F+
n with |k| = k. In particular, B(Hm , Hm+1 ) is spanned by Sj |Hm for j = 1, . . . , n. Recalling
that Sj is the shift by the basis vector ej ∈ H, we see that
X
(k)
Rr Sj Rr∗ H
ιm,l (Sj |Hm ) =
l
|r|=l−m
=
X
Sj Rr Rr∗
Hl
|r|=l−m
= S j |H l .
We can identify a sequence X• = (Xm )m∈N0 of operators Xm ∈ B(Hm , Hm+k ) with an operator on
Fock space HN . The effect of the quotient map π (k) on such a sequence X• is to take it to its image
in the Calkin algebra B(HN )/K. From (4.5) we therefore have (for m ≥ 1)
ς (m,1) (Sj |Hm ) = ς (1,1) (Sj |H ) = π (1) (Sj Hm )m∈N0 = π (1) (Sj ) = Zj ,
where Z1 , . . . , Zn are the generators of OH . Similarly one gets that ς (l,k) (Sk |Hm ) is just Zk for all
(−m)
∗
k ∈ F+
) for |k| = m. So
n with |k| = k and all l ≥ m. The adjoints Sk define elements of Γb (E•
(k)
O
holds
for
all
k
∈
Z.
E (k) ∼
= H
4.4 Formulas for covariant symbols
Our discussion about inductive limits associated to subproduct system has been based on shift
operators on Fock space. We now observe that what we are doing is in fact a generalization of
Berezin quantization. First we show that there is a very simple expression for the maps ς (m) .
Theorem 4.15. Let Z1 , . . . , Zn be the images of the shifts of the subproduct system H• . Then the
(0)
covariant symbol map ς (m) : B(Hm ) → OH defined in (4.3) can be expressed as
X
Aj,k Zj Zk∗ .
ς (m) (A) =
|j|=m=|k|
Proof. From (4.3) we see that we need to express ιm in terms of the Toeplitz operators S1 , . . . , Sn ;
applying π transforms these into Z1 , . . . , Zn . But that is easily done using Lemma 4.8: the “second
quantization” of A,
X
Aj,k Sj Sk∗ ∈ B(HN ),
|j|=m=|k|
acts as ιm,l (A) on Hl for l ≥ m and as 0 on Hl for l < m. Applying the quotient map π to it, we
obtain
X
Aj,k Sj Sk∗ = π((ιm,l (A))l≥m ) = ς (m) (A),
π
|j|=m=|k|
and on the other hand,
π
X
|j|=m=|k|
Aj,k Sj Sk∗ =
X
Aj,k Zj Zk∗ .
|j|=m=|k|
Corollary 4.16. For all j, k ∈ F+
n with |j| = |k| = m and all l ≥ m we have
ς (l) (Sj Sk∗ |Hl ) = Zj Zk∗ .
Example 4.17. Let H• be a commutative subproduct system and let M be the compact manifold it
(0)
defines (see Corollary 3.7). Then ς (m) : B(Hm ) → OH coincides with the Berezin covariant symbol
map ς (m) : B(Hm ) → C(M ) (mentioned in §3.2).
24
Notation 4.18. Fix a faithful representation of OH on a Hilbert space H and let
u ∈ Mn (C) ⊗ B(H)
be a unitary n × n matrix with values uj,k ∈ B(H) such that the first row of u is given by the
generators Z1 , . . . , Zn of OH . Let uc be the matrix obtained from u by taking adjoints of each entry
uj,k . Denote by um the restriction of u⊗m from H⊗m to Hm and similarly for uc . Finally,
α(m) : B(Hm ) → B(Hm ) ⊗ B(H)
will be the map which takes A ∈ B(Hm ) to um (A ⊗ 1)u∗m .
The following formulas are known from the classical case to define the “Berezin covariant symbol”
in case we quantize a coadjoint orbit M = G/K (cf. [Per1], [Lan2]).
Proposition 4.19. Assume that u⊗m preserves the subspace Hm ⊂ H⊗m , in the sense that um =
⊗m
⊗m
u⊗m (pm ⊗ 1). Let |e⊗m
1 ihe1 | be the rank-1 projection onto the line spanned by e1 , where e1 is the
first basis vector in H. Then for all A ∈ B(Hm ) we have
⊗m
⊗m
c
(4.9)
ς (m) (A) = (Tr ⊗ id) (A ⊗ 1)uc∗
m (|e1 ihe1 | ⊗ 1)um
⊗m
= (Tr ⊗ id) α(m) (A)(|e⊗m
(4.10)
1 ihe1 | ⊗ 1) .
Proof. We have
⊗m
⊗m
c
uc∗
m (|e1 ihe1 | ⊗ 1)um =
X
|j|=m=|k|
so (4.9) is clear. For (4.10) we can use the formula
X
Ar,s Sj Sk∗
α(m) (A) =
Sj Sk∗ |Hm ⊗ Zk Zj∗
Hm
|j|=m=|k|
⊗ uj,r u∗k,s .
4.5 Commutative case and the Arveson conjecture
One of the most striking applications of our results is the Arveson conjecture (see Remark 2.17).
Corollary 4.20. Arveson’s conjecture holds for all homogeneous ideals I ⊂ C[z1 , . . . , zn ], i.e. the
Cuntz–Pimsner algebra OH of the subproduct system H• associated to I (as in Lemma 2.9) is commutative.
Proof. Lemma 2.15 together with [KeSh1, Prop. 4.14] gives the result.
In [Vas1], [Vas3] it was shown that for any continuous line bundle L → M , the Cuntz–Pimsner
algebra OE (defined in [Pims1]) of the Hilbert C(M )-bimodule E of continuous sections of L is
isomorphic to the C ∗ -algebra C(SL ) of continuous functions on the total space of the circle bundle
SL associated to L∗ . Recall that in the definition of OE (which is Pimsner’s original one) the tensor
products are taken over the coefficient algebra C(M ). We shall now see that, in the case (M, L) is a
polarized manifold, from a projectively induced quantization we can also obtain C(SM ) := C(SL ) as
the Cuntz–Pimsner algebra OH of the associated subproduct system H• .
Proposition 4.21. Let (M, L) be a polarized (not necessarily smooth) variety and let H• be a projectively induced quantization of (M, L) (the definition still makes sense in the non-smooth case), and
endow M with the complex (Hausdorff ) topology [Serre1, §2]. Then
C(SM ) ∼
= OH ,
and for all k ∈ Z,
(k)
Γ(M ; L⊗k ) ∼
= OH
as a Hilbert C(M )-bimodule. In particular, k = 0 gives C(M ) as an inductive limit.
25
(4.11)
Proof. As in the general noncommutative case, the commutative algebra OH is built up from Hilbert
(0)
modules over OH and the latter is generated by the images of the covariant symbol maps ς (m) . An
argument given in [CGR1, §4] shows that the ς (m) (A)’s (for all m ∈ N0 and all A ∈ Bm ) separate
points. The Stone–Weierstrass theorem gives that they form a dense subalgebra of C(M ). The
inclusion of Hm into Hm+1 coincide by construction with our ιm,m+1 , and the supremum norm
on C(M ) is seen to coincide with the norm on the inductive limit B∞ . Therefore, C(M ) ∼
= B∞ .
The result now follows from Theorem 4.12 and the well-know decomposition of C(SM ) into the
Γ(M ; L⊗m )’s (see §3.3).
In fact, given Corollary 4.20 we get Proposition 4.21 from the calculation of the space of multiplicative U(1)-valued functionals on OH done in [KeSh1, Prop. 2.4]. Indeed, the circle bundle SL
can be identified with the preimage of M under the map Cn \ {0} → CPn−1 which defined projective
space, and this is the boundary of the analytic variety discussed [KeSh1].
Note that it is not so obvious that we couldL
recover M completely (as a topological space) from
0
⊗m
) of M ⊂ P[H∗ ] is not a ring of
H• because the homogeneous coordinate ring
m∈N0 H (M ; L
functions on M . The choice of basis on H 0 (M ; L⊗m ) corresponds to a choice of algebra structure on
the ring C(M ) and the inner product on H 0 (M ; L⊗m ) to a choice of C ∗ -algebra structure on C(M ),
but all possible C ∗ -algebra structures obtain in this way are isomorphic.
In the following we use the terminology from Lemma 3.8.
Corollary 4.22. Let (M, L) be a polarized manifold and let H• be a projectively induced quantization
of (M, L). Then the Fubini–Study metric FS(h·|·i) on L associated with the inner product on H
coincides with the ∗-operation which defines the C ∗ -algebra OH , and is thus equal to the inductive
limit of the Hermitian pairings
B(Hm+1 , Hm ) × B(Hm+1 , Hm ) → B(Hm ),
(A, B) → A∗ B.
Consequently, (M, L) is balanced if and only if the limit state ω coincides with the the unique FS(h·|·i)balancing state. If M = G/K is a coadjoint orbit one also has (using Notation 2.12) for all j, k ∈ F+
n
with |j| = m = |k| that
ˆ
pj,k
ωd
1
=
,
(4.12)
Zj∗ Zk
vol(M, L) M
d!
Tr(pm )
where pm : H⊗m → Hm is the orthogonal projection and H• ⊂ H∨• is the subproduct system of
(M, L, h·|·i) and we denote by ω also the Fubini–Study Kähler form on M ⊂ CPn−1 .
Proof. Recall that the generators Z1 , . . . , Zn of OH satisfy the relation of the ideal which
defines M ,
P
so they can be identified with the homogeneous coordinates on M . Recall also that nk=1 Zk Zk∗ = 1,
which says that the adjoint operation on the operator system in OH spanned by Z1 , . . . , Zn is the
Fubini–Study metric h on L. Since products of m of the generators Z1 , . . . , Zn identify with elements
of H 0 (M ; Lm ) and since the ∗-operation on OH is given by (Zk1 · · · Zkm )∗ = Zk∗m · · · Zk∗1 , we get that
it is induced by the tensor-product metric hm on H 0 (M ; Lm ) for all m ∈ N.
If M = G/K is a coadjoint orbit then the limit state ω = ωp1 : C(M ) → C satisfies ω(Zj∗ Zk ) =
δj,k /n. The formula for ω on the products Zj∗ Zk then gives (4.12).
5 Projective limits
(0)
We now want to realize the same algebra OH as a projective limit. For this we need some background
information from [Hawk1, §B2].
26
5.1 Relaxing the notion of projective limit
In this paper, a “projective limit” will always refer to the following object which, in comparison
to more conventional C ∗ -algebraic projective limits, is defined in terms of completely positive maps
instead of C ∗ -homomorphisms.
Definition 5.1 ([Hawk1, §B2]). A projective system (B• , •,• ) is a sequence of finite-dimensional
matrix algebras Bm and norm-contracting completely positive mappings l,m : Bl → Bm for m ≤ l
satisfying l,m = r,m ◦ l,r for all m ≤ r ≤ l. The projective limit of (B• , •,• ) is the vector space
defined by
B ∞ := {A• = (Am )m∈N0 ∈ Γb (B• )| Am−1 = m,m−1 (Am ) for all m ∈ N},
equipped with the norm
kA• k := lim kAm k.
m→∞
∞
Remark 5.2. The intersection of B with Γ0 is {0}. We always identify B ∞ with its embedding into
Γb (B• )/Γ0 (B• ) because it is more likely that B ∞ is an algebra when multiplication is taken modulo
Γ0 . If we do so and then pull back B ∞ via the quotient map π : Γb → Γb /Γ0 , we obtain a vector
space π −1 (B ∞ ) which is much larger than B ∞ , namely
π −1 (B ∞ ) = B ∞ ∪ Γ0 (B• ).
(5.1)
Importantly, B ∞ is an algebra (hence a C ∗ -algebra) if and only if (5.1) is.
Remark 5.3. We could also define B ∞ as the set of elements
f = π (∞,m (f ))m∈N0
where the components ∞,m (f ) ∈ Bm satisfy ∞,m (f ) = l,m ◦ ∞,l (f ). We can regard ∞,m :=
liml→∞ l,m as the map from B ∞ to Bm which evaluates A• = (Am )m∈N0 ∈ B ∞ at m,
∞,m (A• ) = Am .
5.2 Changing the inner products
From now on Q ∈ B(H) is a positive invertible n × n matrix and let
Qm := pm Q⊗m |Hm
be the compression of Q⊗m ∈ B(H⊗m ) to the subspace Hm = pm H⊗m . We choose the orthonormal
basis e1 , . . . , en for H1 such that Q is diagonal and, as before, we let S1 , . . . , Sn be the shifts on HN
by these basis vectors. We shall write
Qj,k := (Q−1
m )j,k .
Qj,k := (Qm )j,k ,
We associate to each Qm a density matrix
(m)
ρ(m) = ρQ
:=
Qm
,
Tr(Qm )
and denote by φm the state on B(Hm ) defined by
φm (A) := Tr(ρ(m) A),
∀A ∈ B(Hm ).
Sometimes it will be useful to change the inner product on Hm to
hψ|φiρ(m) := hψ|ρ(m) φi,
∀φ, ψ ∈ Hm .
We stress that (unless H• = H⊗• )
ρ(m) 6=
Qm
= pm (ρ(1) )⊗m pm .
Tr(Q)m
27
(5.2)
Assumption 5.4. From now on Q ∈ B(H) is a positive invertible n × n matrix such that
(i) the operator Q⊗m on H⊗m preserves the subspace Hm for all m ∈ N0 , i.e. Qm = Q⊗m |Hm , and
(ii) ιm,l preserves the Q-traces, i.e.
φl ◦ ιm,l = φm ,
for all m ≤ l ∈ N.
Property (i) allows the subproduct condition to be maintained with the new inner products
h·|·iρ(m) (see Proposition 5.8 below) while property (ii) allows the construction of a state on OH
with very nice properties. We will show that in the examples of subproduct systems coming from
compact quantum groups there is always a matrix Q which satisfies these assumptions (i) and (ii). It
is possible to drop either (or both) of the assumptions (i) and (ii) and many of the constructions in
the next section will carry over; in particular there will be a limit state but with weaker quantization
properties. We shall elaborate on this slightly in the commutative case, where assumptions (i) and
(ii) hold with Qm = pm for all m ∈ N precisely when H• is a regular quantization (in the sense of
Remark 3.10).
Remark 5.5. The property pm Q⊗m pm = Q⊗m pm ensures that Qm is invertible; its inverse is
pm (Q⊗m )−1 pm because
(Q⊗m )−1 pm Q⊗m pm = (Q⊗m )−1 Q⊗m pm = pm ,
Q⊗m pm (Q⊗m )−1 pm = Q⊗m (Q⊗m )−1 pm = pm ,
where we used the fact that the inverse A−1 of any invertible matrix A preserves every A-invariant
subspace. We denote by Q−1
m this inverse of Qm . For l ≥ m we have
⊗l −1
Q−1
((Q⊗m ) ⊗ (Q⊗(l−m) ))pl = pl ,
l (Qm ⊗ Ql−m ) = (Q )
where we regard pl as an operator from Hm ⊗ Hl−m onto Hl .
5.3 The isometries
Now that Hm is endowed with the inner product (5.2) we discuss how B(Hm ) can be mapped into
B(Hl ) when m < l and calculate the explicit isometries Hl ֒→ Hl−m ⊗ Hm and Hl ֒→ Hm ⊗ Hl−m .
This construction would fail without the assumption that Q⊗m preserves Hm .
Proposition 5.6. The isometry from Hl into Hl−m ⊗ Hm is given by
s
Tr(Qm ) Tr(Ql−m ) X
pl−r er ⊗ Sr∗ ψ,
V̄m,l ψ =
Tr(Ql )
|r|=l−m
and its adjoint by
∗
V̄m,l
=
s
∀ψ ∈ Hl ,
(5.3)
Tr(Qm ) Tr(Ql−m ) (l) −1 (l−m)
(ρ ) (ρ
⊗ ρ(m) ).
Tr(Ql )
Proof. We proceed by first calculating the adjoint of the given operator (5.3). Let λm,l :=
For all ξ ∈ Hl−m , η ∈ Hm and all ψ ∈ Hl we have
∗
hV̄m,l
(ξ ⊗ η)|ψiρ(l) = hξ ⊗ η|V̄m,l ψiρ(l−m) ⊗ρ(m)
X
hξ ⊗ η|pl (er ⊗ Sr∗ ψ)iρ(l−m) ⊗ρ(m)
= λm,l
|r|=l−m
= λm,l
X
|r|=l−m
hξ ⊗ η|Sr Sr∗ ψiρ(l−m) ⊗ρ(m)
= λm,l h(ρ(l) )−1 (ρ(l−m) ξ ⊗ ρ(m) η)|ψiρ(l)
28
(5.4)
q
Tr(Qm ) Tr(Ql−m )
.
Tr(Ql )
and hence (5.4) holds. Finally, using Remark 5.5,
Tr(Qm ) Tr(Ql−m ) X
(ρ(l) )−1 (ρ(l−m) er ⊗ ρ(m) Sr∗ ψ)
Tr(Ql )
|r|=l−m
X
X
−1
pl (er ⊗ Sr∗ ψ)
Ql (Ql−m er ⊗ Qm Sr∗ ψ) =
=
∗
V̄m,l
V̄m,l ψ =
|r|=l−m
|r|=l−m
=
X
Sr Sr∗ ψ = ψ,
|r|=l−m
so V is the desired isometry. Let us also calculate the final projection:
X
∗
Sr Sr∗ Q−1
=
V̄m,l V̄m,l
l (Ql−m ⊗ Qm ) = pl (pl−m ⊗ pm ).
|r|=l−m
Now let A ∈ B(Hl ) and define
∗
ῑm,l (A) := V̄m,l
(1Hl−m ⊗ A)V̄m,l .
(5.5)
We have
Tr(Qm ) Tr(Ql−m ) X
(ρ(l) )−1 (ρ(l−m) er ⊗ ρ(m) ASr∗ ψ)
Tr(Ql )
|r|=l−m
X
∗
Q−1
=
l (Ql−m er ⊗ Qm ASr ψ)
∗
V̄m,l
(1Hl−m ⊗ A)V̄m,l ψ =
|r|=l−m
=
X
|r|=l−m
=
X
pl (er ⊗ ASr∗ ψ)
Sr ASr∗ ψ,
|r|=l−m
so ῑm,l is a “chirality-flipped” version of ιm,l , i.e. the Rr ’s are replaced by Sj ’s (cf. (4.7)). If we
use the inductive system ῑ•,• instead of ι•,• , the roles of the left and right Toeplitz algebras are
interchanged.
Proposition 5.7. Define a coherent system of maps ῑm,l : B(Hm ) → B(Hl ) by
∗
ῑm,l (A) := V̄m,l
(1Hl−m ⊗ A)V̄m,l
X
Ss ASs∗ H .
=
l
(5.6)
(5.7)
|s|=l−m
Then (B(H• ), ῑ•,• ) is an inductive system, and the C ∗ -subalgebra of Γb (B• ) consisting of norm limits
of the eventually constant sequences for ῑ•,• is equal to the U(1)-invariant part of the right Toeplitz
algebra C ∗ (R1 , . . . , Rn ).
Proof. The proof is very similar to the case of the ιm,l ’s (the “left case”).
P
We furthermore note that if we expand A ∈ B(Hm ) as A = |j|=m=|k| Aj,k Rj Rj∗ |Hm then
X
Aj,k Rj Rk∗ H ,
ῑm,l (A) =
l
|j|=m=|k|
again similar to the left case. More will be said on the “chiral duality” between ῑ•,• and ι•,• in
Remark 5.13.
We now want to find an isometric implementation of ιm,l similar to (5.5). For this we need to flip
the tensor factors.
29
Proposition 5.8. The isometry from Hl into Hm ⊗ Hl−m is given by
s
Tr(Qm ) Tr(Ql−m ) X
Vm,l ψ =
Rr∗ ψ ⊗ pl−r er ,
Tr(Ql )
|r|=l−m
and its adjoint by
∗
Vm,l
=
s
∀ψ ∈ Hl ,
Tr(Qm ) Tr(Ql−m ) (l) −1 (m)
(ρ ) (ρ
⊗ ρ(l−m) ).
Tr(Ql )
(5.8)
(5.9)
Proof. For all ξ ∈ Hl−m , η ∈ Hm and all ψ ∈ Hl we have
∗
hVm,l
(η ⊗ ξ)|ψiρ(l) = hη ⊗ ξ|Vm,l ψiρ(m) ⊗ρ(l−m)
X
hη ⊗ ξ|Rr Rr∗ ψiρ(m) ⊗ρ(l−m)
= λm,l
|r|=l−m
= λm,l h(ρ(l) )−1 (ρ(m) η ⊗ ρ(l−m) ξ)|ψiρ(l) ,
and the rest is similar to the proof of Proposition 5.8.
Corollary 5.9. The inductive system ι•,• is implemented by the system V•,• of isometries:
∗
ιm,l (A) = Vm,l
(A ⊗ 1Hl−m )Vm,l ,
∀A ∈ B(Hm ).
Proof. Follows from formula (4.7) and the calculation (with ψ ∈ Hl )
s
Tr(Qm ) Tr(Ql−m ) X
∗
Vm,l (A ⊗ 1Hl−m )Vm,l ψ =
(ρ(l) )−1 (ρ(m) ARr∗ ψ ⊗ ρ(l−m) er )
Tr(Ql )
|r|=l−m
X
−1
Ql (Qm ARr∗ ψ ⊗ Ql−m er )
=
|r|=l−m
=
X
|r|=l−m
=
X
pl (ARr∗ ψ ⊗ er )
Rr ARr∗ ψ.
|r|=l−m
5.4 Projective system for subproduct systems
Let H• be a subproduct system. We let Q ∈ B(H) be as in Assumption 5.4 and endow Hm with the
inner product defined by the density matrix ρ(m) := Qm / Tr(Qm ).
Lemma 5.10. Define maps l,m : B(Hl ) → B(Hm ) for m ≤ l by
l,m (A) :=
Tr(Qm )
Tr(Ql )
X
(Q⊗m )k,k Rk∗ ARk
Hm
,
|k|=l−m
∀A ∈ B(Hl ).
(5.10)
Then, with Vm,l as in Proposition 5.8, we have the formula
∗
)
l,m (A) = (idBm ⊗φl−m )(Vm,l AVm,l
and •,• is a projective system.
30
(5.11)
Proof. First of all, for all A ∈ B(Hl ) we have
r,m ◦ l,r (A) =
=
=
X
Tr(Qr ) Tr(Qm )
Tr(Ql ) Tr(Qr )
Tr(Qm )
Tr(Ql )
Tr(Qm )
Tr(Ql )
(Q⊗m )j,j (Q⊗m )k,k Rj∗ Rk∗ ARk Rj
Hm
|j|=r−m,|k|=l−r
X
∗
(Q⊗m )kj,kj Rkj
ARkj
Hm
|kj|=l−m
X
(Q⊗m )r,r Rr∗ ARr
Hm
= l,m (A),
|r|=l−m
and it is obvious that each l,m is completely positive. The norm-contracting property holds because
l,m is in fact unital. To see this we first prove the alternative formula (5.11). We have
l,m (A) =
=
=
=
Tr(Qm )
Tr(Ql )
Tr(Qm )
Tr(Ql )
Tr(Qm )
Tr(Ql )
Tr(Qm )
Tr(Ql )
X
(Q⊗m )r,r Rr∗ ARr
Hm
|r|=l−m
X
X
(Q⊗m )r,r hej |Rr∗ ARr ek iSj Sk∗
Hm
|r|=l−m |j|=m=|k|
X
X
(Q⊗m )r,r hej ⊗ er |A(ek ⊗ er )iSj Sk∗
Hm
|r|=l−m |j|=m=|k|
X
X
(Q⊗m )r,r (A)jr,kr Sj Sk∗
Hm
|r|=l−m |j|=m=|k|
∗
= (id ⊗φl−m )(Vm,l AVm,l
),
where in the last equality we used that, for all ξ1 , ξ2 ∈ Hm , η1 , η2 ∈ Hl−m ,
∗
∗
hVm,l
(ξ1 ⊗ η1 )|AVm,l
(ξ2 ⊗ Ql−m η2 )i
=
Tr(Qm ) Tr(Ql−m ) (l)−1 (m)
hρ
(ρ ξ1 ⊗ ρ(l−m) η1 )|Aρ(l)−1 (ρ(m) ξ2 ⊗ ρ(l−m) Ql−m η2 )i
Tr(Ql )
= hξ1 ⊗ η1 |Aρ(l)−1 (ρ(m) ξ2 ⊗ ρ(l−m) Ql−m η2 )i
=
Tr(Ql )
hξ1 ⊗ η1 |A(ξ2 ⊗ Ql−m η2 )i,
Tr(Ql−m ) Tr(Qm )
m)
∗
so that summing such inner products over a basis for Hm ⊗Hl−m and multiplying with Tr(Q
Tr(Ql ) Sj Sk Hm
∗
is the same thing as partially tracing Vm,l AVm,l
with Ql−m / Tr(Ql−m ).
The formula (5.11) shows that l,m is the adjoint of ιm,l with respect to φl and φm (see details in
5.11 below). Our assumption φl ◦ ιm,l = φm is then equivalent to the unitality
l,m (pl ) = pm ,
and hence l,m is contractive.
Notice that since l,m intertwines the normalized traces, the unitality assumption on l,m is
equivalent to assuming that l,m is contractive.
Proposition 5.11. Let m, l ∈ N0 with m ≤ l. Then l,m is the adjoint of ιm,l : for all A ∈ B(Hl )
and all B ∈ B(Hm ) we have
φl (Aιm,l (B) = φm l,m (A)B .
(5.12)
In particular, taking A = pl respectively B = pm we obtain the equivalences
φl ◦ ιm,l = φm ⇐⇒ l,m (pl ) = pm ,
31
(5.13)
φl = φm ◦ l,m ⇐⇒ ιm,l (pm ) = pl ,
(5.14)
where (5.14) holds for any subproduct system while (5.13) is our standing assumption for this section.
Proof. We have
∗
φl (Aιm,l (B)) = φl (AVm,l
(B ⊗ pl−m )Vm,l )
∗
= (φm ⊗ φl−m )(Vm,l AVm,l
(B ⊗ pl−m ))
∗
(B ⊗ pl−m ))
= φm (idBm ⊗φl−m )(Vm,l AVm,l
∗
)B ,
= φm (idBm ⊗φl−m )(Vm,l AVm,l
which equals φm l,m (A)B by (5.11).
Corollary 5.12. The states φm satisfy the “right invariance” condition
φm (l,m ◦ ιm,l (A)) = φm (A),
∀A ∈ B(Hm ).
(5.15)
Proof. Just use (5.14) and then (5.13).
Remark 5.13. Similarly one shows that
φm = φl ◦ ῑm,l .
(5.16)
for the “right” inductive system ῑ•,• . The adjoint of ῑm,l is
̄l,m (A) =
Tr(Qm )
Tr(Ql )
X
(Q⊗m )r,r Sr∗ ASr
Hm
|r|=l−m
∗
= (φl−m ⊗ id)(V̄m,l AV̄m,l
).
A “left invariance” condition similar to (5.15) is also deduced using the ῑm,l ’s and their adjoints.
5.5 The state on OH
Corollary 5.14. The limit
φ∞ := lim φm
m→∞
is a well-defined state on
(0)
TH .
It annihilates
(0)
TH
(0)
∩ K, so it descends to a state ωQ on OH = B∞ .
(0)
Proof. For well-definedness we use (5.13) and recall that the elements of TH are norm limits of
(0)
(0)
eventually constant under ι•,• . The fact that φ∞ descends to TH /(TH ∩ K) follows from
|φm (A)| ≤ kAk,
∀A ∈ B(Hm ), m ∈ N0 ,
(0)
since this shows that limm kAm k = 0 implies φ∞ (A• ) = 0 for all A• = (Am )m∈N0 ∈ TH .
Proposition 5.15. The state ωQ : B∞ → C is KMS, with modular automorphism group σ• = (σt )t∈R
given by
−it
σt ◦ ς (m) (A) = ς (m) (Qit
m AQm )
for all A ∈ B(Hm ) and all m ∈ N0 , and ωQ satisfies
ωQ (Zj Zk∗ ) =
Qk,j
Tr(Qm )
(5.17)
(m)
for all j, k ∈ F+
: B(Hm ) → B∞
n with |j| = |k| = m. Moreover, the covariant symbol map ς
intertwines ωQ and φm :
ωQ ◦ ς (m) = φm .
(5.18)
32
Proof. Due to (5.13) we have, if |j| = |k| = m,
ωQ (Zj Zk∗ ) = φ∞ (Sj Sk∗ )
= lim φl (Sj Sk∗ pl )
m≤l→∞
= φm (Sj Sk∗ pm ),
and so the first formula in (5.17) follows from
X
(Q⊗m )r,r her |pm ej ihek |pm er i
Tr(Qm )φm (Sj Sk∗ ) =
|r|=m
= hek |Qm ej i.
The definition of ς (m) immediately gives (5.18), again using (5.13).
That ωQ is KMS follows from (5.18), in view of the fact that the ∗-algebra generated by the
covariant symbols ς (m) (A) is dense in B∞ and that each φm is KMS. Finally, for t ∈ R the modular
−it
automorphism σtφm of φm takes A ∈ B(Hm ) to (ρ(m) )it A(ρ(m) )−it = Qit
m AQm .
We can extend ωQ to a state, still denoted by ωQ , on the whole Cuntz–Pimsner algebra by defining
(k)
(0)
it to be zero on each spectral subspace OH except OH .
Remark 5.16. The property (5.18) of the limit state relies on Assumption 5.4(ii) and ensures that
ω is faithful. Without this assumption we could still obtain a limit ωQ of the states φm but it is not
clear what would guarantee its faithfulness. We could go the GNS representation of B∞ associated
with ωQ and use the faithful state induced by ωQ on the image of B∞ , which is a quotient of B∞
(recall that ωQ is KMS). Then analogous results hold for the image of B∞ in the GNS representation.
Example 5.17. For the product system H⊗• , the Cuntz–Pimsner algebra OH is the Cuntz algebra
On and ωQ is the quasi-free state on On defined by the density matrix Q/ Tr(Q) [Ev1].
5.6 Contravariant symbols
Proposition 5.18. The adjoint ς˘(m) : B∞ → B(Hm ) of the covariant symbol map ς (m) : B(Hm ) →
B∞ , defined by the relation
∀A ∈ B(Hm ), f ∈ B∞ ,
ωQ (ς (m) (A)∗ f ) = φm A∗ ς˘(m) (f ) ,
is given by
X
ς˘(m) (f ) = Tr(Qm )
∗
∗
(Q⊗m )−1
j,j ωQ (Zj Zk f )Sk Sj
Hm
(5.19)
|j|,|k|=m
Proof. Let ς˘(m) (f ) be defined by (5.19). Then
X
(A∗ )j,k ωQ (Zj Zk∗ f )
ωQ (ς (m) (A)∗ f ) =
|j|=m=|k|
=
X
∗
∗
(Q⊗m )j,j (Q⊗m )−1
j,j (A )j,k ωQ (Zj Zk f )
|j|=m=|k|
= φm A∗ ς˘(m) (f ) .
Corollary 5.19. We have
(m)
and, moreover, each ς˘
ωQ = φm ◦ ς˘(m)
is unital.
33
(5.20)
Proof. Equation (5.20) is a direct consequence of the fact that ς˘(m) is adjoint to the unital map ς (m) .
Unitality of ς˘(m) follows from ωQ (ς (m) (A)∗ 1) = φm (A∗ ), which we know from (5.18).
We can now assembly the ς˘(m) ’s to a map
Y
Y
B(Hm ),
ς˘(m) : B∞ →
ς˘ :=
(5.21)
m∈N0
m∈N0
which is a noncommutative generalization of the total Toeplitz map (3.5). We can recover its components as
ς˘(m) (f ) = ς˘(f )pm .
The following result which relies on the fact that ωQ is faithful (cf. Remark 5.16).
(0)
Lemma 5.20. No nonzero element of B∞ is mapped to Γ0 = TH ∩ K under the map ς˘.
Proof. We have φm ς˘(m) (f ) = ωQ (f ), so if ς˘(m) (f ) → 0 as m → ∞ then ωQ (f ) = 0. Hence if f ≥ 0
then f = 0 and the result follows.
Let M = πωQ (B)′′ be the von Neumann algebra generated by the inductive limit B∞ in the GNS
representation of the limit state ωQ . Then we can define ς˘(f ) ∈ Γb also for elements in M, and
Lemma 5.20 extends to M.
Lemma 5.21. For all f ∈ M and all l ≥ m,
ς˘(m) (f ) = l,m ◦ ς˘(l) (f ).
(5.22)
Hence the image of M under the total Toeplitz map ς˘ is contained in the projective limit B ∞ , and in
fact we have equality
ς˘(M) = B ∞ .
Therefore B ∞ can be identified with the weak-∗-closed operator system of elements of the form
ς˘(f ) = (˘
ς (m) (f ))m∈N0 ,
f ∈M
and, as in Remark 5.3,
ς˘(m) = ∞,m
is the map which evaluates (Xm )m∈N0 ∈ B ∞ at m ∈ N0 . The norm-closed subset ς˘(B∞ ) equals the
anti-normally ordered part of the Toeplitz core.
Proof. We know that ς˘ is injective (Lemma 5.20). Since we have shown that l,m is adjoint to ιm,l ,
we obtain (5.22) by taking adjoints of
ς (l) ◦ ιm,l = ς (m) .
From (5.22) follows that ς˘(f ) ∈ B ∞ for all f ∈ M. Moreover, ς˘ : M → Γb is onto B ∞ because each
ς˘(m) is onto. Thus B ∞ is in bijection with M via ς˘.
(0)
We need to show that ς˘(B∞ ) equals the anti-normally part of the Toeplitz core TH . Firstly,
since the left and right shifts commute outside the vacuum subspace, for all r, s = 1, . . . , n we have
l,m (Sr∗ Ss pl ) =
=
Tr(Qm )
Tr(Ql )
Tr(Qm )
Tr(Ql )
X
(Q⊗m )k,k Rk∗ Sr∗ Ss Rk
Hm
|k|=l−m
X
(Q⊗m )k,k Sr∗ Rk∗ Rk Ss
|k|=l−m
Hm
= Sr∗ Ss |Hm
(0)
(where we used that l,m (pl ) = pm ), which shows that the anti-normally ordered elements of TH are
constant under •,• . Secondly, an explicit calculation using (5.19) shows that ς˘(f ) is anti-normally
ordered for each f ∈ B∞ .
34
Remark 5.22. Now we can give an alternative proof for the fact that the contravariant symbol map
ς˘(l) : B∞ → B(Hm ) intertwines ωQ with φl ,
ωQ = φl ◦ ς˘(l) .
Recall that ωQ denotes the limit state φ∞ := limm→∞ φm when regarded as a state on the quotient
(0)
B∞ of TH . Then ωQ = φl ◦ ς˘(l) follows from the compatibility φm = φl ◦ m,l (see (5.13)) and the
fact that ∞,l = ς˘(l) .
5.7 The asymptotic multiplication
We now endow the projective limit B ∞ with a multiplication which is the m → ∞ limit of the
multiplication on B(Hm ).
Definition 5.23. The projective-limit multiplication on B ∞ ⊂ Γb is defined by
ς˘(f ) · ς˘(g) := lim ς˘(m) (f g)
m→∞
(5.23)
for all f, g ∈ B∞ .
The projective limit B ∞ is not an algebra under the projective-limit multiplication, but we shall
see that the subset ς˘(B∞ ) is.
The multiplication on B ∞ taken modulo Γ0 is the one where sequences (˘
ς (m) (f ))m∈N0 and
(m)
(˘
ς (g))m∈N0 are multiplied componentwise but the finite-m part is ignored. That is,
π(˘
ς (f )˘
ς (g)) = lim ς˘(m) (f )˘
ς (m) (g).
m→∞
(5.24)
We will see momentarily that the products (5.23) and (5.24) coincide for f, g ∈ B∞ . Comparing the
two formulas one then concludes that the Toeplitz maps ς˘(m) are “asymptotically multiplicative”.
Again the projective limit B ∞ is not an algebra under the multiplication modulo compacts, while
ς˘(B∞ ) will be shown to be so.
Remark 5.24 (Filters). A projective-limit multiplication can be defined using any filter ω on N.
On the C ∗ -level this corresponds to considering not a subalgebra of Γb /Γ0 but a subalgebra of Γb /Γω
where Γω is the ideal consisting of the sequences A• with
lim kAm k = 0.
ω
We recover Γ0 if ω is the free filter of all cofinite subsets of N0 . Confer [RoSt1, §6.2].
5.8 The adjoint of the total Toeplitz map
Lemma 5.25. For A ∈ B(Hl ) we have
l,0 (A) = φl (A)p0 .
(5.25)
(0)
and hence, if ε̂ denotes the vacuum state restricted to TH ,
ε̂ ◦ l,0 = φl .
The vacuum state ε̂ restricted to B ∞ is equal to ε̂ ◦ ∞,0 and coincides with the limit state φ∞ :=
liml→∞ φl ,
ε̂ ◦ ∞,0 = φ∞ .
35
Proof. We use φm ◦ l,m = φl for m = 0. This gives (5.25). Alternatively, note that for each positive
operator A on B(Hm ) one has
X
Ak,k
Tr(A) =
|k|=m
where Ak,k := hpm ek |Apm ek i. Consequently,
φm (A) =
X
1
hΩ|Sk∗ ASk Ωi = φ0 (m,0 (A)).
dim Hm
|k|=m
Letting m → ∞ one obtains
ω(f ) = φ0 (˘
ς (0) (f )),
∀f ∈ C 0 (M), f ≥ 0.
The rest is obvious.
We can therefore regard ε̂ as a state on the projective limit modulo compact operators as well,
i.e. on the algebra π(B ∞ ) ⊂ Γb /Γ0 .
Recall that the covariant-symbol map ς (m) is the adjoint of ς˘(m) , for each m ∈ N. We now show
that the total Toeplitz map ς˘ has an adjoint as well. This should be compared with [INT1, Lemma
2.3].
Proposition 5.26. There exists a completely positive map
ς˘∗ : ς˘(B∞ ) → B∞
such that, for all X ∈ ς˘(B∞ ) and all f ∈ B∞ ,
ωQ (˘
ς ∗ (X ∗ )f ) = ε̂(π(X ∗ )π(˘
ς (f ))).
(5.26)
Explicitly, this map is given by the point-norm limt ς˘∗ = limm→∞ ς (m) ,
ς˘∗ (X) = lim ς (m) (Xpm ),
m→∞
∀X ∈ ς˘(B∞ ),
and will be denoted by ς.
Proof. We identify X ∈ ς˘(B∞ ) with a bounded sequence (Xm )m∈N0 of operators Xm = Xpm ∈
B(Hm ). Using the formula (5.24) for the multiplication in π(B ∞ ) we have, by norm-continuity of the
vacuum state, the norm limits
∗ (m)
ε̂(π(X ∗ )π(˘
ς (f )))p0 = Ω lim Xm
ς˘ (f )Ω p0
m→∞
∗ (m)
ς˘ (f )
= ∞,0 lim Xm
m→∞
∗ (m)
ς˘ (f )
= lim m,0 Xm
m→∞
∗ (m)
= lim φm Xm
ς˘ (f ) p0
m→∞
∗
)f p0
= lim ωQ ς (m) (Xm
m→∞
∗
)f p0 .
= ωQ lim ς (m) (Xm
m→∞
Being a point-norm limit of completely positive maps, ς is completely positive.
Since B ∞ = ς˘(M) contains no compact operator, it is clear from the definition of the covariant
symbol that ς restricts to a bijection from B ∞ onto M and that ς ◦ ς˘ is the identity on M. Hence ς˘
is an isometry. So we have a decomposition of the identity map on B∞ ,
id = ς ◦ ς˘ = lim ς (m) ◦ ς˘(m) ,
m→∞
36
making Remark 4.7 explicit. We have now seen that ς˘ : B∞ → ς˘(B∞ ) is a complete order isomorphism,
i.e. a bijective unital completely positive map with completely positive inverse.
There is also a version of this result on the level of von Neumann algebras. As we shall see in
§6.5.2, for any subproduct system H• , the weak-∗-closed operator system B ∞ becomes a von Neumann
algebra when equipped with a SOT-version of the projective-limit multiplication (5.23). When H•
is the G-subproduct system (see §6 below), B ∞ is an operator system in the group-von Neumann
algebra R(G).
Corollary 5.27. The total Toeplitz map ς˘ intertwines the state ωQ on M = πωQ (B∞ )′′ with the
vacuum state ε̂ on B ∞ ⊂ B(HN ),
ωQ = ε̂ ◦ ς˘,
(5.27)
and similarly
ωQ ◦ ς = ε̂,
(5.28)
Proof. Take X = 1 respectively f = 1 in (5.26) to get (5.27) respectively (5.28).
(0)
5.9 OH as a projective limit
Theorem 5.28. The operator system ς˘(B∞ ) ⊂ B ∞ is a C ∗ -algebra with the multiplication taken
(0)
modulo compacts; indeed π(˘
ς (B∞ )) is isomorphic as a C ∗ -algebra to the inductive limit B∞ ∼
= OH .
Remark 5.29. We saw in Lemma 5.20 that the image of ς˘ does not contain Γ0 . On the other hand,
ς˘(B∞ ) + Γ0 is generated by Ran ς˘ alone as a C ∗ -algebra. The theorem implies that
ς˘(f )˘
ς (g) − ς˘(f g) ∈ Γ0
even for nontrivial f, g ∈ B∞ . These elements do not belong to B ∞ , which is why we need to apply
the quotient map π : Γb → Γb /Γ0 in order to obtain an algebra.
Proof. This is a well-known consequence of the fact that ς˘ : B∞ → ς˘(B∞ ) is a complete order
isomorphism from a C ∗ -algebra B∞ onto the operator system ς˘(B∞ ); see [Arv10, Prop. 2.2].
Corollary 5.30. The projective-limit multiplication on ς˘(B∞ ) ⊂ Γb coincides with the multiplication
on ς˘(B∞ ) taken modulo Γ0 .
Proof. By uniqueness of the C ∗ -algebraic structure we know that any two multiplications on ς˘(B∞ )
compatible with the norm must be isomorphic. But as remarked in §5.7, the exact equality of the
two products at hand is equivalent to the statement that ς˘ is multiplicative modulo Γ0 , whence the
result.
(0)
Recall that it is the normally ordered elements of TH which are constant under the inductive
system ι•,• . The partial inverse ς(m) : ς (m) (Bm ) → Bm Q
of the covariant symbol map ς (m) gives a
normally ordered “quantization” of B∞ , and the image of m∈N0 ς(m) is the normally ordered part of
(0)
the Toeplitz core, which is all of TH . In contrast, the projective limit B ∞ contains (as an operator
(0)
space) only the anti-normally ordered elements in TH , so “Toeplitz quantization” gives the anti−1
normal ordering. Lemma 2.15 shows that π (π(˘
ς (B∞ ))) = ς˘(B∞ ) + Γ0 nevertheless gives all of
(0)
TH .
37
5.10 Strict quantization
Some authors ([Hawk3], [Rie1], [Sain]) do not require commutativity of the “classical limit algebra”
in an axiomatic approach to “strict quantization”. Adapting such a definition, we can show that
what we have done here is a strict quantization.
Let 0 B∞ denote the ∗-algebra generated by the ς (m) (A)’s for all A ∈ B(Hm ) and all m ∈ N0 ; thus
0
B∞ is a dense ∗-subalgebra (the “algebraic part”) of the inductive-limit C ∗ -algebra B∞ .
Definition 5.31. The Berezin product on 0 B∞ is defined for all f, g ∈ 0 B∞ by
(m)
f ⋆ g := ς (m) ς(m) (f )ς(m) (g) ,
where ς(m) : ς (m) (Bm ) → Bm denotes the partial inverse of ς (m) . The Poisson bracket on 0 B∞ is
defined by (cf. [Hawk1, §D.1])
(m)
(m)
m
{f, g} := lim √ (f ⋆ g − g ⋆ f ).
m→∞
−1
(5.29)
Of course we do not expect (5.29) to be a Poisson bracket in the ordinary sense, and {·, ·} is not
likely to be interesting unless H• is commutative.
Corollary 5.32. The sequence (B• , ς˘(•) ) = (Bm , ς˘(m) )m∈N0 is a strict quantization of 0 B∞ in the
sense that each ς˘(m) is surjective and, for all f, g ∈ 0 B∞ ,
lim k˘
ς (m) (f g) − ς˘(m) (f )˘
ς (m) (g)k = 0,
(5.30)
lim k˘
ς (m) (f )k = kf k,
(5.31)
lim km−1 [˘
ς (m) (f ), ς˘(m) (g)] − {f, g}]k = 0.
(5.32)
m→∞
m→∞
m→∞
(0)
Proof. We have seen that π ◦ ς˘ : B∞ → TH /Γ0 is injective, so
ς˘(f )˘
ς (g) − ς˘(f g) ∈ Γ0 ,
which is equivalent to von Neumann’s condition (5.30). Rieffel’s condition (5.31) coincides with the
definition of the norm on B ∞ . Similarly, the Dirac condition (5.32) is tautology in view of our
definition of {·, ·} in (5.29).
We have seen that a subproduct system H• comes with a sequence B• = (Bm )m∈N0 of finite(0)
dimensional algebras Bm := B(Hm ), to which we can add B∞ ∼
= OH , and two sequences ς (•) =
(•)
(m)
(m)
ς )m∈N0 of positive unital maps
(ς )m∈N0 and ς˘ = (˘
ς (m) : Bm → B∞ ,
ς˘(m) : B∞ → Bm
such that ς (m) ◦ ς˘(m) converges to the identity map on B∞ . As in [Sain, Prop. 2.2] we can associate
to this data a continuous field of C ∗ -algebras, making explicit the assertion in Remark 4.5.
(0)
Corollary 5.33. The C ∗ -algebra TH can be identified with the space of continuous sections of the
continuous field N0 ∪ {∞} ∋ m → Bm , i.e.
o
n
Y
(0)
Bm X∞ = lim ς (m) (Xm ) .
TH ∼
= (Xm )m∈N0 ∪{∞} ∈
mN0 ∪{∞}
m→∞
Proof. We have seen that ς : ς˘(B∞ ) → B∞ can be obtained as
ς(X) = lim ς (m) (Xpm ).
m→∞
∗
Thus, the C -algebra of continuous sections of B• consists of the image of ς˘(B∞ ) under ς together
with the sequences (Xm )m∈N0 ∪{∞} such that X∞ = 0. Hence the result follows from the facts that
(0)
ς˘(B∞ ) + Γ0 = TH
and that ς is an isomorphism.
38
5.11 OH assembled from projective limits
We now define modules over the projective limit B ∞ .
(k)
Recall that we defined in §4.3 an inductive system ιm,l : B(Hm , Hm+k ) → B(Hl , Hl+k ). Define the
(k)
(k)
adjoint l,m : B(Hl+k , Hl ) → B(Hm+k , Hm ) of ιm,l by the property that
(k)
(k)
φm l,m (X)Y = φl Xιm,l (Y )
for all X ∈ B(Hl+k , Hl ) and all Y ∈ B(Hm , Hm+k ). We deduce that
Tr(Qm ) X
(k)
l,m (X) =
∀X ∈ B(Hl+k , Hl ),
(Q⊗m )r,r Rr∗ XRr Hm ,
Tr(Ql )
|r|=l−m
(k)
and that the opertors on HN which are constant with respect to the system •,• are precisely those
(l)
of the form ς˘k (f ) = (˘
ςk (f ))l∈N0 for some f ∈ B∞ , where
X
(l)
−1
(Q⊗m )j,j
ωQ (Zj Zk∗ f )Sj Sk∗ H .
ς˘k (f ) = Tr(Ql )
l
|j|=l,|k|=l+k
Define E(k) to be the set of ς˘k (f )’s for all f ∈ B∞ . Then E(k) is an operator system.
(k)
Let Γ0
l → ∞.
be the vector space of sequences of operators in B(Hl+k , Hl ) which converge to zero as
(k)
(−k)
Proposition 5.34. The vector space E(k) + Γ0 coincides with the subspace TH
of the Toeplitz
(k)
∗
∞
algebra. Hence E(k) is a module over the C -algebra B , the module action taken modulo Γ0 . In
∼ O(−k) .
fact, E(k) is isomorphic as a Hilbert bimodule to E (−k) =
H
Proof. The first statements are proven in the same way as for k = 0. For the last assertion, note
that the algebras of compact module operators KB∞ (E (−k) ) and KB∞ (E(k) ) are isomorphic, namely
to B∞ ∼
= B ∞ . Hence the modules E(k) and E (−k) are isomorphic [Frank1].
5.12 Commutative case
If H• is associated with the radical homogeneous ideal which defines a projective variety M ⊂ CPn−1 ,
Assumption 5.4 is satisfies if and only if M is a coadjoint orbit G/K under some Lie group G, and
then one may take Q = p1 (the identity operator on H1 ). However, there are many commutative and
noncommutative subproduct systems H• for which a projective limit can be constructed in a weaker
sense.
Let us focus on the case when commutative H• . Then we know that the right shifts R1 , . . . , Rn
coincide with the left shifts S1 , . . . , Sn , and from our discussion about inductive limits we know
that [Sj∗ , Sk ] is compact for all j, k = 1, . . . , n. Therefore, the maps l,m , and hence the Toeplitz
maps ς˘(m) , become asymptotically multiplicative. That is, ς˘ is a homomorphism modulo compacts.
Furthermore, the asymptotic unitality of the maps m+1,m allows us to define a norm on B ∞ by the
same formula as before. The map ς˘ then becomes isometric, and it is adjoint to the map ς which
(0)
identifies TH /Γ0 with B∞ = C 0 (M ). One easily sees that no Toeplitz operator is in Γ0 and that
ς ◦ ς˘ = id .
The main difference from the special case M = G/K is that for general M one has
ς˘(1) 6= 1
P
and the passage from S to a spherical isometry is more involved since nk=1 Sk∗ Sk pm is not just a
scalar multiple of pm for each m.
As we have seen (recall Proposition 3.6), from the version of Berezin quantization with prequantum condition one obtains a strict quantization of C(M ). With projectively induced quantization we
obtain from Corollary 5.32 a strict quantization of C(M ), and we do not require M to be smooth.
39
Corollary 5.35. For any projective variety M , the sequence (Bm , ς˘(m) )m∈N0 is a strict quantization
of the dense ∗-subalgebra of C(M ) generated by the ς (m) (Bm )’s.
Corollary 5.35 was inspired [Hawk2], where the Toeplitz operators where defined geometrically in
the case of a prequantum quantization. When M is smooth we see from the proof of [Hawk2, Lemma
4.2] that the limit state on B∞ = C(M ) is faithful and hence C(M ) is also isomorphic to the subset
ς˘(B∞ ) of the projective limit B ∞ with the multiplication taken modulo compacts. For non-smooth
M we do not know if the limit state on B∞ is faithful.
6 Application to compact matrix quantum groups
6.1 Compact matrix quantum groups
For the theory of compact quantum groups we refer to [KlS], [MaVD], [Timm1]. We shall restrict
attention to compact matrix quantum groups, defined as follows.
Definition 6.1 ([Wang3], [Wor1]). A compact matrix quantum group G is defined by a C ∗ algebra C(G) generated by the entries uj,k of a single unitary matrix u ∈ Mn (C) ⊗ C(G) (for some
n ∈ N) such that the map ∆ : C(G) → C(G) ⊗ C(G) defined by
∆(uj,k ) :=
n
X
r=1
uj,r ⊗ ur,k
is a ∗-homomorphism, and such that the transpose ut is invertible.
We refer to the generating matrix u as the defining representation of the “group” G.
The Haar state on C(G) (or the Haar measure on G) is the unique state on C(G) which is
left G-invariant, in the sense that
(id ⊗h) ◦ ∆(f ) = h(f )1.
The Haar state is always faithful on the ∗-algebra generated by the uj,k ’s but not necessarily so on
the norm closure C(G). There is a canonical construction of a “reduced version” of G, which is a
compact quantum group with faithful Haar state [BMT1, §2] and has the same dense Hopf ∗-algebra.
We shall always work with the reduced version or, what amounts to the same thing, assume that h
is faithful on all of C(G). Then h is a KMS state [Wor4].
6.1.1
Representations and actions
Definition 6.2. A representation of G is a corepresentation of C(G), i.e. an invertible element
v ∈ B(Hv ) ⊗ C(G), for some Hilbert space Hv , satisfying (in leg-numbering notation)
(id ⊗∆)(v) = v13 v23
as elements of B(Hv ) ⊗ B(Hv ) ⊗ C(G). A representation v is irreducible if the set
HomG (v, v) := {T ∈ B(Hv )|(T ⊗ 1)v = v(T ⊗ 1)}
is trivial.
Definition 6.3. Two representations v ∈ B(Hv ) ⊗ C(G) and w ∈ B(Hw ) ⊗ C(G) are equivalent,
denoted v ≃ w, if there is a unitary U : Hv → Hw such that
(U ⊗ 1)v = w(U ⊗ 1)
(in particular, this requires dim Hv = dim Hw ). We denote by Irrep G the (countable) set of equivalence classes of irreducible representations of G. We choose a representative u(λ) ∈ B(Hλ ) ⊗ C(G)
for each λ ∈ Ĝ.
40
Definition 6.4. The tensor product of two representations u ∈ B(H) ⊗ C(G) and v ∈ B(K) ⊗ C(G)
is the representation
u ⊗ v := u13 v23 ∈ B(H ⊗ K) ⊗ C(G).
In particular, u⊗m = u1,m+1 · · · um,m+1 is the matrix whose entries in the product basis for H⊗m is
given by
uj,k = uj1 ,k1 · · · ujm ,km .
Now let us explain the motivation for the invertible operator Q ∈ B(H) that we incorporated in
the Berezin quantization (recall §5.2).
It is a crucial consequence of the axioms of compact matrix quantum groups that for any finitedimensional representation v ∈ B(Hv ) ⊗ C(G) of G, one can find an invertible matrix Fv ∈ B(Hv )
such that
v̄ := (Fv ⊗ 1)v c (Fv−1 ⊗ 1)
(6.1)
is unitary, where v c = (v t )∗ is the matrix whose coefficients are the adjoints of those of v. The
equivalence class of v̄ is the conjugate of the equivalence class of v (we shall also say that v̄ is a
conjugate of v). The matrix Fv in (6.1) is usually chosen such that Qv := Fv∗ Fv satisfies
Tr(Q−1
v ) = Tr(Qv ) ≡ dimq (v),
and this quantity is the “quantum dimension” of v. We have Qv̄ = (Qtv )−1 . We shall write Qλ :=
Qu(λ) etc. for irreducibles λ ∈ Irrep G and we denote by λ̄ the conjugate of λ.
Every representation of G decomposes completely into a direct sum of irreducibles. Hence, for
each pair of irreps λ, µ ∈ G there are integers mult(ν, λ ⊗ µ) ∈ N0 such that
M
u(λ) ⊗ u(µ) ≃
mult(ν, λ ⊗ µ)u(ν) .
(6.2)
ν∈Irrep G
Definition 6.5. The equations (6.2) dictate the fusion rules of G. The fusion rules are commutative if
mult(ν, λ ⊗ µ) = mult(λ ⊗ µ, ν),
∀λ, µ, ν ∈ Irrep G.
Example 6.6. Compact groups G = G have commutative fusion rules. More generally, q-deformations
of compact Lie groups have commutative fusion rules because the equivalence class of an irreducible
representation is determined by the highest weight of the representation.
The quantum groups in the next two examples are introduced in Definition 6.10 below.
Example 6.7. For any F , the fusion rules of the quantum group Bu (F ) are identical to those of
SU(2); in particular this is true for SUq (2). These fusion rules in fact characterize the Bu (F )’s among
compact quantum groups [Ban3, Théorème 2].
Example 6.8. The fusion rules of Au (Q) are far from commutative, see [Ban4].
Definition 6.9 ([Wang1, Def. 3.1]). A left action of a compact matrix quantum group G on a
C ∗ -algebra B is a unital ∗-homomorphism α : B → B ⊗ C(G) such that
(i) (α ⊗ id) ◦ α = (id ⊗∆) ◦ α,
(ii) (id ⊗ε) ◦ α = id, where ε is the counit on the dense Hopf-∗-subalgebra of C(G), and
(iii) there is a dense ∗-subalgebra 0 B of B such that α(0 B) ⊂ 0 B ⊗ C ∞ (G).
Similarly, a right action of G on B is a unital ∗-homomorphism α : B → C(G) ⊗ B satisfying the
obvious analogues of the properties (i), (ii) and (iii).
If B is a von Neumann algebra then we replace C(G) by its weak closure L∞ (G) in the GNS
representation of the Haar state, and only condition (i) is required in the definition of an action.
Every unitary representation v ∈ B(H) ⊗ C(G) of G induces a left action of G on B(H) given by
Adv : B(H) → B(H) ⊗ L∞ (G),
41
Adv (A) := v(A ⊗ 1)v ∗ .
(6.3)
6.1.2
Universal quantum groups
In the following, for a matrix u with entries in C(G), we write uc for the transpose of u∗ , i.e.
(uc )j,k := u∗j,k where u∗j,k is the adjoint of uj,k in C(G).
Definition 6.10 ([Wang3], [Ban4, Déf. 1]). Let F ∈ GL(n, C) be an invertible matrix and write
Q := F ∗ F . The universal unitary quantum group G = Au (Q) is the compact matrix quantum
group G whose algebra of continuous functions C(G) is generated by the entries of a unitary n × n
matrix u satisfying the relations making (F ⊗ 1)uc (F −1 ⊗ 1) a unitary matrix.
The universal orthogonal quantum group G = Bu (F ) is the compact matrix quantum group
whose algebra C(G) is the quotient of that of Au (Q) by the relation u = (F ⊗ 1)uc (F −1 ⊗ 1).
The prototype example of a Bu (F ) is the quantum SUq (2) group G := SUq (2). In general, Bu (F )
is some kind of higher-dimensional quantum SU(2) group which has no classical counterpart.
Suppose that H and G are compact matrix quantum groups such that C(H) is a quotient of
C(G). If the quotient map π : C(G) → C(H) fulfills (π ⊗ π) ◦ ∆G = ∆H ◦ π, i.e. if π intertwines the
comultiplication of G with that of H, then H is a quantum subgroup G. We have seen that Bu (F )
is a quantum subgroup of Au (Q) when F ∗ F = Q.
The name “universal” is motivated by the following fact, which we should anticipate from (6.1).
Lemma 6.11 ([VaDW]). Any compact matrix quantum group G is a quantum subgroup of Au (Q)
for some Q. If G in addition has a self-conjugate defining representation, then C(G) is a quantum
subgroup of some Bu (F ). We write G ⊂ Au (Q) and G ⊂ Bu (F ) ⊂ Au (Q) for these cases respectively.
Let u be the fundamental representation of G ⊂ Au (Q), with Q ∈ GL(n, C). Then the elements
z1 := uk,1 , . . . , zn := uk,n of the first row of u satisfy the Q-sphere relations
(Q−1 )1,1 1 =
n
X
(Q−1 )r,s zr∗ zs ,
r,s=1
6.1.3
1=
n
X
zs zs∗ .
(6.4)
s=1
The dual discrete quantum group
Let G be a compact matrix quantum group such that the GNS representation C(G) → B(L2 (G)) of
the Haar state is faithful. We shall identify C(G) with its image in B(L2 (G)) and denote by L∞ (G)
the von Neumann algebra generated by C(G) in B(L2 (G)).
In perfect analogy to the case of ordinary compact groups, the C ∗ -algebra C(G) has a Peter–
Weyl decomposition
M
C(G) =
B(Hλ )∗ ,
(6.5)
λ∈Irrep G
2
and the completion L (G) of C(G) in the inner product defined by the Haar state then allows for a
similar decomposition. The comultiplication ∆ on C(G) ⊂ B(L2 (G)) takes the form
∆(f ) = W (f ⊗ 1)W ∗ ,
∀f ∈ C(G)
for a unitary operator W on L2 (G) ⊗ L2 (G) referred to as the multiplicative unitary of G. The
dual of G is then defined via the (multiplier) Hopf C ∗ -algebra
M
c0 (Ĝ) :=
B(Hλ ),
(6.6)
λ∈Irrep G
with the comultiplication given by
ˆ
∆(X)
:= W ∗ (1 ⊗ X)W,
∀X ∈ c0 (Ĝ).
We denote by pλ the identity in B(Hλ ), regarded as an element of c0 (Ĝ). Then the counit ε̂ on c0 (Ĝ)
is characterized by (cf. (2.4))
ε̂(X)p0 = Xp0 ,
42
∀X ∈ c0 (Ĝ),
where 0 ∈ Irrep G is the trivial representation. Every irreducible representation u(λ) of G is obtained
from W by means of
u(λ) = W (pλ ⊗ 1).
The object Ĝ is referred to as a discrete quantum group. In the general theory of “locally
compact quantum groups”, there is a canonical dual quantum group associated also to Ĝ, and this
quantum group is precisely G. In particular, the dual of an ordinary compact group G is a discrete
quantum group, which is an honest group only if G is abelian.
The C ∗ -algebra c0 (Ĝ) is contained in the C ∗ -algebra K of compact operators on L2 (G). The
multiplier algebra of c0 (Ĝ) can be identified with the C ∗ -direct product
Y
cb (Ĝ) :=
B(Hλ ),
λ∈Irrep G
ˆ is a map from c0 (Ĝ) into cb (Ĝ) ⊗ cb (Ĝ). Continuing the analogy with the
and the comultiplication ∆
theory of honest groups, we shall denote by
M
cc (Ĝ) :=
B(Hλ )
λ∈Irrep G
the algebraic sum, and we denote the weak closure of c0 (Ĝ) in B(L2 (G)) by
Y
R(G) = ℓ∞ (Ĝ) := ℓ∞ B(Hλ ).
λ∈Irrep G
This “group-von Neumann algebra” R(G) is contained in the dual C(G)∗ of C(G).
Finally, the algebra of operators on L2 (G) affiliated with R(G) can be identified with the product
Q
b
λ∈Irrep G B(Hλ ), containing all (not necessarily bounded) sequences of elements in the B(Hλ )’s. Of
particular importance is the operator QG := (Qλ )λ∈Irrep G , where Qλ is as in §6.1.1.
From (6.6) we see that the irreducible representations of the C ∗ -algebra c0 (Ĝ) (also referred to
as the irreducible “corepresentations” of Ĝ) are parameterized by λ ∈ Irrep G. In fact, if u(λ) ∈
B(Hλ ) ⊗ C(G) is the irreducible representation of G with label λ (as in §6.1.1) then
πλ (X) := (id ⊗X)(u(λ) )
(6.7)
is the corresponding irreducible corepresentation of Ĝ, where X ∈ c0 (Ĝ) is regarded as a functional
on C(G). In general, if v is a unitary representation of G then (6.7) defines a representation πv of Ĝ
by substituting u(λ) with v. Then the commutant of πv (c0 (Ĝ)) in B(Hv ) equals
πv (c0 (Ĝ))′ = B(Hv )G ,
the fixed-point subalgebra under the G-action Adv (recall (6.3)).
The same equation (6.7) represents any (possibly unbounded) operator X affiliated to R(G) on
B(Hλ ). For each finite-dimensional representation v of G,
πv (QG ) = Qtv ,
where Qv is as in §6.1.1 and Qtv denotes its transpose. In particular, Qtv commutes with the projection
ˆ G ) = QG ⊗ QG , and it gives
onto any irreducible subrepresentation of v. It is well known that ∆(Q
Qv⊗w = Qv ⊗ Qw
for all finite-dimensional representations v and w.
43
6.2 First-row and first-column algebras
Throughout this section, G is a compact matrix quantum group. We denote by (u, H) the defining
representation of G. Thus the C ∗ -algebra C(G) is generated by the matrix coefficients of the unitary
matrix u ∈ B(H) ⊗ C(G). Set n := dim(H) and fix an orthonormal basis e1 , . . . , en of H so that
H∼
= Cn , and let uj,k be the matrix coefficients of u in this basis.
Definition 6.12. The first-row algebra of G is the C ∗ -algebra C(SG ) generated by the first row
z1 := u1,1 , . . . , zn := u1,n . This defines the quantum homogeneous space SG .
There is a Z-grading on C(SG ) obtained by letting the zk ’s have degree 1 while their adjoints
are given degree −1. We write the decomposition into spectral subspaces for the corresponding
U(1)-action as
k·k
M
C(SG ) =
C(SG )(k) .
k∈Z
Definition 6.13. The quantum homogeneous space G/K is defined as the noncommutative manifold
corresponding to the C ∗ -subalgebra of fixed points in C(SG ) for the U(1)-action:
C(G/K) := C(SG )(0) .
It is clear that C(G/K) is generated by the n2 elements {zj∗ zk }nj,k=1 (but this is obviously not a
minimal set of generators).
Example 6.14. If G = G is a compact semisimple Lie group then G/K is a coadjoint orbit and SG is
a principal U(1)-bundle over G/K. Indeed, let (U−1 , H−1 ) be the irreducible unitary representation
of G with highest weight (−1, 0, . . . , 0) and let K be the stabilizer of the complex line spanned by the
highest-weight vector ξ−1 . The action U−1 of G on H−1 is unitary, and hence Hamiltonian for the
symplectic form given by the imaginary part of the inner product on H−1 . The orbit U−1 (G) · [ξ−1 ]
of the line spanned by ξ−1 is then a Hamiltonian G-homogeneous space, and a characterization of
coadjoint orbits shows that G/K ∼
= U−1 (G) · [ξ−1 ] is a coadjoint orbit.
The hyperplane bundle over the projectivization P[H−1 ] restricts to a holomorphic line bundle L
over G/K and we let SG be the total space of the principal U(1)-bundle over G/K associated with
L∗ (cf. §3.3). A basis for the space of holomorphic sections of L generates C(SG ) and, after fixing
a basis e∗1 , . . . , e∗n for H−1 ∼
= Cn , such a basis is provided by the coordinate functions z1 , . . . , zn of
n−1
CP
restricted to G/K. Choosing the basis such that e∗1 ∝ ξ−1 we get G/K = G/K and SG = SG .
For example, if G = SU(n) then K = U(1) × SU(n − 1) and G/K = G/K is complex projective
n-space CPn−1 , whereas SSU(n) = S2n−1 is the unit sphere in Cn .
Example 6.15. The preceding discussion carries over to q-deformations of G, and we get that G/K
is a quantum flag manifold. For G = SUq (n) we obtain quantum projective n-space G/K = CPqn−1 ,
and SG is the q-deformed (2n − 1)-sphere S2n−1
(cf. [DDL1]).
q
Example 6.16 ([BaGo1, Thm. 3.3]). Let G := O∗ (n) be the half-liberated orthogonal group. Then
C(G/K) is in fact commutative and G/K is just the ordinary complex projective n-space CPn−1 ,
C(G/K) ∼
= C(CPn−1 ).
The spectral subspaces C(SG )(k) for the U(1) action on C(SG ) are Hilbert bimodules over the
fixed-point subalgebra C(G/K), where the multiplication in the ambient algebra C(SG ) defines left
and right C(G/K)-valued inner products
hξ|ηiright := ξ ∗ η,
hξ|ηileft := ξη ∗
(6.8)
between elements ξ and η in C(SG )(k) .
Remark 6.17 (Row vs column). We can also consider the C ∗ -algebra generated by the first column
elements wj := uj,1 of u (the “first-column algebra”). Everything proven about the first-row algebra
C(SG ) in this paper has a version where one instead uses the generators of the first-column algebra.
44
Lemma 6.18. The first-row algebra C(SG ) carries an ergodic action of G which contains every
irreducible representation of G with multiplicity one.
Proof. We define a left action C(SG ) → C(SG ) ⊗ C(G) by restriction of the comultiplication. The
Peter–Weyl decomposition (6.5) of C(G) gives the decomposition
M
C(SG ) ∼
Hλ ,
=
λ∈Irrep G
and the comultiplication restricts to the irreducible G-representation u(λ) on each Hλ .
In view of Lemma 6.18 and the above examples, the quantum homogeneous G-space G/K is a
natural generalization of the q-deformed projective spaces (in particular the standard Podleś sphere
S2q = CP1 ). In all cases the unique invariant state under the G-action is the restriction to C(SG ) of
the Haar state on C(G).
6.3 Subproduct systems of G-representations
The subproduct system associated with a compact quantum group G will be a subproduct H• in which
the Hilbert space Hm is contained in the mth tensor power H⊗m of the fundamental representation
H of G.
The idea is based on the observation that if u ∈ Mn (C) ⊗ C(G) is a representation of a compact matrix quantum group G then the first row z1 := u1,1 , . . . , zn := u1,n of u transforms as the
representation u under the “left translation” action λ of G given by restricting the comultiplication
∆:
n
X
zj ⊗ uj,k .
λ(zk ) := ∆(zk ) =
j=1
The constructions below can easily be made more general but we shall always assume that u is
irreducible. In fact we shall, for simplicity and concreteness, from now on assume that u is the
defining representation of G. Let H be the n-dimensional Hilbert space with basis vectors z1 , . . . , zn .
In most cases there are subrepresentations of G contained in the tensor product H ⊗ H. Keeping only
the largest G-invariant subspace H2 of H ⊗ H we obtain another irreducible representation u(2) of G.
Indeed, H2 can be identified with the span of zj zk for j, k = 1, . . . , n, and then u(2) is obtained by
the restricting the comultiplication, just as for u. Continuing like this we obtain a family (Hm )m∈N0
of Hilbert spaces satisfying the subproduct condition (2.1).
Definition 6.19. Let G be a compact matrix quantum group. The subproduct system H• just
described will be referred to as the the G-subproduct system.
If we dropped the requirement that u generates C(G) then Definition 6.19 makes no use of the
fact that C(G) is finitely generated, and hence it works for all compact quantum groups. On the
other hand, if we do not require irreducibility but G is a matrix group, we may always find a selfconjugate unitary finite-dimensional representation u whose coefficients generate C(G). Indeed, if u
generates C(G) then so does u⊕ ū, and the latter is self-conjugate. If u is self-conjugate and generates
C(G), every irreducible representation of G is obtained as Hm for a unique m ∈ N0 . However, for
definiteness we shall always suppose that G is a compact matrix quantum group and that u is the
defining representation, assumed irreducible.
Example 6.20. For G = G a classical compact Lie group, it is well known that the representation
Hλ+µ with dominant weight λ + µ occurs exactly once in the tensor product Hλ ⊗ Hµ ; this is the
“Cartan product” of Hλ and Hµ [East1]. In particular, if H• denotes the G-subproduct system then
Hm+1 is the Cartan product of Hm and H. This subproduct system is commutative, i.e. H• ⊆ H∨• ,
and the associated projective variety mentioned in §3.1 is a coadjoint orbit, isomorphic to G/K for
some closed subgroup K of G.
45
Example 6.21. The SU(n)-subproduct system coincides with the fully symmetric subproduct system
H∨• (Example 2.2).
Example 6.22. The G-subproduct system of the universal quantum group G = Au (Q) with positive
Q ∈ GL(n, C) is the product system H⊗• because each power u⊗m of the defining representation is
irreducible [Ban4].
Example 6.23. For G = Bu (F ) we have self-conjugacy u ≃ ū [Ban3]. So if u is irreducible then, as
mentioned above, every λ ∈ Irrep G occurs as Hλ ∼
= Hm for some m ∈ N0 .
The Hm ’s are all irreducible and usually pairwise inequivalent. The only examples where they
are not all inequivalent are those where C(G) is finite-dimensional.
6.4 Berezin quantization of C(G/K)
6.4.1
Covariant symbols as first-row matrix coefficients
Recall the formula (4.9) for the covariant symbol derived in Proposition 4.19. We now observe
that for any compact matrix quantum groups G, the same formula (with u now being the defining
representation of G) appears when taking matrix coefficients of the representation Ad(u(m) ) ≃ u(m) ⊗
ū(m) on B(Hm ).
Definition 6.24. We say that an element of C(G/K) is normally ordered if all zj ’s occur to the
left of the zk∗ ’s.
Lemma 6.25. Suppose that every element in C(G/K) can be written in normally ordered form.
Then the C ∗ -algebra C(G/K) can be generated by covariant symbols
(m)
⊗m
(m)c
ςG (A) := (Tr ⊗ id) (A ⊗ 1)u(m)c∗ (|e⊗m
(6.9)
1 ihe1 | ⊗ 1)u
for all A ∈ B(Hm ) and all m ∈ N0 .
Proof. Write φA := Tr(A·). Then the elements
(m)
⊗m
(m)c
,
ςG (A) = (φA ⊗ id) u(m)c∗ (|e⊗m
1 ihe1 | ⊗ 1)u
A ∈ B(Hm )
form a set which consists exactly of those matrix coefficients of the representation u(m)c∗ ⊗ u(m)c ≃
u(m) ⊗ ū(m) which are contained in C(G/K).
6.4.2
Intertwining the actions
Let W be the multiplicative unitary of G, i.e. the unitary on L2 (G) ⊗ L2 (G) implementing the
comultiplication ∆ (cf. §6.1.3). Then the restriction of W to L2 (G/K) ⊗ L2 (G) identifies with
u(N) := (u(m) )m∈N0 ∈ cb (Ĝ) ⊗ C(G).
We have an action of G on R(G) given by Ad(W ) (i.e. the same formula as for ∆ but now applied
to elements of a different algebra; of course this action extends to all of B(L2 (G))). The restriction
ˆ to B(L2 (G)) restricts to a
of this G-action to B(HN ) is just Ad(u(N) ). Similarly, the extension of ∆
∞
right Ĝ-action on L (G).
So we have the following actions of G and Ĝ on C(G/K).
Definition 6.26. The left G-action αG : C(G/K) → C(G/K) ⊗ C(G) is defined by
(m)
(m)
αG (ςG (A)) := (ςG
⊗ id)(u(m) (A ⊗ 1)u(m)∗ )
for all A ∈ B(Hm ) and all m ∈ N0 . The right Ĝ-action α̂G : C(G/K) → cb (Ĝ) ⊗ C(G/K) is defined
by
(m)
(m)
α̂G (ςG (A)) := (id ⊗ςG )(u(m)∗ (1 ⊗ A)u(m) ).
46
(m)
Lemma 6.27. The map ςG
is G- and Ĝ-equivariant,
(m)
(m)
⊗ id)(u(m) (A ⊗ 1)u(m)∗ ),
(6.10)
ˆ ◦ ς (m) (A) = (id ⊗ς (m) )(u(m)∗ (1 ⊗ A)u(m) ).
∆
G
G
(6.11)
∆ ◦ ςG (A) = (ςG
G
G
∞
The actions α and α̂ therefore coincide on C(G/K) ⊂ L (G) with the left regular action of G and
the right regular action of Ĝ, respectively.
Proof. Let α(m) (A) := u(m) (A ⊗ 1)u(m)∗ and α̂(m) (A) := u(m)∗ (1 ⊗ A)u(m) . We use the defining
property of a left action,
(α(m) ⊗ id) ◦ α(m) = (id ⊗∆) ◦ α(m) .
Formula (4.9) gives
(m)
(ςG
m
⊗ id)(α(m) (A)) = (Tr ⊗ id ⊗ id) (α(m) ⊗ id) ◦ α(m) (A)(|em
1 ihe1 | ⊗ 1)
m
= (Tr ⊗ id ⊗ id) (id ⊗∆) ◦ α(m) (A)(|em
1 ihe1 | ⊗ 1)
(m)
m
= ∆(Tr ⊗ id) α(m) (A)(|em
1 ihe1 | ⊗ 1) = ∆ ◦ ςG (A).
ˆ ⊗ id) ◦ α̂(m) .
The proof of (6.11) is identical, using (id ⊗α̂(m) ) ◦ α̂(m) = (∆
6.4.3
C(G/K) as an inductive limit
Now we will, for certain compact matrix quantum groups G, realize the first-row algebra C(G/K) as a
projective limit B ∞ . From our previous results we have then obtained C(G/K) as the U(1)-invariant
part of the Cuntz–Pimsner algebra of the G-subproduct system.
Let G ⊂ Au (Q) be a compact matrix quantum group (with faithful Haar measure), with Q ∈
GL(n, C), and let C(G/K) be the U(1)-invariant part of the first-row algebra C(SG ).
Let H• = HG
• be the G-subproduct system and let OH be its Cuntz–Pimsner algebra. In §5 we
defined the Toeplitz quantization ς˘(m) as a map from OH to B ∞ . In that way we could realize an
(0)
isomorphism between OH ∼
= B∞ and the projective limit B ∞ . In this section we shall use the same
strategy but with a map
(m)
ς˘G : C(G/K) → B(Hm ),
(6.12)
(0)
(m)
i.e. we quantize C(G/K) instead of OH . We define (6.12) to be the adjoint of the map ςG appearing
in (6.9). We shall in this way obtain an isomorphism between C(G/K) and B ∞ , and hence, due to
B∞ ∼
= B ∞ , we will arrive at the following result.
Theorem 6.28. Assume that normal ordering is possible in C(G/K). Then there is a C ∗ -isomorphism
between C(G/K) and the U(1)-invariant part of the Cuntz–Pimsner algebra OH of the G-suproduct
system:
(0)
C(G/K) ∼
= OH .
Q (m)
This isomorphism is given by the total Toeplitz map ς˘G = m ς˘G , and it intertwines the ergodic Gand Ĝ-actions as well as the G-invariant states.
Remark 6.29 (Normal ordering). For general G, the map ς˘G maps the normally ordered part of
(0)
C(G/K) onto OH . For instance, normal ordering is possible for G = Bu (F ) but not for Au (Q). In
fact, Au (Q) is our only example where ς˘G is not an isomorphism. Commutative fusion rules implies
normal ordering (recall Definition 6.5).
So let us begin by defining B ∞ to be the projective limit of the system (B(H• ), •,• ), where H• is
the G-subproduct system. Here the operator Q on H which appears in the construction of B ∞ (§5.2)
is taken to be the same as the matrix defining Au (Q) ⊃ G, assuming Q is equal to its transpose.
(m)
Thus, B(Hm ) is equipped with the φm -inner product, where φm = Tr(ρQ ·) is the state defined by
47
(m)
(0)
the density matrix ρQ := Qm / Tr(Qm ). We may therefore regard B ∞ ∩ TH
having trivial intersection with c0 (Ĝ).
as a subset of cb (Ĝ)
Lemma 6.30. The G-subproduct system H• and the matrix Q satisfy Assumption 5.4.
Proof. That Q⊗m preserves the subspace Hm ⊂ H⊗m follows from the fact, mentioned in §6.1.3, that
Q⊗m belongs to the commutant of the G-action on H⊗m .
Let L2 (S) be the GNS Hilbert space of the restriction of the Haar state to the subalgebra C(S) ⊂
C(G). Then the generators z1 , . . . , zn of C(S) are represented ∗-homomorphically as operators on
L2 (S). If we denote these by Z1 , . . . , Zn we thus have (assuming Q1,1 = 1) from (6.4) that
n
X
Qk,k Zk∗ Zk = 1.
k=1
Let H 0 (S) be the closed subspace of L2 (S) spanned by z1 , . . . , zn . Then H 0 (S) is invariant under
Z1 , . . . , Zn . Denote by Tk the restriction of Zk to H 0 (S) for each k ∈ {1, . . . , n}. If P is the orthogonal
projection of L2 (S) onto H 2 (S), we get
n
X
Qk,k Tk∗ Tk = P
n
X
Qk,k Zk∗ Zk
H 2 (S)
= 1.
(6.13)
k=1
k=1
Now the tuple T1 , . . . , Tn is unitarily equivalent to an operator tuple T̃1 , . . . , T̃n on the Fock space
HN satisfying
Tr(Qm+1 ) ∗
T̃j∗ T̃k |Hm =
∀m ∈ N0 , j, k ∈ {1, . . . , n}.
S S k |H m ,
Tr(Qm ) j
where S1 , . . . , Sn are the standard shifts on HN and we used again that Qm is simply the restriction
of Q⊗m to an invariant subspace. Equation (6.13) then says that the maps l,m are unital, i.e. that
ιm,l preserves the Q-traces.
Remark 6.31 (Q-spherical isometries and Q-subnormality). One may say that a tuple of operators
T1 , . . . , Tn satisfying Equation (6.13) is a Q-spherical isometry’. For Q = 1 we obtain the usual
notion of a spherical isometry. The particular Q-spherical isometry T1 , . . . , Tn obtained in the
proof of Lemma 6.30 is moreover Q-subnormal
P in that it is the restriction to an invariant subspace
of a tuple of operators Z1 , . . . , Zn satisfying nk=1 Zk Zk∗ = 1. In the commutative case (where we
must have Q = 1) we know from [Atha3, Prop. 2] that a subnormal operator tuple is a spherical
isometry if and only if it is subnormal and its normal extension has joint spectrum contained in the
unit sphere S2n−1 .
From our general results we have an isomorphism ς˘(B∞ ) ∼
= B∞ which realizes the projective limit
as an inductive limit. We stress again that this isomorphism ς˘ is not the same as the map ς˘G which
we now try to prove is an isomorphism.
The matrix coefficients of the operator ς˘(m) (f ) are of the form
h(zj zk∗ f ),
j, k ∈ F+
n with |j| = m = |k|.
(m)
(6.14)
(m)
Proof of Theorem 6.28. Since the “coefficient map” ςG in (6.9) is injective, its adjoint ς˘G is surjective. As in the case of B∞ , we get that the image of L∞ (G/K) under ς˘G is exactly B ∞ as a
set.
We cannot use the reasoning in the proof of Lemma 5.20 to deduce that ς˘G is an injection of
C(G/K) into the operator system B ∞ . On the other hand, we see directly that if ς˘(f ) = 0 then,
using that the matrix coefficients of ς˘(f ) are given by (6.14) for all m ∈ N, we get that f must be
orthogonal to the whole normally ordered part of C(G/K) ⊂ L2 (G). Since we have assumed that
each f ∈ C(G/K) can be normally ordered and that that the Haar state is faithful, this means that
f = 0.
48
(0)
Moreover, π −1 (˘
ς (B∞ )) is again equal to TH . Namely, the proof in §5.9 carries over completely.
As before we get that ς˘(m) = ∞,m . Since we know that ς˘(B∞ ) is a C ∗ -algebra (using that it is a
quotient of the Toeplitz algebra) with a unique multiplication, we obtain the von Neumann condition,
i.e. ς˘(m) is asymptotically a homomorphism (see Corollary 5.30). Thus ς˘ is an isomorphism for the
projective-limit multiplication.
We also know that ς˘G intertwines the vacuum state ε̂ with the Haar state h restricted to C(G/K).
Composing with the isomorphism ς˘ : B∞ → ς˘(B∞ ) we get that h is intertwined with the limit state
ωQ . Finally, Lemma 6.27 shows that ς˘G is G-Ĝ-equivariant.
6.4.4
C(SG ) as an inductive limit
Corollary 6.32. Let G be a compact matrix quantum group with faithful Haar measure h : C(G) → C
such that normal ordering is possible in C(G/K). Then there is a G-Ĝ-equivariant isomorphism
between the first-row algebra C(SG ) and the Cuntz–Pimsner algebra OH of the G-subproduct system,
C(SG ) ∼
= OH .
In particular, OH carries an ergodic action of G in which each irreducible representation of G occurs
exactly once.
(0)
(k)
Proof. For notation simplicity we identify OH with the inductive limit B∞ and the modules OH
with the modules E (k) .
Since B∞ ∼
= C(G/K), we know that there is a basis e1 , . . . , en for H such that the Q-sphere
condition (6.4) is satisfied by the generators Z1 , . . . , Zn of OH , just as it is for the generators z1 , . . . , zn
−1/2
−1/2
of C(SG ). This says precisely that Z1 , . . . , Zn and Q1,1 Z1 , . . . , Qn,n Zn are standard right and left
tight normalized frames for the B∞ -bimodule E (1) , respectively; for all ξ ∈ E (1) ,
n
X
k=1
n
X
k=1
−1/2
hξ|Zk iright hZk |ξiright = hξ|ξiright ,
−1/2
hξ|Qk,k Zk ileft hQk,k Zk |ξileft =
n
X
k=1
−1 ∗
ξQk,k
Zk Zk ξ ∗ = ξξ ∗ = hξ|ξileft ,
and identically for C(SG )(1) and the zj ’s. If we identify C(G/K) with B∞ , this means that the
projection P (1) ∈ Mn (C) ⊗ C(G/K) which defines the module C(SG )(1) coincides with the projection
which defines the module E (1) . So the modules are the same and the isomorphism C(SG ) ∼
= OH is
clear.
For the G-Ĝ-equivariance, we must first define actions on OH . But since we know that C(SG ) ∼
=
OH we can just specify these action on generators Z1 , . . . , Zn by the same formulas as for C(SG ).
The last statement is due to Lemma 6.18.
6.5 Comparison with Poisson and Martin boundaries
6.5.1
Poisson integral versus total Toeplitz map
Let G be a compact matrix quantum group with commutative fusion rules (see Definition 6.5) and
faithful Haar measure. The Poisson boundary to be discussed here is the one defined in [Iz1], so if
we were phrasing things in terms of random walks (we shall not), there would in the background be
a representation u of G whose coefficients generate C(G) (without any need of the adjoints u∗j,k ).
Izumi defines [Iz1, Lemma 3.8] the Poisson integral to be the unital completely positive map
Θ : L∞ (G) → R(G) given by
Θ(f ) := (id ⊗h)(W ∗ (1 ⊗ f )W ),
(6.15)
49
where W is the fundamental unitary (§6.1.3). Similar to the projective limit B ∞ which is the image
of our total Toeplitz map ς˘, the image of map Θ is an operator system, usually denoted by H ∞ (Ĝ),
which can be made into a von Neumann algebra by replacing the operator multiplication by the
new one. Moreover, Θ is a complete order isomorphism onto its image. A possible definition of the
Poisson boundary of Ĝ is then as the preimage, say L∞ (G/T), of Θ in L∞ (G). We then refer to
the abstract object G/T as the Poisson boundary of Ĝ. The notation G/T is chosen to indicate that
G and Ĝ act ergodically on L∞ (G/T).
The Poisson boundary G/T is defined in terms of a von Neumann algebra. In order to compare
G/T with what we have denoted G/K, note that ς˘ extends to a normal completely positive map
(denoted by the same symbol)
ς˘ : L∞ (G/K) → B ∞
(6.16)
and this is the “first-row” version of the Poisson integral. Using it one can carry out Berezin quantization on the level of von Neumann algebras. Inspiring work here is [INT1].
The Poisson integral (6.15) can be decomposed into components Θλ : L∞ (G) → B(Hλ ) for
λ ∈ Irrep G, and doing so one easily calculates the adjoints Θ∗λ : B(Hλ ) → L∞ (G). Noticing
the similarity to Berezin quantization, [INT1] referred to the composition Θ∗λ ◦ Θλ as the “Berezin
transform”. This terminology is not entirely fortunate because Θ∗λ ◦ Θλ does not coincide with the
usual notion of Berezin transform when G = G is an ordinary group. The issue is that Θ∗λ is obtained
by tracing against the invertible operator Qλ (which is of full rank) instead of a rank-1 projection.
The distinction is the use of “first-row” versus all of G. This distinction persists even if we, as Izumi
does, assume that every irreducible representation of G is contained in some power of u.
It is therefore interesting that the final results (G/K and G/T) are not very different. For SUq (2)
they even coincide. In general, we should view G/K as a (noncommutative) non-maximal flag variety
(prototype example being CPqn−1 ) while G/T is the maximal flag variety (so T is the “maximal
torus”); cf. [Tom1].
The transition between classical and quantum Poisson boundaries is rather involved [NT1]. In
fact, if G = G is an ordinary compact group then the Poisson boundary is trivial: L∞ (G/T) = C1
[Iz1, Cor. 3.9]. In contrast, Berezin quantization carries over in perfect analogy with the commutative
case.
6.5.2
Markov operator
The set H ∞ (Φ) of fixed points of a normal completely positive map Φ on a von Neumann algebra is
an ultraweakly closed operator system which can be made into a von Neumann algebra by replacing
the operator multiplication by the so-called “Choi–Effros multiplication” [Arv10, Thm. 3.1], [Iz4].
The new multiplication on the Poisson boundary H ∞ (Ĝ) mentioned above is just an example
of a Choi–Effros multiplication. The completely positive map on R(G) whose fixed-point set equals
H ∞ (Ĝ) takes the role of Markov operator for the “noncommutative random walk” on Ĝ.
The following can be summarized by saying that with Berezin quantization one ends up with a
random walk on the “dual” of G/K instead of the dual of G. Note however that it works for any
subproduct system H• . Fix thus a subproduct system H• and denote as usual by Γb = Γb (B• ) the
von Neumann-algebraic direct sum of the matrix algebras Bm := B(Hm ).
Definition 6.33. The Markov operator on B• is the unital normal completely positive map
Φ : Γb → Γb defined by
Φ(X• ) := X•−1 = (m,m−1 (Xm ))m∈N .
Proposition 6.34. The set H ∞ (Φ) of Φ-fixed points in Γb (B• ) is equal to the projective limit B ∞ . In
particular, B ∞ is a von Neumann algebra. On the subset ς˘(B∞ ) ⊂ B ∞ , the Choi–Effros multiplication
coincides with the projective-limit multiplication.
Proof. The first statement is clear, so B ∞ is a von Neumann algebra. To prove the last statement
we use the result [Iz3, Cor. 5.2] that the Choi–Effros product of X, Y ∈ H ∞ (Φ) is given by
X ⋄ Y = lim Φr (XY ),
r→∞
50
where XY is the multiplication in B(HN ) and the limit is in the strong operator topology. Now, the
mth component of X ⋄ Y is
(X ⋄ Y )m = lim Φr (XY ) m = lim m+r,m ((XY )m+r )
r→∞
r→∞
= ∞,m lim (XY )m+r
r→∞
= lim ς˘(l) (XY ) m ,
l→∞
where we used that Xm = ∞,m (X) = ς˘(m) (X). This shows that ⋄ is the projective-limit multiplication (5.23) whenever we have convergence in norm. Since norm-convergence holds for X = ς˘(f ) and
Y = ς˘(g) with f, g ∈ B∞ , the claim holds.
6.5.3
Martin boundaries
While the Poisson boundary is a measure-theoretic object defined via a von Neumann algebra, the
Martin boundary is specified in terms of a C ∗ -algebra [NT1]. Its relation to Berezin quantization is
the same ”first-row versus all-of-G“ story as with the Poisson boundary but we shall discuss only a
special case in which G/K agrees with the Martin boundary of Ĝ. The reason for this coincidence is
that the defining representation of the chosen G is self-conjugate and irreducible, so that H• contains
all irreducible representations.
Our approach here via inductive limits was partially inspired by [VVer1], where they construct a
“Martin boundary” of the dual of G for G = Bu (F ) in the same way. Our notation B∞ is chosen to
make comparison with that paper easy. Let F ∈ GL(n, C) such that F̄ F = ±1; this ensures that the
defining representation of G = Bu (F ) is irreducible. By construction, the Martin boundary of Ĝ is
equal to the inductive limit B∞ of the G-subproduct system.
In [VaVe1], another realization of the Martin boundary was accomplished. First define Bu (F, Fq )
to be the universal C ∗ -algebra generated by the entries of a unitary 2 × n matrix Y satisfying
Y = Fq Y c F −1 .
where Fq :=
0
|q|1/2
±|q|−1/2
0
(6.17)
, with q defined by |q + q −1 | = Tr(F ∗ F ) and F F̄ = ±q and we wrote
q = ∓|q|. It is shown in [VaVe1] that the U(1)-action ρ on Bu (F, Fq ) given by
λ 0
ρλ (Y ) :=
Y,
∀λ ∈ C, |λ| = 1
0 λ̄
allows recovering the Martin boundary as the fixed-point algebra Bu (F, Fq )U(1) .
Since our inductive limit B∞ coincides with the Martin boundary of the dual of Bu (F ), we know
(using Theorem 6.28) that Bu (F, Fq )U(1) must coincide with C(G/K). This can be seen directly. Let
z1 , . . . , zn and w1 , . . . , wn be the elements of the first and second row of Y respectively. Then (6.17)
reads
n
X
−1
zk = |q|1/2
ws∗ Fs,k
,
∀k = 1, . . . , n.
(6.18)
s=1
The action ρz is given by
ρλ (zk ) = λzk ,
ρλ (wk ) = λ̄wk ,
so the fixed-point algebra consists of elements of the form zj wk , and their adjoints, as well as zj zk∗
and wj wk∗ . Now (6.18) shows that Bu (F, Fq )U(1) = C(G/K), as asserted.
Note however that Bu (F, Fq ) is not at all the same as the first-row algebra C(SG ).
Thus we have one example where the Martin boundary of a discrete quantum group identifies
with the object G/K defined in this paper. Also, let q ∈ (0, 1]. Then for the G = SUq (2)-subproduct
system H• we have an equivariant isomorphism
(0)
OH ∼
= C(S2q ),
51
where G/K = S2q is the Podleś sphere. For q < 1, this shows again that our inductive limit B∞
coincides with the C ∗ -algebra referred to as the Martin boundary of the dual of SUq (2) in [VVer1]
(because the same boundary is known to be the Podleś sphere [NT1]). For q = 1, we have agreement
with Biane’s Martin boundary of the dual of SU(2) [Biane1].
7 Concluding remarks
We have seen that the inductive limit B∞ associated with a subproduct system H• is a sensible
generalization of the C ∗ -algebra C(M ) of continuous functions on a quantizable Kähler manifold,
in the case the Kähler structure is projectively induced (so M is embeddable as a submanifold of
projective space). We may therefore write
C(M) := B∞ ,
and say that M is the noncommutative projective variety associated to H• . We may also refer
to elements of
C(SM ) := OH .
as functions on the total space SM of a noncommutative circle bundle over M. Thus the notation
M := G/K and SM := SG would be consistent with that in §6 when H• is the G-subproduct system.
In [An5] we refer to M as the “dequantization manifold”.
By defining C(M) to be equal (and not just isomorphic) to the inductive limit B∞ , the noncommutative space M comes with more structure than just its topology. Namely, if M = M is
commutative then the inductive system gives an embedding into projective space CPn−1 and, if M
is non-singular, endows M with a complex-analytic (in particular smooth) structure, a polarization
L (choice of ample line bundle), an inner product on H 0 (M ; L) (the one we started with, making
H 0 (M ; L) into the Hilbert space H), a Hermitian metric on L (the one defining the ∗-structure on
OH ; this is just the Fubini–Study metric associated with the inner product on H) and a volume
form on M (viz. the limit state, which need not be the same as Fubini–Study volume form). As we
have seen, these structures have perfect generalizations to the noncommutative setting. Note that
the quantum homogeneous spaces G/K are “balanced” in the sense that the limit state on C(G/K)
coincides with the state induced by the Haar state on C(G).
The covariant symbols ς (m) (A) and the Toeplitz operators ς˘(m) (f ) can be expressed in terms of
the projections P (m) ∈ B(Hm ) ⊗ C(M) which define the modules E (m) . In this way one generalizes
the Rawsnely coherent-state projections in [RCG1], and in particular the coherent states in [Per1].
We will use this when we discuss the (fuzzy) geometry of M in another paper.
There are also projections P (−m) and maps ς (−m) , ς˘(−m) etc. associated with the modules E (−m) .
In this way one quantizes instead the anti-normal part of C(G/K).
8 References
11-124 (2000).
[An4] Andersson A. Detailed balance as a quantum-group symmetry of Kraus operators. arXiv:1506.00411
(2015).
[An5] Andersson A. Dequantization via quantum channels. Lett. Math. Phys. Vol 106, Issue 10, pp.
1397-1414 (2016).
[ArMa1] Ara P, Mathieu M. Local multipliers of C ∗ -algebras. Springer (2003).
[ArLo1] Arezzo C, Loi A. Quantization of Kähler manifolds and the asymptotic expansion of Tian–
Yau–Zelditch. J. Geom. Phys. Vol 47, Issue1, pp. 87-99 (2003).
[ArZh1] Artin M, Zhang JJ. Noncommutative projective schemes. Adv. in Math. Vol 109, Issue 2, pp.
228-287 (1994).
52
[Arv6c] Arveson W. Subalgebras of C ∗ -algebras III: Multivariable operator theory. Acta Math. Vol
181, pp. 159-228 (1998).
[Arv8] Arveson W. p-Summable commutators in dimension d. J. Operator Theory. Vol 54, pp. 101-117
(2005).
[Arv9] Arveson W. Quotients of standard Hilbert modules. Trans. Amer. Math. Soc. Vol 359, pp.
6027-6055 (2007).
[Arv10] Arveson W. Noncommutative Poisson boundaries. Unpublished. Available at http://math.berkeley.edu/ arve
son/Dvi/290F04/22Sept.pdf (2004).
[Atha3] Athavale A. On the intertwining of joint isometries. J. Oper. Theory. Vol 23, pp. 339-350
(1990).
[BDLMC] Balachandran AP, Dolan BP, Lee JH, Martin X, O’Connor D. Fuzzy complex projective
spaces and their star-products. J. Geom. Phys. Vol 43, pp. 184 (2002).
[Ban3] Banica T. Théorie des représentations du groupe quantique compact libre O(n). C. R. Acad.
Sci. Paris Sér. I Math. Vol 322, pp. 241-244 (1996).
[Ban4] Banica T. Le groupe quantique compact libre U(n). Comm. Math. Phys. Vol 190, pp. 143–172
(1997).
[BaGo1] Banica T, Goswami D. Quantum isometries and noncommutative spheres. Comm. Math.
Phys. Vol 298, Issue 2, pp. 343-356 (2010).
[BMT1] Bédos E, Murphy GJ, Tuset L. Co-amenability of compact quantum groups. J. Geom. Phys.
Vol 40, Issue 2, pp. 129-153 (2001).
[BeSl1] Berceanu S, Schlichenmaier M. Coherent state embeddings, polar divisors and Cauchy formulas.
J. Geom. Phys. Vol, Issue 34, pp. 336-358 (2000).
[Bere1] Berezin FA. Quantization. Math. USSR-Izv. Vol 38, pp. 1116-1175 (1974).
[Bere2] Berezin FA. General concept of quantization. Comm. Math. Phys. Vol 40, pp. 153-174 (1995).
[Bere3] Berezin FA. Covariant and contravariant symbols of operators. Math. USSR-Izv. Vol 6, Issue
5, p. 1117 (1972).
[Bere4] Berezin FA. Some remarks about the associative envelope of a Lie algebra. Funct. Anal. Appl.
Vol 1, pp. 91-102 (1967).
[Biane1] Biane P. Introduction to random walks on noncommutative spaces. In: Quantum Potential
Theory. pp. 61-116. Springer (2008).
[Bied1] Biedenharn LC. The quantum group SUq (2) and a q-analogue of the boson operators. J. Phys.
A: Math. Gen. Vol 22, pp. L873-L878 (1989).
[Blac1] Blackadar B. Operator algebras: theory of C ∗ -algebras and von Neumann algebras. Springer
(2006).
[BlKi1] Blackadar B, Kirchberg E. Generalized inductive limits of finite-dimensional C ∗ -algebras. Math.
Ann. Vol 307, pp. 343-380 (1997).
[BMS] Bordemann M, Meinrenken E, Schlichenmaier M. Toeplitz quantization of Kähler manifolds and
gl(n), n → ∞ limits. Comm. Math. Phys. Vol 165, Issue 2, pp. 281–296 (1994).
[BHSS] Bordemann M, Hoppe J, Schaller P, Schlichenmaier M. gl(∞) and geometric quantization.
Comm. Math. Phys. Vol 138, Issue 2, pp. 209-244 (1991).
[BLY1] Bourguignon J-P, Li P, Yau ST. Upper bound for the first eigenvalue of algebraic submanifolds.
Comment. Math. Helv. Vol 69, pp. 199-207 (1994).
[Brow1] Brown NP. Invariant means and finite representation theory of C ∗ -algebras. Mem. Amer.
Math. Soc. Vol 184, Issue 865 (2006).
[BrOz1] Brown NP, Ozawa N. C ∗ -algebras and finite-dimensional approximations. Amer. Math. Soc.
(2008).
[CGR1] Cahen M, Gutt S, Rawnsley J. Quantization of Kähler manifolds I: geometric interpretation of
Berezin’s quantization. J. Geom. Phys. Vol 7, Issue 1 (1990).
53
[CGR2] Cahen M, Gutt S, Rawnsley J. Quantization of Kähler manifolds II. Trans. Amer. Math. Vol
337, Issue 1 (1993).
[DDL1] D’Andrea F, Dabrowski L, Landi G. The noncommutative geometry of the quantum projective
plane. Reviews in Mathematical Physics. Vol 20, Issue 08, pp. 979-1006 (2008).
[DRS1] Davidson KR, Ramsey C, Shalit OM. The isomorphism problem for some universal operator
algebras. Adv. Math. Vol 228, Issue 1, pp. 167-218 (2011).
[DRS2] Davidson KR, Ramsey C, Shalit OM. Operator algebras for analytic varieties. Trans. Amer.
Math. Soc. Vol 367, Issue 2, pp. 1121-1150 (2015).
[Don1] Donaldson SK. Scalar curvature and projective embeddings, I J. Differential Geom. Vol 59, pp.
479-522 (2001).
[Don2] Donaldson SK. Scalar curvature and projective embeddings, II. Quart. J. Math. Vol 56, pp.
345-356 (2005).
[Don3] Donaldson SK. Some numerical results in complex differential geometry. Pure Appl. Math. Q.
Vol 5, Issue 2, pp. 571-618 (2009).
[Don9] Donaldson SK. Scalar curvature and stability of toric varieties. J. Differential Geom. Vol 62,
pp. 289-349 (2002).
[Don12] Donaldson SK. Lower bounds on the Calabi functional. J. Differential Geom. Vol 70, pp.
453-472 (2005).
[Doug4] Douglas RG. Essentially reductive Hilbert modules. J. Operator Theory. Vol 55, Issue 1, pp.
117-133 (2006).
[DoWa2] Douglas RG, Wang K. Geometric Arveson–Douglas conjecture and holomorphic extension.
arXiv:1511.00782v1 (2015).
[DoWa4] Douglas RG, Wang K. Essential normality of cyclic submodule generated by any polynomial.
J. Funct. Anal. Vol 261, pp. 3155-3180 (2011).
[East1] Eastwood M. The Cartan product. Bull. Belg. Math. Soc. Simon Stevin. Vol. 11, Issue 5, pp.
641-651 (2005).
[EnEs1] Engliš M, Eschmeier J. Geometric Arveson–Douglas conjecture. Adv. Math. Vol 274, pp.
606-630 (2015).
[Esch1] Eschmeier J. Essential normality of homogeneous submodules. Integr. Equ. Oper. Theory.
Vol 69, Issue 2, pp. 171-182 (2011).
[Ev1] Evans DE. On On . Publ. RIMS, Kyoto Univ. Vol 16, pp. 915-927 (1980).
[Frank1] Frank M. Isomorphisms of Hilbert C ∗ -modules and*-isomorphisms of related operator C ∗ algebras. Math. Scand. Vol 80, pp. 313-319 (1997).
[Hawk1] Hawkins E. Quantization of equivariant vector bundles. Comm. Math. Phys. Vol 202, Issue
3, pp. 517-546 (1999).
[Hawk2] Hawkins E. Geometric quantization of vector bundles and the correspondence with deformation
quantization. Comm. Math. Phys. Vol 215, Issue 2, pp. 409-432 (2000).
[Hawk3] Hawkins E. An obstruction to quantization of the sphere. Comm. Math. Phys. Vol 283, Issue
3, pp. 675-699 (2008).
[Huy] Huybrechts D. Complex geometry. An introduction. Springer (2005).
[Iz1] Izumi M. Non-commutative Poisson boundaries and compact quantum group actions. Adv. Math.
Vol 169, Issue 1, pp. 1-57 (2002).
[Iz2] Izumi M. Non-commutative Markov operators arising from subfactors. Operator algebras and
applications. Adv. Stud. Pure Math. Vol 38, pp. 201-217 (2004).
[Iz3] Izumi M. E0 -semigroups: around and beyond Arveson’s work. J. Operator Theory. Vol 68, Issue
2, pp. 335-363 (2012).
[Iz4] Izumi M. Non-commutative Poisson boundaries, Discrete geometric analysis. Contemp. Math.
Vol 347, pp. 69-81 (2004).
54
[INT1] Izumi M, Neshveyev S, Tuset L. Poisson boundary of the dual of SUq (n). Comm. Math. Phys.
Vol 262, Issue 2, pp. 505-531 (2006).
[KaSh1] Kakariadis E, Shalit O. On operator algebras associated with monomial ideals in noncommuting
variables. arXiv:1501.06495 (2015).
[KaSc1] Karabegov AV, Schlichenmaier M. Identification of Berezin–Toeplitz deformation quantization.
J. Reine Angew. Math. Issue 540, p.49-76 (2001).
[Kenn1] Kennedy M. Essential normality and the decomposability of homogeneous submodules. Trans.
Amer. Math. Socl. Vol 367, Issue 1, pp. 293-311 (2015).
[KeSh1] Kennedy M, Shalit OM. Essential normality and the decomposability of algebraic varieties.
New York J. Math. Vol 18, pp. 877-890 (2012).
[KeSh2] Kennedy M, Shalit OM. Essential normality, essential norms and hyperrigidity. New York J.
Math. Vol 18, pp. 877-890 (2012).
[KlS] Klimyk AU, Schmüdgen K. Quantum groups and their representations. Vol. 552, Springer, Berlin
(1997).
[Lan1] Landsman NP. Mathematical topics between classical and quantum mechanics. Springer (1998).
[Lan2] Landsman NP. Strict quantization of coadjoint orbits. J. Math. Phys. Vol 39, Issue 12, pp.
6372-6383 (1998).
[LMS1] Lazaroiu CI, McNamee D, Sämann C. Generalized Berezin quantization, Bergman metrics and
fuzzy Laplacians. J. High Energy Phys. Issue 9, pp. 1-59 (2008).
[LuTe1] Lübke M, Teleman A. The Kobayashi–Hitchin correspondence. World Scienlific Publishing
(1995).
[MaVD] Maes A, Van Daele A. Notes on compact quantum groups. arXiv: math/9803122 (1998).
[McTr1] McCullough S, Trent TT. Invariant subspaces and Nevanlinna–Pick kernels. J. Funct. Anal.
Vol 178, Issue 1, pp. 226-249 (2000).
[NT1] Neshveyev S, Tuset L. The Martin boundary of a discrete quantum group, J. Reine Angew.
Math. Vol 568, pp. 23-70 (2004).
[Per1] Perelomov AM. Generalized coherent states and applications. Springer (1986).
[Pims1] Pimsner MV. A class of C ∗ -algebras generalizing both Cuntz-Krieger algebras and crossed
products by Z. Fields Inst. Commun. Vol 12, pp. 189–212 (1997).
[Pop1] Popescu G. Operator theory on noncommutative varieties. Indiana Univ. Math. J. Vol 56, Issue
2, pp. 389-442 (2006).
[Pop2] Popescu G. Operator theory on noncommutative varieties II. Proc. Amer. Math. Soc. Vol 135,
Issue 7, pp. 2151-2164 (2007).
[Raj1] Rajeev SG. New classical limits of quantum theories. In: Infinite Dimensional Groups and
Manifolds. Vol 5, 213 (2004).
[RCG1] Rawnsley J, Cahen M, Gutt S. Quantization of Kähler manifolds I: geometric interpretation of
Berezin’s quantization. J. Geom. Phys. Vol 7, Issue 1, pp. 45-62 (1990).
[RCG2] Rawnsley J, Cahen M, Gutt S. Quantization of Kähler manifolds II. Trans. Amer. Math. Soc.
Vol 337, Issue 1 (1993).
[Rie1] Rieffel M. Deformation quantization for actions of Rd . Mem. Amer. Math. Soc. Vol 506 (1993).
[Rie2] Rieffel M. Matrix algebras converge to the sphere for quantum Gromov–Hausdorff distance.
Mem. Amer. Math. Soc. Vol 168, Issue 796, pp. 67-91 (2004).
[RoSt1] Rørdam M, Størmer E (Eds.). Classification of nuclear C ∗ -algebras. Entropy in operator
algebras. Vol 7. Springer (2002).
[Sain] Sain J. Berezin quantization from ergodic actions of compact quantum groups, and quantum
Gromov–Hausdorff distance. PhD thesis, University of California (2009).
[Schl1] Schlichenmaier M. Berezin-Toeplitz quantization for compact Kähler manifolds. A review of
results. Adv. Math. Phys. Article ID 927280, 38 pages (2010).
55
[Schl2] Schlichenmaier M. Singular projective varieties and quantization. In: Quantization of Singular
Symplectic Quotients. pp. 259-282 (2001).
[Seg1] Segal G. Lectures on Lie groups. In: Lectures on Lie groups and Lie algebras. Cambridge
University Press (1995).
[Serre1] Serre JP. Géométrie algébrique et géométrie analytique. Ann. Inst. Fourier, Grenoble. Vol 6,
pp. 1-42 (1956).
[Serre2] Serre JP. Faisceaux algébriques cohérents. Ann. Math. Vol 61, Issue 2, pp. 197-278 (1955).
[Sha3] Shalit OM. Stable polynomial division and essential normality of graded Hilbert modules. J.
London Math. Soc. Vol 83, Issue 2, pp. 273-289 (2011).
[ShSo1] Shalit OM, Solel B. Subproduct systems. Doc. Math. Vol 14, pp. 801-868 (2009).
[Sten1] Stenstrom B. Rings of quotients – An introduction to methods of ring theory. Springer (1975).
[Tian1] Tian G. Kähler–Einstein metrics with positive scalar curvature. Invent. Math. Vol 137, pp.
1-37 (1995).
[Tian2] Tian G. Canonical metrics in Kähler geometry. Birkhauser (2000).
[Timm1] Timmermann T. An invitation to quantum groups and duality: from Hopf algebras to multiplicative unitaries and beyond. Eur. Math. Soc. (2008).
[Tom1] Tomatsu R. A characterization of right coideals of quotient type and its application to classification of Poisson Boundaries. Comm. Math. Phys. Vol 275, Issue 1, pp 271-296 (2007).
[UnUp1] Unterberger A, Upmeier H. The Berezin transform and invariant differential operators. Comm.
Math. Phys. Vol 164, pp. 563-597 (1994).
[VaVe1] Vaes S, Vennet N. Identification of the Poisson and Martin boundaries of orthogonal discrete
quantum groups. J. Inst. Math. Jussieu. Issue 7, pp. 391-412 (2008).
[VVer1] Vaes S, Vergnioux R. The boundary of universal discrete quantum groups, exactness and factoriality. Duke Math. J. Vol 140, Issue 1, pp. 35-84 (2006).
[VaDW] Van Daele A, Wang SZ. Universal quantum groups. Int. J. Math. Vol 7, Issue 2, pp. 255-264
(1996).
[Vas1] Vasselli E. Continuous fields of C ∗ -algebras arising from extensions of tensor C ∗ -categories. J.
Funct. Anal. Vol 199, pp. 122-152 (2003).
[Vas3] Vasselli E. The C ∗ -algebra of a vector bundle and fields of Cuntz algebras. J. Funct. Anal. 222,
Issue 2, pp. 491-502 (2005).
[Vis2] Viselter A. Cuntz–Pimsner algebras for subproduct systems. Int. J. Math. Vol 23, Issue 8, p.
1250081 (2012).
[Wang1] Wang SZ. Ergodic actions of universal quantum groups on operator algebras. Comm. Math.
Phys. Vol 203, pp. 481-498 (1999).
[Wang3] Wang SZ. Structure and isomorphism classification of compact quantum groups Au (q) and
Bu (q). arXiv:math/9807095v2 (2000).
[Wa1] Wang X. Balance point and stability of vector bundles over a projective manifold. Math. Res.
Lett. Vol 9, Issues 2-3, pp. 393-411 (2002).
[Wor1] Woronowicz SL. Compact matrix pseudogroups. Comm. Math. Phys. Vol 111, pp. 613-665
(1987).
[Wor3] Woronowicz SL. Tannaka–Krein duality for compact matrix pseudogroups. Twisted SU(n)
groups, Inv. Math. Vol 93, pp. 35-76 (1988).
[Wor4] Woronowicz SL. Compact quantum groups. Symétries quantiques, Les Houches 1995, pp. 845884 (1998).
[Yau1] Yau S-T. On the Ricci curvature of a compact Kähler manifold and the complex Monge–Amère
equation, I. Commun. Pure Appl. Math. Vol 31, pp. 339-411 (1978).
[Zeld1] Zelditch S. Szegö kernels and a theorem of Tian. Int. Math. Res. Not. Issue 6, pp. 317-331
(1998).
56
| 4math.GR
|
arXiv:1801.06733v2 [cs.DS] 29 Jan 2018
Probabilistic Tools for the Analysis of
Randomized Optimization Heuristics
Benjamin Doerr
Laboratoire d’Informatique (LIX)
École Polytechnique
Palaiseau
France
January 31, 2018
Abstract
This chapter collects several probabilistic tools that proved to be useful in the analysis of randomized search heuristics. This includes classic
material like Markov, Chebyshev and Chernoff inequalities, but also lesser
known topics like stochastic domination and coupling or Chernoff bounds
for geometrically distributed random variables and for negatively correlated
random variables. Almost all of the results presented here have appeared
previously, some, however, only in recent conference publications. While the
focus is on collecting tools for the analysis of randomized search heuristics,
many of these may be useful as well in the analysis of classic randomized
algorithms or discrete random structures.
1
Introduction
Unlike in the field of classic randomized algorithms for discrete optimization problems, where theory has always supported (and, in fact, often led) the development
and understanding of new algorithms, the theoretical analysis of nature-inspired
search heuristics is much younger than the use of these heuristics. The use of
nature-inspired heuristics can easily be traced back to the 1960s, their rigorous
analysis with proven performance guarantees only started in the late 1990s. Propelled by impressive results, most notably from the German computer scientist
Ingo Wegener (*1950–†2008) and his students, theoretical works became quickly
accepted in the nature-inspired algorithms field and now form an integral part of it.
1
They now help to understand these methods, guide the choice of their parameters,
and even (as in the classic algorithms field) suggest new promising algorithms.
It is safe to say that Wegener’s vision that nature-inspired heuristics are nothing
more than a particular class of randomized algorithms, which therefore should be
analyzed with the same rigor as other randomized algorithms, has come true.
After around 20 years of theoretical analysis of nature-inspired algorithms,
however, we have to note that the methods used here are different from those in
the analysis of classic randomized algorithms. This is most visible for particular methods like the fitness level method or drift analysis, but applies even to the
elementary probabilistic tools employed throughout the field. The aim of this chapter is to collect those elementary tools which often have been used in the past 20
years. This includes classic material like expectations, variances, the coupon collector process, Markov’s inequality, Chebyshev’s inequality and Chernoff-Hoeffding
bounds for sums of independent random variables, but also topics that are used
rarely outside the analysis of nature-inspired heuristics like stochastic domination,
Chernoff-Hoeffding bounds for sums of independent geometrically distributed random variables, and Chernoff-Hoeffding bounds for sums of random variables which
are not fully independent. For all results, we also sketch a typical application or
refer to applications in the literature. The large majority of the results and applications presented in this work have appeared previously, some in textbooks, some
in recent conference publications.
This chapter aims to serve both as introduction for newcomers to the field
and as reference book for regular users of these methods. With both addressees
in mind, we did not shy away from stating also elementary reformulations of the
results. We hope that this saves all users of this chapter some time, which is better
spent on understanding the challenging random processes that arise in the analysis
of nature-inspired heuristics.
2
Notation
All notation in this chapter is standard and should need not much additional
explanation. We use N := {1, 2, . . . } to denote the positive integers. We write
N0 := N ∪ {0}. For intervals of integers, we write [a..b] := {x ∈ Z | a ≤ x ≤ b}.
We use the standard definition 00 := 1 (and not 00 = 0).
3
Elementary Probability Theory
We shall assume that the reader has some basic understanding of the concepts
of probability spaces, events and random variables. As usual in probability theory
2
and very convenient in analysis of algorithms, we shall almost never explicitly state
the probability space we are working in. Hence an intuitive understanding of the
notion of a random variable should be enough to follow this exposition.
While many results presented in the following naturally extend to continuous
probability spaces, in the interest of simplicity and accessability for a discrete
optimization audience, we shall assume that all random variables in this book
will be discrete, that is, they take at most a countable number of values. As
a simple example, consider the random experiment of independently rolling two
distinguishable dice. Let X1 denote the outcome of the first roll, that is, the
number between 1 and 6 which the first die displays. Likewise, let X2 denote
the outcome of the second roll. These are already two random variables. We
formalize the statement that with probability 16 the first die shows a one by saying
Pr[X1 = 1] = 16 . Also, the probability that both dice show the same number is
Pr[X1 = X2 ] = 61 . The complementary event that they show different numbers,
naturally has a probability of Pr[X1 6= X2 ] = 1 − Pr[X1 = X2 ] = 65 .
We can add random variables (defined over the same probability space), e.g.,
X := X1 + X2 is the sum of the numbers shown by the two dice, and we can
multiply a random variable by a number, e.g., X := 2X1 is twice the number
shown by the first die.
The most common type of random variable we shall encounter in this book is an
extremely simple one called binary random variable or Bernoulli random variable.
It takes the values 0 and 1 only. In consequence, the probability distribution of a
binary random variable X is fully described by its probability Pr[X = 1] of being
one, since Pr[X = 0] = 1 − Pr[X = 1].
Binary random variables often show up as indicator random variables for random events. For example, if the random experiment is a simple roll of a die, we
may define a random variable X by setting X = 1, if the die shows a 6, and X = 0
otherwise. We say that X is the indicator random variable for the event “die shows
a 6.”
Indicator random variables are useful for counting. If we roll a die n times and
X1 , . . . , Xn are the indicator random variables for the events that the correspondP
ing roll showed a 6 (considered as a success), then ni=1 Xi is a random variable
describing the number of times we saw a 6 in these n rolls. In general, a random
variable X that is the sum of n independent binary random variables being one
all with equal probability p, is called a binomial random variable (with success
probability p). We denote this distribution by Bin(n, p) and write X ∼ Bin(n, p)
to denote that X has this distribution. We have
!
n k
p (1 − p)n−k
Pr[X = k] =
k
for all k ∈ [0..n]. See Section 4.3 for the definition of the binomial coefficient.
3
A different question is how long we have to wait until we roll a 6. Assume that
we have an infinite sequence of die rolls and X1 , X2 , . . . are the indicator random
variables for the event that the corresponding roll showed a 6 (success). Then we
are interested in the random variable Y = min{k ∈ N | Xk = 1}. Again for the
general case of all Xi being one independently with probability p > 0, this random
variable Y is called geometric random variable (with success probability p). We
denote this distribution by Geom(p) and write Y ∼ Geom(p) to indicate that Y
is geometrically distributed (with parameter p). We have
Pr[Y = k] = (1 − p)k−1 p
for all k ∈ N. We note that an equally established definition is to count only the
failures, that is, to regard the random variable Y − 1. So some care is necessary
when comparing results from different sources.
4
Useful Inequalities
Before starting our presentation of probabilistic tools useful in the analysis of
randomized search heuristics, let us briefly mention a few inequalities that are
often needed to estimate probabilities arising naturally in this area.
4.1
Switching Between Exponential and Polynomial Terms
When dealing with events occurring with small probability ε > 0, we often encounter expressions like (1 − ε)n . Such a mix of a polynomial term (1 − ε) with
an exponentiation is often hard to work with. It is therefore very convenient that
1 − ε ≈ e−ε , so that the above expression becomes approximately the purely exponential term e−εn . In this section, we collect a few estimates of this flavor. All
are well-known and can be derived via elementary arguments.
Lemma 1. For all x ∈ R,
1 + x ≤ ex .
We give a canonic proof as an example for a proof method that is often useful
for such estimates.
Proof. Define a function f : R → R by f (x) = ex − 1 − x for all x ∈ R. Since
f ′ (x) = ex − 1, we have f ′ (x) = 0 if and only if x = 0. Since f ′′ (x) = ex > 0 for
all x, we see that x = 0 is the unique minimum of f . Since f (0) = 0, we have
f (x) ≥ 0 for all x, which is equivalent to the claim of the lemma.
4
f (x) = 1 + x + x2
x2
f (x) = 1 + x + 1−x
f (x) = ex
f (x) = 1 + x
f (x)
3
2
1
0
−1
−0.5
0
x
0.5
1
Figure 1: Plot of the estimates of Lemma 1 and 2.
Applying Lemma 1 to −x and taking reciprocals, we immediately derive the
first of the following two upper bounds for the exponential function. The second
bound again follows from elementary calculus. Obviously, the first estimate is
better for x < 0, the second one is better for x > 0.
Lemma 2.
(a) For all x < 1,
ex ≤
x
x2
1
=1+
=1+x+
.
1−x
1−x
1−x
(1)
In particular, for 0 ≤ x ≤ 1, we have e−x ≤ 1 − x2 .
(b) For all x < 1.79,
ex ≤ 1 + x + x2 .
(2)
As visible also from Figure 1, these estimates are strongest for x close to zero.
x
Replacing x with 1+x
in the first inequality of Lemma 2 gives the following bounds.
Corollary 3. For all x > −1, we have
x
e 1+x ≤ 1 + x ≤ ex .
For all x, y > 0,
xy
e x+y ≤ (1 + xy )y ≤ ex .
(3)
(4)
A reformulation often useful on the context of standard-bit mutation (mutating
a bit-string by flipping each bit independently with a small probability like n1 ) is
5
f (x) = (1 − 1/x)x
f (x) = (1 − 1/x)x−0.5
f (x) = (1 − 1/x)x−1
f (x) = ((1 − 1/x)x−1 + (1 − 1/x)x )/2
0.8
f (x)
0.6
0.4
0.2
2
3
4
5
6
7
8
9
10
x
Figure 2: Plots related to Corollary 4.
the following. Note that the first bound holds for all r ≥ 1, while it is often only
stated for r ∈ N. For the (not so interesting) boundary case r = 1, recall that we
use the common convention 00 := 1.
Corollary 4. For all r ≥ 1 and 0 ≤ s ≤ r,
(1 − 1r )r ≤ 1e ≤ (1 − 1r )r−1 ,
(1 − rs )r ≤ e−s ≤ (1 − rs )r−s .
(5)
(6)
Proof.
Occasionally, it is useful to know that (1 − 1r )r is monotonically increasing and
that (1 − 1r )r−1 is monotonically decreasing in r (and thus both converge to 1e ).
Lemma 5. For all 1 ≤ s ≤ r, we have (1 − 1s )s ≤ (1 − 1r )r and (1 − 1s )s−1 ≥
(1 − 1r )r−1 .
Finally, we mention Bernoulli’s inequality and a related result. The lower
Q
bound on ni=1 (1 − pi ) in Lemma 6 (b) below is also valid when all pi ∈ (−∞, 0].
The upper bound will be proven at the end of the Section 5.2, both to show how
probabilistic arguments can be used to prove non-probabilistic results and because
we have not found a proof for it in the literature.
Lemma 6. (a) Bernoulli’s inequality: Let x ≥ −1 and r ∈ {0} ∪ [1, ∞). Then
(1 + x)r ≥ 1 + rx.
6
(b) Weierstrass product inequality: Let p1 , . . . , pn ∈ [0, 1]. Let P :=
Then
n
1−P ≤
4.2
Y
(1 − pi ) ≤ 1 − P +
i=1
X
i<j
Pn
i=1
pi .
pi pj ≤ 1 − P + 12 P 2 .
Harmonic Number
Quite frequently in the analysis of randomized search heuristics we will encounter
P
the harmonic number Hn . For all n ∈ N, it is defined by Hn = nk=1 k1 . Approximating this sum via integrals, namely by
Z
n+1
1
Z n
1
1
dx ≤ Hn ≤ 1 +
dx,
x
1 x
we obtain the estimate
ln n < Hn ≤ 1 + ln n
(7)
valid for all n ≥ 1. Sharper estimates involving the Euler-Mascheroni constant
γ ≈ 0.5772156649 are known, e.g.,
Hn = ln n + γ ± O( n1 ),
1
Hn = ln n + γ + 2n
± O( n12 ).
For non-asymptotic statements, it is helpful to know that Hn −ln n is monotonically
decreasing (with limit γ obviously). In most cases, however, the simple estimate (7)
will be sufficient.
4.3
Binomial Coefficients and Stirling’s Formula
Since discrete probability is strongly related to counting, we often encounter the
binomial coefficients defined by
!
n!
n
:=
k
k! (n − k)!
for all n ∈ N0 , k ∈ [0..n]. The binomial coefficient nk equals the number of kelement subsets ofa given n-element set. For this reason, the above definition is
often extended to nk := 0 for k > n.
In this section, we give several useful estimates for binomial coefficients. We
start by remarking that, while very precise estimates are available, in the analysis
of randomized search heuristics often crude estimates are sufficient.
7
The following lemma lists some estimates which all can be proven by elementary
P
kk
ki
means. To prove the second inequality of (11), note that ek = ∞
i=0 i! ≥ k! gives
the elementary estimate
!k
k
≤ k! ≤ k k .
(8)
e
Qn/2
n!
= i=1 2i(2i−1)
=
(n/2)!(n/2)!q
i2
P
Q
n/2
n/2
2n i=1 (1 − 2i1 ) ≤ 2n exp(− 12 i=1 1i ) ≤ 2n exp(− 12 ln n2 ) = 2n n2 , see Lemma 1
q
n
n+1
2
and (7), while for odd n we have ⌊n/2⌋
= 21 (n+1)/2
≤ 2n n+1
.
To prove (12), note that for even n we have
n
n/2
=
Lemma 7. For all n ∈ N and k ∈ [1..n], we have
!
(9)
!
(10)
n
≤ 2n ,
k
n
k
!k
n
≤ nk ,
≤
k
nk
n
≤
≤
k!
k
!
ne
k
!
!
!k
,
(11)
s
2
n
n
≤ 2n
≤
.
⌊n/2⌋
n
k
(12)
q
n
2
of (12),
Stronger estimates, giving also the well-known version ⌊n/2⌋
≤ 2n πn
+
can be obtained from Stirling’s formula below. See [HPR 14] for an analysis of
randomized search heuristics which clearly required Stirling’s formula.
Theorem 8 (Robbins [Rob55]). For all n ∈ N,
√
n! = 2πn( ne )n Rn ,
1
1
) < Rn < exp( 12n
) < 1.08690405.
where 1 < exp( 12n+1
Corollary 9. For all n ∈ N and k ∈ [1..n − 1],
!
1
n
=√
k
2π
s
n
k(n − k)
n
k
!k
n
n−k
1
1
where 0.88102729... = exp(− 16 + 25
) ≤ exp(− 12k
−
1
1
1
− 12(n−k)+1
+ 12n
) < 1.
exp(− 12k+1
!n−k
Rnk ,
1
12(n−k)
+
1
)
12n+1
< Rnk <
Stirling’s formula was also used in [DW14, proof of Lemma 8] to√compute
another useful fact, namely that all binomial coefficients that are O( n) away
from the middle one have the same asymptotic order of magnitude of Θ(2n n−1/2 ).
Here the upper bound is simply (12).
8
Corollary 10. Let γ ≥ 0. Let n ∈ N and ℓ =
n
2
o(1)) 2√2 πn e−4γ .
n
2
√
± γ n. Then nℓ ≥ (1 −
When working with mutation rates different from classic choice of n1 , the following estimates can be useful.
Lemma 11. Let n ∈ N, k ∈ [0..n], and p ∈ [0, 1]. Let X ∼ Bin(n, p).
(a) Let Y ∼ Bin(n, nk ). Then Pr[X = k] ≤ Pr[Y = k]. This inequality is strict
except for trivial case p = nk .
(b) For k ∈ [1..n − 1], Pr[X = k] ≤
√1
2π
q
n
.
k(n−k)
Proof. The first part follows from Pr[X = k] = nk pk (1 − p)n−k and noting that
p 7→ pk (1−p)n−k has a unique maximum in the interval [0, 1], namely at p = nk . The
second part follows from the first and using
Corollary 9 to estimate the binomial
coefficient in the expression Pr[Y = k] = nk ( k1 )k (1 − nk )n−k .
For the special case that np = k, the second part of the lemma above was
already shown in [SW16]. For k ∈ {⌊np⌋, ⌈np⌉} but np 6= k, a bound larger than
ours by a factor of e was shown there as well.
To estimate sums of binomial coefficients, large deviations bounds (to be discussed in Section
10) can be an elegant tool. Imagine we need an upper bound
Pn n
for S = k=a k , where a > n2 . Let X be a random variable with distribution
Bin(n, 21 ). Then Pr[X ≥ a] = 2−n S. Using the additive Chernoff bound of The-
orem 53, we also see Pr[X ≥ a] = Pr[X ≥ E[X] + (a − n2 )] ≤ exp(−
2(a− n )2
2(a− n
)2
2
).
n
Consequently, S ≤ 2n exp(− n 2 ).
The same argument can even be used to estimate single binomial coefficients,
in particular,
to close to the middle one. Note that by Lemma 82,
those
not
P
S = nk=a nk and na are quite close when a is not too close to n2 . Hence
2(a − n2 )2
n
≤ 2n exp −
a
n
!
!
(13)
is a good estimate in this case.
5
Union Bound
The union bound, sometimes called Boole’s inequality, is a very elementary consequence of the axioms of a probability space, in particular, the σ-additivity of the
probability measure.
9
Lemma 12 (Union bound). Let E1 , . . . , En be arbitrary events in some probability
space. Then
#
"
Pr
n
[
i=1
Ei ≤
n
X
Pr[Ei ].
i=1
Despite its simplicity, the union bound is a surprisingly powerful tool in the
analysis of randomized algorithms. It draws its strength from the fact that does
not need any additional assumptions. In particular, the events Ei are not required
to be independent. Here is an example of such an application of the union bound.
5.1
Example: The (1 + 1) EA Solving the Needle Problem
The needle function is the fitness function f : {0, 1}n → Z defined by f (x) = 0
for all x ∈ {0, 1}n \ {(1, . . . , 1)} and f ((1, . . . , 1)) = 1. It is neither surprising
nor difficult to prove that all reasonable randomized search heuristics need time
exponential in n to find the maximum of the needle function. To give a simple
example for the use of the simplified drift theorem, it was shown in [OW11] that
the classic (1 + 1) EA within a sufficiently small exponential time does not even
get close to the optimum of the needle function (see Theorem 13 below). We now
show that the same result (and in fact a stronger one) can be shown via the union
bound.
The (1 + 1) EA is the simple randomized search heuristic that starts with a
random search point x ∈ {0, 1}n . Then, in each iteration, it generates from x a
new search point y by copying x into y and flipping each bit independently with
probability n1 . If the new search point (“offspring”) y is at least as good as the
parent x, that is, if f (y) ≥ f (x) for an objective function to be maximized, then
x is replaced by y, that is, we set x := y. Otherwise, y is discarded.
Algorithm 1: The (1 + 1) EA for maximizing f : {0, 1}n → R.
n
1 Choose x ∈ {0, 1} uniformly at random;
2 for t = 1, 2, 3, . . . do
3
y ← x;
4
for i ∈ [1..n] do
5
with probability n1 do yi ← 1 − yi;
6
if f (y) ≥ f (x) then x ← y;
The precise result of [OW11] is the following.
Theorem 13. For all η > 0 there are c1 , c2 > 0 such that with probability 1 − 2c1 n
the first 2c2 n search points generated in a run of the (1 + 1) EA on the needle
function all have a Hamming distance of more than ( 21 − η)n from the optimum.
10
The proof of this theorem in [OW11] argues as follows. Denote by x(0) , x(1) , . . .
the search points generated in a run of the (1 + 1) EA. Denote by x∗ the optimum
(i)
of the needle function. For all i ≥ 0, let Xi := H(x(i) , x∗ ) := |{j ∈ [1..n] | xj 6=
x∗j }| be the Hamming distance of x(i) from the optimum. The random initial
search point x(0) has an expected Hamming distance of n2 from the optimum. By
a simple Chernoff bound argument (Theorem 53), we see that with probability
1 − exp(−2η 2 n), we have X0 = H(x(0) , x∗ ) > ( 21 − η)n. Now a careful analysis of
the random process (Xi )i≥0 via a new “simplified drift theorem” gives the claim.
We now show that the Chernoff bound argument plus a simple union bound are
sufficient to prove the theorem. We show the following more explicit bound, which
also applies to all other unbiased algorithms in the sense of Lehre and Witt [LW12]
(roughly speaking, all algorithms which treat the bit-positions [1..n] in a symmetric
fashion as well as the bit-values {0, 1}).
Theorem 14. For all η > 0 and c > 0 we have that with probability at least
2
1−2(c−2 ln(2)η )n the first L := 2cn search points generated in a run of the (1+1) EA
(or any other unbiased black-box optimization algorithm) on the needle function
all have a Hamming distance of more than ( 12 − η)n from the optimum.
Proof. The key observation is that as long as the (1 + 1) EA has not found the
optimum, any search point x generated by the (1 + 1) EA is uniformly distributed
in {0, 1}n . Hence Pr[H(x, x∗ ) ≤ ( 21 − η)n] ≤ exp(−2η 2 n) by Theorem 53. By
the union bound, the probability that one of the first L := 2cn search points
generated by the (1 + 1) EA has a distance H(x, x∗ ) of at most ( 12 − η)n, is at
2
most L exp(−2η 2 n) = 2(c−2 ln(2)η )n .
To be more formal, let x(0) , x(1) , . . . be the search points generated in a run of
the (1 + 1) EA. Let T = min{t ∈ N0 | x(t) = x∗ }. Define a sequence y (0) , y (1) , . . .
of search points by setting y (t) := x(t) for all t ≤ T . For all t > T , let y (t) be
obtained from y (t−1) by flipping each bit independently with probability n1 . With
this definition, and since x(t) = x∗ for all t ≥ T , we have
{x(t) | t ∈ [0..L − 1]} = {x(t) | t ∈ [0.. min{T, L − 1}]}
= {y (t) | t ∈ [0.. min{T, L − 1}]} ⊆ {y (t) | t ∈ [0..L − 1]}.
Consequently,
Pr[∃t ∈ [0..L−1] : H(x(t) , x∗ ) ≤ ( 12 −η)n] ≤ Pr[∃t ∈ [0..L−1] : H(y (t) , x∗ ) ≤ ( 12 −η)n].
By the union bound,
Pr[∃t ∈ [0..L − 1] : H(y (t) , x∗ ) ≤ ( 12 − η)n] ≤
11
L−1
X
t=0
Pr[H(y (t), x∗ ) ≤ ( 21 − η)n].
Note that when y (t) is a search point uniformly distributed in {0, 1}n , then so is
y (t+1) . Since y (0) is uniformly distributed, all y (t) are. Hence by Theorem 53 we
have Pr[H(y (t) , x∗ ) ≤ ( 12 − η)n] ≤ exp(−2η 2 n) for all t and thus
L−1
X
t=0
Pr[H(y (t) , x∗ ) ≤ ( 21 − η)n] ≤ L exp(−2η 2 n) = 2(c−2 ln(2)η
2 )n
.
This proof immediately extends to all algorithms which, when optimizing the
needle function, generate uniformly distributed search points until the optimum is
found. These are, in particular, all unbiased algorithms in the sense of Lehre and
Witt [LW12].
Note that the yt in the proof above are heavily correlated. For all t, the
search points yt and yt+1 have an expected Hamming distance of exactly one.
Nevertheless, we could apply the union bound to the events “H(yt , x∗ ) < ( 21 − η)n”
and from this obtain a very elementary proof of Theorem 14.
5.2
Lower Bounds, Bonferroni Inequalities
The union bound is tight, that is, holds with equality, when the events Ei are
disjoint. In this case, the union bound simply reverts to the σ-additivity of the
probability measure. The second Bonferroni inequality gives a lower bound for the
probability of a union of events also when they are not disjoint.
Lemma 15. Let E1 , . . . , En be arbitrary events in some probability space. Then
Pr
"
n
[
i=1
#
Ei ≥
n
X
i=1
Pr[Ei ] −
n−1
X
n
X
i=1 j=i+1
Pr[Ei ∩ Ej ].
As an illustration, let us regard the performance of blind random search on
the needle function, that is, we let x(1) , x(2) , . . . be independent random search
points from {0, 1}n and ask ourselves what is the first hitting time T = min{t ∈
N | x(t) = (1, . . . , 1)} of the maximum x∗ = (1, . . . , 1) of the needle function (any
other function f : {0, 1}n → R with unique global optimum would do as well).
This is easy to compute directly. We see that T has geometric distribution with
success probability 2−n , so the probability that L iterations do not suffice to find
the optimum is Pr[T > L] = (1 − 2−n )L .
Let us nevertheless see what we can derive from union bound and second Bonferroni inequality. Let Et be the event x(t) = x∗ . Then the union bound gives
Pr[T ≤ L] = Pr
"
L
[
t=1
12
#
Et ≤ L2−n ,
the second Bonferroni inequality yields
Pr[T ≤ L] = Pr
"
L
[
t=1
#
Et ≥ L2−n −
L(L − 1) −2n
2 .
2
Hence if L = o(2n ), that is, L is of smaller asymptotic order than 2n , then Pr[T ≤
L] = (1 − o(1))L2−n , that is, the union bound estimate is asymptotically tight.
For reasons of completeness, we state the full set of Bonferroni inequalities.
Note that the case k = 1 is the union bound and the case k = 2 is the lemma
above.
Lemma 16. Let E1 , . . . , En be arbitrary events in some probability space. For all
k ∈ [1..n], let
X
Sk :=
Pr[Ai1 ∩ · · · ∩ Aik ].
1≤i1 <···<ik ≤n
Then for all k ∈ [1..n] we have
• Pr
"
• Pr
"
n
[
Ei ≤
n
[
#
i=1
i=1
#
Ei ≥
k
X
(−1)j−1 Sj for k ∈ [1..n] odd,
k
X
(−1)j−1 Sj for k ∈ [1..n] even.
j=1
j=1
In simple terms, the Bonferroni inequalities state that when we omit the terms
for j > k in the inclusion-exclusion formula
Pr
"
n
[
i=1
#
Ei =
n
X
(−1)j−1 Sj ,
j=1
then first of the omitted terms (that is, the one for j = k + 1) dominates the error.
So if k is odd and thus the first omitted term is negative, then we obtain a ≤
inequality, and inversely for k even.
We now use the Bonferroni inequalities to prove two of the inequalities given
in the previous section.
Proof of Lemma 6 (b). Consider some probability space with independent events
E1 , . . . , En having Pr[Ei ] = pi . Due to the independence,
n
Y
(1 − pi ) = Pr[∀i ∈ [1..n] : ¬Ei ] = 1 − Pr[∃i ∈ [1..n] : Ei ].
i=1
13
(14)
By the union bound, the right-hand side of (14) is at least 1 − ni=1 pi = 1 − P . By
the Bonferroni inequality for k = 2 and again the independence, the right-hand
side of (14) is at most
P
1−P +
X
i<j
Pr[Ei ∩ Ej ] = 1 − P +
X
i<j
pi pj ≤ 1 − P +
1
2
n X
n
X
i=1 j=1
pi pj = 1 − P + 21 P 2.
Note that the slack in the last inequality is only the term ni=1 p2i , so there is
P
not much reason to prefer the stronger upper bound 1 − P + i<j pi pj over the
bound 1 − P + 12 P 2 .
P
6
6.1
Expectations,
Markov’s
Chebyshev’s Inequality
Inequality,
and
Expectation
The expectation (or mean) of a random variable X taking values in some set Ω ⊆ R
P
is defined by E[X] = ω∈Ω ω Pr[X = ω], where we shall always assume that the
sum exists and is finite. As a trivial example, we immediately see that if X is a
binary random variable, then E[X] = Pr[X = 1].
For non-negative integral random variables, the expectation can also be computed by the following formula (which is valid also when E[X] is not finite).
Lemma 17. Let X be a random variable taking values in the non-negative integers.
Then
∞
E[X] =
X
i=1
Pr[X ≥ i].
This lemma, among others, allows to conveniently transform information about
the tail bound of a distribution into a bound on its expectation. This was done,
e.g., in [DJW02, proof of Lemma 10] for lower bounds and in [DG13] in the simplified proof of the multiplicative drift theorem (which is an upper bound on an
expectation). For exponential tails, we easily obtain the following general result.
Corollary 18 (Expectations from exponential tail bounds). Let α, β > 0 and
T ≥ 0. Let X be an integer random variable and Y be a non-negative integer
random variable.
(a) If Pr[X ≥ T + λ] ≤ α exp(− βλ ) for all λ ∈ N, then E[X] ≤ T + αβ.
(b) If Pr[Y ≤ T − λ] ≤ α exp(− βλ ) for all λ ∈ [1..T ], then E[Y ] ≥ T − αβ.
14
(c) If Pr[X ≥ (1 + ε)T ] ≤ α exp(− βε ) for all ε > 0, then E[X] ≤ (1 + αβ)T .
(d) If Pr[X ≤ (1 − ε)T ] ≤ α exp(− βε ) for all ε ∈ (0, 1], then E[X] ≥ (1 − αβ)T .
∞
i−T
Proof. We compute E[X] ≤ ∞
i=1 Pr[X ≥ i] ≤ T +
i=T +1 α exp(− β ) = T − α +
1
≤ T + αβ, where the last estimate uses Lemma 2.
α 1−exp(−1/β)
P
P
Similarly, we compute E[Y ] ≥ Ti=1 Pr[Y ≥ i] ≥ Ti=1 (1 − Pr[Y ≤ i − 1]) ≥
P∞
PT
λ
λ
λ=1 α exp(− β ) ≥ T − αβ.
λ=1 (1 − α exp(− β )) ≥ T −
The last two claims are simple reformulations of the first two.
P
P
In a similar vein, Lemma 17 yields an elegant analysis of the expectation of a
geometric random variable. Let X be a geometric random variable with success
probability p. Intuitively, we feel that the expected waiting time for a success
is 1p . This intuition is guided by the fact that after 1p repetitions of the underlying
binary random experiment, the expected number of successes is exactly one. This
intuition led to the right result, the “proof” however is not correct. The correct
proof either uses standard results in Markov chain theory, or elementary but nontrivial calculations, or (as done below) the above lemma.
Lemma 19 (Waiting time argument). Let X be a geometric random variable with
success probability p > 0. Then E[X] = p1 .
Proof. We have Pr[X ≥ i] = (1 − p)i−1 , since X ≥ i is the event of having no
success in the first i − 1 rounds of the random experiment. Now Lemma 17 gives
E[X] =
∞
X
i=1
Pr[X ≥ i] =
∞
X
i=1
(1 − p)i−1 =
1
1
= .
1 − (1 − p)
p
An elementary, but very useful property is that expectation is linear.
Lemma 20 (Linearity of expectation). Let X1 , . . . , Xn be arbitrary random variables and a1 , . . . , an ∈ R. Then
E
"
n
X
i=1
#
ai Xi =
n
X
ai E[Xi].
i=1
This fact is very convenient when we can write a complicated random variable
as sum of simpler ones. For example, let X be a binomial
random variable with
n k
parameters n and p, that is, we have Pr[X = k] = k p (1 −p)n−k . Since X counts
P
the number of successes in n (independent) trials, we can write X = ni=1 Xi as
the sum of (independent) binary random variables X1 , . . . , Xn , each with Pr[Xi =
15
1] = p. Here Xi is the indicator random variable for the event that the i-th trial
is a success. Using linearity of expectation, we compute
E[X] = E
"
n
X
i=1
#
Xi =
n
X
E[Xi ] = np.
i=1
Note that we did not need that the Xi are independent.
We just proved the following.
Lemma 21 (Expectation of binomial random variables). Let X be a binomial
random variable with parameters n and p. Then E[X] = pn.
In the same fashion, we can compute the following elementary facts.
Lemma 22. Let x, y, x∗ ∈ {0, 1}n . Denote by H(x, y) := |{i ∈ [1..n] | xi 6= yi }|
the Hamming distance of x and y.
(a) Let z be obtained from x via standard-bit mutation with rate p ∈ [0, 1],
that is, by flipping each bit of x independently with probability p. Then
E[H(x, z)] = pn and E[H(z, x∗ )] = H(x, x∗ ) + 2p( n2 − H(x, x∗ )).
(b) Let z be obtained from x and y via uniform crossover, that is, for each
i ∈ [1..n] independently, we have Pr[zi = xi ] = 21 = Pr[zi = yi ]. Then
E[H(x, z)] = 12 H(x, y) and E[H(z, x∗ )] = 12 (H(x, x∗ ) + H(y, x∗)).
(c) Let z be obtained from the unordered pair {x, y} via 1-point crossover, that
is, we choose r uniformly at random from [0..n] and then with probability 21
each
• define z by zi = xi for i ≤ r and zi = yi for i > r, or
• define z by zi = yi for i ≤ r and zi = xi for i > r.
Then E[H(x, z)] = 12 H(x, y) and E[H(z, x∗ )] = 12 (H(x, x∗ ) + H(y, x∗)).
The fact that the results for the two crossover operators are identical shows
again that linearity of expectation does not care about possible dependencies. We
have Pr[zi = xi ] = 12 in both cases, and this is what is important for the result,
whereas the fact that the events “zi = xi ” are independent for uniform crossover
and strongly dependent for 1-point crossover has no influence on the results.
16
6.2
Markov’s Inequality
Markov’s inequality is an elementary large deviation bound valid for all nonnegative random variables.
Lemma 23 (Markov’s inequality). Let X be a non-negative random variable.
Then for all λ > 0,
Pr[X ≥ λE[X]] ≤ λ1 ,
Pr[X ≥ λ] ≤
Proof. We have E[X] =
X
ω
ω Pr[X = ω] ≥
E[X]
.
λ
X
ω≥λ
(15)
(16)
λ Pr[X = ω] = λ Pr[X ≥ λ].
It is important to note that Markov’s inequality, without further assumptions,
only gives information about deviations above the expectation. If X is a (not
necessarily non-negative) random variable taking only values not larger than some
u ∈ R, then the random variable u − X is non-negative and Markov’s inequality
gives the bound
u − E[X]
Pr[X ≤ λ] ≤
,
u−λ
which is sometimes called reverse Markov’s inequality. An equivalent formulation
of this bound is
E[X] − λ
.
Pr[X > λ] ≥
u−λ
Markov’s inequality is useful if not much information is available about the
random variable under consideration. Also, when the expectation of X is very
small, then the following elementary corollary is convenient and, in fact, often
quite tight.
Corollary 24 (First moment method). If X is a non-negative random variable,
then Pr[X ≥ 1] ≤ E[X].
Corollary 24 together with linearity of expectation often gives the same results
as the union bound. For an example, recall that in Section 5.2 we observed that
in a run of the randomized search heuristic, the probability that the t-th search
point xt is the unique optimum of a given function f : {0, 1}n → R, is 2−n .
Denote this event by Et and let Xt be the indicator random variable for this event.
Then the probability that one of the first L search points is the optimum, can be
estimated equally well via the union bound or the above corollary and linearity of
17
expectation:
6.3
Pr
"
Pr
"
L
[
t=1
L
X
t=1
#
Et ≤
L
X
Pr[Et ] = L2−n ,
t=1
#
Xt ≥ 1 ≤ E
"
L
X
#
Xt =
t=1
L
X
E[Xt ] = L2−n .
t=1
Chebyshev’s Inequality
The second elementary large deviation bound is Chebyshev’s inequality, sometimes
called Bienaymé-Chebyshev inequality as it was first stated in Bienaymé [Bie53]
and later proven in Chebyshev [Tch67]. It seems rarely used in the theory of
randomized search heuristics (exceptions being [NSW10, DJWZ13]).
Recall that the variance of a discrete random variable X is
Var[X] = E[(X − E[X])2 ] = E[X 2 ] − E[X]2 .
Just by definition, the variance already is a measure of how well X is concentrated
around its mean. Applying Markov’s inequality to the random variable (X −
E[X])2 easily yields the following inequality.
Lemma 25 (Chebyshev’s inequality). For all λ > 0,
q
h
i
Pr |X − E[X]| ≥ λ Var[X] ≤
h
i
Pr |X − E[X]| ≥ λ ≤
1
,
λ2
Var[X]
.
λ2
(17)
(18)
Note that Chebyshev’s inequality automatically yields a two-sided tail bound
(that is, for both cases that the random variable is larger and smaller than its
expectation), as opposed to Markov’s inequality (giving just a bound for exceeding
the expectation). There is a one-sided version of Chebyshev’s inequality that is
often attributed to Cantelli, though Hoeffding [Hoe63] sees Chebyshev [Tch74] as
its inventor.
Lemma 26 (Cantelli’s inequality). For all λ > 0,
h
q
i
Pr X ≥ E[X] + λ Var[X] ≤
1
,
λ2 +1
(19)
Pr X ≤ E[X] − λ Var[X] ≤
1
.
λ2 +1
(20)
h
q
i
In most applications, the slightly better bound of Cantelli’s inequality is not
very interesting. Cantelli’s inequality, however, has the charm that the right-hand
side is always less than one.
18
While Markov’s inequality can be used to show that a non-negative random variable X rarely is positive (first moment method), Chebyshev’s inequality can serve the opposite purpose, namely showing that X is positive with
good probability. By taking λ = E[X] in (18), we obtain the first estimate
of the following lemma. Using the Cauchy-Schwarz inequality and computing
E[X]2 = E[X1X6=0 ]2 ≤ E[X 2 ]E[1X6=0 ] = E[X 2 ] Pr[X 6= 0], we obtain the second
2
. Since
estimate, which has the nice equivalent formulation Pr[X 6= 0] ≥ E[X]
E[X 2 ]
2
2
E[X ] ≥ E[X] , the second estimate gives a stronger bound for Pr[X = 0] than
the first. While the lemma below does not require that X is non-negative, the
typical application of showing that X is positive requires that X is non-negative
in the second bound, so that Pr[X 6= 0] = Pr[X > 0].
Lemma 27 (Second moment method). For a random variable X with E[X] 6= 0,
Pr[X = 0] ≤ Pr[X ≤ 0] ≤
Pr[X = 0] ≤
Var[X]
,
E[X]2
(21)
Var[X]
.
E[X 2 ]
(22)
In the (purely academic) example of finding a unique global optimum via blind
random search (see Section 5.2), let Xt be indicator random variable for the event
P
that the t-th search point is the optimum. Let X = Lt=1 Xt . Then the probability
that the optimum is found within the first L iterations is
Pr[X > 0] = 1 − Pr[X = 0] ≥ 1 −
Var[X]
.
E[X]2
The variance for a sum of binary random variables is
Var[X] =
L
X
t=1
Var[Xt ] +
X
s<t
Cov[Xs , Xt ] ≤ E[X] +
X
Cov[Xs , Xt ],
s<t
where we recall the definition of the covariance Cov[U, V ] := E[UV ] − E[U]E[V ]
of two arbitrary random variables U and V . Here we have Cov[Xs , Xt ] = 0, since
the Xt are independent. Consequently,
Pr[X > 0] ≥ 1 −
1
.
E[X]
Hence the probability to find the optimum within L iterations is Pr[T ≤ L] =
Pr[X > 0] ≥ 1 − L21−n . Note that this estimate is, for interesting case that E[X]
2−2n which we
is large, much better than the bound Pr[T ≤ L] ≥ L2−n − L(L−1)
2
obtained from the second Bonferroni inequality.
19
7
Conditioning
In the analysis of randomized heuristics, we often want to argue that a certain
desired event C already holds and the continue arguing under this condition. Formally, this gives rise to a new probability space where each of the original events
A now has a probability of
Pr[A | C] :=
Pr[A ∩ C]
.
Pr[C]
Obviously, this only makes sense for events C with Pr[C] > 0. In an analogous
fashion, we define the expectation of a random variable X conditional on C by
P
E[X | C] = ω∈C X(ω) Pr[ω | C]. Conditioning as a proof technique has many
faces, among them the following.
Decomposing events: If we can write some event A as the intersection of two
events A1 and A2 , then it can be useful to first compute the probability of A1 and
then the probability of A2 conditional on A1 . Right from the definition, we have
Pr[A1 ∩ A2 ] = Pr[A1 ] Pr[A2 | A1 ]. Of course, this requires that we have some direct
way of computing Pr[A1 | A2 ].
Case distinctions: Let C1 , . . . , Ck be a partition of our probability space. If it
is easy to analyse our problem conditional on each of these events (“in the that case
Ci holds”), then the following law of total probability and law of total expectation
are useful.
Lemma 28 (Laws of total probability and total expectation). Let C1 , . . . , Ck be
a partition of our probability space. Let A be some event and X be some random
variable. Then
Pr[A] =
E[X] =
k
X
i=1
k
X
i=1
Pr[A | Ci ] Pr[Ci],
E[X | Ci ] Pr[Ci ].
Excluding rare events: Quite often in the analysis of nature-inspired search
heuristics, we would like to exclude some rare unwanted event. For example,
assume that we analyze an evolutionary algorithm using standard-bit mutation
with mutation rate n1 . Then it is very unlikely that in an application of this
mutation operator more than n1/4 bits are flipped. So it could be convenient to
1/4
exclude this rare event, say by stating that “with probability 1 − 2−Ω(n ) , in none
20
of the first n2 applications of the mutation operator more than n1/4 bits are flipped;
let us in the following condition on this event”. See the proofs of Theorem 7 and 8
in [DJW02] for examples where such an argument is useful.
What could be a problem with this approach is that as soon as we condition
on such an event, we change the probability space and thus arguments valid in
the unconditional setting are not valid anymore. As a simple example, note that
once we condition on that we flip at most n1/4 bits, the events Ei that the i-th
bit flips are not independent anymore. Fortunately, we can safely ignore this in
most cases (and many authors do so without saying a word on this affair). The
reason is that when conditioning on an almost sure event, then the probabilities
of all events change only very little (see the lemma below for this statement made
precise). Hence in our example, we can compute the probability of some event
assuming that the bit flips are independent and then correct this probability by a
minor amount.
Lemma 29. Let C be some event with probability 1 − p. Let A be any event. Then
Pr[A]
Pr[A] − p
≤ Pr[A | C] ≤
.
1−p
1−p
In particular, for p ≤ 21 , we have Pr[A] − p ≤ Pr[A | C] ≤ Pr[A] + 2p.
The proof of this lemma follows right from the definition of conditional probabilities and the elementary estimate Pr[A] − p ≤ Pr[A \ C] = Pr[A ∩ C] ≤ Pr[A],
where C denotes the complement of C. From this, we also observe the natural fact
that when A ⊆ C, that is, the event A implies C, then conditioning on C rather
increases the probability of A:
Pr[A ∩ C]
Pr[A]
=
≥ Pr[A].
Pr[C]
Pr[C]
(23)
Pr[A] − p
Pr[A \ C]
=
≤ Pr[A].
Pr[C]
1−p
(24)
Pr[A | C] =
Likewise, when A ⊇ C, then
Pr[A | C] =
For example, if X is the number of bits flipped in an application of standard-bit
mutation, then Pr[X ≤ 10 | X ≤ n2 ] ≥ Pr[X ≥ 10] and Pr[X ≥ 10 | X ≤ n2 ] ≤
Pr[X ≥ 10].
Conditional binomial random variables: We occasionally need to know the
expected value of a binomially distributed random variable X ∼ Bin(n, p) conditional on that the variable has at least a certain value k. An intuitive (but wrong)
21
argument is that E[X | X ≥ k] should be around k + p(n − k), because we know
already that k of the n independent trials are successes and the remaining (n − k)
trials still have their independent success probability of p. While this argument
is wrong (as we might need more than k trials to have k successes), the result is
correct as an upper bound as shown in this lemma from [DD17].
Lemma 30. Let X be a random variable with binomial distribution with parameters n and p ∈ [0, 1]. Let k ∈ [0..n]. Then
E[X | X ≥ k] ≤ k + (n − k)p ≤ k + E[X].
Pn
Proof. Let X = i=1 Xi with X1 , . . . , Xn being independent binary random variables with Pr[Xi = 1] = p for all i ∈ [1..n]. Conditioning on X ≥ k, let
P
ℓ := min{i ∈ [1..n] | ij=1 Xj = k}. Then
E[X | X ≥ k] =
n
X
i=1
Pr[ℓ = i | X ≥ k]E[X | ℓ = i].
Note that ℓ ≥ k by definition. Note also that (X | ℓ = i) = k + nj=i+1 Xj with
unconditioned Xj . In particular, E[X | ℓ = i] = k + (n − i)p. Consequently,
P
E[X | X ≥ k] =
≤
n
X
i=1
n
X
i=k
Pr[ℓ = i | X ≥ k]E[X | ℓ = i]
Pr[ℓ = i | X ≥ k](k + (n − k)p) = k + (n − k)p.
We note that, in the language introduced in the following section, we have
actually shown the stronger statement that (X | X ≥ k) is dominated by k +
Bin(n − k, p). This stronger version can be useful to obtain tail bounds for (X |
X ≥ k).
8
Stochastic Domination and Coupling
In this section, we discuss two concepts that are not too often used explicitly, but
where we feel that mastering them can greatly help in the analysis of randomized
search heuristics. The first of these is stochastic domination, which is a very strong
way of saying that one random variable is better than another even when they are
not defined on the same probability space. The second concept is coupling, which
means defining two random variables suitably over the same probability space to
facilitate comparing them. These two concepts are strongly related: If the random
variable Y dominates X, then X and Y can be coupled in a way that Y is pointwise not smaller than X, and vice versa. The results of this section and some
related ones have appeared in [Doe18], however, in a more condensed form.
22
8.1
The Notion of Stochastic Domination
Definition 31 (Stochastic domination). Let X and Y be two random variables
not necessarily defined on the same probability space. We say that Y stochastically
dominates X, written as X Y , if for all λ ∈ R we have Pr[X ≤ λ] ≥ Pr[Y ≤ λ].
If Y dominates X, then the cumulative distribution function of Y is point-wise
not larger than the one of X. The definition of domination is equivalent to
∀λ ∈ R : Pr[X ≥ λ] ≤ Pr[Y ≥ λ],
which is maybe a formulation making it more visible why we feel that Y is at least
as large as X.
Concerning nomenclature, we remark that some research communities in addition require that the inequality is strict for at least one value of λ. Hence,
intuitively speaking, Y is strictly larger than X. From the mathematical perspective, this appears not very practical. Consequently, our definition above is
more common in computer science. We also note that stochastic domination is
sometimes called first-order stochastic domination. For an extensive treatment of
various forms of stochastic orders, we refer to [MS02].
The usual way of explaining stochastic domination is via games. Let us
consider the following three games.
Game A: With probability 12 , each you win 500 and 1500.
Game B: With probability 31 , you win 500, with probability 61 , you win 800, and
with probability 12 , you win 1500.
1
Game C: With probability 1000
, you win 2,000,000. Otherwise, you win nothing.
Which of these games is best to play? It is intuitively clear that you prefer
Game B over Game A. However, it is not clear whether you should prefer Game C
over Game B. Clearly, the expected win in Game C is 2000 compared to only 1050
in Game B. However, the chance of winning something at all is really small in
Game C. If you do not like to go home empty-handed, you might prefer Game B.
The mathematical take on these games is that the random variable XB describing the win in Game B stochastically dominates the one XA for Game A.
This captures our intuitive feeling that it cannot be wrong to prefer Game B over
Game A. For Games B and C, neither of XB and XC dominates the other. Consequently, it depends on the precise utility function of the player which game he
prefers. This statement is made precise in the following lemma.
Lemma 32. The following two conditions are equivalent.
(a) X Y .
23
(b) For all monotonically non-decreasing functions f : R → R, we have
E[f (X)] ≤ E[f (Y )].
As a simple corollary we note the following.
Corollary 33. If X Y , then E[X] ≤ E[Y ].
We note another simple, but useful property.
Lemma 34. Let X1 , . . . , Xn be independent random variables defined over some
common probability space. Let Y1 , . . . , Yn be independent random variables defined
over a possibly different probability space. If Xi Yi for all i ∈ [1..n], then
n
X
i=1
Xi
n
X
Yi .
i=1
For discrete random variables, this result is a special case of Lemma 38 stated
further below.
Finally, we note two trivial facts.
Lemma 35. Let X and Y be random variables.
(a) If X and Y are defined on the same probability space and X ≤ Y , then
X Y.
(b) If X and Y are identically distributed, then X Y .
8.2
Stochastic Domination in Runtime Analysis
From the perspective of algorithms analysis, stochastic domination allows to very
clearly state that one algorithm is better than another. If the runtime distribution
XA of algorithms A dominates the one XB of Algorithm B, then from the runtime
perspective Algorithm B is always preferable to Algorithm A.
In a similar vein, we can use domination also to give more detailed descriptions
of the runtime of an algorithm. For almost all algorithms, we will not be able to
precisely determine the runtime distribution. However, via stochastic domination
we can give a lot of useful information beyond, say, just the expectation. We
demonstrate this via an extension of the classic fitness level method, which is
implicit in Zhou, Luo, Lu, and Han [ZLLH12].
Theorem 36 (Domination version of the fitness level method). Consider an iterative randomized search heuristic A maximizing a function f : Ω → R. Let
A1 , . . . , Am be a partition of Ω such that for all i, j ∈ [1..m] with i < j and all
24
x ∈ Ai , y ∈ Aj , we have f (x) < f (y). Set A≥i := Ai ∪ · · · ∪ Am . Let p1 , . . . , pm−1
be such that for all i ∈ [1..m − 1] we have that if the best-so-far search point is
in Ai , then regardless of the past A has a probability of at least pi to generate a
search point in A≥i+1 in the next iteration.
Denote by T the (random) number of iterations A takes to generate a search
point in Am . Then
T
m−1
X
Geom(pi ),
i=1
where this sum is to be understood as a sum of independent geometric distributions.
To prove this theorem, we need a technical lemma which we defer to the subsequent subsection to ease reading this part.
Proof. Consider a run of the algorithm A. For all i ∈ [1..m], let Ti be the first
time (iteration) when A has generated a search point in A≥i . Then T = Tm =
Pm−1
i=1 (Ti+1 − Ti ). By assumption, Ti+1 − Ti is dominated by a geometric random
variable with parameter pi regardless what happened before time Ti . Consequently,
Lemma 38 gives the claim.
Note that a result like Theorem 36 implies various statements on the runtime.
P
1
By Corollary 33, the expected runtime satisfies E[T ] ≤ m−1
i=1 pi , which is the common version of the fitness level theorem [Weg01]. By using tail bounds for sums
of independent geometric random variables (see Section 10.4), we also obtain runtime bounds that hold with high probability. This was first proposed in [ZLLH12].
We defer a list of examples where previous results can profitably be turned into a
domination statement to Section 10.4, where we will also have the large deviation
bounds to exploit such statements.
8.3
Domination by Independent Random Variables
A situation often encountered in the analysis of algorithms is that a sequence of
random variables is not independent, but that each member of the sequence has a
good chance of having a desired property no matter what was the outcome of its
predecessors. In this case, the random variables in some sense can be treated as if
they were independent.
Lemma 37. Let X1 , . . . , Xn be arbitrary binary random variables and let
X1∗ , . . . , Xn∗ be independent binary random variables.
(a) If we have
Pr[Xi = 1 | X1 = x1 , . . . , Xi−1 = xi−1 ] ≤ Pr[Xi∗ = 1]
25
for all i ∈ [1..n] and all x1 , . . . , xi−1 ∈ {0, 1} with Pr[X1 = x1 , . . . , Xi−1 =
xi−1 ] > 0, then
n
X
i=1
Xi
n
X
Xi∗ .
i=1
(b) If we have
Pr[Xi = 1 | X1 = x1 , . . . , Xi−1 = xi−1 ] ≥ Pr[Xi∗ = 1]
for all i ∈ [1..n] and all x1 , . . . , xi−1 ∈ {0, 1} with Pr[X1 = x1 , . . . , Xi−1 =
xi−1 ] > 0, then
n
X
i=1
Xi∗
n
X
Xi .
i=1
Note that here and in the following, we view “X1 = x1 , . . . , Xi−1 = xi−1 ” for
i = 1 as an empty intersection of events, that is, an intersection over an empty
index set. As most textbooks, we define this to be the whole probability space.
Both parts of the lemma are simple corollaries from the following, slightly
technical, general result, which might be of independent interest.
For two sequences (X1 , . . . , Xn ) and (X1∗ , . . . , Xn∗ ) of random variables, we say
that (X1∗ , . . . , Xn∗ ) unconditionally sequentially dominates (X1 , . . . , Xn ) if for all
i ∈ [1..n] and all x1 , . . . , xi−1 ∈ R with Pr[X1 = x1 , . . . , Xi−1 = xi−1 ] > 0, we have
(Xi | X1 = x1 , . . . , Xi−1 = xi−1 ) Xi∗ . Analogously, we speak of unconditional
sequential subdomination if the last condition is replaced by Xi∗ (Xi | X1 =
x1 , . . . , Xi−1 = xi−1 ).
The following lemma shows that unconditional sequential (sub-)domination
and independence of the Xi∗ imply (sub-)domination for the sums of these random
variables. Note that unconditional sequential (sub-)domination is inherited by
subsequences, so the following lemma immediately extends to sums over arbitrary
subsets I of the index set [1..n].
Lemma 38. Let X1 , . . . , Xn be arbitrary discrete random variables.
X1∗ , . . . , Xn∗ be independent discrete random variables.
Let
(a) If (X1∗ , . . . , Xn∗ ) unconditionally sequentially dominates (X1 , . . . , Xn ), then
Pn
Pn
∗
i=1 Xi
i=1 Xi .
(b) If (X1∗ , . . . , Xn∗ ) unconditionally sequentially subdominates (X1 , . . . , Xn ),
P
P
then ni=1 Xi∗ ni=1 Xi .
Proof. The two parts of the lemma imply each other (as seen by multiplying the
random variables with −1), so it suffices to prove the first statement.
26
Since the statement of the theorem is independent of the correlation between
the Xi and the Xi∗ , we may assume that they are independent. Let λ ∈ R. Define
Pj := Pr
X
j
n
X
Xi +
i=j+1
i=1
Xi∗ ≥ λ
for j ∈ [0..n]. We show Pj+1 ≤ Pj for all j ∈ [0..n − 1].
For m ∈ R, let Ωm denote the set of all (x1 , . . . , xj , xj+2 , . . . , xn ) ∈ Rn−1 such
P
that Pr[X1 = x1 , . . . , Xj = xj ] > 0 and i∈[1..n]\{j+1} xi = λ − m. Let M := {m ∈
R | Ωm 6= ∅}. Then
Pj+1 = Pr
j+1
X
n
X
Xi∗ ≥ λ
Xi +
Xi∗
Xi +
i=j+2
i=1
=
X
Pr
i=1
m∈M
=
X
j
n
X
i=j+2
X
X
m∈M (x1 ,...,xj ,xj+2 ,...,xn )∈Ωm
= λ − m ∧ Xj+1 ≥ m
h
Pr X1 = x1 , . . . , Xj = xj ·
h
i
Pr Xj+1 ≥ m X1 = x1 , . . . , Xj = xj ·
≤
X
Pr
Xi +
X
j
i=1
Xi +
n
X
i=j+2
i=1
m∈M
= Pr
X
j
n
X
i=j+1
Xi∗
n
Y
h
Pr Xi∗ = xi
i=j+2
h
i
i
∗
Xi∗ = λ − m · Pr Xj+1
≥m
≥λ
= Pj .
Thus, we have
Pr
X
n
i=1
8.4
Xi ≥ λ = Pn ≤ Pn−1 ≤ · · · ≤ P1 ≤ P0 = Pr
X
n
i=1
Xi∗
≥λ .
Coupling
Coupling is an analysis technique that consists of defining two unrelated random
variables over the same probability space to ease comparing them. As an example,
let us regard standard-bit mutation with rate p and with rate q, where p < q.
Intuitively, it seem obvious that we flip more bits when using the higher rate q.
27
We could make this precise by looking at the distributions of the random variables
Xp and Xq describing the numbers of bits that flip and computing that Xp Xq .
For that, we would need to show that for all k ∈ [0..n], we have
k
X
i=0
k
X
n i
n i
p (1 − p)n−i ≥
q (1 − q)n−i .
i
i
i=0
!
!
Coupling is a way to get the same result in a more natural manner.
Consider the following random experiment. For each i ∈ [1..n], let ri be a
random number chosen independently and uniformly distributed in [0, 1]. Let X̃p
be the number of ri that are less than p and let Xq be the number of ri that are less
than q. We immediately see that X̃p ∼ Bin(n, p) and X̃q ∼ Bin(n, q). However,
we know more. We defined X̃p and X̃q in a common probability space in a way
that we have X̃p ≤ X̃q with probability one: Xp and Xq , viewed as functions
on the (hidden) probability space Ω = {(r1 , . . . , rn ) | r1 , . . . , rn ∈ [0, 1]} satisfy
X̃p (ω) ≤ X̃q (ω) for all ω ∈ Ω. Consequently, by the trivial Lemma 35, we have
Xp X̃p X̃q Xq and hence Xp Xq .
The same argument works for geometric distributions. We summarize these
findings (and two more) in the following lemma. Part (b) follows from the obvious
embedding (which is a coupling as well) of the Bin(n, p) probability space into the
one of Bin(m, p). The first inequality of part (c) is easily computed right from the
definition of domination (and holds in fact for all random variables), the second
part was proved and used in [KW17].
Lemma 39. Let X and Y be two random variables. Let p, q ∈ [0, 1] with p ≤ q.
(a) If X ∼ Bin(n, p) and Y ∼ Bin(n, q), then X Y .
(b) If n ≤ m, X ∼ Bin(n, p) and Y ∼ Bin(m, p), then X Y .
(c) If X ∼ Bin(n, p) and x ∈ [0..n], then X (X | X ≥ x) (X + x).
(d) If p > 0, X ∼ Geom(p), and Y ∼ Geom(q), then X Y .
Let us now formally define what we mean by coupling. Let X and Y be two
random variables not necessarily defined over the same probability space. We
say that (X̃, Ỹ ) is a coupling of (X, Y ) if X̃ and Ỹ are defined over a common
probability space and if X and X ′ as well as Y and Y ′ are identically distributed.
This definition itself is very weak. (X, Y ) have many couplings and most of
them are not interesting. So the art of coupling as a proof and analysis technique
is to find a coupling of (X, Y ) that allows to derive some useful information.
It is not a coincidence that we could use coupling to prove stochastic domination. The following theorem is well-known.
28
Theorem 40. Let X and Y be two random variables. Then the following two
statements are equivalent.
(a) X Y .
(b) There is a coupling (X̃, Ỹ ) of (X, Y ) such that X̃ ≤ Ỹ .
We remark without further details that coupling as a proof technique found
numerous powerful applications beyond its connection to stochastic domination.
8.5
Domination in Fitness or Distance
So far we have used stochastic domination to compare runtime distributions. We
now show that stochastic domination is a very elegant proof tool also when applied
to other distributions. To do so, we give a short and elegant proof for the result of
Witt [Wit13] that compares the runtimes of mutation-based algorithms. The main
reason why our proof is significantly shorter than the one of Witt is that we use
the notion of stochastic domination also for the distance from the optimum. This
will also be an example where we heavily exploit the connection between coupling
and stochastic domination (Theorem 40).
To state this result, we need the notion of a (µ, p) mutation-based algorithm
introduced in [Sud13]. This class of algorithms is called only mutation-based
in [Sud13], but since (i) it does not include all adaptive algorithms using mutation only, e.g., those regarded in [JW06, OLN09, BDN10, BLS14, DGWY17],
(ii) it does not include all algorithms using a different mutation operator than
standard-bit mutation, e.g., those in [DDY16a, DDY16b, LOW17, DLMN17], and
(iii) this notion collides with the notion of unary unbiased black-box complexity algorithms (see [LW12]), which without greater justification could also be called the
class of mutation-based algorithms, we feel that a notion making these restrictions
precise is more appropriate.
The class of (µ, p) mutation-based algorithms comprises all algorithms which
first generate a set of µ search points uniformly and independently at random from
{0, 1}n and then repeat generating new search points from any of the previous ones
via standard-bit mutation with probability p. This class includes all (µ + λ) and
(µ, λ) EAs which only use standard-bit mutation with static mutation rate p.
Denote by (1 + 1) EAµ the following algorithm in this class. It first generates
µ random search points. From these, it selects uniformly at random one with
highest fitness and then continues from this search point as a (1 + 1) EA, that
is, repeatedly generates a new search point from the current one via standard-bit
mutation with rate p and replaces the previous one by the new one if the new one
is not worse (in terms of the fitness). This algorithm was called (1 + 1) EA with
BestOf(µ) initialization in [dPdLDD15].
29
For any algorithm A from the class of (µ, p) mutation-based algorithms and
any fitness function f : {0, 1}n → R, let us denote by T (A, f ) the runtime of the
algorithm A on the fitness function f , that is, the number of the first individual
generated that is an optimal solution. Usually, this will be µ plus the number
of the iteration in which the optimum was generated. To cover also the case
that one of the random initial individuals is optimal, let us assume that these
initial individuals are generated sequentially. As a final technicality, for reasons
of convenience, let us assume that the (1 + 1) EAµ in iteration µ + 1 does not
choose as parent a random one of the previous search points with maximal fitness,
but the last one with maximal fitness. Since the first µ individuals are generated
independently, this modification does not change the distribution of this parent.
In this language, Witt [Wit13] shows the following remarkable result.
Theorem 41. For any (µ, p) mutation-based algorithm A and any f : {0, 1}n → R
with unique global optimum,
T ((1 + 1) EAµ , OneMax) T (A, f ).
This result significantly extends results of a similar flavor in [BE08, DJW12,
Sud13]. The importance of such types of results is that they allow to prove lower
bounds for the performance of many algorithm on essentially arbitrary fitness
functions by just regarding the performance of the (1 + 1) EAµ on OneMax.
Let us denote by |x|1 the number of ones in the bit string x ∈ {0, 1}n . In
other words, |x|1 = kxk1 , but the former is nicer to read. Then Witt [Wit13] has
shown the following natural domination relation between offspring generated via
standard-bit mutation.
Lemma 42. Let x, y ∈ {0, 1}n . Let p ∈ [0, 21 ]. Let x′ , y ′ be obtained from x, y via
standard-bit mutation with rate p. If |x|1 ≤ |y|1, then |x′ |1 |y ′|1 .
We are now ready to give our alternate proof for Theorem 41. While it is
clearly shorter that the original one in [Wit13], we also feel that it is more natural.
In very simple words, it shows that T (A, f ) dominates T ((1 + 1) EAµ , OneMax)
because the search points generated in the run of the (1 + 1) EAµ on OneMax
always are at least as close to the optimum (in the domination or coupling sense)
as the ones in the run of A on f .
Proof. Since A treats bit-positions and bit-values in a symmetric fashion, we may
without loss of generality assume that the unique optimum of f is (1, . . . , 1).
Let x(1) , x(2) , . . . be the sequence of search points generated in a run of A
on the fitness function f . Hence x(1) , . . . , x(µ) are independently and uniformly
distributed in {0, 1}n and all subsequent search points are generated from suitably
chosen previous ones via standard-bit mutation with rate p. Let y (1) , y (2) , . . . be
30
the sequence of search points generated in a run of the (1 + 1) EAµ on the fitness
function OneMax.
We show how to couple these random sequences of search points in a way that
(t)
|x̃ |1 ≤ |ỹ (t) |1 for all t ∈ N. We take as common probability space Ω simply the
space that (x(t) )t∈N is defined on and let x̃(t) = x(t) for all t ∈ N.
We define the ỹ (t) inductively as follows. For t ∈ [1..µ], let ỹ (t) = x(t) . Note
that this trivially implies |x̃(t) |1 ≤ |ỹ (t) |1 for these search points. Let t > µ and
′
′
assume that |x̃(t ) |1 ≤ |ỹ (t ) |1 for all t′ < t. Let s ∈ [1..t − 1] be maximal such that
ỹ (s) has maximal OneMax-fitness among ỹ (1) , . . . , ỹ (t−1) . Let r ∈ [1..t − 1] be such
that x(t) was generated from x(r) in the run of A on f . By induction, we have
|x(r) |1 ≤ |ỹ (r) |1 . By the choice of s we have |ỹ (r) |1 ≤ |ỹ (s) |1 . Consequently, we have
|x(r) |1 ≤ |ỹ (s)|1 . By Lemma 42 and Theorem 40, there is a random ỹ (t) (defined on
Ω) such that ỹ (t) has the distribution of being obtained from ỹ (s) via standard-bit
mutation with rate p and such that |x(t) |1 ≤ |ỹ (t) |1 .
With this construction, the sequence (ỹ (t) )t∈N has the same distribution as
(y (t) )t∈N . This is because the first µ elements are random and then each subsequent
one is generated via standard-bit mutation from the current-best one, which is just
the way the (1 + 1) EAµ is defined. At the same time, we have |x̃(t) |1 ≤ |ỹ (t) |1
for all t ∈ N. Consequently, we have min{t ∈ N | |ỹ (t) |1 = n} ≤ min{t ∈
N | |x(t) |1 = n}. Since T ((1 + 1) EAµ , OneMax) and min{t ∈ N | |ỹ (t) |1 = n} are
identically distributed and also T (A, f ) and min{t ∈ N | |x(t) |1 = n} are identically
distributed, we have T ((1 + 1) EAµ , OneMax) T (A, f ).
While not explicitly using the notion of stochastic domination, the result and
proof in [BE08] bear some similarity to those above. In very simple words and
omitting many details, the result [BE08, Theorem1] states the following. Assume
that you run the (1 + 1) EA and some other algorithm A (from a relatively large
class of algorithms) to maximize a function f . Denote by x(t) and y (t) the best
individual produced by the (1 + 1) EA and A up to iteration t. Assume that
for all t and all possible runs of the algorithms up to iteration t we have that
f (x(t) ) ≥ f (y (t) ) implies f (x(t+1) ) f (y (t+1) ). Assume further that the random
initial individual of the (1 + 1) EA is at least as good (in terms of f ) as all initial
individuals of algorithm A. Then f (x(t) ) f (y (t) ) for all t.
The proof of this result (like the one of the fitness domination statement in
our proof of Theorem 41) uses induction over the time t. Since [BE08] do not use
the notion of stochastic domination explicitly, they cannot simply couple the two
processes, but have to compare the two distributions manually using an argument
they call Abel transform. While not fully exploiting the power of stochastic domination, [BE08] might be the first paper in the theory of evolutionary algorithms
to formulate results and proofs via stochastic domination.
31
9
The Coupon Collector Process
The coupon collector process is one of the central building blocks in the analysis of
randomized algorithms. It is particularly important in the theory of randomized
search heuristics, since the unorganized progress of these algorithms often leads to
situations very similar to the coupon collector process.
The coupon collector process is the following simple randomized process. Assume that there are n types of coupons available. Whenever you buy a certain
product, you get one coupon with a type chosen uniformly at random from the
n types. How long does it take until you have a coupon of each type? We denote,
in this section, the random variable describing the first round after which we have
all types by Tn and call it coupon collecting time. Hence in simple words, this is
the number of rounds it takes to obtain all types.
As an easy example showing the coupon collector problem arising in the theory
of randomized search heuristics, let us regard how the randomized local search
(RLS) heuristic optimizes strictly monotonically increasing functions. The RLS
heuristic when maximizing a given function f : {0, 1}n → R starts with a random
search point. Then, in each iteration of the process, a single random bit is flipped
in the current solution. If this gives a solution worse than the current one (in terms
of f ), then the new solution is discarded. Otherwise, the process is continued from
this new solution.
Assume that f is strictly monotonically increasing, that is, flipping any 0-bit
to 1 increases the function value. Then the optimization process of RLS on f
strongly resembles a coupon collector process. In each round, we flip a random
bit. If this bit was 1 in our current solution, then nothing changes (we discard
the new solution as it has a smaller f -value). If this bit was 0, then we keep the
new solution, which now has one extra 1. Hence taking the 1- bits as coupons, in
each round we obtain a random coupon. This has no effect if we had this coupon
already, but is good if not.
We observe that the optimization time (number of solutions evaluated until
the optimal solution is found) of RLS on strictly monotonic functions is exactly
the coupon collecting time when we start with an initial stake of coupons that
follows a Bin(n, 21 ) distribution. This shows that the optimization time is at most
the classic coupon collector time (where we start with no coupons). See [DD16]
for a very precise analysis of this process.
The expectation of the coupon collecting time is easy to determine. Recall
P
from Section 4.2 the definition of the harmonic number Hn := nk=1 k1 .
Theorem 43 (Coupon collector, expectation). The expected time to collect all n
coupons is E[Tn ] = nHn = (1 + o(1))n ln n.
32
Proof. Given that we already have k different coupons for some k ∈ [0..n − 1], the
. By
probability that the next coupon is one that we do not already have, is n−k
n
the waiting time argument (Lemma 19), we see that the time Tn,k needed to obtain
n
a new coupon given that we have exactly k different ones, satisfies E[Tn,k ] = n−k
.
Pn−1
Clearly, the total time Tn needed to obtain all coupons is k=0 Tn,k . Hence, by
Pn−1
linearity of expectation (Lemma 20), E[Tn ] = k=0
E[Tn,k ] = nHn .
We proceed by trying to gain more information on Tn than just the expectation.
The tools discussed so far (and one to come in a later section) lead to the following
results.
• Markov’s inequality (Lemma 23) gives Pr[Tn ≥ λnHn ] ≤
1
λ
for all λ ≥ 1.
• Chebyshev’s inequality (Lemma 25) can be used to prove Pr[|Tn − nHn | ≥
6
π2
This builds on the fact (implicit in
εn] ≤ 6ε
2 for all ε ≥ π 2 ≈ 0.6079.
the proof above) that the coupon collector time is the sum of independent
Pn−1
geometric random variables Tn = k=0
Geom( n−k
). Hence the variance is
n
π 2 n2
Var[Tn ] = 6 .
n−1
), Witt’s Chernoff bound for geomet• Again exploiting Tn = k=0
Geom( n−k
n
ric random variables (Theorem 78) gives
P
Pr[Tn ≥ E[Tn ] + εn] ≤
exp(− 3ε2 )
π2
exp(− ε )
4
if ε ≤
if ε >
π2
6
π2
6
2
Pr[Tn ≤ E[Tn ] − εn] ≤ exp(− 3ε
)
π2
for all ε ≥ 0. See [Wit14] for the details.
Interestingly, asymptotically much stronger bounds can be derived by fairly
elementary means. These are the basis for Theorem 79, hence for that reason we
did not mention this result in the list above. The key idea is to not regard how the
number of coupons increases over time, but instead to regard the event that we
miss a particular coupon for some period of time. Note that the probability that
a particular coupon is not obtained for t rounds is (1 − n1 )t . By a union bound
argument (see Lemma 12), the probability that there is a coupon not obtained
within t rounds, and equivalently, that Tn > t, satisfies
Pr[Tn > t] ≤ n(1 − n1 )t .
Using the simple estimate of Lemma 1, we obtain the following (equivalent)
bounds.
33
Theorem 44 (Coupon collector, upper tail). For all ε ≥ 0,
Pr[Tn ≥ (1 + ε)n ln n] ≤ n−ε ,
Pr[Tn ≥ n ln n + εn] ≤ exp(−ε).
(25)
(26)
Surprisingly, prior to the following result from [Doe11], no good lower bound
for the coupon collecting time was published.
Theorem 45 (Coupon collector, lower tail). For all ε ≥ 0,
Pr[Tn ≤ (1 − ε)(n − 1) ln n] ≤ exp(−nε ),
Pr[Tn ≤ (n − 1) ln n − ε(n − 1)] ≤ exp(−eε ).
(27)
(28)
Theorem 45 was proven in [Doe11] by showing that the events of having a
coupon after a certain time are 1-negatively correlated. The following proof defers
this work to Lemma 71.
Proof. Let t = (1 − ε)(n − 1) ln n. For i ∈ [1..n], let Xi be the indicator random
variable for the event that a coupon of type i is obtained within the first t rounds.
Then Pr[Xi = 1] = 1 − (1 − n1 )t ≤ 1 − exp(−(1 − ε) ln n) = 1 − n−1+ε , where the
estimate follows from Corollary 4.
Since in the coupon collector process in each round j we choose a random set
Sj of cardinality 1, by Lemma 71 the Xi are negatively correlated. Consequently,
Pr[Tn ≤ (1 − ε)(n − 1) ln n] = Pr[∀i ∈ [1..n] : Xi = 1]
≤
n
Y
Pr[Xi = 1]
i=1
≤ (1 − n−1+ε )n ≤ exp(−nε )
by Lemma 1.
We may remark that good mathematical understanding of the coupon collector process not only is important because such processes directly show up in some
randomized algorithms, but also because it might give us the right intuitive understanding of other processes. Consider, for example, a run of the (1 + 1) EA on
some pseudo-Boolean function f : {0, 1}n → R with a unique global maximum.
The following intuitive consideration leads us to believe that the (1 + 1) EA
with high probability needs at least roughly n ln n2 iterations to find the optimum
of f : By the strong concentration of the binomial distribution, the initial search
point differs in at least roughly n2 bits from the global optimum. To find the global
optimum, it is necessary (but clearly not sufficient) that each of these missing bits
is flipped at least once in some mutation step. Now that the (1 + 1) EA in average
34
flips one bit per iteration, this looks like a coupon collector process started with an
initial stake of n2 coupons, so we expect to need at least roughly n ln n2 iterations
to perform the n ln n2 bit-flips necessary to have each missing bit flipped at least
once. Clearly, this argumentation is not rigorous, but it suggested to us the right
answer.
Theorem 46. The optimization time T of the (1 + 1) EA on any function f :
{0, 1}n → R with unique global maximum satisfies
Pr[T ≤ (1 − ε)(n − 1) ln n2 ] ≤ exp(−nε ).
Proof. By symmetry, we may assume that the unique global optimum of f is
(1, . . . , 1). Let t = (1 − ε)(n − 1) ln n2 . For all i ∈ [1..n], let Yi denote the event
that the i-th bit was zero in the initial search point and was not flipped in any
application of the mutation operator in the first t iterations. Let Xi = 1 − Yi.
Then Pr[Xi = 1] = 1 − 12 (1 − n1 )t ≤ 1 − n−1+ε . The events Xi are independent, so
we compute
Pr[T ≤ t] ≤ Pr[∀i ∈ [1..n] : Xi = 1]
=
n
Y
Pr[Xi = 1]
i=1
= (1 − n−1+ε )n = exp(−nε ).
We stated the above theorem to give a simple example how understanding the
coupon collector process can help understanding also randomized search heuristics
that do not directly simulate a coupon collecting process. We remark that the
theorem above is not best possible, in particular, it does not outrule an expected
optimization time of n ln n2 . In contrast it is known that the optimization time of
the (1 + 1) EA on the OneMax function is E[T ] ≥ en ln n − O(n), see [DFW11],
improving over the slightly weaker E[T ] ≥ en ln n − O(n log log n) from, independently, [DFW10] and [Sud13], where the latter holds for a much broader
class of mutation-based algorithms. Since it has been observed that OneMax
is the easiest function with unique global optimum for many mutation-based algorithms [DJW12, Sud13, Wit13], these results take over to all pseudo-Boolean
functions with unique global optimum.
10
Large Deviation Bounds
Often, we are not only interested in the expectation of some random variable, but
we need a bound that holds with high probability. We have seen in the proof of
35
Theorem 14 that such high-probability statements can be very useful: If a certain
bad event occurs in each iteration with a very small probability only, then a simple
union bound is enough to argue that this event is unlikely to occur even over a large
number of iterations. The better the original high-probability statement is, the
more iterations we can cover. For this reason, the tools discussed in this chapter
are among the most employed in the theory of randomized search heuristics.
Since computing the expectation often is easy, a very common approach is to
first compute the expectation of a random variable and then bound the probability
that the random variable deviates from this expectation by a too large amount.
The tools for this second step are called tail inequalities or large deviation inequalities, and this is the topic of this section. In a sense, Markov’s and Chebyshev’s
inequality discussed in Section 6 can be seen as large deviation inequalities as well,
but usually the term is reserved for exponential tail bounds.
A large number of large deviation bounds have been developed in the past.
They differ in the situations they are applicable to, but also in their sharpness.
Often, the sharpest bounds give expressions for the tail probability that are very
difficult to work with. Hence some experience is needed to choose a tail bound
that is not overly complicated, but sharp enough to give the desired result.
To give the novice to this topic some orientation, here is a short list of results
that are particularly useful and which are sufficient in many situations.
(a) The simple multiplicative Chernoff bounds (33) and (40) showing that for
sums of independent [0, 1] random variables, a constant factor deviation from
the expectation occurs only with probability negatively exponential in the
expectation.
(b) The additive Chernoff bounds of Theorems 53 and 55 showing that a sum
of n independent random variables deviates from the expectation by more
2
than an additive term of λ only with probability exp(−Ω( λn )).
(c) The fact that essentially all large deviation bounds can be used also with
a pessimistic estimate for the expectation instead of the precise expectation
(Section 10.1.8).
(d) The method of bounded differences (Theorem 72), which states that the
additive Chernoff bounds remain valid if X is functionally dependent of
independent random variables each having a small influence on X.
For the experienced reader, the following results may be interesting as they go
beyond what most introductions to tail bounds cover.
(a) In Section 10.2.2 we show that essentially all large deviation bounds usually
stated for sums of independent random variables are also valid for negatively correlated random variables. An important application of this result
36
are distributions arising from sampling without replacement or with partial
replacement.
(b) In Section 10.4, we present a number of large deviation bounds for sums of
independent geometrically distributed random variables. These seem to be
particularly useful in the analysis of randomized search heuristics, whereas
they are rarely used in classic randomized algorithms.
(c) In Theorem 73, we present a version of the bounded differences method which
only requires that the t-th random variable has a bounded influence on the
expected outcome resulting from variables t + 1 to n. This is much weaker
than the common bounded differences assumption that each random variable, regardless how we condition on the remaining variables, has a bounded
influence on the result. We feel that this new version (which is an easy consequence of known results) may be very useful in the analysis of iterative
improvement heuristics. In particular, it may lead to elementary proofs for
results which so far could only be proven via tail bounds for martingales.
10.1
Chernoff Bounds for Sums of Independent Bounded
Random Variables
In this longer subsection, we assume that our random variable of interest is the
sum of independent random variables, each taking values in some bounded range.
The bounds presented below are all known under names like Chernoff or
Hoeffding bounds, referring to the seminal papers by Chernoff [Che52] and
Hoeffding [Hoe63]. Since the first bounds of this type were proven by Bernstein [Ber24]—via the exponential moments method that is used in essentially all
proofs of such results, see Section 10.1.7—the name Bernstein inequalities would be
more appropriate. We shall not be that precise and instead use the most common
name Chernoff inequalities for all such bounds.
For the readers’ convenience, we shall not be shy to write out also minor reformulations of some results. We believe that it helps a lot to have seen them and
we think that it is convenient, both for using the bounds or for referring to them,
if all natural version are visible in the text.
10.1.1
Multiplicative Chernoff Bounds for the Upper Tail
The multiplicative Chernoff bounds presented in this and the next section bound
the probability to deviate from the expectation by at least a given factor. Since
in many algorithmic analyses we are interested only in the asymptotic order of
magnitude of some quantity, a constant-factor deviation can be easily tolerated
37
and knowing that larger deviations are very unlikely is just what we want to
know. For this reason, the multiplicative Chernoff bounds are often the right tool.
The following theorem collects a number of bounds for the upper tail, that is,
for deviations above the expectation.
Theorem 47. Let X1 , . . . , Xn be independent random variables taking values in
P
[0, 1]. Let X = ni=1 Xi . Let δ ≥ 0. Then
Pr[X ≥ (1 + δ)E[X]]
≤
1
1+δ
!(1+δ)E[X]
!E[X]
n − E[X]
n − (1 + δ)E[X]
!n−(1+δ)E[X]
(29)
eδ
≤
= exp(−((1 + δ) ln(1 + δ) − δ)E[X])
(1 + δ)1+δ
!
δ 2 E[X]
≤ exp −
2 + 32 δ
!
min{δ 2 , δ}E[X]
≤ exp −
,
3
where the bound in (29) is read as 0 for δ >
For δ ≤ 1, equation (32) simplifies to
n−E[X]
E[X]
Pr[X ≥ (1 + δ)E[X]] ≤ exp
and as ( E[X]
)n for δ =
n
δ 2 E[X]
.
−
3
!
(30)
(31)
(32)
n−E[X]
.
E[X]
(33)
The first and strongest bound (29) was first stated explicitly by
Hoeffding [Hoe63]. It improves over Chernoff’s [Che52] tail bounds in particular by not requiring that the Xi are identically distributed. Hoeffding also shows
that (29) is the best bound that can be shown via the exponential moments methods under the assumptions of Theorem 47.
For E[X] small, say E[X] = o(n) when taking a view asymptotic in n → ∞,
the second bound (30) is easier to use, but essentially as strong as (29). More
precisely, it is larger only by a factor of (1 + o(1))E[X] , since we estimated
n − E[X]
n − (1 + δ)E[X]
!n−(1+δ)E[X]
δE[X]
= 1+
n − (1 + δ)E[X]
!n−(1+δ)E[X]
≤ eδE[X] (34)
using Lemma 1.
3δ2
Equation (31) is derived from (30) by noting that (1 + δ) ln(1 + δ) − δ ≥ 6+2δ
holds for all δ ≥ 0, see Theorem 2.3 and Lemma 2.4 in McDiarmid [McD98].
Equations (32) and (33) are trivial simplifications of (31).
38
f (x) = (1 + x) ln(1 + x) − x
x2
f (x) = 2+2x/3
f (x) = min{x2 , x}/3
f (x)
1.5
1
0.5
0
0.5
1
1.5
2
2.5
x
Figure 3: Visual comparison of the bounds (30), (31), (32). Depicted is the term
f (x) leading to the bound Pr[X ≥ (1 + x)E[X]] ≤ exp(−f (x)E[X]).
In general, to successfully use Chernoff bound in one’s research, it greatly
helps to look a little behind the formulas and understand their meaning. Very
roughly speaking, we can distinguish three different regimes relative to δ, namely
that the tail bound is of order exp(−Θ(δ log(δ)E[X])), exp(−Θ(δE[X])), and
exp(−Θ(δ 2 E[X])).
Superexponential regime: Equation (30) shows a tail bound of order
−Θ(δE[X])
δ
= exp(−Θ(δ log(δ)E[X])), where the asymptotics are for δ → ∞. In
this regime, the deviation δE[X] from the expectation E[X] is much larger than
the expectation itself. It is not very often that we need to analyze such large deviations, so this Chernoff bound is rarely used. It can be useful in the analysis of
evolutionary algorithms with larger offspring populations, where the most extreme
behavior among the offspring can deviate significantly from the expected behavior.
See [DK15, DD17] for examples how to use Chernoff bounds for large deviations
occurring in the analysis of the (1 + λ) EA. Note that in [JJW05], the first theoretical work on the (1 + λ) EA, such Chernoff bounds could have been used as
well, but the authors found it easier to directly estimate the tail probability by
estimating binomial coefficients.
Weaker forms of (30) are
Pr[X ≥ (1 + δ)E[X]] ≤
e
(1 + δ)
Pr[X ≥ (1 + δ)E[X]] ≤
e
δ
39
!(1+δ)E[X]
!δE[X]
,
,
(35)
(36)
where the first one is stronger for those values of δ where the tail probability is
less than one (that is, δ > e − 1).
It is not totally obvious how to find a value for δ ensuring that ( δe )δ is less than
a desired bound. The following lemma solves this problem.
Lemma 48. Let t ≥ ee
1/e
≈ 4.24044349... . Let δ =
Proof. We compute δ ln δe = δ ln
ln t
ln t
e ln( e ln
)
ln t
ln t
.
ln t
ln( e ln
)
ln t
Then ( eδ )δ ≤ 1t .
≥ δ ln( e lnlnlnt t ) = ln t.
We use this estimate to bound the number of bits flipped in an application of
the standard-bit mutation operator. The standard-bit mutation operator takes as
input a bit-string x ∈ {0, 1}n of length n and returns the bit-string y ∈ {0, 1}n
obtained from flipping each bit of x independently at random with probability n1 ,
that is, for each i ∈ [1..n] independently we have Pr[yi = 1 − xi ] = n1 and Pr[yi =
xi ] = 1 − n1 . Let us denote by H(x, y) := |{i ∈ [1..n] | xi 6= yi } the Hamming
distance of x and y, that is, the number of bits x and y differ in. By linearity
of expectation, it is clear that the expected Hamming distance of x and y is
E[H(x, y)] = 1, see Lemma 21. Using Chernoff bounds, we now give an upper
bound on how far we can exceed this value. Such arguments are often useful
in the analysis of evolutionary algorithms, see, e.g., Lemma 26 in [DK15] for an
example.
Lemma 49. (a) Let x ∈ {0, 1}n and y be obtained from x via standard-bit mutation with mutation rate αn . Then Pr[H(x, y) ≥ k] ≤ ( eα
)k .
k
(b) Let 0 < p ≤ exp(−α exp( 1e )).
Let k ≥ kp :=
ln(1/p)
ln
Pr[H(x, y) ≥ k] ≤ p.
ln(1/p1/α )
e ln ln(1/p1/α )
.
Then
(c) Let T ∈ N and 0 < p ≤ T1 exp(−α exp( 1e )). Let y1 , . . . , yT be obtained from
ln(T /p)
.
x1 , . . . , xT , respectively, via standard-bit mutation. Let k ≥ ln((T
/p)1/α )
ln
e ln ln((T /p)1/α )
Then
Pr[∃i ∈ [1..T ] : H(xi , yi) ≥ k] ≤ p.
Proof. Note that H(x, y) ∼ Bin(n, αn ), hence H(x, y) can be written as a sum of n
independent random variables X1 , . . . , Xn with Pr[Xi = 1] = αn and Pr[Xi = 0] =
1 − αn for all i ∈ [1..n]. Since E[H(x, y)] = α, we can apply equation (35) with
(δ + 1) = αk . This proves (a).
For part (b), we use Lemma 48 and compute Pr[H(x, y) ≥ k] ≤ Pr[H(x, y) ≥
kp ] ≤ (( kpe/α )kp /α )α ≤ (p1/α )α = p. Similarly, for (c) we obtain Pr[H(xi, yi ) ≥ k] ≤
p
and use the union bound (Lemma 12).
T
40
Observe that the bounds in Lemma 49 are independent of n. Also, the bounds
in parts (b) and (c) depend only mildly on α. By applying part (c) with p = n−c1
and T = nc2 , we see that the probability that an evolutionary algorithm using
standard-bit mutation with rate αn , α a constant, flips more than (c1 +c2 +o(1)) lnlnlnnn
bits in any of the first nc2 applications of the mutation operator, is at most n−c1 .
Exponential regime: When δ = Θ(1), then all bounds give a tail probability of order exp(−Θ(δE[X])). Note that the difference between these bounds
often is not very large. For δ = 1, the bounds in (30), (31), and (32) become
(0.67957...)E[X], (0.68728...)E[X] , and (0.71653...)E[X] , respectively. So there is often no reason to use the unwieldy equation (30).
We remark that also for large δ, where the bound (30) gives the better asymptotics exp(−Θ(δ log(δ)E[X])), one can, with the help of Section 10.1.8 resort to
the easier-to-use bounds (31) and (32) when the additional logarithmic term is
not needed. For example, when X is again the number of bits that flip in an
application of the standard-bit mutation operator with mutation rate p = αn , then
for all c > 0 and n ∈ N with c ln n ≥ α equation (32) with E[X] ≤ µ+ := c ln n
and the argument of Section 10.1.8 gives Pr[X ≥ 2c ln n] = Pr[X ≥ (1 + 1)µ+ ] ≤
exp(− 31 µ+ ) = n−c/3 , which in many applications is fully sufficient.
A different way of stating an exp(−Θ(δE[X])) tail bound, following directly
from applying (35) for δ ≥ 2e − 1, is the following.
Corollary 50. Under the assumptions of Theorem 47, we have
Pr[X ≥ k] ≤ 2−k
(37)
for all k ≥ 2eE[X].
Sub-exponential regime: When δ ≤ 1, the tail probability is of order
exp(−Θ(δ 2 E[X])). We need δ to be at least of order (E[X])−1/2 to obtain useful statements. Note that for E[X] close to n2 , Theorem 53 below gives slightly
stronger bounds.
10.1.2
Multiplicative Chernoff Bounds for the Lower Tail
In principle, of course, there is no difference between bounds for the upper and
lower tail. If in the situation of Theorem 47 we set Yi := 1 − Xi , then the Yi
are independent random variables taking values in [0, 1] and any upper tail bound
P
for X turns into a lower tail bound for Y := ni=1 Yi via Pr[Y ≤ t] = Pr[X ≥
n − t]. However, since this transformation also changes the expectation, that is,
E[Y ] = n − E[X], a convenient bound like (33) becomes Pr[Y ≤ (1 − δ)E[Y ]] ≤
E[Y ] 2
exp(− 13 (1 + δ n−E[Y
) (n − E[Y ])).
]
For this reason, usually the tail bounds for the lower tail are either proven
completely separately (however, using similar ideas) or are derived from simplifying
41
f (x) = −x − (1 − x) ln(1 − x)
2
f (x) = x2
0.8
f (x)
0.6
0.4
0.2
0
0.2
0.4
0.6
0.8
x
Figure 4: Visual comparison of the bounds (39) and (40). Depicted is the term
f (x) leading to the bound Pr[X ≤ (1 − x)E[X]] ≤ exp(−f (x)E[X]).
the result of applying the above symmetry argument to (29). Either approach can
be used to show the following bounds. As a visible result of the asymmetry of
the situation for upper and lower bounds, note the better constant of 21 in the
exponent of (40) as compared to the 13 in (33).
Theorem 51. Let X1 , . . . , Xn be independent random variables taking values in
P
[0, 1]. Let X = ni=1 Xi . Let δ ∈ [0, 1]. Then
Pr[X ≤ (1 − δ)E[X]] ≤
1
1−δ
!(1−δ)E[X]
!E[X]
e−δ
≤
(1 − δ)1−δ
!
δ 2 E[X]
≤ exp −
.
2
n − E[X]
n − (1 − δ)E[X]
!n−(1−δ)E[X]
(38)
(39)
(40)
For the not so interesting boundary cases, recall our definition 00 := 1. The
first bound (38) follows from (29) by regarding the random variables Yi := 1 − Xi.
P
, we compute
Setting Y = ni=1 Yi and δ ′ = δ E[X]
E[Y ]
Pr[X ≤ (1 − δ)E[X]] = Pr[Y ≥ (1 + δ ′ )E[Y ]]
≤
1
1 + δ′
!(1+δ′ )E[Y ]
42
n − E[Y ]
n − (1 + δ ′ )E[Y ]
!n−(1+δ′ )E[Y ]
=
n − E[X]
n − (1 − δ)E[X]
!n−(1−δ)E[X]
1
1−δ
!(1−δ)E[X]
.
Obviously, in an analoguous fashion, (29) can be derived from (38), so both
bounds are equivalent. Equation (39) follows from (38) using an elementary estimate analoguous to (34). Equation (40) follows from (39) using elementary
calculus, see, e.g., the proof of Theorem 4.5 in [MU05].
Theorems 47 and 51 in particular show that constant-factor deviations from
the expectation appear only with exponentially small probability.
Corollary 52. Let X1 , . . . , Xn be independent random variables taking values in
P
[0, 1]. Let X = ni=1 Xi . Let δ ∈ [0, 1]. Then
h
i
Pr |X − E[X]| ≥ δE[X] ≤ 2 exp
10.1.3
δ 2 E[X]
−
.
3
!
Additive Chernoff Bounds
We now present a few bounds for the probability that a random variable deviates
from its expectation by an additive term independent of the expectation. The
advantage of such bounds is that they are identical for upper and lower tails and
that they are invariant under additive rescalings.
Already from Theorem 47, equation (29), by careful estimates (see, e.g.,
Hoeffding [Hoe63]) and exploiting the obvious symmetry, we obtain the following
estimates. When E[X] is close to n2 , this additive Chernoff bound gives (slightly)
stronger results than the simplified bounds of Theorems 47 and 51.
Theorem 53. Let X1 , . . . , Xn be independent random variables taking values in
P
[0, 1]. Let X = ni=1 Xi . Then for all λ ≥ 0,
Pr[X ≥ E[X] + λ] ≤ exp
Pr[X ≤ E[X] − λ] ≤ exp
2λ2
,
−
n
!
2λ2
−
.
n
!
(41)
(42)
A second advantage of additive Chernoff bounds is that they are often very
easy to apply. As a typical application in evolutionary computation, let us regard
the number of ones kxk1 in a random search point x ∈ {0, 1}, which could, e.g.,
be an initial solution in an evolutionary algorithm.
Lemma 54. Let x ∈ {0, 1}n be chosen uniformly at random. Then
Pr
n
2λ2
kxk1 −
.
≥ λ ≤ 2 exp −
2
n
43
!
Proof. Note that if x ∈ {0, 1}n is uniformly distributed, then the xi are independent random variables uniformly distributed in {0, 1}. Consequently, kxk1 =
Pn
n
i=1 xi is a sum of independent random variables and E[kxk1 ] = 2 . The claim now
follows immediately from applying Theorem 53 to the events “kxk1 ≥ E[kxk1 ]+ λ”
and “kxk1 ≤ E[kxk1 ] − λ”.
The following theorem, again due to Hoeffding [Hoe63], non-trivially extends
Theorem 53 by allowing the Xi to take values in arbitrary intervals [ai , bi ].
Theorem 55. Let X1 , . . . , Xn be independent random variables. Assume that each
P
Xi takes values in a real interval [ai , bi ] of length ci := bi − ai . Let X = ni=1 Xi .
Then for all λ > 0,
Pr[X ≥ E[X] + λ] ≤ exp
Pr[X ≤ E[X] − λ] ≤ exp
2λ2
!
− Pn
,
(43)
.
(44)
2
i=1 ci
!
2
2λ
− Pn
2
i=1 ci
For comparison, we now reformulate Theorems 47 and 51 as additive bounds.
There is no greater intellectual challenge hidden, but we feel that it helps to have
seen these bounds at least once. Note that, since the resulting bounds depend on
the expectation, we need that the Xi take values in [0, 1]. In other words, different
from the bounds presented so far in this subsection, the following bounds are not
invariant under additive rescaling and are not symmetric for upper and lower tails.
Theorem 56 (equivalent to Theorem 47). Let X1 , . . . , Xn be independent random
P
variables taking values in [0, 1]. Let X = ni=1 Xi . Let λ ≥ 0. Then
Pr(X ≥ E[X] + λ)
≤
E[X]
E[X] + λ
≤ eλ
≤ exp
≤ exp
!E[X]+λ
E[X]
E[X] + λ
n − E[X]
n − E[X] − λ
!E[X]+λ
= exp
!n−E[X]−λ
(45)
!
λ
+λ
− (E[X] + λ) ln 1 +
E[X]
λ2
−
2E[X] + 23 λ
(
)!
1
λ2
− min
,λ ,
3
E[X]
!
!
(46)
(47)
(48)
)n for λ =
where the bound in (45) is read as 0 for λ > n − E[X] and as ( E[X]
n
n − E[X]. For λ ≤ E[X], equation (48) simplifies to
Pr[X ≥ E[X] + λ] ≤ exp
44
λ2
−
.
3E[X]
!
(49)
Theorem 57 (equivalent to Theorem 51). Let X1 , . . . , Xn be independent random
P
variables taking values in [0, 1]. Let X = ni=1 Xi . Let δ ∈ [0, 1]. Then
Pr[X ≤ E[X] − λ] ≤
E[X]
E[X] − λ
!E[X]−λ
!n−E[X]+λ
!E[X]−λ
E[X]
≤e
E[X] − λ
!
λ2
.
≤ exp −
2E[X]
−λ
10.1.4
n − E[X]
n − E[X] + λ
(50)
(51)
(52)
Chernoff Bounds Using the Variance
There are several versions of Chernoff bounds that take into account the variance.
In certain situations, they can give significantly stronger bounds than the estimates
discussed so far. Hoeffding [Hoe63] proves essentially the following result.
Theorem 58. Let X1 , . . . , Xn be independent random variables such that Xi ≤
P
P
E[Xi ] + 1 for all i = 1, . . . , n. Let X = ni=1 Xi . Let σ 2 = ni=1 Var(Xi ) = Var[X].
Then for all λ ≥ 0,
Pr[X ≥ E[X] + λ]
≤
≤ exp
= exp
≤ exp
≤ exp
2
σ
) n+σ
2
!− 1− λ n !n
λ ( n ) n+σ2
1−
n
!
!
!!
2
λ
σ
ln 1 + 2 − 1
−λ 1+
λ
σ
!
!
!!
λ
λ
λ
2
1 + 2 ln 1 + 2 − 2
−σ
σ
σ
σ
!
2
λ
− 2 2
2σ + 3 λ
(
)!
1
λ2
− min
,λ ,
3
σ2
λ
1+ 2
σ
!−(1+
λ
σ2
(53)
(54)
(55)
(56)
2
σ
n
where the term in (53) is understood to mean 0 when λ > n and ( n+σ
when
2)
λ = n.
The estimate from (53) to (54) is non-trivial and can be found, e.g., in
Hoeffding [Hoe63]. From (54) we derive (55) in the same way that was used
to derive (31) from (30).
By replacing Xi with −Xi , we obtain the analoguous bounds for the lower tail.
45
Corollary 59. If the condition Xi ≤ E[Xi ] + 1 in Theorem 58 is replaced by
Xi ≥ E[Xi ] − 1, then Pr[X ≤ E[X] − λ] satisfies the estimates of equations (53)
to (56).
As discussed in Hoeffding [Hoe63], the bound (54) is the same as inequality (8b)
in Bennett [Ben62], which is stronger than the bound (55) due to Bernstein [Ber24]
and the bound of exp(− 12 λ arcsinh( 2σλ2 )) due to Prokhorov [Pro56].
In comparison to the additive version of the usual Chernoff bounds for the upper
tail (Theorem 56), very roughly speaking, we see that the Chernoff bounds working
with the variance allow to replace the expectation of X by its variance. When the
Xi are binary random variables with Pr[Xi = 1] small, then E[X] ≈ Var[X]
and there is not much value of using Theorem 58. For this reason, Chernoff
bounds taking into account the variance have not been used a lot in the theory
of randomized search heuristics. They can, however, be convenient. For example,
when a search point y ∈ {0, 1}n is obtained from a given x ∈ {0, 1}n via standardbit mutation with probability p. Assume for simplicity that we are interested
in estimating the number of ones in y (the same argument would hold for the
Hamming distance of y to some other search point z ∈ {0, 1}∗, e.g., a unique
P
optimum). Now the number of ones in y is simply X = ni=1 yi and X is a sum of
independent binary random variables. The expectation of X may be big, namely
if x has many ones, but the variance is small (assuming that p is small), namely
Var[X] = np(1 − p). Consequently, here the Chernoff bounds of this subsection
are very useful. See, e.g., [DGWY17] for an example where this problem appeared
in a recent research paper.
When not too precise bounds are needed, looking separately at the number
of zeros and ones of x that flip (and bounding these via simple Chernoff bounds)
is a way to circumvent the use of Chernoff bounds taking into account the variance. Several research works follow this approach despite the often more technical
computations.
Chernoff bounds using the variance can also be useful in ant colony algorithms
and estimation of distribution algorithms, where again pheromone values or frequencies close to 0 or 1 can lead to a small variance. See [Wit17] for a recent
example.
The bounds of Theorem 58 can be written in a multiplicative form, e.g.,
Pr[X ≥ (1 + δ)E[X]]
!−(1+ δE[X] )
δE[X]
≤
1+
σ2
!
δ 2 E[X]2
≤ exp − 2 2
.
2σ + 3 δE[X]
σ2
σ2
n+σ 2
δE[X]
1−
n
!−(1− δE[X] )
n
n
n+σ 2
!n
(57)
(58)
46
This is useful when working with relative errors, however, it seems that unlike for
some previous bounds (compare, e.g., (30) and (46)) the multiplicative forms are
not much simpler here.
Obviously, the case that all Xi satisfy Xi ≤ E[Xi ] + b for some number b
(instead of 1) can be reduced to the case b = 1 by dividing all random variable
by b. For the reader’s convenience, we here state the resulting Chernoff bounds.
Theorem 60 (equivalent to Theorem 58 and Corollary 59). Let X1 , . . . , Xn be
independent random variables. Let b be such that Xi ≤ E[Xi ] + b for all i =
P
P
1, . . . , n. Let X = ni=1 Xi . Let σ 2 = ni=1 Var(Xi ) = Var[X]. Then for all λ ≥ 0,
!−(1+ bλ )
σ2
nb2 +σ 2
2
≤ exp
!−(1− λ ) nb !n
nb nb2 +σ 2
λ
1−
nb
!
!
!!
σ2
bλ
1+
ln 1 + 2 − 1
bλ
σ
!
2
λ
− 2
σ (2 + 23 σbλ2 )
≤ exp
1
λ2 λ
− min
,
3
σ2 b
bλ
Pr[X ≥ E[X] + λ] ≤
1+ 2
σ
λ
≤ exp −
b
σ2
(
)!
,
(59)
(60)
(61)
(62)
2
where the term in (59) is understood to mean 0 when λ > nb and ( nb2σ+σ2 )n when
λ = nb.
When we have Xi ≥ E[Xi ] − b instead of Xi ≤ E[Xi ] + b for all i = 1, . . . , n,
then the above estimates hold for Pr[X ≤ E[X] − λ].
10.1.5
Relation Between the Different Chernoff Bounds
We proceed by discussing how the bounds presented so far are related. The main
finding will be that the Chernoff bounds depending on the variance imply all
other bounds discussed so far with the exception of the additive Chernoff bound
for different variable ranges (Theorem 55).
Surprisingly, this fact is not stated that explicitly in Hoeffding’s paper [Hoe63].
More precisely, in [Hoe63] the analogue of Theorems 58 and 60 uses the additional
assumption that all Xi have the same expectation. Since this assumption is not
made for the theorems not involving the variance, Hoeffding explicitely states that
the latter are stronger in this respect (see the penultimate paragraph of Section 3
of [Hoe63].
It is however quite obvious that the common-expectation assumption can be
easily removed. From random variables with arbitrary means we can obtain random variables all having mean zero by subtracting their expectation. This operation does not change the variance and does not change the distribution of X −E[X].
47
Consequently, Hoeffding’s result for variables with identical expectations immediately yields our version of this result (Theorem 58 and 60). Theorem 58 implies
Theorem 47 via the equivalent version of Theorem 56, see again the penultimate
paragraph of Section 3 of [Hoe63].
Consequently, the first (strongest) bound in Theorem 58 (equivalently the first
bound of Theorem 60) implies the first (strongest) bound of Theorem 47, which
is equivalent to the first (strongest) bound in Theorem 51. Essentially all other
bounds presented so far can be derived from these main theorems via simple,
sometimes tedious, estimates. The sole exception is Theorem 55, which can lead to
significantly stronger estimates when the random variables have ranges of different
size.
As an example, let X1 , . . . , Xn be independent random variables such that
X1 , . . . , Xn−1 take the values 0 and (n − 1)−1/2 with equal probability 12 and such
P
that Xn takes
the values 0 and 1 with equal probability 21 . Let X = ni=1 Xi . Then
√
E[X] = 12 ( n − 1 + 1). Theorem 55, taking ci = (n − 1)−1/2 for i ∈ [1..n − 1] and
cn = 1, yields the estimate
Pr[X ≥ E[X] + λ] ≤ exp
Note that Var[X] =
1
2
2λ2
− Pn
2
i=1 ci
!
= exp(−λ2 ).
(63)
=: σ 2 . Consequently, the strongest Chernoff bound of
λ
nσ 2
Theorem 58, equation (53), gives an estimate larger than (1 + σλ2 )−(1+ σ2 ) n+σ =
exp(−Θ(λ log λ)). Consequently, in this case Theorem 55 gives a significantly
stronger estimate than Theorem 58.
10.1.6
Tightness of Chernoff Bounds, Lower Bounds for Deviations
As a very general and not at all precise rule of thumb, we can say that often the
sharpest Chernoff bound presented so far gives an estimate for the tail probability
that is near-tight. This is good to know from the perspective of proof design, since
it indicates that failing to prove a desired statement usually cannot be overcome by
trying to invent sharper Chernoff bounds. We shall not try to make this statement
precise.
However, occasionally, we also need lower bounds for the deviation from the
expectation as a crucial argument in our analysis. For example, when generating
several offspring independently in parallel, as, e.g., in a (1 + λ) EA, we expect the
best of these to be significantly better than the expectation and the efficiency of
the algorithm relies on such large deviations from the expectation.
Such lower bounds for deviations seem to be harder to work with. For this
reason, we only briefly give some indications how to handle them and refer the
reader to the literature. We note that there is a substantial body of mathematics
48
literature on this topic, see, e.g., [Nag01] and the references therein, which however
is not always easy to use for algorithmic problems.
Estimating binomial coefficients: While more advanced methods are available, surprisingly often estimating the sum of binomial coefficients arising in the
expression of the tail probability is a good approach. In the theory of randomized search heuristics, this approach was used, among others, in the analysis of the
(1+λ) EA in [JJW05, DK15, GW17, DGWY17] and the (1+(λ, λ)) GA in [DD17].
The following elementary bound was shown in [Doe14].
Lemma 61. Let n ∈ N and X ∼ Bin(n,
1
).
2
Then Pr X ≤ E[X] −
1
2
q
E[X] ≥ 18 .
Approximation via the normal distribution: The generic approach of approximating binomial distributions via normal distributions, surprisingly, is not
often used in the theory of randomized search. In [OW15], it was used to give
an elegant proof of the fact that a binomial random variable
X with p ≤ 12 with
q
constant probability deviates from its expectation by Ω( E[X]). This matches
the case of δ = Θ(E[X]−1/2 ) in (40).
Lemma 62. Let X ∼ Bin(n, p) with p ≤ 21 . Then Pr[X ≤ E[X] −
Pr[X ≥ E[X] +
q
1
E[X] ]
2
are both Ω(1).
q
1
E[X] ]
2
and
In [dPdLDD15],
the normal approximation was used to show that the best of
√
n
k ∈ ω(n) ∩ o( n) independent random
q initial search points in {0, 1} with probability 1 − o(1) has a distance of n2 − n2 (ln k − 12 ln ln k ± ck ) from the optimum,
where ck is an arbitrary sequence tending to infinity.
Order statistics: The last example actually says something about the maximum order statistics of k independent Bin(n, 12 ) random variables. In general, the
maximum order statistics is strongly related to lower bounds for tail probabilities
as the following elementary argument (more or less explicit in all works on the
(1 + λ) EA) shows: Let X1 , . . . , Xλ be independent random variables following the
same distribution. Let Xmax = max{Xi | i ∈ [1..λ]}. Then
Pr[X ∗ ≥ D] ≤ λ Pr[X1 ≥ D];
Pr[X ∗ ≥ D] = 1 − (1 − Pr[X1 ≥ D])λ ≥ 1 − exp(λ Pr[X1 ≥ D]).
Consequently, Pr[X ∗ ≥ D] is constant if and only if Pr[X1 ≥ D] = Θ( λ1 ).
For the maximum order statistics of binomially distributed random variables
with small success probability, Gießen and Witt [GW17] proved the following result
and used it in the analysis of the (1 + λ) EA.
49
Lemma 63. Let α ≥ 0 and c > 0 be constants. Let n ∈ N and let all of the
following asymptotics be for n → ∞. Let k = n(ln n)−α and λ = ω(1). Let Xmax
be the maximum of λ independent random variables with distribution Bin(k, nc ).
1
ln λ
Then E[Xmax ] = (1 ± o(1)) 1+α
.
ln ln λ
Extremal situations: Occasionally, it is desirable to understand which situation gives the smallest or the largest deviations. For example, let X1 , . . . , Xn be
independent binary random variables with expectations E[Xi ] = pi . Then it could
P
be useful to know that X = ni=1 Xi deviates most (in some suitable sense) from
its expectation when all pi are 12 . Such statements can be made formal and can be
proven with the notions of majorization and Schur-convexity. We refer to [Sch00]
for a nice treatment of this topic. Such arguments have been used to analyze
estimation-of-distribution algorithms in [SW16].
Staying on one side of the expectation and Feige’s inequality: When it
suffices to know that with reasonable probability we stay (more or less) on one
side of the expectation, then the following results can be useful.
A very general bound is Feige’s inequality [Fei06], which has found applications
in the analysis of randomized search heuristics, among others, in [ST12, DL15,
CDEL16, LN17].
Lemma 64 (Feige’s inequality). Let X1 , . . . , Xn be independent non-negative ranP
dom variables with expectations µi := E[Xi ] satisfying µi ≤ 1. Let X = ni=1 Xi .
Then
δ
1
, δ+1
}.
Pr[X ≤ E[X] + δ] ≥ min{ 13
For binomial distributions, we have stronger guarantees. Besides bounds comparing the binomial distribution with its normal approximation [Slu77], the following specific bounds are known.
Lemma 65. Let n ∈ N, p ∈ [0, 1], and k = ⌊np⌋. Let X ∼ Bin(n, p).
(a) If
1
n
< p, then Pr[X ≥ E[X]] > 41 .
(b) If 0.29/n ≤ p < 1, then Pr[X > E[X]] ≥ 14 .
1
,
n
(c) If
1
n
≤p≤1−
(d) If
1
n
≤ p < 1, then Pr[X > E[X]] >
then Pr[X ≥ E[X]] ≥
50
1
2
−
√
1 √ np(1−p)
√
.
2 2
np(1−p)+1+1
q
n
.
2πk(n−k)
(e) If
1
n
≤ p < 1 − n1 , then Pr[X > E[X] + 1] ≥ 0.037.
Surprisingly, all these results are quite recent. Bound (a) from [GM14] appears
to be the first general result of this type at all.1 It was followed up by estimate (c)
from [PR16], which gives stronger estimates when np(1 − p) > 8. Result (d)
from [Doe17] is the only one to give a bound tending to 12 for both np and n(1 − p)
tending to infinity. Estimates (b) and (e) are also from [Doe17]. A lower bound
for the probability of exceeding the expectation by more than one, like (e), was
needed in the analysis of an evolutionary algorithm with self-adjusting mutation
rate (proof of Lemma 3 of the extended version of [DGWY17]).
10.1.7
Proofs of the Chernoff Bounds
As discussed in the previous subsection, all Chernoff bounds stated so far can
be derived from the strongest bounds of Theorem 55 or 58 via elementary estimates that have nothing to do with probability theory. We shall not detail these
estimates—the reader can find them all in the literature, e.g., in [Hoe63]. We shall,
however, sketch how to prove these two central inequalities (43) and (53). One reason for this is that we can then argue in Section 10.2.2 that these proofs (and thus
also all Chernoff bounds presented so far) in fact not only hold for independent
random variables, but also negatively correlated ones.
A second reason is that occasionally, it can be profitable to have this central
argument ready to prove Chernoff bounds for particular distributions, for which
the classic bounds are not applicable or do not give sufficient results. This has
been done, e.g., in [DFW11, OW11, OW12, DJWZ13, LW14, Wit14, BLS14].
The central step in almost all proofs of Chernoff bounds, going back to Bernstein [Ber24], is the following one-line argument. Let h > 0. Then
Pr[X ≥ t] ≤ Pr[ehX ≥ eht ] ≤
n
Y
E[ehX ]
−ht
E[ehXi ].
=
e
eht
i=1
Here the first inequality simply stems from the fact that the function x 7→
ehx is monotonically increasing. The second inequality is Markov’s inequality
(Lemma 23) applied to the (non-negative) random variable ehX . The last equality
exploits the independence of the Xi , which carries over to the ehXi .
It now remains to estimate E[ehXi ] and choose h as to minimize the resulting
expression. We do this exemplarily for the case that all Xi take values in [0, 1] and
that E[X] < t < n. Since the exponential function is convex, E[ehXi ] is maximized
(which is the worst-case for your estimate) when Xi is concentrated on the values
For p ∈ [ n1 , 12 ], this result follows from the proof of Lemma 6.4 in [RT11]. The lemma itself
only states the bound Pr[X ≥ E[X]] ≤ min{p, 14 } for p ≤ 21 . The assumption p ≤ 12 appears to
be crucial for the proof.
1
51
0 and 1, that is, we have Pr[Xi = 1] = E[Xi ] and Pr[Xi = 0] = 1 − E[Xi ]. In
this case, E[ehXi ] = (1 − E[Xi ])e0 + E[Xi ]eh . By the inequality of arithmetic and
geometric means, we compute
n
Y
i=1
E[ehXi ] ≤
≤
≤
n
Y
(1 − E[Xi] + E[Xi ]eh )
i=1
n
1X
(1 − E[Xi ] + E[Xi ]eh )
n i=1
1
(n − E[X] + E[X]eh )
n
!n
!n
.
This gives the tail estimate Pr[X ≥ t] ≤ e−ht ( n1 (n − E[X] + E[X]eh ))n , which is
minimized by taking
!
(n − E[X])t
h = ln
,
(n − t)E[X]
which then gives the strongest multiplicative Chernoff bound (29) by rewriting
t = (1 + δ)E[X].
10.1.8
Chernoff Bounds with Estimates for the Expectation
Often we do not know the precise value of the expectation or it is tedious to compute it. In such cases, we can exploit the fact that all Chernoff bounds discussed
in this work are valid also when the expectation is replaced by an upper or lower
bound for it. This is obvious for many bounds, e.g., from equations (41), (42),
and (40) we immediately derive the estimates
2λ2
Pr[X ≥ µ + λ] ≤ Pr[X ≥ µ + λ] ≤ exp −
,
n
!
2
2λ
Pr[X ≤ µ− − λ] ≤ Pr[X ≤ µ − λ] ≤ exp −
,
n
!
δ2µ
−
≤ exp
Pr[X ≤ (1 − δ)µ ] ≤ Pr[X ≤ (1 − δ)µ] ≤ exp −
2
!
+
δ 2 µ−
−
2
!
for all µ+ ≥ E[X] =: µ and µ− ≤ E[X].
This is less obvious for a bound like Pr[X ≥ (1 + δ)µ+ ] ≤ exp(− 31 δ 2 µ+ ), since
now also the probability of the tail event decreases for increasing µ+ . However,
also for such bounds we can replace E[X] by an estimate as the following argument
shows.
52
Theorem 66. (a) Upper tail: Let X1 , . . . , Xn be independent random variables
P
taking values in [0, 1]. Let X = ni=1 Xi . Let µ+ ≥ E[X]. Then for all
δ ≥ 0,
+
Pr[X ≥ (1 + δ)µ ] ≤
1
1+δ
!(1+δ)µ+
n − µ+
n − (1 + δ)µ+
+
!n−(1+δ)µ+
+
,
(64)
+
and as ( µn )n for δ = n−µ
.
where this bound is read as 0 for δ > n−µ
µ+
µ+
Consequently, all Chernoff bounds of Theorem 47 (including equations (35)
and (36) and Corollary 50) as well as those of Theorem 56 are valid when all
occurrences of E[X] are replaced by µ+ . The additive bounds (41) and (43)
as well as those of Theorem 58 and 60 are trivially valid with the expectation
replaced by an upper bound for it.
(b) Lower tail: All Chernoff bounds of Theorem 51 and 57, the ones in equations (42) and (44), as well as those of Corollary 59 are valid when all
occurrences of E[X] are replaced by µ− ≤ E[X].
Proof. We first show (64). There is nothing to do when (1 + δ)µ+ > n, so let
+ −E[X]
us assume (1 + δ)µ+ ≤ n. Let γ = µn−E[X]
. For all i ∈ [1..n], define Yi by
Yi = Xi + γ(1 − Xi ). Since γ ≤ 1, Yi ≤ 1. By definition, Yi ≥ Xi , and thus also
P
Y ≥ X for Y := ni=1 Yi . Also, µ+ = E[Y ]. Hence
Pr[X ≥ (1 + δ)µ+ ] ≤ Pr[Y ≥ (1 + δ)µ+ ] = Pr[Y ≥ (1 + δ)E[Y ]].
Now (64) follows immediately from Theorem 47, equation (29). Since (64) implies
all other Chernoff bounds of Theorem 47 (including equations (35) and (36) and
Corollary 50) via elementary estimates, all these bounds are valid with E[X] replaced by µ+ as well. This extends to Theorem 56, since it is just a reformulation
of Theorem 47. For the remaining (additive) bounds, replacing E[X] by an upper bound only decreases the probability of the tail event, so clearly these remain
valid.
To prove our claim on lower tail bounds, it suffices to note that all bounds in
Theorem 51 are monotonically decreasing in E[X]. So replacing E[X] by some
µ− < E[X] makes the tail event less likely and increases the probability in the
statement. Similarly, the additive bounds are not affected when replacing E[X]
with µ− .
We note without proof that the variance of the random variable Y constructed
above is at most the one of X. Since the tail bound in Theorem 58 is increasing in σ 2 , the same argument as above also shows that multiplicative versions of
Theorem 58 such as equation (57) remain valid when all occurrences of E[X] are
replaced by an upper bound µ+ ≥ E[X].
53
10.2
Chernoff Bounds for Sums of Dependent Random
Variables
In the previous subsection, we discussed large deviation bounds for the classic
setting of sums of independent random variables. In the analysis of algorithms,
often we cannot fully satisfy the assumption of independence. The dependencies
may appear minor, maybe even in our favor in some sense, so we could hope for
some good large deviations bounds.
In this section, we discuss three such situations which all lead to (essentially)
the known Chernoff bounds being applicable despite perfect independence missing.
The first of these was already discussed in Section 8.3, so we just note here how it
also implies the usual Chernoff bounds.
10.2.1
Unconditional Sequential Domination
In the analysis of sequential random processes such as iterative randomized algorithms, we rarely encounter that the events in different iterations are independent,
simply because the actions of our algorithm depend on the results of the previous
iterations. However, due to the independent randomness used, we can often say
that, independent of what happened in iterations 1, . . . , t − 1, in iteration t we
have a particular event with at least some probability p.
This property was made precise in the definition of unconditional sequential
domination before Lemma 38. This lemma then showed that unconditional sequential domination can lead to domination between the sums of the random
variables. This, however, immediately yields that tail bounds for the one set take
over to the other. We make this elementary insight precise in the following lemma.
This type of argument was used, among others, in works on shortest path problems [DHK11, DHK12]. Here one can show that, in each iteration, independent
of the past, with at least a certain probability an extra edge of a desired path is
found.
Lemma 67. Let (X1 , . . . , Xn ) and (X1∗ , . . . , Xn∗ ) be finite sequences of discrete
random variables. Assume that X1∗ , . . . , Xn∗ are independent.
(a) If (X1∗ , . . . , Xn∗ ) unconditionally sequentially dominates (X1 , . . . , Xn ), then
P
P
for all λ ∈ R, we have Pr[ ni=1 Xi ≥ λ] ≤ Pr[ ni=1 Xi∗ ≥ λ] and the latter expression can be bounded by Chernoff bounds for the upper tail of independent
random variables.
(b) If (X1∗ , . . . , Xn∗ ) unconditionally sequentially subdominates (X1 , . . . , Xn ),
P
P
then for all λ ∈ R, we have Pr[ ni=1 Xi ≤ λ] ≤ Pr[ ni=1 Xi∗ ≤ λ] and
the latter expression can be bounded by Chernoff bounds for the lower tail of
independent random variables.
54
10.2.2
Negative Correlation
Occasionally, we encounter random variables that are not independent, but that
display an intuitively even better negative correlation behavior. Take as example
the situation that we do not flip bits independently with probability nk , but that
we flip a set of exactly k bits chosen uniformly at random among all sets of k out
of n bits. Let X1 , . . . , Xn be the indicator random variables for the events that
bit 1, . . . , n flips. Clearly, the Xi are not independent. If X1 = 1, then Pr[X2 =
k−1
1] = n−1
, which is different from the unconditional probability nk . However, things
feel even better than independent: Knowing that X1 = 1 actually reduces the
probability that X2 = 1. This intuition is made precise in the following notion of
negative correlation.
Let X1 , . . . , Xn binary random variables. We say that X1 , . . . , Xn are
1-negatively correlated if for all I ⊆ [1..n] we have
Pr[∀i ∈ I : Xi = 1] ≤
Y
Pr[Xi = 1].
i∈I
We say that X1 , . . . , Xn are 0-negatively correlated if for all I ⊆ [1..n] we have
Pr[∀i ∈ I : Xi = 0] ≤
Y
Pr[Xi = 0].
i∈I
Finally, we call X1 , . . . , Xn negatively correlated if they are both 0-negatively correlated and 1-negatively correlated.
In simple words, these conditions require that the event that a set of variables
is all zero or all one, is at most as likely as in the case of independent random
variables. It seems natural that sums of such random variables are at least as
strongly concentrated as independent random variable, and in fact, Panconesi and
Srinivasan [PS97] were able to prove that negatively correlated random variables
admit Chernoff bounds. To be precise, they only proved that 1-negative correlation implies Chernoff bounds for the upper tail, but it is not too difficult to show
(see below) that the main argument works for all bounds proven via Bernstein’s
exponential moments method. In particular, for sums of 1-negatively correlated
random variable we obtain all Chernoff bounds for the upper tail that were presented in this work for independent random variables (as far as they can be applied
to binary random variables). We prove a slightly more general result as this helps
arguing that we can also work with upper bounds for the expectation instead of the
precise expectation. Further below, we then use a symmetry argument to argue
that 0-negative correlation implies all lower tail bounds presented so far.
Theorem 68 (1-negative correlation implies upper tail bounds). Let X1 , . . . , Xn
be 1-negatively correlated binary random variables. Let a1 , . . . , an , b1 , . . . , bn ∈ R
with ai ≤ bi for all i ∈ [1..n]. Let Y1 , . . . , Yn be random variables with Pr[Yi =
P
ai ] = Pr[Xi = 0] and Pr[Yi = bi ] = Pr[Xi = 1]. Let Y = ni=1 Yi .
55
(a) If a1 , . . . , an , b1 , . . . , bn ∈ [0, 1], then Y satisfies the Chernoff bounds given
in equations (29) to (33), (35) to (37), (41), (45) to (49), and (53) to (58),
P
where in the latter we use σ 2 := ni=1 Var[Yi ].
(b) Without the restriction to [0, 1] made in the previous paragraph, Y satisfies
the Chernoff bound of (43) with ci := bi − ai and the bounds of (59) to (62)
P
with σ 2 := ni=1 Var[Yi].
All these results also hold when E[Y ] is replaced by some µ+ ≥ E[Y ].
Proof. Let X1′ , . . . , Xn′ be independent binary random variables such that for all
i ∈ [1..n], the random variables Xi and Xi′ are identically distributed. Let ci :=
bi −ai for all i ∈ [1..n]. Note that Yi = ai +ci Xi for all i ∈ [1..n]. Let Yi′ = ai +ci Xi′ .
P
Let Y ′ = ni=1 Yi′ .
We first show that the 1-negative correlation of the Xi implies E[Y ℓ ] ≤ E[(Y ′ )ℓ ]
for all ℓ ∈ N0 . There is nothing to show for ℓ = 0 and
ℓ = 1, so let ℓ ≥ 2. Since
Pn
Pn
Pℓ
P
ℓ Pn
ℓ
Y = ( i=1 ai ) + ( i=1 ci Xi ), we have Y = k=0 k ( i=1 ai )ℓ−k ( ni=1 ci Xi )k . By
P
P
linearity of expectation, it suffices to show E[( ni=1 ci Xi )k ] ≤ E[( ni=1 ci Xi′ )k ].
P
P
Q
We have ( ni=1 ci Xi )k = (i1 ,...,ik )∈[1..n]k kj=1 cij Xij . Applying the definition of
1-negative correlation to the set I = {i1 , . . . , ik }, we compute
E
"
k
Y
j=1
#
cij Xij =
k
Y
j=1
≤
=
k
Y
!
cij Pr[∀j ∈ [1..k] : Xij = 1]
cij
j=1
k
Y
cij
j=1
!
!
Y
!
Pr[Xi = 1]
i∈I
Y
!
Pr[Xi′ = 1] = E
i∈I
"
k
Y
#
cij Xi′j .
j=1
Consequently, by linearity of expectation, E[( ni=1 ci Xi )k ] ≤ E[( ni=1 ci Xi′ )k ] for
all k ∈ N and thus E[Y ℓ ] ≤ E[(Y ′ )ℓ ].
We recall from Section 10.1.7 that essentially all large deviation bounds were
proven via upper bounds on the exponential moment E[ehY ]. Since the random
variable Y is bounded, by Fubini’s theorem we have
P
hY
E[e
∞
X
∞
X
hℓ Y ℓ
hℓ E[Y ℓ ]
]=E
=
.
ℓ!
ℓ=0 ℓ!
ℓ=0
"
#
′
P
(65)
Since E[Y ℓ ] ≤ E[(Y ′ )ℓ ], we have E[ehY ] ≤ E[ehY ]. Consequently, we obtain for Y
all Chernoff bounds which we could prove with the classic methods for Y ′ .
It remains to show that we can also work with an upper bound µ+ ≥ E[Y ].
For this, note that when we apply the construction of Theorem 66 to our random
56
variables Yi, that is, we define Zi = Yi + γ(1 − Yi ) for a suitable γ ∈ [0, 1], then the
resulting random variables Zi satisfy the same properties as the Yi, that is, there
are a′i and c′i such that Zi = a′i + c′i Xi . Consequently, we have Pr[Y ≥ (1 + δ)µ+ ] ≤
P
Pr[Z ≥ (1 + δ)µ+ ] = Pr[Z ≥ (1 + δ)E[Z]] for the sum Z = ni=1 Zi and the last
expression can bounded via the results we just proved.
Theorem 69 (0-negative correlation implies lower tail bounds). Let X1 , . . . , Xn
be 0-negatively correlated binary random variables. Let a1 , . . . , an , b1 , . . . , bn ∈ R
with ai ≤ bi for all i ∈ [1..n]. Let Y1 , . . . , Yn be random variables with Pr[Yi =
P
ai ] = Pr[Xi = 0] and Pr[Yi = bi ] = Pr[Xi = 1]. Let Y = ni=1 Yi .
(a) If a1 , . . . , an , b1 , . . . , bn ∈ [0, 1], then Y satisfies the Chernoff bounds given
in equations (38) to (40), (42), (50) to (52), and those in Corollary 59 with
P
σ 2 := ni=1 Var[Yi].
(b) Without the restriction to [0, 1] made in the previous paragraph, Y satisfies
the Chernoff bound of (44) with ci := bi − ai and those of the last paragraph
P
of Theorem 60 with σ 2 := ni=1 Var[Yi].
All these results also hold when E[Y ] is replaced by some µ− ≤ E[Y ].
Proof. Let Ỹi := 1 − Yi . Then the Ỹi satisfy the assumptions of Theorem 68 (with
ãi = 1 − bi , b̃i = 1 − ai , and X̃i = 1 − Xi ; note that the latter are 1-negatively
correlated since the Xi are 0-negatively correlated, note further that ãi , b̃i ∈ [0, 1] if
ai , bi ∈ [0, 1]). Hence Theorem 68 gives the usual Chernoff bounds for the Ỹi. As in
Section 10.1.2, these translate into the estimates (38) to (40) and these imply (50)
to (52). Bound (41) for the Ỹi immediately translates to (42) for the Yi . Finally, the
results of Theorem 58 imply those of Corollary 59. All these results are obviously
weaker when E[Y ] is replaced by some µ− ≤ E[Y ].
10.2.3
Hypergeometric Distribution
It remains to point out some situations where we encounter negatively correlated
random variables. One typical situations (but by far not the only one) is sampling
without replacement, which leads to the hypergeometric distribution.
Say we choose randomly n elements from a given N-element set S without
replacement. For a given m-element subset T of S, we wonder how many of
its elements we have chosen. This random variable is called hypergeometrically
distributed with parameters N, n and m.
More formally, let S be any N-element set. Let T ⊆ S have exactly m elements. Let U be a subset of S chosen uniformly among all n-element subsets of
57
S. Then X = |U ∩ T | is a random variable with hypergeometric distribution (with
parameters N, n and m). By definition,
Pr[X = k] =
m
k
N −m
n−k
N
n
for all k ∈ [max{0, n + m − N}.. min{n, m}].
||T |
It is easy to see that E[X] = |U|S|
= mn
: Enumerate T = {t1 , . . . , tm } in an
N
arbitrary manner (before choosing U). For i = 1, . . . , m, let Xi be the indicator
|
= Nn . Since
random variable for the event ti ∈ U. Clearly, Pr[Xi = 1] = |U
|S|
Pm
X = i=1 Xi , we have E[X] = mn
by linearity of expectation (Lemma 20).
N
It is also obvious that the Xi are not independent. If n < m and X1 = . . . =
Xn = 1, then necessarily we have Xi = 0 for i > n. Fortunately, however, these
dependencies are of the negative correlation type. This is intuitively clear, but
also straight-forward to prove.
Let I ⊆ [1..m], W = {ti | i ∈ I}, and w = |W | = |I|. Then Pr[∀i ∈ I : Xi =
1] = Pr[W ⊆ U]. Since U is uniformly
chosen, it suffices to count the number of
|S\W |
U that contain W , these are |U \W | , and to compare them with the total number
of possible U. Hence
N −w
Pr[W ⊆ U] =
n−w
!
N
n
!
n
n · . . . · (n − w + 1)
<
=
N · . . . · (N − w + 1)
N
w
=
Y
Pr[Xi = 1].
i∈I
In a similar fashion, we have
Pr[∀i ∈ I : Xi = 0] = Pr[U ∩ W = ∅]
!
!
N
N −w
=
n
n
(N − n) . . . (N − n − w + 1)
=
N . . . (N − w + 1)
N −n w Y
≤
=
Pr[Xi = 0],
N
i∈I
where we read N −w
= 0 when n < N − w.
n
Together with Theorems 68 and 69, we obtain the following theorem.
Theorem 70. Let N ∈ N. Let S be some set of cardinality N, for convenience,
S = [1..N]. Let n ≤ N and let U be a subset of S having cardinality n uniformly
chosen among all such subsets. For i ∈ [1..N] let Xi be the indicator random
variable for the event i ∈ U. Then X1 , . . . , XN are negatively correlated.
58
Consequently, if X is a random variable having a hypergeometric distribution
with parameters N, n, and m, then the usual Chernoff bounds for sums of n
independent binary random variables (listed in Theorems 68 and 69) hold.
Note that for hypergeometric distributions we have symmetry in n and m, that
is, the hypergeometric distribution with parameters N, n, and m is the same as
the hypergeometric distribution with parameters N, m, and n. Hence for Chernoff
bounds depending on the number of random variables, e.g., Theorem 53, we can
make this number being min{n, m} by interpreting the random experiment in the
right fashion.
That the hypergeometric distribution satisfies the Chernoff bounds of Theorem 53 has recently in some works been attributed to Chvátal [Chv79], but this is
not correct. As Chvátal writes, the aim of his note is solely to give an elementary
proof of the fact that the hypergeometric distribution satisfies the strongest Chernoff bound of Theorem 47 (which implies the bounds of Theorem 53), whereas the
result itself is from Hoeffding [Hoe63].
For a hypergeometric random variable X with parameters N, n, and m,
N
and z ≥ n2 , then Pr[X = z] ≤ ( 2em
)z . With
[BLS14] show that if m < 2e
N
Theorem 70, we can use the usual Chernoff bound of equation (46) and obtain the
stronger bound
z−E[X]
Pr[X ≥ z] ≤ e
E[X]
z
!z
≤
eE[X]
z
!z
enm
=
zN
z
,
(66)
)z for z ≥ n/2.
which is at most ( 2em
N
Theorem 70 can be extended to point-wise maxima of several families like (Xi )
in Theorem 70 if these are independent. This result was used in the analysis of a
population-based genetic algorithm in [DD17], but might be useful also in other
areas of discrete algorithmics.
Lemma 71. Let k, N ∈ N. For all j ∈ [1..k], let nj ∈ [1..N]. Let S be some
set of cardinality N, for convenience, S = [1..N]. For all j ∈ [1..k], let Uj be a
subset of S having cardinality nj uniformly chosen among all such subsets. Let
the Uj be stochastically independent. For all i ∈ S, let Xi be the indicator random
variable for the event that i ∈ Uj for some j ∈ [1..k]. Then the random variables
X1 , . . . , XN are negatively correlated.
Note that the situation in the lemma above can be seen as sampling with partial
P
replacement. We sample a total of j nj elements, but we replace the elements
chosen only after round n1 , n1 + n2 , ... We expect that other partial replacement
scenarios also lead to negatively correlated random variables, and thus to the usual
Chernoff bounds.
59
We remark that negative correlation can be useful also without Chernoff
bounds. For example, in Section 9 we used the lemma above to prove a lower
bound on the coupon collector time (equivalently, on the runtime of the randomized local search heuristic on monotonic fuctions).
10.3
Chernoff Bounds for Functions of Independent Variables, Martingales, and Bounds for Maxima
So far we discussed tail bounds for random variables which can be written as sum
of (more or less) independent random variables. Sometimes, the random variable
we are interested in is determined by the outcomes of many independent random
variables, however, not simply as a sum of these. Nevertheless, if each of the
independent random variables only has a limited influence on the outcome, then
bounds similar to those of Theorem 55 can be proven. Such bounds can be found
under the names Azuma’s inequality, Martingale inequalities, or method of bounded
differences.
We shall try to avoid the use of martingales, which need some deeper understanding of probability theory, and first present two bounds due to McDiarmid [McD98] that need martingales in their proof, but not in their statement.
Theorem 72 (Method of bounded differences). Let X1 , . . . , Xn be independent
random variables taking values in the sets Ω1 , . . . , Ωn , respectively. Let Ω := Ω1 ×
. . . × Ωn . Let f : Ω → R. For all i ∈ [1..n] let ci > 0 be such that for all ω, ω̄ ∈ Ω
we have that if for all j 6= i, ωj = ω̄j , then |f (ω) − f (ω̄)| ≤ ci .
Let X = f (X1 , . . . , Xn ). Then for all λ ≥ 0,
Pr[X ≥ E[X] + λ] ≤ exp
Pr[X ≤ E[X] − λ] ≤ exp
2λ2
− Pn
!
2
i=1 ci
!
2
2λ
− Pn
2
i=1 ci
,
.
The version of Azuma’s inequality given above is due to McDiarmid [McD98]
P
and is stronger than the bound exp(−λ2 /2 ni=1 c2i ) given by several other authors.
Theorem 72 found numerous applications in discrete mathematics and computer science, however, only few in the analysis of randomized search heuristics
(the only one we are aware of is [BD17]). All other analyses of randomized search
heuristics that needed Chernoff-type bounds for random variables that are determined by independent random variables, but in a way other than as simple sum,
resorted to the use of martingales.
One reason for this might be that the bounded differences assumption is easily proven in discrete mathematics problems like the analysis of random graphs,
60
whereas in algorithms the sequential nature of the use of randomness may make
it hard to argue that a particular random variable sampled now has a bounded
influence on the final result regardless of how we condition on all future random
variables. A more natural condition might be that the outcome of the current
random variable has only a limited influence on the expected result determined by
the future random variables. For this reason, we are optimistic that the following
result might become useful in the analysis of randomized search heuristics. This
result is a weak version of Theorem 3.7 in [McD98].
Theorem 73 (Method of bounded conditional expectations). Let X1 , . . . , Xn be
independent random variables taking values in the sets Ω1 , . . . , Ωn , respectively.
Let Ω := Ω1 × . . . × Ωn . Let f : Ω → R. For all i ∈ [1..n] let ci > 0 be such that
for all ω1 ∈ Ω1 , . . . , ωi−1 ∈ Ωi−1 and all ωi , ω̄i ∈ Ωi we have
|E[f (ω1 , . . . , ωi−1 , ωi , Xi+1 , . . . , Xn )] − E[f (ω1 , . . . , ωi−1, ω̄i , Xi+1 , . . . , Xn )]| ≤ ci .
Let X = f (X1 , . . . , Xn ). Then for all λ ≥ 0,
Pr[X ≥ E[X] + λ] ≤ exp
Pr[X ≤ E[X] − λ] ≤ exp
λ2
!
− Pn 2 ,
2 i=1 ci
!
λ2
− Pn 2 .
2 i=1 ci
Here is an example of how the new theorem can be helpful. The compact
genetic algorithms (cGA) maximizes a function f : {0, 1}n → R as follows. There
is a (hypothetical) population size K ∈ N, which we assume to be an even integer.
The cGA sets the initial frequency vector τ (0) ∈ [0, 1]n to τ (0) = ( 21 , . . . , 12 ). Then,
in each iteration t = 1, 2, . . . it generates two search points x(t,1) , x(t,2) ∈ {0, 1}n
randomly such that, independently for all j ∈ {1, 2} and i ∈ [1..n], we have
(t,j)
(t)
Pr[xi = 1] = τi . If f (x(t,1) ) < f (x(t,2) ), then we swap the two variables, that
is, we set (x(t,1) , x(t,2) ) ← (x(t,2) , x(t,1) ). Finally, in this iteration, we update the
frequency vector by setting τ (t+1) ← τ (t) + K1 (x(t,1) − x(t,2) ).
(t)
Let us analyze the behavior of the frequency τi of a neutral bit i ∈ [1..n], that
(t)
is, one that has no influence on the fitness. To ease reading, let Xt := τi . We
have X0 = 12 with probability one. Once Xt is determined, we have
Pr[Xt+1 = Xt + K1 ] = Xt (1 − Xt ),
Pr[Xt+1 = Xt − K1 ] = Xt (1 − Xt ),
Pr[Xt+1 = Xt ] = 1 − 2Xt (1 − Xt ).
In particular, we have E[Xt+1 | X0 , . . . , Xt ] = E[Xt+1 | Xt ] = Xt . By induction,
we have E[XT | Xt ] = Xt for all T > t.
61
2
K
, XT has
Our aim is to show that with probability at least 1 − 2 exp − 32T
not yet converged to one of the absorbing states 0 and 1. We first write the
frequencies as results of independent random variables. For convenience, these
will be continuous random variables, but it is easy to see that instead we could
have used discrete ones as well. For all t = 1, 2, . . . let Rt be a random number
uniformly distributed in the interval [0, 1]. Define Y0 , Y1 , . . . as follows. We have
Y0 = 12 with probability one. For t ∈ N0 , we set
Pr[Yt+1 = Yt + K1 ] if Rt ≥ 1 − Yt (1 − Yt ),
Pr[Yt+1 = Yt − K1 ] if Rt ≤ Yt (1 − Yt ),
Pr[Xt+1 = Xt ] else.
It is easy to see that (X0 , X1 , . . . ) and (Y0 , Y1, . . . ) are identically distributed.
Note that YT is a function g of (R1 , . . . , RT ). For concrete values r1 , . . . , rt ∈
[0, 1], we have E[g(r1, . . . , rt , Rt+1 , . . . , RT )] = E[YT | Yt ] = Yt . Consequently,
for all r t ∈ [0, 1], the two expectations E[g(r1, . . . , rt−1 , rt , Rt+1 , . . . , RT )] and
E[g(r1, . . . , rt−1 , rt , Rt+1 , . . . , RT )] are two possible outcomes of Yt given a common value for Yt−1 (which is determined by r1 , . . . , rt−1 ), and hence differ by at
most ct = K2 . We can thus apply Theorem 73 as follows.
h
Pr[YT ∈ {0, 1}] = Pr |YT − 12 | ≥
h
1
2
i
= Pr |g(R1 , . . . , RT ) − E[g(R1, . . . , RT )]| ≥
( 1 )2
≤ 2 exp − 2 2 2
2T ( K )
!
K2
= 2 exp −
32T
!
1
2
i
.
Note that it is not obvious how to obtain this result with the classic method
of bounded differences (Theorem 72). In particular, the above construction does
not satisfy the bounded differences condition, that is, there are values r1 , . . . , rT
and rt such that g(r1 , . . . , rT ) and g(r1 , . . . , rt−1 , r t , rt+1 , . . . , rT ) differ by significantly more than K2 . For this reason, we are optimistic that Theorem 73 will find
applications in the theory of randomized search heuristics.
Both Theorem 72 and 73 are special cases of the following martingale result,
which is often attributed to Azuma [Azu67] despite the fact that it was proposed
already in Hoeffding [Hoe63].
Theorem 74 (Azuma-Hoeffding inequality). Let X0 , X1 , . . . , Xn be a martingale.
Let c1 , . . . , cn > 0 with |Xi − Xi−1 | ≤ ci for all i ∈ [1..n]. Then for any λ ≥ 0,
Pr[Xn − X0 ≥ λ] ≤ exp
62
λ2
!
− Pn 2 .
2 i=1 ci
This result has found several applications in the theory of randomized search
heuristics, e.g., in [DK15, Köt16, DDY16b].
We note that the theorem above is a direct extension of Theorem 55 to Martingales. In a similar vein, there are Martingale versions of most other Chernoff
bounds presented in this work. We refer to McDiarmid [McD98] for more details.
Tail bounds for maxima and minima of partial sums: We end this section
with a gem already contained in Hoeffding. It builds on the following elementary
observation: If X0 , X1 , . . . , Xn is a martingale, then Y0 , Y1 , . . . , Yn defined as follows
also forms a martingale. Let λ ∈ R. Let i ∈ [0..n] minimal with Xi ≥ λ, if such
an Xi exists, and i = n + 1 otherwise. Let Yj = Xj for j ≤ i and Yj = Xi
for j > i. Then Yn ≥ λ if and only if maxi∈[1..n] Xi ≥ λ. Since Y0 , . . . , Yn is a
martingale with martingales differences bounded as least as well as for X0 , . . . , Xn
(and also all other variation measures at least as good as for X0 , . . . , Xn ), all large
deviation bounds provable for the martingale X0 , . . . , Xn via the Bernstein method
are valid also for Y0 , . . . , Yn , that is, for maxi∈[1..n] Xi . Since we did not introduce
martingales here, we omit the details and only state some implications of this
observation. The reader finds more details in [Hoe63, end of Section 2] and [McD98,
end of Section 3.5]. It seems that both authors do not see this extension as very
important (see also the comment at the end of Section 2 in [McD98]). We feel that
this might be different for randomized search heuristics. For example, to prove
that a randomized search heuristic has at least some optimization time T , we need
to show that the distance of each of the first T − 1 solutions from the optimum is
positive, that is, that the minimum of these differences is positive.
Theorem 75 (Tail bounds for maxima and minima). Let X1 , . . . , Xn be indepenP
dent random variables. For all i ∈ [1..n], let Si = ij=1 Xi . Assume that one of
the results in Section 10.1 yields the tail bound Pr[Sn ≥ E[Sn ] + λ] ≤ p. Then we
also have
Pr[∃i ∈ [1..n] : Si ≥ E[Si ] + λ] ≤ p.
(67)
In an analogous manner, each tail bound Pr[Sn ≤ E[Sn ] − λ] ≤ p derivable from
Section 10.1 can be strengthened to Pr[∃i ∈ [1..n] : Si ≤ E[Si ] − λ] ≤ p.
Note that if the Xi in the theorem are non-negative, then trivially equation (67)
implies the uniform bound
Pr[∃i ∈ [1..n] : Si ≥ E[Sn ] + λ] ≤ p.
(68)
Note also that the deviation parameter λ does not scale with i. In particular, a
bound like Pr[∃i ∈ [1..n] : Si ≥ (1 + δ)E[Si ]] ≤ p cannot be derived.
63
10.4
Chernoff Bounds for Geometric Random Variables
As visible from Lemma 80 below, sums of independent geometric random variables
occur frequently in the analysis of randomized search heuristics. Surprisingly, it
was only in 2007 that a Chernoff-type bound was used to analyze such sums
in the theory of randomized search heuristics [DHK07] (for subsequent uses see,
e.g., [BBD+ 09, DHK11, ZLLH12, DJWZ13, DD17]). Even more surprisingly, only
recently Witt [Wit14] proved good tail bounds for sums of geometric random
variables having significantly different success probabilities. Note that geometric
random variables are unbounded. Hence the Chernoff bounds presented so far
cannot be applied directly.
We start this subsection with simple Chernoff bounds for sums of identically
distributed geometric random variables as these can be derived from Chernoff
bounds for sums of independent 0, 1 random variables discussed so far. We remark
that a sum X of n independent geometric distributions with success probability
p > 0 is closely related to the negative binomial distribution NB(n, 1 − p) with
parameters n and 1 − p: We have X ∼ NB(n, 1 − p) + n.
Theorem 76. Let X1 , . . . , Xn be independent geometric random variables with
P
common success probability p > 0. Let X := ni=1 Xi and µ := E[X] = np .
(a) For all δ ≥ 0,
δ2 n − 1
Pr[X ≥ (1 + δ)µ] ≤ exp −
2 1+δ
!
1
≤ exp − min{δ 2 , δ}(n − 1) .
4
(69)
(b) For all 0 ≤ δ ≤ 1,
(1 − δ)(µ − n)
Pr[X ≤ (1 − δ)µ] ≤ (1 − δ)
(1 − δ)µ − n
≤ (1 − δ)n exp(δn)
n
2
≤ exp −
!(1−δ)µ−n
!
δ n
,
2 − 43 δ
(70)
(71)
(72)
where the first bound reads as pn for (1 − δ)µ = n and as 0 for (1 − δ)µ < n.
For 0 ≤ δ < 1, we also have
2δ 2 pn
,
Pr[X ≤ (1 − δ)µ] ≤ exp −
(1 − δ)
!
2p3 λ2
Pr[X ≤ µ − λ] ≤ exp −
.
n
!
64
(73)
(74)
Bounds (73) and (74) are interesting only for relatively large values of p. Since
part (a) has been proven in [BBD+ 09], we only prove (b). The main idea in both
cases is exploiting the natural relation between a sum of independent identically
distributed geometric random variables and a sequence of Bernoulli events.
Proof. Let Z1 , Z2 , . . . be independent binary random variables with Pr[Zi = 1] = p
P
for all i ∈ N. Let n < K ≤ np . Let YK = K
i=1 Zi . Then X ≤ K if and only if
YK ≥ n. Consequently, by Theorem 47,
Pr[X ≤ K] = Pr[YK ≥ n]
"
n
= Pr YK ≥ 1 +
−1
Kp
!!
#
E[YK ]
Kp n K − Kp K−n
n
K−n
Kp n
exp(n − Kp)
≤
n
!
pλ2
≤ exp
,
2K + 32 λ
≤
where we used the shorthand λ := µ − K for the absolute deviation. From Theorem 53, we derive
Pr[X ≤ K] = Pr[YK ≥ n]
= Pr[YK ≥ E[YK ] + (n − Kp)]
2(n − Kp)2
≤ exp −
K
!
!
2p2 λ2
2p2 λ2
≤ exp −
.
= exp −
µ−λ
µ
!
Replacing K by (1 − δ)µ and λ by δµ in these equations gives the claim.
When the geometric random variables have different success probabilities, the
following bounds can be employed.
Theorem 77. Let X1 , . . . , Xn be independent geometric random variables with
success probabilities p1 , . . . , pn > 0. Let pmin := min{pi | i ∈ [1..n]}. Let X :=
Pn
Pn 1
i=1 Xi and µ = E[X] =
i=1 pi .
65
(a) For all δ ≥ 0,
1
(1 − pmin)µ(δ−ln(1+δ))
1+δ
≤ exp(−pmin µ(δ − ln(1 + δ)))
Pr[X ≥ (1 + δ)µ] ≤
!n
δµpmin
exp(−δµpmin )
n
!
(δµpmin)2
.
≤ exp −
2n(1 + δµpnmin )
≤ 1+
(75)
(76)
(77)
(78)
(b) For all 0 ≤ δ ≤ 1,
Pr[X ≤ (1 − δ)µ] ≤ (1 − δ)pminµ exp(−δpmin µ)
(79)
≤ exp −
!
(80)
≤ exp(− 12 δ 2 µpmin ).
(81)
δ 2 µpmin
2 − 34 δ
Estimates (75) and (76) are from [Jan17], bound (77) is from [Sch00], and (78)
follows from the previous by standard estimates. Consequently, these bounds
are not only more general, but also stronger than our elementary bound (69) for
identically distributed geometric random variables.
For the lower tail bounds, (79) from [Jan17] is identical to (71) for identically distributed variables. Hence (70) is the strongest estimate for identically
distributed geometric random variables. Equation (79) gives (80) via the same
estimate that gives (72) from (71). Estimate (81) appeared already in [Sch00].
Overall, it remains surprising that such useful bounds were proven only relatively late and have not yet appeared in a scientific journal.
The bounds of Theorem 77 allow the geometric random variables to have different success probabilities, however, the tail probability depends only on the smallest
of them. This is partially justified by the fact that the corresponding geometric
random variable has the largest variance, and thus might be most detrimental to
the desired strong concentration. If the success probabilities vary significantly,
however, then this approach gives overly pessimistic tail bounds. Witt [Wit14]
proves the following result, which can lead to stronger estimates.
Theorem 78. Let X1 , . . . , Xn be independent geometric random variables with
P
P
success probabilities p1 , . . . , pn > 0. Let X = ni=1 Xi , s = ni=1 ( p1i )2 , and pmin :=
66
min{pi | i ∈ [1..n]}. Then for all λ ≥ 0,
λ2
1
, λpmin
Pr[X ≥ E[X] + λ] ≤ exp − min
4
s
!
λ2
Pr[X ≤ E[X] − λ] ≤ exp −
.
2s
(
)!
,
(82)
(83)
In the analysis of randomized search heuristics, apparently we often encounter
sums of independent geometrically distributed random variables X1 , . . . , Xn with
success probabilities pi proportional to i. For this case, the following result
from [DD17] gives stronger tail bounds than the previous result. See Section 4.2
for the definition of the harmonic number Hn .
Theorem 79. Let X1 , . . . , Xn be independent geometric random variables with
success probabilities p1 , . . . , pn . Assume that there is a number C ≤ 1 such that
P
pi ≥ C ni for all i ∈ [1..n]. Let X = ni=1 Xi . Then
E[X] ≤
1
nHn
C
Pr[X ≥ (1 +
≤
1
n(1
C
+ ln n),
δ) C1 n ln n]
−δ
≤n
(84)
for all δ ≥ 0.
(85)
As announced in Section 8.2, we now present a few examples where the existing
literature only gives an upper bound for the expected runtime, but where a closer
look at the proofs easily gives more details about the distribution, which in particular allows to obtain tail bounds for the runtime. We note that similar results so
far have only been presented for the (1 + 1) EA optimizing the LeadingOnes test
function [DJWZ13] and for RLS optimizing the OneMax test function [Wit14].
Zhou et al. [ZLLH12] implicitly give several results of this type, however, the resulting runtime guarantees are not optimal due to the use of an inferior Chernoff
bound for geometric random variables.
Lemma 80. (a) The runtime T of the (1 + 1) EA on the OneMax function is
P
i
dominated by the independent sum ni=1 Geom( en
) [DJW02]. Hence E[T ] ≤
−δ
enHn and Pr[T ≥ (1 + δ)en ln n] ≤ n for all δ ≥ 0.
(b) For any function f : {0, 1}n → R, the runtime T of the (1 + 1) EA is
dominated by Geom(n−n ) [DJW02]. Hence E[T ] ≤ nn and Pr[T ≥ γnn ] ≤
n
(1 − n−n )γn ≤ e−γ for all γ ≥ 0.
(c) The runtime T of the (1 + 1) EA finding Eulerian
graphs using perfect matchings in the adjacency lists
ing the edge-based mutation operator is dominated by
Pm/3
i
Hence E[T ] ≤ 2emHm/3
i=1 Geom( 2em ) [DJ07].
δ)em ln m3 ] ≤ ( m3 )−δ for all δ ≥ 0.
67
cycles in undirected
as genotype and usthe independent sum
and Pr[T ≥ 2(1 +
(d) The runtime of the (1 + 1) EA sorting an array of length n by minimizing the number of inversions is dominated by the independent sum
P(n2 )
4e n
3i
H(n) ≤ 2e
n2 (1 + 2 ln n)
)
[STW04].
Hence
E[T
]
≤
Geom(
i=1
3 2
3
4e(n
2
2)
and Pr[T ≥ (1 + δ) 4e
n2 ln n] ≤
3
−δ
n
2
.
(e) The runtime of the multi-criteria (1 + 1) EA for the single-source shortest
path problem in a graph G can be described as follows. Let ℓ be such that
there is a shortest path from the source to any vertex having at most ℓ edges.
Then there are random variables Gij , i ∈ [1..ℓ], j ∈ [1..n − 1], such that
(i) Gij ∼ Geom( en12 ) for all i ∈ [1..ℓ] and j ∈ [1..n − 1], (ii) for all j ∈
[1..n−1] the variables G1j , . . . , Gℓj are independent, and (iii) T is dominated
P
by max{ ℓi=1 G
| j ∈ [1..n − 1]} [STW04, DHK11]. Consequently, for δ =
qij
4 ln(n−1)
4 ln(n−1)
1
} and T0 := (1 + δ) pℓ , we have E[T ] ≤ (1 + ln(n−1)
)T0
max{ ℓ−1 ,
ℓ−1
−ε
and Pr[T ≥ (1 + ε)T0 ] ≤ (n − 1) for all ε ≥ 0.
Proof. We shall not show the domination statements as these can be easily derived
from the original analyses cited in the theorem. Given the domination result,
parts (a), (c), and (d) follow immediately from Theorem 79. Part (b) follows
directly from the law of the geometric distribution.
To prove part (e), let X1 , . . . , Xℓ be independent geometrically distributed
Pℓ
random variables with parameter p = en1 2 . Let X =
Let δ =
i=1 Xi .
q
2
δ
max{ 4 ln(n−1)
, 4 ln(n−1)
}. Then, by (69), Pr[X ≥ (1 + δ)E[X]] ≤ exp(− 12 1+δ
(ℓ −
ℓ−1
ℓ−1
1 4 ln(n−1)
1
1
2
1)) ≤ exp(− 4 min{δ , δ}(ℓ − 1)) ≤ exp(− 4 ℓ−1 (ℓ − 1)) = n−1 . For all ε > 0,
again by (69), we compute
1 (δ + ε + δε)2
Pr[X ≥ (1 + ε)(1 + δ)E[X]] ≤ exp −
(ℓ − 1)
2 (1 + δ)(1 + ε)
!
1 δ 2 (1 + ε)2
(ℓ − 1)
≤ exp −
2 (1 + δ)(1 + ε)
!
!1+ε
1 δ2
≤ exp −
(ℓ − 1)
21+δ
≤ (n − 1)−(1+ε) .
Let Y1 , . . . , Yn−1 be random variables with distribution equal to the one of X. We
do not make any assumption on the correlation of the Yi , in particular, they do
not need to be independent. Let Y = max{Yi | i ∈ [1..n − 1]} and recall that
the runtime T is dominated by Y . Let T0 = (1 + δ)E[X] = (1 + δ) pℓ . Then
Pr[Y ≥ (1 + ε)T0 ] ≤ (n − 1) Pr[X ≥ (1 + ε)T0 ] ≤ (n − 1)−ε by the union bound
68
(Lemma 12). By Corollary 18,
!
1
T0 .
E[Y ] ≤ 1 +
ln(n − 1)
We note that not all classic proofs reveal details on the distribution. For
results obtained via random walk arguments, e.g., the optimization of the short
path function SPCn [JW01], monotone polynomials [WW05], or vertex covers on
paths-like graphs [OHY09], as well as for results proven via additive drift [HY01],
the proofs often give little information about the runtime distribution (an exception
is the analysis of the needle and the OneMax function in [GKS99]).
For results obtained via the average weight decrease method [NW07] or multiplicative drift analysis [DG13], the proof also does not give information on
the runtime distribution. However, the probabilistic runtime bound of type
Pr[T ≥ T0 + λ] ≤ (1 − δ)λ obtained from these methods implies that the runtime is dominated by T T0 − 1 + Geom(1 − δ).
10.5
Tail Bounds for the Binomial Distribution
For binomially distributed random variables, slightly stronger tail bounds exist.
The difference is small, but since they have been used in the analysis of randomized
search heuristics, we briefly describe them here.
In this section, let X always be a binomially distributed random variable with
P
parameters n and p, that is, X = ni=1 Xi with independent Xi satisfying Pr[Xi =
1] = p and Pr[Xi = 0] = 1 − p. The following estimate seems well-known (e.g., it
was used in [JJW05] without proof or reference). Gießen and Witt [GW17] give an
elementary proof via estimates of binomial coefficients and the binomial identity.
We find the proof below more intuitive.
Lemma 81. Let X ∼ Bin(n, p). Let k ∈ [0..n]. Then
!
n k
p .
Pr[X ≥ k] ≤
k
Proof. For all T ⊆ [1..n] with |T | = k let AT be the event that Xi = 1 for all
i ∈ T . Clearly, Pr[AT ] = pk . The event “X ≥ k” is the
union of the events AT
P
with T as above. Hence Pr[X ≥ k] ≤ T Pr[AT ] = nk pk by the union bound
(Lemma 12).
)k , which often is an apWhen estimating the binomial coefficient by nk ≤ ( en
k
propriate way to derive more understandable expressions, the above bound reverts
69
to equation (35), a slightly weaker version of the classic multiplicative bound (30).
Since we are not aware of an application of Lemma 81 that does not estimate the
binomial coefficient in this way, its main value might be its simplicity.
The following tail bound for the binomial distribution was shown by
Klar [Kla00], again with elementary arguments. In many cases, it is significantly
stronger than Lemma 81. However, again we do not see an example where this tail
bound would have improved an existing analysis of a randomized search heuristics.
Lemma 82. Let X ∼ Bin(n, p) and k ∈ [np..n]. Then
Pr[X ≥ k] ≤
(k + 1)(1 − p)
Pr[X = k].
k + 1 − (n + 1)p
Note that, trivially, Pr[X = k] ≤ Pr[X ≥ k], so it is immediately clear that this
(k+1)(1−p)
). With elementary
estimate is quite tight (the gap is at most the factor k+1−(n+1)p
arguments, Lemma 82 gives the slightly weaker estimate
Pr[X ≥ k] ≤
k − kp
Pr[X = k],
k − np
(86)
which appeared also in [Fel68, equation (VI.3.4)]. For p = n1 , the typical mutation
rate in standard-bit mutation, Lemma 82 gives
Pr[X ≥ k] ≤
k+1
Pr[X = k].
k
(87)
1 n−k
)
Pr[X = k] and
Writing Lemma 81 in the equivalent form Pr[X ≥ k] ≤ ( 1−p
1 n−k
noting that ( 1−p )
≥ exp(p(n − k)), we see that in many cases Lemma 82 gives
substantially better estimates.
References
[Azu67]
Kazuoki Azuma. Weighted sums of certain dependent variables. Tohoku Mathematical Journal, 3:357–367, 1967.
[BBD+ 09]
Surender Baswana, Somenath Biswas, Benjamin Doerr, Tobias
Friedrich, Piyush P. Kurur, and Frank Neumann. Computing single
source shortest paths using single-objective fitness. In 10th Workshop
on Foundations of Genetic Algorithms, FOGA 2009, pages 59–66.
ACM, 2009.
[BD17]
Maxim Buzdalov and Benjamin Doerr. Runtime analysis of the
(1 + (λ, λ)) genetic algorithm on random satisfiable 3-CNF formulas. In Genetic and Evolutionary Computation Conference, GECCO
70
2017, pages 1343–1350. ACM, 2017.
http://arxiv.org/abs/1704.04366.
Full version available at
[BDN10]
Süntje Böttcher, Benjamin Doerr, and Frank Neumann. Optimal
fixed and adaptive mutation rates for the LeadingOnes problem. In
11th International Conference on Parallel Problem Solving from Nature, PPSN 2010, pages 1–10. Springer, 2010.
[BE08]
Pavel A. Borisovsky and Anton V. Eremeev. Comparing evolutionary
algorithms to the (1+1)-EA. Theoretical Computer Science, 403:33–
41, 2008.
[Ben62]
George Bennett. Probability inequalities for the sum of independent
random variables. Journal of the American Statistical Association,
57:33–45, 1962.
[Ber24]
Sergey N. Bernstein. On a modification of Chebyshev’s inequality
and of the error formula of Laplace. Ann. Sci. Inst. Sav. Ukraine,
Sect. Math. 1, 4:38–49, 1924.
[Bie53]
Irénée-Jules Bienaymé. Considérations à l’appui de la découverte de
Laplace. Comptes Rendus de l’Académie des Sciences, 37:309–324,
1853.
[BLS14]
Golnaz Badkobeh, Per Kristian Lehre, and Dirk Sudholt. Unbiased
black-box complexity of parallel search. In Parallel Problem Solving
from Nature, PPSN 2014, pages 892–901. Springer, 2014.
[CDEL16]
Dogan Corus, Duc-Cuong Dang, Anton V. Eremeev, and Per Kristian Lehre. Level-based analysis of genetic algorithms and other
search processes. CoRR, abs/1407.7663v2, 2016.
[Che52]
Herman Chernoff. A measure of asymptotic efficiency for tests of a
hypothesis based on the sum of observations. Annals of Mathematical
Statistics, 23:493–507, 1952.
[Chv79]
Vasek Chvátal. The tail of the hypergeometric distribution. Discrete
Mathematics, 25:285–287, 1979.
[DD16]
Benjamin Doerr and Carola Doerr. The impact of random initialization on the runtime of randomized search heuristics. Algorithmica,
75:529–553, 2016.
71
[DD17]
Benjamin Doerr and Carola Doerr. Optimal static and self-adjusting
parameter choices for the (1+(λ, λ)) genetic algorithm. Algorithmica,
2017. To appear.
[DDY16a]
Benjamin Doerr, Carola Doerr, and Jing Yang. k-bit mutation with
self-adjusting k outperforms standard bit mutation. In 14th International Conference on Parallel Problem Solving from Nature, PPSN
2016, pages 824–834. Springer, 2016.
[DDY16b]
Benjamin Doerr, Carola Doerr, and Jing Yang. Optimal parameter
choices via precise black-box analysis. In Genetic and Evolutionary
Computation Conference, GECCO 2016, pages 1123–1130. ACM,
2016.
[DFW10]
Benjamin Doerr, Mahmoud Fouz, and Carsten Witt. Quasirandom
evolutionary algorithms. In Genetic and Evolutionary Computation
Conference, GECCO 2010, pages 1457–1464. ACM, 2010.
[DFW11]
Benjamin Doerr, Mahmoud Fouz, and Carsten Witt. Sharp bounds
by probability-generating functions and variable drift. In Genetic
and Evolutionary Computation Conference, GECCO 2011, pages
2083–2090. ACM, 2011.
[DG13]
Benjamin Doerr and Leslie Ann Goldberg. Adaptive drift analysis.
Algorithmica, 65:224–250, 2013.
[DGWY17]
Benjamin Doerr, Christian Gießen, Carsten Witt, and Jing
Yang. The (1+λ) evolutionary algorithm with self-adjusting mutation rate.
In Genetic and Evolutionary Computation Conference, GECCO 2017. ACM, 2017. Full version available at
http://arxiv.org/abs/1704.02191.
[DHK07]
Benjamin Doerr, Edda Happ, and Christian Klein. A tight bound
for the (1 + 1)-EA for the single source shortest path problem.
In Congress on Evolutionary Computation, CEC 2007, pages 1890–
1895. IEEE, 2007.
[DHK11]
Benjamin Doerr, Edda Happ, and Christian Klein. Tight analysis
of the (1+1)-EA for the single source shortest path problem. Evolutionary Computation, 19:673–691, 2011.
[DHK12]
Benjamin Doerr, Edda Happ, and Christian Klein. Crossover can
provably be useful in evolutionary computation. Theoretical Computer Science, 425:17–33, 2012.
72
[DJ07]
Benjamin Doerr and Daniel Johannsen. Adjacency list matchings:
an ideal genotype for cycle covers. In Genetic and Evolutionary
Computation Conference, GECCO 2007, pages 1203–1210. ACM,
2007.
[DJW02]
Stefan Droste, Thomas Jansen, and Ingo Wegener. On the analysis
of the (1+1) evolutionary algorithm. Theoretical Computer Science,
276:51–81, 2002.
[DJW12]
Benjamin Doerr, Daniel Johannsen, and Carola Winzen. Multiplicative drift analysis. Algorithmica, 64:673–697, 2012.
[DJWZ13]
Benjamin Doerr, Thomas Jansen, Carsten Witt, and Christine
Zarges. A method to derive fixed budget results from expected optimisation times. In Genetic and Evolutionary Computation Conference, GECCO 2013, pages 1581–1588. ACM, 2013.
[DK15]
Benjamin Doerr and Marvin Künnemann. Optimizing linear functions with the (1+λ) evolutionary algorithm—different asymptotic
runtimes for different instances. Theoretical Computer Science,
561:3–23, 2015.
[DL15]
Duc-Cuong Dang and Per Kristian Lehre. Simplified runtime analysis of estimation of distribution algorithms. In Genetic and Evolutionary Computation Conference, GECCO 2015, pages 513–518.
ACM, 2015.
[DLMN17]
Benjamin Doerr, Huu Phuoc Le, Régis Makhmara, and Ta Duy
Nguyen. Fast genetic algorithms. In Genetic and Evolutionary Computation Conference, GECCO 2017. ACM, 2017. Full version available at http://arxiv.org/abs/1703.03334.
[Doe11]
Benjamin Doerr. Analyzing randomized search heuristics: Tools
from probability theory. In Anne Auger and Benjamin Doerr, editors, Theory of Randomized Search Heuristics, pages 1–20. World
Scientific, 2011.
[Doe14]
Benjamin Doerr. A lower bound for the discrepancy of a random
point set. Journal of Complexity, 30:16–20, 2014.
[Doe17]
Benjamin Doerr. An elementary analysis of the probability that a
binomial random variable exceeds its expectation. ArXiv e-prints,
arXiv:1712.00519, 2017.
73
[Doe18]
Benjamin Doerr. Better runtime guarantees via stochastic domination. In Evolutionary Computation in Combinatorial Optimization,
EvoCOP 2018. Springer, 2018.
[dPdLDD15] Axel de Perthuis de Laillevault, Benjamin Doerr, and Carola Doerr.
Money for nothing: Speeding up evolutionary algorithms through
better initialization. In Genetic and Evolutionary Computation Conference, GECCO 2015, pages 815–822. ACM, 2015.
[DW14]
Benjamin Doerr and Carola Winzen. Ranking-based black-box complexity. Algorithmica, 68:571–609, 2014.
[Fei06]
Uriel Feige. On sums of independent random variables with unbounded variance and estimating the average degree in a graph.
SIAM Journal of Computing, 35:964–984, 2006.
[Fel68]
William Feller. An Introduction to Probability Theory and Its Applications, volume I. Wiley, third edition, 1968.
[GKS99]
Josselin Garnier, Leila Kallel, and Marc Schoenauer. Rigorous hitting times for binary mutations. Evolutionary Computation, 7:173–
203, 1999.
[GM14]
Spencer Greenberg and Mehryar Mohri. Tight lower bound on the
probability of a binomial exceeding its expectation. Statistics and
Probability Letters, 86:91–98, 2014.
[GW17]
Christian Gießen and Carsten Witt. The interplay of population size
and mutation probability in the (1 + λ) EA on OneMax. Algorithmica, 78:587–609, 2017.
[Hoe63]
Wassily Hoeffding. Probability inequalities for sums of bounded
random variables. Journal of the American Statistical Association,
58:13–30, 1963.
[HPR+ 14]
Hsien-Kuei Hwang, Alois Panholzer, Nicolas Rolin, Tsung-Hsi Tsai,
and Wei-Mei Chen. Probabilistic analysis of the (1+1)-evolutionary
algorithm. CoRR, abs/1409.4955, 2014.
[HY01]
Jun He and Xin Yao. Drift analysis and average time complexity of
evolutionary algorithms. Artificial Intelligence, 127:51–81, 2001.
[Jan17]
Swante Janson. Tail bounds for sums of geometric and exponential
variables. ArXiv e-prints, arXiv:1709.08157, 2017.
74
[JJW05]
Thomas Jansen, Kenneth A. De Jong, and Ingo Wegener. On the
choice of the offspring population size in evolutionary algorithms.
Evolutionary Computation, 13:413–440, 2005.
[JW01]
Thomas Jansen and Ingo Wegener. Evolutionary algorithms - how to
cope with plateaus of constant fitness and when to reject strings of
the same fitness. IEEE Transactions on Evolutionary Computation,
5:589–599, 2001.
[JW06]
Thomas Jansen and Ingo Wegener. On the analysis of a dynamic
evolutionary algorithm. Journal of Discrete Algorithms, 4:181–199,
2006.
[Kla00]
Bernhard Klar. Bounds on tail probabilities of discrete distributions.
Probability in the Engineering and Informational Sciences, 14:161–
171, 2000.
[Köt16]
Timo Kötzing. Concentration of first hitting times under additive
drift. Algorithmica, 75:490–506, 2016.
[KW17]
Martin S. Krejca and Carsten Witt. Lower bounds on the run time
of the univariate marginal distribution algorithm on OneMax. In
Foundations of Genetic Algorithms, FOGA 2017, pages 65–79. ACM,
2017.
[LN17]
Per Kristian Lehre and Phan Trung Hai Nguyen. Improved runtime bounds for the univariate marginal distribution algorithm via
anti-concentration. In Genetic and Evolutionary Computation Conference, GECCO 2017, pages 1383–1390. ACM, 2017.
[LOW17]
Andrei Lissovoi, Pietro Simone Oliveto, and John Alasdair Warwicker. On the runtime analysis of generalised selection hyperheuristics for pseudo-boolean optimisation. In Genetic and Evolutionary Computation Conference, GECCO 2017, pages 849–856.
ACM, 2017.
[LW12]
Per Kristian Lehre and Carsten Witt. Black-box search by unbiased
variation. Algorithmica, 64:623–642, 2012.
[LW14]
Per Kristian Lehre and Carsten Witt. Concentrated hitting times
of randomized search heuristics with variable drift. In 25th International Symposium on Algorithms and Computation, ISAAC 2014,
pages 686–697. Springer, 2014.
75
[McD98]
Colin McDiarmid. Concentration. In Probabilistic Methods for Algorithmic Discrete Mathematics, volume 16, pages 195–248. Springer,
Berlin, 1998.
[MS02]
Alfred Müller and Dietrich Stoyan. Comparison Methods for Stochastic Models and Risks. Wiley, 2002.
[MU05]
Michael Mitzenmacher and Eli Upfal. Probability and Computing—
Randomized Algorithms and Probabilistic Analysis. Cambridge University Press, 2005.
[Nag01]
S. V. Nagaev. Lower bounds on large deviation probabilities for
sums of independent random variables. Theory of Probability and
Its Applications, 46:79–102, 2001.
[NSW10]
Frank Neumann, Dirk Sudholt, and Carsten Witt. A few ants are
enough: ACO with iteration-best update. In Genetic and Evolutionary Computation Conference, GECCO 2010, pages 63–70. ACM,
2010.
[NW07]
Frank Neumann and Ingo Wegener. Randomized local search, evolutionary algorithms, and the minimum spanning tree problem. Theoretical Computer Science, 378:32–40, 2007.
[OHY09]
Pietro Simone Oliveto, Jun He, and Xin Yao. Analysis of the (1+1)EA for finding approximate solutions to vertex cover problems. IEEE
Transactions on Evolutionary Computation, 13:1006–1029, 2009.
[OLN09]
Pietro Simone Oliveto, Per Kristian Lehre, and Frank Neumann.
Theoretical analysis of rank-based mutation - combining exploration
and exploitation. In Congress on Evolutionary Computation, CEC
2009, pages 1455–1462. IEEE, 2009.
[OW11]
Pietro Simone Oliveto and Carsten Witt. Simplified drift analysis
for proving lower bounds in evolutionary computation. Algorithmica,
59:369–386, 2011.
[OW12]
Pietro Simone Oliveto and Carsten Witt. Erratum: Simplified
drift analysis for proving lower bounds in evolutionary computation.
CoRR, abs/1211.7184, 2012.
[OW15]
Pietro Simone Oliveto and Carsten Witt. Improved time complexity analysis of the simple genetic algorithm. Theoretical Computer
Science, 605:21–41, 2015.
76
[PR16]
Christos Pelekis and Jan Ramon. A lower bound on the probability
that a binomial random variable is exceeding its mean. Statistics
and Probability Letters, 119:305–309, 2016.
[Pro56]
Yuri Prokhorov. Convergence of random processes and limit theorems in probability theory. Theory of Probability and Its Applications, 1:157–214, 1956.
[PS97]
Alessandro Panconesi and Aravind Srinivasan. Randomized distributed edge coloring via an extension of the Chernoff–Hoeffding
bounds. SIAM Journal on Computing, 26:350–368, 1997.
[Rob55]
Herbert Robbins. A remark on Stirling’s formula. American Mathematical Monthly, 62:26–29, 1955.
[RT11]
Philippe Rigollet and Xin Tong. Neyman-Pearson classification, convexity and stochastic constraints. Journal of Machine Learning Research, 12:2831–2855, 2011.
[Sch00]
Christian Scheideler. Probabilistic Methods for Coordination Problems. University of Paderborn, 2000. Habilitation thesis. Available at
http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.70.1319.
[Slu77]
Eric V. Slud. Distribution inequalities for the binomial law. Annals
of Probability, 5:404–412, 1977.
[ST12]
Dirk Sudholt and Christian Thyssen. A simple ant colony optimizer
for stochastic shortest path problems. Algorithmica, 64:643–672,
2012.
[STW04]
Jens Scharnow, Karsten Tinnefeld, and Ingo Wegener. The analysis
of evolutionary algorithms on sorting and shortest paths problems.
Journal of Mathematical Modelling and Algorithms, 3:349–366, 2004.
[Sud13]
Dirk Sudholt. A new method for lower bounds on the running time of
evolutionary algorithms. IEEE Transactions on Evolutionary Computation, 17:418–435, 2013.
[SW16]
Dirk Sudholt and Carsten Witt. Update strength in EDAs and ACO:
how to avoid genetic drift. In Genetic and Evolutionary Computation Conference, GECCO 2016, pages 61–68. ACM, 2016. Extended
version CoRR abs/1607.04063.
[Tch67]
Pafnuty Tchebichef.
Des valeurs moyennes.
Journal de
mathématiques pures et appliquées, série 2, 12:177–184, 1867.
77
[Tch74]
Pafnuty Tchebichef. Sur les valeurs limites des intégrales. Journal
de mathématiques pures et appliquées, série 2, 19:157–160, 1874.
[Weg01]
Ingo Wegener. Theoretical aspects of evolutionary algorithms. In International Colloquium on Automata, Languages and Programming,
ICALP 2001, pages 64–78. Springer, 2001.
[Wit13]
Carsten Witt. Tight bounds on the optimization time of a randomized search heuristic on linear functions. Combinatorics, Probability
& Computing, 22:294–318, 2013.
[Wit14]
Carsten Witt. Fitness levels with tail bounds for the analysis of randomized search heuristics. Information Processing Letters, 114:38–
41, 2014.
[Wit17]
Carsten Witt. Upper bounds on the runtime of the univariate
marginal distribution algorithm on onemax. In Genetic and Evolutionary Computation Conference, GECCO 2017, pages 1415–1422.
ACM, 2017.
[WW05]
Ingo Wegener and Carsten Witt. On the optimization of monotone
polynomials by simple randomized search heuristics. Combinatorics,
Probability & Computing, 14:225–247, 2005.
[ZLLH12]
Dong Zhou, Dan Luo, Ruqian Lu, and Zhangang Han. The use of
tail inequalities on the probable computational time of randomized
search heuristics. Theoretical Computer Science, 436:106–117, 2012.
78
| 9cs.NE
|
arXiv:1709.06534v1 [cs.CR] 19 Sep 2017
BIOS ORAM: Improved Privacy-Preserving Data Access
for Parameterized Outsourced Storage
Michael T. Goodrich
University of California, Irvine
Dept. of Computer Science
[email protected]
Abstract
Algorithms for oblivious random access machine (ORAM) simulation allow a client, Alice, to
obfuscate a pattern of data accesses with a server, Bob, who is maintaining Alice’s outsourced
data while trying to learn information about her data. We present a novel ORAM scheme that
improves the asymptotic I/O overhead of previous schemes for a wide range of size parameters
for client-side private memory and message blocks, from logarithmic to polynomial. Our method
achieves statistical security for hiding Alice’s access pattern and, with high probability, achieves
an I/O overhead that ranges from O(1) to O(log2 n/(log log n)2 ), depending on these size
parameters, where n is the size of Alice’s outsourced memory. Our scheme, which we call BIOS
ORAM, combines multiple uses of B-trees with a reduction of ORAM simulation to isogrammic
access sequences.
1
Introduction
In outsourced storage applications, a client, Alice, outsources her data to a server, Bob, who stores
her data and provides her with an interface to access it via a network. We assume that Bob is
“honest-but-curious,” meaning that Bob is trusted to keep Alice’s data safe and available, but he
wants to learn as much as he can about Alice’s data. For privacy protection, we also assume that
Alice encrypts her data by a semantically secure encryption scheme so that each time she accesses an
item then she securely re-encrypts it before returning it to Bob’s storage. The remaining problem,
then, is for Alice to obscure her data access pattern so that Bob can learn nothing from her access
sequence.
Fortunately, there is a large and growing literature on algorithms for oblivious random access
machine (ORAM) simulation to obfuscate Alice’s access sequence (e.g., see [1,2,6,9–12,17,19,22,23,
25]). Such ORAM simulation methods provide ways for Alice to replace each of the data accesses
in her algorithm, A, with a sequence of accesses to her data stored with Bob so as to obfuscate
her original access sequence. Ideally, such an ORAM scheme achieves statistically security, which
intuitively means that Bob is unable to determine any information about Alice’s original access
sequence beyond its length, N , and the size of her data set, n. In addition to the parameters, n
and N , the following parameters are also important in this context:
• B: The maximum number of words in a message block sent from/to Alice in one I/O
operation.
1
• M : The number of words in Alice’s client-side private memory.
In this paper, we are interested in scenarios where B and M can be set to arbitrary values that
are at least logarithmic in n, and at most a constant fraction of n, so as to apply to a wide range
of scenarios, while still applying to cases where Alice would still be motivated to outsource her
memory to Bob. Formally, in an ORAM framework, we assume that Alice’s RAM algorithm, A,
indexes data using integer addresses in the range [0, n − 1], using the following operations:
• write(i, v): Write the value v into the memory cell indexed by the integer, i. Since A is a
RAM algorithm, we assume i and v each fit in a single memory word.
• read(i): Read and return the value, v, stored in the cell with integer address, i.
These operations are the low-level accesses that are issued during Alice’s execution of her RAM
algorithm, A. The goal of an ORAM simulation scheme is to allow Alice to perform her algorithm,
A, but to replace each individual read or write operation with a sequence of input/output (I/O)
messages exchanged between Alice and Bob so that Alice effectively hides A’s pattern of data
accesses, i.e., A’s sequence of read/write operations. Moreover, we would like to minimize the
amortized number of such messages exchanged between Alice and Bob while still preserving Alice’s
privacy.
A related concept is that of an oblivious storage (OS), e.g., see [1, 2, 12, 19, 23, 24]. In this
framework, Alice stores a dictionary at the server, Bob, of size at most n, and her algorithm, A,
accesses this dictionary using the following operations:
• put(k, v): Add the key-value item, (k, v). An error occurs if there is already an item with this
key. We assume here that each key, k, fits in a single memory word and each value, v, fits in
a message block of size B.
• get(k): Return and remove the value, v, associated with the key, k. If there is no item in the
collection with key k, then return a special “not found” value.
Note that OS includes ORAM as a special case. For example, we can initialize A’s memory,
A[0..n−1], as put(i, A[i]), for i = 0, . . . , n−1. Then we can perform any write(i, v) as get(i), put(i, v),
and we can perform any read(i) as v = get(i), put(i, v). We assume that Alice’s original access
sequence has a given length, N , where N is at most polynomial in n, and an OS scheme replaces each
such operation with a sequence of operations that obfuscate the original access sequence. Ideally,
Bob should learn nothing about Alice’s original access sequence, which we formalize in terms of a
security game. Let σ denote a sequence of N read/write operations (or get/put operations). An
ORAM (resp., OS) scheme transforms σ into into a sequence, σ 0 , of access operations. As mentioned
above, we assume that each item is stored using a semantically-secure encryption scheme, so that
independent of whether Alice wants to keep a key-value item unchanged, the sequence σ 0 involves
always replacing anything Alice accesses with a new encrypted value so that Bob is unable to tell
if the underlying plaintext value has changed. The security for an ORAM (or OS) simulation is
defined in terms of the following security game. Let σ1 and σ2 be two different RAM-algorithm or
dictionary access sequences, of length N , for a key/index set of size n, that are chosen by Bob and
given to Alice. Alice chooses uniformly at random one of these sequences and transforms it into the
access sequence σ 0 according to her ORAM (or OS) scheme, which she then executes. Her ORAM
(resp., OS) scheme is statistically secure if Bob can determine which sequence, σ1 or σ2 , Alice chose
2
with probability at most 1/2. This assumes that Bob learns nothing from the encryption of Alice’s
data.
The I/O overhead for such an OS or ORAM scheme is a function, T (n), such that the total
number of messages sent between Alice and Bob during the simulation of all N of her accesses from
σ is O(N · T (n)) with high probability (i.e., with probability at least 1 − 1/nc , for some constant
c > 1). That is, the I/O overhead is the amortized expected number of accesses to Bob’s storage
that are done to hide each of Alice’s original accesses in her algorithm, A. Of course, we would like
T (n) to be as small as possible.
In this paper, we provide methods for improving the asymptotic I/O overhead for ORAM
simulations by more than just constant factors. The approach we take to achieve this goal is to
first transform the original RAM access sequence, σ, into an intermediate OS sequence, σ̂, which
has a restricted structure that we refer to as it being isogrammic, and we then efficiently implement
an oblivious storage for this isogrammic sequence, σ̂, transforming it into a final access sequence,
σ 0 , by taking advantage of this restricted structure. We define a sequence, σ = (σ1 , σ2 , . . . , σN )
of put and get operations to be isogrammic 1 , for an underlying set of size, n, if it satisfies the
following conditions:
• For every get(k) operation, there is a previous put(k, v) operation, with the same key, k.
• No put(k, v) operation attempts to add an item (k, v) when the key k is already in the set.
• The key, k, used in each get(k) (resp., put(k, v)) operation contains random component chosen
independently and uniformly of at least dlog ne bits. Thus, keys are unlikely to be repeated.
That is, each time we issue a put(k, v) operation, the key, k, contains a random nonce that is
chosen independently from any of the previous random nonces we chose for previous keys. At a
high level, then, our two-phase ORAM simulation scheme is surprisingly simple, in that it combines
the classic well-known B-tree data structure, which is ubiquitous in database applications, with
isogrammic OS. That is, at a high level our scheme can be summarized as
B-trees + Isogrammic-OS =⇒ ORAM.
Thus, we call our scheme BIOS ORAM.2 This approach allows us to achieve the main goals of this
paper, which is the design of ORAM simulation methods that work for a wide range of values to
the parameters B and M . Moreover, we are able to use this approach to design schemes that are
statistically secure and, w.h.p., have efficient I/O overhead bounds.
1.1
Previous Related Results
Work on ORAM simulation methods traces its origins to seminal work of Goldreich and
Ostrofsky [9], who achieve an I/O overhead of O(log3 n) with M and B being O(1) using a scheme
1
An isogram is a word, like “copyrightable,” without a repeated letter.
E.g., see wikipedia.org/wiki/Isogram.
2
Our scheme implements an ORAM by a reduction to an isogrammic OS, i.e., by replacing a sequence of read
and write operations with a sequence of get and put operations. If one desires a scheme that works entirely in the
ORAM framework, one can, for example, implement these get and put operations using a cuckoo hash table, with
O(1) lookup times in the worst case and O(1) amortized insertion times with high probability (w.h.p.). This will
result in a sequence of read and write operations and it does not reveal any information about Alice’s original access
sequence, since the get and put operations come from an OS.
3
Method
Parameterizable?
Statistically Secure?
B
M
I/O Overhead
Goldreich-Ostrofsky [9]
No
No
Θ(1)
Θ(1)
O(log3 n)
Kushilevitz et al. [17]
No
No
Θ(1)
Θ(1)
O(log2 n/ log log n)
Damgård et al. [6]
No
Yes
Θ(1)
Θ(1)
O(log3 n)
Supermarket ORAM [3]
No
Yes
Θ(1)
Θ(polylog n)
O(log2 n log log n)
Somewhat
No
Θ(1)
Θ(n )
O(log n)
No
Θ(n )
O(1)
Goodrich-Mitzenmacher [10]
Melbourne shuffle [19]
Somewhat
Θ(n )
Path ORAM [25]
Yes
Yes
ω(log n)
ω(B log n)
O(log n/ log B)†
BIOS ORAM
Yes
Yes
Ω(log n)
Ω(log n)
O(log2 n/ log2 B)
2
Table 1: Our BIOS ORAM bounds (in boldface), compared to some of the asymptotically
best previous ORAM methods, which are distinguished between those results have parameterized
message/memory sizes or not, and those that are statistically secure or not. The parameter
0 < ≤ 1/2 is a fixed constant. The above results that are not statistically secure assume the
existence of random one-way hash functions, i.e., they assume the existence of random oracles.
† The Path ORAM method [25] claims an I/O bandwidth of O(log n) blocks, but this assumes that
blocks can contain the responses of multiple back-and-forth messages; hence, we use the above
bound to characterize the I/O overhead for Path ORAM, which counts the actual number of
messages, each of size at most B.
that fails with polynomial probability and is not statistically secure. Kushilevitz et al. [17] improve
the I/O overhead for ORAM with a constant-size client-side memory to be O(log2 n/ log log n),
albeit while still not achieving statistical security. Damgård et al. [6] introduce an ORAM scheme
that is statistically secure with an I/O overhead that is O(log3 n), with M and B being O(1),
that is, there method is not parameterized for general values of M and B. Chung et al. [3]
provide a statistically secure ORAM scheme, which we are calling “supermarket” ORAM (due to
its reliance on an interesting “supermarket” analysis), for the case when B is O(1) and M is at least
polylogarithmic, which has an I/O overhead of O(log2 n log log n). Unfortunately, these previous
schemes do not apply to parameterized scenarios with larger values for B, such as when B is at
least logarithmic, let alone for cases when B is Ω(n ), for some constant 0 < ≤ 1. Interestingly,
Goldreich and Ostrofsky [9] give a lower-bound argument that the I/O overhead for an ORAM
scheme must be Ω(log n) when M is O(1), but their lower bound does not apply to larger values
of M and B.
There is previous work that is parameterized for larger values of M and B, however. Goodrich
and Mitzenmacher [10] provide an ORAM simulation scheme that achieves an O(log n) I/O overhead
and constant-sized messages, but their method requires M to be Ω(n ), for some fixed constant
0 < ≤ 1, which is not fully parameterized, e.g., for when M is polylogarithmic. Also, their
method is not statistically secure. Stefanov et al. [25] introduce the Path ORAM method, which is
statistically secure and parameterized for values of B that are super-logarithmic and values of M
that are at least a logarithmic factor larger than B, to achieve an I/O overhead (in terms of the
number of messages exchanged between Alice and Bob) of O(log2 n/ log B). That is, their method
also can match the logarithmic lower bound of Goldreich and Ostrofsky, but with a scheme that
requires both M and B to be Ω(n ), for a fixed constant 0 < ≤ 1. Ohrimenko et al. [19] present
4
an oblivious storage (OS) scheme (which, as we observed, can also be used for ORAM simulation)
that achieves an I/O overhead of O(1), but it is not statistically secure and it requires M and B
to be at least Ω(n ).
Wang et al. [27] introduce an interesting “oblivious data structure” framework, which applies
to bounded-degree data structures, such as search trees, to achieve an O(log n) I/O overhead for
data-structure access sequences. Their algorithms are based on (non-recursive) Path ORAM [25],
however, which requires that M and B be super-logarithmic, and, even then, their method does
not achieve an O(1) I/O overhead, even for larger values of B and M .
In addition, there is a growing literature on other ORAM and OS solutions, which is too large to
review here (e.g., see [1,7,8,11–13,20–24,26]). These results also optimize I/O overhead, but we are
not aware of any previous results that beat the asymptotic I/O bounds for the Path ORAM scheme
for a wide range of values of B and M while also achieving statistical security for the simulation
method.
1.2
Our Results
We provide a method for ORAM simulation, which we call BIOS ORAM, that achieves statistical
security and has efficient asymptotic I/O overheads for a wide range of values for the parameters
B and M . In particular, we show how to perform an ORAM simulation of a polynomial number of
accesses to an outsourced storage of size n with an I/O overhead that is O(log2 n/ log2 B), w.h.p.,
for B and M ranging from logarithmic to a fraction of n. For example, we can achieve the following
specific bounds, depending on the values of B and M :
• When B and M are logarithmic or polylogarithmic in n, we achieve an I/O overhead that is
O(log2 n/(log log n)2 ), w.h.p.
√
• When B and M are only Ω(2
log n ),
we achieve an I/O overhead that is O(log n), w.h.p.
• When B and M are O(n ), for some constant 0 < ≤ 1/2, we achieve an I/O overhead that
is O(1), w.h.p.
We summarize our results in Table 1, comparing them to some of best-known previous ORAM
results. Note, for example, that our results apply to a wider range of values of the parameters B and
M than the Path ORAM scheme [25] and improves the I/O overhead for ORAM simulation over
this entire range. For example, the best I/O overhead that Path ORAM can achieve is O(log n) even
when B and M are O(n ), whereas our BIOS ORAM scheme achieves a constant I/O overhead
in such scenarios. In addition, our I/O overhead bounds match those of the Melbourne shuffle
for values of B that are Θ(n ) while also extending to values of B that are smaller than those
possible √using the Melbourne shuffle approach. For example, as mentioned above, if B and M are
just Ω(2 log n ), then we achieve an I/O overhead of O(log n), w.h.p., which is a result that is not
achievable using previous ORAM methods for such values of B and M .
Our methods are remarkably simple and make multiple uses of the ubiquitous B-tree data
structure (e.g., see [4,5,16]), along with a reduction of ORAM simulations to isogrammic OS access
sequences, as well as efficient ways of obliviously simulating isogrammic access sequences (again,
using B-trees). In particular, we show how to implement an oblivious storage (OS) scheme for
any isogrammic access sequence so as to achieve a simulation that achieves statistical security and
has an I/O overhead that is O(log n/ log B) with high probability. In addition, we show how to
5
apply this result to improve the I/O overhead for oblivious tree-structured data structures, which
improves an oblivious data-structure bound of Wang et al. [27] and may be of independent interest.
2
An Overview of B-Trees
As is well-known in database circles, a B-tree is a multi-way search tree, which stores internal nodes
as blocks so that its depth is O(logB n), e.g., see [4, 5, 16]. In the B-trees we use in this paper, we
choose a branching factor of
B 0 = B 1/4 ,
where B is our message-size parameter. Such a B-tree supports searching and updates (insertions
and deletions) in O(logB 0 n) = O(log n/ log B 0 ) = O(log n/ log B) I/Os of blocks of size B 0 . That
is, each search or update involves accessing O(1) nodes on each level of the B-tree, in a root-to-leaf
search followed (for updates) by a leaf-to-root set of updates, for which we refer the interested reader
to known methods for searching and updating B-trees (e.g., see [4,5,16]). From the perspective of the
server, Bob, the I/Os for searching a B-tree would simply look like Alice accessing O(log n/ log B 0 ) =
O(log n/ log B) blocks of storage. (See Figure 1.)
An example B-tree:
a
b
c
e
f
g
h
d
i
j
Figure 1: An example B-tree with branching factor 3.
permission.
3
k
l
m
c 2017 Michael Goodrich. Used with
B-Trees + Isogrammic OS = ORAM
In this section, we describe the first component of our BIOS ORAM scheme, which is a reduction of
ORAM simulation to isogrammic OS, at the cost of increasing the I/O overhead of Alice’s accesses
by a factor of O(log n/ log B). That is, we show how to transform an arbitrary sequence of read and
write operations into an isogrammic sequence of get and put operations, with a blow-up in length
of O(log n/ log B). By then showing how to obliviously simulate an isogrammic access sequence
with an I/O overhead of O(log n/ log B), we get the main result of this paper, that is, that we can
achieve an ORAM scheme with an I/O overhead of O((log n/ log B)2 ).
Suppose, then, that Alice’s RAM algorithm, A, which she wishes to perform on her data
outsourced to Bob, uses a memory of n cells indexed by integers in the range [0, n − 1]. Let R be
a B-tree having each cell of Alice’s storage stored in a sub-block of size B 0 = B 1/4 associated with
a block at a leaf of R (ordered in the standard left-to-right fashion). Furthermore, let R have a
branching factor of B 0 = B 1/4 , so each internal node in R can be stored in a single sub-block of
size B 0 and the depth of R is O(log n/ log B 0 ) = O(log n/ log B). Intuitively, the main idea of our
reduction is that, for each write(i, v) or read(i) operation in A, we perform a search in R for the
6
index i, to find the sub-block containing memory cell, i, and then we replace this sub-block of size
B 0 and all the nodes of size B 0 that we just traversed with new nodes.
Initially, we construct R in a bottom-up fashion so that each node u in R is assigned a random
nonce, ru , of dlog ne bits. For each leaf, u, of R, which is storing some block, V , of B values for
the cells in Alice’s storage for some set of indices, {i, i + 1, . . . , i + B 0 − 1}, we issue a put(k, v)
operation, where k = (ru , u) and v = (i, V ). In addition, we (obliviously) store ru at u in R on
the server. Note that v fits in a single block of size B. For each internal node, u, of R, which we
construct level-by-level, so that we can obliviously read in the random nonce, rui , associated with
each child, ui , of u in R, then we assign u a random nonce, ru , and we issue a put(k, v) operation,
where k = (ru , u) and
v = (I, ru1 , u1 , ru2 , u2 , . . . , ruB0 −1 , uB 0 −1 ),
where I is the block of key values needed to decide for any search which child, ui , of u to access
next. Note that v fits in O(1) sub-blocks of size B 0 . This initialization phase establishes the B-tree,
R, in Bob’s storage and issues O(n/B 0 ) put operations that identify each node, u, of R using a key
that comprises a random nonce, ru , of dlog ne bits. We also keep a global variable, that maintains
the random nonce for the root.
For each read(i) or write(i, v) operation after this initialization, we traverse a root-to-leaf path,
π, in R to the leaf associated with the sub-block holding the cell i. This involves performing a
sequence of O(log n/ log B) get(k) operations, where each k is a pair of a node name in R and the
random nonce for that node. We cache the nodes returned by these get operations in Alice’s private
memory. Then, processing the nodes in π in reverse order, we give each node u in this sequence a
new random nonce, ru , and we issue a put(k, v) operation, where k = (ru , u) and
v = (I, ru1 , u1 , ru2 , u2 , . . . , ruB0 −1 , uB 0 −1 ),
where u1 , . . . , uB 0 −1 are the children of u in R. We do this processing in reverse order so that when
we issue such a put(k, v) operation, we will have available the new random nonce for the child, uj ,
of u that we previously processed as well as the old (and still unused) random nonces for the other
children of u. In this case, the nodes are B-tree nodes, which are of size B 1/4 , as we have defined
this to be the branching factor for our B-tree. Fortunately, our message block size, B, is large
enough to store up to B 3/4 such nodes in a single message block.
Each read or write involves issuing O(log n/ log B) get and put operations; hence, this increases
the total number of I/Os for Alice’s access sequence by an O(log n/ log B) factor. The important
observation is that this process results an isogrammic access sequence, since each key used in a
put(k, v) operation comprises a random nonce of at least dlog ne bits, each such key is not already in
our set (since each key also comprises a unique node name), and each get(k) operation is guaranteed
to match up with a previous put(k, v) operation. Thus, we have the following.
Theorem 1. Given a RAM algorithm, A, with memory size, n, we can simulate the memory
accesses of A using an isogrammic access sequence that initially creates O(n/B 0 ) put operations
and then creates O(log n/ log B) get and put operations for each step of A. Each key used in a get
or put operation comprises a random nonce of at least dlog ne bits and each value used in a put
operation is a sub-block of size O(B 0 ) words.
Proof. For the security claim, consider a simulation of the security game mentioned in the
introduction, assuming the statistical security for our isogrammic OS. Suppose, then, that Bob
7
R:
a
b
e
f
c
g
h
i
O(log n/log B)
I/O overhead
d
j
k
l
m
Our Isogrammic OS method
Another
O(log n/log B)
I/O overhead
Figure 2: A high-level view of our BIOS ORAM scheme. c 2017 Michael Goodrich. Used with
permission.
creates two access sequences, σ1 and σ2 , and gives them to Alice, who then chooses one at random
and simulates it, as described above. For each access to a memory index, i, in the RAM simulation
for her chosen σj , the memory cell for i is read and written to by doing a search in R. The
important observation is that this access consists of O(log n/ log B) accesses a root-to-leaf sequence
of nodes of R, indexed by newly-generated independent random numbers each time. Thus, nothing
is revealed to Bob about the index, i. That is, the number of accesses in Alice’s simulation is the
same for σ1 and σ2 , and the sequence of keys used is completely independent of the choice of σ1 or
σ2 . Thus, Bob is not able to determine which of these sequences she chose with probability better
than 1/2.
As we show in the remainder of this paper, we can simulate an isogrammic access sequence
obliviously in a statistically secure manner with an I/O overhead that is O(log n/ log B) with
high probability. Combining this result with Theorem 1 gives us our claimed result that we can
perform statistically-secure ORAM simulation with an I/O overhead that, with high probability, is
O(log2 n/ log2 B). (See Figure 2.)
4
B-tree OS for Small Sets
In this section, we present an OS solution for small sets, that is, sets whose size, n, is O(B 3/2 ), and
small items, that is, items whose size is O(B 0 ) words, where B 0 = B 1/4 . This is admittedly a fairly
restrictive scenario, but it nevertheless is a critical component of our isogrammic OS scheme. We
show how, in this scenario, to use a B-tree to achieve statistical security for an OS simulation that
works for general sequences of put and get operations, not just isogrammic sequences. That is, in
this subsection, we allow for keys that are not necessarily random (so long as they are unique) and
we allow get(k) operations to return “not found” responses.
Suppose, then, that we wish to support an oblivious storage (OS) for a set of items that can
be as large as n = Θ(B 3/2 ). In this case, we utilize a B-tree, F , with branching factor, B 0 = B 1/4 ,
so its height is O(log n/ log B 0 ) = O(log n/ log B) = O(1), since we are restricting ourselves here to
sets of size at most O(B 3/2 ) and items of size O(B 0 ). Put another way, our restriction on the set
size, n, implies that B is Ω(n2/3 ) and B 0 is Ω(n1/6 ), that is, that B is Ω(B 1/4 n1/2 ).
8
√
Our small-set B-tree OS method is a modification and adaptation of the “ n ” solution of
Damgård et al. [6] to B-trees and the OS setting. Let F be our B-tree with capacity n and
branching factor B 0 ; hence, F has depth D = 4dlog n/ log Be, with n items stored in its leaves,
and randomly shuffled in an array of Bob’s storage of size O(nB 0 ) memory words. Initially, we
construct F using an initial set of items, which could even be empty. We pad the nodes of F with
empty dummy nodes, as necessary, to make every node of F have the same depth, D. An internal
B-tree node consists of B 0 keys; hence, a single message block can fit B 3/4 B-tree nodes or items
(since items are also of size B 0 here), that is, Ω(n1/2 ) B-tree nodes or items. In addition, we also
√
store a singly linked list, `, of Dd ne nodes, which are the same size as B-tree nodes and items
and are randomly shuffled in with the B-tree nodes of F .
We can initialize F in this way using the oblivious shuffling method of Goodrich and
Mitzenmacher [10]. In this context, where we are obliviously sorting nodes and items that are
themselves of size B 0 , their method involves a multi-way merging of sorted lists with a branching
factor of (M/B 0 )1/3 = Θ(n1/6 ), where the merging step involves a simple oblivious scanning of each
of these arrays. The merge in their method requires, in this context (where we are merging nodes
and items of size B 0 = Θ(n1/6 )) that there be
Ω((M/B 0 )2/3 B 0 + (M/B 0 )1/3 (B 0 )2 )
values stored in memory at any given time, which we can bound as Ω(n1/2 ) using the fact that, in
this case,
(M/B 0 )2/3 B 0 + (M/B 0 )1/3 (B 0 )2
is at least (n1/2 )2/3 n1/6 + (n1/2 )1/3 n1/3 , which equals 2n1/2 . Thus, we can scan each of the Θ(n1/6 )
sorted lists by reading Θ(n1/2 ) elements from each sorted list at a time (and stopping the recursion
when we reach lists of this size). This means that the total number of I/Os needed for this oblivious
shuffling is O((n/n1/2 ) log2M/B 0 (n/B 0 )) I/Os. Given our other assumptions about B and M , this
shuffling therefore requires at most O(n1/2 ) I/Os. In addition, we maintain D caches, C1 , . . . , CD ,
√
of size B 0 n each, one for each level of F . Thus, each cache can be read or written in O(1) I/Os.
square-root size
F
C3
C2
C1
a
a
b
c
f
b
c
d
h
e
f
g
h
i
j
k
l
m
Figure 3: A B-tree, F , and a hierarchy of Θ(log n/ log B) caches, C1 , C2 . . . . We shade the nodes
that have been accessed previously in grey. Each cache is associated with a specific level of F .
c 2017 Michael Goodrich. Used with permission.
Let us consider each type of access, with the design of making it impossible for Bob to determine
even if we are performing a get or put operation. In either a get(k) or put(k, v) operation, we begin
9
with a search for the key k in F . To perform such a search in F , we read each of the nodes in a
path, π, from the root of F to the leaf containing our search key, k, or its predecessor (if k is not in
our set, S). For each level, i, of F during this search, we first read (as one I/O), the cache, Ci , to
see if the i-th node, vi , of π is in Ci . If vi is in Ci , then we examine it and determine the location
of the next node, vi+1 , in π, and we read the next dummy node in ` (for the sake of obliviousness,
since the location of this dummy node looks random to Bob and has not been previously accessed).
If vi is not in Ci , then we read it in (this is the first time we are accessing vi and this location looks
to Bob to be random). We note that each such node is of size B 0 ; hence, it and each cache can be
read or written in O(1) I/Os. (See Figure 3.)
After we have completed the reading of all the nodes in π, we can perform any updating as
necessary for these nodes so as to perform the functionality of our get(k) or put(k, v) operation.
Without going into details (e.g., see [4, 5, 16]), either of these operations will either involve no
structural changes to F or will involve our adding O(1) nodes per level of F . Let π 0 denote the
set of updated nodes in F (which we can determine using Alice’s private memory). Note that each
internal node of F in π 0 will including pointers to existing nodes in F , but these have not yet been
accessed yet and we have not revealed any information about them to Bob. We then write the
nodes of π 0 out to Bob’s storage, placing each node, vi , on level of i of π, in the cache, Ci , using
O(1) I/Os for each level of F . In fact, we pad this set so that we always write the same number of
O(1) nodes to each Ci , based on standard update rules for B-trees (e.g., see [4, 5, 16]). We perform
this process in a bottom-up leaf-to-root fastion, so that we can inductively always be able to know
the locations for the child nodes for any node in F (even if that node is in a cache and some of its
children are in the lower-level cache). Thus, we can determine the locations in Bob’s storage for
any root-to-leaf path in F by reading the nodes and caches in Bob’s storage sequentially starting
√
from the root. After we have completed d ne accesses of F in this manner, we rebuild and reshuffle
F and a new linked list, `, and repeat this B-tree access procedure. This rebuilding and reshuffling
requires O(n1/2 ) I/Os, as described above. Thus, we have the following.
Theorem 2. Suppose we have a set, S, of up to n items, where each item is of size at most B 0 ,
where B 0 = B 1/4 , and n is O(B 3/2 ). Then our B-tree OS solution can implement an oblivious
storage for S that has an I/O overhead of O(log n/ log B) = O(1), with high probability. This
simulation is statistically secure, even for non-isogrammic access sequences.
Proof. With respect to the security of this method, note that each access involves a search of all
√
the caches of size n and an access to a distinct random location (which is either a real node or a
dummy node that Bob cannot tell apart) for each level of a shuffled B-tree, F , which was shuffled
obliviously and has every possible permutation of its nodes on that level as being equally likely. We
then access each cache for every level of F in a bottom-up fashion. Moreover, both the top-down
and bottom-up phases of this computation involve the same form of access irrespective of whether
we are performing a get or put operation. Thus, the adversary, Bob, can learn nothing about Alice’s
access sequence based on observing her access pattern. That is, in terms of the security game, Bob
is unable to distinguish between two access sequences, σ1 and σ2 , of length N for sets of up to n
items.
Since D is O(log n/ log B), and B is large enough for us to read an entire cache with one I/O, it
is easy to see that each access in this simulation requires O(log n/ log B) = O(1) I/Os. In addition,
√
after we have performed O( n) such accesses, we do a rebuilding action that requires O(n1/2 ) I/Os;
√
hence, this adds an amortized O(1) number of I/Os for each of the previous O( n) accesses. Thus,
10
the total I/O overhead is O(log n/ log B) = O(1), with high probability (where this probability is
dependent only on the algorithm we use for oblivious shuffling, e.g., see [10]).
This solution is a crucial component of our general isogrammic OS solution, which we describe
next.
5
Isogrammic Oblivious Storage
In this section, we describe our isogrammic OS scheme, which is able to obfuscate any isogrammic
access sequence with statistical security, achieving an I/O overhead that is O(log n/ log B) with high
probability. Our construction involves yet another use of B-trees, as a primary search structure, as
well as repeated uses of our B-tree OS for small sets from Section 4.
H:
a
b
1
2
c
3
4
5
d
6
7
8
9
Figure 4: An illustration of our isogrammic OS. The B-tree tree, H, is shown in black. Each
bucket, which implements our B-tree OS for small sets, is shown as a blue triangle. That is, each
triangle represents a bucket of up to B 3/2 items, which are accessed according to our B-tree OS
method of Section 4. c 2017 Michael Goodrich. Used with permission.
Let n be the size of a set for which we wish to support an oblivious storage (OS), and let H
be a static B-tree of height O(log n/ log B) with branching factor B 0 = B 1/4 , such that H has n/B
leaves. For every node, u in H, including both the internal nodes and leaves, we store a “bucket,”
bu , which maintains an instance of our B-tree OS, as described above in Section 4, and let each
such bucket have capacity 4L, where L = B 3/2 , except for leaves, which each have capacity 8L.
These buckets are used to store (k, v) items that are in the current set, i.e., items for which we
have processed a put(k, v) operation and have yet to perform a get(k) operation. See Figure 4.
Recall that in an isogrammic access sequence we are given a sequence of put(k, v) and get(k) such
that get(k) operations always have an item to return (i.e., there is a previous matching put(k, v)
operation) and put(k, v) operations never try to insert an item whose key matches the key of an
existing item. More importantly, every key contains a random nonce component of at least dlog ne
bits. We use these random nonces as the addresses for where items should go in H. Namely, we
maintain the following invariant throughout our OS simulation:
• For each item, (k, v), in our current set of items, (k, v) is stored in exactly one bucket, bu , for
a node, u, on the root-to-leaf search path in H for the random part of k.
11
Given this invariant, let us describe how we process put and get operations.
For a put(k, v) operation, we add (k, v) to the bucket, br , for the root, r, of H, using the B-tree
OS method described in Section 4. Note that this satisfies our invariant for storing items in H,
that is, storing an item in bucket for the root implies that it is stored in the root-to-leaf search path
for the random part of its key. (We will describe later what we do when the root bucket becomes
full, but that too will satisfy our invariant.) Then, for the sake of obliviousness (so Bob cannot tell
whether this operation is a get or put), we uniformly and independently choose a random key, k 0 ,
and traverse the root-to-leaf path in H for k 0 , performing a search for k 0 in the bucket, bu , for each
node u on this path, using the B-tree OS method described in Section 4. Alice just “throws away”
the results of these searches, but, of course, Bob doesn’t know this.
For any given get(k) operation, we begin, for the sake of obliviousness, by inserting a dummy
item, (k 0 , e), in the bucket, br , for the root, r, of H, where e is a special “empty” value (that
nevertheless has the same size as any other value) and k 0 is a random key, using the fusion-tree OS
method described in Section 4. So as to distinguish this type of dummy item from others, we refer
to each such dummy item as an original dummy item. We then traverse the root-to-leaf path, π,
for (the random part of) k in H, and, for each node, u, in π, we search in the bucket, bu , for u,
to see if the key-vaue pair for k is in this bucket, using the B-tree OS scheme described above in
Section 4. By our invariant, the item, (k, v), must be stored in the bucket for one of the nodes
in the path π. Note that we search in the bucket for every node in π, even after we have found
and removed the key-value pair, (k, v). Because we are simulating an isogrammic access sequence,
there will be one bucket with this item, but we search all the buckets for the sake of obliviousness.
An important consequence of the above methods and the fact that we are simulating an
isogrammic access sequence is that each traversal of a path in H is determined by a random
nonce that is chosen uniformly at random and is independent of every other nonce used to do a
search in H. Thus, the server, Bob, learns nothing about Alice’s access pattern from these searches.
In addition, as we will see shortly, the server cannot determine where any item, (k, v), is actually
stored, because the random part of the key k is only revealed when we do a get(k) operation and
put operations never reveal the locations of their keys. Moreover, we maintain the fact that the
server doesn’t know the actual location of any item, along with our invariant, even as bucket for a
node, u, becomes full and needs to have its items distributed to its children.
Periodically, so as to avoid overflowing buckets, we move items from a bucket, bu , stored at a
node u in H to u’s children, in a process we call a flush operation. In particular, we flush the root
node, r, every L put or get operations. We flush each internal node, u, after u has received B 0 flushes
from its parent, which each involve inserting exactly 4L/B 0 real and dummy items (including new
dummy items) into the bucket for u. Because of this functionality, and the fact that we are moving
items based on random keys, the number of real and original dummy items in the bucket, bu , at
a time when we are flusing a node u at depth i is expected to be L, and we maintain it to be at
most 4L. Also, note that we will periodically perform flush operations across all the nodes on a
given level of H at any given time when flush operations occur, which is the main reason why our
I/O overhead bounds are amortized. We don’t flush the leaf nodes in H, however. Instead, after
every leaf, u, in H has received B 0 flushes, we perform an oblivious compression to compress out
a sufficient number of dummy items so that the number of real and dummy items in u’s bucket is
4L. Thus, the bucket for a leaf never grows to have more than 8L real and dummy items. If, at
the time we are compressing the contents of a leaf bucket, we determine that there are more than
4L real items being stored in such a bucket, which, as we show, is an event that occurs with low
12
probability, then we restart the entire OS simulation. Such an event doesn’t compromise privacy,
since it depends only on random keys, not Alice’s data or access sequence. Thus, doing a restart
just impacts performance, but because restarts are so improbable, our I/O bounds still hold with
high probability.
At a high-level, our method for doing a flush operation at a node, u, in H has a similar structure
to an analogous operation in the Path ORAM scheme [25], as well as in the paper mentioned above
that is currently under submission for ORAM simulation when B and M are both very small. The
details for our flush operation here are different than both of these works, however, in that our
flush method depends crucially on the B-tree OS method of Section 4.
1. We obliviously shuffle the real and original dummy items of bu into an array, A, of size 4L,
stored at the server. This step will never overflow A (because of how we perform the rest
of the steps in a flush operation). This step can be done using known oblivious shuffling
methods (e.g., see [10]), which add just a constant I/O overhead factor.
2. For each child, xi , i = 1, 2, . . . , B 0 , of u, we create an array, Ai , of size 4L/B 0 .
3. We obliviously sort the real and original dummy items from A into the arrays, A1 , . . . , A` ,
according the keys for these items, so that the item, (k, v), goes to the array Ai if the next
O(log B 0 ) bits of the random part of key k would direct a search for k to the child xi . We
perform this oblivious sorting step so that if there are fewer than 4L/B 0 items destined for any
array, Ai , we pad the array with (new) dummy items to bring the number of items destined
to each array, Ai , to be exactly 4L/B 0 . However, if we determine from this oblivious sorting
step that there are more than 4L/B 0 real and original dummy items destined for any array,
Ai , which (as we show) is an event that occurs with low probability, then we restart the entire
OS simulation. Because this step is done obliviously and search keys are random (hence, they
never depend on Alice’s data values or access pattern), even if we restart, Bob learns nothing
about Alice’s access sequence during this step. So, let us assume that we don’t restart. This
step can be done using known oblivious sorting, padding, and partitioning methods (e.g.,
see [10]), which add only a constant I/O overhead factor.
4. For each real and dummy item (including both original and new dummy items), (k, v), in each
Ai , we insert (k, v) into the bucket bxi using the B-tree OS method of Section 4. This method
works only for small sets, but, of course, the number of items in each bucket determines such
a small set.
The first important thing to note about a flush operation is that it is guaranteed to preserve
our invariant that each item, (k, v), is stored in the bucket of a node in H on the root-to-leaf path
determined by the random part of k. Moreover, because we move real and original dummy items to
children nodes obliviously, in spite of our invariant, the server never knows where an item, (k, v),
is stored; hence, the server can never differentiate two access sequences more than at random.
Let us analyze the complexity of a flush operation. Since we flush the root every L steps, and
we flush every other node, u, at depth i, after it has received B 0 flushes, and both real and original
dummy items are mapped to u only if the first i log B 0 bits of each of their random keys matches
u’s address, the expected number of real and original dummy items stored in the bucket for u is
at most L at the time we flush u. In fact, this is a rather conservative estimate, since it assumes
that none of these items were removed as a result of get operations. More importantly, we have
the following.
13
Lemma 3. The number, f , of real and original dummy items flushed from a node, u, to one of
its children, xi , is never more than 4L/B 0 , with high probability. Likewise, a leaf in H will never
receive more than 4L real items, with high probability.
Proof. The expected value of f , which can be expressed as a sum of independent indicator random
variables, is at most
L/B 0 = B 3/2−1/4 = B 5/4 ≥ d log5/4 n,
for a constant, d ≥ 3, since we are assuming that B is Ω(log n). Thus, by a Chernoff bound (e.g.,
see [18]),
5/4
1/4
Pr(f ≥ 4L/B 0 ) ≤ e−L/B ≤ e−d log n ≤ n−3 log n .
The probability bound argument for a leaf in H is similar. The lemma follows, then, by a union
bound across all nodes of H and the polynomial length of access sequences.
Thus, with high probability, we never need to do a restart as a result of a potential overflow
during a flush operation.
Theorem 4. We can obliviously simulate an isogrammic sequence of a polynomial number of
put(k, v) and get(k) operations, for a data set of size n, with an I/O overhead of O(log n/ log B),
with high probability. Moreover, this simulation is statistically secure.
Proof. The height of the tree, H, is O(log n/ log B). Thus, by Theorem 2, with high probability, the
I/O overhead is proportional to a constant times O(log n/ log B), which is itself O(log n/ log B). For
the security claim, consider an instance of the simulation game, where Bob chooses two isogrammic
access sequences, σ1 and σ2 , of length N for a key set of size n, and gives them to Alice, who then
chooses one uniformly at random and simulates it according to the isogrammic OS scheme. Each
access that she does involves accessing a sequence of nodes of H determined by random keys and
for each node doing a lookup in an OS scheme that is itself statistically secure, by Theorem 2.
In addition, put operations add items at the top bucket and are obfuscated with data-oblivious
flush operations. Therefore, Bob is not able to distinguish between σ1 and σ2 any better than at
random.
6
Our BIOS ORAM Algorithm
Putting the above pieces together, then, gives us the following theorem, which is the main result
of this paper.
Theorem 5. Given a RAM algorithm, A, with memory size, n, where n is a power of 2, we can
simulate the memory accesses of A in an oblivious fashion that achieves statistical security, such
that, with high probability, the I/O overhead is O(log2 n/ log2 B) for a client-side private memory
of size M ≥ B and messages of size B ≥ 3 log n.
Proof. By Theorem 1, each access in A gets expanded into O(log n/ log B) operations in an
isogrammic access sequence, and, with high probability, each such operation has an overhead of
O(log n/ log B), by Theorem 4. The security claim follows from the security claims of Theorems 1
and 4.
14
7
Isogrammic Algorithm Design
In this section, we study the expressive power of the isogrammic access sequences, showing that it
subsumes some previous specialized design patterns for implementing algorithms in the cloud in a
privacy-preserving way. Thus, by Theorem 4, any algorithm designed in this framework, to give
rise to an isogrammic access sequence, can be simulated in an oblivious fashion to have an I/O
overhead that is O(log n/ log B), with high probability.
There are a number of previous algorithm-engineering design paradigms that can facilitate
privacy-preserving data access in the cloud, which, as we show, can be reduced to isogrammic
access sequences at only a constant cost per operation. Thus, the observations made in this section
may be of independent interest for these specialized applications.
7.1
Simulating Oblivious Data Structures
The first application we explore is for bounded-degree directed data structures in the oblivious data
structure framework of Wang et al. [27]. This framework applies to any data structure that has a
small number of “root” nodes for tree structures with bounded out-degree, such that updates and
accesses are done as a sequence of linked nodes starting from a root. Using a heuristic similar to
that used by Wang et al. [27], we can make any such access sequence isogrammic. Namely, let us
keep a random nonce, ru , of dlog ne bits for each node, u, in our data structure, and let us assign
the key for accessing a node u to be the pair, (ru , u). That is, any other node that points to u
will identify u using the pair (ru , u). The important observation is that any access sequence can
inductively be able to access nodes with their nonces, because bounded-degree data structures are
accessed starting from a root node; hence, any set of node updates performs its operations on a
path from a root and we can update each node on such a path in reverse order to have its new
random nonce. More importantly, for each node, x, that points to node u, we can also update x to
change its pointer to u to have u’s new nonce. Using such nonces as keys, therefore, gives rise to
an isogrammic access sequence; hence, our result from Theorem 4 applies to such scenarios. This
implies the existence of efficient oblivious simulations for access sequences involving stacks, queues,
and deques (which have just Θ(1) node updates per operation), as well as binary trees, such as
AVL trees and red-black trees (whose updates and searches can be padded to have Θ(log n) node
updates per operation). We summarize as follows.
Theorem 6. Any algorithm, A, written in the oblivious data structure framework, with boundeddegree tree nodes reachable from a constant number of “root” nodes, can be implemented as an
algorithm, A0 , in the isogrammic algorithm design paradigm such that each data access in A is
translated into O(1) accesses in A0 .
For example, we can create an isogrammic queue, Q, by using an array, A, of size n and a
dummy array, B, of size n, together with three global “root” indices, front, rear, and dummy.
Each time we access Q, for an enqueue, dequeue, or no-op operation, we read the front, rear, and
dummy variables. If this is a no-op operation (or this would be an error operation, like doing a
dequeue from an empty queue), then we next read the next dummy slot in B and we increment
the dummy variable. If this is a valid enqueue, then we increment rear and add the new element to
that location in A. If this is a valid dequeue, then we increment front and read the element from
the previous front location in A. Anytime we wrap around A or B, we increment a version counter
15
(associated with the global variables), so that we access the cells in A or B by index and version
counter. This implies that we always access the queue using an isogrammic access sequence.
7.2
Compressed Scanning
In addition, the compressed-scanning paradigm [14,15] also falls into our framework for isogrammic
access sequences. A compressed-scanning algorithm consists of t rounds, where each round involves
accessing each of the elements of a set, S, of n data items exactly once in a read-compute-write
operation. By introducing random nonces and assigning them to items for each round, we can easily
transform any algorithm in the compressed-scanning model into an isogrammic access sequence.
Thus, all of the graph algorithms presented in these papers [14, 15] can be simulated obliviously
with our isogrammic OS scheme. We summarize as follows.
Theorem 7. Any algorithm, A, written in the compressed-scanning framework can be implemented
as an algorithm, A0 , in the isogrammic algorithm design paradigm such that each data access in A
is translated into O(1) accesses in A0 .
8
Conclusion
In this paper we have shown how to improve the I/O overhead for statistically secure ORAM
simulation for a wide range of parameterized values for the client-side private memory size, M ,
and the size, B, of message blocks. Our results imply I/O overhead bounds that range from O(1)
to O(log2 n/(log log n)2 ), with high probability. For example, we can achieve an I/O overhead of
O(log
√ n), with high probability, for statistically secure ORAM simulation if B and M are at least
Ω(2 log n ), which is asymptotically smaller than n , for any constant 0 < ≤ 1.
For future work, it would be interesting to see if there is a super-logarithmic lower bound for the
I/O overhead for ORAM simulation for cases when B and M are small. Also, it would interesting
to see if it is possible to achieve √
an I/O overhead for statistically secure ORAM
simulation that is
√
O(log n) when B and M are o(2 log n ), i.e., asymptotically smaller than 2 log n .
Acknowledgments
This research was supported in part by the National Science Foundation under grant 1228639.
This article also reports on work supported by the Defense Advanced Research Projects Agency
(DARPA) under agreement no. AFRL FA8750-15-2-0092. The views expressed are those of the
authors and do not reflect the official policy or position of the Department of Defense or the
U.S. Government. We would like to thank Eli Upfal and Marina Blanton for helpful discussions
regarding the topics of this paper.
References
[1] D. Apon, J. Katz, E. Shi, and A. Thiruvengadam. Verifiable oblivious storage. In 17th Int.
Conf. on Practice and Theory in Public-Key Cryptography (PKC), volume 8383 of LNCS,
pages 131–148. Springer, 2014.
16
[2] D. Boneh, D. Mazières, and R. A. Popa. Remote oblivious storage: Making oblivious RAM
practical. Technical report, MIT CSAIL, 2011.
http://dspace.mit.edu/handle/1721.1/62006.
[3] K.-M. Chung, Z. Liu, and R. Pass. Statistically-secure ORAM with Õ(log2 n) overhead. In
20th ASIACRYPT, pages 62–81, 2014.
[4] D. Comer. Ubiquitous B-tree. ACM Comput. Surv., 11(2):121–137, June 1979.
[5] T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein. Introduction to Algorithms. The
MIT Press, 3rd edition, 2009.
[6] I. Damgård, S. Meldgaard, and J. B. Nielsen. Perfectly secure oblivious RAM without
random oracles. In 8th Theory of Cryptography Conference (TCC), volume 6597 of LNCS,
pages 144–163. Springer, 2011.
[7] S. Devadas, M. Dijk, C. W. Fletcher, L. Ren, E. Shi, and D. Wichs. Onion ORAM: A
constant bandwidth blowup oblivious RAM. In 13th Int. Conf. on Theory of Cryptography
(TCC), volume 9563 of LNCS, pages 145–174. Springer, 2016.
[8] C. Gentry, K. A. Goldman, S. Halevi, C. Julta, M. Raykova, and D. Wichs. Optimizing
ORAM and using it efficiently for secure computation. In 13th Int. Symp. on Privacy
Enhancing Technologies (PETS), pages 1–18. Springer, 2013.
[9] O. Goldreich and R. Ostrovsky. Software protection and simulation on oblivious RAMs. J.
ACM, 43(3):431–473, May 1996.
[10] M. T. Goodrich and M. Mitzenmacher. Privacy-preserving access of outsourced data via
oblivious RAM simulation. In 28th Int. Colloq. on Automata, Languages and Programming
(ICALP), volume 6756 of LNCS, pages 576–587. Springer, 2011.
[11] M. T. Goodrich, M. Mitzenmacher, O. Ohrimenko, and R. Tamassia. Oblivious RAM
simulation with efficient worst-case access overhead. In 3rd Cloud Computing Security
Workshop, pages 95–100, 2011.
[12] M. T. Goodrich, M. Mitzenmacher, O. Ohrimenko, and R. Tamassia. Practical oblivious
storage. In 2nd ACM Conf. on Data and Application Security and Privacy (CODASPY),
pages 13–24, 2012.
[13] M. T. Goodrich, M. Mitzenmacher, O. Ohrimenko, and R. Tamassia. Privacy-preserving
group data access via stateless oblivious RAM simulation. In 23rd ACM-SIAM Symposium
on Discrete Algorithms (SODA), pages 157–167, 2012.
[14] M. T. Goodrich, O. Ohrimenko, and R. Tamassia. Graph drawing in the cloud: Privately
visualizing relational data using small working storage. In 20th Int. Symp. on Graph Drawing
(GD), volume 7704 of LNCS, pages 43–54. Springer, 2013.
[15] M. T. Goodrich and J. A. Simons. Data-oblivious graph algorithms in outsourced external
memory. In 8th Int. Conf. on Combinatorial Optimization and Applications (COCOA),
volume 8881 of LNCS, pages 241–257. Springer, 2014.
17
[16] M. T. Goodrich and R. Tamassia. Algorithm Design and Applications. Wiley Publishing, 1st
edition, 2014.
[17] E. Kushilevitz, S. Lu, and R. Ostrovsky. On the (in)security of hash-based oblivious RAM
and a new balancing scheme. In 23rd ACM-SIAM Symposium on Discrete Algorithms
(SODA), pages 143–156, 2012.
[18] M. Mitzenmacher and E. Upfal. Probability and computing: Randomized algorithms and
probabilistic analysis. Cambridge University Press, 2005.
[19] O. Ohrimenko, M. T. Goodrich, R. Tamassia, and E. Upfal. The Melbourne shuffle:
Improving oblivious storage in the cloud. In Int. Colloq. on Automata, Languages, and
Programming (ICALP), volume 8573 of LNCS, pages 556–567. Springer, 2014.
[20] L. Ren, C. W. Fletcher, A. Kwon, E. Stefanov, E. Shi, M. van Dijk, and S. Devadas. Ring
ORAM: Closing the gap between small and large client storage oblivious RAM, 2014.
Cryptology ePrint Archive, Report 2014/997, http://eprint.iacr.org/.
[21] L. Ren, X. Yu, C. W. Fletcher, M. van Dijk, and S. Devadas. Design space exploration and
optimization of path oblivious RAM in secure processors. In 40th ACM Int. Symp. on
Computer Architecture (ISCA), pages 571–582, 2013.
[22] E. Shi, T.-H. H. Chan, E. Stefanov, and M. Li. Oblivious RAM with O((log N )3 ) worst-case
cost. In 17th Int. Conf. on the Theory and Application of Cryptology and Information
Security (ASIACRYPT), volume 7073 of LNCS, pages 197–214. Springer, 2011.
[23] E. Stefanov and E. Shi. Multi-cloud oblivious storage. In 2013 ACM Conf. on Computer and
Communications Security (CCS), pages 247–258, 2013.
[24] E. Stefanov and E. Shi. ObliviStore: High performance oblivious cloud storage. In IEEE
Symp. on Security and Privacy (SP), pages 253–267, May 2013.
[25] E. Stefanov, M. van Dijk, E. Shi, C. Fletcher, L. Ren, X. Yu, and S. Devadas. Path ORAM:
An extremely simple oblivious RAM protocol. In ACM Conf. on Computer and
Communications Security (CCS), pages 299–310, 2013.
[26] X. Wang, H. Chan, and E. Shi. Circuit ORAM: On tightness of the Goldreich-Ostrovsky
lower bound. In 22nd ACM Conf. on Computer and Communications Security (CCS), pages
850–861, 2015.
[27] X. S. Wang, K. Nayak, C. Liu, T.-H. H. Chan, E. Shi, E. Stefanov, and Y. Huang. Oblivious
data structures. In ACM Conf. on Computer and Communications Security (CCS), pages
215–226, 2014.
18
| 8cs.DS
|
arXiv:1701.06849v1 [math.AC] 24 Jan 2017
SYMMETRIES AND CONNECTED COMPONENTS OF THE
AR-QUIVER
TONY J. PUTHENPURAKAL
Abstract. Let (A, m) be a commutative complete equicharacteristic Gorenstein isolated singularity of dimension d with k = A/m algebraically closed.
Let Γ(A) be the AR (Auslander-Reiten) quiver of A. Let P be a property
of maximal Cohen-Macaulay A-modules. We show that some naturally defined properties P define a union of connected components of Γ(A). So in this
case if there is a maximal Cohen-Macaulay module satisfying P and if A is
not of finite representation type then there exists a family {Mn }n≥1 of maximal Cohen-Macaulay indecomposable modules satisfying P with multiplicity
e(Mn ) > n. Let Γ(A) be the stable quiver. We show that there are many
symmetries in Γ(A). Furthermore Γ(A) is isomorphic to its reverse graph. As
an application we show that if (A, m) is a two dimensional Gorenstein isolated singularity with multiplicity e(A) ≥ 3 then for all n ≥ 1 there exists an
indecomposable self-dual maximal Cohen-Macaulay A-module of rank n.
1. introduction
Let us recall that a commutative Noetherian local ring (A, m) is called an isolated
singularity if AP is a regular local ring for all prime ideals P 6= m. We note
that with this definition if A is Artinian then it is an isolated singularity. This is
not a usual practice, nevertheless in this paper Artin rings will be considered as
isolated singularities. Also recall that if a local Noetherian ring (B, n) is Henselian
then it satisfies Krull-Schmidt property, i.e., every finitely generated B-module is
uniquely a direct sum of indecomposable B-modules. Now assume that B is CohenMacaulay. Then we say B is of finite (Cohen-Macaulay) representation type if B
has only finitely many indecomposable maximal Cohen-Macaulay B-modules. In a
remarkable paper Auslander proved that in this case B is an isolated singularity,
for instance see [23, Theorem 4.22].
To study (not necessarily commutative) Artin algebra’s Auslander and Reiten introduced the theory of almost-split sequences. These are now called AR-sequences.
The AR-sequences are organized to form the AR-quiver. Later Auslander and Reiten extended the theory of AR-sequences to the case of commutative Henselian
isolated singularities.
If (A, m) is a Henselian Cohen-Macaulay isolated singularity then we denote its
AR-quiver by Γ(A). A good reference for this topic is [23]. The motivation for
this paper comes from the following crucial fact about AR quivers (under some
conditions on A), see [23, 6.2]:
Date: January 25, 2017.
1991 Mathematics Subject Classification. Primary 13 C14, Secondary 13H10, 14B05.
Key words and phrases. Artin-Reiten Quiver, Hensel rings, indecomposable modules, Ulrich
modules, periodic modules, Non-periodic modules with bounded betti numbers.
1
2
TONY J. PUTHENPURAKAL
If C is a non-empty connected component of Γ(A) and if A is not of finite
representation type then there exist a family {Mn }n≥1 of maximal Cohen-Macaulay
indecomposable modules in C with multiplicity e(Mn ) > n.
Let P be a property of maximal Cohen-Macaulay A-modules. We show that
some naturally defined properties P define a union of connected components of the
AR quiver of A. Thus the above mentioned observation still holds. Therefore if
there is a maximal Cohen-Macaulay module satisfying P then there exists a family
{Mn }n≥1 of maximal Cohen-Macaulay indecomposable modules satisfying P with
multiplicity e(Mn ) > n.
1.1. For the rest of the paper let us assume that (A, m) is a complete equicharacteristic Gorenstein isolated singularity of dimension d. Assume k = A/m is
algebraically closed. Some of our results are applicable more generally. However
for simplicity we will make this hypothesis throughout this paper. We will also
assume that A does not have finite representation type. This is automatic if A is
not a hypersurface, see [23, 8.15]. Furthermore if A is a hypersurface ring with
dim A ≥ 2 and e(A) ≥ 3 then also A is not of finite representation type.
Now we describe our results. We first describe our results on
Connected Components of the AR-quiver:
I: Modules with periodic resolution:
Let M be a maximal Cohen-Macaulay non-free A-module. Let Syzn (M ) be the
nth -syzygy module of M . We say M has periodic resolution if there exists a nonnegative integer n and a positive integer p with Syzn+p (M ) ∼
= Syzn (M ). The
smallest p for which this holds is called the period of the resolution. We say M has
property P if it has a periodic resolution.
If A is a hypersurface ring then any non-free maximal Cohen-Macaulay A-module
has periodic resolution with period ≤ 2 and in fact Syz3 (M ) ∼
= Syz1 (M ). There
exits maximal Cohen-Macaulay modules with periodic resolutions if A is a complete
intersection of any codimension c ≥ 1. Again it can be shown that in this case the
period is ≤ 2 and in fact Syz3 (M ) ∼
= Syz1 (M ).
For general Gorenstein local rings there is no convenient criterion to determine
when A has a module with perodic resolution (however see [11, 5.8] for a criterion).
It was conjectured by Eisenbud that if a module M has a periodic resolution then
the period is ≤ 2, see [13, p. 37]. This was disproved by Gasharov and Peeva, see
[14, Theorem 1.3].
Our first result is
Theorem 1.2. (with hypotheses as in 1.1.) If A is not a hypersurface ring then P
defines a union of connected components of Γ(A).
We now give more refined versions of Theorem 1.2:
1.3. Assume A = Q/(f ) where (Q, n) is a Gorenstein local ring and f ∈ n2 is a
Q-regular element.
Let M be a maximal Cohen-Macaulay non-free A-module. We say M has property PQ if projdimQ M finite. In this case it is easy to prove that M has a periodicresolution over A with period ≤ 2. There is essentially a unique method to construct
non-free maximal Cohen-Macaulay modules over A having finite projective dimension over Q. This is essentially due to Buchweitz et al, see [10, 2.3]. Also see the
paper [17, 1.2] by Herzog et al. Our next result is:
SYMMETRIES AND CONNECTED COMPONENTS OF THE AR-QUIVER
3
Theorem 1.4. [ with hypotheses as in 1.3.] If Q is not regular then PQ defines a
union of connected components of Γ(A).
Again our results implies existence of indecomposable maximal Cohen-Macaulay
A-modules with arbitrarily high multiplicity and satisfying property PQ . However
our method does not give a way to construct these modules.
1.5. Eisenbud’s conjecture (as stated above) is valid if M has the so-called finite
CI-dimension [3, 7.3]. We say M has property PO if M has finite CI-dimension
over A and has a periodic resolution over A. We say M has property PE if M has
periodic resolution over A but it has infinite CI dimension over A. Our next result
is
Theorem 1.6. [with hypotheses as in 1.1.] Assume A is not a hypersurface.
(1) PO defines a union of connected components of Γ(A).
(2) PE defines a union of connected components of Γ(A).
We note that in [14, 3.1] a family Aα of an Artininian Gorenstein local ring
is constructed with each having a single module Mα having periodic resolution
of period > 2 is given. As the period of Mα is greater than two it cannot have
finite CI-dimension over Aα . Thus our result implies existence of indecomposable
modules with arbitrary length, having a periodic resolution and having infinite CI
dimension over Aα .
Note that till now our results does not give any information regarding period’s.
In dimension two we can say something, see Theorem 1.15.
1.7. Now assume that A is a complete intersection of codimension c ≥ 2. There
is a theory of support varieties for modules over A. Essentially for every finitely
generated module E over A an algebraic set V (E) in the projective space Pc−1 is
attached, see [4, 6.2]. Conversely it is known that if V is an algebraic set in Pc−1
then there exists a finitely generated module E with V (E) = V , see [7, 2.3]. It is
known that V (Syzn (E)) = V (E) for any n ≥ 0. Thus we can assume E is maximal
Cohen-Macaulay. If E has periodic resolution over A then V (E) is a finite set of
points. The converse is also true. If further E is indecomposable then V (E) is
a singleton set, see [7, 3.2]. Let a ∈ Pc−1 . We say a maximal Cohen-Macaulay
A-module M satisfies property Pa if V (M ) = {a}. We prove:
Theorem 1.8. [ with hypotheses as in 1.7.] Let a ∈ Pc−1 . Then Pa defines a
union of connected components of Γ(A). Conversely if C is a non-empty connected
component of Γ(A) containing a periodic module M then for any [N ] ∈ C we have
V (N ) = V (M )(= {p}). In particular Γ(A) has at least |k| connected components.
II Modules with bounded betti-numbers but not having a periodic resolution:
For a long time it was believed that if a module M has a bounded resolution (i.e.,
there exists c with βi (M ) ≤ c for all i ≥ 0) then it is periodic. If A is a complete
intersection then modules having bounded resolutions are periodic [13, 4.1]. In [14,
3.2] there are examples of modules M having a bounded resolution but M is not
periodic.
If M is a maximal Cohen-Macaulay A-module having a bounded resolution but
M is not periodic then we say that M has property BN P . We prove
Theorem 1.9. [with hypotheses as in 1.1.] BN P defines a union of connected
components of Γ(A).
4
TONY J. PUTHENPURAKAL
We note that if M has bounded resolution but not periodic then there exists c
with e(Syzn (M )) ≤ c for all n ≥ 0. Our result implies the existence of modules
with bounded but not periodic resolution of arbitrary multiplicity.
III: Ulrich modules:
Let M be a maximal Cohen-Macaulay A-module. It is well-known that e(M ) ≥
µ(M ) (here µ(M ) denotes the cardinality of a minimal generating set of M ) A maximal Cohen-Macaulay module M is said to be an Ulrich module if its multiplicity
e(M ) = µ(M ). In this case we say M has property U.
If dim A = 1 then A has a Ulrich module. It is known that if A is a strict
complete intersection of any dimension d then it has a Ulrich module, see [17,
2.5]. In particular if A is a hypersurface ring then it has a Ulrich module. There
are some broad class of examples of Gorenstein normal domain (but not complete
intersections) of dimension two that admit an Ulrich module see [8, 4.8]. However
there are no examples of Gorenstein local rings R ( but not complete intersections)
with dim R ≥ 3 such that R admits an Ulrich module (note we are not even insisting
that R is reduced).
Even if A is a hypersurface there is essentially a unique way to construct an
Ulrich modules. We show
Theorem 1.10. [ with hypotheses as in 1.1.] Further assume that either A is a
hypersurface ring of even dimension d ≥ 2 and multiplicity e(A) ≥ 3 OR A is
Gorenstein of dimension two. Then U defines a union of connected components of
Γ(A).
The reason we cannot say anything about Ulrich modules over hypersurface rings
of odd dimension is due to a peculiar nature of AR-sequences, see remark 8.2. Also
note that if e(A) = 2 then any non-free MCM A-module is an Ulrich module.
We now describe our result on:
Symmetries of AR-quiver:
Let Γ0 (A) be the connected component of Γ(A) containing the vertex [A]. Set
e
Γ(A)
= Γ(A) \ Γ0 (A). Let Γ(A) denote the stable AR-quiver of A, i.e., we delete
the vertex [A] from Γ(A) and all arrows connecting to [A]. Also set Γ0 (A) to be
the stable part of Γ0 (A).
Our starting point is the observation that for Klienian singularities Γ(A) is trivially isomorphic to its reverse graph (see [23, p. 95]). Recall if G is a directed graph
then it’s reverse graph Grev is a graph with the same vertices as G and there is an
arrow from vertex u to v in Grev if and only if there is an arrow from vertex v to
u in G. In fact we construct
Theorem 1.11. [with hypotheses as in 1.1.] There exists isomorphisms
D, λ : Γ(A) → Γ(A)rev as graphs. If A is not a hypersurface ring then
(1) D 6= λ.
(2) There exists indecomposable maximal Cohen-Macaulay modules M, N with
λ(M ) 6= M and D(N ) 6= N .
The first isomorphism D is just the dual functor. The next map λ arises in the
theory of horizontal linkage defined by Martsinkovsky and Strooker, see [20, p. 592].
We note that the assumption A not a hypersurface is essential for the later part of
Theorem 1.11, for in the case of Klienian singularities it is known that λ(M ) = M
for each non-free indecomposable M , see [20, Theorem 3].
SYMMETRIES AND CONNECTED COMPONENTS OF THE AR-QUIVER
5
1.12. For n ≥ 0 let Syzn be the nth syzygy functor. As A is Gorenstein we can
also define for integers n ≤ −1 the nth cosyzygy functor which we again denote
with Syzn . By the definition of horizontal linkage we have Syz−1 ◦D = λ. Thus
Syz−1 = λ ◦ D−1 and Syz1 = D ◦ λ−1 . So under the assumptions as in 1.1 we get
that Syzn : Γ(A) → Γ(A) is an isomorphism for all n ∈ Z. We prove:
Theorem 1.13. [ with hypotheses as in 1.1] Let C be a connected component of
Γ(A). For [M ] ∈ C, set I(M ) = {n | [Syzn (M )] ∈ C}. Then
(1) I(M ) is an ideal in Z (possibly zero).
(2) I(N ) = I(M ) for all [N ] ∈ C.
If A is not of finite representation type then there is practically no information on
connected components of Γ(A). The only case known is when A is a hypersurface
e
there is information on connected components of Γ(A),
see [12, Theorem I]. It is
easy to show that Γ0 (A) has only finitely many components. As an application of
Theorem 1.13 we show:
Corollary 1.14. [with hypotheses as in 1.1.] Assume further that A is not a
hypersurface ring. Let D be a connected component of Γ0 (A). Then
(1) D has infinitely many vertices.
(2) The function [M ] → e(M ) is unbounded on D.
Structure of the AR-quiver:
If A is of finite representation type then the structure of the AR-quiver is known,
see [23]. For hypersurface rings which are not of finite representation type there is
some information regarding connected components of A not containing the vertex
[A]. For two dimensional Gorenstein rings we show:
Theorem 1.15. [with hypothesis as in 1.1.] Assume dim A = 2 and e(A) ≥ 3. Let
C be a non-empty component of Γ(A). Then C is of the form
M1 ⇆ M2 ⇆ M3 ⇆ M4 ⇆ · · · ⇆ Mn ⇆ · · ·
where e(Mn ) = ne(M1 ) for all n ≥ 1. Furthermore
(1) If C = Γ0 (A) then M1 = A. Furthermore Mn∗ ∼
= Mn for all n ≥ 1.
(2) Assume now that A is not a hypersurface ring. Then
(a) If Mj is periodic with period c for some j then Mn is periodic with period
c for all n ≥ 1.
(b) Let C denote the stable part of C. Let [Mi ] ∈ C. If the Poincare series of
Mi is rational then the Poincare series of M is rational for all [M ] ∈ C.
Furthermore all of them share a common denominator.
In the Theorem above the Poincare series PM (z) of a module M is
A
n
n≥0 dimk Torn (M, k)z . We also note that the structure of all components of
Γ(A) \ Γ0 (A) is already known, see [6, 4.16.2].
We have several interesting consequences of Theorem 1.15. A direct consequence
of this theorem is that if dim A = 2 and e(A) ≥ 3 then for all n ≥ 1 there exist’s an
indecomosable maximal Cohen-Macaulay A-module Mn of rank n with Mn∗ ∼
= Mn .
I do not know whether such a result holds for higher dimensional rings.
A simple consequence of Theorem’s 1.15 and 1.13 is the following:
P
Corollary 1.16. [with hypotheses as in 1.1.] Assume A is not a hypersurface ring.
Also assume dim A = 2. Then Syzn (Γ0 (A)) are distinct for all n ∈ Z, n 6= 0.
6
TONY J. PUTHENPURAKAL
We now describe in brief the contents of this paper. In section two we discuss
some preliminary results that we need. In section three we discuss lifts of irreducible
maps. In the next section we dicuss non-free indecomosable summands of maximal
Cohen-Macaulay approximation of the maximal ideal. In section five we give proof’s
of Theorem’s 1.2, 1.4 and 1.6. In the next section we give a proof of Theorem 1.8. In
section seven we discuss our notion of quasi AR-sequences and in the next section
we give a proof of Theorem 1.10. In section nine we prove Theorem 1.11. In
the next section we give a few obstructions to existence of quasi-AR sequences.
In section 11 we describe the describe Γ0 (A) when A is a two dimensional with
e(A) ≥ 3. In section twelve we give a proof of Theorem 1.15 and Corollary 1.16. In
the last section we discuss curvature and complexity of MCM modules and as an
application give a proof of Theorem 1.9
Remark 1.17. Srikanth Iyengar informed me about the excellent paper [15] where
the authors considered AR-quiver of self-injective Artin algebra’s. Note that commutative Artin Gorenstein rings is an extremely special case of self-injective Artin
Algebra’s. So our results in this case is sharper than that of [15]. I do not beleive
that the results of this paper when A is commutative Artin Gorenstein ring will
hold for the more general case of self-injective Artin algebra’s. I also thank Dan
Zacharia and Lucho Avramov for some useful discussions.
2. Some Preliminaries
In this paper all rings will be Noetherian local. All modules considered are finitely
generated. Let (A, m) be a local ring and let k = A/m be its residue field. If M is
an A-module then µ(M ) = dimk M/mM is the number of a minimal generating set
of M . Also let ℓ(M ) denote its length. In this section we discuss a few preliminary
results that we need.
th
2.1. Let M be an A-module.
βi (M ) = dimk TorA
i (M, k) be its i
P For i ≥ 0 let
n
betti-number. Let PM (z) = n≥0 βn (M )z , the Poincare series of M . Set
cx(M ) = inf{d | lim sup
βn (M )
< ∞}
nd−1
and
1
curv M = lim sup(βn (M )) n .
It is possible that cx(M ) = ∞, see [5, 4.2.2]. However curv(M ) is finite for any
module M [5, 4.1.5]. It can be shown that if cx(M ) < ∞ then curv(M ) ≤ 1.
2.2. It can be shown that for any A-module M we have
cx(M ) ≤ cx(k)
and
curv(M ) ≤ curv(k).
see [5, 4.2.4].
2.3. If A is a complete intersection of co-dimension c then for any A-module M
we have cx(M ) ≤ c. Furthermore for each i = 0, . . . , c there exists an A-module
Mi with cx(Mi ) = i. Also note that cx(k) = c. [5, 8.1.1(2)]. If A is a complete
intersection and M is a module with cx(M ) = cx(k) then we say M is extremal.
2.4. If A is not a complete intersection then curv(k) > 1. [5, 8.2.2]. In this case
we say a module M is extremal if curv(M ) = curv(k).
SYMMETRIES AND CONNECTED COMPONENTS OF THE AR-QUIVER
7
2.5. Let M, N be A-modules and let f : M → N be A-linear. Let FM be a minimal
resolution of M and let FN be a minimal resolution of N . Then f induces a lift
fe: FM → FN . This map fe is unique up to homotopy. The chain map fe induces an
A-linear map fn : Syzn (M ) → Syzn (N ) for all n ≥ 1.
2.6. Denote by β(M, N ) the set of A-homomorphisms of M to N which pass
through a free module. That is, an A-linear map f : M → N lies in β(M, N ) if and
only if it factors as
M❇
❇❇
❇❇f
u
❇❇
❇
/N
F
v
where F ∼
= A for some n ≥ 1.
n
Remark 2.7. If f : M → N is A-linear and if f1 , f1′ : Syz1 (M ) → Syz1 (N ) are two
lifts of f then it is well known and easily verified that f1 −f1′ ∈ β(Syz1 (M ), Syz1 (N )).
Recall that an A-module M is called stable if M has no free summands. We
need the following:
Proposition 2.8. Let (A, m) be a Noetherian local ring and let M, N be A-modules.
Set Λ = EndA (M ) and r = rad Λ. Let f ∈ Λ. Also suppose there exists A-linear
maps u : M → N and v : N → M . Set g = v ◦ u.
(1) If f (M ) ⊆ mM then f ∈ r.
(2) If M is stable and f ∈ β(M, M ) then f ∈ r.
(3) If 1 − g ∈ r then u is a split mono and v is a split epi.
Proof. (1) This is well known.
(2) Assume f = β ◦ α where α : M → F and β : F → M and F ∼
= An . As M
is stable it follows that α(M ) ⊆ mF and so f (M ) ⊆ mM . The result follows from
(1).
(3) Let 1 − g = h where h ∈ r. Then g = 1 − h is invertible in Λ. So there exists
τ ∈ Λ with τ ◦ g = g ◦ τ = 1M . The result follows.
2.9. Now assume A is Cohen-Macaulay. Let M, N be maximal Cohen-Macaulay
A-modules and let f : M → N be A-linear. Recall f is said to be irreducible if
(1) f is not a split epimorphism and not a split monomorphism.
(2) If X is a maximal Cohen-Macaulay A-module and if there is a commutative
diagram
M❇
❇❇
❇❇f
u
❇❇
❇
X v /N
then either u is a split monomorphism or v is a split epimorphism.
Remark 2.10. Let (A, m) be a Cohen-Macaulay local ring. If M is a maximal
Cohen-Macaulay A-module then it is easy to verify that Syzn (M ) is stable for each
n ≥ 1.
8
TONY J. PUTHENPURAKAL
2.11. Suppose M is a maximal Cohen-Macaulay over a local Gorenstein ring A.
Then M ∗ = HomA (M, A) is also a maximal Cohen-Macaulay module. Furthermore
ExtiA (M, A) = 0 for i > 0. We also have (M ∗ )∗ ∼
= M . Notice if
α
α
2
1
M3 → 0,
M2 −→
0 → M1 −→
is a short exact sequence of maximal Cohen-Macaulay A-modules then we have the
following short exact sequence
α∗
α∗
2
1
0 → M3∗ −−→
M2∗ −−→
M1∗ → 0,
of maximal Cohen-Macaulay A-modules.
Remark 2.12. Let (A, m) be a Gorenstein local ring and let M be a maximal
Cohen-Macaulay A-module. Then there exists exact sequences of the form
0 → M → F → M1 → 0,
with F free and M1 maximal Cohen-Macaulay.
Remark 2.13. Due to 2.11 the following assertions hold:
Let M1 , M2 and N1 , N2 are maximal Cohen-Macaulay A-modules and let F, G be
free A-modules. Suppose we have exact sequences:
0 → M1 → F → M2 → 0 and 0 → N1 → G → N2 → 0.
If there exists an A-linear map ψ1 : M1 → N1 then:
(1) there exists A-linear maps ψ2 : M2 → N2 and φ : F → G such that the
following diagram commutes:
0
/ M1
0
/ N1
ψ1
′
/F
/ M2
φ
/G
/0
ψ2
/ N2
/0
′
(2) If ψ : M2 → N2 and φ : F → G are another pair of maps such that the
above commutative diagram holds then ψ2 − ψ2′ ∈ β(M2 , N2 ).
Definition 2.14. (with hypotheses as in 2.13.) We call ψ2 to be a pre-lift of ψ1 .
The following is an easy consequence of 2.11.
Proposition 2.15. Let (A, m) be a local Gorenstein ring and let M, N be maximal
Cohen-Macaulay A-modules. Let f : M → N be A-linear and let f ∗ : N ∗ → M ∗ be
the induced map. We have:
(1) f is a split mono if and only if f ∗ is a split epi.
(2) f is a split epi if and only if f ∗ is a split mono.
(3) f is irreducible if and only if f ∗ is irreducible.
2.16. Let (A, m) be a Gorenstein local ring. Let us recall the definition of CohenMacaulay approximation from [1]. A Cohen-Macaulay approximation of an Amodule M is an exact sequence
0 −→ Y −→ X −→ M −→ 0,
where X is a maximal Cohen-Macaulay A-module and Y has finite projective dimension. Such a sequence is not unique but X is known to unique up to a free
SYMMETRIES AND CONNECTED COMPONENTS OF THE AR-QUIVER
9
summand and so is well defined in the stable category CM(A). We denote by X(M )
the maximal Cohen-Macaulay approximation of M
2.17. If M is Cohen-Macaulay then maximal Cohen-Macaulay approximation of
M are very easy to construct. We recall this construction from [1]. Let n =
codim M = dim A − dim M . Let M ∨ = ExtnA (M, A). It is well-known that M ∨ is
Cohen-Macaulay module of codim n and M ∨∨ ∼
= M . Let F be any free resolution
of M ∨ with each Fi a finitely generated free module. Note F need not be minimal
∂n
Fn−1 ). Then note Sn (F) is
free resolution of M . Set Sn (F) = image(Fn −→
maximal Cohen-Macaulay A-module. It can be easily proved that X(M ) ∼
= Sn (F)∗
in CM(A).
The following result is well-known and easy to prove.
Proposition 2.18. Let (A, m) be a Gorenstein local ring and let M be an A-module.
Then
X(Syz1 (M )) ∼
= Syz1 (X(M ))
in CM(A).
3. Lifts of irreducible maps
In this section (A, m) is a Gorenstein local ring, not necessarily an isolated singularity. Also A need not be Henselian.
The following is the main result of this section.
Theorem 3.1. (with hypotheses as above.) Let M, N be stable maximal CohenMacaulay A-modules and let f : M → N be A-linear. Let f1 : Syz1 (M ) → Syz1 (N )
be any lift of f . If f is irreducible then f1 is also an irreducible map.
We need a few preliminaries to prove Theorem 3.1.
We first prove:
Lemma 3.2. Let (A, m) be a local Gorenstein ring and let M, N be stable maximal
Cohen-Macaulay A-modules. Let f : M → N be A-linear and let δ ∈ β(M, N ). If
f is irreducibe then so is f + δ.
∼ An . As M is
Proof. Assume δ = v ◦ u where u : M → F and v : F → N and F =
stable it follows that u(M ) ⊆ mF and so δ(M ) ⊆ mN .
Claim-1: f + δ is not a split mono.
Suppose it is so. Then there exists σ : N → M with σ ◦ (f + δ) = 1M . So
σ ◦ f + σ ◦ δ = 1M . As δ(M ) ⊆ mN we get σ ◦ δ(M ) ⊆ mM . It follows that
σ ◦ δ ∈ rad EndA (M ). By 2.8(3) it follows that f is a split mono. This is a
contradiction as f is irreducible.
Claim-2: f + δ is not a split epi.
Suppose it is so. Then there exists σ : N → M with (f +δ)◦σ = 1N . So f ◦σ+δ◦σ =
1N . Notice
δ ◦ σ(N ) ⊂ δ(M ) ⊆ mN.
It follows that δ ◦ σ ∈ rad EndA (N ). By 2.8(3) it follows that f is a split epimorphism. This is a contradiction as f is irreducible.
10
TONY J. PUTHENPURAKAL
Claim 3: Suppose X is maximal Cohen-Macaulay and we have a commutative
diagram
M❇
❇❇
❇❇f +δ
g
❇❇
❇
/N
X
h
then either g is a split monomorphism or h is a split epimorphism.
Proof of Claim 3: Notice we have a commutative diagram
M●
●●
●●f
●●
(g,−u)
●●
#
/N
X ⊕F
h+v
As f is irreducible either (g, −u) is a split mono or h + v is a split epi. We assert:
Subclaim-1: If (g, −u) is a split mono then g is a split mono.
Subclaim-2: If h + v is a split epi then h is a split epi.
Notice that Subclaim 1 and 2 will finish the proof of Claim 3. Also Claims 1,2,3
implies the assertion of the Lemma. We now give:
Proof of Subclaim-1: As (g, −u) is a split mono there exits σ : X ⊕ F → M such
that σ ◦(g, −u) = 1M . Write σ = σ1 +σ2 where σ1 : X → M and σ2 : F → M . Thus
we have σ1 ◦ g − σ2 ◦ u = 1M . As M is stable u(M ) ⊆ mF . So σ2 ◦ u(M ) ⊆ mM .
Thus σ2 ◦ u ∈ rad EndA (M ). It follows from 2.8(3) that g is a split mono.
We now give:
Proof of Subclaim-2: As h + v is a split epi there exists σ : N → X ⊕ F with
(h + v) ◦ σ = 1N . Write σ = (σ1 , σ2 ) where σ! : N → X and σ2 : N → F . It follows
that h ◦ σ1 + v ◦ σ2 = 1N .
As N is stable σ2 (N ) ⊆ mF . So v ◦ σ2 (N ) ⊆ mN . Thus v ◦ σ2 ∈ rad EndA (N ).
It follows from 2.8(3) that h is a split epi.
We also need
Lemma 3.3. Let (A, m) be a Gorenstein local ring and let M be a stable maximal
Cohen-Macaulay A-module. Let F = Aµ(M) and let ǫ : F → M be a minimal map.
Set M1 = ker ǫ ∼
= Syz1 (M ). Let X be another maximal Cohen-Macaulay A-module
(not necessarily stable) and let η : G → X be a projective cover (not necessarily
minimal). Let X1 = ker η. Let α : M → X be A-linear and let α1 : M1 → X1 be
any lift of α. If α is a split mono then α1 is a split mono.
Proof. We note that M1 is also stable. Let φ : X → M be such that φ ◦ α = 1M .
Let φ1 : X1 → M1 be a lift of φ. Then note that φ1 ◦ α1 is a lift of 1M . Thus
φ1 ◦ α1 − 1 ∈ β(M1 , M1 ). The result now follows from 2.8.
The following is a dual version of Lemma 3.3 and can be proved similarly.
Lemma 3.4. Let (A, m) be a Gorenstein local ring and let N be a stable maximal
Cohen-Macaulay A-module. Let F = Aµ(N ) and let ǫ : F → N be a minimal map.
Set N1 = ker ǫ ∼
= Syz1 (N ). Let X be another maximal Cohen-Macaulay A-module
(not necessarily stable) and let η : G → X be a projective cover (not necessarily
minimal). Let X1 = ker η. Let β : X → N be A-linear and let β1 : X1 → N1 be any
lift of α. If β is a split epi then β1 is a split epi.
SYMMETRIES AND CONNECTED COMPONENTS OF THE AR-QUIVER
11
We now give:
Proof of Theorem 3.1. Set M1 = Syz1 (M ) and N1 = Syz1 (N ).
Claim-1: f1 is not a split mono.
Suppose if possible f1 is a split mono. Then there exists σ1 : N1 → M1 with
σ1 ◦ f1 = 1M1 . Let σ : N → M be a pre-lift of σ (see 2.14 for this notion). Then
σ ◦ f is a pre-lift of 1M1 . Notice 1M is a pre-lift of of 1M1 . Then by 2.13 we get
that 1M − σ ◦ f ∈ β(M, M ). As M is stable we get by 2.8 that f is a split mono.
This is a contradiction as f is irreducible.
Claim-2: f1 is not a split epi.
Suppose if possible f1 is a split epi. Then there exists σ1 : M1 → N1 with f1 ◦ σ1 =
1N1 . Let σ : N → M be a pre-lift of σ. Then f ◦ σ is a pre-lift of 1N1 . Notice 1N is
a pre-lift of of 1N1 . Then by 2.13 we get that 1N − f ◦ σ ∈ β(N, N ). As N is stable
we get by 2.8 that f is a split epi. This is a contradiction as f is irreducible.
Claim-3: If X1 is a maximal Cohen-Macaulay A-module and if there is a commutative diagram
M1 ❈
❈❈ f
❈❈ 1
u1
❈❈
❈!
X1 v1 / N1
then either u1 is a split monomorphism or v1 is a split epimorphism.
Proof of Claim-3: By 2.12 there exists an exact sequence
0 → X1 → L → X → 0,
with L1 free and X maximal Cohen-Macaulay.
Let u : M → X be a pre-lift of u1 and let v : X → N be a pre-lift of v1 . Then
notice v ◦ u is a pre-lift of f1 = v1 ◦ u1 . As f by definition is a pre-lift of f1 we get
that v ◦ u = f + δ for some δ ∈ β(M, N ).
By 3.2 we get that f + δ is irreducible. So u is a split mono or v is a split epi.
By Lemma’s 3.3 and 3.4 we get that u1 is a split mono or v1 is a split epi.
By Claims 1, 2 and 3 the result follows.
4. Indecomposable non-free summands of
maximal Cohen-Macaulay approximation of the maximal ideal
In this section (A, m) is a Henselian Gorenstein local ring. Let X(m) be a
maximal Cohen-Macaulay approximation of the maximal ideal. In this section we
are concerned with non-free indecomposable summands of X(m). Our results are:
Theorem 4.1. Let (A, m) be a Henselian Gorenstein local ring of dimension d and
let X(m) be a maximal approximation of m. Let M be an indecomposable non-free
summand of X(m). Then M is extremal, i.e.,
(1) If A is a complete intersection of codimension c then cx(M ) = cx(k) = c.
(2) If A is not a complete intersection then curv(M ) = curv(k) > 1.
We also prove:
Theorem 4.2. Let (A, m) be a Henselian Gorenstein local ring of dimension d ≥ 1
and infinite residue field k. Let e(A) ≥ 3. Assume either dim A = 2 or A is a
hypersurface ring (with no restriction on dimension) with multiplicity e(A) ≥ 3.
12
TONY J. PUTHENPURAKAL
Let M be an indecomposable Ulrich A-module. Then neither M or Syz1 (M ) is a
summand of X(m).
4.3. Motivation: Our motivation to prove the above results is the following: Assume A is a Gorenstein Henselian isolated singularity. If M is a maximal CohenMacaulay non-free indecomposable module then there exists an irreducible morphism from M → A only if M is a summand of X(m), see [23, 4.2.1]. In our
Theorems we have to show that the vertex [A] does not belong to certain components of Γ(A).
We first give:
Proof of Theorem 4.1. By Proposition 2.18 we get that X(m)⊕F = Syz1 (X(k))⊕G
for some free modules F, G. Thus it suffices to prove that if M is a direct summand
of X(k) then it is extemal. By 2.17 it suffices to prove that if M is a summand of
Syzd (k)∗ then M is extremal. We prove by induction on d that if M is a summand
of Syzn (k)∗ for some n ≥ d then M is extremal.
We first consider the case d = 0. We note that as k is indecomposable Syzn (k)
is indecomposable for all n ≥ 0. So Syzn (k)∗ is indecomposable. Therefore M =
Syzn (k)∗ . Notice Syzn (Syzn (k)∗ ) = k. It follows that M is extremal.
We now assume that d ≥ 1 and the result has been proved for Gorenstein
Henselian rings of dimension d − 1. Let x ∈ m \ m2 be a non-zero divisor on
A. Set B = A/(x) and for any A-module N set N = N/xN . We note that
B
B
∼
SyzA
d (k) = Syzd (k) ⊕ Syzd−1 (k).
It follows that
B
B
∗
∗
∗ ∼
SyzA
d (k) = Syzd (k) ⊕ Syzd−1 (k) .
A ∗
∗
If M is a summand of SyzA
d (k) then M is a summand of Syzd k . Let E be an
irreducible summand of M . Then by Krull-Schmidt it is an irreducible summand
B
∗
∗
of either SyzB
d (k) or of Syzd−1 (B) . By induction hypothesis we get that E is
extremal. It follows that M is extremal.
∗
We prove by induction on n ≥ d that if M is an irreducible summand of SyzA
n (k)
then M is extremal. We just proved the result for n = d. We now assume that
n ≥ d + 1 and the result has been proved for n − 1. There is an exact sequence
A
0 → SyzA
n (k) → F → Syzn−1 (k) → 0,
with F free.
As n ≥ d + 1 we have that SyzA
j (k) is maximal Cohen-Macaulay for all j ≥ n − 1.
So there is an exact sequence
A
∗
∗
∗
0 → SyzA
n−1 (k) → F → Syzn (k) → 0.
A
∗
Therefore SyzA
1 (M ) is a summand of Syzn−1 (k) . By induction hypothesis Syz1 (M )
is extremal. So M is extremal.
We now give:
Proof of Theorem 4.2. Case 1: We first consider the case when A is a hypersurface
of dimension d ≥ 1 and multiplicity e(A) ≥ 3.
Let M be an indecomposable Ulrich A-module. Suppose if possible M is a
summand of X(m). By Proposition 2.18 we get that X(m) ⊕ F = Syz1 (X(k)) ⊕ G
for some free modules F, G. It follows that Syz−1 (M ) is a summand of X(k). As
SYMMETRIES AND CONNECTED COMPONENTS OF THE AR-QUIVER
13
M has no free summands we get Syz1 (M ) = Syz−1 (M ). By 2.17 we get Syz1 (M )
is a summand of Syzd (k)∗ . It follows that Syz1 (M )∗ is a summand of Syzd (k). But
Syz1 (M )∗ ∼
= Syz1 (M ∗ ).
= Syz−1 (M ∗ ) ∼
Notice if M is Ulrich then M ∗ is also Ulrich. Similarly if we set N = Syz1 (M ) is a
summand of X(m) then M ∗ is a summand of Syzd (k).
By the arguments in the previous paragraph it suffices to prove that if E is an
Ulrich A-module then neither E nor Syz1 (E) is a summand of Syzd (k). This we
prove by induction on d.
We first consider the case d = 1. Then as e(A) ≥ 3 we have that m = Syz1 (k) is
indecomposable [22, Theorem A].
If E is a summand of m then m = E. Let x be E-superficial. Then as E = m
is Ulrich we get that mm = xm. So m2 = xm. So A has minimal multiplicity. It
follows that e(A) = 2. This is a contradiction.
If Syz1 (E) is a summand of m then m = Syz1 (E). Using [21, Theorem 2] we get
the h-polynomial of m
hm (z) = 2(1 + z + z 2 + · · · + z e−2 ) where e = e(A) ≥ 3.
It follows that the h-polynomial of A is
hA (z) = 1 + z(hm (z) − 1) = 2z e−1 + lower terms in z.
This is a contradiction as A is a hypersurface.
Now assume that d ≥ 2 and the result has been proved for hypersurface rings
of dimension d − 1. Let x ∈ m \ m2 be sufficiently general. Then x is A-regular
and A ⊕ E ⊕ SyzA
1 (E)-superficial. Set B = A/(x) and if V is an A-module set
V = V /xV . Then notice
B
B
SyzA
d (k) = Syzd (k) ⊕ Syzd−1 (k).
Note E is an Ulrich B-module. Let E = U1 ⊕ U2 ⊕ · · · ⊕ Us where Ui are indecomposable B-modules. Then each Ui is an Ulrich B-module. We also have
B
B
B
∼
∼
SyzA
1 (E) = Syz1 (E) = Syz1 (U1 ) ⊕ · · · ⊕ Syz1 (Us ).
By [23, 8.17], SyzB
1 (Ui ) is an indecomposable B-module for i = 1, . . . , s.
A
If E is a summand of SyzA
d (k) then E is a summand of Syzd (k). So U1 is
B
B
a summand of Syzd (k) or Syzd−1 (k). By our induction hypothesis U1 is not a
B
summand of SyzB
d−1 (k). It follows that U1 is a summand of Syzd (k). Therefore
B
B
∼
SyzB
1 (U1 ) is a summand of Syzd+1 (k) = Syzd−1 (k), a contradiction. A similar
A
argument will show that Syz1 (E) is not a summand of SyzA
d (k).
Case 2: We now consider the case when A is a Gorenstein local ring of dimension
2 but not a hypersurface.
Notice e(A) ≥ 4. Let M be an indecomposable non-free, maximal Cohen-Macaulay
A-module. If M is a summand of X(m) then SyzA
−1 (M ) is a summand of X(k).
Notice SyzA
(M
)
is
an
indecomposable
non-free
A-module.
Therefore SyzA
−1
−1 (M )
A
A
∗
is a summand of Syz2 (k) . By [22, Theorem B], Syz2 (k) is an indecomposable
A
∗
∼
maximal Cohen-Macaulay A-module. It follows that SyzA
−1 (M ) = Syz2 (k) . Notice
(4.3.1)
A
∗ ∼
Syz1 (M ∗ ) ∼
= SyzA
−1 (M ) = Syz2 (k).
14
TONY J. PUTHENPURAKAL
Let x1 , x2 be a A ⊕ M ∗ ⊕ Syz1 (M ∗ )-superficial sequence. Set C = A/(x1 , x2 ) and
if E is an A-module, set E = E/(x1 , x2 ). We note that
C
C
C
2
SyzA
2 (k) = Syz2 (k) ⊕ Syz1 (k) ⊕ Syz0 (k).
(4.3.2)
C
∗ ∼
Set N = M ∗ . Then SyzA
1 (M ) = Syz1 (N ). We now consider two cases.
Case 1: M is Ulrich.
Then notice M ∗ is also Ulrich. Then N = k a where a = µ(N ). Let n be the
maximal ideal of C. Note it is indecomposable as a C-module. By 4.3.2 and 4.3.1
we get
C
2
na = SyzC
2 (k) ⊕ Syz1 (k) ⊕ k.
By Krull-Schmidt we get that k = n. So n2 = 0. It follows that e(A) = e(C) = 2,
a contradiction.
Case 2: M = SyzA
1 (D) where D is Ulrich.
A
A
∗
Then D = SyzA
−1 (M ). It follows that Syz2 (k) is Ulrich. Therefore Syz2 (k) is
∼ l
also Ulrich. Therefore SyzA
2 (k) = k for some l. So by 4.3.2 it follows that n the
maximal ideal of C is isomorphic to k. Therefore n2 = 0 and so e(A) = e(C) = 2 a
contradiction.
5. Proofs of Theorem’s 1.2, 1.4 and 1.6
We first give
Proof of Theorem 1.2. Let M be an indecomposable periodic maximal
Cohen-Macaulay A-module. Let N, L be indecomposable non-free maximal CohenMacaulay A-modules and assume there exists irreducible maps u : N → M and
v : M → L. Let
0 → τ (M ) → EM → M → 0,
0 → M → VM → τ −1 (M ) → 0
be the AR-sequences starting and ending at M .
We first note that Syzi (u) : Syzi (N ) → Syzi (M ) and Syzi (v) : Syzi (M ) →
Syzi (L) are irreducible, see 3.1. Let p be the period of M . Note Syzp (M ) ∼
= M.
We have irreducible maps Syzip : Syzip (N ) → M for all i ≥ 1. As Syzip (N )
is indecomposable maximal Cohen-Macaulay A-module we get that Syzip (N ) are
factors of EM for all i ≥ 1. By Krull-Schmidt theorem we get that N is periodic.
A dual argument gives that L is periodic.
If there exists an irreducible map from M → A then M is factor of X(m) the
maximal Cohen-Macaulay approximation of m. So by 4.1 we get that M is extremal.
As A is not a hypersurface this is a contradiction.
Notice τ (M ) = Syz−d+2 (M ) as A is Gorenstein. If there is an irreducible map
from A → M then A is a factor of EM and so there exist’s an irreducible map
from τ (M ) → A. By previous argument we get that τ (M ) is extremal. So M is
extremal. As A is not a hypersurface this is a contradiction,
Thus P defines a union of connected components of Γ(A).
Remark 5.1. In Proposition 5.2 of [2] it is shown that if A is a self-injective Artin
algebra and M, N are non-projective A-modules such that there is an irreducible
map f : M → N then M is a periodic module if and only if N is. To the best of
my knolwedge this proof does not generalize to higher dimensions.
We now give
SYMMETRIES AND CONNECTED COMPONENTS OF THE AR-QUIVER
15
Proof of Theorem 1.4. Let M be an indecomposable maximal Cohen-Macaulay Amodule such that projdimQ M is finite. Then M is a periodic A-module with period
≤ 2. As A is not a hypersurface ring then by proof of previous Theorem there is
no irreducible map from A → M or M → A.
Let N, L be indecomposable non-free maximal Cohen-Macaulay A-modules and
assume there exists irreducible maps u : N → M and v : M → L. Let
0 → τ (M ) → EM → M → 0,
0 → M → VM → τ −1 (M ) → 0
be the AR-sequences starting and ending at M . Notice τ (M ) = SyzA
−d+2 (M ) and
−1
τ −1 (M ) = SyzA
(M ) is
d−2 (M ). It follows that projdimQ τ (M ) and projdimQ τ
finite. Therefore projdimQ EM and projdimQ VM is finite.
By [23, 5.5, 5.6], we get that N, L are direct summands of EM and VM respectively. It follows that projdimQ N and projdimQ L are finite. Thus PQ determines
a union of connected components of Γ(A).
5.2. Let us recall that a quasi-deformation A → B < − − Q of A is a flat local map
η
A → B and a deformation Q −
→ B (i.e., ker η is generated by a Q-regular sequence.
We say CI-dimension of an A-module M is finite if there is a quasi-deformation
A → B < − − Q with projdimQ M ⊗A B is finite. We say CI-dimension of M is
infinite if it is not finite.
We now give:
Proof of Theorem 1.6. Let M be an indecomposable maximal Cohen-Macaulay Amodule which is periodic and has a finite CI-dimension over A. Let A → B <
− − Q be a quasi-deformation of A with projdimQ M ⊗A B finite. As A is not a
hypersurface by proof of Theorem 1.2 there is no irreducible map from A → M or
M → A.
Let N, L be indecomposable non-free maximal Cohen-Macaulay A-modules and
assume there exists irreducible maps u : N → M and v : M → L. Let
0 → τ (M ) → EM → M → 0,
0 → M → VM → τ −1 (M ) → 0
be the AR-sequences starting and ending at M . Notice τ (M ) = SyzA
−d+2 (M ) and
A
−1
τ (M ) = Syzd−2 (M ).
Notice τ (M ) ⊗A B = SyzB
−d+2 (M ⊗A B). Thus CI-dimension of τ (M ) is finite.
As the quasi-deformations involved are the same we get that EM also has finite
CI-dimension over A. A similar argument yields that VM has finite CI dimension
over A. As N and L are summands of EM and VM respectively we get that CI
dimension of N and L are finite. It follows that PO defines a union of connected
components of Γ(A).
Let C be the union of connected components of Γ(A) consisting of periodic indecomposable maximal Cohen-Macaulay A-modules and let CO be the union of connected components of Γ(A) consisting of periodic indecomposable maximal CohenMacaulay A-modules having finite CI-dimension over A. Then as CO ⊆ C we get
that C \ CO is a union of connected components of Γ(A). Notice C \ CO consists
of precisely those periodic maximal Cohen-Macaulay A-modules which has infinite
CI-dimension over A.
16
TONY J. PUTHENPURAKAL
6. Proof of Theorem 1.8
We need to recall some preliminaries regarding support varieties. This is relatively simple in our case since A is complete with algebraically closed residue field.
6.1. Let A = Q/(u) where (Q, n) is regular local and u = u1 , . . . , uc ∈ n2 is a
regular sequence. We need the notion of cohomological operators over a complete
intersection ring; see [16] and [13]. The Eisenbud operators, [13] are constructed as
follows:
∂
∂
→ Fi → · · · be a complex of free A-modules.
→ Fi+1 −
Let F : · · · → Fi+2 −
Step 1: Choose a sequence of free Q-modules Fei and maps ∂e between them:
e
e
∂
∂
e : · · · → Fei+2 −
→ Fei → · · ·
→ Fei+1 −
F
e
so that F = A ⊗ F
Pc
Step 2: Since ∂e2 ≡ 0 modulo (u), we may write ∂e2 = j=1 uj e
tj where tej : Fei →
Fei−2 are linear maps for every i.
Step 3: Define, for j = 1, . . . , c the map tj = tj (Q, f , F) : F → F(−2) by tj =
A⊗e
tj .
6.2. The operators t1 , . . . , tc are called Eisenbud’s operator’s (associated to u) . It
can be shown that
(1) ti are uniquely determined up to homotopy.
(2) ti , tj commute up to homotopy.
6.3. Let R = A[t1 , . . . , tc ] be a polynomial ring over A with variables t1 , . . . , tc
of degree 2. Let M, N be finitely generated A-modules. By considering a free
resolution F of M we get well defined maps
for 1 ≤ j ≤ c and all n,
tj : ExtnA (M, N ) → Extn+2
R (M, N )
L
i
∗
which turn ExtA (M, N ) = i≥0 ExtA (M, N ) into a module over R. Furthermore
these structure depend on u, are natural in both module arguments and commute
with the connecting maps induced by short exact sequences.
6.4. Gulliksen, [16, 3.1], proved that Ext∗A (M, N ) is a finitely generated R-module.
We note that Ext∗ (M, k) is a finitely generated graded module over T = k[t1 , . . . , tc ].
Define V ∗ (M ) = V ar(annT (Ext∗ (M, k)) in the projective space Pc−1 . We call
V ∗ (M ) the support variety of a module M .
We need the following
Lemma 6.5. Let (Q, n) be a complete regular local ring with algebraically closed
residue field k. Let f = f1 , . . . , fc ∈ n2 be a regular sequence. Assume c ≥ 2. Set
A = Q/(f ) and let d = dim A. Let W be an irreducible non-empty sub-variety
of Pc−1 . Then there exists an indecomposable non-free maximal Cohen-Macaulay
A-module M with V ∗ (M ) = W .
Proof. By [7, 2.3], there exists an A-module E with V ∗ (E) = W . Then N =
Syzd+1 (E) is a maximal Cohen-Macaulay A-module and V ∗ (N ) = V ∗ (E) = W .
Notice N has no free summands. If N is indecomposable then we are done. Otherwise N = N1 ⊕ N2 where N1 and N2 are maximal Cohen-Macaulay A-modules
with no free-summands. Let T = k[t1 , . . . , tc ] and let Ext∗A (N, k), Ext∗A (N1 , k) and
SYMMETRIES AND CONNECTED COMPONENTS OF THE AR-QUIVER
17
Ext∗A (N2 , k) be given T -module structure as above. As Ext∗ (N, k) = Ext∗A (N1 , k)⊕
Ext∗A (N2 , k) we get that
annT Ext∗ (N, k) = annT Ext∗ (N1 , k) ∩ annT Ext∗ (N2 , k).
It follows that
W = W1 ∩ W2
where Wi = V ∗ (Ni ) for i = 1, 2.
As W is irreducible we get that W = W1 or W = W2 . Iterating this procedure we
get our result.
The following result yields Theorem 1.8 as an easy corollary.
Theorem 6.6. Let Q = k[[x1 , . . . , xn ]] be the formal power series over an algebraically closed field k. Let u = u1 , . . . , uc ∈ n2 be a regular sequence. Assume
c ≥ 2. Set A = Q/(u) and let d = dim A. Let W be a proper sub-variety of Pc−1 .
Let
CW = {M | M is indecomposable MCM A-module with V ∗ (M ) = W }.
Then CW defines an union of connected components of Γ(A). If W is irreducible
then CW is non-empty.
Proof. Let M ∈ CW . Notice M is not free. As W is a proper subset of Pc−1 we
get that dim W ≤ c − 2. So cx M = dim W + 1 ≤ c − 1. In particular M is not
extremal. So there is no irreducible map M → A. As τ (M ) = Syz−d+2 (M ) is also
not extremal there is no irreducible map τ (M ) → A. So there is no irreducible map
A → M.
Let N, L be indecomposable non-free maximal Cohen-Macaulay A-modules and
suppose there exists an irreducible map u : N → M and an irreducible map v : M →
L.
Claim: V ∗ (N ) = V ∗ (M ) = W and V ∗ (L) = V ∗ (M ) = W .
Suppose there exists a ∈ V ∗ (N ) \ V ∗ (M ). Let D be a maximal Cohen-Macaulay
A-module with V ∗ (D) = {a}. As {a} ∩ W = ∅ we get that ExtA
i (D, M ) = 0
for all i ≫ 0. As V ∗ (τ (M )) = V ∗ (M ) = W we also get ExtA
(D,
τ
(M )) = 0 for
i
i ≫ 0. Thus ExtA
(D,
E
)
=
0
for
i
≫
0.
As
N
is
a
summand
of
EM we get
M
i
∗
that ExtA
(D,
N
)
=
0
for
all
i
≫
0.
This
implies
a
∈
/
V
(N
),
a
contradiction.
i
Thus V ∗ (N ) ⊆ V ∗ (M ). We now notice that there exists an irreducible map M →
τ −1 (N ). By the previous argument we get that V ∗ (M ) ⊆ V ∗ (τ −1 (N )) = V ∗ (N ).
Thus V ∗ (N ) = V ∗ (M ) = W .
As there exist’s an irreducible map v : M → N . Then there exists an irreducible map from v ′ : N → τ −1 (M ). By the previous argument we get V ∗ (N ) =
V ∗ (τ −1 (M )). As V ∗ (τ −1 (M )) = V ∗ (M ) = W we get V ∗ (N ) = W . Thus we have
proved our Claim.
By our claim and as there are no irreducible maps from M → A and A → M we
get that CW is a union of connected components of Γ(A). If W is irreducible then
by 6.5 we get that CW is non-empty.
We now give
Proof of Theorem 1.8. By 6.6 the result follows.
18
TONY J. PUTHENPURAKAL
7. quasi AR-sequences
7.1. Setup: In this section (A, m) is a Henselian Gorenstein local ring with algebraically closed residue field k. We also assume A is an isolated singularity. For
the notion of AR -sequences see [23, Chapter 2]. In this section we introduce the
notion of quasi AR-sequences.
Definition 7.2. Let M be an indecomposable non-free maximal Cohen-Macaulay
A-module. By a quasi-AR sequence ending at M we mean an exact sequence
φ
s: 0 → K → E −
→ M such that
(1) E is a stable maximal Cohen-Macaulay A-module.
(2) φ is irreducible.
(3) If L is a stable maximal Cohen-Macaulay A-module and if there is an Alinear map σ : L → M which is not a split epi then there exist’s a map
ξ : L → E such that φ ◦ ξ − σ ∈ β(L, M ).
Remark 7.3. Unlike AR-sequences the module K need not be maximal CohenMacaulay. Also the map φ need not be surjective.
A consequence of the definition of quasi AR-sequence is the following:
Proposition 7.4. [with hypothesis as in 7.1.] Suppose σ is irreducible. Then ξ is
a split monomorphism.
Proof. By 3.2, φ ◦ ξ is irreducible. Now φ is irreducible. So in particular it is not a
split epi. It follows that ξ is a split mono.
We need the following analogue to Corollary 2.12 from [23].
Lemma 7.5. Let (A, m) be a Henselian Gorenstein local ring with algebraically
closed residue field k. Let M, L be indecomposable non-free maximal
φ
Cohen-Macaulay A-modules and let s : 0 → K → E −
→ M be a quasi AR-sequence
ending at M . Then the following two conditions are equivalent:
(i) There is an irreducible morphism from L to M .
(ii) L is isomorphic to a direct summand of E.
Proof. (i) =⇒ (ii). This follows from 7.4.
(ii) =⇒ (i). Assume the decomposition of E is given by E = L ⊕ Q. Denote
φ = (f, g) along this decomposition. We claim that f is irreducible.
Clearly f is not a split epi as φ is not a split epi. If f is a split mono then as L
and M are indecomposable we get that f is an isomorphism and so a split epi, a
contradiction.
The rest of the proof is similar to the proof of (ii) =⇒ (i) of Corollary 2.12
from [23].
We give two constructions of quasi AR-sequences. The first one comes from AR
sequences.
Proposition 7.6. Let (A, m) be a Henselian Gorenstein local ring with algebraically
closed residue field k. Let M be an indecomposable non-free maximal Cohenp
→ M → 0 be an AR-sequence
Macaulay A-modules and let l : 0 → N → EM −
ending at M . Suppose EM = E ⊕ F where F is a free A-module and E has no
free summands. Denote p = (φ, ψ) along this decomposition. Assume E 6= 0. Let
φ
K = ker φ. Then s : 0 → K → E −
→ M is a quasi AR-sequence ending at M .
SYMMETRIES AND CONNECTED COMPONENTS OF THE AR-QUIVER
19
Proof. By Corollary 2.12 from [23] we get that φ is an irreducible map. Now let
L be a stable maximal Cohen-Macaulay A-module and let f : L → M be A-linear
which is not a split epi. Then as l is an AR sequence ending at M there exist’s an
A-linear map g : L → EM with f = p ◦ g. Suppose
σ
g=
where σ : L → E and δ : L → F.
δ
So we get f = φ ◦ σ + ψ ◦ δ. Notice ψ ◦ δ ∈ β(L, M ). It follows that s is a quasi AR
sequence ending at M .
p
→ M → 0 be an AR-sequence ending at M then it
7.7. Let l : 0 → N → EM −
is not true that a lift of p; q : Syz1 (E) → Syz1 (M ) is surjective and defines a AR
sequence ending at Syz1 (M ). The great advantage of quasi AR sequences is that
it behaves well under lifting (and also pre-lifting).
Theorem 7.8. Let (A, m) be a Henselian Gorenstein local ring with algebraically
closed residue field k. Let M be an indecomposable non-free maximal Cohenφ
Macaulay A-modules and let s : 0 → K → E −
→ M be a quasi AR-sequence ending
ψ
at M . Let ψ be any lift of φ. Set K ′ = ker ψ. Then s′ : 0 → K ′ → Syz1 (E) −
→
Syz1 (M ) is a quasi AR sequence ending at Syz1 (M ). Similarly if θ is any pre-lift
θ
e → Syz−1 (E) −
of φ. Then se: 0 → K
→ Syz−1 (M ) is a quasi AR sequence ending at
Syz−1 (M ).
Proof. As E is a stable maximal Cohen-Macaulay A-module we get that Syz1 (E)
is also a stable maximal Cohen-Macaulay A-module. By Theorem 3.1 we get that
ψ is an irreducible map.
Let L be a stable maximal Cohen-Macaulay A-module and let f : L → Syz1 (M )
be an A-linear map which is not a split epi. Let g : Syz−1 (L) → M be any prelift of f . Then by 3.3 we get that g is not a split epi. It follows that there
exists ξ : Syz−1 (L) → M such that φ ◦ ξ − g = δ where δ ∈ β(Syz−1 (L), M ). Let
ξ ′ : L → Syz1 (M ) be a lift of ξ. Then notice by construction ψ ◦ ξ ′ − g is a lift of δ.
It follows that ψ ◦ ξ ′ − g ∈ β(L, Syz1 (M )). Thus s′ is a quasi AR-sequence ending
at Syz1 (M ).
The assertion regarding se can be proved similarly.
7.9. Till now we have not used the fact that k, the residue field of A is algebraically
closed. We will now use this fact. Let us recall the following definition from [23,
Chapter 5].
Let M, N be maximal
Cohen-MacaulayL
A-modules. Set (M, N ) = HomA (M, N ).
Lm
n
Decompose M = i=1 Mi and N = j=1 Nj where Mi , Nj are indecomposable
A-modules for all i, j For g ∈ (M, N ) decompose g = (gij ) where gij : Mi → Nj .
Definition 7.10. We say g ∈ (M, N )∗ if no gij is an isomorphism.
7.11. We define the following descending chain {(M, N )n }n≥1 of A-submodules of
(M, N ) as follows:
(M, N )n consists of those f ∈ (M, N ) such that there is a sequence X0 , . . . , Xn
of maximal Cohen-Macaulay A-modules with X0 = M and Xn = N and fi ∈
(Xi−1 , Xi )∗ such that f = fn ◦ fn−1 ◦ · · · ◦ f1 .
It is easy to see that (M, N )n are A-submodules of (M, N ) and that
(M, N ) ⊇ (M, N )1 ⊇ · · · ⊇ (M, N )n ⊇ (M, N )n+1 ⊇ · · · .
20
TONY J. PUTHENPURAKAL
It is not difficult to see that (M, N )1 /(M, N )2 is a k = A/m vector space. It is
finite dimensional since it is finitely generated as an A-module. Set
(M, N )1
.
irr(M, N ) = dimk
(M, N )2
Let us restate the following basic result from [23, 5.5].
Lemma 7.12. [with hypothesis as in 7.1.] Let M, N be indecomposable maximal
Cohen-Macaulay A-modules. Assume there is an AR-sequence ending at M
0 → τ (M ) → EM → M → 0.
Let n be the number of copies of N in direct summands of EM (note that n = 0 is
possible). Then the following equality holds:
irr(N, M ) = n.
We note that the assumption k is algebraically closed is used in the proof of
Lemma 7.12. The following is a basic result in our theory of quasi AR-sequences.
Theorem 7.13. [with hypothesis as in 7.1.] Let M, N be indecomposable nonφ
free maximal Cohen-Macaulay A-modules. Let 0 → K → E −
→ M be a quasi AR
sequence ending at M . Let n be the number of copies of N in direct summands of
E (note that n = 0 is possible). Then the following equality holds:
irr(N, M ) = n.
Proof. Set S(N, E) = (N, E)/(N, E)1 . Then by proof of Lemma 5.5. in [23] it
follows that S(N, E) ∼
= k n . Define
(N, M )1
,
(N, M )2
[f ] → [φ ◦ f ].
θ : S(N, E) →
By 7.5 we get that θ is a well-defined k-linear map.
We first show that θ is surjective. Let σ : N → M be an irreducible map. Denote
by [σ] it’s class in (N, M )1 /(N, M )2 . By our definition of quasi AR sequence there
exist’s ξ : N → E such that φ ◦ ξ − σ = g ∈ β(N, M ). By 7.4 we get that ξ is a
split monomorphism. As N, M are both non-free indecomposable A-modules we
get that g ∈ (N, M )2 . Therefore we get θ([ξ]) = [σ].
Next we show that θ is injective . Let [h] ∈ S(N, E) be non-zero. Thus h : N → E
is a split mono. Then by 7.5 we get that φ ◦ h : N → M is an irreducible map.
Thus θ([h]) = [φ ◦ h] 6= 0. Therefore θ is injective.
A consequence of the previous two results is the following:
Corollary 7.14. [with hypothesis as in 7.1.] Let M be indecomposable non-free
maximal Cohen-Macaulay A-module. Suppose the following is an AR-sequence ending at M :
p
→ M → 0.
t : 0 → τ (M ) → EM −
Further assume EM has no free summnads. Let
φ
s:0→K →E−
→M
∼ EM and φ is surjective. Furtherbe a quasi AR sequence ending at M . Then E =
more s is also an AR-sequence ending at M (and so K ∼
= τ (M )).
SYMMETRIES AND CONNECTED COMPONENTS OF THE AR-QUIVER
21
Proof. By 7.12 and 7.13 it follows that E ∼
= EM . As p : EM → M is an indecomosable map, by defining property of quasi AR-sequences there exists a map
ξ : EM → E such that φ ◦ ξ − p = δ ∈ β(EM , M ). As p is indecomposable, by 7.4 we
get ξ is a split mono. As E ∼
= EM , by Krull-Schmidt we get ξ is an isomorphism.
It follows that
φ − p ◦ ξ −1 = δ ◦ ξ −1 := η ∈ β(E, M ).
Set ψ = p◦ξ −1 : E → M . Notice ψ is surjective. As E, M has no free summands we
get that η(E) ⊆ mM . It follows that the maps φ, ψ : E/mE → M/mM are equal.
As ψ is surjective, it follows that φ is surjective. So by Nakayama’s Lemma we get
that φ is surjective.
We now use that t is an AR-sequence. As φ is irreducible, it is not a split epi.
Therefore there exists θ : E → EM such that p ◦ θ = φ. As φ is irreducible we get
that θ is a split mono. Since E ∼
= EM , by Krull-Schmidt we get θ is an isomorphism.
Note there exists f : K → τ (M ) which makes the following diagram commute:
s: 0
/K
t: 0
/ τ (M )
φ
/E
f
/M
/0
1M
θ
/M
p
/ EM
/0
By Snake Lemma we get that f is an isomorphism. So K ∼
= τ (M ). In the terminology of [23, 2.3] we get s ∼ t. So s is an AR-sequence ending at M .
The following consequence of Corollary 7.14 is significant:
Lemma 7.15. [with hypothesis as in 7.1.] Let M be an indecomposable maximal
p
→ M → 0 be an ARCohen-Macaulay non-free A-module. Let t : 0 → τ (M ) → EM −
sequence ending at M . If there is no irreducible maps A → M and A → Syz1 (M )
then we have
µ(EM ) = µ(M ) + µ(τ (M )).
Proof. Set N = τ (M ), M1 = Syz1 (M ) and E1 = Syz1 (E). As there is no irreducible
maps from A → M we get that EM is stable. In particular t is a quasi AR-sequence,
see 7.6. Let φ : E1 → M1 be any lift of p. Then we have a quasi AR-sequence
φ
→ M1 .
s : 0 → K1 → E1 −
As there are no irreducible maps from A → M1 we get that EM1 is stable. Therefore
by Corollary 7.14 we get that φ is surjective and s is an AR-sequence ending at
M1 . Furthermore E1 ∼
= τ (M1 ).
= EM1 and K1 ∼
Let F → EM and G → M be projective covers. Then we have an exact sequence
0
/ E1
0
/ M1
φ
/F
/ EM
θ
/G
/0
p
/M
/0
Set N = τ (M ). As φ, p are surjective we get θ is surjective. As G is free it is in
fact a split epi. Set H = ker θ. Then H is free and µ(H) = µ(EM ) − µ(M ).
As φ is surjective we get by Snake Lemma that the induced map H → N is
surjective. It follows that µ(N ) ≤ µ(EM ) − µ(M ). As there is an exact sequence
22
TONY J. PUTHENPURAKAL
0 → N → EM → M → 0 it follows (after tensoring with A/m) that µ(N ) ≥
µ(EM ) − µ(M ). The result follows.
8. Proof of Theorem 1.10
In this section we give a proof of Theorem 1.10.
8.1. [with hypotheses as in 1.1.] Recall if M is an indecomposable non-free Amodule then it’s AR-translate is τ (M ) = Syz−d+2 (M ). So if d = 2 or if A is a
hypersurface of even dimension then τ (M ) = M .
We now give:
Proof of Theorem 1.10. By 8.1 we get that τ (M ) = M . Let M be an indecomosable
Ulrich A-module. Then by 4.2 there is no irreducible map from M → A and
Syz1 (M ) → A. By our assumptions on the ring there is no irreducible map from
A → M and A → Syz1 (M ). Let s : 0 → M → EM → M → 0 be the ARsequence ending at M . Then by 7.15 we get µ(EM ) = 2µ(M ). Also note that
e(EM ) = 2e(M ). So e(EM ) = µ(EM ). Therefore EM is Ulrich A-module.
Let N be a non-free indecomposable maximal Cohen-Macaulay A-module. If
there is an irreducible morphism N → M then N is a summand of EM . As EM is
Ulrich we also get N is Ulrich. If there is an irreducible morphism from M → N
then by our assumptions on the ring there is also an irreducible morphism from
N → M . By our earlier argument we get N is Ulrich.
As there is no irreducible map from A → M or from M → A it follows that U
defines a union of connected components of Γ(A).
Remark 8.2. If A = Q/(f ) is a hypersurace ring with dim A is odd then note
that Auslander-Reiten translate τ (M ) = SyzA
1 (M ) which in general is not an Ulrich module (even if M is Ulrich), see [21, Theorem 2]. So for odd dimensions
our technique to produce an infinite family of indecomosable Ulrich modules with
unbounded multiplicities; fails.
9. Proof of Theorem 1.11
In this section we give a proof of Theorem 1.11.
9.1. We recall the definition of linkage of modules as given in [20]. Throughout
(A, m) is a Gorenstein local ring of dimension d.
φ
→ F0 → M → 0
9.2. Let us recall the definition of transpose of a module. Let F1 −
be a minimal presentation of M . Let (−)∗ = Hom(−, A). The transpose Tr(M ) is
defined by the exact sequence
φ∗
0 → M ∗ → F0∗ −→ F1∗ → Tr(M ) → 0.
Set λ(M ) = Syz1 (Tr(M )). We note that if M is a maximal Cohen-Macaulay Amodule then Tr(M ) = (Syz2 (M ))∗ .
Definition 9.3. Two A-modules M and N are said to be horizontally linked if
M∼
= λ(N ) and N ∼
= λ(M ).
If E is a stable maximal Cohen-Macaulay A-module then it is known that E is
linked to λ(E), i.e., λ2 (E) = E see [20, Corollary 7]. Note if M is an indecomposable
non-free maximal Cohen-Macaulay A-module then so is λ(M ).
SYMMETRIES AND CONNECTED COMPONENTS OF THE AR-QUIVER
23
Proof. We prove the result only for λ. The proof for D is in fact simpler.
Let M, N be indecomposable non-free maximal Cohen-Macaulay A-modules. Using terminology from 7.11 it suffices to prove that there exists an isomorphism
φ:
(λ(N ), λ(M ))1
(M, N )1
→
.
(M, N )2
(λ(N ), λ(M ))2
Let f ∈ (M, N )1 . Then as M, N are indecomposable we get that f is not an
isomorphism. In particular it is not a split mono. Let f2 : Syz2 (M ) → Syz2 (M ) be
a lift of f . By 3.3 and 2.15 it follows that f2 is not a split mono. Let f2∗ : Tr(N ) →
Tr(M ) be the dual of f2 . Then f2∗ is not a split epi. Let gf : λ(N ) → λ(M ) be any
lift of f2∗ . Define
(λ(N ), λ(M ))1
φe : (M, N )1 →
,
(λ(N ), λ(M ))2
f 7→ gf + (λ(N ), λ(M ))2 .
We first show that this map is independent of the choices we made. If f2′ is another
lift of f then f2 − f2′ ∈ β(Syz2 (M ), Syz2 (N )). So f2∗ − (f2′ )∗ ∈ β(Tr(N ), Tr(M )).
We know that if σ ∈ β(Tr(N ), Tr(M )) then any lift of σ is in β(λ(N ), λ(M )). Thus
we have gf − gf′ = δ ∈ β(λ(N ), λ(M )). As λ(N ), λ(M ) are indecomposable and
non-free we get that δ ∈ (λ(N ), λ(M ))2 . Thus φe is well-defined. It is elementary
to show that φe is A-linear.
Now let f ∈ (M, N )2 . Then there exists a maximal Cohen-Macaulay A-module
X and a commutative diagram
M❇
❇❇
❇❇f
u
❇❇
❇
/N
X
v
such that u is not a split mono and v is not a split epi. Let u2 : Syz2 (M ) → Syz2 (X)
be a lift of u and v2 : Syz2 (X) → Syz2 (N ) be a lift of v. By 3.3 and 3.4 we get
that u2 is not a split mono and v2 is not a split epi. Then f2 = v2 ◦ u2 is a lift of
f . Then f2∗ = u∗2 ◦ v2∗ . Also u∗2 is not a split epi and v2∗ is not a split mono. Let
Syz1 (u∗2 ) be a lift of u∗2 and Syz1 (v2∗ ) be a lift of v2∗ . Then gf = Syz1 (u∗2 ) ◦ Syz1 (v2∗ )
is a lift of f2∗ . By 3.3 and 3.4 we get that Syz1 (u∗2 ) is not a split epi and Syz1 (v2∗ )
is not a split mono. So gf ∈ (λ(N ), λ(M ))2 . Thus we have a well-defined A-linear
map
(λ(N ), λ(M ))1
(M, N )1
→
.
φ:
(M, N )2
(λ(N ), λ(M ))2
As λ2 (M ) = M and λ2 (N ) = N we have a well defined A-linear map
ψ:
(λ(N ), λ(M ))1
(M, N )1
→
.
(λ(N ), λ(M ))2
(M, N )2
Finally it is tautological that φ and ψ are inverses of each other. Thus λ : Γ(A) →
Γ(A)rev is an isomorphism.
Now assume that A is not a hypersurface ring. We first note that Syz1 (λ(M )) =
M ∗ when M is stable maximal Cohen-Macaulay A-module. If λ(M ) = D(M ) for all
indecomposable maximal non-free A-modules then M ∗ (and so M ) has a periodic
resilution with period 1. It follows that A is a hypersurface ring, a contradiction.
24
TONY J. PUTHENPURAKAL
Next we show that there exist’s E with D(E) 6= E. As A is not a hypersurface
there exists an MCM module M which is not periodic. Let M1 = Syz1 (M ). As M
is not periodic either M 6= M ∗ or M1 6= M1∗ .
If λ(M ) = M for all indecomposable maximal Cohen-Macaulay non-free M then
note that Syz1 (M ) = M ∗ for all such M . We now note that
Syz (M ∗ ) ∼
= Syz (Syz (M )) = Syz (M )
= (Syz (M ))∗ ∼
−2
2
1
2
3
We now note that
Syz−2 (M ∗ ) = Syz−2 (Syz1 (M )) = Syz−1 (M )
It follows that M is periodic for all indecomposable maximal Cohen-Macaulay nonfree M . Thus A is a hypersurface, a contradiction.
Thus λ, D : Γ(A) → Γ(A)rev are distinct isomorphism’s if A is not a hypersurface
ring. Furthermore λ 6= 1 and D 6= 1.
The following result is immediate:
Corollary 9.4. (with hypotheses as in 1.1). For all n ∈ Z the map Syzn : Γ(A) →
Γ(A) is an isomorphism.
Proof. We have Syz1 ◦λ = D. So Syz1 = D ◦ λ and Syz−1 = λ ◦ D. The result
follows.
We now give
Proof of Theorem 1.13. Let [M ] in C. Recall
I(M ) = {n | [Syzn (M )] ∈ C}.
We first show that I(M ) is an ideal in Z. As M = Syz0 (M ) we get 0 ∈ I(M ).
Now let n ∈ I(M ). The isomorphism Syz−n : Γ(A) → Γ(A) maps C to itself since
Syz−n (Syzn (M )) = M . In particular we have [Syz−n (M )] = Syz−n ([M ]) ∈ C. If
m, n ∈ C then note that the isomorphism Syzn : Γ(A) → Γ(A) maps C to itself
as Syzn ([M ]) = [Syzn (M )] ∈ C. Therefore [Syzn+m (M )] = Syzn ([Syzm (M )]) ∈ C.
Thus I(M ) is an ideal in Z. In particular there exist’s a unique non-negative integer
i(M ) such that I(M ) = i(M )Z.
To prove rest of the assertion of the theorem we first make a convention: if
[X], [Y ] ∈ C then write [X] < −− > [Y ] if there is an irreducible map from X to
Y OR there is an irreducible map from [Y ] to [X].
As [M ], [N ] are in C there is a sequence
[M = X0 ] < −− > [X1 ] < −− > · · · < −− > [Xn−1 ] < −− > [Xn = N ],
in C. Set a = i(M ) and b = i(N ). By 9.4 we have the following sequence in Γ(A):
[Syza (M )] < −− > [Syza (X1 )] < −− > · · · < −− > [Syza (N )].
As [Syza (M )] ∈ C we get [Syza (N )] ∈ C. So a ∈ I(N ) and therefore I(M ) ⊆ I(N ).
Similarly we get I(N ) ⊆ I(M ). Thus I(M ) = I(N ).
We now give
Proof of Corollary 1.14. (1) Suppose if possible D has only finitely many vertices.
e
Then Syzn (D) cannot be a component of Γ(A)
= Γ(A) \ Γ0 (A). As Γ0 (A) has
only finitely many components we get Syzn (D) = Syzm (D) for some n > m. Set
c = n − m. Then Syzc (D) = D. Therefore Syzlc (D) = D for all l ∈ Z.
SYMMETRIES AND CONNECTED COMPONENTS OF THE AR-QUIVER
25
We note that the function Syzlc permutes vertices of D among itself. As D is
finite it follows that all modules in D is periodic.
As D is a connected component of Γ0 (A) it follows that there exists [M ] ∈ D
such that there is an irreducible map either from M to A or an irreducible map
from A to M . In the first case M is a component of X(m) the maximal CohenMacaulay approximation of m. By 4.1 we get that M is extremal. As M is periodic
we get that A is a hypersurface ring, a contradiction. In the second case there
is an irreducible map from Syz−d+2 (M ) → A. Note as M is periodic then so is
Syz−d+2 (M ). An argument similar to the earlier case yields that A is a hypersurface
ring, a contradiction.
(2) Suppose if possible the function f : V ert(D) → Z given by f ([M ]) = e(M ) is
bounded. As e(M ) ≥ µ(M ) and e(Syz1 (M )) = e(A)µ(M ) − e(M ), it follows that
the multiplicity function on V ert(Syz1 (D)) is bounded. Iterating we get that the
multiplcity function on V ert(Syzn (D)) is bounded for each n ≥ 1. Then Syzn (D)
e
cannot be a component of Γ(A)
= Γ(A) \ Γ0 (A). As Γ0 (A) has only finitely many
components we get Syzn (D) = Syzm (D) for some n > m. Set c = n − m. Then
Syzc (D) = D. Therefore Syzlc (D) = D for all l ∈ Z. In particular there exists c
such that
(∗∗)
βil (M ) ≤ c
for all l ≥ 0 and all [M ] in D.
As D is a connected component of Γ0 (A) it follows that there exists [M ] ∈ D
such that there is an irreducible map either from M to A or an irreducible map
from A to M . In the first case M is a component of X(m) the maximal CohenMacaulay approximation of m. By 4.1 we get that M is extremal. As A is not an
hypersurface we get the following
(1) If A is a complete intersection of codimension c ≥ 2 then cx(M ) = c.
Furthermore lim βi (M ) = ∞. In particular the sequence {βil (M )} is unbounded. Thus (**) is not possible in this case.
(2) If A is Gorenstein but not a complete intersection then
curv(M ) = curv(k) > 1. So there exists r > 1 such that βi (M ) > ri for all
i ≫ 0. Thus (**) is not possible in this case too.
In the second case note that there is an irreducible map from N = Syz−d+2 (M )
to A. We then have that for all i ≥ 0
βil+d−2 (N ) ≤ c
Then an argument similar to above gives a contradiction.
10. Obstruction to quasi AR-sequences
Let the setup be as in 1.1. Let M be a non-free maximal Cohen-Macaulay
indecomosable A-module.
10.1. Let s : 0 → τ (M ) → EM → M → 0 be the AR-sequence ending at M . Then
using Proposition 7.6 we get the following:
There is no quasi-AR sequence ending at M ⇐⇒ EM is free.
The next result gives an essential obstruction to non-existence of quasi ARsequences.
Lemma 10.2. [with hypothesis as in 1.1] Further assume d 6= 1. Suppose there is
a non-free indecomposable maximal Cohen-Macaulay module M such that there is
no quasi AR-sequence ending at M . Then A is a hypersurface ring.
26
TONY J. PUTHENPURAKAL
Proof. By 10.1 it follows that τ (M ) = Syz1 (M ). By construction
τ (M ) = Syz−d+2 (M ). Therefore we get that M ∼
= Syz−d+1 (M ). As d 6= 1 we get
that M (and so τ (M )) is periodic.
EM is non-zero and free. In particular it has A as a summand. So there is an
irreducible map from τ (M ) → A. It follows that τ (M ) is a summand of X(m),
the maximal Cohen-Macaulay approximation of m. By 4.1 we get that τ (M ) is
extremal A-module. It is also periodic. So A is a hypersurface ring.
We now analyze hypersuface rings having perhaps modules M such that there
is no quasi AR-sequence ending at M . Let < M > denote the isomorphism class
of a module M . Set
Qc (A) = {< M >| [M ] ∈ Γ(A) with no quasi AR-sequence ending at M }.
We show
Proposition 10.3. Let (A, m) be a complete equicharacteristic hypersurface isolated singularity. Assume d = dim A is even and non-zero. Also assume that
k = A/m is algebraically closed. Then
(1) Qc (A) is a finite set (possibly empty).
(2) If A is not of finite representation type and Syzd (k) is indecomposable then
Qc (A) is empty.
(3) If Qc (A) is non-empty and if < M > ∈ Qc (A) then
(a) Syzn (M ) = M for all n ∈ Z.
(b) [M ] is an isolated component of Γ(A).
Proof. We note that as A is a hypersurface and d is even we get that τ (M ) = M
for any non-free maximal Cohen-Macaulay indecomosable A-module M .
(1) If < M >∈ QC (A) then there is an irreducibe map from M = τ (M ) → A.
So M is a component of X(m). It follows that Qc (A) is a finite set.
(2) As Syzd (k) is indecomposable there is a unique non-free component of X(m).
It follows that ♯Qc (A) ≤ 1. If < M >∈ Qc (A) then note that [M ] ⇆ [A] is a
connected component of Γ(A). It follows that A is of finite representation type, a
contradiction.
(3)(a) As there is no quasi AR-sequence ending at M we get that EM is free. So
τ (M ) = Syz1 (M ). As dim A is even we get M = τ (M ). As A is a hypersurface we
get Syzn (M ) = M for all n ∈ Z.
(3)(b) Notice [M ] is only connected to [A]. So we get that [M ] is an isolated
component in Γ(A).
11. Structure of Γ0 (A)
In this section we completely determine the structure of Γ0 (A) when dim A = 2
and its multiplcity e(A) ≥ 3.
Theorem 11.1. [with hypothesis as in 1.1.] Assume dim A = 2 and e(A) ≥ 3. Set
M1 = A. Then Γ0 (A) is of the form
M1 ⇆ M2 ⇆ M3 ⇆ M4 ⇆ · · · ⇆ Mn ⇆ · · ·
where e(Mn ) = ne(M1 ) for all n ≥ 1. Furthermore
(1) X(m) = M2 ⊕ F where F is free.
(2) Mn∗ = Mn for all n ≥ 1.
SYMMETRIES AND CONNECTED COMPONENTS OF THE AR-QUIVER
27
Remark 11.2. We do not have any idea of the structure of Γ0 (A) when e(A) = 2
(so necessarily A is a hypersurface) and A is of infinite representation type. The
reason is that Proposition 11.3 given below breaks down in the case e(A) = 2.
The following result is essential in our proof of Theorem 11.1.
Proposition 11.3. (with hypotheses as in Theorem 11.1). Let X(m) be a MCM
approximation of m. Write X(m) = M ⊕ F where F is free and M has no free
summands. Then
(1) M is indecomposable.
(2) rank M = 2.
(3) M ∼
= M ∗.
A
∗
Proof. (1) By [22, Theorem B]; SyzA
2 (k) is indecomposable. So X(k) = Syz2 (k)
A
is indecomposable. As X(m) = Syz1 (X(k)) ⊕ G where G is free we get that
A
∗
M = SyzA
1 (Syz2 (k) ) is indecomposable by [23, 8.17].
A
A
∗
(2) We get M = SyzA
−1 (Syz2 (k)). Let x, y be a A ⊕ M ⊕ Syz2 (k)-superficial
sequence. Set C = A/(x, y). If E is an A-module then set E = E/(x, y)E. Notice
A
A
C
C
C
2
∼
M∗ ∼
= SyzC
−1 (Syz2 (k)) and Syz2 (k) = Syz2 (k) ⊕ Syz1 (k) ⊕ Syz0 (k).
Therefore
C
C
2
M∗ ∼
= SyzC
1 (k) ⊕ Syz0 (k) ⊕ Syz−1 (k).
We note that as we have an exact sequence 0 → k = soc(C) → C → C/ soc(C) → 0.
Thus SyzC
−1 (k) = C/ soc(C). Let n be the maximal ideal of C. Thus we have
M∗ ∼
= n ⊕ k 2 ⊕ C/ soc(C).
Thus ℓ(M ∗ ) = 2ℓ(C). Therefore
e(M ) = e(M ∗ ) = e(M ∗ ) = ℓ(M ∗ ) = 2ℓ(C) = 2e(C) = 2e(A).
It follows that rank M = 2.
(3) As there is an irreducible map M → A there exists an irreducible map
A → M ∗ . As dim A = 2 we have τ (M ∗ ) = M ∗ . So there is an irreducible map
from M ∗ → A. Thus M ∗ is an irreducible component of X(m). By (1) we have
M∗ ∼
= M.
We now give
Proof of Theorem 11.1. Set X(m) = M2 ⊕ F where F is free and M2 has no free
summands. By Proposition 11.3 we get that M2 is indecomposable of rank 2. We
have the AR-sequence
0 → M2 → M1a ⊕ X → M2 → 0.
Thus a + rank X = 4. By Lemma 10.2 and Proposition 10.3(2) we get that X 6= 0.
Thus 1 ≤ a ≤ 3. We assert a = 1. We prove this by showing that the cases a = 2
or 3 do not occur.
Claim 1: a 6= 3.
Suppose if possible a = 3 then rank X is one. So X is indecomposable. As
dim A = 2 and there is an irreducible map from X to M2 , there is an irreducible
map from M2 → X. By rank considerations we get that the AR-quiver ending at
X is
0 → X → M2 → X → 0.
28
TONY J. PUTHENPURAKAL
It follows that M1 , M2 and X constitute a connected component of Γ0 (A) and so
it is equal to Γ0 (A). Therefore A has finite representation type, a contradiction.
Claim 2: a 6= 2.
If possible assume a = 2. It follows that rank X = 2. We assert:
Subclaim 3: X is indecomposable.
Suppose if possible X = X1 ⊕ X2 where rank Xi = 1. As dim A = 2 and there is
an irreducible map from Xi to M2 , there is an irreducible map from M2 → Xi . By
rank considerations we get that the AR-quiver ending at Xi for i = 1, 2 is
0 → Xi → M2 → Xi → 0.
It follows that M1 , M2 , X1 and X2 constitute a connected component of Γ0 (A)
and so it is equal to Γ0 (A). It follows that A has finite representation type, a
contradiction. Thus X is indecomposable.
The AR-sequence ending at X is
0 → X → M2 ⊕ X1 → X → 0.
By an argument similar to Subclaim-3 we get that X1 is indecomposable of rank
2. Set X0 = X.
For i ≥ 1, by an argument similar to Subclaim-3 we get that there exists indecomposable module Xi+1 of rank 2 such that the AR-sequence ending at Xi
is
0 → Xi → Xi−1 ⊕ Xi+1 → Xi → 0.
Thus Γ0 (A) consists of the modules {M1 , M2 , Xi | i ≥ 0}, Also rank Xi = 2. This
implies that A is of finite representation type (see [23, 6.2]), a contradiction.
By claims 1, 2 we get a = 1. Thus rank X = 3.
Claim 4: X is indecomposable.
Suppose if possible this is not so. Then either
Subcase 5: X = X1 ⊕ X2 ⊕ X3 where rank Xi = 1 for 1 ≤ i ≤ 3, OR
Subcase 6: X = X1 ⊕ X2 where rank Xi = i for i = 1, 2.
We show that subcase 5, 6 are not possible. If subcase 5 occurs then by rank
considerations the AR-quiver ending at Xi is
0 → Xi → M 2 → Xi → 0
for i = 1, 2, 3.
Thus the vertices of Γ0 (A) will be
{M1 , M2 , X1 , X2 , X3 }.
This implies that A has finite representation type, a contradiction.
If subcase 6 occurs then by rank considerations the AR-quiver ending at X1 is
0 → X1 → M2 → X1 → 0.
Furthermore the AR-quiver ending at X2 is
0 → X2 → M2 ⊕ X3 → X2 → 0.
Note rank X3 = 2. By an argument similar to that of subcase 5 we get that X3
is indecomposable. Iterating we obtain rank two indecomposable modules Xi for
i ≥ 4 such that the AR-quiver ending at Xi is
0 → Xi → Xi+1 ⊕ Xi−1 → Xi → 0.
It follows that the vertices of Γ0 (A) is
{M1 , M2 , Xi | i ≥ 1}.
SYMMETRIES AND CONNECTED COMPONENTS OF THE AR-QUIVER
29
As there is a bound on the ranks of vertices of Γ0 (A) it follows that A is of finite
representation type, a contradiction.
Set M3 = X. We have rank M3 = 3 and that M3 is indecomposable. Inductively
assume that we have indecomposable MCM A-modules M1 , . . . , Mn with n ≥ 3 and
rank Mi = i such that the AR-sequence ending at Mj for j ≤ n − 1 is
0 → Mj → Mj ⊕ Mj+1 → Mj → 0.
Let the AR-sequence ending at Mn be
0 → Mn → Mn−1 ⊕ Y → Mn → 0.
Clearly rank Y = n + 1. If we prove that Y is indecomposable then we can set
Mn+1 = Y and we will be done by induction.
Let Z be an indecomposable summand of Y . Then the AR-sequence ending at
Z is
0 → Z → Mn ⊕ W → Z → 0,
where W is an MCM A-module (possibly zero). Nevertheless we get that rank Z ≥
n/2.
As n ≥ 3, rank Y = n + 1 and an indecomposable summand Z of Y has rank
atleast n/2 it follows that Y has at most two indecomposable summnads.
We want to prove that Y is indecomposable. Suppose it is not so. Then by our
previous argument it has two indecomposable summands say Y1 and Y2 . Suppose
rank Y1 ≤ rank Y2 . Then we have
n+1
n
≤ rank Y1 ≤
.
2
2
We consider two cases:
Case 1: n = 2m + 1 is odd.
We get rank Y1 = m + 1. So rank Y2 = m + 1 also. Let the AR-sequence ending at
Y1 be
0 → Y1 → Mn ⊕ T → Y1 → 0.
Thus T has rank 1. The AR-sequence ending at T is
0 → T → Y1 ⊕ L → T → 0.
As m + 1 ≤ 2 we get m ≤ 1. As m ≥ 1 we get m = 1. Therefore n = 2m + 1 = 3.
Now consider the case n = 3. We get rank Yj = 2 for j = 1, 2 and rank T = 1.
Furthermore L = 0. Similarly the AR-sequence ending at Y2 will be
0 → Y2 → M3 ⊕ T ′ → Y2 → 0,
where T ′ has rank 1. The AR-sequence ending at T ′ is
0 → T ′ → Y2 → T ′ → 0.
It follows that the vertices of Γ0 (A) will be
{M1 , M2 , M3 , Y1 , Y2 , T, T ′ }.
It follows that A has finite representation type, a contradiction.
Case 2: n = 2m is even.
We get rank Y1 = m and rank Y2 = m + 1. The AR sequence ending at Y1 is
0 → Y1 → Mn → Y1 → 0.
30
TONY J. PUTHENPURAKAL
The AR sequence ending at Y2 is
0 → Y2 → Mn ⊕ T → Y2 → 0.
It follows that rank T = 2. We have to consider two sub cases:
Subcase-1: T is decomposable. In this case T = T1 ⊕ T2 where rank Ti = 1 for
i = 1, 2. The AR-sequence ending at T1 is
0 → T1 → Y2 ⊕ L → T1 → 0.
We have 2 = m + 1 + rank L. As m ≥ 1 we get m = 1 and L = 0. So n = 2. We
have already dealt with this case.
Subcase-2: T is indecomposable. The AR-sequence ending at T is
0 → T → Y2 ⊕ W → T → 0.
We have 4 = m + 1 + rank W . As m ≥ 1 the possibilities for m is 1, 2, 3. If m = 1
then n = 2. This case has been discussed earlier. Next we consider the case m = 3.
In this case W = 0. So the vertices of Γ0 (A) will be
{Mi , Y1 , Y2 , T | 1 ≤ i ≤ n}.
It follows that A has finite representation type, a contradiction.
Finally we consider the case when m = 2. So n = 4. Thus rank W = 1. The
AR-sequence ending at W is
0 → W → T → W → 0.
Thus the vertices of Γ0 (A) will be
{Mi , Y1 , Y2 , T, W | 1 ≤ i ≤ n}.
It follows that A has finite representation type, a contradiction.
(2) We note that the dual map D : Γ(A) → Γ(A)rev is an isomorphism of graphs.
As D(M2 ) = M2∗ ∼
= M2 and as Γ0 (A) is connected we get that D maps Γ0 (A) to
itself. Comparing ranks we get Mn∗ ∼
= Mn for all n ≥ 3.
12. Proof of Theorem 1.15 and Corollary 1.16
In this section of the results as stated in the title of this section. Throughout
(A, m) is an equicharacteristic Gorenstein isolated singularity of dimension two.
We also assume that A is complete and the residue field k is algebraically closed.
Furthermore we assume that e(A) ≥ 3.
We first give
Proof of Corollary 1.16. It suffices to show that Syzn (M ) ∈
/ Γ0 (A) for all M ∈
Γ0 (A) and for all n 6= 0. Using the terminology of Theorem 1.13 we need to
show I(M ) = 0 for all M in Γ0 (A). We also recall that I(M ) = I(N ) for all
M, N ∈ Γ0 (A). We denote this common value by c.
We want to show c = 0. If possible assume c > 0. Set
V = {|i − j| | Mj = Syzn Mi for some n 6= 0} and r = min V.
Notice c 6= 0 if and only if V 6= ∅.
We first consider the case when r = 0. Say Mi = Syzn Mi for some n 6= 0. We
may assume n > 0. Then Mi is periodic. As A is not a hypersurface this is a
contradiction by Theorem 1.2 and Theorem 4.1.
SYMMETRIES AND CONNECTED COMPONENTS OF THE AR-QUIVER
31
We now assume r ≥ 1. Say Mi+r = Syzn (Mi ) for some r > 0 and for some
n 6= 0. Note we are not assuming n > 0. As we have an irreducible morphism from
Mi+r−1 → Mi+r we have an irreducible map from
Syz−n (Mi+r−1 ) → Mi .
So we have Mi+1 = Syz−n (Mi+r−1 ) or Mi−1 = Syz−n (Mi+r−1 ). The first case
cannot occur as r = min V . So Mi−1 = Syz−n (Mi+r−1 ) and therefore Mi+r−1 =
Syzn (Mi−1 ). Iterating this procedure we get that M2+r = Syzn (M2 ). We have
irreducible maps from M2+r−1 and M2+r+1 to M2+r = Syzn (M2 ). So we have an
irreducible map from Syz−n (M2+r−1 ) and Syz−n (M2+r+1 ) to M2 . It follows that
atleast one of Syz−n (M2+r−1 ) and Syz−n (M2+r+1 ) is A. This is a contradiction.
Next we give
Proof of Theorem 1.15. The assertion on the structure of C follows from Theorem
11.1 and [6, 4.16.2].
(1) This follows from Theorem 11.1.
(2)(a) Let C be a connected component of Γ(A) such that [M ] ∈ V ert(C) is a
periodic module. Then by Theorem 1.2 all the modules N in V ert(C) is periodic.
We note that Syzn (C) consists of periodic modules and so [A] ∈
/ V ert(Syzn (C)
for all n ∈ Z (see Theorem 4.1). Using Theorem 7.8 and Corollary 7.14 we get
that if [M ] ∈ V ert(C) and if 0 → M → EM → M → 0 is an AR sequence
ending at M then for all n ∈ Z the AR-sequence ending at Syzn (M ) is of the form
0 → Syzn (M ) → Syzn (E) → Syzn (M ) → 0.
Now consider the structure of C as given in (1). Let period of M1 be c. We
first show that I(M1 ) = cZ (notation as in Theorem 1.13). Note c ∈ I(M1 ). If
I(M1 ) 6= cZ then there exists a with 1 < a < c such that [Syza (M1 )] ∈ V ert(C).
We note that 0 → Syza (M1 ) → Syza (M2 ) → Syza (M1 ) → 0 is the AR sequence
ending at Syza (M1 ). As M1 is the unique vertex in C which is connected to only
one other vertex we get that Syza (M1 ) = M1 . This contradicts the fact that period
of M1 is c.
We show by induction on n ≥ 2 that the period of Mn is c. We first consider
the case n = 2. As period of M1 is c we get that 0 → M1 → Syzc (M2 ) → M1 → 0
is also an AR-sequence ending at M1 . By uniqueness of AR sequences we get
M2 ∼
= Syzc (M2 ). Suppose for some a with 1 ≤ a < c we have Syza (M2 ) = M2 then
note that a ∈ I(M2 ) = I(M1 ) = cZ, a contradiction. Thus period of M2 is c.
Now assume that period of M1 , . . . , Mn is c. We prove that period of Mn+1 is also
c. As the period of Mn−1 and Mn is c we get that 0 → Mn → Mn−1 ⊕Syzc (Mn+1 ) →
Mn → 0 is another AR-sequence ending at Mn . By uniqueness of AR-sequences
we get that Mn+1 ∼
= Syzc (Mn+1 ). Suppose for some a with 1 ≤ a < c we have
Mn+1 = Syza (Mn+1 ). Then a ∈ I(Mn+1 ) = I(M1 ) = cZ, a contradiction. Thus
period of Mn+1 is c. The result follows.
(2)(b) By 1.16 there exists atmost one m0 ≥ 1 such that Syzm0 (C) = Γ0 (A). Thus
for n > m0 we have that [A] ∈
/ V ert(Syzn (C)). Set M0 = 0. We have that for all
n > m0 the sequence 0 → Syzn (Mi ) → Syzn (Mi−1 )⊕Syzn (Mi+1 ) → Syzn (Mi ) → 0
is the AR quiver ending at Mi for all i ≥ 1. By Lemma 7.15 we get that for all
n > m and for all i ≥ 1
2βn (Mi ) = βn (Mi−1 ) + βn (Mi+1 ).
32
TONY J. PUTHENPURAKAL
As M0 = 0 an easy recurssion yields that βn (Mi ) = iβn (M1 ). The result follows.
13. curvature and complexity
If (A, m) is a complete intersection of codimension c then it is known that for
any non-zero module M we have 0 ≤ cx M ≤ c. Furthermore for any integer i with
0 ≤ i ≤ c there exists an A-module M with complexity i. If A is not a complete
intersection then cx k = ∞. To deal with this situation the notion of curvature was
introduced. It can be shown that 1 < curv k < ∞ and for any non-zero module
with infinite projective dimension we have 1 ≤ curv M ≤ curv k. Furthermore if
cx M < ∞ then curv M = 1. We first prove
Proposition 13.1. Let (A, m) be an equicharactersitic complete Gorenstein isolated singularity with algebraically closed residue field k. Assume A is not a complete intersection. Then
(1) For any i ≥ 1 the modules M with complexity i form a union of connected
components of Γ(A).
(2) For any α ∈ [1, curv k) the modules M with curvature α form a union of connected components of Γ(A).
We first show
Lemma 13.2. [with hypotheses as in Proposition 13.1] Let 1 ≤ α < curv k. Let
Vα be the collection of all indecomposable modules M with curv M ≤ α. Then Vα
is a union of connected components of Γ(A). Furthermore Γ0 (A) * Vα .
Proof. Let M ∈ Vα . Note that τ (M ) = Syz−d+2 (M ) ∈ Vα . As α < curv k it follows
that there is no irreducible map from M to A or from A to M , see 4.1.
Clearly Syzn (M ) ∈ Vα for all n ∈ Z. By a similar argument as before there is
no irreducible map from Syzn (M ) to A or from A to Syzn (M ) for all n ∈ Z.
Let 0 → τ (M ) → EM → M → 0 be the AR-sequence ending at M . By 7.13 and
7.14 we get that
(1) 0 → Syzn (τ (M )) → Syzn (EM ) → Syzn (M ) → 0 is the AR-sequence ending
at Syzn (M ) for all n ≥ 0.
(2) βn (EM ) = βn (M ) + βn (τ (M )) for all n ≥ 0.
Thus we have curv(E) ≤ α. If there is an irreducible map from N to M then N is
a factor of EM and so curv(N ) ≤ curv(E) ≤ c. Thus N ∈ Vα . In a similar fashion
if there is an irreducible map from M to N then also N ∈ Vα . Thus Vα is a union
of connected components of Γ(A). Also clearly Γ0 (A) * Vα .
As an immediate consequence we get
Corollary 13.3. [with hypotheses as in Proposition 13.1] Let 1 < β < curv k. Let
Uβ be the collection of all indecomposable modules M with curv M < β. Then Uβ
is a union of connected components of Γ(A). Furthermore Γ0 (A) * Uβ .
Proof. Let 1 = α1 < α2 < · · · < αn < αn+1 < · · · be any strictly monotonically
increasing sequence converging to β. Notice
[
Uβ =
Vαn
n≥1
The result now follows from Lemma 13.2.
SYMMETRIES AND CONNECTED COMPONENTS OF THE AR-QUIVER
33
We now give
Proof of Proposition 13.1. We first prove (2). Let Cα = the collection of modules
with complexity α. Notice (with notation as in Lemma 13.2 and Corollary 13.3
(a) C1 = V1 .
(b) For 1 < α < curv(k) we have Cα = Vα − Uα .
Thus (2) follows.
(1) This is similar to (2). We have to prove results analogous to Lemma 13.2
and Corollary 13.3 first.
We now give
Proof of Theorem 1.9. Suppose A has a module M with bounded betti-numbers
but not periodic. Then note that A is not a complete intersection. We note that a
MCM A-module M will have bounded betti-numbers if and only if cx(M ) ≤ 1. By
Proposition 13.1, D the collection of all such modules defines a union of connected
components of Γ(A). We note that modules M having a periodic resolution will
form a subset C of D. By Theorem 1.2 we get that C is a union of connected
components of Γ(A). It follows that D \ C is a union of connected components of
Γ(A). If M is not periodic but has a bounded resolution then [M ] ∈ D \ C. The
result follows.
References
[1] M. Auslander and R-O. Buchweitz, The homological theory of maximal Cohen-Macaulay
approximations, Colloque en l’honneur de Pierre Samuel (Orsay, 1987). Mem. Soc. Math.
France (N.S.) No. 38 (1989), 5-37.
[2] M. Auslander and I. Reiten, Representation Theory of artin algebra V: Methods for computing
almost split sequences and irreducible morphisms, Communications in Algebra, 5(5) (1977),
519–554
[3] L. L. Avramov, V. N. Gasharov and I. V. Peeva, Complete intersection dimension, Inst.
Hautes Études Sci. Publ. Math. (1997), no. 86, 67–114 (1998).
[4] L. L. Avramov, Modules of finite virtual projective dimension, Invent. math. 96 (1989), 71101.
, Infinite free resolutions, Six lectures on commutative algebra (Bellaterra, 1996),
[5]
1118, Progr. Math., 166, Birkhuser, Basel, 1998.
[6] D. J. Benson, Representations and cohomology. I. Basic representation theory of finite groups
and associative algebras, Second edition. Cambridge Studies in Advanced Mathematics, 30.
Cambridge University Press, Cambridge, 1998
[7] P. A. Bergh, On support varieties for modules over complete intersections, Proc. Amer. Math.
Soc. 135 (2007), no. 12, 3795-3803.
[8] J. P. Brennan, J. Herzog and B. Ulrich, Maximally generated Cohen-Macaulay modules,
Math. Scand. 61 (1987), no. 2, 181-203.
[9] W. Bruns and J. Herzog, Cohen-Macaulay Rings, revised edition, Cambridge Stud. Adv.
Math., vol. 39, Cambridge University Press, Cambridge, (1998).
[10] R.-O. Buchweitz, G.-M. Greuel and F.-O. Schreyer, Cohen-Macaulay modules on hypersurface
singularities. II, Invent. Math. 88 (1987), no. 1, 165-182.
[11] A. Croll, Periodic modules over Gorenstein local rings, J. Algebra 395 (2013), 47-62.
[12] E. Dieterich, Reduction of isolated singularities, Comment. Math. Helv. 62 (1987), 654–676.
[13] David Eisenbud, Homological algebra on a complete intersection, with an application to group
representations, Trans. Amer. Math. Soc. 260 (1980), no. 1, 35–64.
[14] V. Gasharov and I. Peeva, Boundedness versus periodicity over commutative local rings,
Trans. Amer. Math. Soc. 320 (1990), no. 2, 569-580.
[15] E. L. Green and D. Zacharia, Auslander-Reiten components containing modules with bounded
Betti numbers, Trans. Amer. Math. Soc, 361 (2009), no. 8, 4195–4214.
34
TONY J. PUTHENPURAKAL
[16] T. H. Gulliksen, A change of ring theorem with applications to Poincaré series and intersection multiplicity, Math. Scand. 34 (1974), 167–183.
[17] J. Herzog, B. Ulrich and J. Backelin, Linear maximal Cohen-Macaulay modules over strict
complete intersections, J. Pure Appl. Algebra 71 (1991), no. 2-3, 187-202.
[18] S. B. Iyengar; G. J. Leuschke; A. Leykin; C. Miller; E. Miller; A. K. Singh and U. Walther,
Twenty-four hours of local cohomology. Graduate Studies in Mathematics, 87. American
Mathematical Society, Providence, RI, 2007.
[19] H. Matsumura, Commutative ring theory, second ed., Cambridge Studies in Advanced Mathematics, vol. 8, Cambridge University Press, Cambridge, 1989.
[20] A. Martsinkovsky and J. R. Strooker, Linkage of modules, J. Algebra 271 (2004), no. 2,
587-626.
[21] T. J. Puthenpurakal, The Hilbert function of a maximal Cohen-Macaulay module, Math. Z.
251 (2005), no. 3, 551-573.
[22] R. Takahashi, Direct summands of syzygy modules of the residue class field, Nagoya Math.
J. 189 (2008), 1-25.
, Cohen-Macaulay modules over Cohen-Macaulay rings, London Mathematical Soci[23]
ety Lecture Note Series, 146. Cambridge University Press, Cambridge, 1990.
Department of Mathematics, Indian Institute of Technology Bombay, Powai, Mumbai
400 076, India
E-mail address: [email protected]
| 0math.AC
|
1
On Coded Caching in the Overloaded MISO
Broadcast Channel
arXiv:1702.01672v3 [cs.IT] 24 Jun 2017
Enrico Piovano, Hamdi Joudeh and Bruno Clerckx
Department of Electrical and Electronic Engineering, Imperial College London, United Kingdom
Email: {e.piovano15, hamdi.joudeh10, b.clerckx}@imperial.ac.uk
Abstract—This work investigates the interplay of coded caching
and spatial multiplexing in an overloaded Multiple-Input-SingleOutput (MISO) Broadcast Channel (BC), i.e. a system where
the number of users is greater than the number of transmitting
antennas. On one hand, coded caching uses the aggregate global
cache memory of the users to create multicasting opportunities.
On the other hand, multiple antennas at the transmitter leverage
the available CSIT to transmit multiple streams simultaneously.
In this paper, we introduce a novel scheme which combines both
the gain derived from coded-caching and spatial multiplexing
and outperforms existing schemes in terms of delivery time and
CSIT requirement.
I. I NTRODUCTION
Caching is a promising technique proposed to improve the
throughput and reduce the latency in communication networks
[1]–[3]. In a seminal work by Maddah-Ali and Niesen [2], the
fundamental limits of cache-aided networks were explored by
considering a setting in which a single transmitter (server)
communicates with multiple users over a shared medium. The
analysis revealed that although users cannot cooperate, there
exists a (hidden) global caching gain that scales with the
aggregated memory distributed across the network, alongside
the more obvious local caching gain.
In the context of wireless networks, recent efforts have
been made to combine coded caching with the conventional
interference management techniques [4], [5] and proved to
provide performance gains in different settings [6]–[10]. In
this work, we focus on the cache-aided multiple-input-singleoutput (MISO) broadcast channel (BC), in which a transmitter equipped with multiple antennas serves multiple singleantenna users equipped with cache memories.
Overloading the MISO BC: The ability to simultaneously
serve a large number of users is a key feature envisioned for
future wireless networks, pushing them towards overloaded
regimes in which the number of users exceeds the number
of transmit antennas. In this context, the setting considered
in [2] is intrinsically an overloaded BC where all nodes are
equipped with a single antenna. In this work, we consider
a more general overloaded BC in which the transmitter is
allowed to have multiple antennas K > 1, which is yet
still smaller than the number of users Kt . A recent work on
overloaded system can be found in [11]. In this paper, we make
progress towards characterizing the optimum caching strategy
This work has been partially supported by the EPSRC of UK, under grant
EP/N015312/1.
and cache-aided performance in such setting, which remain
unknown. To simplify the analysis, we consider scenarios in
which the number of scheduled users is an integer multiple of
the number of antennas, i.e. Kt = GK, where G is a positive
integer denoted as an overloading factor.
First, we consider the above setting under the assumption of
perfect CSIT. A natural way to serve the GK users is to divide
them into G groups of K users, each served independently in
an orthogonal manner (e.g. over time). Caching is carried out
independently for each group in which only the local caching
gain is relevant. This is denoted by the Orthogonal Scheme
(OS). An alternative scheme is the one proposed by MaddahAli and Niesen (MAN) [2], in which spatial multiplexing gains
from zero-forcing are completely ignored. We show that both
strategies are in fact suboptimal for the considered setting by
proposing a scheme that outperforms both in terms of the
delivery time, i.e. the time required to deliver the requested
information during the delivery phase.
Proposed scheme (PS): We propose to partition the library
of files at the transmitter into two parts by dividing each file
into two subfiles. One part of the library is stored in the
global memory of all Kt users in the MAN manner (cached
part), and the remaining part is never cached (uncached part).
During the delivery phase, information requested from the
cached part is transmitted through a single coded multicasting
stream. This is superposed on top of a zero-forcing layer
shared in an orthogonal manner between the G groups, and
carrying information requested from the uncached part. From
its structure, it can be seen that such scheme can exploit both
the zero-forcing gains of the MISO BC, and the global cache
memory of all users. We show that under adequate partitioning
of the library, PS outperforms both OS and MAN.
Partial CSIT: We relax the assumption of perfect CSIT. We
show that the delivery time achieved by the PS with perfect
CSIT is in fact maintained under partial CSIT, up to a certain
quality. We then extend the PS to deal with any partial CSIT
level. This is implemented by caching a fraction of the library
tailored to the actual CSIT and we show, through numerical
results, the gain compared to the MAN scheme. We then
compare with the OS. However, we consider now the coded
caching strategy in [8] applied independently to each group
of K users. We show that this allows to serve the overloaded
system by achieving the same delivery time as in perfect CSIT
with a reduced quality. We prove then that the PS achieves the
same delivery time with a further reduced CSIT requirement.
2
II. S YSTEM M ODEL
A. Overloaded cache-aided MISO BC
We consider the overloaded scenario described before. The
transmitter has access to a library with Nf distinct files, denoted as W1 , W2 , . . . , WNf , each of size f bits and we assume
Nf ≥ Kt . Each user k ∈ Kt is equipped with a cache-memory
Zk of size M f bits, where M < Nf . For tractability, we
is an integer. This allows for
assume that the ratio Γ = KNt M
f
a closed-form expression of the achievable delivery time for
the MAN scheme. The communication takes place over two
phases: the placement phase and the delivery phase. During
the placement phase, before actual user demands are revealed,
the caches Zk are pre-filled with information from the Nf
files W1 , W2 , . . . , WNf . During the delivery phase, each user
k requests a single file Fk , for some Fk ∈ {1, 2, · · · , Nf }.
The requested files WF1 , . . . , WFKt are jointly mapped into
the transmitted signal x ∈ CK×1 , which satisfies the power
constraint E(xH x) ≤ P . At the t-th discrete channel use, the
k-th user’s received signal is given by
yk (t) = hH
k (t)x(t) + nk (t),
hH
k (t)
k = 1, 2, . . . , Kt
(1)
1×K
where
∈C
is the channel between the transmitter
and the k-th user and nk (t) ∼ CN (0, 1) is the Additive White
Gaussian Noise (AWGN). The duration of the delivery phase
is given by T , which is a normalized measure as explained in
the next subsection. The discrete channel uses are indexed by
t, where t ∈ [0, T ] as in [9]. In the remainder of the paper, the
index t of the channel use will be omitted for ease of notation.
At the end of the delivery phase, each user combines the signal
yk with its own cached information Zk in order to retrieve the
requested file WFk .
B. Performance measure
In this paper, the analysis is restricted to the high-SNR
regime. Such analysis gives insight into the role of coded
caching in interference management. The metric of evaluation
is the duration T needed to complete the delivery phase for
every request by the users. T is measured in time slots per
file served. As in [8], the time is normalized such that one
time slot is the amount of time required to transfer a single
file to a single receiver without caching and interference.
As a consequence, since the single-stream capacity scales as
log2 (P ) for the high-SNR regime, we can impose the size of
each file f = log2 (P ) to align our measure of performance
with [2] and make the two comparable. Finally, it is worth
noting that similar to [2], [8], T represents the worst case
scenario, i.e. the case when each user requests a different file.
For ease of notation, and without loss of generality, we assume
that user k requests file Wk for all k ∈ Kt .
which is assumed to have a covariance matrix σk2 I. The
variance σk2 is parametrized as a function of the SNR P in
the form of P −αk , where αk is the CSIT quality exponent
defined as
log(σk2 )
, k = 1, 2, . . . , Kt .
(3)
αk = lim −
P →∞
log(P )
The exponent is restricted to αk ∈ [0, 1] which captures the
entire range of CSIT (from unknown to perfect) in the highSNR regime [12], [13].
III. M ULTIPLEXING AND M ULTICASTING
In this section, we present what we consider to be the two
most obvious ways to deal with the cached-aided scenario at
hand. We further assume that CSIT is perfectly known.
A. The Orthogonal Scheme (OS)
The OS transmits K interference-free streams due to the
presence of K antennas and perfect CSIT. Each user stores
the same fraction from each file in the library, hence during
M
f bits from
the placement phase its memory is filled with N
f
M
each file. This leaves (1 − Nf )f bits to be delivered during the
delivery phase. The Kt users are divided into G groups of K
users each, where groups are orthogonalized in time and each
group is served using zero-forcing. It follows that the delivery
M
from which the total
time of each group is given by 1 − N
f
delivery time is given by
M
.
(4)
TDOS = G 1 −
Nf
M
corresponds to the local caching gain achieved
The term 1− N
f
due to the locally stored content in the memory of each user. It
is evident that the OS only exploits the local caching gain and
no coded multicasting is needed. In fact multiuser interference
is eliminated through time-sharing and zero-forcing.
B. Maddah-Ali Niesen scheme (MAN)
Alternatively, all Kt users can be served jointly using the
MAN scheme proposed in [2]. While this scheme is designed
for a system with a single transmitting antenna, it can be used
here by ignoring the multiplexing gains of the antenna array.
Following the same placement and delivery procedure in [2],
the total delivery time is given by
M
1
·
(5)
TDMAN = Kt 1 −
Nf 1 + Γ
M
The term 1 − N
corresponds to the local caching gain as in
f
1
·
the OS, while the global caching gain is captured by 1+Γ
C. Motivations for a new scheme
C. Partial Instantaneous CSIT
To study the influence of CSIT imperfections, the channel
is modelled by
hk = ĥk + h̃k
(2)
where ĥk denotes the instantaneous channel estimate available
at the transmitter for user k, and h̃k denotes the CSIT error,
We compare the performance of OS and MAN using the
OS
TD
ratio T MAN
= 1+Γ
K . The value of Γ, which is an integer,
D
belongs to the set {1, . . . , Kt }. It is evident from the ratio that
the OS performs comparatively better or equal to the MAN
scheme for Γ ∈ {1, . . . , K −1}. On the other hand, for Γ ≥ K,
MAN scheme performs better than the OS scheme. The OS
3
and MAN schemes can be seen as the two extreme cases of
utilizing the available CSIT on one side and caching on the
other side. We propose a scheme that combines elements of
OS and MAN by incorporating both spatial multiplexing and
global cache gains. This is shown to outperform both OS and
MAN in the region Γ ∈ {1, . . . , ζ}, where ζ ≥ K − 1 is some
integer. For the region Γ ∈ {ζ + 1, . . . , Kt }, the proposed
scheme matches the performance of MAN. In this paper we
focus on Γ ≤ K − 1, i.e. Γ belongs to the set {1, . . . , K − 1}.
IV. P ROPOSED S CHEME
The design of the PS involves a partition of the library
into two parts: one treated in the MAN manner, while the
other is never cached, hence delivered using zero-forcing. Such
partition is implemented by caching a fraction p ∈ [0, 1] of
each file, so that the MAN scheme consists of pNf files, while
the remaining (1 − p)Nf files correspond to the zero-forcing
part. Since we treat the cached part in a MAN manner, the
tM
ratio η = KpN
is assumed to be an integer, where η indicates
f
the number of times the cached part is repeated inside the
memories. In case of p = 1, we cache all the library and
M
, the cached part fits
η = Γ. On the other hand, for p = N
f
exactly in every single memory and in this case η = Kt .
tM
We can write p in terms of η, i.e. p = KηN
, where
f
η ∈ [Γ, Kt ] ∩ Z. Increasing η translates into a decreased
cached portion from each file and higher replication in the
memories which, in turn, is translated into multicasting messages intended for more receivers. However, a larger uncached
part needs to be delivered using zero-forcing. Selecting the
right parameter η (or equivalently the right p), is crucial to
achieve an optimum trade-off between spatial multiplexing and
coded multicasting gains as we see in what follows. Next, the
placement and delivery phase of the PS is described in detail.
B. Delivery Phase
During the delivery phase, the actual demands from users
are revealed. A given file Wk requested by the corresponding
user k consists of three parts: a) a part that is contained in
the cache Zk and hence it is not requested during the delivery
phase, b) a part that is contained in the memories of other users
Zm for m ∈ Kt \ k, which is termed as the requested cached
(c)
part and given by the subfiles Wk,τ , for all k ∈
/ τ c) a part
that has never been cached termed as the uncached part which
(p)
is denoted by Wk . The delivery phase is carried out using a
superposed transmission of the requested cached part, where
the coded multicasting information is delivered by a common
symbol decoded by all Kt users, and the uncached parts,
delivered through private symbols in a zero-forcing fashion.
Since the users share the zero-forcing layer in a orthogonal
manner, the transmission is carried out over G sub-phases
indexed by g = 1, . . . , G. Without loss of generality, we
assume that in the g-th sub-phase, the private symbols are
intended to the group Kg = {(g − 1)K + 1, . . . , gK}.
We denote the common symbols transmitted over the G
(c)
sub-phases as {xg }G
g=1 . They carry the coded messages
(c)
(c)
WS = (⊕k∈S Wk,S\{k} : S ∈ Θ)
(7)
where Θ = {S ⊂ Kt : |S| = η + 1}. These coded messages
are a combination of the requested cached part as they were
first proposed in [2] and each of them contains information for
η + 1 users. Similarly, at the g-th sub-phase, private symbols
(p)
(p)
xk carry the uncached parts Wk , where k ∈ Kg . During
the g-th delivery sub-phase, the transmitted signal is
X √
√
(p) (p)
(8)
P β vk xk +O(1)
xg = P v(c) x(c)
g +
|
{z
} k∈Kg
MAN
{z
}
|
ZF
(p)
A. Placement Phase
We first describe the placement and delivery for a given
tM
factor η, or equivalently p = KηN
. Then η is optimized to
f
minimize the delivery time. We divide each of the Nf files
W1 , W2 , . . . , WNf into two subfiles:
Wn = (Wn(c) , Wn(p) )
(6)
(c)
Wn ,
where
of size pf , is cached into the memories while
(p)
Wn , of size (1 − p)f , is never cached. We apply the idea in
(c)
(c)
(c)
[2] to the subfiles W1 , W2 , . . . , WNf by jointly encoding
all Kt users together considering their total global cache
(c)
memory. Consequently,
according to [2], each subfile Wn
(c)
Kt
is split into η smaller subfiles Wn,τ , for all τ ∈ Ω, where
(c)
Ω = {τ ⊂ Kt : |τ | = η}. It follows that each subfile Wn,τ has
. From each file n ∈ {1, 2, . . . , Nf }, user k caches
size pf
(Kηt )
(c)
the subfiles Wn,τ such that k ∈ τ . Hence each user caches Mf
Nf
(c)
bits from any Wn . Since we assume different demands across
users, the total number of files to be delivered from the cached
(c)
M
tM
part of the library is given by Qη = Kt ( KηN
−N
) 1 ·
f
f 1+η
Additionally, the number of files to be delivered
through zero
(p)
Kt M
forcing is given by Qη = Kt 1 − ηNf ·
where v(c) ∈ CM×1 and vk ∈ CM×1 are unitary precoding
vectors, and β ∈ [0, 1] is the power partitioning factor [11].
Note that the term O(1) guarantees that the power constraint is
not violated. Such term has no influence on high-SNR analysis.
(c)
Since xg has to be decoded by all Kt users, v(c) is chosen as
a generic precoding vector. On the other hand, private symbols
are precoded using the well-known zero-forcing precoders.
The received signal for each user k ∈ Kt is given by
X√
√
(p) (p)
(c) (c)
y k = P hH
+ O(1).
P β hH
k v xg +
k vi xi
i∈Kg
(c)
All users in Kt decode xg by treating the interference
(c)
from all the other symbols as noise. Hence, xg can carry
(1 − β) log2 (P ) bits per time slot. Users in Kg proceed to
(c)
remove the contribution of xg to the received signal and
(p)
(p)
decode their private symbols xk . The symbol xk is decoded
β
with a SINR of O(P ), hence carrying β log2 (P ) bits per
time slot. It follows that the cached part of the library is
delivered with a rate of (1 − β) log2 (P ) bits per time slot and
the uncached part of the library with a rate of Kβ log2 (P ) bits
(c)
per time slot. From {xg }G
g=1 each user k ∈ Kt retrieves the
(c)
messages WS and consequently its cached part. In fact any
4
(c)
missing sub-files Wk,τ , so that k ∈
/ τ , can be recovered by
(c)
(c)
combining WS with the pre-stored sub-files Wi,S\{i} , where
(p)
S = τ ∪ {k} and i ∈ S \ {k}. On the other hand, from xk
(p)
each user k ∈ Kg retrieves the uncached part Wk .
C. Delivery time
During the communication both the cached and uncached
∗
parts must be delivered. We indicate as TD,η
the delivery
time of such communication. The power partitioning factor,
indicated as β ∗ , is chosen to let both parts to be delivered
(p)
over this delivery time. Hence, the two equalities Qη f =
(c)
∗
∗
Kβ ∗ f TD,η
and Qη f = (1 − β ∗ )f TD,η
must be simultaneously satisfied. It can be verified that the solution is given by
Q(p)
(c)
η
K + Qη and
(p)
(c)
Qη and Qη , we obtain
∗
TD,η
=
∗
TD,η
β∗ =
Q(p)
η /K
(p)
(c)
Qη /K+Qη
. By expanding
M Kt
M Kt
M
Kt
=G 1−
−
+
.
ηNf
η + 1 ηNf
Nf
(9)
The value of η ∈ [Γ, Kt ] ∩ Z must be chosen to minimize the
delivery time in (9). This is given by
TD∗ =
min
η∈[Γ,Kt ]∩Z
∗
TD,η
.
(10)
where TD∗ denotes the minimum delivery time. The following
result holds for which proof can be found in the appendix.
Proposition 1. In case of perfect CSIT, the PS achieves a
delivery time TD∗ given by
TD∗ =
min
η∈{⌊x⌋,⌈x⌉}
∗
TD,η
x=
∗
which scales as O(P β ), where β ∗ =
(p)
Qη∗ /K
(p)
(c)
Qη∗ /K+Qη∗
. If the
channel estimation error decays as O(P −α ), where α ≥ β ∗ ,
each user can still decode its own private symbol as interference from other symbols is at (or below) the noise floor.
[14]. Hence the SINR still scales as in case of perfect CSIT. It
follows that, the delivery time TD∗ is mantained for any CSIT
quality no less than β ∗ . Since this represents a CSIT quality
threshold, by indicating it as α∗ , we obtain
t
G 1 − MK
η ∗ Nf
.
(13)
α∗ = β ∗ =
MKt
Kt
M
t
G 1 − η∗ Nf + η∗ +1 MK
η ∗ Nf − Nf
B. General Case
The case α < α∗ is treated by caching a larger fraction
of the library which reduces the load carried by the private
symbols. We proceed by considering a given η which will be
then optimized based on the actual CSIT. For the delivery of
the uncached part, zero-forcing is replaced by rate-splitting,
which is well suited to operate under partial CSIT [14]. The
placement phase and delivery phase follow the same steps
as in perfect CSIT. In rate-splitting, at the g-th delivery sub(p)
phase, the uncached parts Wk with k ∈ Kg are divided
(p)
and partially mapped to the private symbols xk , while
the remaining parts are jointly encoded into the symbol x̃g ,
decoded by all users in Kg . The transmitted signal is
X √
√
√
(p) (p)
β (c) x̃ +
P a vk xk +O(1)
xg = P v(c) x(c)
g
g + P v
{z
}
|
k∈Kg
MAN
{z
}
|
RS
(11)
where
G(K − 1) +
A. CSIT Relaxation
In Section IV-B, perfect CSIT was exploited to deliver K
interference-free private symbols, transmitted with a power
p
G2 (K − 1)2 + G(G + 1)(K − 1)
.
G+1
(12)
The value of η that minimizes the delivery time in (11) is
∗
denoted by η ∗ , i.e. TD∗ = TD,η
∗ . By comparing the result in
Preposition 1 with the delivery time in (4) and (5), it is verified
that TD∗ < TDOS ≤ TDMAN . Hence, simultaneously exploiting
the global cache memory of all users and the multiplexing
gain, the PS strictly outperforms both OS and MAN.
V. PARTIAL CSIT
In this section, we relax the constraint of perfect CSIT
and consider a setting with partial CSIT. The scheme in turn
is generalized to account for such imperfection. We start by
showing that the delivery time in (11) can be maintained, while
relaxing the CSIT quality α. Then we modify the PS to be
adaptable for any α, by properly tailoring the cached part of
the library based on the available CSIT.
where symbols, precoding vectors and powers are defined as in
the previous section, with the difference that the power of the
private symbols scales as O(P a ), with a ≤ β. Furthermore,
to force interference between private symbols to the noise
floor level, we need a ≤ α. Hence, we take a = min(α, β).
(c)
All users in Kt decode xg by treating the other symbols as
noise, hence it can carry (1 − β) log2 (P ) bits per time slot.
(c)
Then, all users in Kg proceed to cancel xg and decode x̃g ,
which is then canceled before decoding their respective private
symbols. It can be easily verified that rate-splitting achieves a
total rate of (β+(K −1) min(α, β)) log 2 (P ) bits per time slot.
Further details on the rate-splitting procedure can be found in
(c)
[11], [14]. As before, from {xg }G
g=1 each user can retrieve
(p)
its desired cached part, while from x̃g and xk users k ∈ Kg
retrieve their uncached parts.
C. Delivery Time
As in Section IV-C, the power partitioning factor β ∗ is cho(p)
sen such that Qη f = (β ∗ + (K − 1) min(α, β ∗ ))f TD,η (α)
(c)
and Qη f = (1 − β ∗ )f TD,η (α), where TD,η (α) denotes the
delivery time. The delivery time is given by
( ∗
TD,η , α ≥ α∗η
(c)
(14)
TD,η (α) = Q(p)
η +Qη
∗
1+(K−1)α , α < αη
5
(c)
TD (α) =
min
η∈[Γ,Kt ]∩Z
TD,η (α).
8
7
MAN
PS
6
TD
5
4
3
2
1
0
0.2
0.4
0.6
0.8
1
α
Fig. 1. TD as function of the CSIT quality α for K = 8 and G = 2
(15)
The following result holds for which the proof is given in the
appendix.
0.8
0.8
α th
Proposition 2. Given a CSIT quality α, we can achieve the
following delivery time:
(
TD∗ , α ≥ α∗
TD (α) =
1+(K−1)α∗
η+1
∗
∗
∗
min{TD,η
, TD,η+1
1+(K−1)α }, α < α
(16)
where
η = arg ′ max∗ {η ′ : α∗η′ ≤ α}.
(17)
1
1
OS
PS
0.6
OS, K=4
PS, K=4
OS, K=8
PS, K=8
0.6
α th
Q(p)
Q(p) /K
η
∗
where α∗η = (p) η
(c) and TD,η =
K + Qη . It can be
Qη /K+Qη
seen that given a certain partition η, the same delivery time
as in perfect CSIT can be maintained as long as α ≥ α∗η . On
the other hand, in case of α < α∗η , the delivery time is given
by the ratio between the total number of transmitted files and
the total communication rate. In case of η = η ∗ , we reduce to
∗
∗
α∗η∗ = α∗ and TD,η
∗ = TD , in agreement with Section V-A.
For a given quality α, the minimum achievable delivery time,
denoted by TD (α), is given as
0.4
0.4
0.2
0.2
0
1
2
3
4
G
OS
a) αPS
th and αth with respect to G
0
0.5
1
1.5
M
2
2.5
3
OS
b) αPS
th and αth with respect to M
OS
Fig. 2. CSIT qualities for the PS and OS to achieve TD
η ∈[Γ,η )∩Z
∗
For α ≥ α we can achieve the same delivery time as perfect
CSIT given in Proposition 1. Examining the delivery time from
the expression in (16) for the entire range of α is not an easy
task as we were not able to derive a more tractable expression.
However, it can be seen that for α = 0, we obtain η = Γ
in (17) and the scheme reduces to MAN. The influence of
increasing α is examined through simulations in the following
section where it is shown that the PS leverages the available
CSIT and outperforms the MAN scheme.
Looking at the PS from a different perspective, we investigate the minimum CSIT quality required to achieve a given
delivery time T . This is given through the following result for
which the proof is given in the appendix.
Proposition 3. A delivery time T ≥ TD∗ can be achieved by
a CSIT quality α equal to
∗
TD,η
1 + (K − 1)α∗η − T
(18)
α=
(K − 1)T
where
η = arg
min
η ′ ∈[Γ,η ∗ ]∩Z
∗
{η ′ : TD,η
′ ≤ T }.
(19)
Since the MAN scheme does not require CSIT, it is suitable
to compare to the OS in this context. We further assume that
each orthogonal group of K users is served using the strategy
in [8] to account for imperfect CSIT. It follows that the delivery time TDOS can be achieved while applying coded caching
N −
Nf
−M
f
K
and relaxing the CSIT quality to αOS
th = M +N − Nf −M . In
f
K
K
comparision, by applying Proposition 3, we can characterize
the minimum CSIT quality needed by the PS to achieve TDOS,
OS
∗
denoted as αPS
th . We observe that TD = TD,K−1 . Hence the
PS
value of αth in (18) is given by
αPS
th =
Nf −
M
K
Nf
K
+ Nf −
− MG
Nf
K
−M
.
(20)
PS
PS
By comparing αOS
th and αth , we can see that αth is strictly
lower for the overloaded case (G ≥ 2).
VI. N UMERICAL RESULTS
In this section, we numerically evaluate the PS. First, we
compare the delivery time TD (α) achieved under different
CSIT qualities to the one achieved using the MAN scheme, i.e.
TDMAN . This is given in Fig. 1 for a setup with K = 8 antennas,
overloading factor G = 2 and memories of size M f bits, with
M = 1. We consider a library of Nf = Kt throughout the
simulations. We observe that the PS leverages the available
CSIT quality to reduce the delivery time compared to MAN.
In particular, for α = 1, the delivery time is four times lower.
Next, we compare the PS with OS in terms of the minimum
CSIT quality needed to achieve TDOS in (4). In Fig. 2a we plot
OS
αPS
th and αth as a function of G. We consider a setup with
K = 4, 8 and M = 2. As we can see, the OS suffers when
G increases, while the PS, in addition to outperforming OS,
benefits from a larger G. Finally, in Fig. 2b we plot αPS
th and
αOS
as
a
function
of
M
with
K
=
4
and
G
=
2.
We
can
see
th
OS
for
increasing
M
.
decreases
much
faster
than
α
that αPS
th
th
VII. C ONCLUSION
In this paper, we studied the complementary gain of coded
caching and spatial multiplexing gain for an overloaded
MISO BC. We investigated how to simultaneously combine
the gain offered from the aggregate global cache memory,
with the gain offered by the presence of available CSIT.
We introduced a novel scheme which superimposes MaddahAli Niesen scheme, to exploit the coded caching gain, to a
multiuser transmission scheme, to leverage the available CSIT.
Analytical and numerical results showed a significant gain
compared to the existing schemes, both in terms of delivery
time and CSIT quality requirement.
6
A PPENDIX
C. Proof of Proposition 3
A. Proof of Proposition 1
In order to prove Proposition 1, we need to evaluate the
∗
minimum of the function TD,η
in (9) with respect to η over
the interval [Γ, Kt ] ∩ Z, and we remember that Kt = GK.
The minimum value is indicated as TD∗ in (10). We start by a
relaxation, where η is allowed to take any value in [Γ, Kt ].
∗
We can notice that TD,η
is convex over such interval and
∗
∗
∗
TD,K−1 = TD,GK , hence TD,η
has a minimum in [K − 1, Kt ].
∗
The first derivative of TD,η with respect to η is equal to zero
for η = x, given in (12). By restricting η to be an integer, i.e.
∗
η ∈ [Γ, Kt ] ∩ Z, we have that TD,η
is minimum in either ⌊x⌋
or ⌈x⌉, from which the expression in (11) follows. It can be
∗
∗
easily verified that TD,K−1
= TD,GK
= TDOS , where TDOS is
∗
OS
MAN
given in (4). Hence, TD < TD ≤ TD .
B. Proof of Proposition 2
In order to prove Proposition 2, we need to find the value
of TD (α) in (15). We have
TD (α) =
min
η ′ ∈[Γ,Kt ]∩Z
TD,η′ (α)
(21)
where TD,η′ (α) is defined as in (14). As explained in
Section V-A, in case α ≥ α∗ , it is possible to achieve the
same delivery time as in the case of perfect CSIT. Hence,
TD (α) = TD (1) = TD∗ . We now focus on the case α < α∗ .
We start by considering α∗η′ in Section V-C
(p)
Qη′ /K
α∗η′ =
(p)
Qη′ /K
+
(c)
Qη ′
=
1
(c)
(p)
1 + KQη′ /Qη′
.
(22)
(c)
It can be shown that Qη′ is a decreasing function of η ′ while
(p)
Qη′ is an increasing function of η ′ . It follows that α∗η′ is
also an increasing function of η ′ . Then, by denoting the total
number of files to be transmitted as Qη′ , we have
(c)
(p)
Qη ′ = Qη ′ + Qη ′ = Kt −
M 1
(K 2 + Kt )
Nf η ′ + 1 t
(23)
which is also an increasing function of η ′ . Moreover, from
(14), the following relationship is verified
∗
∗
Qη′ = TD,η
′ (1 + (K − 1)αη ′ ).
(24)
Given that we are considering α < α∗ and we know that
α∗η∗ = α∗ and α∗Γ = 0, there must exist an η such that
η = arg maxη′ ∈[Γ,η∗ )∩Z {η ′ : α∗η′ ≤ α}. Moreover, from the
∗
proof of Proposition 1, we know that TD,η
′ is a decreasing
′
∗
function with respect to η in [Γ, η ) ∩ Z. From (14), consid∗
∗
ering η ′ ≤ η, we have TD,η′ (α) = TD,η
′ ≥ TD,η . In case of
′
η ≥ η + 1, we have
(p)
(c)
Qη ′ + Qη ′
(p)
(c)
Qη+1 + Qη+1
= TD,η+1 (α).
1 + (K − 1)α
1 + (K − 1)α
(25)
∗
Hence, TD (α) is the minimum between TD,η
and TD,η+1 (α),
from which Proposition 2 follows.
TD,η′ (α) =
≥
∗
From the proof of Proposition 1, we know that TD,η
′ is
′
∗
a decreasing function with respect to η in [Γ, η ] ∩ Z and
increasing in [η ∗ , Kt ] ∩ Z. Hence, there must exist
η = arg
min
∗
{η ′ : TD,η
′ ≤ T}
(26)
∗
{η ′ : TD,η
′ ≤ T }.
(27)
η ′ ∈[Γ,η ∗ ]∩Z
and
ζ = arg
max
η ′ ∈[η ∗ ,Kt ]∩Z
∗
In case of η ′ < η or η ′ > ζ, we have TD,η
′ > T , thus
the delivery time T is not achievable. On the other hand, by
considering η ′ ∈ [η, ζ] and applying (14), the delivery time T
is achieved by a CSIT quality given by
(p)
α
η′
=
(c)
Qη ′ + Qη ′ − T
(K − 1)T
.
(28)
(p)
(c)
As expressed in the proof of Proposition 2, Qη′ + Qη′
increases with η ′ . It follows that η = arg minη′ ∈[η,ζ]∩Z αη′ .
(p)
(c)
∗
From Qη + Qη = TD,η
(1 + (K − 1)α∗η ), we obtain (18).
R EFERENCES
[1] N. Golrezaei, K. Shanmugam, A. G. Dimakis, A. F. Molisch, and
G. Caire, “Femtocaching: Wireless video content delivery through
distributed caching helpers,” in Proc. IEEE INFOCOM, pp. 1107–1115,
Mar. 2012.
[2] M. A. Maddah-Ali and U. Niesen, “Fundamental limits of caching,”
IEEE Trans. Inf. Theory, vol. 60, pp. 2856–2867, May 2014.
[3] M. A. Maddah-Ali and U. Niesen, “Decentralized coded caching attains
order-optimal memory-rate tradeoff,” IEEE/ACM Trans. Net., vol. 23,
pp. 1029–1040, Aug. 2015.
[4] D. Gesbert, S. Hanly, H. Huang, S. S. Shitz, O. Simeone, and W. Yu,
“Multi-cell MIMO cooperative networks: A new look at interference,”
IEEE J. Sel. Areas Commun., vol. 28, pp. 1380–1408, Dec. 2010.
[5] S. A. Jafar, “Interference alignment a new look at signal dimensions in
a communication network,” Found. Trends Commun. Inf. Theory, vol. 7,
no. 1, pp. 1–134, 2011.
[6] M. A. Maddah-Ali and U. Niesen, “Cache-aided interference channels,”
in Proc. IEEE ISIT, pp. 809–813, Jun. 2015.
[7] N. Naderializadeh, M. A. Maddah-Ali, and A. S. Avestimehr, “Fundamental limits of cache-aided interference management,” in Proc. IEEE
ISIT, pp. 2044–2048, Jul. 2016.
[8] J. Zhang, F. Engelmann, and P. Elia, “Coded caching for reducing CSITfeedback in wireless communications,” in Proc. Allerton, pp. 1099–1105,
Sep. 2015.
[9] J. Zhang and P. Elia, “Fundamental limits of cache-aided wireless
BC: Interplay of coded-caching and CSIT feedback,” arXiv:1511.03961,
2015.
[10] S. Yang, K. H. Ngo, and M. Kobayashi, “Content delivery with coded
caching and massive MIMO in 5G,” in Proc. ISTC, pp. 370–374, Sep.
2016.
[11] E. Piovano, H. Joudeh, and B. Clerckx, “Overloaded multiuser MISO
transmission with imperfect CSIT,” arXiv:1612.00628, 2016.
[12] N. Jindal, “MIMO broadcast channels with finite-rate feedback,” IEEE
Trans. Inf. Theory, vol. 52, pp. 5045–5060, Nov. 2006.
[13] S. Yang, M. Kobayashi, D. Gesbert, and X. Yi, “Degrees of freedom
of time correlated MISO broadcast channel with delayed CSIT,” IEEE
Trans. Inf. Theory, vol. 59, pp. 315–328, Aug. 2013.
[14] H. Joudeh and B. Clerckx, “Sum-rate maximization for linearly precoded
downlink multiuser MISO systems with partial CSIT: A rate-splitting
approach,” IEEE Trans. Commun., vol. 64, pp. 4847–4861, Aug. 2016.
| 7cs.IT
|
Characterizations of the Logistic and Related Distributions
Chin-Yuan Hu and Gwo Dong Lin
arXiv:1803.06620v1 [math.ST] 18 Mar 2018
National Changhua University of Education and Academia Sinica
Abstract. It is known that few characterization results of the logistic distribution were
available before, although it is similar in shape to the normal one whose characteristic properties have been well investigated. Fortunately, in the last decade, several authors have made
great progress in this topic. Some interesting characterization results of the logistic distribution have been developed recently. In this paper, we further provide some new results by
the distributional equalities in terms of order statistics of the underlying distribution and
the random exponential shifts. The characterization of the closely related Pareto type II
distribution is also investigated.
AMS subject classifications: Primary 62E10, 62G30, 60E10.
Key words and phrases: Characterization, order statistics, stochastic order, logistic distribution, exponential distribution, Pareto type II distribution.
Short title: The logistic and related distributions
Postal addresses: Chin-Yuan Hu, Department of Business Education, National Changhua
University of Education, Changhua 50058, Taiwan. (E-mail: [email protected])
Gwo Dong Lin, Institute of Statistical Science, Academia Sinica, Taipei 11529, Taiwan.
(E-mail: [email protected])
1
1. Introduction
The logistic distribution is similar to a normal distribution in shape (Mudholkar and
George 1978) and has an explicit closed form, so it has some advantages in practical applications. As remarked by Kotz (1974), few characterizations of the logistic distribution were
available before, but recently, some interesting results have been developed. In this paper,
we will further provide some more new results by properties of order statistics.
We first introduce some notations. Let X obey the distribution F , denoted by X ∼ F.
Let {Xj }nj=1 be a random sample of size n from distribution F and denote the corresponding
order statistics by X1,n ≤ X2,n ≤ · · · ≤ Xn,n . The distribution function of Xk,n is denoted
by Fk,n . It is known that Fk,n is the composition of Bk,n−k+1 and F (see, e.g., Hwang and
Lin 1984), where Bα,β is the beta distribution with parameters α, β > 0, namely,
Z F (x)
n
Fk,n (x) = Bk,n−k+1(F (x)) = k
tk−1 (1 − t)n−k dt, x ∈ R ≡ (−∞, ∞), (1)
k
0
Z
Γ(α + β) u α−1
Bα,β (u) =
t (1 − t)β−1 dt, u ∈ [0, 1].
(2)
Γ(α)Γ(β) 0
On the other hand, for Y ∼ G, we say that X is less than or equal to Y in the usual
stochastic order, denoted by X ≤st Y, if F ≤ G, where F (x) = 1 − F (x) = Pr(X > x).
Let us start with an interesting simple example. Clearly, for general distribution F , we
2
have X1,2 ≤st X because F1,2 = 1−F ≥ F by (1). One possibility to adjust this ”inequality”
is to choose a nonnegative random variable Z, independent of X and Xj ’s, such that
d
X = X1,2 + Z
(3)
or
d
X − Z = X1,2 ,
(4)
d
where = means equality in distribution. One might think that the solutions of the distributional equations (3) and (4) are the same, but this is not true in general, because
the characteristic function of Z is not equal to the reciprocal of that of −Z, namely,
E[eitZ ] 6= (E[e−itZ ])−1 , t ∈ R, in general. For example, if Z has the standard exponential
distribution E, then the solution of (3) is a logistic distribution F (x) = 1/[1+e−(x−µ) ], x ∈ R,
2
where µ ∈ R is a constant, while the solution of (4) is a negative (or reversed) exponential
distribution F (x) = e(x−µ)/2 , x ≤ µ, where µ ∈ R is a constant (see, e.g., George and Mudholkar 1982, Lin and Hu 2008, and Ahsanullah et al. 2011, and note that the smoothness
conditions on F therein are redundant due to Lemmas 1–3 below).
Throughout the paper, let U and ξ obey the uniform distribution U on [0, 1] and the
standard exponential distribution E, respectively. Moreover, let {Uj }nj=1 and {Uj′ }nj=1 be two
random samples of size n from U, and let {ξj }nj=1 and {ξj′ }nj=1 be two random samples of size
n from E. All the above random variables X, U, ξ, Xj , Uj , Uj′ , ξj and ξj′ , j = 1, 2, . . . , n, are
assumed to be independent from now on.
Mimicking the above characterization approaches (3) and (4), several authors have considered the general stochastic inequality Xk,n ≤st Xk+1,n and solved the distributional
d
d
equations (a) Xk+1,n = Xk,n + aξ and (b) Xk,n = Xk+1,n − bξ, or, more generally, (c)
d
Xk,n + aξ1 = Xk+1,n − bξ2 , where a and b are nonnegative constants. In particular, the
equality
Xk,n +
1
1
d
ξ1 = Xk+1,n − ξ2
n−k
k
also characterizes the logistic distribution. (See AlZaid and Ahsanulla 2003, Wesolowski and
Ahsanulla 2004, and Ahsanulla et al. 2010, 2012 for equations arising from Xk,n ≤st Xk+r,n
with 1 ≤ r ≤ n − k.) Besides, the distributional equations arising from (a) X ≤st Xn,n ,
(b) X1,n ≤st Xn,n and (c) Xm,m ≤st Xn,n , where m < n, were investigated by Zykov and
Nevzorov (2011), Ananjevskii and Nevzorov (2016) as well as Berred and Nevzorov (2013),
respectively.
In this paper we will solve the distributional equations arising from stochastic inequalities:
(i) Xk,n ≤st Xk,n−1 , (ii) Xk,n−1 ≤st Xk+1,n , (iii) Xk,n ≤st Xk,k , (iv) X1,k ≤st Xn−k+1,n , (v)
Xk,n ≤st Xk,n−m, and (vi) Xm−k,n−k ≤st Xm,n .
To do this, some useful lemmas are given in the next section. The main characterization
results are stated and proved in Sections 3 and 4. For simplicity, we first deal with the closely
related Pareto type II distribution in Section 3, and then the logistic distribution in Section
4. Finally, we pose an open problem in Section 5.
2. Lemmas
3
We need some lemmas in the sequel. Lemma 1(i) was given without proof in Lukacs (1970,
p. 38), but has been ignored in the literature. We provide a proof here for completeness.
Lemma 1. (i) Let Y and Z be two independent random variables. If Y has an absolutely
continuous distribution, then so does Y + Z, regardless of the distribution of Z.
(ii) If, in addition to the assumptions in (i), both Y and Z are positive random variables,
then the product Y Z has an absolutely continuous distribution.
Proof. Let F , G and H be the distributions of Y + Z, Y and Z, respectively. Then
Z ∞
F (x) =
G(x − z)dH(z), x ∈ R.
−∞
Since G is absolutely continuous, we have that for each ε > 0, there exists a δ > 0 such
P
Pn
that nj=1 [G(bj ) − G(aj )] < ε if
j=1 (bj − aj ) < δ, where aj < bj ≤ aj+1 < bj+1 , j =
1, 2, . . . , n − 1. This in turn implies that for the above {(aj , bj )}nj=1 ,
n
X
[F (bj ) − F (aj )] =
Z
∞
n
X
[G(bj − z) − G(aj − z)]dH(z) <
−∞ j=1
j=1
Z
∞
ε dH(z) = ε.
−∞
Hence, part (i) is proved. To prove part (ii), we recall first that both the logarithmic and
exponential functions are absolutely continuous, and that the composition preserves the
property of absolute continuity. Then consider log(Y Z) = log Y + log Z and use part (i)
to conclude that log(Y Z) has an absolutely continuous distribution, and hence so does Y Z.
This completes the proof.
It is known that the inverse function of an absolutely continuous function with positive
derivative almost everywhere is not necessarily absolutely continuous. However, we have the
following useful result.
Lemma 2. Let F be an absolutely continuous distribution on [0, 1] and F ′ (x) = f (x) > 0
on (0, 1). Then the inverse function of F is itself an absolutely continuous distribution.
Proof. By the assumptions, F is a strictly increasing and continuous function from [0, 1]
to [0, 1], so is its inverse function F −1 . This implies that F −1 is a continuous distribution on [0, 1]. Moreover,
d
F −1 (t)
dt
= 1/f (F −1(t)) is a positive measurable function on (0, 1)
(see, e.g., Shorack and Wellner 1986, pp. 8-9). By changing variables x = F −1 (t), we have
R 1 d −1
R1
[ F (t)]dt = 0 [1/f (x)] · f (x)dx = 1. Therefore, the distribution function F −1 has no
0 dt
singular part and is absolutely continuous. The proof is complete.
4
Lemma 3. If the distribution Fk,n of order statistic Xk,n is absolutely continuous, then so
is the underlying distribution F of X.
Proof. Recall (see (1)) that Fk,n (x) = Bk,n−k+1(F (x)), x ∈ R, where Bk,n−k+1, defined in
(2), is the beta distribution Bα,β with parameters α = k and β = n − k + 1, and has a
−1
positive continuous density function on (0, 1). Therefore, F = Bk,n−k+1
◦ Fk,n is absolutely
continuous by Lemma 2. The proof is complete.
Lemma 4. Let 1 ≤k
< n. Then we have the following identities:
n
n−1
n−k
k n−k
F F
(i) Fk,n − Fk+1,n =
F kF
, (ii) Fk,n − Fk,n−1 =
, and
k
k−1
n−k
n−1
(iii) Fk,n−1 − Fk+1,n =
F kF
.
k
Proof. For parts (i) and (ii), see, e.g., David and Shu (1978) as well as David and Nagaraja
(2003, p. 23), while part (iii) follows from parts (i) and (ii) and the identity:
n
n−1
n−1
=
+
.
k
k−1
k
Denote the left and the right extremities of F by ℓF and rF , respectively. It is known that
if the absolutely continuous F satisfies the functional equation F ′ (x) = F (x)(1 − F (x)), x ∈
(ℓF , rF ), then F is a logistic distribution. In the next lemma we extend this result.
Lemma 5. Let r, θ > 0, a ∈ [0, 1], and let F be an absolutely continuous distribution
function satisfying xa F ′ (x) = θF (x)(1 − F r (x)) for x ∈ (ℓF , rF ).
(i) If a = 1, then ℓF = 0, rF = +∞ and
1/r
λxrθ
, 0 ≤ x < ∞,
F (x) =
1 + λxrθ
where λ is a positive constant.
(ii) If a ∈ [0, 1), then ℓF = −∞, rF = +∞ and
F (x) =
rθ 1−a
λ exp( 1−a
x )
rθ 1−a
x )
1 + λ exp( 1−a
!1/r
, −∞ < x < ∞,
where λ is a positive constant.
Proof. Define the increasing function G(x) = (1 − F r (x))−1 − 1 from (ℓF , rF ) onto (0, ∞).
Then G′ (x) = rF r−1 (x)F ′ (x)(1 − F r (x))−2 , and hence
xa G′ (x) = rθG(x), x ∈ (ℓF , rF ).
5
(a) If a = 1, solving the above equation leads to G(x) = λxrθ , x ∈ (ℓF , rF ), for some constant
λ > 0. On the other hand, we have, by the definition of G, that
1/r
1/r
G(x)
λxrθ
F (x) =
=
, x ∈ (ℓF , rF ),
1 + G(x)
1 + λxrθ
and hence, ℓF = 0 and rF = +∞, because F is a distribution function. This proves part (i).
rθ 1−a
x ), x ∈ (ℓF , rF ), for some constant
(b) If a ∈ [0, 1), we have instead G(x) = λ exp( 1−a
λ > 0. The required result then follows from both the definition of the function G and the
fact that F is a distribution function. The proof is complete.
Some equalities (in distribution) of the next lemma are essentially due to Nevzorov (2001,
Lecture 3), but we provide here an alternative and possibly simpler proof.
Lemma 6. Let ξk,n (Uk,n , resp.) be the k-th smallest order statistic of a random sample
of size n from the standard exponential distribution E (the uniform distribution U, resp.).
Then the following statements are true.
n−k+1
(i) The Laplace transform of ξk,n is Lξk,n (s) = n−k+1+s
·
d P
1
ξk−j+1, where 1 ≤ k ≤ n.
(ii) ξk,n = kj=1 n−j+1
n−k+2
n−k+2+s
n
· · · n+s
, s ≥ 0.
d
′
(iii) ξm,n = ξm−k,n−k + ξk,n
, where 1 ≤ k < m ≤ n.
d Q
1/(n−j+1)
(iv) Un−k+1,n = kj=1 Uj
, where 1 ≤ k ≤ n.
d
′
(v) Uk,n = Uk,m−1 · Um,n
, where 1 ≤ k < m ≤ n.
n R E(x) k−1
Proof. Recall that the distribution of ξk,n is Fξk,n (x) = k
t (1 − t)n−k dt, x ≥ 0,
k 0
where E(x) = Fξ (x) = 1 − e−x , x ≥ 0. Then the Laplace transform of ξk,n is
Z ∞
Z ∞
n
−sξk,n
−sx
Lξk,n (s) = E[e
]=
e dFξk,n (x) = k
e−(n−k+1+s)x (1 − e−x )k−1 dx, s ≥ 0.
k
0
0
By integration by parts, it follows from the above integral that
Z ∞
k−1
n
Lξk,n (s) = k
e−(n−k+2+s)x (1 − e−x )k−2 dx = · · ·
k n−k+1+s 0
Z ∞
k−2
1
k−1
n
·
···
·
e−(n+s)x dx
= k
k n−k+1+s n−k+2+s
n−1+s 0
n−k+2
n−1
n
n−k+1
·
···
·
, s ≥ 0.
=
n−k+1+s n−k+2+s
n−1+s n+s
This proves part (i), which in turn implies parts (ii) and (iii) by the fact that E[e−s(ξ/k) ] =
d
k/(k + s), s ≥ 0. Part (iv) follows from part (ii) because Uj = exp(−ξk−j+1 ) and the order
6
d
d
′
statistic Un−k+1,n = exp(−ξk,n ). To prove part (v), we have Un−m+1,n = Un−m+1,n−k · Un−k+1,n
by using part (iii), and then reset k = n − m + 1. The proof is complete.
Lemma 7. Let Y obey the Pareto type II (or log-logistic) distribution G(y) = y/(1+y), y ≥
0. Let {Yj }nj=1 , independent of U and {Uj }nj=1 , be a random sample of size n from G. Then
we have the following equalities in distribution:
d
d
(i) 1/Y = Y and in general, 1/Yk,n = Yn−k+1,n , where 1 ≤ k ≤ n.
d
(ii) Yk,n−1 = Yk,n /U 1/(n−k) , where 1 ≤ k ≤ n − 1.
d
(iii) Yk,n−m = Yk,n/Un−k−m+1,n−k , where 1 ≤ k ≤ n − m.
d
(iv) Yk,n−1 = Yk+1,n · U 1/k , where 1 ≤ k ≤ n − 1.
d
(v) Ym−k,n−k = Ym,n · Um−k,m−1 , where 2 ≤ k + 1 ≤ m ≤ n.
Proof. It is easy to check part
To prove the remaining parts, recall that the distri (i).
n R G(y) k−1
bution of Yk,n is Gk,n (y) = k
t (1 − t)n−k dt, y ≥ 0. Then we have H(y) ≡
k 0
R1
Pr(Yk,n /U 1/(n−k) ≤ y) = 0 Gk,n (yu1/(n−k))du. By changing variables,
Z 1
Z 1
n−k
H(y) =
Gk,n (yt)dt
= (n − k)
Gk,n (yt)tn−k−1dt, y ≥ 0.
0
0
Therefore, Gk,n−1 = H iff, by differentiation,
n Z 1
n+1
1
1
=
tn−1 dt, y ≥ 0,
n
1+y
1
+
yt
0
which is, however, a special case of the well-known identity
β1
β1 +β2
Z
1
Γ(β1 + β2 ) 1
1
=
tβ1 −1 (1 − t)β2 −1 dt, y ≥ 0
1+y
Γ(β1 )Γ(β2 ) 0 1 + yt
(see Gradshteyn and Ryzhik 2014, p. 314). This proves part (ii). Part (iii) follows from part
(ii) by iteration and Lemma 6(iv), while part (iv) follows from either parts (i) and (ii) (letting
d
d
k∗ = n − k) or Lemma 8(iii) below because Y = exp(X) and U = exp(−ξ). Finally, we prove
part (v) by using part (iv), iteration and Lemma 6(iv) again. The proof is complete.
Lemma 8. Let X obey the standard logistic distribution F (x) = 1/[1 + exp(−x)], x ∈ R.
Then we have the following equalities in distribution:
d
(i) Xk,n−1 = Xk,n +
1
ξ,
n−k
where 1 ≤ k ≤ n − 1.
d
(ii) Xk,n−m = Xk,n + ξm,n−k , where 1 ≤ k ≤ n − m.
d
(iii) Xk,n−1 = Xk+1,n − k1 ξ, where 1 ≤ k ≤ n − 1.
7
d
(iv) Xm−k,n−k = Xm,n − ξk,m−1 , where 2 ≤ k + 1 ≤ m ≤ n.
d
d
Proof. The results follow from Lemma 7 by noting that (a) X = log Y, (b) ξ = − log U and
d
(c) − log Uk,n = ξn−k+1,n for all 1 ≤ k ≤ n. The proof is complete.
For the proof of the next lemma, see Lin and Hu (2008, Lemma 5).
Lemma 9. Let f and g be two functions real analytic and strictly monotone in [0, ∞).
Assume that for each n ≥ 1, the n-th derivatives f (n) and g (n) are strictly monotone in some
interval [0, δn ). Let {xn }∞
n=1 be a sequence of positive real numbers converging to zero. If
f (xn ) = g(xn ), n = 1, 2, . . . , then f = g.
3. Characterizations of the Pareto type II distribution
We start with the Pareto type II distribution which is easier to handle, and recall that
the uniform order statistic Uk,n ∼ Bk,n−k+1.
Theorem 1. Let Y ∼ G be a positive random variable and let 1 ≤ k ≤ n − 1 be fixed
integers. Let Y1 , Y2 , . . . , Yn be n independent copies of Y, and let U, independent of {Yi }ni=1 ,
be a random variable with uniform distribution on [0, 1]. Then the distributional equality
d
Yk,n−1 = Yk,n /U 1/(n−k)
(5)
holds iff G is a Pareto distribution G(y) = λy/(1 + λy), y ≥ 0, where λ > 0 is a constant.
d
Proof. The sufficiency part follows from Lemma 7(ii) because (λY )ℓ,m = λYℓ,m for λ > 0 and
for all 1 ≤ ℓ ≤ m. To prove the necessity part, we note, by Lemma 1(ii), that the distribution
Gk,n−1 of Yk,n−1 is absolutely continuous, and so is G by Lemmas 2 and 3. Rewrite (5) as
Z 1
Gk,n−1(y) =
Gk,n (yu1/(n−k) )du, y ≥ 0.
0
By changing variables t = yu1/(n−k), we have
Gk,n−1(y) = (n − k)y
−(n−k)
Z
y
Gk,n (t)tn−k−1 dt, y > 0.
0
Taking differentiation leads to
yG′k,n−1(y) = (n − k)[Gk,n (y) − Gk,n−1(y)], y > 0.
With the help of (1) and Lemma 4(ii), (6) is equivalent to
yG′(y) = G(y)[1 − G(y)], y ∈ (ℓG , rG ).
8
(6)
Finally, Lemma 5(i) with r = θ = 1 completes the proof.
Corollary 1.
Under the same assumptions of Theorem 1, the distributional equality
d
Yk,n−1 = Yk,n/U 1/α holds for some α > 0, iff G is a Pareto distribution with G(y) =
1/[1 + λy α/(n−k) ], y ≥ 0, where λ is a positive constant.
α/(n−k)
Proof. Consider Yi′ = Yi
α/(n−k)
′
for i = 1, 2, . . . . Then Yk,n
= Yk,n
, and apply Theorem
d
′
′
1 to the case: Yk,n−1
= Yk,n
/U 1/(n−k) .
Corollary 2. Under the same assumptions of Theorem 1, the distributional equality
d
Yk,n−1 = Yk+1,n · U 1/k
holds iff G is a Pareto distribution G(y) = λy/(1 + λy), y ≥ 0, where λ > 0 is a constant.
Proof. The sufficiency part is a consequence of Lemma 7(iv). To prove the necessity
d
∗
part, denote Y ∗ = 1/Y. Then Yℓ,m
= (1/Y )ℓ,m = 1/Ym−ℓ+1,m for all 1 ≤ ℓ ≤ m. By
d
d
∗
assumptions, we have the equality 1/Yk,n−1 = 1/Yk+1,n · 1/U 1/k , or, equivalently, Yn−k,n−1
=
∗
Yn−k,n
· 1/U 1/(n−(n−k)) . It follows from Theorem 1 (letting k∗ = n − k) that Y ∗ has a Pareto
distribution G∗ (y) = λ∗ y/(1 + λ∗ y), y ≥ 0, for some constant λ∗ > 0. This in turn implies
that Y has the Pareto distribution G(y) = λy/(1 + λy), y ≥ 0, where λ = 1/λ∗. We can
d
prove the last claim directly, or by using Lemma 7(i), because λ∗ Y ∗ = 1/(λ∗ Y ∗ ) = Y /λ∗
having the standard Pareto type II distribution. The proof is complete.
Theorem 2. Let Y ∼ G be a positive random variable and let 1 ≤ k ≤ n − 1 be fixed
integers. Let Y1 , Y2 , . . . , Yn be n independent copies of Y, and let B, independent of {Yi }ni=1 ,
be a random variable having beta distribution Bα,β with parameters α = 1 and β = n − k,
that is, FB (u) = 1 − (1 − u)n−k , u ∈ [0, 1]. Assume further that limy→0+ G(y)/y = λ > 0.
Then the distributional equality
d
Yk,k = Yk,n /B
(7)
holds iff G is the Pareto distribution G(y) = λy/(1 + λy), y ≥ 0.
d
Proof. The sufficiency part follows from Lemma 7(iii) with n − m = k and the fact B =
U1,n−k . To prove the necessity part, we note first that G is absolutely continuous by Lemmas
1–3, and then rewrite (7) as the functional equation:
Z 1 Z G(yu)
n k−1
k
G (y) =
k
t (1 − t)n−k dtdFB (u), y ≥ 0.
k
0
0
9
(8)
Now, it suffices to prove that the solution of equation (7) is unique under the smoothness
condition on the distribution. Namely, if the absolutely continuous distribution F on (0, ∞)
satisfies limy→0+ F (y)/y = λ > 0 and
k
F (y) =
Z
1
0
Z
F (yu)
0
n k−1
t (1 − t)n−k dtdFB (u), y ≥ 0,
k
k
then we will prove that F = G. From (8) and (9) it follows that
Z 1
1
k
k
|F (y) − G (y)| ≤
|F k (yu) − Gk (yu)|dFB (u), y ≥ 0,
E[B k ] 0
n
k
where E[B ] = 1/
. Define the bounded function
k
g(y) =
(9)
(10)
F k (y) − Gk (y)
, y > 0, and g(0) = lim+ g(y) = 0,
y→0
yk
and the increasing function
h(y) = sup g(t), y > 0, and h(0) = lim+ h(y) = 0.
y→0
0≤t≤y
By (10), we see that
g(y) ≤
Z
1
g(uy)dH(u), y ≥ 0,
(11)
0
where H(u) = (1/E[B k ])
Ru
0
tk dFB (t), u ∈ [0, 1]. Now, by (11) and the definition of the
increasing function h, we have
Z 1
Z
h(y) ≤
h(uy)dH(u) ≤ h(y)
0
1
dH(u) = h(y), y ≥ 0.
0
This in turn implies that h is a constant function and hence h(y) = 0, y ≥ 0, because
h(0) = h(0+ ) = 0. Consequently, g(y) = 0, y ≥ 0, and F = G. The proof is complete.
The next result is the counterpart of Theorem 2 for the minimum order statistics.
Corollary 3. Under the same setting in Theorem 2 with the condition on G replaced by
limy→0+ G(1/y)/y = 1/λ > 0 (equivalently, limy→+∞ yG(y) = 1/λ > 0), the distributional
d
equality Y1,k = Yn−k+1,n · B holds iff G is the Pareto distribution G(y) = λy/(1 + λy), y ≥ 0.
d
Proof. Use Lemma 7(v), Theorem 2 and the fact (1/Y )ℓ,m = 1/Ym−ℓ+1,m for all 1 ≤ ℓ ≤ m.
We now further extend Theorem 2 under some stronger smoothness conditions.
10
Theorem 3. Let Y ∼ G be a positive random variable and let n, m, k be three fixed positive
integers with 1 ≤ k ≤ n − m. Let Y1 , Y2 , . . . , Yn be n independent copies of Y, and let B1 ,
independent of {Yi }ni=1 , be a random variable having beta distribution Bα,β with parameters
α = n − m − k + 1 and β = m. Assume further that the distribution function G satisfies the
following conditions:
(i) G is real analytic and strictly increasing in [0, ∞) and for each i ≥ 1, its i-th derivative
G(i) is strictly monotone in some interval [0, δi ).
(ii) limy→0+ [Gk (y) − (λy)k ]/(λy)k+1 = −k for some positive constant λ.
Then the distributional equality
d
Yk,n−m = Yk,n /B1
(12)
holds iff G is the Pareto distribution G(y) = λy/(1 + λy), y ≥ 0.
d
Proof. The sufficiency part follows from Lemma 7(iii) and the fact B1 = Un−m−k+1,n−k . To
prove the necessity part, we note first that G is absolutely continuous as before, and then
we rewrite (12) as the functional equation:
Z 1
Gk,n−m(y) =
Gk,n (yu)dFB1 (u), y ≥ 0.
(13)
0
Now, it suffices to prove that the solution of equation (12) is unique under the smoothness
condition on the distribution. Namely, if the absolutely continuous distribution F on (0, ∞)
satisfies the above conditions (i) and (ii) and
Z 1
Fk,n−m(y) =
Fk,n (yu)dFB1 (u), y ≥ 0,
(14)
0
then we will prove that F = G.
Define the increasing function H(y) = max{F (y), G(y)}, y ≥ 0. From (1) it follows that
for any a > 0 and 0 ≤ y ≤ a,
Z F (y)
n−m
|Fk,n−m (y) − Gk,n−m (y)| = k
tk−1 (1 − t)n−m−k dt
k
G(y)
1
n−m
≥ k
(1 − H(y))n−m−k F k (y) − Gk (y)
k
k
n−m
≥
(1 − H(a))n−m−k F k (y) − Gk (y) .
k
11
(15)
On the other hand, we have
n
F k (y) − Gk (y) , y ≥ 0.
|Fk,n (y) − Gk,n (y)| ≤
k
Combing (12)–(16) leads to
n−m
(1 − H(a))n−m−k F k (y) − Gk (y) ≤ |Fk,n−m(y) − Gk,n−m(y)|
k
Z 1
Z 1
n
≤
|Fk,n (yu) − Gk,n (yu)| dFB1 (u) ≤
F k (yu) − Gk (yu) dFB1 (u).
k
0
0
(16)
(17)
Define the bounded increasing function
h(y) = sup
0<t≤y
F k (t) − Gk (t)
, y > 0,
tk+1
and h(0) = lim+ h(y) = 0.
y→0
Then from the inequality (17) it follows that for any a > 0,
Z 1
n−m
n
n−m−k
(1 − H(a))
h(y) ≤
h(yu)uk+1dFB1 (u)
k
k
0
Z 1
n
n
h(y)
≤
h(y)E[B1k+1], 0 ≤ y ≤ a.
uk+1dFB1 (u) =
k
k
0
n−m
n
Recall that E[B1k ] =
/
. Then rewrite the inequality (18) as follows:
k
k
E[B1k ](1 − H(a))n−m−k h(y) ≤ h(y)E[B1k+1], 0 ≤ y ≤ a, a > 0.
(18)
(19)
We claim that there exists a y0 > 0 such that h(y0 ) = 0. Otherwise, we have, by (19),
E[B1k ](1 − H(a))n−m−k ≤ E[B1k+1 ], ∀ a > 0,
which in turn implies, by letting a → 0+ , that E[B1k ] ≤ E[B1k+1 ], a contradiction. Therefore,
h(y0 ) = 0 for some y0 > 0 and hence F (y) = G(y) for y ∈ [0, y0]. By Lemma 9 and the
assumptions on F and G, we conclude that F = G. The proof is complete.
Corollary 4. Let Y ∼ G be a positive random variable and let n, m, k be three fixed positive
integers with k + 1 ≤ m ≤ n. Let Y1 , Y2 , . . . , Yn be n independent copies of Y, and let B2 ,
independent of {Yi }ni=1 , be a random variable having beta distribution Bα,β with parameters
α = m − k and β = k. Assume further that the distribution function G∗ of 1/Y satisfies the
following conditions:
12
(i) G∗ is real analytic and strictly increasing in [0, ∞) and for each i ≥ 1, its i-th derivative
(i)
G∗ is strictly monotone in some interval [0, δi ).
(ii) limy→0+ [Gk∗∗ (y) − (y/λ)k∗ ]/(y/λ)k∗ +1 = −k∗ for some positive constant λ, where k∗ =
n − m + 1.
d
Then the distributional equality Ym−k,n−k = Ym,n · B2 holds iff G is the Pareto distribution
G(y) = λy/(1 + λy), y ≥ 0.
d
Proof. Use Lemma 7(v), Theorem 3 and the fact (1/Y )ℓ,m = 1/Ym−ℓ+1,m for all 1 ≤ ℓ ≤ m.
In summary, for a positive random variable Y ∼ G, we have the following characteristic
properties of the Pareto distribution G(y) = λy/(1 + λy), y ≥ 0, where λ is a positive
constant (compare with Lemma 7).
d
1. Yk,n−1 = Yk,n/U 1/(n−k) .
d
2. Yk,n−1 = Yk+1,n · U 1/k .
d
d
3. Yk,k = Yk,n /B (equivalently, Yk,k = Yk,n /U1,n−k ).
d
d
4. Y1,k = Yn−k+1,n · B (equivalently, Y1,k = Yn−k+1,n · U1,n−k ).
d
d
5. Yk,n−m = Yk,n /B1 (equivalently, Yk,n−m = Yk,n /Un−m−k+1,n−k ).
d
d
6. Ym−k,n−k = Ym,n · B2 (equivalently, Ym−k,n−k = Ym,n · Um−k,m−1 ).
Here, the random variables U ∼ U, B ∼ B1,n−k , B1 ∼ Bn−m−k+1,m , B2 ∼ Bm−k,k , and on
the RHS of each equality, the two random variables are independent.
4. Characterizations of the logistic distribution
We are now ready to provide characterization results of the logistic distribution.
Theorem 4. Let X ∼ F and let 1 ≤ k ≤ n − 1 be fixed integers. Then the distributional
equality
d
Xk,n−1 = Xk,n +
1
ξ
n−k
holds iff F is a logistic distribution F (x) = 1/[1 + e−(x−µ) ], x ∈ R, where µ ∈ R is a constant.
Proof. Take Yi = exp(Xi ), U = exp(−ξ) and λ = e−µ . Then the result follows immediately
from Theorem 1.
Corollary 5. Let X ∼ F, α > 0 and let 1 ≤ k ≤ n − 1 be fixed integers. Then the
d
distributional equality Xk,n−1 = Xk,n + α1 ξ holds iff F is a logistic distribution F (x) =
1/{1 + e−[α/(n−k)](x−µ) }, x ∈ R, where µ ∈ R is a constant.
13
Corollary 6. Let X ∼ F, α > 0 and let 1 ≤ k ≤ n − 1 be fixed integers. Then the
distributional equality
d
Xk,n−1 = Xk+1,n −
1
ξ
α
(20)
holds iff F is a logistic distribution F (x) = 1/[1 + e−(α/k)(x−µ) ], x ∈ R, where µ ∈ R is a
constant.
d
Proof. Use Corollary 5 and the fact that Xℓ,m = −(−X)m−ℓ+1,m for all 1 ≤ ℓ ≤ m.
d
The counterpart of (20), namely, Xk+1,n = Xk,n−1 + α1 ξ, and the two-sided case: Xk,n +
d
aξ1 = Xk,n−1 + bξ2 (see Corollary 5), where α, a, b > 0, were investigated by Wesolowski and
Ahsanullah (2004). All the solutions of these two equations are exponential distributions.
The next result improves and extends Theorem 6 of Lin and Hu (2008) by an approach
different from the previous method of intensively monotone operator (Kakosyan et al. 1984).
Theorem 5.
Let 1 ≤ k ≤ n − 1 be fixed integers.
Assume that X ∼ F satisfies
limx→−∞ F (x)/ex = e−µ for some constant µ ∈ R. Then the distributional equality
d
Xk,k = Xk,n + ξn−k,n−k
holds iff F is the logistic distribution F (x) = 1/[1 + e−(x−µ) ], x ∈ R.
Proof. The sufficiency part follows from Lemma 8(ii), while the necessity part is a consequence of Theorem 2.
The next result is the counterpart of Theorem 5 for the minimum order statistics.
Corollary 7. Let 1 ≤ k ≤ n − 1 be fixed integers.
Assume that X ∼ F satisfies
limx→−∞ F (−x)/ex = eµ for some constant µ ∈ R. Then the distributional equality
d
X1,k = Xn−k+1,n − ξn−k,n−k
holds iff F is the logistic distribution F (x) = 1/[1 + e−(x−µ) ], x ∈ R.
d
Proof. Use Lemma 8(iv), Theorem 5 and the fact Xℓ,m = −(−X)m−ℓ+1,m for all 1 ≤ ℓ ≤ m.
Using Theorem 3 and Lemma 8(ii), we further extend Theorem 5 to the following.
Theorem 6. Let n, m, k be three fixed positive integers with 1 ≤ k ≤ n − m and let X ∼ F
satisfy limx→−∞ [e−k(x−µ) F k (x) − 1]/ex−µ = −k for some constant µ ∈ R. Assume further
that the distribution function G of exp(X1 ) is real analytic and strictly increasing in [0, ∞)
14
and that for each i ≥ 1, its i-th derivative G(i) is strictly monotone in some interval [0, δi ).
Then the distributional equality
d
Xk,n−m = Xk,n + ξm,n−k
holds iff F is the logistic distribution F (x) = 1/[1 + e−(x−µ) ], x ∈ R.
As before, Theorem 6 and Lemma 8(iv) together lead to the following.
Corollary 8. Let n, m, k be three fixed positive integers with k + 1 ≤ m ≤ n and let
X ∼ F satisfy limx→−∞ [e−k∗ (x+µ) (F (−x))k∗ − 1]/ex+µ = −k∗ for some constant µ ∈ R, where
k∗ = n−m+1. Assume further that the distribution function G∗ of exp(−X1 ) is real analytic
(i)
and strictly increasing in [0, ∞) and that for each i ≥ 1, its i-th derivative G∗ is strictly
monotone in some interval [0, δi ). Then the distributional equality
d
Xm−k,n−k = Xm,n − ξk,m−1
(21)
holds iff F is the logistic distribution F (x) = 1/[1 + e−(x−µ) ], x ∈ R.
In summary, for a random variable X ∼ F, we have the following characteristic properties
of the logistic distribution F (x) = 1/[1 + e−(x−µ) ], x ∈ R (compare with Lemma 8). Here,
µ ∈ R, ξ ∼ E, and on the RHS of each equality, the two random variables are independent.
d
1. Xk,n−1 = Xk,n +
1
ξ.
n−k
d
2. Xk,n−1 = Xk+1,n − k1 ξ.
d
3. Xk,k = Xk,n + ξn−k,n−k .
d
4. X1,k = Xn−k+1,n − ξn−k,n−k .
d
5. Xk,n−m = Xk,n + ξm,n−k .
d
6. Xm−k,n−k = Xm,n − ξk,m−1 .
5. Open problem
Finally, we would like to pose the following problem in which part (i) is the counterpart
of (21) for exponential distribution, and in part (ii), the first two cases, m = k + 1, k + 2,
have been solved by AlZaid and Ahsanullah (2003) and Ahsanullah et al. (2010).
Problem. Let X ∼ F and let 1 ≤ k < m ≤ n be fixed integers. Then solve the general
d
d
distributional equations: (i) Xm,n = Xm−k,n−k + ξk,n and (ii) Xm,n = Xk,n + ξm−k,n−k .
15
Acknowledgments. The authors would like to thank the Editor-in-Chief and the Referee
for helpful comments and constructive suggestions which improve the presentation of the
paper.
References
Ahsanullah, M., Berred, A. and Nevzorov, V.B. (2011). On characterizations of the exponential distributions. J. Appl. Statist. Sci., 19, 37–43.
Ahsanullah, M., Nevzorov, V.B. and Yanev, G.P. (2010). Characterizations of distributions
via order statistics with random exponential shifts. J. Appl. Statist. Sci., 18, 297–305.
Ahsanullah, M., Yanev, G.P. and Onica, C. (2012). Characterizations of logistic distribution through order statistics with independent exponential shifts. Economic Quality
Control, 27, 85–96.
AlZaid, A.A. and Ahsanullah, M. (2003). A characterization of the Gumbel distribution
based on record values. Commun. Statist. - Theory and Methods, 32, 2101–2108.
Ananjevskii, S.M. and Nevzorov, V.B. (2016). On families of distributions characterized
by certain properties of ordered random variables. Vestnik St. Petersburg University:
Mathematics, 49, 197–203.
Berred, A. and Nevzorov, V.B. (2013). Characterizations of families of distributions, which
include the logistic one, by properties of order statistics. Journal of Mathematical
Sciences, 188, 673–676.
David, H.A. and Nagaraja, H.N. (2003). Order Statistics, 3rd edn. Wiley, New Jersey.
David, H.A. and Shu, V.S. (1978). Robustness of location estimators in presence of an
outlier. In: Contributions to Survey Sampling and Applied Statistics, ed. H.A. David,
pp. 235–250. Academic Press, New York.
George, E.O. and Mudholkar, G.S. (1982). On the logistic and exponential laws. Sankhyā,
Ser. A, 44, 291–293.
16
Gradshteyn, I.S. and Ryzhik, I.M. (2014). Table of Integrals, Series, and Products, 8th edn.
Elsevier, New York.
Hwang, J.S. and Lin, G.D. (1984). Characterizations of distributions by linear combinations
of moments of order statistics. Bulletin of the Institute of Mathematics, Academia
Sinica, 12, 179–202.
Kakosyan, A.V., Klebanov, L.B. and Melamed, J.A. (1984). Characterization of Distributions by the Method of Intensively Monotone Operators. Springer, New York.
Kotz, S. (1974). Characterizations of statistical distributions: a supplement to recent
surveys. Int. Statist. Rev., 42, 39–65.
Lin, G.D. and Hu, C.-Y. (2008). On characterizations of the logistic distribution. J. Statist.
Plann. Infer., 138, 1147–1156.
Lukacs, E. (1970). Characteristic Functions, 2nd edn. Hafner Pub. Co., New York.
Mudholkar, G.S. and George, E.O. (1978). A remark on the shape of the logistic distribution. Biometrika, 65, 667–668.
Nevzorov, V.B. (2001). Records: Mathematical Theory. Translations of Mathematical
Monographs, Vol. 194, Amer. Math. Soc., Rhode Island.
Shorack, G.R. and Wellner, J.A. (1986). Empirical Processes with Applications to Statistics.
Wiley, New York.
Wesolowski, J. and Ahsanullah, M. (2004). Switching order statistics through random
power contractions. Aust. N.Z. J. Statist., 46, 297–303.
Zykov, V.O. and Nevzorov, V.B. (2011). Some characterizations of families of distributions,
including logistic or exponential ones, by properties of order statistics. Journal of
Mathematical Sciences, 176, 203–206.
17
| 10math.ST
|
arXiv:1611.05911v2 [cs.DS] 24 Feb 2017
Computing Absolutely Normal Numbers in
Nearly Linear Time
Jack Lutz∗
Elvira Mayordomo†
February 27, 2017
Abstract
A real number x is absolutely normal if, for every base b ≥ 2, every
two equally long strings of digits appear with equal asymptotic frequency
in the base-b expansion of x. This paper presents an explicit algorithm
that generates the binary expansion of an absolutely normal number x,
with the nth bit of x appearing after npolylog(n) computation steps. This
speed is achieved by simultaneously computing and diagonalizing against
a martingale that incorporates Lempel-Ziv parsing algorithms in all bases.
Keywords: algorithms, computational complexity, Lempel-Ziv parsing, martingales, normal numbers
1
Introduction
In 1909 Borel [4] defined a real number α to be normal in base b (b ≥ 2) if, for
every m ≥ 1 and every length-m sequence w of base-b digits, the asymptotic,
empirical frequency of w in the base-b expansion of α is b−m . Borel defined α to
be absolutely normal if it is normal in every base b ≥ 2. (This clearly anticipated
the fact, proven a half-century later, that a real number may be normal in one
base but not in another [8, 23].) Borel’s proof that almost every real number
(i.e., every real number outside a set of Lebesgue measure 0) is absolutely normal
was an important milestone in the prehistory of Kolmogorov’s development
of the rigorous, measure-theoretic foundations of probability theory [17]. For
example, it is section 1 of Billingsley’s influential textbook [3]. The recent book
[7] provides a good exposition of the many aspects of current research on normal
numbers.
Borel’s proof shows that absolutely normal numbers are commonplace, i.e.,
that a “randomly chosen” real number is absolutely normal with probability 1.
Rational numbers cannot be normal in even a single base b, since their base-b
∗ Department of Computer Science, Iowa State University, Ames, IA 50011 USA.
[email protected].
† Departamento de Informática e Ingenierı́a de Sistemas, Instituto de Investigación en Ingenierı́a de Aragón, Universidad de Zaragoza, 50018 Zaragoza, SPAIN. [email protected].
1
expansions
are eventually periodic, but computer analyses of the expansions of
√
π, e, 2, ln 2, and other irrational numbers that arise in common mathematical
practice suggest that these numbers are absolutely normal [5]. Nevertheless,
no such “natural” example of a real number has been proven to be normal in
any base, let alone absolutely normal. The conjectures that every algebraic
irrational is absolutely normal and that π is absolutely normal are especially
well known open problems [5, 7, 29].
This paper concerns an old problem, namely, the complexity of explicitly
computing a real number that is provably absolutely normal, even if it is not
natural in the above informal sense. Sierpinski and Lebesgue gave explicit constructions of absolutely normal numbers in 1917 [25, 18], but these were intricate
limiting processes that offered no complexity analyses (coming two decades before the theory of computing) and little insight into the nature of the numbers
constructed. In a 1936 note that was not published in his lifetime, Turing [27]
gave a constructive proof that almost all real numbers are absolutely normal
and then derived constructions of absolutely normal numbers from this proof.
Moreover, although Turing does not mention Turing machines or computability,
the note is typed, with equations handwritten by him, on the back of a draft
of his paper on computable real numbers [26], so it is reasonable to interpret
“constructively” in a computability-theoretic sense. And in fact his proof, with
2007 corrections by Becher, Figueira, and Picchi [1], explicitly computes an absolutely normal number. In 2013, Becher, Heiber, and Slaman [2] published
an algorithm that computes an absolutely normal number in polynomial time.
Specifically, this algorithm computes the binary expansion of an absolutely normal number x, with the nth bit of x appearing after O(n2 f (n)) steps for any
computable unbounded nondecreasing function f . (Unpublished polynomialtime algorithms for computing absolutely normal numbers were also announced
independently by Mayordomo [22] and Figueira and Nies [12, 13] at about the
same time.)
In this paper we present a new algorithm that provably computes an absolutely normal in nearly linear time. Our algorithm computes the binary expansion of an absolutely normal number x, with the nth bit of x appearing
after O(npolylog(n)) steps. The term “nearly linear time” was introduced by
Gurevich and Shelah [14]. In that paper they showed that, while linear time
computability is very model-dependent, nearly linear time is very robust. For
example, they showed that random access machines, Kolmogorov-Uspensky machines, Schoenhage machines, and random-access Turing machines share exactly
the same notion of nearly linear time.
Our algorithm uses the Lempel-Ziv parsing algorithm to achieve its nearly
linear time bound. For each base b ≥ 2, we use a martingale (betting strategy)
that employs the Lempel-Ziv parsing algorithm and is implicit in the work
of Feder [11]. This base-b Lempel-Ziv martingale succeeds exponentially when
betting on the successive digits of the base-b expansion of any real number that is
not normal in base b. Our algorithm simultaneously computes and diagonalizes
against (limits the winnings of) a martingale that incorporates efficient proxies
of all these martingales, thereby efficiently computing a real number that is
2
normal in every base.
The rest of this paper is organized as follows. Section 2 presents the baseb Lempel-Ziv martingales and their main properties. Section 3 shows how to
transform a base-b Lempel-Ziv martingale into a base-b martingale with an efficiently computable nondecreasing savings account that is unbounded whenever
the base-b Lempel-Ziv martingale succeeds exponentially. Section 4 develops
an efficient method for converting a base-b martingale with an efficiently computable savings account to a base-2 martingale that succeeds whenever the base
b savings account is unbounded. Section 5 presents an algorithm that exploits
the uniformity of these constructions to efficiently and simultaneously compute
(a) a single base-2 martingale d that succeeds on the binary expansion of every
real number x for which some base-b martingale succeeds on the base-b expansion of x and (b) a particular real number x on which binary expansion d does
not succeed. This x is, perforce, absolutely normal. Section 6 presents an open
problem related to our work.
2
Lempel-Ziv Martingales
For each base b ≥ 1 we let Σb = {0, 1, . . . , b − 1} be the alphabet of base-b digits.
We write Σ∗b for the set of all (finite) strings over Σb and Σ∞
b for the set of all
(infinite) sequences over Σb . We write |x| for the length of a string or sequence
x, and we write λ for the empty string, the string of length 0. For x ∈ Σ∗b ∪ Σ∞
b
and 0 ≤ i ≤ j < |x|, we write x[i..j] for the string consisting of the ith through
jth digits in x. For x ∈ Σ∗b ∪ Σ∞
b and 0 ≤ n < |x|, we write x ↾ n = x[0..n − 1].
For w ∈ Σ∗b and x ∈ Σ∗b ∪ Σ∞
b , we say that w is a prefix of x, and we write
w ⊑ x, if x ↾ |w| = w.
A (base-b) martingale is a function d : Σ∗b → [0, ∞) satisfying
1 X
d(wa)
(1)
d(w) =
b
a∈Σb
Σ∗b .
for all w ∈
(This is the original martingale notion introduced by Ville
[28] and implicit in earlier papers of Lévy [20, 21]. Its relationship to Doob’s
subsequent modifications [10], which are the “martingales” of probability theory,
is explained in [16] along with the reason why Ville’s original notion is still
essential for algorithmic information theory.) Intuitively, a base-b martingale d
is a strategy for betting on the subsequent digits in a sequence S ∈ Σ∞
b , with
the strategy encoded in such a way that d(S ↾ n) is the amount of money that
a gambler using the strategy d has after the first n bets. The condition (1)
says that the payoffs for these bets are fair in the sense that the conditional
expectation of d(wa), given that w has occurred (and assuming that the digits
a ∈ Σb are equally likely), is d(w).
A function g : Σ∗b → [0, ∞) (which may or may not be a martingale) succeeds
on a sequence S ∈ Σ∞
b if
lim sup g(S ↾ n) = ∞,
n→∞
3
(2)
i.e., if its winnings on S are unbounded. The success set of a function g : Σ∗b →
[0, ∞) is
S ∞ [g] = {S ∈ Σ∞
b | g succeeds on S } .
A function g : Σ∗b → [0, ∞) succeeds exponentially on a sequence S ∈ Σ∞
b if
lim sup
n→∞
log g(S ↾ n)
> 0,
n
(3)
i.e., if its winnings on S grow at some exponential rate, perhaps with recurrent
setbacks. The exponential success set of a function g : Σ∗b → [0, ∞) is
S exp[g] = {S ∈ Σ∞
b | g succeeds exponentially on S } .
The f (n) success set of a function g : Σ∗b → [0, ∞) is
log g(S ↾ n)
≥
1
.
S f (n) [g] = S ∈ Σ∞
lim
sup
b
log f (n)
n→∞
ǫn
Note that S exp [g] = ∪ǫ>0 S 2 [g].
A function g : Σ∗b → [0, ∞) succeeds strongly on a sequence S ∈ Σ∞
b if (2)
holds with the limit superior replaced by a limit inferior i.e., if the winnings
converge to ∞. A function g : Σ∗b → [0, ∞) succeeds strongly exponentially
on a sequence S ∈ Σ∞
b if (3) holds with the limit superior replaced by a limit
inferior i.e., if the winnings grow at exponential rate. The strong success sets
exp
∞
Sstr
[g] and Sstr
[g] of a function g : Σ∗b → [0, ∞) are defined in the now-obvious
manner. It is clear that the inclusions
exp
Sstr
[g] ⊆ S exp [g] ⊆ S ∞ [g]
and
exp
∞
Sstr
[g] ⊆ Sstr
[g] ⊆ S ∞ [g]
hold for all g : Σ∗b → [0, ∞).
For each base b ≥ 2 the base-b Lempel-Ziv martingale is a particular martingale dLZ(b) based on the Lempel-Ziv parsing algorithm [19], as we now explain.
Formally, dLZ(b) is computed by the algorithm in Figure 1, but some explanation is appropriate here. The algorithm is written with several instances of
parallel assignment. For example, the second line initializes x, L(x), and d to
the values λ, 1, and 1, respectively. The items T , j, and x(j) are not needed
for the computation of dLZ(b) (w), but they are useful for understanding and
analyzing the algorithm.
4
i n p u t w ∈ Σ∗b ;
x, L(x), d = λ, 1, 1 ;
T, j = {λ}, 0 ;
while true do
begin
i f w = λ then o utput d and h a l t ;
i f L(x) = 1 then
begin
L(x) = b ;
for ea ch 0 ≤ a < b do L(xa) = 1 ;
T, x(j), j = T ∪ {x}Σb , x, j + 1 ;
x = λ;
end
else
begin
a, w = head(w), tail(w) ;
L(x), x, d = L(x) + b − 1, xa, bL(xa)
L(x) d
end
end
Figure 1: Algorithm for computing dLZ(b) (w).
The growing set T of strings in Σ∗b always contains all the prefixes of all its
elements, so it is a tree. We envision this tree as being oriented with its root at
the top and the immediate children v0, v1, . . . , v(b − 1) of each interior vertex
v of T displayed left-to-right below T . The dictionary of the algorithm is the
current set of leaves of T .
The string x in the algorithm is always an element of (i.e., location in) the
tree T , and L(x) is always the number of leaves of T that are descendants of x.
We regard x as a descendant of itself, so x is a leaf if and only if L(x) = 1.
It is clear that dLZ(b) (λ) = 1. In fact, the algorithm’s successive values of
d are the values dLZ(b) (u) for successive prefixes u of the input string w. More
precisely, if wt and dt are the values of w and d after t executions of the elseblock, then w = (w ↾ t)wt and dt = dLZ(b) (w ↾ t).
For w ∈ Σ∗b we define the tree T (w) as follows. If w = λ, then T (w) = {λ}.
If w = w′ a, where w′ ∈ Σ∗b and a ∈ Σb , then T (wa) is the value of T when
the algorithm terminates on input w′ . (Note that this is one step before it
terminates on input w′ a.) For w ∈ Σ∗b we define D(w) to be the number of
leaves in T (w).
The computation is divided into “epochs”. At the beginning of each epoch,
the string x is λ, i.e., it is located at the root of T . The string x then takes
successive digits from whatever is left of w (because a, w = head(w), tail(w)
removes the first digit of w and stores it in a), following this path down the
tree and updating d at each step, until w is empty (the end of the last epoch)
5
or x is a leaf if T . In the latter case, the jth epoch is over, the b children
x0, x1, . . . , x(b − 1) are added to T as new leaves, x is the jth phrase x(j) of w,
and x is reset to the root λ of T .
When the algorithm terminates, it is clear that exactly one of the following
things must hold.
(a) w = λ.
(b) w = x(1) . . . x(j).
(c) w = x(1) . . . x(j)u for some nonempty interior vertex u of T (w).
In case (a) or (b) we call w a full parse. In case (b) or (c) we call x(1), . . . , x(j)
the full phrases of w. In case (c) we call u the partial phrase of w.
Define the set Ab = {1 + (b − 1)r | r ∈ N } and the generalized factorial function f actb : Ab → Ab by
f actb (1 + (b − 1)r) =
r
Y
k=1
(1 + (b − 1)k)
for all r ∈ N.
Observation 2.1 For all n ∈ Ab ,
1≤
f actb (n)
n ≤ n.
1
n b−1
e b−1
(4)
e
Lemma 2.2 (Feder [11]) Let w ∈ Σ∗b .
1. If w is a full parse, then
dLZ(b) (w) =
b|w|
.
f actb (D(w))
2. If w is not a full parse and u is its partial phrase, then
dLZ(b) (w) =
b|w|
L(u),
f actb (D(w))
where L(u) is the number of leaves below u in T (w).
Lemma 2.3 For S ∈ Σ∞
b and α ∈ (0, 1) the following three conditions are
equivalent.
(1−α)n
(a) S ∈ S b
[dLZ(b) ].
(b) There exist infinitely many full parses w ⊑ S for which
D(w) logb |w| < α(b − 1)|w|.
6
(c) There exist infinitely many full parses w ⊑ S for which
D(w) log b D(w) < α(b − 1)|w|.
Corollary 2.4 For S ∈ Σ∞
b the following three conditions are equivalent.
(a) S ∈ S exp [dLZ(b) ].
(b) There exist α < 1 and infinitely many full parses w ⊑ S for which
D(w) logb |w| < α(b − 1)|w|.
(c) There exist α < 1 and infinitely many full parses w ⊑ S for which
D(w) log b D(w) < α(b − 1)|w|.
We conclude this section by explaining the connection between the LempelZiv martingales and normality. First, Schnorr and Stimm [24] defined (implicitly) a notion of finite-state base-b martingale and proved that every sequence
S ∈ Σ∞
b obeys the following dichotomy.
1. If S is normal, then no finite-state base-b martingale succeeds on S. (In
fact, every finite-state base-b martingale decays exponentially on S.)
2. If S is not normal, then some finite-state base-b martingale succeeds exponentially on S.
Some twenty years later, Feder [11] defined (implicitly) the Lempel-Ziv martingale dLZ(b) and proved (implicitly) that dLZ(b) is at least as successful on
every sequence as every finite-state base-b martingale. That is, for finite-state
base-b martingale d, the inclusions
∞
S ∞ [d] ⊆ S ∞ [dLZ(b) ], Sstr
[d] ⊆ S ∞ [dLZ(b) ],
exp
exp
S exp [d] ⊆ S exp [dLZ(b) ], Sstr
[d] ⊆ Sstr
[dLZ(b) ]
all hold. This, together with Schnorr and Stimm’s dichotomy result, implies
that dLZ(b) succeeds exponentially on every non-normal sequence in Σ∞
b . Hence
a real number x is absolutely normal if none of the martingales dLZ(b) succeed
exponentially on the base-b expansion of x.
3
Savings Accounts
In this section we construct a conservative version of the the Lempel-Ziv martingale dLZ(b) consisting of a new martingale d′ that can be smaller than dLZ(b)
but that has a savings account in the following sense.
Definition. A nondecreasing function g : Σ∗b → [0, ∞) is a savings account of
a martingale d : Σ∗b → [0, ∞) if, for every w ∈ Σ∗b , d(w) ≥ g(w).
7
Construction 3.1 Let d = dLZ(b) be the base-b-Lempel-Ziv martingale. We
define a new martingale d′ = e′ + g ′ as follows.
We first define e′ . e′ (λ) = b, goal(λ) = −1, taken(λ) = −1. For w ∈ Σ∗b ,
let w = x(1) . . . x(j)u, for z = x(1) . . . x(j) a full parse and u a partial or full
phrase of w. Let
goal(w)
= |w| − ⌈D(z)(logb (D(z)))/(b − 1)⌉ + ⌊D(z)(logb (e))/(b − 1)⌋
1
−⌈logb (D(z))⌉ − ⌈logb e b−1 ⌉ − 1,
taken(w) =
taken(z) if goal(w) ≤ taken(z)
goal(w) if goal(w) > taken(z)},
e′ (w) = e′ (z) ·
d(w) −taken(w)+taken(z)
b
.
d(z)
Let g ′ be defined as follows. g ′ (λ) = 1 and for y ∈ Σ∗b , a ∈ Σb , if ya = w, let
w = x(1) . . . x(j)u, for z = x(1) . . . x(j) a full parse and u a partial phrase of w.
′
g (y)
if goal(w) ≤ taken(z)
g ′ (ya) =
if
goal(w) > taken(z)
g ′ (y) + e′ (y) b−1
b
Theorem 3.2 Let d′ and g ′ be as defined in Construction 3.1. Then d′ is a
martingale and g ′ is its savings account,
S exp [dLZ(b) ] ⊆ S ∞ [g ′ ],
d′ is bounded by a polynomial, and d′ is computable in a nearly linear time bound
that does not depend on b.
Proof of Theorem 3.2.
Claim 3.3 d′ = e′ + g ′ is a martingale and g ′ is a savings account for d′ .
Proof. Let us prove that d′ is a martingale. When goal(w) ≤ taken(z) we have
that
′
1 X ′
(z) 1 P
e (wa) = ed(z)
a∈Σb d(wa) =
b
b
a∈Σb
e′ (z)
d(z)
· d(w) = e′ (w)
with that last equality holding both when w = z and when z is a proper substring
of w. Since g ′ (wa) is constant the martingale equality holds in this case.
In the second case, when goal(w) > taken(z), we have that
e′ (ya) = e′ (y)d(ya)/d(y)1/b, so
X
P
P
d′ (wa)
= a∈Σb e′ (ya) + a∈Σb g ′ (ya)
a∈Σb
′
′
′
= e′ (y) + b(g ′ (y) + e′ (y) b−1
b ) = b(e (y) + g (y)) = bd (w).
8
Since e′ is nonnegative, by definition g ′ is a nondecreasing function. Therefore g ′ is a savings account of d′ .
Claim 3.4 There is a C > 0 such that for every w ∈ Σ∗b ,
d(w)b−taken(w) ≤ C · D(z) · L(u).
d(w)b−goal(w) ≥ b
Proof. Use that taken(w) ≥ goal(w), Lemma 2.2, and Observation 2.1.
Claim 3.5 If w ∈ Σ∗b then e′ (w) = d(w)b−taken(w) .
Proof. By induction on |w|, let w = x(1) . . . x(j)u, z = x(1) . . . x(j)x(j + 1)
e′ (w)
d(w) −taken(w)+taken(z)
b
d(z)
d(w) −taken(w)+taken(z)
b
= d(w)b−taken(w) .
d(z)b−taken(z) ·
d(z)
e′ (z) ·
=
=
∞ ′
Claim 3.6 If y ∈ Σ∞
b and goal(y ↾ n) is unbounded then y ∈ S [g ].
Proof.
If goal(y ↾ n) is unbounded then infinitely often we use the second case
in the definitions of taken and g ′ and have that taken(y ↾ n) = goal(y ↾ n),
e′ (y ↾ n) = d(y ↾ n)b−goal(y↾n) , and g ′ (y ↾ na) = g ′ (y ↾ n) + e′ (y ↾ n) b−1
b .
For those n, by Claim 3.4, e′ (y ↾ n) ≥ b, therefore g ′ (y ↾ na) ≥ g ′ (y ↾
n) + b − 1.
Since g ′ is monotonic, y ∈ S ∞ [g ′ ].
(1−α)n
Claim 3.7 S b
[dLZ(b) ] ⊆ S ∞ [g ′ ].
(1−α)n
Proof. If y ∈ S b
[dLZ(b) ] then by Lemma 2.3 for infinitely many n, D(y ↾
n) logb (D(y ↾ n)) < α(b − 1)n.
Notice that therefore goal(y ↾ n) is unbounded and by Claim 3.6 y ∈ S ∞ [g ′ ].
Claim 3.8 There is a C > 0 such that for everyP
w, e′ (w) ≤ C · D(z) · L(u) (for
′
u a partial or the last full phrase of w). g (w) ≤ v⊑w e′ (v)(b − 1)/b. Therefore
there is a c > 0 such that for every w, d′ (w) ≤ |w|c .
Claim 3.9 d′ can be computed in nearly linear time (in time n logc n for c not
depending on b.
9
Proof.
Again, for w ∈ Σ∗b , let w = x(1) . . . x(j)u, for z = x(1) . . . x(j) a full parse
and u a partial or full phrase of w. If goal(w) > taken(z), let t be such that
t ⊑ u and goal(zt) = taken(z).
Notice that
′
g (w) =
g ′ (z) P
if goal(w) ≤ taken(z)
g ′ (z) + v⊑u,|zv|>|zt| e′ (zt) b−1
L(v)
if goal(w) > taken(z),
b
where L(v) isPthe number of leaves below v in T (w). Given precomputed values
′
for f (u) =
v⊑u L(v), the value of g (w) can be easily computed in nearly
linear time. The precomputed value of f (u) will be stored in the u node of the
LZ tree for z.
This completes the proof of Theorem 3.2.
4
Base change
We use infinite sequences over Σb to represent real numbers in [0,1). For this,
we associate each string w ∈ Σ∗b with the half-open interval [w]b defined by
|w|
P
[w]b = [x, x + b−|w| ), for x =
w[i − 1]b−i . Each real number α ∈ [0, 1) is then
i=1
represented by the unique sequence seqb (α) ∈ Σ∞
b satisfying
w ⊑ seqb (α) ⇐⇒ α ∈ [w]b
for all w ∈ Σ∗b . We have
α=
∞
X
i=1
seqb (α)[i − 1]b−i
and the mapping seqb : [0, 1) → Σ∞
b is a bijection. (Notice that [w]b being
half-open prevents double representations.) We define realb : Σ∞
b → [0, ∞) to
be the inverse of seqb . A set of real numbers A ⊆ [0, 1) is represented by the set
seqb (A) = {seqb (α) | α ∈ A}
of sequences. If X ⊆ Σ∞
b then
realb (X) = {realb (x) | x ∈ X}.
For a positive number x we use the notation logb (x) = ⌈logb (x)⌉.
Construction 4.1 Let d : Σ∗b → [0, ∞) be a polynomially-bounded martingale
with a savings account g.
We define γ : Σ∗b → [0, 1] a probability measure γ(w) := b−|w| d(w).
10
Using the Carathéodory extension to Borel sets, γ can be extended to any
interval [a, c]; we denote with γ̂ this extension. (In fact if we consider all U ⊆ Σ∗b
such that all u, v ∈ U,
Pu 6= v are incomparable and [u]b ⊆ [a, c] for all u ∈ U ,
then γ̂([a, c]) = supU u∈U γ(u)).
We define µ : {0, 1}∗ → [0, 1] by µ(y) = γ̂([y]2 ).
Finally we define d(2) : {0, 1}∗ → [0, ∞) by d(2) (y) = 2|y| µ(y).
Theorem 4.2 Assume that d is a base-b martingale that is polynomially bounded
and g is a savings account of d, and let d(2) be defined from d and g as in Construction 4.1. Then
realb (S ∞ [g]) ⊆ real2 (S ∞ [d(2) ]).
Moreover, if d is computable in a nearly linear time bound not depending on b,
then so is d(2) .
Proof of Theorem 4.2.
Property 4.3 Let α ∈ [0, 1). If seqb (α) ∈ S ∞ [g], then seq2 (α) ∈ S ∞ [d(2) ].
Proof. Let x = seqb (α) ∈ S ∞ [g]. Let y = seq2 (α).
We use here that d has a savings account g, so if g(x ↾ n) > m then for all
w with x ↾ n ⊑ w, g(w) > m.
Let m ∈ N and choose n such that g(x ↾ n) > m. Let q be such that
[y ↾ q]2 ⊆ [x ↾ n]b . Let us see that d(2) (y ↾ q) > m.
Let r ∈ N. Let Aqr = {w ∈ Σ∗b | |w| = r and [w]b ⊆ [y ↾ q]2 }. Then
d(2) (y ↾ q) =
≥
2q γ̂([y ↾ q]2 ) = 2q lim
r
2q m lim
r
X
X
d(w)b−|w|
w∈Aqr
b−|w| = 2q m2−q = m.
w∈Aqr
The last chain of equations holds because [y ↾ q]2 ⊆ [x ↾ n]b and for every
w ∈ Aqr , [w]b ⊆ [y ↾ q]2 , so x ↾ n ⊑ w for any w ∈ Aqr .
We next compute d(2) . For each m ∈ N we define µm : {0, 1}∗ → [0, 1] by
X
γ(w).
µm (y) =
|w|=m,[w]b∩[y]2 6=∅
Claim 4.4 For every y ∈ {0, 1}∗ and m ∈ N, |µ(y) − µm (y)| ≤ 2b−m mc .
Proof. Let c be such that d(w) ≤ |w|c for every w (using that d is polynomially
bounded). Then since at most two strings w with |w| = m have the property
that [w]b ∩ [y]2 6= ∅ and [w]b 6⊆ [y]2 , we have
|µ(y) − µm (y)| ≤ 2b−m mc .
(2)
(2)
For each m ∈ N we define dm : {0, 1}∗ → [0, ∞) by dm (y) = 2|y| µm (y).
11
Claim 4.5 For some c > 0, for every y ∈ {0, 1}∗, for every m ∈ N,
|y| −m log b+c log m
|d(2) (y) − d(2)
.
m (y)| ≤ 2 2
Corollary 4.6 For some c′ > 0, for every y ∈ {0, 1}∗,
(2)
|d(2) (y) − d|y|/ log b+c′ log |y| (y)| ≤ 1/|y|3 .
Proof. Take c′ = (4 + c)/ log b and use the previous claim.
Notice that the approximation of d(2) is slow, but it is enough for the diagonalization performed in the next section.
(2)
Property 4.7 For m ∈ N, y ∈ {0, 1}∗, dm (y) can be computed by considering
a maximum of 2b neighbor strings w ∈ Σrb for r = |y|/ log b to m, computing
d(w) for each of them and doing an addition and a multiplication for each.
Proof. Consider P , the smallest prefix free set of strings w ∈ Σ∗b such that
[w]b ⊆ [y]2 , and notice that |w| ≥ |y|/ log b for each such string. For each r
there are at most 2b − 2 strings of length r in P (otherwise we can replace some
of them by a single string of length r − 1). For length m we may need two more
strings |w| = m, [w]b ∩ [y]2 6= ∅.
(2)
Corollary 4.8 For y ∈ {0, 1}∗, d|y|/ log b+c′ log |y| (y) can be computed by considering a maximum of 2b neighbor strings w ∈ Σrb for r = |y|/ log b to |y|/ log b +
c′ log |y|, computing d(w) for each of them and doing an addition and a multiplication for each.
(2)
By Corollary 4.6 f (y) = d|y|/ log b+c′ log |y| (y) approximates d(2) (y) within a
1/|y|3 bound, and by the last corollary f can be computed in nearly linear time.
In order to have a nearly linear time bound independent of b consider only y for
which log|y| ≥ 2b.
This concludes the proof of Theorem 4.2.
5
Absolutely Normal Numbers
In this section we give an algorithm that diagonalizes against the Lempel-Ziv
martingales for all bases in nearly linear time.
We use the following theorem
Theorem 5.1 Let (dk )k∈N be a sequence of base-2 martingales such that for
ck : {0, 1}∗ → [0, ∞) with the following two
each of them there exists a function d
properties
1. For every y ∈ {0, 1}∗
ck (y)| ≤ 1 .
|dk (y) − d
|y|3
12
ck is computable in a nearly linear time bound that does not depend on k.
2. d
Then we can compute in nearly linear time a binary sequence x such that, for
every k, x 6∈ S ∞ [dk ].
Proof. Without loss of generality we assume that dk (λ) = 1 for all k.
Let d : {0, 1}∗ → [0, ∞) be defined by
log2 |w|
X
d(w) =
√
2−k−2
k
k=1
ck (w).
d
Our algorithm will diagonalize against d, constructing a binary sequence x
as follows. If x ↾ n has been defined then choose the next bit of x as c ∈ {0, 1}
that minimizes d((x ↾ n)c).
Claim 5.2 If x 6∈ S ∞ [d] then for every k, x 6∈ S ∞ [dk ].
Let n ∈ N, k ∈ N, n ≥ 2
√
k
.
dk (x ↾ n) ≤ c
dk (x ↾ n) + 1/n3 ≤
√
≤ 2k+2
k
d(x ↾ n) + 1/n3
Claim 5.3 x 6∈ S ∞ [d].
Let w ∈ {0, 1}∗. We prove that there exists c ∈ {0, 1} such that d(wc) ≤
d(w) + 1/|w|2 .
log2 (|w|+1)
d(wc)
X
=
√
k
ck (wc)
d
√
k
(dk (wc) + 1/(|w| + 1)3 )
√
k
(dk (w) + 1/(|w| + 1)3 )
2−k−2
k=1
log2 (|w|+1)
X
≤
2−k−2
k=1
log2 (|w|+1)
X
≤
2−k−2
k=1
log2 (|w|)
X
≤
2
√
−k−2
k=1
log2 (|w|+1)
+
X
log2 (|w|+1)
k
ck (w) + 1/|w| ) +
(d
3
√
2−k−2
k
X
√
2−k−2
k
1/(|w| + 1)3 )
k=1
(dk (w))
k=log2 (|w|)+1
∞
X
≤
d(w) +
≤
d(w) + 1/|w|2 .
2
√
−k−2
log2 (|w|+1)
k
3
2/(|w|) ) +
X
k=log2 (|w|)+1
k=1
13
√
k
2−k−2
2|w|
Therefore
and x 6∈ S ∞ [d].
6
d(x ↾ n) ≤ d(x ↾ (n − 1)) + 1/n2 ≤ C
Open Problem
Many questions arise naturally from this work, but the following problem appears to be especially likely to demand new and useful methods.
As we have seen, normal numbers are closely connected to the theory of
finite automata. Schnorr and Stimm [24] proved that normality is exactly the
finite-state case of randomness. That is, a real number α is normal in a base
b ≥ 2 if and only if no finite-state automaton can make unbounded money
betting on the successive digits of the base-b expansion of α with fair payoffs.
The theory of finite-state dimension [9], which constrains Hausdorff dimension
[15] to finite-state automata, assigns each real number α a finite-state dimension
(b)
dimFS (α) ∈ [0, 1] in each base b. A real number α then turns out to be normal
(b)
in base b if and only if dimFS (α) = 1 [6]. Do there exist absolutely dimensioned
(b)
numbers, i.e., real numbers α for which dimFS (α) = dimFS (α) does not depend
on b, and 0 < dimFS (α) < 1?
Acknowledgments
The first author’s research was supported in part by National Science Foundation Grants 0652569, 1143830, 1247051, and 1545028. Part of this author’s
work was done during a sabbatical at Caltech and the Isaac Newton Institute
for Mathematical Sciences at the University of Cambridge, part was done during three weeks at Heidelberg University, with support from the Mathematics
Center Heidelberg and the Heidelberg University Institute of Computer Science,
and part was done during the workshop “Normal Numbers: Arithmetical, Computational and Probabilistic Aspects” at the Erwin Schrödinger International
Institute for Mathematics and Physics at the University of Vienna.
The second author’s research was supported in part by Spanish Government MEC Grants TIN2011-27479-C04-01 and TIN2016-80347-R. Part of this
author’s work was done during a research stay at the Isaac Newton Institute
for Mathematical Sciences at the University of Cambridge, part was done during three weeks at Heidelberg University, with support from the Mathematics
Center Heidelberg and the Heidelberg University Institute of Computer Science,
and part was done during the workshop “Normal Numbers: Arithmetical, Computational and Probabilistic Aspects” at the Erwin Schrödinger International
Institute for Mathematics and Physics at the University of Vienna.
14
References
[1] V. Becher, S. Figueira, and R. Picchi. Turing’s unpublished algorithm for
normal numbers. Theoretical Computer Science, 377:126–138, 2007.
[2] V. Becher, P. A. Heiber, and T. A. Slaman. A polynomial-time algorithm
for computing absolutely normal numbers. Information and Computation,
232:1–9, 2013.
[3] P. Billingsley. Probability and Measure, third edition. John Wiley and Sons,
New York, N.Y., 1995.
[4] E. Borel.
Sur les probabilités dénombrables et leurs applications
arithmétiques. Rendiconti del Circolo Matematico di Palermo, 27(1):247–
271, 1909.
[5] J. Borwein and D. Bailey. Mathematics by Experiment: Plausible Reasoning
in the 21st Century, second edition. A. K. Peters, 2008.
[6] C. Bourke, J. M. Hitchcock, and N. V. Vinodchandran. Entropy rates
and finite-state dimension. Theoretical Computer Science, 349(3):392–406,
2005.
[7] Y. Bugeaud. Distribution modulo one and Diophantine approximation, volume 193. Cambridge University Press, 2012.
[8] J. W. S. Cassels. On a problem of Steinhaus about normal numbers. Colloquium Mathematicum, 7:95–101, 1959.
[9] J. J. Dai, J. I. Lathrop, J. H. Lutz, and E. Mayordomo. Finite-state dimension. Theoretical Computer Science, 310:1–33, 2004.
[10] J. Doob. Regularity properties of certain families of chance variables.
Transactions of the American Mathematical Society, 47:455–486, 1940.
[11] M. Feder. Gambling using a finite state machine. IEEE Transactions on
Information Theory, 37:1459–1461, 1991.
[12] S. Figueira and A. Nies. Feasible analysis and randomness. Manuscript,
2013.
[13] S. Figueira and A. Nies. Feasible analysis, randomness, and base invariance.
Theory of Computing Systems, 56:439–464, 2015.
[14] Y. Gurevich and S. Shelah. Nearly linear time. In Proceedings of the First
Symposium on Logical Foundations of Computer Science. Springer, 1989.
[15] F. Hausdorff. Dimension und äußeres Maß. Math. Ann., 79:157–179, 1919.
[16] J. Hitchcock and J. Lutz. Why computational complexity requires stricter
martingales. Theory of Computing Systems, 39:277–296, 2006.
15
[17] A. N. Kolmogorov. Foundations of the Theory of Probability. Chelsea,
1950.
[18] H. Lebesgue. Sur certaines demonstrations d’existence. Bull. Soc. Math.
de France, 45:132–144, 1917.
[19] A. Lempel and J. Ziv. Compression of individual sequences via variable
rate coding. IEEE Transaction on Information Theory, 24:530–536, 1978.
[20] P. Lévy. Propriétés asymptotiques des sommes de variables indépendantes
ou enchainées. Journal des mathématiques pures et appliquées. Series 9.,
14(4):347–402, 1935.
[21] P. Lévy. Théorie de l’Addition des Variables Aleatoires. Gauthier-Villars,
1937 (second edition 1954).
[22] E. Mayordomo. Construction of an absolutely normal real number in polynomial time. Manuscript, 2013.
[23] W. M. Schmidt. On normal numbers. Pacific J. Math, 10(2):661–672, 1960.
[24] C.-P. Schnorr and H. Stimm. Endliche automaten und zufallsfolgen. Acta
Informatica, 1(4):345–359, 1972.
[25] W. Sierpinski. Démonstration élémentaire du théorème de M. borel sur les
nombres absolument normaux et détermination effective d’une tel nombre.
Bull. Soc. Math. France, 45:125–132, 1917.
[26] A. M. Turing. On computable numbers with an application to the Entscheidungsproblem. Proceedings of the London Mathematical Society, 42:230–
265, 1936-1937.
[27] A. M. Turing. A note on normal numbers. In S. B. Cooper and J. van
Leeuwen, editors, Alan Turing: His Work and Impact. Elsevier Science,
2013.
[28] J. Ville. Étude Critique de la Notion de Collectif. Gauthier–Villars, Paris,
1939.
[29] S. Wagon. Is π normal? Math. Intelligencer, 7:65–67, 1985.
16
| 8cs.DS
|
Efficient Nearest-Neighbor Search for Dynamical
Systems with Nonholonomic Constraints
Valerio Varricchio, Brian Paden, Dmitry Yershov, and Emilio Frazzoli
arXiv:1709.07610v1 [cs.CG] 22 Sep 2017
Massachusetts Institute of Technology
Abstract. Nearest-neighbor search dominates the asymptotic complexity of sampling-based motion planning algorithms and is often addressed
with k-d tree data structures. While it is generally believed that the expected complexity of nearest-neighbor queries is O(log(N )) in the size of
the tree, this paper reveals that when a classic k-d tree approach is used
with sub-Riemannian metrics, the expected query complexity is in fact
Θ(N p log(N )) for a number p ∈ [0, 1) determined by the degree of nonholonomy of the system. These metrics arise naturally in nonholonomic
mechanical systems, including classic wheeled robot models. To address
this negative result, we propose novel k-d tree build and query strategies
tailored to sub-Riemannian metrics and demonstrate significant improvements in the running time of nearest-neighbor search queries.
1
Introduction
Sampling-Based algorithms such as Probabilistic Roadmaps (PRM) [9], Rapidly
exploring Random Trees (RRT) [11] and their asymptotically optimal variants
(PRM∗ , RRT∗ ) [7] are widely used in motion planning. These algorithms build a
random graph of motions between points on the robot’s configuration manifold.
During the graph expansion, nearest-neighbor search is used to limit the
computation to regions of the graph close to the new configurations and it is
shown to dominate the asymptotic complexity of randomized planners. The notion of closeness appropriate for motion planning is induced by the length of the
shortest paths between configurations, or in general, by the minimum cost of
controlling a system between states.
As compared to exhaustive linear search, efficient algorithms with reduced
complexity have been studied in computational geometry and their use in motion
planning has been highlighted as a factor of dramatic performance improvement
[12]. Among a variety of approaches, k-d trees [1] are ideal due to their remarkable efficiency in low-dimensional spaces, typical of motion planning problems.
Classic k-d trees are shown to have logarithmic average case complexity for
distance functions strongly equivalent to L-p metrics [4]. While this requirement
is reasonable for many applications, it does not apply to distances induced by
the shortest paths of nonholonomic systems. Therefore, identifying nearest neighbors in the sense of a generic control cost remains an important open problem
in sampling-based motion planning. In both literature and practical implementations, when searching for neighbors, randomized planners resort to distance
functions that only approximate the true control cost. Arguably, the most common choices are Euclidean distance or quadratic forms [12]. This ad-hoc approach
1
2. GEOMETRY OF NONHOLONOMIC SYSTEMS
can significantly slow down the convergence rate of sampling-based algorithms
if an inappropriate metric is selected.
A number of heuristics have been proposed to resolve this issue, such as
the reachability and utility guided RRTs [17,3] which bias the tree expansion
towards promising regions of the configuration space. Other approaches use the
cost of linear quadratic regulators [5] and learning techniques [15,2]. In specific
examples, these heuristics can significantly reduce the negative effects of finding
nearest neighbors according to a metric inconsistent with the minimum cost path
between configurations. However, the underlying inconsistency is not directly
addressed. In contrast, a strong motivation to address it comes from recent
research [8], which shows that a major speedup of sampling-based kynodinamic
planners can be achieved by considering nonholonomy at the stage of nearestneighbor search.
Specialized k-d tree algorithms have been proposed to account for nonstandard topologies of some configuration manifolds [10,19,6].
However, no effort is known towards generalizing such algorithms to differential constraints.
In this work, we investigate the use of k-d trees for exact nearest-neighbor
search in the presence of differential constraints. The main contributions can be
summarized as follows: (i) we derive the expected complexity of nearest-neighbor
queries with k-d trees built according to classic techniques and reveal that it is
super-logarithmic (ii) we propose novel k-d tree build and query procedures tailored to sub-Riemannian metrics (iii) we provide numerical trials which verify
our theoretical analysis and demonstrate the improvement afforded by the proposed algorithms as compared with popular open source software libraries, such
as FLANN [14] and OMPL [18].
In Section 2, we review background material on sub-Riemannian geometries
and show connections with nonholonomic systems, providing asymptotic bounds
to their reachable sets. Based on these bounds, in Section 3 we propose a query
procedure specialized for nonholonomic systems, after a brief review of the k-d
tree algorithm. In Section 4, we study the expected complexity of m-nearestneighbor queries on a classic k-d tree with sub-Riemannian metrics. Inspired
by this analysis, in Section 5 we propose a novel incremental build procedure.
Finally, in Section 6 we show positive experimental results for a nonholonomic
mobile robot, which confirm our theoretical predictions and the effectiveness of
the proposed algorithms.
2
Geometry of Nonholonomic Systems
Nonholonomic constraints are frequently encountered in robotics and describe
mechanical systems whose local mobility is, in some sense, limited. Basic concepts from differential geometry, reviewed below, are used to clarify these limitations and discuss them quantitatively.
2
2. GEOMETRY OF NONHOLONOMIC SYSTEMS
2.1
Elements of Differential Geometry
A subset M of Rn is a smooth k-dimensional manifold if for all p ∈ M there
exists a neighborhood V of M such that V ∩ M is diffeomorphic to an open
subset of Rk . A vector v ∈ Rn is said to be tangent to M at point p ∈ M if
there exists a smooth curve γ : [0, 1] → M, such that γ̇(0) = v and γ(0) = p.
The tangent space of M at p, denoted Tp M, is the subspace of vectors tangent
to M at p. A map Y : M → Rn is a vector field on M if for each p ∈ M,
Y (p) ∈ Tp M. A smooth euclidean vector bundle of rank r over M is defined as
a set E ⊂ M × Rl such that the set Ep : v ∈ Rl , (p, v) ∈ E is an r-dimensional
vector space for all p ∈ M. Ep is called the fiber of bundle E at point p. Any set
of h linearly independent smooth vector fields Y1 , . . . Yh such that for all p ∈ M,
span(Y1 (p), ..., Yh (p)) = Ep is called a basis of E and Yi are called generator
vector fields of E. The tangent bundle of M is defined as the vector bundle
T M whose fiber at each point p is the tangent space at that point, Tp M. A
distribution H on a manifold is a subbundle of the tangent bundle, i.e., a vector
bundle such that its fiber Hp at all points is a vector subspace of Tp M.
Connection with nonholonomic systems. Consider a nonholonomic system described by the differential constraint:
ẋ(t) = f (x(t), u(t)),
(1)
where the configuration x(t) and control u(t) belong to the smooth manifolds
X and U respectively. Each ui ∈ U defines a vector field gi (z) = f (z, ui ) on
X . Therefore, for a fixed configuration z, f (z, u) has values in a vector space
Hz := Span({gi (z)}). In other words, the dynamics described by equation (1)
define a distribution H on the configuration manifold.
Throughout the paper, we will make use of the Reeds-Shepp vehicle [16] as
an illustrative example.
Example (Configuration manifold of a Reeds-Shepp vehicle). The configuration
manifold for this vehicle model is X = SE(2) with coordinates x = (x1 , x2 , x3 ).
The mobility of the system is given by ẋ1 = u1 cos(x3 ), ẋ2 = u1 sin(x3 ), ẋ3 =
u1 u2 , where u1 , u2 ∈ [−1, 1]. The inputs u1 = (1, 0) and u2 = (1, 1) define the
vector fields g1 (x) = (cos(x3 ), sin(x3 ), 0) and g2 (x) = (cos(x3 ), sin(x3 ), 1). At
each z ∈ SE(2) the fiber of the Reeds-Shepp distribution HRS is Span{g1 (z),
g2 (z)}. Let fˆ(x) = (cos(x3 ), sin(x3 ), 0), ˆl(x) = (− sin(x3 ), cos(x3 ), 0) and θ̂(x) =
(0, 0, 1). These vector fields indicate the body frame of the vehicle, i.e., its front fˆ,
lateral ˆl and rotation θ̂ axes. The Reeds-Shepp distribution is then equivalently
defined by the fibers HzRS = Span{fˆ(z), θ̂(z)}.
2.2
Distances in a Sub-Riemannian Geometry
A sub-Riemannian geometry G on a manifold M is a tuple G = (M, H, h·, ·iH ),
where H is a distribution on M whose fibers at all points p are equipped with
the inner product h·, ·iH : Hp × Hp → R. The distribution H is referred to as the
3
2. GEOMETRY OF NONHOLONOMIC SYSTEMS
horizontal distribution. A smooth curve γ : [0, 1] → M is said to be horizontal
if γ̇(t) ∈ Hγ(t) for all t ∈ [0, 1].
R1p
The length of smooth curves is defined `G (γ) := 0 hγ̇(t), γ̇(t)iH dt. If
Γab denotes the set of horizontal curves between a, b ∈ M, then dG (a, b) :=
inf γ∈Γab `G (γ) is a sub-Riemannian metric defined by the geometry. The ball
centered at p of radius r with respect to the metric dG is denoted B(p, r).
In control theory, the attainable set A(x0 , t) is the subset of X such that
for each p ∈ A(x0 , t) there exists a u : [0, τ ] → U for which the solution to (1)
through x(0) = x0 satisfies x(τ ) = p for some τ ≤ t. Solutions to (1) are simply
horizontal curves, so the attainable set A(x0 , t) is equivalent to the ball B(x0 , t)
defined by the sub-Riemannian metric. This equivalence holds for systems with
time-reversal symmetry, under the controllability condition stated in Theorem 1.
Example (Sub-Riemannian geometry of a Reeds-Shepp vehicle). The geometry
associated with the Reeds-Shepp vehicle is G RS = (SE(2), HRS , h·, ·iRS ), with
the standard inner product hv, wiRS = v T w. Horizontal curves for this geometry
are feasible paths satisfying the differential constraints. Geodesics correspond
to minimum-time paths between two configurations and are known in closed
form [16].
2.3
Iterated Lie Brackets and the Ball-box Theorem
A system is said to be controllable if any pair of configurations can be connected
by a feasible (horizontal) path. Determining controllability of a nonholonomic
system is nontrivial. For example, the Reeds-Shepp vehicle cannot move directly
in the lateral direction, but intuition suggests that an appropriate sequence of
motions can result in a lateral displacement (e.g., in parallel parking). Chow’s
Theorem and the Ball-box Theorem, reviewed below, are fundamental tools related to the controllability and the reachable sets of nonholonomic systems.
The Lie derivative of a vector field Y at p ∈ M in the direction v ∈ Tp M is
d
defined as dY (p)v = dt
Y (γ(t)) t=0 , where γ is a smooth curve starting in p =
γ(0) with velocity v = γ̇(0). Given two vector fields Y1 , Y2 on M, the Lie bracket
[Y1 , Y2 ] is a vector field on M defined as [Y1 , Y2 ](p) = dY2 (p)Y1 (p)−dY1 (p)Y2 (p).
From the horizontal distribution H, one can construct a sequence of distributions by iterating the Lie brackets of the generating vector fields, Y1 , Y2 . . . Yh ,
with h ≤ k. Recursively, this sequence is defined as: H1 = H, Hi+1 = Hi ∪
[H, Hi ], where [H, Hi ] denotes the distribution given by the Lie brackets of each
generating vector field of H with those of Hi . Note that the Lie bracket of two vector fields can be linearly independent from the original fields, hence Hi ⊆ Hi+1 .
The Lie hull [13], denoted Lie(H), is the limit of the sequence Hi as i → ∞. A
distribution H is said to be bracket-generating if Lie(H) = T M.
Theorem 1. (Chow’s Theorem [13, p. 44] ). If Lie(H) = T M on a connected manifold M, then any a, b ∈ M can be joined by a horizontal curve.
4
2. GEOMETRY OF NONHOLONOMIC SYSTEMS
Example (Lie hull of a Reeds-Shepp vehicle). Consider the Lie hull of HRS generated by {fˆ(x), θ̂(x)}. For every x ∈ SE(2), these vectors span a two-dimensional
subspace of Tx (SE(2)). The first order Lie bracket is given by
d ˆ
d
[fˆ, θ̂](x) =
f (x) θ̂(x) −
θ̂(x) fˆ(x) = (− sin(x3 ), cos(x3 ), 0) . (2)
dx
dx
This coincides with the body frame lateral axis ˆl(x) of the vehicle. Therefore,
the second order distribution Hx2 = Span{fˆ(x), θ̂(x), [fˆ, θ̂](x)} spans the tangent bundle of SE(2), and thus the Lie hull is obtained in the second step. By
Theorem 1, there is a feasible motion connecting any two configurations which
is consistent with one’s intuition about the wheeled robot.
Above, we have shown that from a basis Y1 . . . Yh of a bracket-generating
distribution H, one can define a basis y1 , . . . yk of T M. Specifically yi = Yi
for i ≤ h, while the remaining d − h fields are obtained with Lie brackets.
The vector fields {yi } are called privileged directions [13]. Define the weight wi
of the privileged direction yi as the smallest order of Lie brackets required to
generate yi from the original basis. More formally, wi is such that yi ∈
/ Hwi −1
wi
and yi ∈ H for all i. The weighted box at p of size , weights w ∈ Nk>0 and
multipliers µ ∈ Rk>0 is defined as:
Boxw,µ (p, ) : {y ∈ Rn : |hy − p, yi (p)iRn | < µi wi for all i ∈ [1, k]}.
(3)
Theorem 2 (The ball-box Theorem). Let H be a distribution on a manifold
M satisfying the assumptions of Chow’s Theorem. Then, there exist constants
0 ∈ R>0 and c, C ∈ Rk>0 such that for all < 0 and for all p ∈ M:
w,c
(p, ) ⊂ B(p, ) ⊂ Boxw,C (p, ) := Boxw
Boxw
o (p, ),
i (p, ) := Box
(4)
w
where Boxw
i and Boxo are referred to as the inner and outer bounding boxes for
the sub-Riemannian ball B, respectively.
Example (Ball-Box Theorem visualized for a Reeds-Shepp vehicle). From the Lie
hull construction in the previous example, we know that fˆ(x), θ̂(x) ∈ H1 so the
corresponding weights are wfˆ = wθ̂ = 1. Conversely, ˆl(x) first appears in H2 , so
its weight is wl̂ = 2. Theorem 2 states the existence of inner and outer bounding
boxes for reachable sets as t → 0 and predicts the infinitesimal order of each
side of these boxes, as shown in figure 1. Higher order Lie brackets correspond
to sides that approach zero at a faster asymptotic rate. The longitudinal and
angular sides — along fˆ and θ̂ — scale with Θ(t), while the lateral one — along
ˆl — with Θ(t2 ). Therefore, both boxes become increasingly elongated along fˆ
and θ̂ and flattened along ˆl as t 0. Intuitively, this geometric feature of boxes
reflects the well known fact that a small lateral displacement of a car requires
more time than an equivalent longitudinal one.
5
3. THE K-D TREE DATA STRUCTURE
2 Cl̂ t2
2 Cfˆ t
2 cfˆ t
1
2 Cθ̂ t
0.6
-0.2
-0.2
fˆ(q )
q
0
-0.5
-1
0
0.2
0.5
t = 0.7
-1
-1
0
q
-0.4
2
1
0
-0.5
-0.6
θ̂(q )
Boxw
i ( q , t)
B (q , t)
2
0.5
0.2
l̂ (q )
Boxw
o (q , t)
2 cl̂ t2
0
0.4
1
0.5
0
-0.5
t=1
0
-2
-2
-1
0
1
2
1
0
-1
-2
t=2
-2
-2
0
0
2
2
t=3
Fig. 1. Reachable sets and bounding boxes for a Reeds-Shepp vehicle around a configuration q and for different values of t. The lengths of the box sides are highlighted.
For both the inner and outer boxes, as t → 0, the sides along the front fˆ and heading
θ̂ axes are linear in t, while the side along the lateral axis ˆ
l is quadratic in t.
For the Reeds-Shepp vehicle, the sides of both boxes can be computed explicitly with geometric considerations on special manoeuvers that maximize
or minimize the displacement along eachp
axis. In particular, we get the values
Cfˆ = Cθ̂ = cθ̂ = 1, Cl̂ = 1/2, cfˆ = 3/2 − 1, cl̂ = 1/8.
3
The k-d tree Data Structure
A k-d tree T is a binary tree organizing a finite subset X ⊂ M, called a database,
with its elements xi ∈ X called data points. We would like to find the m points in
X closest to a given query point q on the manifold. Each point xi ∈ X is put in
relation with a normal vector ni ∈ Rn . Together, the pair vi = (xi , ni ) defines a
vertex of the binary tree. The set of vertices is denoted with V. A vertex defines
a partition of Rn into two halfspaces, referred to as the positive and negative
halfspace, and respectively described algebraically:
h+ (xi , ni ) := {x ∈ Rn : hni , xiRn > hni , xi iRn } ,
h− (xi , ni ) := {x ∈ Rn : hni , xiRn ≤ hni , xi iRn } .
(5)
An edge is an ordered pair of vertices e = (vi , vj ). A binary tree is defined as
T := (V, E − , E + ), with V set of vertices, E − set of left edges and E + set of right
edges. Given one edge e = (vi , vj ), vertex vj is referred to as the left child of vi if
e ∈ E − , or the right child if e ∈ E + . Let parent(vi ∈ V) = vj ∈ V s.t. (vi , vj ) ∈
E − ∪ E + . By convention, parent(vi ) = ∅ if such vj does not exist and vi is called
the root of T , denoted vi = root(T ). Let child(vi ∈ V, s ∈ {−, +}) = vj ∈ V
s.t. (vi , vj ) ∈ E s , or otherwise ∅ if such vj does not exist.
The fundamental property of k-d trees is that left children belong to the negative halfspace defined by their parents and right children to the positive halfspace.
Recursively, a vertex belongs to the parent halfspaces of all its ancestors. As a
result, a k-d tree defines a partition of M into non-overlapping polyhedra, called
buckets [4], that cover the entire manifold. In the sequel, we let BT denote the
set of buckets for a given k-d tree T .
6
3. THE K-D TREE DATA STRUCTURE
Buckets are associated with the leaves of T and with parents of only-child
leaves (i.e., leaves without a sibling). For any given point q ∈ M, we denote with
bq the unique bucket containing q.
3.1
m-nearest-neighbor query algorithm
The computational efficiency afforded by the algorithm comes from the following
observation: let q be the query point and suppose that among the distance evaluations computed thus far, the m closest points to q are within a distance dm
away. Now consider a vertex (x, n) and suppose the query point q is contained in
h+ (x, n) and B(q, dm ) ⊂ h+ (x, n). The data points represented by the sibling of
(x, n) and all of its descendants are contained in h− (r, c) and are therefore at a
distance greater than dm . Thus, the corresponding subtree can be omitted from
the search. The following primitives will be used to define the query procedure
presented in Algorithm 1.
Side of hyperplane. A procedure sideOf(x, p, n) → {−, +}, with x, p ∈ M,
n ∈ Rn . Returns + iff x ∈ h+ (p, n) and − otherwise. For convenience, define
opposite(+) = − and vice-versa.
Queue. For an m-nearest-neighbor query, the algorithm maintains a Bounded
Priority Queue, Q of size m to collect the results. The queue can be thought of
as a sequence Q = [q1 , q2 , . . . qm ], where each element qi is defined as a distancevertex pair qi : (di , vi ), di ∈ R≥0 , vi ∈ V. The property d1 ≤ d2 ≤ · · · ≤ dm is
an invariant of the data structure. When an element (dnew , nnew ) is inserted in
the queue, if dnew < dm , then qm is discarded and the indices of the remaining
elements are rearranged to maintain the order.
Ball-Hyperplane Intersection. A procedure that determines whether a ball
and a hyperplane intersect. Precisely, ballHyperplane(x, R, p, n), with x ∈ M
and R ≥ 0, returns true if B(x, R) ∩ h(p, n) 6= ∅. Note that it does not need to
return false otherwise.
In the classic analysis of k-d trees [4, eq. 14] the distance is assumed to be a
sum of component-wise terms, called coordinate distance functions. Namely:
P
k
d(x, y) = F
(6)
i=1 di (xi , yi ) .
When this holds, e.g., for L-p metrics, the ball-hyperplane intersection procedure
reduces to the direct comparison of two numbers.
However, this property does not hold for other notions of distance and in
general sub-Riemannian balls can have nontrivial shapes. For these cases, the
procedure can be implemented by checking that Vo (x, R) ∩ h(p, n) 6= ∅, where
Vo is an “outer set” with a convenient geometry, i.e., a set such that Vo (x, R) ⊇
B(x, R) for all x ∈ M, R > 0 and such that intersections with hyperplanes are
easy to verify. Clearly, vol(Vo (x, R))/vol(B(x, R)) ≥ 1 and it is desirable that
this ratio stays as small as possible.
A crucial consequence of Theorem 2 is that the choice Vo (x, R) = Boxw
o (x, R)
offers optimal asymptotical behavior, since the volume of the outer set scales with
7
3. THE K-D TREE DATA STRUCTURE
the smallest possible order, namely: vol(Boxw
o (x, R)) ∈ Θ[vol(B(x, R))]. This fact
is fundamental to devise efficient query algorithms for nonholonomic systems.
Example (Ball-hyperplane intersection for a Reeds-Shepp vehicle). The reachable sets shown in figure 1 have a nontrivial geometry. Let us consider two
possible implementations of the ball-hyperplane intersection procedure:
1. Euclidean bound (EB). Define the set C(p, R) = {x ∈ SE(2) |(x1 − p1 )2 +
(x2 − p2 )2 ≤ R2 , |x3 − p3 | ≤ 2R}, i.e., a cylinder in the configuration space
with axis along θ̂. Since the Euclidean distance ka − bk is a lower bound
to dRS (a, b), then B(x, R) ⊂ C(x, R) and ballHyperplane can be correctly
implemented by checking C(x, R) ∩ h(p, n) 6= ∅
2. Outer Box Bound (BB). In this case, the procedure checks Boxw
o (x, R) ∩
h(p, n) 6= ∅. A simple implementation is to test whether all vertices of the
box (8 in this case) are on the same halfspace defined by h(p, n).
While Euclidean bounds (EB) are a simple choice, the Outer Box Bound (BB)
method approximates the sub-Riemannian ball tightly. In fact, as R → 0,
4
vol(C(p, R)) ∈ Θ(R3 ), while vol(Boxw
o (p, R)) ∈ Θ(R ) and therefore the volume of the cylinder tends to be infinitely larger than the volume of the outer
box. As a result, method (BB) yields an asymptotically unbounded speedup in
the algorithm compared to (EB), as confirmed by our experimental results in
Section 6. In addition, any other implementation of the procedure will at most
provide a constant factor improvement with respect to method (BB).
Algorithm 1: k-d tree query
1
2
3
4
5
6
7
8
9
define query (q ∈ M, T = (V, E − , E + )):
Q ←− {q1:m = (∅, ∞)} ;
// initialize queue
define querySubtree (q ∈ M, vi = (xi , ni ) ∈ V):
if vi = ∅ then return;
// caller reached leaf
s ←− sideOf (q, xi , ni );
querySubtree (q, child(vi , s)) ;
// descend to child containing q
Q ←− Q ∪ {(dist(q, xi ), vi )} ;
// add vertex to the queue
if ballHyperplane (q, dk , xi , ni ) then
// check for intersections
querySubtree (q, child(vi , opposite(s))) ;
// visit sibling
11
querySubtree (q, root (T )) ;
return Q;
3.2
k-d tree build algorithms
10
// start recursion from root
The performance of nearest-neighbor queries is heavily dependant on how the
k-d tree is constructed. In the sequel, we describe two popular approaches to
construct k-d trees: batch (or static) and incremental (or dynamic).
In the batch algorithm, all data points are processed at once and statistics
of the database determine the vertices of the tree at each depth. This algorithm
guarantees a balanced tree, but the insertion of new points is not possible without
8
3. THE K-D TREE DATA STRUCTURE
re-building the tree. Conversely, in the incremental version, the tree is updated
on the fly as points are inserted, however, the tree balance guarantee is lost.
Both algorithms find applications in motion planning: the batch version can
be used to efficiently build roadmaps for off-line, multiple-query techniques such
as PRMs, while the incremental is suitable for anytime algorithms such as RRTs.
Batch k-d tree build algorithm
The following primitive procedures will be used to describe the batch construction of k-d trees, defined in Algorithm 2. The range of a set S along direction dˆ
ˆ − inf x∈S hx, di.
ˆ
is defined as rng dˆ(S) = supx∈S hx, di
Maximum range. Let maxRange : 2X → Rn . Given a subset D ∈ 2X of the
data points, maxRange determines a direction l along which the range of data
is maximal. We consider the formulation in [4], where l is chosen among the
cardinal directions {êi }i∈[0,...k−1] , so that maxRange(D) = arg maxêi [rng êi (D)].
Median element. Let median : 2X × Rn → D. Given a subset D ∈ 2X of the
data points and a direction l, median returns the median element of D along
direction l.
Algorithm 2: Batch k-d tree build
1
2
3
4
5
6
7
8
9
10
11
define build (X):
V←−∅; E − ←−∅; E + ←−∅ ;
// Initialize
define buildSubtree (D ∈ 2X , vi = (xi , ni ) ∈ V, s ∈ {−, +}):
if D = ∅ then return;
// caller reached
nnew ←− maxRange (D); xnew ←− median (D, nnew ) ;
vnew ←− (xnew , nnew ) ; V ←− V ∪ vnew ;
if s 6= ∅ then E s ←− E s ∪ (vi , vnew ) ;
L ←− D ∩ h− (xnew , nnew ); R ←− D\(L ∪ xnew );
buildSubtree (L, vnew , -); buildSubtree (R, vnew , +);
tree
leaf
buildSubtree (X, ∅, ∅);
return T = (V, E − , E + );
Incremental k-d tree build algorithm
A key operation to build a k-d tree incrementally is to pick splitting hyperplanes
on the fly as new data points are inserted. This is achieved with a splitting
sequence, which we define as:
Splitting sequence. A map ZM : N≥0 × M → Rn that, given an integer d ≥ 0
and a point p ∈ M, returns a normal vector n associated with them.
When dealing with k-dimensional data, one usually chooses a basis of cardinal
axes {êi }i∈[0...k−1] . The normal of the hyperplane associated with a vertex v is
typically picked by cycling through the cardinal axes based on the depth of v in
the binary tree. Let “a mod b” denote the remainder of the division of integer a
by integer b. Then, the classic splitting (CS) sequence is defined as:
classic
ZM
(d, x) = ê[d mod k] .
9
(7)
4. COMPLEXITY ANALYSIS
Algorithm 3: Incremental k-d tree build.
Insert in k-d tree T = (V, E + , E − ), start recursion with insert(xnew , root(T )).
1
2
3
4
5
6
7
8
9
10
4
define insert (xnew ∈ M, vi = (xi , ni ) ∈ V):
if V = ∅ then
nnew ←− ZM (0, xnew ); V←−(xnew , nnew ) ;
return
// if tree is empty...
// ...add root
side ←− sideOf (xnew , xi , ni ) ; c ←− child (vi , side) ;
if c = ∅ then
nnew ←− ZM (depth(vi ) + 1, xnew ) ; vnew ←− (xnew , nnew );
V ←− V ∪ vnew ; E side ←− E side ∪ (vi , vnew ) ;
return
insert (xnew , c) ;
Complexity Analysis
In this section, we discuss the expected asymptotic complexity of m-nearestneighbor queries in a k-d tree built according to Algorithm 2. The complexity of
the nearest-neighbor query procedure (Algorithm 1) is measured in terms of the
number nv of vertices examined. Let B(q, dm ) be the ball containing the m nearest neighbors to q. For the soundness of the algorithm, all the vertices contained
in this ball must be examined, and therefore, all the buckets overlapping with
it. As we recall from Section 3, buckets are associated with leaves, and therefore
the algorithm will visit as many leaves as the number of such buckets, formally:
nl ∈ Θ(|β|), where β := {b ∈ BT | b ∩ B(q, dm ) 6= ∅} and nl denotes the number
of leaves visited by the algorithm.
If the asymptotic order of visited leaves is known, proofs rely on the following
fact: since the batch algorithm guarantees a balanced tree, descending into nl
leaves from the root requires visiting nv ∈ Θ(nl log N ) vertices, where N is the
cardinality of the database X. Thus the expected query complexity is given by:
E(nv ) ∈ Θ[E(nl ) log N ].
(8)
Lemma 1 reviews the seminal result, originally from [4], that the expected
asymptotic complexity for a Minkowski (i.e., L-p) distance is logarithmic. Our
main theoretical contribution is the negative result described in Theorem 3,
where we reveal that the expected complexity for sub-Riemannian metrics is
in fact super-logarithmic. In the sequel, assume that data points are randomly
sampled from M with probability distribution p(x).
Lemma 1 ([4]). If the distance function used for the nearest-neighbor query is
induced by a p-norm, then the expected complexity of the nearest-neighbor search
on a k-d tree built with Algorithm 2 is O(log(N )) as N → ∞.
For the detailed proof, we suggest reading the original work by Friedman et
al. [4], while here we report key facts used in our next result. In [4, pg. 214], it
10
4. COMPLEXITY ANALYSIS
is shown that:
E[vol(B(q, dm ))] ≈
m
1
,
(N + 1) p(q)
(9)
with no assumptions on the metric and the probability distribution p(x).
In a batch-built k-d tree, the expected asymptotic shape of the bucket bq
containing the query point q is assumed hyper-cubical. From equation (9) with
1
. Thus, hyper-cubes in a neighborhood of q have
m = 1, E(vol(bq )) ≈ (N1+1) p(q)
sides of expected length E(lq ) = [(N + 1)p(q)]−1/k . Using these facts, it is then
shown that the expected number of visited leaves is asymptotically constant
with the size of the database, E(nl ) ∈ Θ(1) and therefore the average query
complexity is logarithmic, as per equation (8).
Theorem 3. Let the distance function used for the nearest-neighbor query be
a sub-Riemannian metric on a smooth connected manifold M with a bracketgenerating horizontal distribution of dimension h and weights {wi }. Then the
expected complexity of the nearest-neighbor query on a k-d tree built with Algorithm 2 is Θ(N p log(N )) as N → ∞, where the expression for p is:
X 1
X
wi
p=
−
, with W :=
wi .
(10)
k W
i
wi ≤W/k
Proof. For N large enough, Theorem 2 ensures the existence of inner and outer
bounding boxes to the sub-Riemannian ball B(q, dm ).
From equations (4) and (9), we get:
E(dm )W ·
k
Y
i=1
ci ≤
k
Y
1
1
≤ E(dm )W ·
Ci ,
N + 1 p(q)
i=1
(11)
P
where W := i wi . Therefore, the expected distance to the m-th nearest neighbor scales asymptotically as E(dm ) ∈ Θ(N −1/W
p ) for N → ∞. Recall that a
bucket bq has sides of expected length lq = k 1/ [(N + 1)p(q)] or equivalently,
lq ∈ Θ(N −1/k ). We are now interested in the asymptotic order of the number
of buckets overlapping with a weighted box of size dm . Let nh (d, lq ) the number
of hypercubes intersected by a segment
of Rk
l √ of lengthm d embedded in a grid
√
−1
with side lq . It can be shown that ( k) (d/lq ) ≤ nh (d, lq ) ≤ k + kdd/lq e.
Then, asymptotically, nh (d, lq ) ∈ Θ(dd/lq e) as (d, lq ) → 0. Since a box has k
orthogonal sides, each with expected length E(dm )wi , the expected number of
visited buckets, is:
!
!
k
k l
m
Y
Y
wi
1
E(dm )wi
.
(12)
E(nl ) ∈ Θ
=Θ
N k− W
l
q
i=1
i=1
In the latter product, only the factors with exponent > 0 contribute to the
complexity of the algorithm, while the other terms tend to 1, as N → ∞. In
other words, the query complexity is determined only by low order Lie brackets
11
5. THE LIE SPLITTING STRATEGY
up until wi ≤ W/k. In fact, along the direction of higher order Lie brackets,
reachable sets shrink to zero faster than the side of a k-d tree bucket. Following
up from equation (12), we get:
X 1
Y
w
1
wi
i
E(nl ) ∈ Θ
N k − W = Θ (N p ) , p =
−
.
(13)
k W
wi ≤W/k
wi ≤W/k
From equation (8) it follows that the expected complexity of the query algorithm
is given by E(nv ) ∈ Θ(N p log N ) as N → ∞.
t
u
Remark. For holonomic systems, the query complexity is logarithmic, in accordance with Lemma 1. In fact, for holonomic systems, w1 = w2 = · · · = wk = 1,
then W = k and p = 0 from equation (10).
Remark. The query complexity is always between logarithmic and linear. Since
W ≥ k by definition, for all i such that wi ≤ W/k, one can state 0 < wi /W ≤
1
i
1/k ≤ 1. It follows that 0 ≤ k1 − w
W < k . Then p can be bounded with:
X 1
X 1
wi
1
<
≤ k · = 1 ⇒ p ∈ [0, 1).
0≤p=
−
(14)
k W
k
k
wi ≤W/k
wi ≤W/k
Example. For the Reeds-Shepp car, k = dim[SE(2)] = 3, wfˆ = wθ̂ = 1, wl̂ = 2
and W = wfˆ + wθ̂ + wl̂ = 4. Only wfˆ and wθ̂ satisfy wi ≤ W/k, therefore,
according
to
Theorem 3, the expected query complexity of a batch kd-tree is
√
6
Θ
N log N . We confirm this prediction with experiments in section 6.
5
The Lie splitting strategy
The basic working principle of k-d trees is the ability to discard subtrees without
loss of information during a query. For each visited node v = (x, n), the algorithm checks whether the current biggest ball in the queue, B(q, dm ) intersects
h(x, n). If such an intersection exists, the algorithm visits both children of v,
otherwise one child is discarded and so is the entire subtree rooted in it. To
reduce complexity, it is thus desirable that during query, a k-d tree presents as
few ball-hyperplane intersections as possible.
In the classic splitting strategy, hyperplanes are chosen cycling through a
globally defined set of cardinal axes. However, we have shown that nonholonomic
systems have a set of locally-defined privileged axes and that the reachable sets
have different infinitesimal orders along each of them. When the metric comes
from a nonholonomic system, the hyperplanes in a classic k-d tree are not aligned
with reachable sets and the buckets have different asymptotic properties than
reachable sets. This can make intersections frequent, as shown in figure 2(a).
Ideally, to minimize the number of ball-hyperplane intersections, the buckets
in a k-d tree should approximate the bounding boxes for the reachable sets of the
12
5. THE LIE SPLITTING STRATEGY
θ
y
y
x
(a) Classic k-d tree construction
θ
x
(b) Lie k-d tree splitting strategy
Fig. 2. Qualitative comparison of a classic k-d tree with its counterpart built with the
proposed Lie splitting strategy. For nonholonomic systems, reachable sets (red, blue,
green) are elongated along configuration-dependent privileged directions. The smaller
the sets, the more pronounced their aspect ratios. The Lie splitting strategy adapts the
hyperplanes locally to the balls and decreases the expected number of ball-hyperplane
intersections, thus reducing the expected asymptotic query complexity.
dynamical system, as depicted in figure 2(b). To achieve this, we propose a novel
splitting rule, named the Lie splitting strategy, which exploits the differential
geometric properties of a system and the asymptotic scaling of its reachable
sets. The Lie splitting strategy is based on the following two principles:
1. The splitting normal associated with each data point xi is along one of the
privileged axes in that point, i.e., ZM (d, xi ) ∈ {y1 (xi ), y2 (xi ) . . . yk (xi )}.
2. The buckets of the k-d tree, in expectation, scale asymptotically according to
the weighted box along all privileged axes as t → 0. Formally, for all q ∈ M
and for all pairs of privileged axes yi (q), yj (q),
h
iwj
h
iwi
E rng yi (q) (bq )
∈ Θ E rng yj (q) (bq )
.
(15)
Requirement 2 prescribes the asymptotic behavior of the sequence ZM as d →
∞, which can be formalized as follows: let ni (d) = |{n < d : ZM (n, x) = yi (x)}|,
i.e., the total number of splits in the sequence ZM along axis yi before index d.
As d h→ ∞, the expected
bucket size along yi after ni (d) splits has asymptotic
i
order E rng yi (q) (bq ) ∈ Θ[e−ni (d) ]. Then, in terms of the number of splits,
equation (15) yields:
ni (d) · wj ∼ nj (d) · wi for all i, j ∈ [1, . . . k] as d → ∞.
(16)
Simply put, each privileged direction should be picked as a splitting normal with a frequency proportional to its weight. Note that this is only relevant
asymptotically, i.e., as d → ∞.
13
6. EXPERIMENTAL RESULTS
Example. For the Reeds-Shepp vehicle, a valid Lie splitting sequence is:
ˆl(x), if d ≡ 0 or d ≡ 1 (mod 4)
Lie
ZRS (d, x) := fˆ(x), if d ≡ 2 (mod 4)
θ̂(x), if d ≡ 3 (mod 4)
(17)
It is easy to verify that this satisfies equation (16). In fact, lateral splits occur
in the sequence twice as often as longitudinal and lateral splits.
6
Experimental results
In this section, we validate our results and compare the performance of our
algorithms with different methods from widely used open-source libraries. Query
performances are averaged over 1000 randomly drawn query points. The same
insertion and query sequences are used across all the analyzed algorithms and
generated from a uniform distribution.
Experiment 1. In this experiment, we confirm the theoretical contributions presented in Section 4. In figure 3(a) we show the average number of leaves visited
when querying a batch k-d tree with Euclidean (blue) and Reeds-Shepp metrics (solid red). While the blue curve settles to a constant value, in accordance
with Lemma 1,√the red curve exhibits exponential growth. When normalizing
this curve by 6 N (dashed red), we observe a constant asymptotic behavior,
consistent with the rate determined by Theorem 3. For comparison, we report
the average time for Euclidean queries on an incremental k-d tree (black), which
tends to visit more vertices than its batch, balanced counterpart (blue).
Experiment 2. In this experiment (figures 3(b-d)), we plot the total number of
distance evaluations and the running times observed with different combinations
of build and query algorithms. Figure 3(d) reveals that the proposed outer Box
Bound (BB) ball-hyperplane intersection method (yellow, purple, green) reduces
the query time significantly as compared to Euclidean bounds (EB) (blue, red),
in accordance with our predictions in Section 3.1.
Additionally, Lie splitting (green) further improves query time, as compared
with the classic splitting (yellow). The corresponding incremental Lie k-d tree
also outperforms a classic batch-built one (purple), guaranteed to be balanced.
More important speedups emerge when comparing the proposed k-d trees
with different techniques, such as Hierarchical Clustering from FLANN and Geometric Near-neighbor Access Tree (GNAT) from OMPL. Interestingly, off-theshelf implementations of k-d trees offered by FLANN and other tools are unusable with nonholonomic metrics altogether, since they are limited to distances
of the form of equation (6). Therefore a comparison is not possible.
All the tested k-d trees visibly outperform Hierarchical Clustering (cyan)
both in build time and query time. In contrast, GNAT offers competitive query
times. However, its insertion is ∼ 100× slower than incremental k-d trees, since a
significant number of distances are evaluated in the build phase, while k-d trees
14
7. CONCLUSION
Batch - dE
Incremental, Z classic - dE
Batch - dRS , BB
Batch - dRS , BB
√
normalized by 6 N
10 2
Incremental, Z classic
- dRS , EB
Batch - dRS , EB
Incremental, Z classic - dRS , BB
10 8
Batch - dRS , BB
Lie
Incremental, ZRS
- dRS , BB
10 7
FLANN Hierarchical clustering
OMPL GNAT
10 6
10 1
10 5
N: number of samples
N: number of samples
10 1
10 2
10 3
10 4
10 5
102
10 6
103
(a) Average leaves visited (query)
10
104
105
106
107
108
(b) Total distance evaluations
8
10 4
10
10
10
6
4
10 3
2
N: number of samples
N: number of samples
10
2
10
3
10
4
10
5
10
6
10
7
10
8
10 2
(c) Build time (µs)
10 3
10 4
10 5
10 6
10 7
10 8
(d) Average query time (µs)
Fig. 3. (a) Experiment 1 : Average number of leaves visited during Euclidean (blue) and
Reeds-Shepp (red, solid) queries of a batch k-d tree. (b-d) Experiment 2 : Performance
of different algorithms. In the legends, dE and dRS indicate Euclidean and ReedsShepp queries, EB and BB indicate Euclidean Bound and outer Box Bound for ballLie
indicate the splitting sequence.
hyperplane intersections, Z classic and ZRS
only evaluate distances during query. This is reflected in a noticeably higher
asymptotic rate of distance evaluations, revealed in figure 3(b).
7
Conclusion
Motivated by applications in sampling-based motion planning, we investigated
k-d trees for efficient nearest-neighbor search with distances defined by the length
of paths of controllable nonholonomic systems. We have shown that for subRiemannian metrics, the query complexity of a classic batch-built k-d tree is
Θ(N p log N ), where p ∈ [0, 1) depends on the properties of the system. In addition, we have proposed improved build and query algorithms for k-d trees that
account for differential constraints. The proposed methods proved superior over
classic ones in numerical experiments carried out with a Reeds-Shepp vehicle.
Future work will analyze whether logarithmic complexity is achieved for nonholonomic systems. In addition, the proposed algorithms are exact and rely
on explicit distance evaluations. Since distances cannot be generally computed
in closed form, we are interested in investigating approximate nearest-neighbor
search algorithms with provable correctness bounds that do not require explicit
distance computations.
15
7. CONCLUSION
References
1. J. L. Bentley. Multidimensional binary search trees used for associative searching.
Communications of the ACM, 18(9):509–517, 1975.
2. M. Bharatheesha, W. Caarls, W. J. Wolfslag, and M. Wisse. Distance metric
approximation for state-space RRTs using supervised learning. In 2014 IEEE/RSJ
International Conference on Intelligent Robots and Systems, pages 252–257, 2014.
3. B. Burns and O. Brock. Single-query motion planning with utility-guided random trees. In Proceedings 2007 IEEE International Conference on Robotics and
Automation, pages 3307–3312, 2007.
4. J. H. Friedman, J. L. Bentley, and R. A. Finkel. An algorithm for finding best
matches in logarithmic expected time. ACM Transactions on Mathematical Software (TOMS), 3(3):209–226, 1977.
5. E. Glassman and R. Tedrake. A quadratic regulator-based heuristic for rapidly exploring state space. In Robotics and Automation (ICRA), 2010 IEEE International
Conference on, pages 5021–5028, 2010.
6. J. Ichnowski and R. Alterovitz. Fast nearest neighbor search in SE(3) for samplingbased motion planning. In Algorithmic Foundations of Robotics XI, pages 197–214.
Springer, 2015.
7. S. Karaman and E. Frazzoli. Sampling-based algorithms for optimal motion planning. The International Journal of Robotics Research, 30(7):846–894, 2011.
8. S. Karaman and E. Frazzoli. Sampling-based optimal motion planning for nonholonomic dynamical systems. In Robotics and Automation (ICRA), 2013 IEEE
International Conference on, pages 5041–5047, 2013.
9. L. E. Kavraki, P. Svestka, J.-C. Latombe, and M. H. Overmars. Probabilistic
roadmaps for path planning in high-dimensional configuration spaces. IEEE transactions on Robotics and Automation, 12(4):566–580, 1996.
10. J. J. Kuffner. Effective sampling and distance metrics for 3D rigid body path
planning. In Robotics and Automation, 2004. Proceedings. ICRA’04. 2004 IEEE
International Conference on, volume 4, pages 3993–3998, 2004.
11. S. M. LaValle. Rapidly-exploring random trees: A new tool for path planning.
Citeseer, 1998.
12. S. M. LaValle and J. J. Kuffner. Randomized kinodynamic planning. The International Journal of Robotics Research, 20(5):378–400, 2001.
13. R. Montgomery. A tour of subriemannian geometries, their geodesics and applications. Number 91. American Mathematical Soc., 2006.
14. M. Muja and D. G. Lowe. Fast approximate nearest neighbors with automatic
algorithm configuration. In International Conference on Computer Vision Theory
and Application VISSAPP’09), pages 331–340. INSTICC Press, 2009.
15. L. Palmieri and K. Arras. Distance metric learning for RRT-based motion planning
for wheeled mobile robots. In 2014 IROS Machine Learning in Planning and
Control of Robot Motion Workshop, 2014.
16. J. Reeds and L. Shepp. Optimal paths for a car that goes both forwards and
backwards. Pacific journal of mathematics, 145(2):367–393, 1990.
17. A. Shkolnik, M. Walter, and R. Tedrake. Reachability-guided sampling for planning
under differential constraints. In Robotics and Automation, 2009. ICRA’09. IEEE
International Conference on, pages 2859–2865, 2009.
18. I. A. Şucan, M. Moll, and L. E. Kavraki. The Open Motion Planning Library.
IEEE Robotics & Automation Magazine, (4):72–82, December.
19. A. Yershova and S. M. LaValle. Improving motion-planning algorithms by efficient
nearest-neighbor searching. IEEE Transactions on Robotics, 23(1):151–157, 2007.
16
| 3cs.SY
|
1
Lossless Analog Compression
Giovanni Alberti1 , Helmut Bölcskei2 , Camillo De Lellis3 ,
Günther Koliander4 , and Erwin Riegler2
1
2
arXiv:1803.06887v1 [math.FA] 19 Mar 2018
3
4
Dept. of Mathematics, University of Pisa, Italy, Email: [email protected]
Dept. IT & EE, ETH Zurich, Switzerland, Email: {boelcskei, eriegler}@nari.ee.ethz.ch
Dept. of Mathematics, University of Zurich, Switzerland, Email: [email protected]
ARI, Austrian Academy of Sciences, Vienna, Austria, Email: [email protected]
Abstract
We establish the fundamental limits of lossless analog compression by considering the recovery of
arbitrary random vectors x ∈ Rm from the noiseless linear measurements y = Ax with measurement
matrix A ∈ Rn×m . Our theory is inspired by the groundbreaking work of Wu and Verdú (2010) on almost
lossless analog compression, but applies to the nonasymptotic, i.e., fixed-m case, and considers zero error
probability. Specifically, our achievability result states that, for Lebesgue-almost all A, the random vector
x can be recovered with zero error probability provided that n > K(x), where the description complexity
K(x) is given by the infimum of the lower modified Minkowski dimensions over all support sets U of
x (i.e., sets U ⊆ Rm with P[x ∈ U] = 1). We then particularize this achievability result to the class
of s-rectifiable random vectors as introduced in Koliander et al. (2016); these are random vectors of
absolutely continuous distribution—with respect to the s-dimensional Hausdorff measure—supported
on countable unions of s-dimensional C 1 -submanifolds of Rm . Countable unions of C 1 -submanifolds
include essentially all signal models used in the compressed sensing literature such as the standard
union of subspaces model underlying much of compressed sensing theory and spectrum-blind sampling,
smooth manifolds, block-sparsity, and low-rank matrices as considered in the matrix completion problem.
Specifically, we prove that, for Lebesgue-almost all A, s-rectifiable random vectors x can be recovered
with zero error probability from n > s linear measurements. This threshold is, however, found not to
be tight as exemplified by the construction of an s-rectifiable random vector that can be recovered with
zero error probability from n < s linear measurements. This leads us to the introduction of the new class
of s-analytic random vectors, which admit a strong converse in the sense of n ≥ s being necessary for
recovery with probability of error smaller than one. The central conceptual tool in the development of
our theory is geometric measure theory.
The material in this paper was presented in part at the 2016 IEEE International Symposium on Information Theory [1].
March 20, 2018
DRAFT
2
I. I NTRODUCTION
Compressed sensing [2]–[6] deals with the recovery of unknown sparse vectors x ∈ Rm from a small
(relative to m) number, n, of linear measurements of the form y = Ax, where A ∈ Rn×m is the
measurement matrix.1 Known recovery guarantees can be categorized as deterministic, probabilistic, and
information-theoretic. The literature in all three categories is abundant and the ensuing overview is hence
necessarily highly incomplete, yet representative.
Deterministic results, such as those in [2], [6]–[10], are uniform in the sense of applying to all ssparse vectors x for a fixed measurement matrix A. Typical guarantees say that s-sparse vectors x can
be recovered through convex optimization algorithms or greedy algorithms provided that s < 21 (1 + 1/µ),
where µ denotes the coherence of A, i.e., the largest (in absolute value) inner product of any two different
columns of A. The Welch bound [11] implies that the minimum number of linear measurements, n,
required for uniform recovery is of order s2 , a result known as the “square-root bottleneck.”
Probabilistic results are either based on random2 measurement matrices A [3]–[5], [12]–[15] or deterministic A and random s-sparse vectors x [13], [14] and typically state that s-sparse vectors can
be recovered, again using convex optimization algorithms or greedy algorithms, with high probability,
provided that n is of order s log m.
An information-theoretic framework for compressed sensing, fashioned as an almost lossless analog
compression problem, was developed by Wu and Verdú [16], [17]. Specifically, [16] derives asymptotic (in
m) achievability results and converses for linear encoders and measurable decoders, measurable encoders
and Lipschitz continuous decoders, and continuous encoders and continuous decoders. For the particular
case of linear encoders and measurable decoders, [16] shows that, asymptotically in m, for Lebesgue
almost all (a.a.) measurement matrices A, the random vector x can be recovered with arbitrarily small
probability of error from n = bRmc linear measurements, provided that R > RB , where RB denotes the
Minkowski dimension compression rate [16, Definition 10] of the random process generating x. For the
special case of x with independent and identically distributed (i.i.d.) discrete-continuous mixture entries,
a matching converse exists. Discrete-continuous mixture distributions ρµc + (1 − ρ)µd are relevant as they
mimic sparse vectors for large m. In particular, if the discrete part µd is a Dirac measure at 0, then the
nonzero entries of x can be generated only by the continuous part µc , and the fraction of nonzero entries
in x converges—in probability—to ρ as m tends to infinity. A nonasymptotic, i.e., fixed-m, statement
in [16] says that a.a. (with respect to a σ -finite Borel measure on Rm ) s-sparse random vectors can be
1
Throughout the paper, we assume that n ≤ m.
2
Random quantities are denoted by sans-serif letters, throughout.
March 20, 2018
DRAFT
3
recovered with zero error probability provided that n > s. Again, this result holds for Lebesgue a.a.
measurement matrices A ∈ Rn×m . A corresponding converse does not seem to be available.
Contributions. We establish the fundamental limits of lossless, i.e., zero error probability, analog
compression in the nonasymptotic, i.e., fixed-m, regime for arbitrary random vectors x ∈ Rm . Specifically,
we show that x can be recovered with zero error probability provided that n > K(x), with the description
complexity K(x) given by the infimum of the lower modified Minkowski dimensions over all support
sets U of x, i.e., all sets U ⊆ Rm with P[x ∈ U] = 1. This statement holds for Lebesgue-a.a. measurement
matrices. Lower modified Minkowski dimension vastly generalizes the notion of s-sparsity and allows
for arbitrary support sets that are not necessarily finite unions of s-dimensional linear subspaces. For
s-sparse vectors, we get the recovery guarantee n > s showing that our information-theoretic thresholds
suffer neither from the square-root bottleneck nor from a log m-factor. We hasten to add, however, that
we do not specify explicit decoders that achieve these thresholds, rather we provide existence results
absent computational considerations. The central conceptual element in the proof of our achievability
result is the probabilistic null-space property first reported in [18]. We emphasize that it is the usage of
modified Minkowski dimension, as opposed to Minkowski dimension [16], [18], that allows us to obtain
achievability results for zero error probability. The asymptotic achievability result for linear encoders in
[16] can be recovered in our framework.
We particularize our achievability result to s-rectifiable random vectors x as introduced in [19]; these
are random vectors supported on countable unions of s-dimensional C 1 -submanifolds of Rm and of
absolutely continuous—with respect to s-dimensional Hausdorff measure—distribution. Countable unions
of C 1 -submanifolds include numerous signal models prevalent in the compressed sensing literature,
namely, the standard union of subspaces model underlying much of compressed sensing theory [20],
[21] and spectrum-blind sampling [22], [23], smooth manifolds [24], block-sparsity [25]–[27], and lowrank matrices as considered in the matrix completion problem [28]–[30]. Our achievability result shows
that s-rectifiable random vectors can be recovered with zero error probability provided that n > s.
Again, this statement holds for Lebesgue-a.a. measurement matrices. Absolute continuity with respect
to s-dimensional Hausdorff measure is a regularity condition ensuring that the distribution is not too
concentrated; in particular, sets of Hausdorff dimension t < s are guaranteed to carry zero probability
mass. One would therefore expect n ≥ s to be necessary for zero error recovery of s-rectifiable random
vectors. It turns out, however, that, perhaps surprisingly, this is not the case in general. An example
elucidating this phenomenon constructs a set G ⊆ R2 of positive 2-dimensional Hausdorff measure that
can be compressed linearly in a one-to-one fashion into R. This will then be seen to lead to the statement
that every 2-rectifiable random vector of distribution absolutely continuous with respect to 2-dimensional
March 20, 2018
DRAFT
4
Hausdorff measure restricted to G can be recovered with zero error probability from a single linear
measurement. What renders this result surprising is that all this is possible even though G contains the
image of a Borel set in R2 of positive Lebesgue measure under a C ∞ -embedding. The picture changes
completely when the embedding is real analytic. Specifically, we show that if a set U ⊆ Rm contains the
real analytic embedding of a Borel set in Rs of positive Lebesgue measure, it can not be compressed
linearly (in fact, not even through a nonlinear real analytic mapping) in a one-to-one fashion into Rn
with n < s. This leads to the new concept of s-analytic random vectors, which allows a strong converse
in the sense of n ≥ s being necessary for recovery of x with probability of error smaller than one. The
qualifier “strong” refers to the fact that recovery from n < s linear measurements is not possible even
if we allow an arbitrary positive error probability strictly smaller than one. The only strong converse
available in the literature applies to random vectors x with i.i.d. discrete-continuous mixture entries [16].
Organization of the paper. In Section II, we present our achievability results, with the central statement
in Theorem II.1. Section III particularizes these results to s-rectifiable random vectors and presents the
2-rectifiable example random vectors that can be recovered with zero error probability from a single linear
measurement. In Section IV, we introduce and characterize the new class of s-analytic random vectors
and we derive a corresponding strong converse, stated in Theorem IV.1. Sections V–VII contain the
proofs of the main technical results stated in Sections II–IV. Appendices A–G contain proofs of further
technical results stated in the main body of the paper. In Appendices H–K, we summarize concepts and
basic results from (geometric) measure theory, the theory of set-valued functions, sequences of functions
in several variables, and real analytic mappings, all needed throughout the paper. The reader not familiar
with these results is advised to consult the corresponding appendices before studying the proofs of our
main results. These appendices also contain new results, which are highly specific and would disrupt the
flow of the paper if presented in the main body.
Notation. We use capital boldface roman letters A, B, . . . to denote deterministic matrices and lowercase boldface roman letters a, b, . . . to designate deterministic vectors. Random matrices and vectors are
set in sans-serif font, e.g., A and x. The m × m identity matrix is denoted by Im . We write rank(A) and
ker(A) for the rank and the kernel of A, respectively. For the matrix A, spark(A) is the smallest number
k such that there exists a set of k column vectors of A that are linearly dependent. The superscript T
√
stands for transposition. The i-th unit vector is denoted by ei . For a vector x ∈ Rm , kxk2 = xT x is its
Euclidean norm and kxk0 denotes the number of nonzero entries of x. For the set A, we write card(A)
for its cardinality, A for its closure, 2A for its power set, and
1A for the indicator function on A. With
A ⊆ Rm and B ⊆ Rn , we let A × B = {(a, b) : a ∈ A, b ∈ B} and A ⊗ B = {a ⊗ b : a ∈ A, b ∈ B},
where ⊗ denotes the Kronecker product. For A, B ⊆ Rm , we write A ( B to express that A ⊆ B
March 20, 2018
DRAFT
5
with A 6= B and let A
B = {a − b : a, b ∈ B}. For the Euclidean space (Rk , k · k2 ), we designate
the open ball of radius ρ centered at u ∈ Rk by Bk (u, ρ), and use V (k, ρ) to refer to its volume. We
write S (X ) for a general σ -algebra on X , B(X ) for the Borel σ -algebra on a topological space X , and
L (Rm ) for the Lebesgue σ -algebra on Rm . The product σ -algebra of S (X ) and S (Y) is denoted by
S (X ) ⊗ S (Y). For measures µ and ν on the same measurable space, we denote absolute continuity
of µ with respect to ν by µ ν . We write µ × ν for the product measure of µ and ν . The Lebesgue
measure on Rk and Rk×l is designated as λk and λk×l , respectively. The distribution of a random vector
x is denoted by µx . If f : Rk → Rl is differentiable, we write Df (v) ∈ Rl×k for its differential at v ∈ Rk
p
and define the min(k, l)-dimensional Jacobian Jf (v) at v ∈ Rk by Jf (v) = det(Df (v)(Df (v))T )
p
if l < k and by Jf (v) = det((Df (v))T Df (v)) if l ≥ k . For an open set U ⊆ Rk , a differentiable
mapping f : U → Rl , where l ≥ k , is called an immersion if Jf (v) > 0 for all v ∈ U . A one-to-one
immersion is referred to as an embedding. For a mapping f , we write f ≡ 0 if it is identically zero
and f 6≡ 0 otherwise. For f : U → V and g : V → W , the composition g ◦ f : U → W is defined as
(g ◦ f )(x) = g(f (x)) for all x ∈ U . For f : U → V and A ⊆ U , f |A denotes the restriction of f to A.
For f : U → V and B ⊆ V , we set f −1 (B) = {x ∈ U : f (x) ∈ B}.
II. ACHIEVABILITY
In classical compressed sensing theory [3]–[6] one typically deals with the recovery of s-sparse vectors
x ∈ Rm , i.e., vectors x that are supported on a finite union of s-dimensional linear subspaces of Rm . The
purpose of this paper is the development of a comprehensive theory of signal recovery in the sense of
allowing arbitrary support sets U , which are not necessarily unions of (a finite number of) linear subspaces
of Rm . Formalizing this idea requires a suitable dimension measure for general nonempty sets. There is a
rich variety of dimension measures available in the literature [31]–[33]. Our choice will be guided by the
requirement of information-theoretic operational significance. Specifically, the dimension measure should
allow the formulation of nonasymptotic, i.e., fixed-m, achievability results with zero error probability.
Modified Minkowski dimension will turn out to meet these requirements.
We first recall the definitions of Minkowski dimension and modified Minkowski dimension, compare
the two concepts, and state the basic properties of modified Minkowski dimension needed in the remainder
of the paper.
Definition II.1. (Minkowski dimension3 ) [32, Section 3.1] For U ⊆ Rm nonempty, the lower and upper
3
Minkowski dimension is sometimes also referred to as box-counting dimension, which is the origin of the subscript B in
the notation dimB (·) used henceforth.
March 20, 2018
DRAFT
6
Minkowski dimension of U are defined as
dimB (U) = lim inf
log NU (ρ)
log ρ1
(1)
dimB (U) = lim sup
log NU (ρ)
,
log ρ1
(2)
ρ→0
and
ρ→0
respectively, where
n
o
[
NU (ρ) = min k ∈ N : U ⊆
Bm (ui , ρ), ui ∈ Rm
(3)
i∈{1,...,k}
is the covering number of U for radius ρ. If dimB (U) = dimB (U), this common value, denoted by
dimB (U), is the Minkowski dimension of U .
Definition II.2. (Modified Minkowski dimension) [32, Section 3.3] For U ⊆ Rm nonempty, the lower
and upper modified Minkowski dimension of U are defined as
(
dimMB (U) = inf sup dimB (Ui ) : U ⊆
i∈N
)
[
i∈N
Ui
(4)
and
)
(
dimMB (U) = inf sup dimB (Ui ) : U ⊆
i∈N
[
i∈N
Ui ,
(5)
respectively, where in both cases the infimum is over all possible coverings {Ui }i∈N of U by nonempty
compact sets Ui . If dimMB (U) = dimMB (U), this common value, denoted by dimMB (U), is the modified
Minkowski dimension of U .
The main properties of modified Minkowski dimension are summarized in Lemma H.12. In particular,
dimMB (·) ≤ dimB (·) and dimMB (·) ≤ dimB (·). Upper modified Minkowski dimension has the advantage
of being countably stable, a key property we will use frequently. In contrast, upper Minkowski dimension
is only finitely stable [32, Section 3.2, Property (iii)]. For example, all countable subsets of Rm have
modified Minkowski dimension zero (as a single point in Rm has Minkowski dimension zero); the
situation is different for Minkowski dimension as it does not respect countable unions:
Example II.1. [32, Example 3.5] Let F = {0, 1/2, 1/3, . . . }. Then, dimMB (F) = 0 < dimB (F) = 1/2.
Minkowski dimension and modified Minkowski dimension also behave differently for unbounded sets.
Specifically, by monotonicity of (upper) modified Minkowski dimension, dimMB (A) ≤ dimMB (A) ≤
dimMB (Rm ) = m for all A ⊆ Rm , in particular also for unbounded sets, whereas dimB (A) =
March 20, 2018
DRAFT
7
dimB (A) = ∞ for all unbounded sets A as a consequence of NA (ρ) = ∞ for all ρ ∈ (0, ∞). Working
with lower modified Minkowski dimension will allow us to consider arbitrary random vectors, regardless
of whether they admit bounded support sets or not.
The following example shows that modified Minkowski dimension agrees with the sparsity notion used
in classical compressed sensing theory.
Example II.2. Let T1 , . . . , Tk be linear subspaces of Rm with their Euclidean dimensions dim(Ti )
satisfying
max dim(Ti ) = s
(6)
i∈{1,...,k}
and consider the union
U=
k
[
i=1
Ti .
(7)
As every linear subspace is a smooth submanifold of Rm , it follows from Properties i), ii), and v) of
Lemma H.12 that dimMB (U) = s. In the union of subspaces model prevalent in compressed sensing
theory [3]–[6], the subspaces Ti correspond to different sparsity patterns, each of cardinality equal to the
sparsity level s.
The aim of the present paper is to develop a theory of lossless analog compression for arbitrary random
vectors x ∈ Rm . An obvious choice for the stochastic equivalent of the sparsity level s is the stochastic
sparsity level S(x) defined as
(
"
S(x) = min s : ∃T1 , . . . , Tk with P x ∈
k
[
i=1
#
Ti = 1 and
)
max dim(Ti ) = s ,
i∈{1,...,k}
(8)
where every Ti is a linear subspace of Rm . This definition is, however, specific to the union of linear
subspaces structure. The theory we develop here requires a more general notion of description complexity,
which we define in terms of lower modified Minkowski dimension according to
K(x) = inf{dimMB (U) : U ⊆ Rm , P[x ∈ U] = 1}.
(9)
Sets satisfying P[x ∈ U] = 1 are hereafter referred to as support sets of x. While the definition of S(x)
involves minimization of Euclidean dimensions of linear subspaces, K(x) is defined by minimizing the
lower modified Minkowski dimensions of general support sets. The definitions (8) and (9) imply directly
that K(x) ≤ S(x) for all random vectors x ∈ Rm (cf. Example II.2). We will see in Section III that this
inequality can actually be strict.
Next, we show that application of a locally Lipschitz map can not increase the random vector’s
description complexity. This result will allow us to construct random vectors with low description
March 20, 2018
DRAFT
8
complexity out of existing ones simply by applying locally Lipschitz mappings. The formal statement is
as follows.
Lemma II.1. Let x ∈ Rk and f : Rk → Rm be locally Lipschitz. Then, K(f (x)) ≤ K(x).
Proof.
K(f (x)) = inf{dimMB (V) : V ⊆ Rm , P[f (x) ∈ V] = 1}
(10)
= inf{dimMB (f (U)) : U ⊆ Rk , P[x ∈ U] = 1}
(11)
≤ inf{dimMB (U) : U ⊆ Rk , P[x ∈ U] = 1}
(12)
= K(x),
(13)
where (12) follows from Property iv) of Lemma H.12.
When the mapping f is locally bi-Lipschitz, the description complexity remains unchanged:
Corollary II.1. Let x ∈ Rm and f : Rm → Rm be locally bi-Lipschitz, i.e., f is invertible and both f
and f −1 are locally Lipschitz. Then, K(x) = K(f (x)).
Proof. K(x) = K((f −1 ◦ f )(x)) ≤ K(f (x)) ≤ K(x), where we applied Lemma II.1 twice.
As a consequence of Corollary II.1, the description complexity K(x) is invariant under a basis change.
Our main achievability result can now be formulated as follows.
Theorem II.1. (Achievability) For x ∈ Rm , n > K(x) is sufficient for the existence of a Borel measurable
mapping g : Rn×m × Rn → Rm , referred to as (measurable) decoder, satisfying
P[g(A, Ax) 6= x] = 0
for λn×m -a.a. A ∈ Rn×m .
(14)
Proof. See Section V.
Theorem II.1 generalizes the achievability result for linear encoders in [16] in the sense of being
nonasymptotic (i.e., it applies for finite m) and guaranteeing zero error probability.
The central conceptual element in the proof of Theorem II.1 is the following probabilistic null-space
property for arbitrary (possibly unbounded) nonempty sets, first reported in [18] for bounded sets and
expressed in terms of lower Minkowski dimension: If the lower modified Minkowski dimension of a
nonempty set U is smaller than n, then, for λn×m -a.a. measurement matrices A ∈ Rn×m , the set U
intersects the (m − n)-dimensional kernel of A at most trivially. What is remarkable here is that the
March 20, 2018
DRAFT
9
notions of Euclidean dimension (for the kernel of the mapping induced by A) and of lower modified
Minkowski dimension (for U ) are compatible. The formal statement is as follows.
Proposition II.1. Let U ⊆ Rm be nonempty with dimMB (U) < n. Then,
for λn×m -a.a. A ∈ Rn×m .
ker(A) ∩ (U \{0}) = ∅
(15)
Proof. By definition of lower modified Minkowski dimension, there exists a covering {Ui }i∈N of U by
nonempty compact sets Ui with dimB (Ui ) < n for all i ∈ N. The countable subadditivity of Lebesgue
measure now implies that
X
λn×m A ∈ Rn×m : ker(A) ∩ (U \{0}) 6= ∅ ≤
λn×m A ∈ Rn×m : ker(A) ∩ (Ui \{0}) 6= ∅ .
i∈N
(16)
The proof is concluded by noting that [18, Proposition 1] with dimB (Ui ) < n for all i ∈ N implies that
every term in the sum on the right-hand side of (16) equals zero.
We close this section by elucidating the level of generality of our theory through particularization of the
achievability result Theorem II.1 to random vectors supported on attractor sets of systems of contractions
as defined below. Such sets include the Cantor sets, Sierpinski gaskets, the modified Koch curves, and
many other fractals. For an excellent in-depth treatment of attractor sets of systems of contractions, the
interested reader is referred to [32, Section 9]. The formal definition is as follows. Let A ⊆ Rm be
closed. For i = 1, . . . , k , consider si : A → A and ci ∈ (0, 1) such that
ksi (u) − si (v)k2 ≤ ci ku − vk2
for all u, v ∈ A.
(17)
Such mappings are called contractions. By [32, Theorem 9.1], there exists a unique compact set K ⊆ A,
referred to as attractor set, such that
K=
k
[
si (K).
(18)
i=1
Thanks to [32, Proposition 9.6], dimB (K) ≤ d, where d > 0 is the unique solution of
k
X
cdi = 1.
(19)
i=1
If, in addition, K satisfies the open set condition [32, Equation (9.11)], then dimB (K) = d by [32,
Theorem 9.3]. Cantor sets, Sierpinski gaskets, and the modified Koch curves all meet the open set
condition [32, Section 9]. As dimMB (K) ≤ dimB (K) by Property iii) of Lemma H.12, every x ∈ Rm
such that P[x ∈ K] = 1 has description complexity K(x) ≤ d.
We finally note that for self-similar distributions on attractor sets K satisfying the open set condition
[32, Equation (9.11)], an achievability result in terms of information dimension was reported in [16].
March 20, 2018
DRAFT
10
III. R ECTIFIABLE S ETS AND R ECTIFIABLE R ANDOM V ECTORS
The signal models employed in classical compressed sensing [3], [4], model-based compressed sensing
[21], and block-sparsity [25]–[27] all fall under the rubric of unions of linear subspaces. More general
prevalent signal models in sparse signal recovery theory include finite unions of smooth manifolds, either
in explicit form as in [24] or implicitly in the context of low-rank matrix recovery [28]–[30]. All these
models are subsumed by the countable unions of C 1 -manifolds structure, formalized next using the notion
of rectifiable sets. We start with the definition of rectifiable sets.
Definition III.1. (Rectifiable sets) [34, Definition 3.2.14] For s ∈ N, a nonempty set U ⊆ Rm is
i) s-rectifiable if there exist a compact set A ⊆ Rs and a Lipschitz mapping ϕ : A → Rm such that
U = ϕ(A),
ii) countably s-rectifiable if it is the countable union of s-rectifiable sets,
iii) countably (H s , s)-rectifiable if it is H s -measurable and there exists a countably s-rectifiable set V
such that H s (U \V) = 0, where H s denotes the s-dimensional Hausdorff measure (cf. Definition
H.7),
iv) (H s , s)-rectifiable if it is countably (H s , s)-rectifiable and H s (U) < ∞.
Our definitions of s-rectifiability and countably s-rectifiability differ from those of [34, Definition
3.2.14] as we require the s-rectifiable set to be the Lipschitz image of a compact rather than a bounded
set. We note, however, that by [35, Theorem 7.2] our definitions of (H s , s)-rectifiability and countably
(H s , s)-rectifiability are nonetheless equivalent to those of [34, Definition 3.2.14]. The more restrictive
definitions i) and ii) above have the advantage of s-rectifiable sets and countably s-rectifiable sets
being guaranteed to be Borel. Therefore, s-rectifiable sets and countably s-rectifiable sets are also H s measurable, which leads to the following chain of implications:
U is s-rectifiable ⇒ U is countably s-rectifiable ⇒ U is countably (H s , s)-rectifiable.
The following result collects properties of (countably) s-rectifiable sets for later use.
Lemma III.1.
i) If U ⊆ Rm is s-rectifiable, then it is t-rectifiable for all t ∈ N with t > s.
ii) For locally Lipschitz ϕi , i ∈ N, the set
V=
[
ϕi (Rs )
(20)
i∈N
is countably s-rectifiable.
March 20, 2018
DRAFT
11
iii) If U ⊆ Rm is countably s-rectifiable and V ⊆ Rn is countably t-rectifiable, then
W = {(u v)T : u ∈ U, v ∈ V} ⊆ Rm+n
(21)
is countably (s + t)-rectifiable.
iv) Every s-dimensional C 1 -submanifold [36, Definition 5.3.1] of Rm is countably s-rectifiable. In
particular, every s-dimensional affine subspace of Rm is countably s-rectifiable.
v) For Ai countably si -rectifiable and si ≤ s, i ∈ N, the set
[
A=
Ai
(22)
i∈N
is countably s-rectifiable.
Proof. See Appendix A.
Countable unions of s-dimensional C 1 -submanifolds of Rm are countably s-rectifiable by Properties
iv) and v) of Lemma III.1. For countably (H s , s)-rectifiable sets we even get an equivalence result,
namely:
Theorem III.1. [34, Theorem 3.2.29] A set U ⊆ Rm is countably (H s , s)-rectifiable if and only if
H s -a.a. of U is contained in the countable union of s-dimensional C 1 -submanifolds of Rm .
We now show that the upper modified Minkowski dimension of a countably s-rectifiable set is upperbounded by s. This will allow us to conclude that, for a random vector x admitting a countably s-rectifiable
support set, K(x) ≤ s. The formal statement is as follows.
Lemma III.2. If U ⊆ Rm is countably s-rectifiable, then dimMB (U) ≤ s.
Proof. Suppose that U ⊆ Rm is countably s-rectifiable. Then, Definition III.1 implies that there exist
nonempty compact sets Ai ⊆ Rs and Lipschitz mappings ϕi : Ai → Rm , i ∈ N, such that
[
U=
ϕi (Ai ).
(23)
i∈N
Thus,
dimMB (U) = sup dimMB (ϕi (Ai ))
(24)
i∈N
≤ sup dimMB (Ai )
(25)
≤ sup dimMB (Rs )
(26)
= s,
(27)
i∈N
i∈N
March 20, 2018
DRAFT
12
where the individual steps follow from Properties of Lemma H.12, namely, (24) is by Property v), (25)
by Property iv), (26) by Property ii), and (27) by Property i).
We next investigate the effect of locally Lipschitz mappings on (countably) s-rectifiable and countably
(H s , s)-rectifiable sets.
Lemma III.3. Let U ⊆ Rm and consider a locally Lipschitz f : Rm → Rn . If U is
i) s-rectifiable, then f (U) is s-rectifiable,
ii) countably s-rectifiable, then f (U) is countably s-rectifiable,
iii) countably (H s , s)-rectifiable and Borel, then f (U) = A ∪ B , where A is a countably (H s , s)rectifiable Borel set and H s (B) = 0.
Proof. See Appendix B.
A slightly weaker version of this statement valid for Lipschitz mappings was derived previously in
[19, Lemma 4].
We are now ready to define rectifiable random vectors.
Definition III.2. (Rectifiable random vectors) [19, Definition 11] A random vector x ∈ Rm is s-rectifiable
if there exists a countably (H s , s)-rectifiable set U ⊆ Rm such that µx H s |U . The corresponding
value s is the rectifiability parameter.
It turns out that an s-rectifiable random vector x always admits a countably s-rectifiable support set
and, therefore, has description complexity K(x) ≤ s by Lemma III.2. The formal statement is as follows.
Lemma III.4. Every s-rectifiable random vector x ∈ Rm admits a countably s-rectifiable support set. In
particular, every s-rectifiable random vector x has description complexity K(x) ≤ s.
Proof. Suppose that x is s-rectifiable. Then, there exists a countably (H s , s)-rectifiable set U ⊆ Rm
such that µx H s |U . As U is countably (H s , s)-rectifiable, by Definition III.1 there exists a countably
s-rectifiable set V ⊆ Rm such that H s (U \V) = 0. Set W = (U \V) ∪ V and note that U ⊆ W implies
H s |U H s |W by monotonicity of H s . Moreover, it follows from the definition of W , the countable
additivity of H s , and H s (U \V) = 0 that H s |W = H s |V . Thus, H s |U H s |V , and µx H s |U
implies µx H s |V . Therefore, as H s |V (Rm \V) = 0, we can conclude that P[x ∈ Rm \V] = 0, which
is equivalent to P[x ∈ V] = 1, that is, the countably s-rectifiable set V is a support set of x. The proof is
now concluded by noting that K(x) ≤ dimMB (V) ≤ s, where the latter inequality follows from Lemma
III.2.
March 20, 2018
DRAFT
13
In light of Theorem III.1, an s-rectifiable random vector x ∈ Rm is supported on a countable union
of s-dimensional C 1 -submanifolds of Rm . Absolute continuity of µx with respect to the s-dimensional
Hausdorff measure is a regularity condition guaranteeing that x can not have positive probability measure
on sets of Hausdorff dimension t < s (cf. Property i) of Lemma H.13). This regularity condition together
with the sharp transition behavior of Hausdorff measures (see Figure 4 in Appendix H) implies uniqueness
of the rectifiability parameter. The corresponding formal statement is as follows.
Lemma III.5. If x ∈ Rm is s-rectifiable and t-rectifiable, then s = t.
Proof. See Appendix C.
We are now ready to particularize our achievability result to s-rectifiable random vectors.
Corollary III.1. For s-rectifiable x ∈ Rm , n > s is sufficient for the existence of a Borel measurable
mapping g : Rn×m × Rn → Rm satisfying
P[g(A, Ax) 6= x] = 0
for λn×m -a.a. A ∈ Rn×m .
(28)
Proof. Follows immediately from the general achievability result, Theorem II.1, and Lemma III.4.
Again, this result is nonasymptotic (i.e., it applies for finite m) and guarantees zero error probability.
Next, we present three examples of s-rectifiable random vectors aimed at illustrating the relationship
between the rectifiability parameter and the stochastic sparsity level defined in (8). Specifically, in the
first example the random vector’s rectifiability parameter will be seen to agree with its stochastic sparsity
level. The second example constructs an (r + t − 1)-rectifiable random vector of stochastic sparsity
level S(x) = rt for general r, t. In this case, the stochastic sparsity level rt can be much larger than
the rectifiability parameter r + t − 1, and hence Corollary III.1 implies that this random vector can be
recovered with zero error probability from a number of linear measurements that is much smaller than
its stochastic sparsity level. The third example constructs a random vector that is uniformly distributed
on a manifold, namely, the unit circle. In this case the random vector’s rectifiability parameter equals
the dimension of the manifold, whereas its stochastic sparsity level equals the dimension of the ambient
space, i.e., the random vector is not sparse at all.
Example III.1. Suppose that x = (ek1 . . . eks )z ∈ Rm , where z ∈ Rs with µz λs and k =
(k1 . . . ks )T ∈ {1, . . . , m}s satisfies k1 < · · · < ks . We first show that S(x) = s. To this end, let
U = {x ∈ Rm : kxk0 ≤ s}.
March 20, 2018
(29)
DRAFT
14
Since P[x ∈ U] = 1 by construction, it follows that S(x) ≤ s. To establish that S(x) ≥ s, and hence
S(x) = s, towards a contradiction, assume that there exists a linear subspace T ⊆ Rm of dimension
d < s such that P[x ∈ T ] > 0. Since
0 < µx (T )
(30)
= P[(ek1 . . . eks )z ∈ T ]
X
=
P[(ei1 . . . eis )z ∈ T ] P[k = (i1 . . . is )T ],
(31)
(32)
1≤i1 <···<is ≤m
there must exist a set of indices {i1 , . . . , is } ⊆ {1, . . . , m} with i1 < · · · < is such that
P[Ez ∈ T ] > 0,
(33)
where E = (ei1 . . . eis ). Next, consider the linear subspace
Te = {z ∈ Rs : Ez ∈ T } ⊆ Rs ,
(34)
which, by (33), satisfies P[z ∈ Te ] > 0. As E is obtained by removing column vectors from the identity
matrix, it follows that dim(Te ) = dim(T ) = d. Since d < s by assumption, there must exist a nonzero
T
e
e
vector b0 ∈ Rs such that bT
0 z = 0 for all z ∈ T . It follows that P[b0 z = 0] ≥ P[z ∈ T ] > 0, which
stands in contradiction to µz λs because λs ({z : bT
0 z = 0}) = 0. Thus, we indeed have S(x) = s. It
follows from Properties iv) and v) in Lemma III.1 that U in (29) is countably s-rectifiable and, therefore,
also countably (H s , s)-rectifiable. We will see in Example IV.3 that x is in fact s-analytic, which, by
Property ii) of Lemma IV.3, implies µx H s . Finally, by P[x ∈ U] = 1, we get µx H s |U , which,
thanks to countable (H s , s)-rectifiability of U establishes s-rectifiability of x.
Example III.2. Let x = a ⊗ b with a ∈ Rk and b ∈ Rl . Suppose that a = (ep1 . . . epr )u and b =
(eq1 . . . eqt )v, where u ∈ Rr and v ∈ Rt with µu × µv λr+t , and p = (p1 . . . pr )T ∈ {1, . . . , k}r
and q = (q1 . . . qt )T ∈ {1, . . . , l}t satisfy p1 < · · · < pr and q1 < · · · < qt , respectively. We first show
that S(x) = rt. Since P[kxk0 ≤ rt] = 1 by construction, it follows that S(x) ≤ rt. To establish that
S(x) ≥ rt, and hence S(x) = rt, towards a contradiction, assume that there exists a linear subspace
T ⊆ Rm of dimension d < rt such that P[x ∈ T ] > 0. Since
0 < µx (T )
(35)
= P[((ep1 . . . epr )u) ⊗ ((eq1 . . . eqt )v) ∈ T ]
(36)
= P[((ep1 . . . epr ) ⊗ (eq1 . . . eqt ))(u ⊗ v) ∈ T ]
X
=
P[((ei1 . . . eir ) ⊗ (ej1 . . . ejt ))(u ⊗ v) ∈ T ]
(37)
(38)
1≤i1 <···<ir ≤k
1≤j1 <···<jt ≤l
March 20, 2018
DRAFT
15
P[p = (i1 . . . ir )T , q = (j1 . . . jt )T ],
(39)
there must exist a set of indices {i1 , . . . , ir } ⊆ {1, . . . , k} with i1 < · · · < ir and a set of indices
{j1 , . . . , jt } ⊆ {1, . . . , l} with j1 < · · · < jt such that
P[(E1 ⊗ E2 )(u ⊗ v) ∈ T ] > 0,
(40)
where E1 = (ei1 . . . eir ) and E2 = (ej1 . . . ejt ). Next, consider the linear subspace
Te = {z ∈ Rrt : (E1 ⊗ E2 )z ∈ T } ⊆ Rrt ,
(41)
which, by (40), satisfies P[u ⊗ v ∈ Te ] > 0. As E1 ⊗ E2 is obtained by removing column vectors from
the identity matrix, it follows that dim(Te ) = dim(T ) = d. Since d < rt by assumption, there must exist
T
e
a nonzero vector b0 ∈ Rrt such that bT
0 z = 0 for all z ∈ T . It follows that P[b0 (u ⊗ v) = 0] ≥ P[u ⊗ v ∈
Te ] > 0. As µu × µv λr+t by assumption, we also have
n
o
(u
⊗
v)
=
0
> 0.
λr+t (uT v T )T : u ∈ Rr , v ∈ Rt , bT
0
(42)
We now view bT
0 (u ⊗ v) as a polynomial in the entries of u and v . Since a polynomial vanishes either
on a set of Lebesgue measure zero or is identically zero (cf. Corollary K.1 and Lemma K.5), it follows
that
bT
0 (u ⊗ v) = 0
for all u ∈ Rr and v ∈ Rt ,
(43)
which stands in contradiction to b0 6= 0. Thus, we indeed have S(x) = rt. We next construct a countably
(r + t − 1)-rectifiable support set U of x. To this end, we let
A = {a ∈ Rk : kak0 ≤ r}
(44)
B = {b ∈ Rl : kbk0 ≤ t}
(45)
and set U = A ⊗ B . Since A is a support set of a and B is a support set of b, U is a support set of
x = a ⊗ b. Note that U = (A\{0}) ⊗ B . For a ∈ A\{0}, let ā denote the first nonzero entry of a. We
can now write
a⊗b=
a
ā
⊗ (ā b)
for all a ∈ A\{0}, b ∈ B.
(46)
This allows us to decompose U according to U = Ã ⊗ B , where
à = {a ∈ A\{0} : kak0 ≤ r, ā = 1}.
March 20, 2018
(47)
DRAFT
16
Now, since à is a finite union of affine subspaces of dimensions r − 1, it is countably (r − 1)-rectifiable
by Properties iv) and v) in Lemma III.1. By the same token, B as a finite union of linear subspaces of
dimensions t is countably t-rectifiable. Therefore, the set
C = {(a b)T : a ∈ Ã, b ∈ B} ⊆ Rk+l
(48)
is countably (r + t − 1)-rectifiable thanks to Property iii) of Lemma III.1. Now, the multivariate mapping
σ : Rk+l → Rkl , (a b)T 7→ a ⊗ b is bilinear and as such locally Lipschitz. Moreover, since U = σ(C)
with C countably (r + t − 1)-rectifiable, it follows from Property ii) of Lemma III.3 that U is countably
(r + t − 1)-rectifiable and, therefore, also countably (H r+t−1 , r + t − 1)-rectifiable. We will see in
Example IV.4 that x is in fact (r + t − 1)-analytic, which, by Property ii) of Lemma IV.3, implies
µx H r+t−1 . With P[x ∈ U] = 1 this yields µx H r+t−1 |U and, in turn, thanks to countable
(H r+t−1 , r + t − 1)-rectifiability of U , establishes (r + t − 1)-rectifiability of x.
Example III.3. Let S 1 denote the unit circle in R2 , z ∈ R with µz λ1 , and g : R → S 1 , z 7→
(cos(z) sin(z))T . Set x = g(z), and note that this implies P[x ∈ S 1 ] = 1. We first establish that
S(x) = 2. Since P[x ∈ R2 ] = 1, it follows that S(x) ≤ 2. To establish that S(x) ≥ 2, and hence
S(x) = 2, towards a contradiction, assume that there exists a linear subspace T ⊆ R2 of dimension one
such that P[x ∈ T ] > 0. Set A = T ∩ S 1 , which consists of two antipodal points on S 1 (cf. Figure
2). Now, 0 < P[x ∈ T ] = P[x ∈ A] = P[z ∈ g −1 (A)], which constitutes a contradiction to µz λ1
because g −1 (A)—as a countable set—must have Lebesgue measure zero. Therefore, S(x) = 2. Finally,
x is 1-rectifiable by [19, Section III.D].
As s-rectifiable random vectors can not have positive probability measure on sets of Hausdorff dimension t < s, it is natural to ask whether taking n ≥ s linear measurements is necessary for zero error
recovery of s-rectifiable random vectors. Surprisingly, it turns out that this is not the case. This will be
demonstrated by first constructing a 2-rectifiable (and therefore also (H 2 , 2)-rectifiable) set G ⊆ R3 of
3
T
strictly positive 2-dimensional Hausdorff measure with the property that eT
3 : R → R, (x1 x2 x3 ) 7→ x3
is one-to-one on G . Then, we show that every 2-rectifiable random vector x satisfying µx H 2 |G can
be recovered with zero error probability from one linear measurement, specifically from y = eT
3 x. What
is more, all this is possible even though G contains the image of a Borel set in R2 of positive Lebesgue
measure under a C ∞ -embedding (see Figure 1 for an illustration).
The construction of our example is based on the following result.
Theorem III.2. There exist a compact set A ∈ B(R2 ) with λ2 (A) = 1/4 and a C ∞ -function κ : R2 → R
such that κ is one-to-one on A.
March 20, 2018
DRAFT
17
A ✓ Rs , L s (A) > 0
U ✓ Rm
h
h(A)
Figure 1. U contains the embedded image of a set A of positive Lebesgue measure.
Proof. See Section VI for the explicit construction of κ and A.
We now proceed to the construction of our example.
Example III.4. Let κ and A be as constructed in the proof of Theorem III.2 and consider the mapping
h : R2 → R3
(49)
z 7→ (z T κ(z))T .
(50)
We set G = h(A) and show the following:
i) h is a C ∞ -embedding;
1
ii) G is 2-rectifiable;
iii) 0 < H 2 (G) < ∞;
3
T
iv) eT
3 : R → R, (x1 x2 x3 ) 7→ x3 is one-to-one on G .
v) For every 2-rectifiable random vector x ∈ R3 with µx H 2 |G , there exists a Borel measurable
mapping g : R3 × R → R3 satisfying P g e3 , eT
3 x 6= x = 0.
It follows immediately that h is one-to-one. Thus, to establish Property i), it suffices to prove that h is
a C ∞ -immersion. Since κ is C ∞ , so is h. Furthermore,
q
Jh(z) = det((Dh(z))T Dh(z))
q
= det(I2 + a(z)a(z)T ) for all z ∈ R2 ,
(51)
(52)
where
a(z) =
∂κ(z)
∂z1 .
∂κ(z)
∂z2
Since a(z)a(z)T is positive semidefinite, [37, Corollary 7.7.4] implies Jh(z) ≥
(53)
p
det(I2 ) = 1 for all
z ∈ R2 , which establishes that h is an immersion and completes the proof of i).
March 20, 2018
DRAFT
18
To prove ii), note that h is C ∞ and as such locally Lipschitz. As A is compact, Lemma H.9 implies
that h|A is Lipschitz. The set G = h(A) is hence the Lipschitz image of a compact set in R2 and as
such 2-rectifiable.
To establish iii), we first note that
H 2 (G) = H 2 (h(A))
(54)
≤ L2 H 2 (A)
(55)
= L2 λ2 (A)
(56)
< ∞,
(57)
where the individual steps follow from Properties of Lemma H.13, namely, (55) from Property ii) with
L denoting the Lipschitz constant of h|A , and (56) from Property iii).
To establish H 2 (G) > 0, consider the linear mapping
π : R3 → R2
(58)
(x1 x2 x3 )T 7→ (x1 x2 )T .
(59)
Clearly, π is Lipschitz with Lipschitz constant equal to 1. Therefore,
H 2 (G) ≥ H 2 (π(G))
(60)
= H 2 (A)
(61)
= λ2 (A)
(62)
1
= ,
4
(63)
where (60) follows from Property ii) of Lemma H.13, (61) from π(G) = A, and (62) is by Property iii)
of Lemma H.13.
To show iv), let x1 , x2 ∈ G with x1 6= x2 . Thus, x1 = (z1T κ(z1 ))T and x2 = (z2T κ(z2 ))T with
z1 , z2 ∈ A and z1 6= z2 . As κ is one-to-one on A, we conclude that eT
3 is one-to-one on G .
It remains to establish v). Since µx H 2 |G by assumption, it follows that P[x ∈ G] = 1. We show
that there exists a Borel measurable mapping g : R3 × R → R3 such that
∈ v ∈ G : aT v = y
if ∃v ∈ G : aT v = y
g(a, y)
= e
else,
March 20, 2018
(64)
DRAFT
19
where e is an arbitrary vector not in G , used to declare a decoding error. Since P[x ∈ G] = 1 and eT
3
T
is one-to-one on G , this then implies P g e3 , e3 x 6= x = 0. To construct g in (64), consider first the
mapping
f : R3 × R × R3 → R
(a, y, u) 7→ |y − aT u|.
(65)
(66)
Since f is continuous, Lemma I.3 implies that f is a normal integrand (see Definition I.4) with respect
to B(R3 × R). Let
T = {(a, y) ∈ R3 × R : ∃u ∈ G with f (a, y, u) ≤ 0}
= {(a, y) ∈ R3 × R : ∃u ∈ G with aT u = y}.
(67)
(68)
Note that G as the Lipschitz image of the compact set A is compact (cf. Lemma H.10). It now follows
from Properties ii) and iii) of Lemma I.5 (with T = R3 × R, α = 0, K = G , and f as in (65)–(66),
which is a normal integrand with respect to B(R3 × R)) that i) T ∈ B(R3 × R) and ii) there exists a
Borel measurable mapping
p : T → R3
a, y 7→ p a, y ∈ u ∈ G : aT u = y .
(69)
(70)
This mapping can then be extended to a mapping g : R3 × R → R3 by setting
g|T = p
g|(R3 ×R)\T = e.
(71)
(72)
Finally, g is Borel measurable thanks to Lemma H.6 as p is Borel measurable and T ∈ B(R3 × R).
IV. S TRONG C ONVERSE
Example III.4 in the previous section demonstrates that n ≥ s is not necessary for zero error recovery
of s-rectifiable random vectors in general. In this section, we introduce the class of s-analytic random
vectors x, which will be shown to allow for a strong converse in the sense of n ≥ s being necessary for
recovery of x with probability of error smaller than one. The adjective “strong” refers to the fact that
n < s linear measurements are insufficient even if we allow a recovery error probability arbitrarily close
to one. We prove that an s-analytic random vector is s-rectifiable if and only if it admits a support set
U that is “not too rich” (in terms of σ -finiteness of H s |U ), show that the s-rectifiable random vectors
considered in Examples III.1–III.3 are all s-analytic, and discuss examples of s-analytic random vectors
March 20, 2018
DRAFT
20
that fail to be s-rectifiable. Random vectors that are both s-analytic and s-rectifiable can be recovered
with zero error probability from n > s linear measurements, and n ≥ s linear measurements are necessary
for recovery with error probability smaller than one. The border case n = s remains open.
We now make our way towards developing the strong converse and the formal definition of s-analyticity.
The following auxiliary result will turn out useful.
Lemma IV.1. For x ∈ Rm and A ∈ Rn×m , consider the following statements:
i) There exists a Borel measurable mapping g : Rn×m × Rn → Rm satisfying P[g(A, Ax) 6= x] < 1.
ii) There exists a U ∈ B(Rm ) with P[x ∈ U] > 0 such that A is one-to-one on U .
Then, i) implies ii).
Proof. See Appendix D.
We first establish a strong converse for the class of random vectors considered in Example III.1. This
will guide us to the crucial defining property of s-analytic random vectors.
Lemma IV.2. Let x = (ek1 . . . eks )z ∈ Rm , where z ∈ Rs with µz λs and k = (k1 . . . ks )T ∈
{1, . . . , m}s satisfies k1 < · · · < ks . If there exist a measurement matrix A ∈ Rn×m and a Borel
measurable mapping g : Rn×m × Rn → Rm such that P[g(A, Ax) 6= x] < 1, then n ≥ s.
Proof. Towards a contradiction, suppose that there exist a measurement matrix A ∈ Rn×m and a Borel
measurable mapping g : Rn×m × Rn → Rm so that P[g(A, Ax) 6= x] < 1 for n < s. By Lemma IV.1,
there then must exist a U ∈ B(Rm ) with P[x ∈ U] > 0 such that A is one-to-one on U . Since
P[x ∈ U] = P[(ek1 . . . eks )z ∈ U]
X
=
P[(ei1 . . . eis )z ∈ U] P[k = (i1 . . . is )T ],
(73)
(74)
1≤i1 <···<is ≤m
there must exist a set of indices {i1 , . . . , is } ⊆ {1, . . . , m} with i1 < · · · < is such that the rank-
s matrix H := (ei1 . . . eis ) satisfies P[Hz ∈ U] > 0. Setting A = {z ∈ Rs : Hz ∈ U } yields
µz (A) = P[Hz ∈ U] > 0. Furthermore, A as the inverse image of the Borel set U under a linear
mapping is Borel. Finally, since µz λs by assumption, we conclude that λs (A) > 0. Summarizing,
there exist a set A ∈ B(Rs ) with λs (A) > 0 and a matrix H ∈ Rm×s such that U contains the one-toone image of A under H . Since A is one-to-one on U and H is one-to-one on Rs , it follows that AH
is one-to-one on A, i.e.,
ker(AH) ∩ (A
March 20, 2018
A) = {0}.
(75)
DRAFT
21
x2
Th
S1
x1
Figure 2. For every vector h ⊆ R2 \{0}, the linear subspace Th := {hz : z ∈ R} intersects the unit circle S1 in the two
antipodal points ±h/khk2 .
Finally, since λs (A) > 0, the Steinhaus Theorem [38] implies the existence of an r > 0 such that
Bs (0, r) ⊆ A
A ⊆ Rs . Since dim ker(AH) ≥ s − n > 0, we conclude that the linear subspace
ker(AH) must have a nontrivial intersection with A
A, which stands in contradiction to (75) and
thereby finishes the proof.
The strong converse just derived hinges critically on the specific structure of the s-rectifiable random
vector x considered. Concretely, we used the fact that, for every U ∈ B(Rm ) with P[x ∈ U] > 0, there
exist a set A ∈ B(Rs ) with λs (A) > 0 and a matrix H ∈ Rm×s such that U contains the one-to-one
image of A under H . The following example demonstrates, however, that this property is too strong for
our purposes as it fails to hold for random vectors on general manifolds like, e.g., the unit circle:
Example IV.1. Let S 1 ⊆ R2 denote the unit circle and consider x ∈ R2 supported on S 1 , i.e., P[x ∈
S 1 ] = 1. Towards a contradiction, suppose that there exist a set A ∈ B(R) with λ1 (A) > 0 and a vector
h ∈ R2 such that {hz : z ∈ A} ⊆ S 1 and h is one-to-one on A. Since h is one-to-one on A and
λ1 (A) > 0, it follows that h 6= 0. Noting that {hz : z ∈ A} ⊆ {h/khk2 , −h/khk2 } (cf. Figure 2), A
necessarily satisfies A ⊆ {1/khk2 , −1/khk2 }. Thus, λ1 (A) = 0, which is a contradiction to λ1 (A) > 0.
The reason for this failure is that every h ∈ R2 maps into a 1-dimensional linear subspace in R2 ,
and 1-dimensional linear subspaces in R2 intersect the unit circle in two antipodal points only. To map
a set A ∈ B(R) to a set in R2 that is not restricted to be a subset of a 1-dimensional linear subspace,
we have to employ a nonlinear mapping. But this puts us into the same dilemma as in Example III.4,
March 20, 2018
DRAFT
22
where we demonstrated that even the requirement of every U ∈ B(Rm ) with P[x ∈ U] > 0 containing
the embedded image—under a C ∞ -mapping—of a set A ∈ B(Rs ) of positive Lebesgue measure is not
sufficient to obtain a strong converse for general x. We therefore need to impose additional constraints
on the mapping. It turns out that requiring real analyticity is enough. Examples of real analytic mappings
include, e.g., multivariate polynomials, the exponential function, or trigonometric mappings. This finally
leads us to the new concept of s-analytic measures and s-analytic random vectors.
Definition IV.1. (Analytic measures) A Borel measure µ on Rm is s-analytic, with s ∈ {1, . . . , m}, if,
for each U ∈ B(Rm ) with µ(U) > 0, there exist a set A ∈ B(Rs ) of positive Lebesgue measure and a
real analytic mapping h : Rs → Rm of s-dimensional Jacobian Jh 6≡ 0 such that h(A) ⊆ U .
Note that the only requirement on the real analytic mappings in Definition IV.1 is that their sdimensional Jacobians do not vanish identically. Since the s-dimensional Jacobian of a real analytic
mapping is a real analytic function, it vanishes either identically or on a set of Lebesgue measure zero
(cf. Lemma K.7). By Lemma K.8, this guarantees that, for an analytic measure µ, every U ∈ B(Rm )
with µ(U) > 0 contains the real analytic embedding of a set A ∈ B(Rs ) of positive Lebesgue measure.
We have the following properties of s-analytic measures.
Lemma IV.3. If µ is an s-analytic measure on Rm , then the following holds:
i) µ is t-analytic for all t ∈ {1, . . . , s − 1};
ii) µ H s ;
iii) there exists a set U ⊆ Rm such that µ = µ|U and H s |U is σ -finite if and only if there exists a
countably (H s , s)-rectifiable set W ⊆ Rm such that µ = µ|W .
Proof. See Appendix E.
We are now ready to define s-analytic random vectors.
Definition IV.2. (Analytic random vectors) A random vector x ∈ Rm is s-analytic if µx is s-analytic.
The corresponding value s is the analyticity parameter.
We have the following immediate consequence of Lemma IV.3.
Corollary IV.1. Let x be s-analytic. Then, x is s-rectifiable if and only if it admits a support set U such
that H s |U is σ -finite.
Proof. Follows from Properties ii) and iii) in Lemma IV.3 and Definition III.2.
March 20, 2018
DRAFT
23
By Corollary IV.1, an s-analytic random vector is s-rectifiable if and only if it admits a support set U
that is “not too rich” (in terms of σ -finiteness of H s |U ). As an example of an s-analytic random vector
that is not s-rectifiable, consider an (s+1)-analytic random vector x with s > 0. By Property i) in Lemma
IV.3, this x is also s-analytic, but it can not be s-rectifiable, as shown next. Towards a contradiction,
suppose that x is s-rectifiable. Then, by Lemma III.4, x has a countably s-rectifiable support set U ,
which by Property i) in Lemma III.1 is also countably (s + 1)-rectifiable. As, by assumption, x is (s + 1)analytic, Property ii) in Lemma IV.3 implies µx H s+1 . Thus, µx H s+1 |U with U countably
(s + 1)-rectifiable, and we conclude that x would also be (s + 1)-rectifiable, which contradicts uniqueness
of the rectifiability parameter, as guaranteed by Lemma III.5.
We just demonstrated that s-analytic random vectors can not be s-rectifiable if they are also (s + 1)analytic. The question now arises whether s-analytic random vectors that fail to be (s + 1)-analytic (and,
therefore, fail to be t-analytic for all t > s by Property i) in Lemma IV.3) are necessarily s-rectifiable.
The next example shows that this is not the case.
Example IV.2. Let C be the middle third Cantor set [32, p. xii] and consider U = {(c t)T : c ∈ C, t ∈
[0, 1]} ⊆ R2 . Since 0 < H ln 2/ ln 3 (C) < ∞ [32, Example 4.5], it follows that the random vector x with
distribution µx = π × (λ1 |[0,1] ), where π = H ln 2/ ln 3 |C /H ln 2/ ln 3 (C), is well defined. We now show
that
i) x is 1-analytic;
ii) x is not 2-analytic;
iii) x is not 1-rectifiable.
To establish i), consider B ∈ B(R2 ) with P[x ∈ B] > 0. Now,
0 < µx (B)
Z
= λ1 t ∈ [0, 1] : (c t)T ∈ B dπ(c)
ZC
≤ λ1 t ∈ R : (c t)T ∈ B dπ(c),
(76)
(77)
(78)
C
where in (77) we applied Corollary H.1 (with the finite measure spaces (R, B(R), π) and (R, B(R), λ1 |[0,1] )
and (78) is by monotonicity of Lebesgue measure. Thus, by Lemma H.2, there must exist a c0 ∈ C such
that A := t ∈ R : (c0 t)T ∈ B satisfies λ1 (A) > 0. Now, define the mapping h : R → R2 , t 7→ (c0 t)T
and note that this mapping is (trivially) real analytic with Jh ≡ 1. Moreover, h(A) ⊆ B by construction,
and A is Borel measurable as the inverse image of the Borel set B under the real analytic and, therefore,
continuous mapping h. Thus, x is 1-analytic.
March 20, 2018
DRAFT
24
We next show that x is not 2-analytic. Towards a contradiction, suppose that x is 2-analytic. Since
µx (U) = 1, by 2-analyticity of x, there must exist a set D ∈ B(R2 ) with λ2 (D) > 0 and a real-analytic
mapping g : R2 → R2 of 2-dimensional Jacobian Jg 6≡ 0 such that g(D) ⊆ U . By Property ii) in Lemma
K.8, we can assume, without loss of generality (w.l.o.g.), that g|D is an embedding. It follows that
Z
2
H (g(D)) =
Jg(z)dλ2 (z)
(79)
D
> 0,
(80)
where in (79) we applied the area formula Corollary H.3 upon noting that g|D is one-to-one as an
embedding and locally Lipschitz by real analyticity of g , and (80) is by Lemma H.2, λ2 (D) > 0, and
Jg(z) > 0 for all z ∈ D. Since g(D) ⊆ U and H 2 (g(D)) > 0, monotonicity of H 2 yields H 2 (U) > 0.
Upon noting that H 1+ln 2/ ln 3 (U) < ∞ [32, Example 4.3], this results in a contradiction to Property i)
in Lemma H.13.
Finally, to establish iii), towards a contradiction, suppose that x is 1-rectifiable. Then, Lemma III.4
implies that x admits a countably 1-rectifiable support set. As every countably 1-rectifiable set is the
countable union of 1-rectifiable sets, the union bound implies that there must exist a 1-rectifiable set V
with P[x ∈ V] > 0. By Definition III.1, there must therefore exist a compact set K ⊆ R and a Lipschitz
mapping f : K → R2 such that V = f (K). It follows that
0 < µx (V)
(81)
= µx (f (K))
Z
= λ1 t ∈ [0, 1] : (c t)T ∈ f (K) dπ(c),
ZC
= λ1 t ∈ [0, 1] : (c t)T ∈ f (Ac ) dπ(c),
(82)
(83)
(84)
C
where in (83) we applied Corollary H.1 (with the finite measure spaces (R, B(R), π) and (R, B(R), λ1 |[0,1] )
and in (84) we set, for every c ∈ C ,
Ac = f −1
(c t)T : t ∈ [0, 1]
⊆ K.
(85)
Note that the sets Ac ⊆ K are pairwise disjoint as inverse images of pairwise disjoint sets. Now, Lemma
H.2 together with (81)–(84) implies that there must exist a set F ⊆ C with π(F) > 0 such that
λ1
March 20, 2018
t ∈ [0, 1] : (c t)T ∈ f (Ac )
>0
for all c ∈ F.
(86)
DRAFT
25
Since π = H ln 2/ ln 3 |C /H ln 2/ ln 3 (C) and 0 < π(F) ≤ 1, the definition of Hausdorff dimension (cf.
Definition H.8) implies dimH (F) = ln 2/ ln 3. As every countable set has Hausdorff dimension zero [32,
p. 29], we conclude that F must be uncountable. Moreover,
λ1 (Ac ) = H 1 (Ac )
(87)
1 1
H (f (Ac ))
L
1
≥ H 1 t ∈ [0, 1] : (c t)T ∈ f (Ac )
L
1
= λ1 t ∈ [0, 1] : (c t)T ∈ f (Ac )
L
≥
>0
(88)
(89)
(90)
for all c ∈ F,
(91)
where (87) and (90) follow from Property iii) in Lemma H.13, (88) is by Property ii) in Lemma H.13
with L the Lipschitz constant of f , (89) is again by Property ii) in Lemma H.13 with the Lipschitz
2
T
constant of the projection eT
2 : R → R, (c t) 7→ t equal to one, and in (91) we used (86). As the sets
Ac are pairwise disjoint subsets of positive Lebesgue measure of the compact set K, it follows that
sup
X
E⊆F :|E|<∞ c∈E
λ1 (Ac ) ≤ λ1 (K) < ∞,
(92)
which, by Lemma H.14, contradicts the uncountability of F . Therefore, x can not be 1-rectifiable.
Our strong converse for analytic random vectors will be based on the following result.
Theorem IV.1. Let A ∈ B(Rs ) be of positive Lebesgue measure, h : Rs → Rm , with s ≤ m, real
analytic of s-dimensional Jacobian Jh 6≡ 0, and f : Rm → Rn real analytic. If f is one-to-one on h(A),
then n ≥ s.
Proof. See Section VII.
With the help of Theorem IV.1, we can now prove the strong converse for s-analytic random vectors.
Corollary IV.2. For x ∈ Rm s-analytic, n ≥ s is necessary for the existence of a measurement matrix
A ∈ Rn×m and a Borel measurable mapping g : Rn×m × Rn → Rm such that P[g(A, Ax) 6= x] < 1.
Proof. Suppose, to the contrary, that there exist a measurement matrix A ∈ Rn×m and a Borel measurable
mapping g : Rn×m × Rn → Rm satisfying P[g(A, Ax) 6= x] < 1 for n < s. Then, by Lemma IV.1, there
must exist a U ∈ B(Rm ) with P[x ∈ U] > 0 such that A is one-to-one on U . As P[x ∈ U] > 0, the
s-analyticity of µx implies the existence of a set A ∈ B(Rs ) of positive Lebesgue measure along with
a real analytic mapping h : Rs → Rm of s-dimensional Jacobian Jh 6≡ 0 such that h(A) ⊆ U . As A is
March 20, 2018
DRAFT
26
one-to-one on h(A) and linear mappings are trivially real-analytic, Theorem IV.1 implies that we must
have n ≥ s, which contradicts n < s.
We next show that the s-rectifiable random vectors considered in Examples III.1–III.3 are all s-analytic
with the analyticity parameter equal to the corresponding rectifiability parameter. We need the following
result, which states that real analytic immersions preserve analyticity in the following sense.
Lemma IV.4. If x ∈ Rm is s-analytic and f : Rm → Rk , with m ≤ k , is a real analytic immersion, then
f (x) is s-analytic.
Proof. See Appendix F.
Example IV.3. We show that x in Example III.1 is s-analytic. To this end, we consider an arbitrary
but fixed U ∈ B(Rm ) with µx (U) > 0 and establish the existence of a set A ∈ B(Rs ) of positive
Lebesgue measure and a real analytic mapping h : Rs → Rm of s-dimensional Jacobian Jh 6≡ 0 such
that h(A) ⊆ U . Since
0 < µx (U)
(93)
= P[(ek1 . . . eks )z ∈ U]
X
=
P[(ei1 . . . eis )z ∈ U] P[k = (i1 . . . is )T ],
(94)
(95)
1≤i1 <···<is ≤m
there must exist a set of indices {i1 , . . . , is } ⊆ {1, . . . , m} with i1 < · · · < is such that
P[u(z) ∈ U] > 0,
(96)
where u : Rs → Rm , z 7→ (ei1 . . . eis )z . As µz λs by assumption, z is s-analytic thanks to Lemma
IV.5 below. The mapping u is linear and, therefore, trivially real analytic. Furthermore,
r
Ju(z) = det (ei1 . . . eis )T (ei1 . . . eis )
=
p
=1
det Is
for all z ∈ Rs ,
(97)
(98)
(99)
where (97) follows from Du(z) = (ei1 . . . eis ) for all z ∈ Rs , which proves that u is an immersion. We
can therefore employ Lemma IV.4 and conclude that u(z) is s-analytic. Hence, Definition IV.2 together
with (96) implies that there must exist a set A ∈ B(Rs ) of positive Lebesgue measure and a real analytic
mapping h : Rs → Rm of s-dimensional Jacobian Jh 6≡ 0 such that h(A) ⊆ U .
Lemma IV.5. If x ∈ Rm with µx λm , then x is m-analytic.
March 20, 2018
DRAFT
27
Proof. We have to show that, for each U ∈ B(Rm ) with µx (U) > 0, we can find a set A ∈ B(Rm )
of positive Lebesgue measure and a real analytic mapping h : Rm → Rm of m-dimensional Jacobian
Jh 6≡ 0 such that h(A) ⊆ U . For given such U ∈ B(Rm ), simply take A = U and h the identity mapping
on Rm .
Example IV.4. We show that x = a ⊗ b ∈ Rkl as in Example III.2 is (r + t − 1)-analytic. To this
end, let U ∈ B(Rkl ) with µx (U) > 0 be arbitrary but fixed. We have to establish that there exist a
set A ∈ B(Rr+t−1 ) of positive Lebesgue measure and a real analytic mapping h : Rr+t−1 → Rkl of
(r + t − 1)-dimensional Jacobian Jh 6≡ 0 such that h(A) ⊆ U . Since
0 < µx (U)
(100)
= P[((ep1 . . . epr )u) ⊗ ((eq1 . . . eqt )v) ∈ U]
(101)
= P[((ep1 . . . epr ) ⊗ (eq1 . . . eqt ))(u ⊗ v) ∈ U]
X
P[((ei1 . . . eir ) ⊗ (ej1 . . . ejt ))(u ⊗ v) ∈ U]
=
(102)
(103)
1≤i1 <···<ir ≤k
1≤j1 <···<jt ≤l
P[p = (i1 . . . ir )T , q = (j1 . . . jt )T ],
(104)
there must exist a set of indices {i1 , . . . , ir } ⊆ {1, . . . , m} with i1 < · · · < ir and a set of indices
{j1 , . . . , jt } ⊆ {1, . . . , m} with j1 < · · · < jt such that
P[v(u ⊗ v) ∈ U] > 0,
(105)
where
v : Rrt → Rkl
w 7→ ((ei1 . . . eir ) ⊗ (ej1 . . . ejt ))w.
(106)
(107)
Since µu ×µv λr+t by assumption, it follows from Lemma IV.6 below that u⊗v is (r +t−1)-analytic.
The mapping v is linear and, therefore, trivially real analytic. Furthermore,
r
Jv(w) = det (E1 ⊗ E2 )T (E1 ⊗ E2 )
r
=
=
p
=1
March 20, 2018
(108)
det (E1T E1 ) ⊗ (E2T E2 )
(109)
det(Ir ⊗ It )
(110)
for all w ∈ Rrt ,
(111)
DRAFT
28
where (108) follows from Dv(w) = E1 ⊗ E2 , with E1 = (ei1 . . . eir ) and E2 = (ej1 . . . ejt ), for all
w ∈ Rrt , which proves that v is an immersion. We can therefore employ Lemma IV.4 and conclude that
v(u ⊗ v) is (r + t − 1)-analytic. Hence, Definition IV.2 together with (105) implies that there must exist
a set A ∈ B(Rr+t−1 ) of positive Lebesgue measure and a real analytic mapping h : Rr+t−1 → Rkl of
(r + t − 1)-dimensional Jacobian Jh 6≡ 0 such that h(A) ⊆ U .
Lemma IV.6. If a ∈ Rk and b ∈ Rl with µa × µb λk+l , then a ⊗ b is (k + l − 1)-analytic.
Proof. See Appendix G.
Example IV.5. Let x, z, and h be as in Example III.3. We first note that sin and cos are real analytic.
In fact, it follows from the ratio test [39, Theorem 3.34] that the power series
sin(z) =
cos(z) =
∞
X
(−1)n z 2n+1
n=0
∞
X
n=0
(2n + 1)!
(−1)n z 2n
(2n)!
(112)
(113)
are absolutely convergent for all z ∈ R. Thus, sin and cos can both be represented by convergent power
series at 0 with infinite convergence radius. Lemma K.1 therefore implies that sin and cos are both real
p
analytic. As each component of h is real analytic, so is h. Furthermore, Jh(z) = sin2 (z) + cos2 (z) = 1
for all z ∈ R, which implies that h is a real analytic immersion. Since z is 1-analytic by Lemma IV.5
and x = h(z), Lemma IV.4 implies that x is 1-analytic.
V. P ROOF OF T HEOREM II.1 (ACHIEVABILITY )
Suppose that K(x) < n. It then follows from (9) that x must admit a support set U ⊆ Rm with
dimMB (U) < n. We first construct a new support set V ⊆ Rm for x as a countable union of compact sets
satisfying dimMB (V) < n. Based on this support set V we will then prove the existence of a measurable
decoder g satisfying P[g(A, Ax) 6= x] = 0. The construction of V starts by noting that, thanks to
Definition II.2, dimMB (U) < n implies the existence of a covering {Ui }i∈N of U by nonempty compact
sets Ui satisfying
sup dimB (Ui ) < n.
(114)
i∈N
For this covering, we set
V=
March 20, 2018
[
i∈N
Ui ,
(115)
DRAFT
29
and note that
)
(
dimMB (V) = inf sup dimB (Vi ) : V ⊆
i∈N
[
i∈N
Vi
(116)
≤ sup dimB (Ui )
(117)
< n,
(118)
i∈N
where (116) follows from Definition II.2 with the infimum taken over all coverings {Vi }i∈N of V by
nonempty compact sets Vi , (117) is by (115), and in (118) we used (114). Since dimMB (Rm ) = m by
Property i) of Lemma H.12, and dimMB (V) < n ≤ m, we must have V ( Rm . Now, V is a support set
because it contains the support set U as a subset. Furthermore, since P[x ∈ V] = 1 and V ( Rm , there
must exist an e ∈ Rm \V such that P[x = e] = 0. This e will be used to declare a decoding error. We
will show in Section V-A that there exists a Borel measurable mapping g : Rn×m × Rn → Rm satisfying
∈ {v ∈ V : Av = y} if ∃v ∈ V : Av = y
g(A, y)
(119)
= e
else.
The mapping g is guaranteed to deliver a v ∈ V that is consistent with (A, y) (in the sense of Av = y ) if
at least one such consistent v ∈ V exists, otherwise an error is declared by delivering the “error symbol”
e. Next, for each A ∈ Rn×m , let pe (A) denote the probability of error defined as
pe (A) = P[g(A, Ax) 6= x].
(120)
It remains to show that pe (A) = 0 for λn×m -a.a. A. Now,
pe (A) = P[g(A, Ax) 6= x, x ∈ V] + P[g(A, Ax) 6= x, x ∈
/ V]
= P[g(A, Ax) 6= x, x ∈ V]
= P[(A, x) ∈ A]
for all A ∈ Rn×m ,
(121)
(122)
(123)
where (122) follows from P[x ∈ V] = 1 and in (123) we set
A = {(A, x) ∈ Rn×m × V : g(A, Ax) 6= x}.
(124)
Since A ∈ B(Rn×m )⊗B(Rm ) by Lemma V.1 below (with X = Rm , Y = Rn×m , f (x, A) = g(A, Ax),
and V ∈ B(Rm ) as a countable union of compact sets), we can apply Corollary H.1 (with the σ -finite
measure spaces (Rm , B(Rm ), µx ) and (Rn×m , B(Rn×m ), λn×m )) to A and get
Z
Z
pe (A)dλn×m (A) =
λn×m ({A : (A, x) ∈ A})dµx (x).
Rn×m
March 20, 2018
(125)
Rm
DRAFT
30
Next, note that for y = Ax with x ∈ V , the vector g(A, y) can differ from x only if there is a v ∈ V \{x}
that is consistent with y , i.e., if y = Av for some v ∈ V \{x}. Thus,
A ⊆ {(A, x) ∈ Rn×m × V : ker(A) ∩ Vx 6= {0}},
(126)
where, for each x ∈ V , we set
Vx = {v − x : v ∈ V}.
(127)
{A ∈ Rn×m : (A, x) ∈ A} ⊆ A ∈ Rn×m : ker(A) ∩ Vx 6= {0} ,
(128)
As (126) yields
monotonicity of λn×m implies
λn×m ({A ∈ Rn×m : (A, x) ∈ A})
≤ λn×m A ∈ Rn×m : ker(A) ∩ Vx 6= {0}
(129)
for all x ∈ V.
(130)
The null-space property Proposition II.1, with U = Vx and dimMB (Vx ) = dimMB (V) < n ((lower)
modified Minkowski dimension is invariant under translation, as seen by translating covering balls
accordingly) now implies that (130) equals zero for all x ∈ V . Therefore, (129) must equal zero as
well for all x ∈ V . We conclude that both sides of (125) must equal zero as the integrand on the righthand side is identically zero (recall that V is a support set of x), which, by Lemma H.2, implies that we
must have pe (A) = 0 for λn×m -a.a. A, thereby completing the proof.
A. Existence of Borel Measurable g
S
Recall that i) V = i∈N Ui ( Rm , where Ui ⊆ Rm is nonempty and compact for all i ∈ N and ii) the
error symbol e ∈ Rm\V . We have to show that there exists a Borel measurable mapping g : Rn×m × Rn →
Rm such that
g(A, y)
∈ {v ∈ V : Av = y}
if ∃v ∈ V : Av = y
= e
else.
(131)
To this end, first consider the mapping
f : Rn×m × Rn × Rm → R
(A, y, v) 7→ ky − Avk2 .
March 20, 2018
(132)
(133)
DRAFT
31
Since f is continuous, Lemma I.3 implies that f is a normal integrand (see Definition I.4) with respect
to B(Rn×m × Rn ). For each i ∈ N, let
Ti = {(A, y) ∈ Rn×m × Rn : ∃u ∈ Ui with f (A, y, u) ≤ 0}
= {(A, y) ∈ Rn×m × Rn : ∃u ∈ Ui with Au = y}.
(134)
(135)
It now follows from Properties ii) and iii) of Lemma I.5 (with T = Rn×m × Rn , α = 0, K = Ui , and f as
in (132)–(133), which is a normal integrand with respect to B(Rn×m × Rn )) that i) Ti ∈ B(Rn×m × Rn )
for all i ∈ N and ii) for every i ∈ N, there exists a Borel measurable mapping
pi : Ti → Rm
(136)
(A, y) 7→ pi (A, y) ∈ {u ∈ Ui : Au = y}.
(137)
For each i ∈ N, the mapping pi can be extended to a mapping gi : Rn×m × Rn → Rm by setting
gi |Ti = pi
(138)
gi |(Rn×m ×Rn )\Ti = e,
(139)
which is Borel measurable thanks to Lemma H.6 as pi is Borel measurable and Ti ∈ B(Rn×m × Rn ).
Based on this sequence {gi }i∈N of Borel measurable mappings gi , we now construct a Borel measurable
mapping satisfying (131). The idea underlying this construction is as follows. For a given pair (A, y),
we first use g1 to try to find a consistent (in the sense of y = Au) u ∈ U1 . If g1 delivers the error
symbol e, we use g2 to try to find a consistent u ∈ U2 . This procedure is continued until a gi delivers a
consistent u ∈ Ui . If no gi yields a consistent u ∈ Ui , we deliver the error symbol e as the final decoder
output. The formal construction is as follows. We set G1 = g1 and, for every i ∈ N \{1}, we define the
mapping Gi : Rn×m × Rn → Rm iteratively by setting
Gi−1 (A, y) if Gi−1 (A, y) 6= e
Gi (A, y) =
gi (A, y)
else.
(140)
Then, G1 (= g1 ) is Borel measurable, and, for each i ∈ N \{1}, the Borel-measurability of Gi follows
from the Borel-measurability of Gi−1 and gi thanks to Lemma V.2 below. Note that by construction
n
o
S
S
∈ v ∈ i Uj : Av = y
if ∃v ∈ ij=1 Uj : Av = y
j=1
Gi (A, y)
(141)
= e
else.
Finally, we obtain g : Rn×m × Rn → Rm according to
g(A, y) = lim Gi (A, y),
i→∞
March 20, 2018
(142)
DRAFT
32
which satisfies (131) by construction. As the pointwise limit of a sequence of Borel measurable mappings,
g is Borel measurable thanks to Corollary H.2.
Lemma V.1. Let X and Y be Euclidean spaces, consider a Borel measurable mapping
f : X × Y → X,
(143)
A = {(x, y) ∈ V × Y : f (x, y) 6= x} ∈ B(X × Y) = B(X ) ⊗ B(Y).
(144)
and let V ∈ B(X ). Then,
Proof. We first note that B(X × X ) = B(X ) ⊗ B(X ) and B(X × Y) = B(X ) ⊗ B(Y), both thanks to
Lemma H.3. Therefore, V × X ∈ B(X × X ). Now, consider the diagonal D = {(x, x) : x ∈ X } and note
that D as the inverse image of {0} under the Borel measurable mapping g : X × X → X , (u, v) 7→ u − v
is in B(X × X ). Let C = (V × X ) ∩ ((X × X )\D). Since D ∈ B(X × X ), it follows that C ∈ B(X × X ).
Define the mapping
F: X ×Y →X ×X
(x, y) 7→ (x, f (x, y)),
(145)
(146)
and note that it is Borel measurable thanks to Lemma H.5 (with f1 : X × Y → X , (x, y) 7→ x and
f2 : X × Y → X , (x, y) 7→ f (x, y)). Finally, A ∈ B(X × Y) as A = F −1 (C).
Lemma V.2. Let X and Y be topological spaces and y0 ∈ Y and suppose that f, g : X → Y are both
Borel measurable. Then, h : X → Y ,
h(x) =
f (x)
g(x)
if f (x) 6= y0
(147)
else
is Borel measurable.
Proof. We have to show that h−1 (U) ∈ B(X ) for all U ∈ B(Y). To this end, consider an arbitrary but
fixed U ∈ B(Y). Now, {y0 } ∈ B(Y) implies U \{y0 } ∈ B(Y). We write h−1 (U) = A ∪ B with
A = {x ∈ X : h(x) ∈ U, f (x) 6= y0 }
(148)
B = {x ∈ X : h(x) ∈ U, f (x) = y0 }
(149)
and show that A and B are both in B(X ), which in turn implies h−1 (U) ∈ B(X ). Since
A = {x ∈ X : f (x) ∈ U, f (x) 6= y0 }
= f −1 (U \{y0 }),
March 20, 2018
(150)
(151)
DRAFT
33
z2
1
Q2,3
Q2,4
Q2,1
Q2,2
1
z1
Figure 3. The set Q2 consists of the four grey squares. The set Q3 consists of the sixteen shaded black squares.
U \{y0 } ∈ B(Y), and f is Borel measurable by assumption, it follows that A ∈ B(X ). Finally, as
B = {x ∈ X : g(x) ∈ U, f (x) = y0 }
= f −1 ({y0 }) ∩ g −1 (U),
(152)
(153)
{y0 } ∈ B(Y), U ∈ B(Y), and f and g are both Borel measurable by assumption, it follows that
B ∈ B(X ). Thus, h−1 (U) = A ∪ B ∈ B(X ). Since U was arbitrary, we conclude that h is Borel
measurable.
VI. P ROOF OF T HEOREM III.2
Construction of A. Consider the sequence {ak }k∈N , where ak = 1/2 + 1/2k , and note that
lim ak =
k→∞
1
2
ak > ak+1
(154)
for all k ∈ N.
(155)
Let Q1 = [0, 1]2 be the unit square of side length one. We define
Q2 = Q2,1 ∪ Q2,2 ∪ Q2,3 ∪ Q2,4 ,
(156)
where every square Q2,i ⊆ Q1 has side length a2 /2 with (0 0)T ∈ Q2,1 , (1 0)T ∈ Q2,2 , (0 1)T ∈ Q2,3 ,
and (1 1)T ∈ Q2,4 . It follows from (155) that the squares Q2,1 , . . . , Q2,4 are pairwise disjoint and
Q2 ( Q1 . To define Q3 , we follow the same procedure and split up every set Q2,i into the disjoint
union of four squares with side length a3 /4. The sets Q1 , Q2 , and Q3 are depicted in Figure 3. We
iterate this construction and obtain a sequence {Qk }k∈N , where Qk is the disjoint union of 4k−1 squares
Qk,i , i = 1, . . . , 4k−1 , of side length ak /2k−1 and
Qk+1 ( Qk
March 20, 2018
for all k ∈ N.
(157)
DRAFT
34
Next, we set
A=
\
k∈N
Qk ,
(158)
which, as the intersection of closed sets, is closed. Since A is also bounded it must be compact by the
Heine-Borel theorem [39, Theorem 2.41]. Finally,
λ2 (A) = lim λ2 (Qk )
(159)
= lim
4k−1 a2k
k→∞ (2k−1 )2
(160)
= lim a2k
(161)
1
= ,
4
(162)
k→∞
k→∞
where (159) follows from Property iii) of Lemma H.1.
Construction of κ. We now construct a C ∞ -function κ : R2 → R that is one-to-one on A as defined
in (158). This will be accomplished by building compactly supported C ∞ -functions ϕk,i : R2 → [0, 1],
i = 1, . . . , 4k−1 , k ∈ N, such that
ϕk,i (z) =
1
0
if z ∈ Qk,i ,
(163)
if z ∈ Qk,j , j ∈
{1, . . . , 4k−1 }\{i}.
The construction of these functions is effected by Lemma VI.1 below with ϕk,i (z) = ψδk,i ,ak,i ,wk,i (z),
where wk,i denotes the center of Qk,i , ak,i equals half the side-length of Qk,i , and δk,i is chosen
sufficiently small for (163) to hold (recall that the squares Qk,i are closed and disjoint). Next, we define
the C ∞ -functions
ϕk =
k−1
4X
i=1
i
ϕk,i
4k−1
(164)
i
4k−1
(165)
and note that
ϕk (z) =
for all z ∈ Qk,i , i = 1, . . . , 4k−1 , and k ∈ N, and
|i − j|
4k−1
1
≥ k−1
4
|ϕk (z) − ϕk (w)| =
March 20, 2018
(166)
(167)
DRAFT
35
for all z ∈ Qk,i , w ∈ Qk,j , 1 ≤ i < j ≤ 4k−1 , and k ∈ N. For l ∈ N and a, b ∈ N0 , consider now the
C ∞ -function
(a,b)
sl
(z) :=
=
l
∂ a+b X
1
ϕk (z)
k
a
b
2
∂z1 ∂z2 k=1 8 (Mk + 1)
l
X
∂ a+b ϕk (z)
1
82 (Mk + 1) ∂z1a ∂z2b
k=1
k
(168)
(169)
with
Mk = max max d(i, j),
1≤i≤k 1≤j<k
(
)
∂ j ϕi (z)
d(i, j) = sup max
: a, b ∈ N0 , a + b = j .
∂z1a ∂z2b
z∈R2
(170)
(171)
We now show that this particular choice for the constants Mk guarantees, for each a, b ∈ N0 , that
(a,b)
the sequence {sl
}l∈N of C ∞ -functions converges uniformly and denote the corresponding limiting
functions by κ(a,b) . Corollary J.1 then implies
∂κ(a,b) (z)
= κ(a+1,b) (z)
∂z1
(172)
∂κ(a,b) (z)
= κ(a,b+1) (z)
∂z2
(173)
for all a, b ∈ N0 and z ∈ R2 , and κ(0,0) must therefore be C ∞ . We set κ = κ(0,0) . It remains to prove
(a,b)
uniform convergence of the sequences {sl
a, b ∈ N0 be arbitrary but fixed and note that
}l∈N of C ∞ -functions for all a, b ∈ N0 . To this end, let
d(k, a + b)
∂ a+b ϕk (z)
≤ 2k
a
b
8 (Mk + 1)
∂z1 ∂z2
k + 1)
1
< 2k
8
1
82k (M
(174)
(175)
for all k > a + b and z ∈ R2 . Furthermore, by the sum formula for the geometric series,
∞
X
k=a+b+1
X 1
1
1
<
= .
k
7
8k
82
(a,b)
We can now conclude from (174) and (176) that the sequence {tl
(a,b)
tl (z)
:=
(176)
k∈N
}l∈N of C ∞ -functions
a+b+l
X
∂ a+b ϕk (z)
1
82 (Mk + 1) ∂z1a ∂z2b
k=a+b+1
k
(177)
satisfies the assumptions of Theorem J.1 and therefore converges uniformly to a function, which we
denote by ρ(a,b) . As
(a,b)
sl (z)
=
a+b
X
k=1
March 20, 2018
1
∂ a+b ϕk (z)
(a,b)
+ tl−a−b (z)
82 (Mk + 1) ∂z1a ∂z2b
k
(178)
DRAFT
36
(a,b)
}l∈N must converge uniformly to
a+b
X
1
∂ a+b ϕk (z)
+ ρ(a,b) (z).
82 (Mk + 1) ∂z1a ∂z2b
for all l > a + b, we conclude that {sl
κ(a,b) (z) :=
k
k=1
(a,b)
Since a and b are arbitrary, this implies that {sl
concluding the proof of κ being C ∞ .
(179)
}l∈N converges uniformly for all a, b ∈ N0 , thereby
It remains to show that κ is one-to-one on A. To this end, consider arbitrary but fixed z0 and w0 in A
with z0 6= w0 . We have to show that κ(z0 ) 6= κ(w0 ). Note that by construction of A (cf. (158)), there
exists a k0 ∈ N such that
i) for every k ≥ k0 , there exist ik and jk in {1, . . . , 4k−1 } with ik 6= jk such that z0 ∈ Qk,ik and
w0 ∈ Qk,jk , and
ii) for every k < k0 , there exists an ik such that z0 and w0 are both in Qk,ik .
We therefore have
|κ(z0 ) − κ(w0 )| =
=
X ϕk (z0 ) − ϕk (w0 )
k∈N
(180)
82k (Mk + 1)
∞
X
ϕk (z0 ) − ϕk (w0 )
82k (Mk + 1)
(181)
k=k0
|ϕk0 (z0 ) − ϕk0 (w0 )|
−
≥
k
82 0 (Mk0 + 1)
≥
|ϕk0 (z0 ) − ϕk0 (w0 )|
−
k
82 0 (Mk0 + 1)
1
≥
Mk0 + 1
∞
X
ϕk (z0 ) − ϕk (w0 )
82k (Mk + 1)
k=k0 +1
∞
X
k=k0 +1
|ϕk (z0 ) − ϕk (w0 )|
82k (Mk + 1)
!
∞
X
|ϕk0 (z0 ) − ϕk0 (w0 )|
|ϕk (z0 ) − ϕk (w0 )|
−
,
k
82k
82 0
(182)
(183)
(184)
k=k0 +1
(0,0)
where (180) follows from the uniform convergence of {sl
}l∈N to κ, in (181) we used (165) and ii)
above, (182) is by the reverse triangle inequality, and (184) follows from Mk ≤ Mk+1 for all k ∈ N (cf.
(170)). Moreover, (166)–(167) imply |ϕk0 (z0 ) − ϕk0 (w0 )| ≥ 1/4k0 −1 and (164)–(165) yield
|ϕk (z0 ) − ϕk (w0 )| ≤ |ϕk (z0 )| + |ϕk (w0 )|
(185)
=
(186)
ik + jk
4k−1
≤2
March 20, 2018
for all k ∈ N.
(187)
DRAFT
37
We can therefore further lower-bound |κ(z0 ) − κ(w0 )| according to
1
|κ(z0 ) − κ(w0 )| ≥
Mk 0 + 1
1
=
Mk 0 + 1
1
≥
Mk 0 + 1
=
(Mk0
1
4k0 −1 82
k0
1
4k0 −1 82
k0
1
4k0 −1 82
1
k
+ 1)82 0
k0
−
−
−
1
4k0 −1
∞
X
2
82k
k=k0 +1
!
2
X
(188)
!
(189)
k+k0 +1
82
k∈N0
2
82
−
k0 +1
X 1
8k
!
X 1
8k
k∈N0
| {z }
!
(190)
k∈N0
2
k0
82
(191)
= 87
>
1
k
82 0 (Mk0 + 1)
4
1
−
4k0 −1 82k0
!
(192)
k
82 0 − 4k0
1
= 2k0
k
8 (Mk0 + 1) 4k0 −1 82 0
(193)
> 0.
(194)
Since z0 and w0 are arbitrary, this establishes that κ is indeed one-to-one, thereby finishing the proof.
Lemma VI.1. For a > 0, δ > 0, and w = (w1 w2 )T ∈ R2 , consider the mapping
ψδ,a,w : R2 → [0, 1]
z 7→ ρδ,a (z1 − w1 )ρδ,a (z2 − w2 ),
(195)
(196)
where
ρδ,a : R → [0, 1]
t 7→
f (a + δ − |t|)
f (a + δ − |t|) + f (|t| − a)
with f (t) = e−1/t 1R+ (t). Then, ψδ,a,w is C ∞ and satisfies
1 if max(|z1 − w1 |, |z2 − w2 |) ≤ a,
ψδ,a,w (z) =
0 if min(|z1 − w1 |, |z2 − w2 |) ≥ a + δ.
(197)
(198)
(199)
Proof. It follows from [40, Lemma 2.22] with H = ρδ,a , r1 = a, and r2 = a + δ that ρδ,a is C ∞ and
satisfies
ρδ,a (t) =
1 if |t| ∈ [0, a],
(200)
0 if |t| ∈ [a + δ, ∞).
The claim now follows from ψδ,a,w = ρδ,a (z1 − w1 )ρδ,a (z2 − w2 ) and the properties of ρδ,a .
March 20, 2018
DRAFT
38
VII. P ROOF OF T HEOREM IV.1 (S TRONG C ONVERSE )
Towards a contradiction, suppose that the statement is false. That is, we can find an s ∈ N such that
there exist
i) an A ∈ B(Rs ) with λs (A) > 0,
ii) a real analytic mapping h : Rs → Rm , with s ≤ m, of s-dimensional Jacobian Jh 6≡ 0, and
iii) an n ∈ N with n < s and a real analytic mapping f : Rm → Rn that is one-to-one on h(A).
Let s0 be the smallest s ∈ N such that i)–iii) hold. The proof will be effected by showing that this implies
validity of i)–iii) for s0 − 1 and n − 1, which contradicts the assumption that s0 is the smallest natural
number for i)–iii) to hold. The reader might wonder what happens to this argument in the case where
n = 1. In fact we establish below that, if i)–iii) is satisfied, then necessarily n ≥ 2, cf. the claims a)–c)
right after (215).
Let A, h, n, and f satisfy i)–iii) for this minimum value s0 . We start by noting that
m ≥ s0 > n ≥ 1
(201)
f (x) = (f1 (x) . . . fn (x))T
(202)
by ii) and iii). Next, we write
and set ψi = fi ◦ h, i = 1, . . . , n, and ψ = f ◦ h, which, by Corollary K.2, are all real analytic as
compositions of real analytic mappings. We now show that there must exist an i0 ∈ {1, . . . , n} and a set
Ai0 ⊆ A such that
λs0 (Ai0 ) > 0,
(203)
Jψi0 (z) > 0, and Jh(z) > 0 for all z ∈ Ai0 . To this end, we first decompose
A = A0 ∪
n
[
i=1
Ai ,
(204)
where
Ai = {z ∈ A : Dψi (z) 6= 0, Jh(z) > 0}
= {z ∈ A : Jψi (z) > 0, Jh(z) > 0},
(205)
i = 1, . . . , n,
A0 = {z ∈ A : Dψ(z) = 0} ∪ {z ∈ A : Jh(z) = 0}.
(206)
(207)
By Lemma VII.1 below (with s = s0 ), Dψ(z) 6= 0 for λs0 -a.a. z ∈ A. Furthermore, Jh(z) 6= 0 for
λs0 -a.a. z ∈ A because of Jh 6≡ 0 and Lemma K.7. Thus, λs0 (A0 ) = 0 by countable subadditivity of
Lebesgue measure. Since λs0 (A) > 0 by assumption, and λs0 (A0 ) = 0, it follows, again by countable
subadditivity of Lebesgue measure, that there must exist an i0 such that (203) holds.
March 20, 2018
DRAFT
39
Now, for each y ∈ R, let
My = ψi−1
({y}).
0
(208)
We show in Section VII-A below that there exist a y0 ∈ R and a z0 ∈ Ai0 ∩ My0 such that
H s0 −1 Bs0 (z0 , r) ∩ Ai0 ∩ My0 > 0
Jψi0 (z) > 0
for all r > 0,
(209)
for all z ∈ My0 ,
(210)
Jh(z0 ) > 0.
(211)
It now follows from (210), real analyticity of ψi0 , My0 6= ∅, and Lemma K.11 that My0 is a (s0 − 1)-
dimensional real analytic submanifold of Rs0 . Therefore, by Lemma K.9, there exist a real analytic
embedding ζ : Rs0 −1 → Rs0 and an η > 0 such that
ζ(0) = z0 ,
Bs0 (z0 , η) ∩ My0 ⊆ ζ Rs0 −1 ,
(212)
(213)
and ζ Rs0 −1 is relatively open in My0 , i.e., there exists an open set V ⊆ Rs0 with ζ Rs0 −1 = V ∩My0 .
Combining (209) and (213) yields
H s0 −1 ζ Rs0 −1 ∩ Ai0 > 0.
(214)
Ci0 = ζ −1 ζ Rs0 −1 ∩ Ai0 .
(215)
Next, let
We now show that
a) Ci0 ∈ B(Rs0 −1 ) with λs0 −1 (Ci0 ) > 0,
b) h̃ = h ◦ ζ : Rs0 −1 → Rm is real analytic and of (s0 − 1)-dimensional Jacobian J h̃ 6≡ 0, and
c) s0 − 1 > n − 1 > 0 and the real analytic mapping
f˜: Rm → Rn−1
x 7→ (f1 (x) . . . fi0 −1 (x) fi0 +1 (x) . . . fn (x))T
(216)
(217)
is one-to-one on h̃(Ci0 ),
which finally yields the desired contradiction to the statement of s0 being the smallest natural number
such that i)–iii) at the beginning of the proof are satisfied.
Proof of a). We first establish that Ci0 ∈ B(Rs0 −1 ). Since ζ Rs0 −1 is relatively open in My0 and,
therefore, a Borel set in Rs0 , it follows from (215) that Ci0 is the inverse image of a finite intersection
March 20, 2018
DRAFT
40
of Borel sets under the real analytic embedding ζ and, therefore, also a Borel set. Next, we show that
λs0 −1 (Ci0 ) > 0. Since
Z
Jζ(w)dλs0 −1 (w) = H s0 −1 (ζ(Ci0 ))
(218)
Ci0
= H s0 −1 ζ(Rs0 −1 ) ∩ Ai0
> 0,
(219)
(220)
where (218) follows from the area formula Corollary H.3 upon noting that ζ is one-to-one as an embedding
and locally Lipschitz by real analyticity, (219) is by (215), and in (220) we applied (214). Using Lemma
H.2, we conclude from (218)–(220) that λs0 −1 (Ci0 ) > 0.
Proof of b). By Corollary K.2, h̃ is real analytic as the composition of real analytic mappings. It
remains to show that J h̃ 6≡ 0. To this end, we establish J h̃(0) > 0. First note that the chain rule yields
Dh̃(0) = (Dh)(ζ(0))Dζ(0)
= Dh(z0 )Dζ(0),
(221)
(222)
where the second equality is by (212). Since ζ : Rs0 −1 → Rs0 is an embedding, it follows that
rank(Dζ(0)) = s0 − 1.
(223)
Moreover, as h : Rs0 → Rm with m ≥ s0 , (211) implies
rank(Dh(z0 )) = s0 .
(224)
Applying Lemma K.12 to Dh̃(0) = Dh(z0 )Dζ(0) therefore yields rank(Dh̃(0)) ≥ s0 − 1, which in
turn implies J h̃(0) > 0 because Dh̃(0) ∈ Rm×(s0 −1) .
Proof of c). s0 − 1 > n − 1 simply follows from s0 > n (cf. (201)). To prove that f˜ is one-to-one
on h̃(Ci0 ), we first show that fi0 is constant on h̃(Ci0 ). In fact, since ζ(Ci0 ) ⊆ My0 by (215), and
My0 = (fi0 ◦ h)−1 ({y0 }), it follows that
fi0 (h̃(w)) = y0
for all w ∈ Ci0 .
(225)
As f is one-to-one on h(A) and fi0 is constant on h̃(Ci0 ) with h̃(Ci0 ) = (h ◦ ζ)(Ci0 ) ⊆ h(A) by (215),
we conclude that f˜ must also be one-to-one on h̃(Ci0 ). It remains to show that we must have n > 1,
which obviously implies n − 1 > 0. Suppose, to the contrary, that n = 1. Since h̃ is real analytic with
J h̃ 6≡ 0 and λs0 −1 (Ci0 ) > 0, it follows from Property ii) of Lemma K.8 that h̃ is one-to-one on a subset
of Ci0 of positive Lebesgue measure. Thus, the set h̃(Ci0 ) is uncountable. Now, since by (225) fi0 is
constant on h̃(Ci0 ), and a constant function can not be one-to-one on a set of cardinality larger than one,
March 20, 2018
DRAFT
41
the uncountability of h̃(Ci0 ) implies that fi0 can not be one-to-one on h̃(Ci0 ). But f = fi0 for n = 1
(with i0 = 1) and f is one-to-one on h̃(Ci0 ) ⊆ h(A) by assumption, which results in a contradiction.
Therefore, we necessarily have n > 1 and we can conclude that a)–c) must hold, which finalizes the
proof of the theorem.
A. Proof of (209)–(211)
We start by noting that
Z
Z
1
s0 −1
H
(Ai0 ∩ My )dλ (y) =
R
Jψi0 (z)dλs0 (z)
(226)
Ai0
Z
=
1A (z)Jψi (z)dλs (z)
0
Rs0
i0
0
> 0,
(227)
(228)
where (226) is by the coarea formula Corollary H.4 upon noting that ψi0 is locally Lipschitz by real
analyticity, and (228) follows from
1A (z)Jψi (z) > 0 for all z ∈ Ai , λs (Ai ) > 0 (cf. (203)), and
0
i0
0
0
0
Lemma H.2. Also by Lemma H.2 and (226)–(228), there must exist a set D ⊆ R with λ1 (D) > 0 such
that
H s0 −1 Ai0 ∩ My > 0
for all y ∈ D.
(229)
Furthermore,
λ1 {y ∈ R : ∃z ∈ My with Jψi0 (z) = 0} = λ1 ψi0 {z ∈ Rs0 : Jψi0 (z) = 0}
= 0,
(230)
(231)
where (230) is by (208) and (231) follows from Property ii) of Theorem H.2 with ψi0 being C ∞ (recall
that ψi0 = fi0 ◦ h is real analytic as a composition of real analytic mappings and, therefore, C ∞ by
Lemma K.3). Now, (230)–(231) together with λ1 (D) > 0 implies the existence of a y0 ∈ D such that
(210) holds. For this y0 , we must have H s0 −1 (Ai0 ∩ My0 ) > 0 by (229). Therefore, Lemma H.7 implies
that there must exist a z0 ∈ Ai0 ∩ My0 such that (209) holds. As this z0 ∈ Ai0 ∩ My0 ⊆ Ai0 , (211)
finally follows from (206).
Lemma VII.1. Let A ∈ B(Rs ) be of positive Lebesgue measure, h : Rs → Rm , with s ≤ m, real
analytic of s-dimensional Jacobian Jh 6≡ 0, and f : Rm → Rn real analytic. If f is one-to-one on h(A),
then D(f ◦ h)(z) 6= 0 for λs -a.a. z ∈ A.
Proof. Towards a contradiction, suppose that f is one-to-one on h(A) and D(f ◦ h) ≡ 0 on a set B ⊆ A
with λs (B) > 0. Since f ◦ h is real analytic by Corollary K.2, D(f ◦ h) is real analytic as a consequence
March 20, 2018
DRAFT
42
of Lemma K.3. It therefore follows from Lemma K.5 that D(f ◦ h) is identically zero. Hence, f ◦ h
must be constant. In particular, f ◦ h must be constant on B and, therefore, f must be constant on h(B).
Since f is one-to-one and constant on h(B), it follows that h must be constant on B . Thus, Jh(z) = 0
for all z ∈ B , and Lemma K.5 implies Jh ≡ 0. This is a contradiction to Jh 6≡ 0.
A PPENDIX A
P ROOF OF L EMMA III.1
Proof of i). Suppose that U is s-rectifiable. Then, U = ϕ(A) with A ⊆ Rs compact and ϕ Lipschitz.
For t ∈ N with t > s, take
z
B=
: z ∈ A ⊆ Rt
0
(232)
and define
ψ : B → Rm
z
7→ ϕ(z).
0
(233)
(234)
Then, U = ψ(B) is t-rectifiable as a consequence of B compact and ψ Lipschitz.
Proof of ii). First note that we can cover Rs according to
[
Ak ,
Rs =
(235)
k∈N
with Ak ⊆ Rs compact for all k ∈ N (take, for example, Ak = Bs (0, k)). Setting, for every i, k ∈ N,
ϕi,k = ϕi |Ak (locally Lipschitz mappings are Lipschitz on compact sets by Lemma H.9), we can write
[
V=
ϕi,k (Ak ),
(236)
i,k∈N
which implies that V is countably s-rectifiable.
Proof of iii). Suppose that U is countably s-rectifiable and V is countably t-rectifiable. Then, we can
write
U=
V=
where the Ai ⊆
Rs
and the Bj ⊆
Rt
[
ϕi (Ai )
(237)
ψj (Bj ),
(238)
i∈N
[
j∈N
are compact and the ϕi : Rs → Rm and the ψj : Rt → Rn are
Lipschitz. Thus,
W=
March 20, 2018
[
θi,j (Ci,j ),
(239)
i,j∈N
DRAFT
43
where we defined
θi,j : Rs+t → Rm+n
(a b)T 7→ (ϕi (a) ψj (b))T
(240)
(241)
and set
Ci,j = {(a b)T : a ∈ Ai , b ∈ Bj }.
(242)
Now, the θi,j are Lipschitz because the ϕi and the ψj are Lipschitz for all i, j ∈ N, and the Ci,j are
compact by Tychonoff’s theorem [41, Theorem 4.42] thanks to Ai and Bj compact for all i, j ∈ N.
Therefore, W is countably (s + t)-rectifiable.
Proof of iv). Let M be a s-dimensional C 1 -submanifold of Rm . By [36, Definition 5.3.1], we can
write
M=
[
ϕx (Ux ),
(243)
x∈M
where, for every x ∈ M, Ux ⊆ Rs is open, and ϕx : Ux → Rm is C 1 and satisfies x ∈ ϕx (Ux ) and
ϕx (Ux ) = Vx ∩ M with Vx ⊆ Rm open. Now, there must exist a countable set {xi : i ∈ N} ⊆ M such
that
M=
[
i∈N
Vxi ∩ M.
(244)
For if such a countable set does not exist, the open set
[
V=
x∈M
Vx
(245)
would not admit a countable subcover, which would contradict that V as an open set in Rm is Lindelöf
[42, Definition 5.6.19, Proposition 5.6.22]. With the countable set {xi : i ∈ N} ⊆ M we can now write
M=
[
ϕi (Ui ),
(246)
i∈N
where we set ϕi = ϕxi and Ui = Uxi . Now fix i ∈ N arbitrarily. As Ui is open, for every ui ∈ Ui , there
exists an rui > 0 such that
B s (ui , rui ) ⊆ Ui .
(247)
We can thus write
Ui =
March 20, 2018
[
ui ∈Ui
ru
B s ui , i .
2
(248)
DRAFT
44
Since Ui as an open set in Rm is Lindelöf [42, Definition 5.6.19, Proposition 5.6.22], there exists a
countable set {ui,j : j ∈ N} ⊆ Ui such that
Ui =
[
j∈N
ri,j
B ui,j ,
,
2
s
(249)
where we set ri,j = rui,j for all j ∈ N. Using B s (ui,j , ri,j ) ⊆ Ui for all j ∈ N (cf. (247)), it follows that
[
ri,j
Ui =
B s ui,j ,
.
(250)
2
j∈N
Since i ∈ N was arbitrary, using (246) and (250), we get
[
ri,j
s
M=
ϕi B ui,j ,
.
2
(251)
i,j∈N
Finally, all the ϕi as C 1 mappings are locally Lipschitz and, therefore, Lipschitz on the compact sets
B s ui,j , ri,j
for all j ∈ N, which establishes countable rectifiability of M.
2
Proof of v). Since every si -rectifiable set, with si ≤ s, is s-rectifiable by i), it follows that A is a
countable union of s-rectifiable sets and, therefore, countably s-rectifiable.
A PPENDIX B
P ROOF OF L EMMA III.3
Proof of i). Suppose that U is s-rectifiable. Then, there exist a compact set A ⊆ Rs and a Lipschitz
mapping ϕ : A → Rm such that U = ϕ(A). As f ◦ ϕ : A → Rn is Lipschitz owing to Lemmata H.9 and
H.11, f (U) = (f ◦ ϕ)(A) is s-rectifiable.
S
Proof of ii). Suppose that U is countably s-rectifiable. We can write U = i∈N Ui , where Ui is sS
rectifiable for all i ∈ N. As f (U) = i∈N f (Ui ) and f (Ui ) is s-rectifiable for all i ∈ N by i), it follows
that f (U) is countably s-rectifiable.
Proof of iii). Suppose that U ∈ B(Rm ) is countably (H s , s)-rectifiable. We have to show that f (U) =
A∪B, where A ∈ B(Rn ) is countably (H s , s)-rectifiable and H s (B) = 0. As U ∈ B(Rm ) is countably
(H s , s)-rectifiable, [35, Lemma 15.5] implies that H s |U is σ -finite. We can therefore employ Lemma
H.8 to decompose
U = V0 ∪
[
j∈N
Vj ,
(252)
where H s (V0 ) = H s |U (V0 ) = 0 and Vj ⊆ Rm is compact for all j ∈ N. This decomposition allows us
to write f (U) = A ∪ B , where
A=
[
f (Vj )
B = f (V0 ).
March 20, 2018
(253)
j∈N
(254)
DRAFT
45
Now, thanks to Lemma H.10, the f (Vj ) are compact. Therefore, A as a countable union of compact sets
is Borel. Furthermore, H s (B) = 0 because of H s (V0 ) = 0 and Lemma B.1 below. It remains to show
that A is countably (H s , s)-rectifiable. Since U is countably (H s , s)-rectifiable, there exists a countably
s-rectifiable set U1 such that H s (U \U1 ) = 0. Furthermore, as U1 is countably s-rectifiable, f (U1 ) is
countably s-rectifiable by ii). Finally,
H s (A\f (U1 )) ≤ H s (f (U)\f (U1 ))
(255)
≤ H s (f (U \U1 ))
(256)
= 0,
(257)
where (255) follows from the monotonicity of H s and f (U) = A ∪ B , (256) is by monotonicity of H s
and f (U)\f (U1 ) ⊆ f (U\U1 ), and (257) follows from Lemma B.1 below. This proves that A is countably
(H s , s)-rectifiable, thereby concluding the proof.
Lemma B.1. If U ⊆ Rm with H s (U) = 0 and f : Rm → Rn is locally Lipschitz, then H s (f (U)) = 0.
Proof. We employ the following chain of arguments:
[
H s (f (U)) = H s f U ∩
Bm (0, l)
(258)
l∈N
=Hs
[
l∈N
≤
X
≤
X
≤
X
l∈N
l∈N
f (Bm (0, l) ∩ U)
(259)
H s (f (Bm (0, l) ∩ U))
(260)
Lsl H s (Bm (0, l) ∩ U)
(261)
Lsl H s (U)
(262)
l∈N
= 0,
(263)
where (260) follows from the countable subadditivity of H s , in (261) we applied Property ii) of Lemma
H.13, where, for every l ∈ N, Ll denotes the Lipschitz constant of f |Bm (0,l) , and in (262) we used the
monotonicity of H s .
A PPENDIX C
P ROOF OF L EMMA III.5
Towards a contradiction, suppose that there exist r, t ∈ N with r 6= t such that x is r-rectifiable and
t-rectifiable. We can assume, w.l.o.g., that r < t. Now, Lemma III.4 implies the existence of
March 20, 2018
DRAFT
46
i) a countably r-rectifiable set U satisfying P[x ∈ U] = 1 and
ii) a countably t-rectifiable set V satisfying P[x ∈ V] = 1.
With Definition III.1, we can conclude that there exist
i) compact sets Ai ⊆ Rr and Lipschitz mappings ϕi : Ai → Rm , i ∈ N, such that
[
U=
ϕi (Ai ),
(264)
i∈N
and
ii) compact sets Bj ⊆ Rt and Lipschitz mappings ψj : Bj → Rm , j ∈ N, such that
[
V=
ψj (Bj ).
(265)
j∈N
The union bound now yields
h
i
[[
P[x ∈ U ∩ V] = P x ∈
ϕi (Ai ) ∩ ψj (Bj )
(266)
i∈N j∈N
≤
XX
i∈N j∈N
P[x ∈ ϕi (Ai ) ∩ ψj (Bj )].
(267)
Since P[x ∈ U ∩ V] = 1, which follows from P[x ∈ U] = 1 and P[x ∈ V] = 1, (266)–(267) guarantee the
existence of an i0 and a j0 , both in N, such that µx (ϕi0 (Ai0 ) ∩ ψj0 (Bj0 )) > 0. As µx H t |V by the trectifiability of x and H t |V H t by the monotonicity of H t , it follows that H t (ϕi0 (Ai0 )∩ψj0 (Bj0 )) >
0. Property i) of Lemma H.13 therefore implies (recall that r < t by assumption) H r (ϕi0 (Ai0 ) ∩
ψj0 (Bj0 )) = ∞. The monotonicity of H r then yields
H r (ϕi0 (Ai0 )) ≥ H r (ϕi0 (Ai0 ) ∩ ψj0 (Bj0 ))
= ∞.
(268)
(269)
But we also have
H r (ϕi0 (Ai0 )) ≤ Lr H r (Ai0 )
(270)
= Lr λr (Ai0 )
(271)
< ∞,
(272)
where (270) follows from Property ii) of Lemma H.13 with L the Lipschitz constant of ϕi0 , (271) is by
Property iii) of Lemma H.13 and the fact that Ai0 as a compact set is in B(Rr ), and finally (272) is a
consequence, again, of Ai0 being compact. This contradicts (269) and thereby concludes the proof.
March 20, 2018
DRAFT
47
A PPENDIX D
P ROOF OF L EMMA IV.1
Suppose that there exists a Borel measurable mapping g : Rn×m × Rn → Rm such that P[g(A, Ax) 6=
x] < 1 and set U = {x ∈ Rm : g(A, Ax) = x}. We first show that U ∈ B(Rm ). To this end, consider the
mapping hA : Rm → Rm , x 7→ g(A, Ax), which as the composition of two Borel measurable mappings,
namely, x 7→ (A, Ax) and g , is Borel measurable. Application of Lemma H.5 (with X = Y = Rm , f1
the identity mapping on Rm , and f2 = hA ) therefore establishes that
F : Rm → Rm × Rm
x 7→ (x, hA (x))
(273)
(274)
is Borel measurable. Now, consider the diagonal D = {(x, x) : x ∈ Rm } and note that D as the
inverse image of {0} under the Borel measurable mapping d : Rm × Rm → Rm , (u, v) 7→ u − v is
in B(Rm × Rm ). Since U = F −1 (D), we conclude that U ∈ B(Rm ). Now, P[x ∈ U] > 0 as P[x ∈
/
U] = P[g(A, Ax) 6= x] < 1 by assumption. It remains to show that A is one-to-one on U . To this end,
consider arbitrary but fixed u, v ∈ U and suppose that Au = Av . This implies g(A, Au) = g(A, Av).
As u, v ∈ U by assumption, this is only possible if u = v . Thus, A is one-to-one on U , which concludes
the proof.
A PPENDIX E
P ROOF OF L EMMA IV.3
Proof of i). The proof is by induction. Suppose that µ is s-analytic for s ∈ N \{1}. We have to show
that this implies (s − 1)-analyticity of µ. Consider C ∈ B(Rm ) with µ(C) > 0. Then, by Definition IV.1,
there exist a set A ∈ B(Rs ) of positive Lebesgue measure and a real analytic mapping h : Rs → Rm
of s-dimensional Jacobian Jh 6≡ 0 such that h(A) ⊆ C . By Property ii) of Lemma K.8, we can assume,
w.l.o.g., that h|A is an embedding. For each z ∈ R, let
Az = {v ∈ Rs−1 : (v z)T ∈ A}.
(275)
We now use Fubini’s Theorem to show that there exists a z0 ∈ R such that λs−1 (Az0 ) > 0. Concretely,
Z
0<
1A (z)Jh(z)dλs (z)
(276)
s
R
Z Z
s−1
T
T
=
1A (v z) Jh (v z) dλ (v) dλ1 (z)
(277)
R
Rs−1
Z Z
s−1
T
(278)
=
1Az (v)Jh (v z) dλ (v) dλ1 (z),
R
March 20, 2018
Rs−1
DRAFT
48
where (276) follows from Lemma H.2 with λs (A) > 0 and Jh(z) > 0 for all z ∈ A, and in (277) we
applied Theorem H.1. We conclude that there must exist a z0 ∈ R such that
Z
1Az0 (v)Jh (v z0 )T dλs−1 (v) > 0.
(279)
Rs−1
Again using Lemma H.2 establishes that λs−1 (Az0 ) > 0. It remains to show that there exists a real
analytic mapping h̃ : Rs−1 → Rm with h̃(Az0 ) ⊆ C and of (s − 1)-dimensional Jacobian J h̃ 6≡ 0. To this
end, consider the mapping
ϕ : Rs−1 → Rs
(280)
v 7→ (v z0 )T
(281)
and set h̃ = h◦ϕ : Rs−1 → Rm . Now, ϕ is real analytic thanks to Corollary K.1. Thus, h̃ as a composition
of real analytic mappings is real analytic by Corollary K.2. Furthermore, since ϕ(Az0 ) ⊆ A by (275), it
follows that h̃(Az0 ) ⊆ C . It remains to establish that J h̃ 6≡ 0. To this end, we first note that the chain
rule implies Dh̃(v) = Dh(v, z0 )Dϕ(v). Therefore,
rank(Dh̃(v)) = rank(Dh(v, z0 )Dϕ(v))
= rank Dh(v, z0 )(Is−1 0)T
≥s−1
(282)
for all v ∈ Az0 ,
(283)
(284)
where (284) is by Lemma K.12 upon noting that Dh(v, z0 ) ∈ Rm×s , with rank(Dh(v, z0 )) = s (recall
that (v z0 )T ∈ A for all v ∈ Az0 by (275) and that the s-dimensional Jacobian Jh(z) > 0 for all
z ∈ A), and (Is−1 0)T ∈ Rs×(s−1) . Since rank(Dh̃(v)) ≥ s − 1 and Dh̃(v) ∈ Rm×(s−1) , we conclude
that J h̃(v) > 0 for all v ∈ Az0 . Thus, J h̃ 6≡ 0, which finalizes the proof.
Proof of ii). Suppose that µ is s-analytic and consider C ∈ B(Rm ) with µ(C) > 0. We have to show
that H s (C) > 0. By Definition IV.1, there exist a set A ∈ B(Rs ) of positive Lebesgue measure and a
real analytic mapping h : Rs → Rm of s-dimensional Jacobian Jh 6≡ 0 such that h(A) ⊆ C . By Property
ii) of Lemma K.8, we can assume, w.l.o.g., that h|A is an embedding. We now use the area formula
Corollary H.3 to conclude that H s (C) > 0. Concretely,
Z
0<
Jh(z)dλs (z)
(285)
A
= H s h(A)
≤ H s (C),
March 20, 2018
(286)
(287)
DRAFT
49
where (285) is by Lemma H.2, λs (A) > 0, and Jh(z) > 0 for all z ∈ A, in (286) we applied the area
formula Corollary H.3 upon noting that h|A is one-to-one as an embedding and locally Lipschitz by real
analyticity of h, and (287) is by monotonicity of H s together with h(A) ⊆ C .
Proof of iii). ,,⇒”: Suppose that µ is s-analytic and there exists a set U ⊆ Rm such that µ = µ|U
and H s |U is σ -finite. Since H s is Borel regular (cf. Definition H.2), we may assume, w.l.o.g., that
U ∈ B(Rm ). By σ -finiteness of H s |U , there exist sets Vi ∈ B(Rm ), i ∈ N, such that
[
U=
U ∩ Vi
(288)
i∈N
and H s (U ∩ Vi ) < ∞ for all i ∈ N. For every i ∈ N, since U ∩ Vi as the intersection of two Borel sets
fi , where Wi is countably
is in B(Rm ) and H s (U ∩ Vi ) < ∞, we can write [43, p. 83] U ∩ Vi = Wi ∪ W
fi is purely H s -unrectifiable, i.e., H s (W
fi ∩ E) = 0 for all countably (H s , s)(H s , s)-rectifiable and W
f with W = S Wi
rectifiable sets E ⊆ Rm . This allows us to decompose U according to U = W ∪ W
i∈N
S
s
f
fi
f
and W = i∈N Wi . Since all the Wi are countably (H , s)-rectifiable, so is W ; and since all the W
f.
are purely H s -unrectifiable, so is W
f, it remains to show that µ(W)
f = 0 to conclude that
As µ = µ|U by assumption, and U = W ∪ W
f > 0.
µ = µ|W for the countably (H s , s)-rectifiable set W . Towards a contradiction, suppose that µ(W)
Analyticity of µ then implies that there exist a set A ∈ B(Rs ) of positive Lebesgue measure and a real
f. By countable
analytic mapping h : Rs → Rm of s-dimensional Jacobian Jh 6≡ 0 such that h(A) ⊆ W
subadditivity of λs , we can assume, w.l.o.g., that A is bounded; and by Property ii) in Lemma K.8, we
can assume, w.l.o.g., that h|A is an embedding. It follows that
Z
0<
Jh(z)dλs (z)
(289)
A
= H s h(A) ,
(290)
where (289) is by Lemma H.2, λs (A) > 0, and Jh(z) > 0 for all z ∈ A, and in (290) we applied the
area formula Corollary H.3 upon noting that h|A is one-to-one as an embedding and locally Lipschitz
by real analyticity of h. Moreover, as H s is Borel regular (cf. Property ii) in Definition H.2), there
must exist a set C ∈ B(Rm ) with h(A) ⊆ C and H s (C \ h(A)) = 0. It follows that C is countably
(H s , s)-rectifiable as H s (C \h(A)) = 0 and h is Lipschitz on the compact set A by Lemma H.9. But
this implies
f ∩ C) ≥ H s (h(A) ∩ C)
H s (W
March 20, 2018
(291)
= H s (h(A))
(292)
> 0,
(293)
DRAFT
50
f is purely H s -unrectifiable, thereby concluding the proof.
which is not possible as W
,,⇐”: Suppose that there exists a countably (H s , s)-rectifiable set W ⊆ Rm such that µ = µ|W . Since
H s is Borel regular (cf. Definition H.2), we may assume, w.l.o.g., that W ∈ B(Rm ). As this W is
countably (H s , s)-rectifiable, [35, Lemma 15.5] implies that H s |W is σ -finite.
A PPENDIX F
P ROOF OF L EMMA IV.4
Suppose that x ∈ Rm is s-analytic and u = f (x), where f : Rm → Rk is a real analytic immersion,
and consider C ∈ B(Rk ) with µu (C) > 0. We have to show that there exist a set A ∈ B(Rs ) of positive
Lebesgue measure and a real analytic mapping g : Rs → Rk of s-dimensional Jacobian Jg 6≡ 0 such that
g(A) ⊆ C . Set D = f −1 (C) ∈ B(Rm ). Since µx (D) = µu (C) > 0 and x is s-analytic, there exist a set
A ∈ B(Rs ) of positive Lebesgue measure and a real analytic mapping h : Rs → Rm of s-dimensional
Jacobian Jh 6≡ 0 such that h(A) ⊆ D. We set g = f ◦ h. Now, g(A) = f (h(A)) ⊆ f (D) ⊆ C .
Furthermore, g as the composition of real analytic mappings is real analytic by Corollary K.2. It remains
to show that Jg 6≡ 0. To this end, we first note that the chain rule implies Dg(z) = (Df )(h(z))Dh(z).
Since Jh 6≡ 0, there exists a z0 ∈ Rs such that Jh(z0 ) 6= 0. Thus
rank(Dh(z0 )) = s.
(294)
Now, as f is an immersion, it follows that k ≥ m and Jf > 0. Thus,
rank((Df )(h(z0 ))) = m.
(295)
Applying Lemma K.12 to (Df )(h(z0 )) ∈ Rk×m and Dh(z0 ) ∈ Rm×s and using (294) and (295)
establishes that rank(Dg(z0 )) ≥ s, which in turn implies Jg(z0 ) 6= 0 as Dg(z0 ) ∈ Rk×s .
A PPENDIX G
P ROOF OF L EMMA IV.6
Suppose that a ∈ Rk and b ∈ Rl with µa × µb λk+l , set x = a ⊗ b, and consider C ∈ B(Rkl )
with µx (C) > 0. We have to show that there exist a set A ∈ B(Rk+l−1 ) of positive Lebesgue measure
and a real analytic mapping h : Rk+l−1 → Rkl of (k + l − 1)-dimensional Jacobian Jh 6≡ 0 such that
h(A) ⊆ C . Let
E = {(aT bT )T : a ∈ Rk , b ∈ Rl , a ⊗ b ∈ C}.
March 20, 2018
(296)
DRAFT
51
Since C ∈ B(Rkl ) and ⊗ is Borel measurable, E as the inverse image of C under ⊗ is Borel measurable.
Furthermore, as x = a ⊗ b, it follows that
(µa × µb )(E) = µx (C)
> 0,
(297)
(298)
which implies λk+l (E) > 0 as µa × µb λk+l by assumption. Using Corollary H.1, we can write
Z
k+l
λ (E) =
λk+l−1 (Ez )dλ1 (z),
(299)
R
where, for each z ∈ R,
Ez =
aT v T
T
: a ∈ Rk , v ∈ Rl−1 , (aT v T z)T ∈ E .
(300)
As λk+l (E) > 0 we can conclude that there must exist a z0 ∈ R\{0} such that λk+l−1 (Ez0 ) > 0. We set
A = Ez0 . Next, we define the mapping
h : Rk+l−1 → Rkl
(c1 . . . ck+l−1 )T 7→ (c1 . . . ck )T ⊗ (ck+1 . . . ck+l−1 z0 )T ,
(301)
(302)
which is real analytic thanks to Corollary K.1, and we write
A=
aT v T
T
: a ∈ Rk , v ∈ Rl−1 , (aT v T z0 )T ∈ E
(303)
=
aT v T
T
: a ∈ Rk , v ∈ Rl−1 , a ⊗ (v T z0 )T ∈ C
(304)
= h−1 (C),
(305)
where (304) follows from (296), and (305) is by (301)–(302). By construction, h(A) ⊆ C . Furthermore,
A as the inverse image of C ∈ B(Rkl ) under a real analytic and, therefore, Borel measurable mapping
is in B(Rk+l−1 ). It remains to show that there exist an a0 ∈ Rk and a v0 ∈ Rl−1 such that
r
T T T T
T
T
T T
= det Dh aT
v
Dh a0 v0
Jh aT
0 v0
0
0
> 0.
This will be accomplished by showing that there exist an a0 ∈ Rk and a v0 ∈ Rl−1 such that
T T
rank Dh aT
v
= k + l − 1.
0
0
March 20, 2018
(306)
(307)
(308)
DRAFT
52
Now,
v
...
0
z0 0
0 v
0 z0
..
.
T T T
Dh a v
= ..
.
0 0
0 0
0 0
0
0
0
a1 Il−1
0
...
0
0
...
0
0
...
..
.
0
..
.
0
..
.
...
v
0
. . . z0
0
...
0
v
...
0
z0
0
a2 Il−1
0
..
.
ak−1 Il−1
0
ak Il−1
(309)
0
for general a ∈ Rk and v ∈ Rl−1 , and
0
...
0
0
Il−1
z0 0
0 0
0 z0
..
T
.
T
Dh aT
= ..
.
0 v0
0 0
0 0
0 0
...
0
0
...
0
0
...
..
.
0
..
.
0
..
.
...
0
0
. . . z0
0
...
0
0
...
0
z0
0
Il−1
0
..
.
Il−1
0
Il−1
0
0
0
0
(310)
for the specific choices a0 = (1 . . . 1)T ∈ Rk and v0 = (0 . . . 0)T ∈ Rl−1 . Since the (kl) × (k + l − 1)
T T has the regular (k + l − 1) × (k + l − 1) submatrix (recall that z 6= 0)
matrix Dh (aT
0
0 v0 )
0
Il−1
,
(311)
z0 I k
0
(308) indeed holds, which concludes the proof.
A PPENDIX H
T OOLS FROM (G EOMETRIC ) M EASURE T HEORY
In this appendix, we state some basic definitions and results from measure theory and geometric
measure theory used throughout the paper. We start with measure theory.
Definition H.1. [36, Definition 1.2.1] A measure (sometimes called pre-measure) on a nonempty set X
is a nonnegative function µ defined on all subsets of X with the following properties:
March 20, 2018
DRAFT
53
i) µ(∅) = 0.
ii) Monotonicity:
µ(A) ≤ µ(B)
for all A ⊆ B ⊆ X .
(312)
iii) Countable subadditivity:
!
[
µ
i∈N
Ai
≤
X
µ(Ai )
(313)
for all E ⊆ X .
(314)
i∈N
for all sequences {Ai }i∈N of sets Ai ⊆ X .
A set A ⊆ X is µ-measurable if it satisfies
µ(E) = µ(E ∩ A) + µ(E \A)
For a probability measure µx , countable subadditivity is equivalent to the union bound
#
"
X
[
P[x ∈ Ai ].
Ai ≤
P x∈
i∈N
(315)
i∈N
The collection of µ-measurable sets forms a σ -algebra [36, Theorem 1.2.4], which we denote by Sµ (X ).
For a topological space X , the smallest σ -algebra containing the open sets is the Borel σ -algebra
B(X ). A measurable space (X , S (X )) is a set X equipped with a σ -algebra S (X ). A measure space
(X , S (X ), µ) is a set X with a measure µ and a σ -algebra S (X ) ⊆ Sµ (X ).
Lemma H.1. [36, Theorem 1.2.5] If (X , S (X ), µ) is a measure space and {An }n∈N a sequence of sets
An ∈ S (X ), then the following properties hold.
i) If the sets An , n ∈ N, are pairwise disjoint, then µ is countably additive on their union. That is,
[
X
µ(An ).
(316)
µ
An =
n∈N
n∈N
ii) If An ⊆ An+1 for all n ∈ N, then
[
An = lim µ(An ).
µ
n∈N
n→∞
iii) If An+1 ⊆ An for all n ∈ N and µ(A1 ) < ∞, then
\
An = lim µ(An ).
µ
n∈N
n→∞
(317)
(318)
Definition H.2. [36, Definition 1.2.10] A measure µ on a topological space X is
i) Borel if all open sets are µ-measurable;
ii) Borel regular if it is Borel and, for each A ⊆ X , there exists a B ∈ B(X ) such that A ⊆ B and
µ(A) = µ(B).
March 20, 2018
DRAFT
54
As a consequence of Carathéodory’s criterion [36, Theorem 1.2.13], Lebesgue and Hausdorff measures
are Borel regular.
Definition H.3. (Measurable mapping) [44, Chapter 2]
i) Let (X , S (X )) and (Y, S (Y)) be measurable spaces. A mapping f : D → Y , with D ∈ S (X ), is
(S (X ), S (Y))-measurable if f −1 (A) ∈ S (X ) for all A ∈ S (Y).
ii) Let (X , S (X )) be a measurable space and Y a topological space. A mapping f : D → Y , with
D ∈ S (X ), is S (X )-measurable if f −1 (A) ∈ S (X ) for all A ∈ B(Y).
iii) Let X and Y be topological spaces. A mapping f : D → Y , with D ∈ B(X ), is Borel measurable
if f −1 (A) ∈ B(X ) for all A ∈ B(Y).
Lemma H.2. [44, Corollary 4.10] Let (X , S (X ), µ) be a measure space and consider a nonnegative
measurable function f : X → R. Then, f (x) = 0 µ-almost everywhere if and only if
Z
f (x)dµ(x) = 0.
(319)
X
Definition H.4. [36, Definition 1.3.25] The measure space (X , S (X ), µ) is σ -finite if X =
S
i∈N Ai ,
with Ai ∈ S (X ) and µ(Ai ) < ∞ for all i ∈ N. The measure space is finite if µ(X ) < ∞. A
Borel measure µ on a topological space X is σ -finite (respectively finite) if (X , B(X ), µ) is a σ -finite
(respectively finite) measure space.
For example, all probability measures are finite and the Lebesgue measure is σ -finite. However, the
s-dimensional Hausdorff measure with s < m is not σ -finite.
Theorem H.1. (Fubini’s theorem) [44, Theorem 10.9] Let (X , S (X ), µ) and (Y, S (Y), ν) be σ -finite
measure spaces and suppose that f : X × Y → R is nonnegative and (S (X ) ⊗ S (Y))-measurable. Then,
Z
ϕ(x) =
f (x, y)dν(y) is S (X )-measurable,
(320)
Y
Z
ψ(y) =
f (x, y)dµ(x) is S (Y)-measurable,
(321)
X
and
Z
Z
ϕ(x)dµ(x) =
X
X ×Y
f (x, y)d(µ × ν)(x × y)
(322)
Z
=
ψ(y)dν(y).
(323)
Y
March 20, 2018
DRAFT
55
Corollary H.1. Let (X , S (X ), µ) and (Y, S (Y), ν) be σ -finite measure spaces and suppose that A ∈
S (X ) ⊗ S (Y). Then,
Z
X
ν({y : (x, y) ∈ A})dµ(x) =
Z
X ×Y
Z
=
Y
1A (x, y)d(µ × ν)(x × y)
µ({x : (x, y) ∈ A})dν(y).
Proof. Follows from Theorem H.1 with f (x, y) = 1A (x, y), noting that
Z
1A (x, y)dν(y) = ν({y : (x, y) ∈ A}),
Y
Z
1A (x, y)dµ(x) = µ({x : (x, y) ∈ A}).
(324)
(325)
(326)
(327)
X
Lemma H.3. [45, Exercise 1.7.19] If X and Y are Euclidean spaces, then B(X × Y) = B(X ) ⊗ B(Y).
Lemma H.4. [44, Corollary 2.10] Let (X , S (X )) be a measurable space and suppose that {fi }i∈N is
a sequence of S (X )-measurable functions fi : X → R converging pointwise to f : X → R. Then, f is
S (X )-measurable.
Lemma H.5. Let X and Y be Euclidean spaces and suppose that fi : X → Y , i = 1, . . . , n, are Borel
measurable mappings. Then,
f : X → Ỹ = Y × · · · × Y
{z
}
|
(328)
n times
x 7→ (f1 (x), . . . , fn (x))
(329)
is Borel measurable.
Proof. Lemma H.3 implies B(Ỹ) = B(Y) ⊗ · · · ⊗ B(Y). Thus, B(Ỹ) is generated by sets U1 × · · · × Un ,
where Ui ∈ B(Y) for i = 1, . . . , n. It is therefore sufficient (cf. Part iii) in Definition H.3) to show
that f −1 (U1 × · · · × Un ) ∈ B(X ) for all Ui ∈ B(X ), i = 1, . . . , n. But f −1 (U1 × · · · × Un ) =
f1−1 (U1 )∩· · ·∩fn−1 (Un ) ∈ B(X ) for all Ui ∈ B(Y), i = 1, . . . , n, as all the fi are Borel measurable.
Corollary H.2. Let X be a Euclidean space and suppose that {fi }i∈N is a sequence of Borel measurable
mappings fi : X → Rn converging pointwise to f : X → Rn . Then, f is Borel measurable.
Proof. We can write f (x) = (f (1) (x), . . . , f (n) (x))T , where f (j) : X → R is Borel measurable for j =
T
(1)
(n)
(j)
1, . . . , n, and fi (x) = fi (x), . . . , fi (x) , where fi : X → R is Borel measurable for j = 1, . . . , n
(j)
and all i ∈ N. Then, for each j = 1, . . . , n, the sequence (fi )i∈N of Borel measurable functions
March 20, 2018
DRAFT
56
converges pointwise to f (j) , which is Borel measurable thanks to Lemma H.4. Finally, Lemma H.5
implies that f is Borel measurable as its individual components f (1) , . . . , f (n) are Borel measurable.
Lemma H.6. Let X and Y be topological spaces, C ∈ B(X ), f : C → Y Borel measurable, and
y0 ∈ Y \f (C). Consider h : X → Y with h|C = f and h|X \C = y0 . Then, h is Borel measurable.
Proof. We have to show that h−1 (A) ∈ B(X ) for all A ∈ B(Y). Consider an arbitrary but fixed
A ∈ B(Y) and suppose first that y0 ∈ A. Now, A ∈ B(Y) and {y0 } ∈ B(Y) imply A\{y0 } ∈ B(Y).
Therefore,
h−1 (A) = h−1 (A\{y0 }) ∪ h−1 ({y0 })
= f −1 (A\{y0 }) ∪ (X \C) ∈ B(X )
(330)
(331)
because f is Borel measurable, A\{y0 } ∈ B(Y), and X \C ∈ B(X ). If y0 ∈
/ A, then
h−1 (A) = f −1 (A) ∈ B(X )
(332)
as f is Borel measurable and A ∈ B(Y).
Lemma H.7. Let µ be a measure on Rk and consider A ⊆ Rk with µ(A) > 0. Then, there exists a
z0 ∈ A such that
µ Bk (z0 , r) ∩ A > 0
for all r > 0.
(333)
Proof. Suppose, to the contrary, that such a z0 does not exist. Then, for every z ∈ A, there must exist
a rz > 0 such that µ Bk (z, rz ) ∩ A = 0. With these rz , we write
[
A=
z∈A
(Bk (z, rz ) ∩ A).
(334)
It now follows from the Lindelöf property of Rk [42, Definition 5.6.19, Proposition 5.6.22] that there
must exist a countable subset {zi : i ∈ N} ⊆ A such that
A=
[
(Bk (zi , rzi ) ∩ A).
(335)
i∈N
Since µ Bk (zi , rzi ) ∩ A = 0 for all i ∈ N, the countable subadditivity of µ implies
µ(A) ≤
X
i∈N
µ Bk (zi , rzi ) ∩ A
= 0,
(336)
(337)
which contradicts the assumption µ(A) > 0 and thereby concludes the proof.
March 20, 2018
DRAFT
57
Lemma H.8. Let (X , B(X ), µ) be a σ -finite measure space and B ∈ B(X ). Then,
B = N ∪ A,
where µ(N ) = 0 and A =
S
i∈N Ai
(338)
with Ai ⊆ Rm compact for all i ∈ N.
Proof. By [43, Proposition 1.43], we can find, for each i ∈ N, a compact set Ki such that Ki ⊆ B and
S
µ(B\Ki ) ≤ 1/i. For j ∈ N, let Aj = ji=1 Ki . It follows that {Aj }j∈N is an increasing (in terms of ⊆)
sequence of compact sets Aj ⊆ B satisfying µ(B) − µ(Aj ) = µ(B \Aj ) ≤ 1/j for all j ∈ N. We set
S
A = j∈N Aj ⊆ B . Thus, µ(A) ≤ µ(B) by monotonicity of µ. Now,
µ(A) = lim µ(Aj )
(339)
j→∞
= µ(B) − lim µ(B\Aj )
j→∞
1
j→∞ j
(340)
≥ µ(B) − lim
(341)
= µ(B),
(342)
where (339) follows from Lemma H.1. We conclude that µ(A) = µ(B), which, together with A ⊆ B ,
yields (338).
Definition H.5. (Locally Lipschitz mapping) [46, Definition 3.118]
i) A mapping f : U → Rl , where U ⊆ Rk , is Lipschitz if there exists a constant L ≥ 0 such that
kf (u) − f (v)k2 ≤ Lku − vk2
for all u, v ∈ U.
(343)
The smallest constant L for (343) to hold is the Lipschitz constant of f .
ii) A mapping f : Rk → Rl is locally Lipschitz if every x ∈ Rk admits an open neighborhood Ux ⊆ Rk
containing x such that f |Ux is Lipschitz.
The following result establishes a necessary and sufficient condition for a mapping to be locally
Lipschitz. Since we could not find a proof for this statement in the literature, we present one here for
completeness.
Lemma H.9. The mapping f : Rk → Rl is locally Lipschitz if and only if f |K is Lipschitz for all
compact sets K ⊆ Rk .
Proof. ,,⇒”: Suppose that f : Rk → Rl is locally Lipschitz and consider a compact set K ⊆ Rk . We
have to show that f |K is Lipschitz. For every x ∈ K, by the local Lipschitz property of f , there exists
S
an open neighborhood Ux containing x such that f |Ux is Lipschitz. As K is compact and K ⊆ x∈K Ux ,
March 20, 2018
DRAFT
58
there must exist Uxi , denoted by Ui , i = 1, . . . , n, such that K ⊆
Sn
i=1 Ui .
For i = 1, . . . , n, let Li denote
the Lipschitz constant of f |Ui . Now, as shown below, there exists a δ > 0 such that, for every x, y ∈ K
with kx − yk2 < δ , we can find a Ui in the finite subcover of U with x, y ∈ Ui . With this δ we let
2∆
L = max L1 , . . . , Ln ,
,
(344)
δ
where ∆ := maxx∈K f (x). Consider an arbitrary pair x, y ∈ K. If kx − yk2 < δ , then this pair
must be in the same Ui , which implies kf (x) − f (y)k ≤ Li kx − yk ≤ Lkx − yk. If kx − yk2 ≥ δ ,
then kf (x) − f (y)k ≤ 2∆ = 2∆δ/δ ≤ (2∆/δ)kx − yk2 ≤ Lkx − yk2 . Thus, f |K is Lipschitz with
Lipschitz constant L. It remains to establish the existence of a δ with the desired properties. Towards a
contradiction, suppose that such a δ does not exist. This implies that, for every m ∈ N, we can find a pair
xm , ym ∈ K such that kxm −ym k2 < 1/m, but there is no Ui containing this pair. Since K×K is compact
by Tychonoff’s Theorem [41, Theorem 4.42], there exists a convergent subsequence {xmj , ymj }j∈N of
the sequence {xm , ym }m∈N [39, Theorem 2.41]. Let (x̄, ȳ) denote the limit point of this subsequence.
Note that x̄ and ȳ can not be in the same Ui , for if they were there would be an M ∈ N such that
(xmj , ymj ) ∈ Ui × Ui for all j ≥ M , which is not possible as xm and ym are in different Ui ’s for all
m ∈ N. But
x̄ − ȳ = lim (xmj − ymj )
j→∞
= 0,
(345)
(346)
where the second equality follows from limj→∞ k(xmj − ymj )k2 < limj→∞ 1/mj = 0 and continuity
of k · k2 . Thus, x̄ = ȳ , which is not possible as there is no Ui containing x̄ and ȳ .
,,⇐”: Suppose that f |K is Lipschitz for all compact sets K ⊆ Rk . It is sufficient to show that f |Bk (x,1)
is Lipschitz for every x ∈ Rk . As Bk (x, 1) is compact, f |Bk (x,1) is Lipschitz. The Lipschitz property of
f |Bk (x,1) therefore follows immediately from Bk (x, 1) ⊆ Bk (x, 1).
Lemma H.10. If f : Rm → Rn is locally Lipschitz and K ⊆ Rm is compact, then f (K) is compact.
Proof. Follows from continuity of f and [47, Theorem 2-7.2].
We will need the following composition property of locally Lipschitz mappings.
Lemma H.11. If f : Rk → Rm and g : Rm → Rn are locally Lipschitz, then h = g ◦ f : Rk → Rn is
locally Lipschitz.
Proof. Suppose that f : Rk → Rm and g : Rm → Rn are both locally Lipschitz. To prove that h = g ◦ f
is locally Lipschitz, by Lemma H.9, it is sufficient to show that h|K is Lipschitz for all compact sets
March 20, 2018
DRAFT
59
K ⊆ Rk . Let K ⊆ Rk be an arbitrary but fixed compact set and note that Q = f (K) is compact owing
to Lemma H.10. Thus, by Lemma H.9, f |K and g|Q are both Lipschitz with Lipschitz constants, say, L
and M , respectively. It follows that
kh(u) − h(v)k2 ≤ M kf (u) − f (v)k2
(347)
≤ LM ku − vk2
(348)
for all u and v in K, which implies that h|K is Lipschitz. As K was arbitrary, we conclude that h is
locally Lipschitz.
Definition H.6. (Differentiable mapping) [48, Definition 1.2] Let U be an open set in Rm . A function
f : U → R is
i) C 0 if it is continuous.
ii) C r with r ∈ N if, for all possible choices of r1 , . . . , rm ∈ N0 with
Pk
i=1 ri
= r, the partial
derivatives
∂r
f (x)
∂xr11 . . . ∂xrmm
(349)
exist and are continuous on U .
iii) C ∞ if it is C r for all r ∈ N.
For r ∈ N0 ∪ {∞}, a mapping f : U → Rn , x 7→ (f1 (x) . . . fn (x))T is C r if every component fi ,
i = 1, . . . , n, is C r .
It follows from the mean value theorem [49, Theorem 3.4] that C 1 mappings are locally Lipschitz.
Conversely, by Rademacher’s Theorem [36, Theorem. 5.1.11], every locally Lipschitz mapping f : Rm →
Rn has a L (Rm )-measurable (but not necessarily continuous) differential Df , which is defined λm -a.e.
Theorem H.2. (Sard’s theorem) [50] Let f : Rm → Rn , x 7→ (f1 (x) . . . fn (x))T be C r and set A =
{x : Jf (x) = 0}. The following statements hold.
i) If m ≤ n, then H m (f (A)) = 0.
ii) If m > n and r ≥ m − n + 1, then λn (f (A)) = 0.
Next, we state two fundamental results from geometric measure theory that are used frequently in the
paper, namely, the area and the coarea formula for locally Lipschitz mappings f : Rm → Rn .
March 20, 2018
DRAFT
60
Theorem H.3. (Area formula) [36, Theorem 5.1.1] If A ∈ B(Rm ), f : Rm → Rn is Lipschitz, and
g : A → R is nonnegative and Lebesgue measurable, and m ≤ n, then
Z
Z
m
Jf (x)dλ (x) =
card A ∩ f −1 ({y}) dH m (y).
A
(350)
Rn
Corollary H.3. (Area formula) If A ∈ B(Rm ), f : Rm → Rn is locally Lipschitz and one-to-one on A,
and m ≤ n, then
Z
Jf (x)dλm (x) = H m (f (A)).
(351)
A
Proof. For i ∈ N, let gi = 1Bm (0,i) . It follows that
Z
Z
m
Jf (x)dλ (x) = lim
gi (x)Jf (x)dλm (x)
i→∞ A
A
Z
= lim
J(f |Bm (0,i) )(x)dλm (x)
i→∞ A
Z
({y})
dH m (y)
= lim
card A ∩ f |−1
Bm (0,i)
i→∞ Rn
= lim H m f A ∩ Bm (0, i)
i→∞
(352)
(353)
(354)
(355)
!
=Hm
[
i∈N
f A ∩ Bm (0, i)
= H m (f (A)),
(356)
(357)
where (352) is by the Lebesgue Monotone Convergence Theorem [44, Theorem 4.6] upon noting that
(gi )i∈N is an increasing sequence of nonnegative Borel measurable functions converging to 1Rm , in (354)
we applied Theorem H.3 to f |Bm (0,i) , which is Lipschitz by Lemma H.9 as Bm (0, i) ⊆ Bm (0, i) and
Bm (0, i) is compact for all i ∈ N, in (355) we used that f is one-to-one, by assumption, which implies
1 if y ∈ f A ∩ Bm (0, i)
−1
(358)
card A ∩ f |Bm (0,i) ({y}) =
0 else,
and (356) is by Property ii) in Lemma H.1.
Theorem H.4. (Coarea formula) [36, Theorem 5.2.1] If f : Rm → Rn is Lipschitz, A ∈ B(Rm ), and
m ≥ n, then
Z
m
Z
Jf (x)dλ (x) =
A
Rn
H m−n A ∩ f −1 ({y}) dλn (y).
(359)
Corollary H.4. (Coarea formula) If f : Rm → Rn is locally Lipschitz, A ∈ B(Rm ), and m ≥ n, then
Z
Z
m
Jf (x)dλ (x) =
H m−n A ∩ f −1 ({y}) dλn (y).
(360)
A
March 20, 2018
Rn
DRAFT
61
Proof. We have
Z
Z
m
Jf (x)dλ (x) = lim
i→∞ A
A
J(f |Bm (0,i) )(x)dλm (x)
(361)
Z
H m−n A ∩ f |−1
({y})
dλn (y)
Bm (0,i)
i→∞ Rn
!
Z
[
n
A ∩ f |−1
=
H m−n
Bm (0,i) ({y}) dλ (y)
= lim
Rn
Z
=
Rn
(362)
(363)
i∈N
H m−n A ∩ f −1 ({y}) dλn (y),
(364)
where (361) follows from (352)–(353), in (362) we applied Theorem H.4 to f |Bm (0,i) , which is Lipschitz
by Lemma H.9 as Bm (0, i) ⊆ Bm (0, i) and Bm (0, i) is compact for all i ∈ N, and (363) is by the
Lebesgue Monotone Convergence Theorem [44, Theorem 4.6] upon noting that, for every i ∈ N, the
function
gi : Rn → R
(365)
y 7→ H m−n
A ∩ f |−1
({y})
Bm (0,i)
(366)
is Lebesgue measurable [36, Lemma 5.2.5] and, therefore, (gi )i∈N is a sequence of nonnegative increasing
Lebesgue measurable functions, with
!
lim gi (y) = H
m−n
i→∞
[
i∈N
A∩
f |−1
Bm (0,i) ({y})
for all y ∈ Rn
(367)
by Property ii) in Lemma H.1.
Lemma H.12. Main properties of modified Minkowski dimension:
i) A smooth s-dimensional submanifold U of Rm has dimMB (U) = s.
ii) dimMB (·) and dimMB (·) are monotonic with respect to ⊆.
iii) dimMB (U) ≤ dimB (U) and dimMB (U) ≤ dimB (U) for all nonempty sets U .
iv) If f is locally Lipschitz, then
dimMB (f (U)) ≤ dimMB (U)
(368)
dimMB (f (U)) ≤ dimMB (U)
(369)
for all nonempty subsets U in the domain of f .
v) dimMB is countably stable, i.e.,
!
dimMB
[
i∈N
Ui
= sup dimMB (Ui )
(370)
i∈N
for all countable collections of nonempty sets Ui , i ∈ N.
March 20, 2018
DRAFT
62
H d (U)
∞
0
dimH (U)
m
d
Figure 4. ([32, Figure 2.3]) Graph of H d (U) as a function of d ∈ [0, m] for a set U ⊆ Rm .
Proof. Properties i)–iv) follow from the definitions of lower and upper modified Minkowski dimension
and the properties of lower and upper Minkowski dimension listed in [32, Section 3.2]. Property v)
follows from [32, Equation 3.26] and [32, Proposition 3.8].
Definition H.7. (Hausdorff measures) [43, Definition 2.46] Let d ∈ [0, ∞) and U ⊆ Rm . The ddimensional Hausdorff measure of U , denoted by H d (U), is limδ→0 Hδd (U), where
(
)
X
[
V
(d,
1)
inf
diam(Ui )d : diam(Ui ) < δ, U ⊆
Ui
Hδd (U) =
2d
i∈N
(371)
i∈N
for all δ > 0, with
diam(U) :=
sup{ku − vk2 : u, v ∈ U}
if U 6= ∅
0
if U = ∅.
(372)
Definition H.8. (Hausdorff dimension) [43, Definition 2.51] The Hausdorff dimension of U ⊆ Rm ,
denoted by dimH (U), is sup{d ≥ 0 : H d (U) = ∞} = inf{d ≥ 0 : H d (U) = 0}, i.e., dimH (U) is the
value of d for which the sharp transition from ∞ to 0 in Figure 4 occurs.
Lemma H.13. Main properties of Hausdorff measures:
i) If a > b ≥ 0, then H a (E) > 0 implies H b (E) = ∞ for all E ⊆ Rm (cf. Fig. 4).
ii) If f : Rm → Rn is Lipschitz with Lipschitz constant L, then
H d (f (E)) ≤ Ld H d (E)
for all E ⊆ Rm .
(373)
iii) H m (E) = λm (E) for all E ∈ B(Rm ).
Proof. See [43, Proposition 2.49] and [43, Theorem 2.53].
March 20, 2018
DRAFT
63
Lemma H.14. Let xi > 0 for all i ∈ I and set
M=
sup
X
xi .
(374)
J ⊆I : |J |<∞ i∈J
If M < ∞, then I is at most countable.
Proof. Suppose that M is finite. For every k ∈ N0 , set Ik = {i ∈ I : 1/(k + 1) ≤ xi < 1/k} and
Mk =
sup
X
xi .
(375)
J ⊆Ik :|J |<∞ i∈J
Since M < ∞ and Mk ≤ M , we must have Mk < ∞ for all k ∈ N0 . But for every k ∈ N0 , by the
S
definition of Ik , Mk can only be finite if |Ik | < ∞. Thus, |Ik | < ∞ for all k ∈ N0 . Since I = k∈N Ik ,
we conclude that I is at most countable.
A PPENDIX I
P ROPERTIES OF S ET-VALUED M APPINGS
A set-valued mapping Φ : T → 2R
m
associates to each x ∈ T a set Φ(x) ⊆ Rm . Many properties of
ordinary mappings such as, e.g., measurability, can be extended to set-valued mappings. In this appendix,
we first briefly review properties of set-valued mappings and then state a result for set-valued mappings
needed in the existence proof of a measurable decoder in Section V-A.
Definition I.1. (Closed-valuedness of set-valued mappings) [51, Section 5] A set-valued mapping Φ : T →
m
2R
is closed-valued if, for every t ∈ T , the set Φ(t) is closed.
Definition I.2. (Inverse image of a set-valued mapping) [51, Section 14] For a set-valued mapping
m
Φ : T → 2R , the inverse image Φ−1 (A) of A ⊆ Rm is
Φ−1 (A) = {t ∈ T : Φ(t) ∩ A =
6 ∅}.
(376)
Definition I.3. (Measurable set-valued mapping) [51, Section 14] Let (T , S (T )) be a measurable space.
A set-valued mapping Φ : T → 2R
image Φ−1 (O) ∈ S (T ).
m
is S (T )-measurable if, for every open set O ⊆ Rm , the inverse
m
Lemma I.1. [51, Theorem 14.3] Let (T , S (T )) be a measurable space and Φ : T → 2R closed-valued.
Then, Φ is S (T )-measurable if and only if Φ−1 (K) ∈ S (T ) for all compact sets K ⊆ Rm .
m
Lemma I.2. [51, Corollary 14.6] Let (T , S (T )) be a measurable space and Φ : T → 2R
a S (T )-
measurable closed-valued mapping. Then, there exists a S (T )-measurable mapping f : Φ−1 (Rm ) → Rm
such that f (t) ∈ Φ(t) for all t ∈ Φ−1 (Rm ).
March 20, 2018
DRAFT
64
Definition I.4. (Normal integrand) [51, Definition 14.27] Let (T , S (T )) be a measurable space. An
extended real-valued function f : T × Rm → R ∪ {∞} ∪ {−∞} is a normal integrand with respect to
S (T ) if its epigraphical mapping
m
Sf : T → 2R
×R
(377)
t 7→ {(x, α) ∈ Rm × R : f (t, x) ≤ α}
(378)
is closed-valued and S (T )-measurable.
Lemma I.3. [51, Example 14.31] Let T = Rn×m × Rn and suppose that f : T × Rm → R ∪{∞}∪{−∞}
is continuous. Then, f is a normal integrand with respect to B(T ).
Lemma I.4. [51, Proposition 14.33] Let (T , S (T )) be a measurable space. An extended real-valued
function f : T × Rm → R ∪ {∞} ∪ {−∞} is a normal integrand with respect to S (T ) if and only if,
for every α ∈ R ∪ {∞} ∪ {−∞}, the level-set mapping
m
Lα : T → 2 R
(379)
t 7→ {x ∈ Rm : f (t, x) ≤ α}
(380)
is S (T )-measurable and closed-valued.
We can now state the result on set-valued mappings needed to prove the existence of a measurable
decoder in Section V-A.
Lemma I.5. Let (T , S (T )) be a measurable space, α ∈ R, and f : T × Rm → R ∪ {∞} ∪ {−∞}.
Suppose that f is a normal integrand with respect to S (T ) and K ⊆ Rm is compact and nonempty.
Then, the following properties hold.
i) The set-valued mapping
m
PK : T → 2R
t 7→ {x ∈ K : f (t, x) ≤ α}
(381)
(382)
is S (T )-measurable and closed-valued.
ii) PK−1 (Rm ) = {t ∈ T : ∃x ∈ K with f (t, x) ≤ α} ∈ S (T ).
iii) There exists a S (T )-measurable mapping
pK : PK−1 (Rm ) → Rm
t 7→ pK (t) ∈ PK (t).
March 20, 2018
(383)
(384)
DRAFT
65
Proof. We start with the proof of i). For each t ∈ T , we can write PK (t) = Lα (t) ∩ K, where Lα is
the S (T )-measurable closed-valued level-set mapping from Lemma I.4. Since Lα is closed-valued and
the intersection of a closed set with a compact set is closed, PK is closed-valued. To prove that PK is
S (T )-measurable it suffices, thanks to Lemma I.1, to show that PK−1 (A) ∈ S (T ) for all compact sets
A ⊆ Rm . To this end, let A ⊆ Rm be an arbitrary but fixed compact set. Since the intersection of two
compact sets is compact it follows that K ∩ A is compact. As Lα is S (T )-measurable and K ∩ A is
−1
−1
compact, L−1
α (K ∩ A) ∈ S (T ) by Lemma I.1. Therefore, as Lα (K ∩ A) = PK (A), we can conclude
that PK−1 (A) ∈ S (T ). Since A was arbitrary, this proves i). Now, S (T )-measurability of PK implies
PK−1 (Rm ) ∈ S (T ), and thereby ii). Finally, the existence of the S (T )-measurable mapping pK in iii)
follows from i) and Lemma I.2.
A PPENDIX J
P ROPERTIES OF S EQUENCES OF F UNCTIONS IN S EVERAL VARIABLES
In this appendix, we summarize properties of sequences of functions in several variables needed in
the proof of Theorem III.2. We start with a result that establishes a sufficient condition for uniform
convergence.
Theorem J.1. (Weierstrass M -test) [39, Theorem 7.10] Consider a sequence {fn }n∈N of functions
fn : Rk → R and suppose that there exists a sequence {Mn }n∈N of nonnegative real numbers Mn
such that
for all x ∈ Rk and n ∈ N
|fn (x)| ≤ Mn
and
P
n∈N Mn
(385)
< ∞. Then, the sequence {sn }n∈N of partial sums
sn =
n
X
fi
(386)
i=1
converges uniformly to a function f :
Rk
→ R.
Next, we state a result that allows us to interchange the order of summation and differentiation for
certain sequences of differentiable functions. We start with the corresponding statement for differentiable
functions in one variable.
Theorem J.2. [39, Theorem 7.17] Consider a sequence {fn }n∈N of functions fn : R → R, each of
which is differentiable on the closed interval [a, b] ⊆ R, such that limn→∞ fn (x0 ) exists and is finite for
March 20, 2018
DRAFT
66
at least one x0 ∈ [a, b]. Let fn0 denote the derivative of fn , n ∈ N. If {fn0 }n∈N converges uniformly on
[a, b], then {fn }n∈N converges uniformly on [a, b] to a function f : R → R satisfying
df (x)
= lim fn0 (x)
n→∞
dx
for all x ∈ [a, b].
(387)
Corollary J.1. Consider a sequence {fn }n∈N of functions fn : Rk → R converging uniformly to f : Rk →
R. Suppose that there exists an i ∈ {1, . . . , k} such that the partial derivatives ∂fn (x)/∂xi , denoted by
fn0 (x), exist and are finite for all x ∈ Rk and n ∈ N. If the sequence {fn0 }n∈N converges uniformly, then
∂f (x)
= lim fn0 (x)
n→∞
∂xi
for all x ∈ Rk .
(388)
Proof. Suppose that the sequence {fn0 }n∈N converges uniformly. Let x = (x1 . . . xk )T ∈ Rk be arbitrary
but fixed and denote by {gn }n∈N the sequence of functions gn (t) = fn ((x1 . . . xi−1 t xi+1 . . . xk )T ).
Since {fn }n∈N converges uniformly to f by assumption, and gn (xi ) = fn (x) for all n ∈ N, it follows
that
lim gn (xi ) = lim fn (x)
n→∞
n→∞
= f (x).
(389)
(390)
For each n ∈ N, let gn0 denote the derivative of gn . Since {fn0 }n∈N converges uniformly, so does {gn0 }n∈N .
In particular, {gn0 }n∈N converges uniformly on a closed interval [a, b] ⊆ R containing xi . Therefore,
Theorem J.2 implies that {gn }n∈N converges uniformly on [a, b] to a function g : R → R satisfying
dg(t)
= lim gn0 (t)
n→∞
dt
for all t ∈ [a, b].
(391)
But g(xi ) = f (x) and gn0 (xi ) = fn0 (x) for all n ∈ N, which implies ∂f (x)/∂xi = limn→∞ fn0 (x). As x
was arbitrary, this finishes the proof.
A PPENDIX K
P ROPERTIES OF R EAL A NALYTIC M APPINGS AND C ONSEQUENCES T HEREOF
In this appendix, we review material on real analytic mappings our strong converse established in
Sections IV and VII is based on. We start with the definition of a real analytic mapping.
Definition K.1. (Real analytic mapping) [52, Definition 2.2.1] Let U be an open set in Rm .
i) A function f : U → R is real analytic on U if, for each x ∈ U , f may be represented by a convergent
power series in some open neighborhood of x; if U = Rm , then f is real analytic.
ii) A mapping f : U → Rn , x 7→ (f1 (x) . . . fn (x))T is real analytic on U if every component fi ,
i = 1, . . . , n, is real analytic on U ; if U = Rm , then f is real analytic.
March 20, 2018
DRAFT
67
Lemma K.1. [52, Corollary 1.2.4] If a power series
f (x) =
∞
X
j=0
aj (x − α)j
(392)
converges on an open interval I ⊆ R, then f is real analytic on I .
Lemma K.2. [52, Proposition 2.2.2] Let U and V be open sets in Rm . If f : U → R is real analytic on
U and g : V → R is real analytic on V , then f + g and f · g are both real analytic on U ∩ V . Furthermore,
if g(x) 6= 0 for all x ∈ U ∩ V , then f /g is real analytic on U ∩ V .
Corollary K.1. All polynomials on Rm are real analytic.
Lemma K.3. [52, Proposition 2.2.3] Let U be an open set in Rm and suppose that f : U → R is real
analytic on U . Then, the partial derivatives—of arbitrary order—of f are real analytic on U .
Lemma K.4. [52, Proposition 2.2.8] Let U be an open set in Rm and V an open set in Rn . Suppose
that f : U → Rn is real analytic on U and g : V → Rk is real analytic on V . Then, for every x ∈ U with
f (x) ∈ V , there exists an open set W ⊆ Rm such that x ∈ W and (g ◦ f )|W : W → Rk is real analytic
on W .
Corollary K.2. If f : Rm → Rn and g : Rn → Rk are both real analytic, then g ◦ f : Rm → Rk is real
analytic.
Lemma K.5. [52, p. 83] A real analytic function f : Rm → R vanishes either identically or on a set of
Lebesgue measure zero.
The following definition extends the notion of a diffeomorphism of class C r [34, Definition 3.1.18]
to a real analytic diffeomorphism.
Definition K.2. (Real analytic diffeomorphism) Let U ⊆ Rn be an open set and ρ : U → Rn real analytic.
The mapping ρ is a real analytic diffeomorphism if V = ρ(U) is an open set and the inverse ρ−1 of ρ
exists on V and is real analytic on V .
Theorem K.1. [52, Theorem 2.5.1] Let U ⊆ Rm be open and f : U → Rm real analytic. If Jf (x0 ) > 0
for some x0 ∈ U , then f −1 exists on an open set V ⊆ Rm containing f (x0 ) and is real analytic on V .
Corollary K.3. Let U be an open set in Rm and suppose that f : U → Rm is real analytic on U . If there
exists an x0 ∈ U such that Jf (x0 ) > 0, then there exists an r > 0 such that f |Bm (x0 ,r) is a real analytic
diffeomorphism.
March 20, 2018
DRAFT
68
Proof. Suppose that there exists an x0 ∈ U such that Jf (x0 ) > 0. Since Jf (x0 ) > 0, Theorem K.1
implies that there exists an open set V ⊆ Rm such that f (x0 ) ∈ V , and f −1 exists on V and is real
analytic on V . Since V is open and f (x0 ) ∈ V , there must exist an ε > 0 such that Bm (f (x0 ), ε) ⊆ V . By
continuity of f , which follows from real analyticity, there must exist an r > 0 such that f (Bm (x0 , r)) ⊆
Bm (f (x0 ), ε) ⊆ V . We set W = f (Bm (x0 , r)). Summarizing, f is real analytic on Bm (x0 , r) and f −1
exists on W = f (Bm (x0 , r)) and is real analytic on W . It remains to show that W is open. This follows
by noting that thanks to W = (f −1 )−1 (Bm (x0 , r)), W is the inverse image of an open set under a real
analytic (and therefore continuous) mapping, and is hence open. We conclude that f |Bm (x0 ,r) is a real
analytic diffeomorphism.
Corollary K.4. Let U be an open set in Rm and suppose that f : U → Rn with n ≥ m is real analytic on
U . If there exists an x0 ∈ U such that Jf (x0 ) > 0, then there exists an r > 0 such that f is one-to-one
on Bm (x0 , r).
Proof. Suppose that there exists an x0 ∈ U such that Jf (x0 ) > 0. Then, Jf (x0 ) > 0 and n ≥ m imply
rank(Df (x0 )) = m. Thus, the n × m matrix Df (x0 ) has m linearly independent row vectors. Denote
the indices of m such linearly independent row vectors by {i1 , . . . , im } and consider the mapping
g : U → Rm
(393)
x 7→ (fi1 (x) . . . fim (x))T .
(394)
Since f is real analytic on U , so is g . Furthermore, rank(Dg(x0 )) = m and hence Jg(x0 ) > 0. Corollary
K.3 therefore implies the existence of an r > 0 such that g|Bm (x0 ,r) is a real analytic diffeomorphism. In
particular, g is one-to-one on Bm (x0 , r), which in turn implies that f is one-to-one on Bm (x0 , r).
Next, we show that the square root is a real analytic diffeomorphism on the set of positive real numbers.
Lemma K.6. The function
√
: R+ → R+ , x 7→
√
x, where R+ = {x ∈ R : x > 0}, is a real analytic
diffeomorphism.
Proof. The function (·)2 : R+ → R+ , x 7→ x2 is real analytic by Corollary K.1. Let y ∈ R+ be arbitrary
√
but fixed and set x = y . Since dx2 /dx = 2x > 0, Theorem K.1 implies that there exists an r > 0
√
such that the inverse of (·)2 , given by ·, exists on (y − r, y + r) and is real analytic on (y − r, y + r).
√
√
As y was arbitrary, it follows that · is real analytic on R+ . Finally, since (·)2 is the inverse of · and
p
√
R+ = R+ is open, · is a real analytic diffeomorphism.
March 20, 2018
DRAFT
69
Lemma K.7. If f : Rm → Rn is real analytic, so is Jf . In particular, Jf vanishes either identically or
on a set of Lebesgue measure zero.
p
Proof. Suppose that f : Rm → Rn is real analytic. Recall that Jf (x) = g(x), where g : Rm → R,
det(Df (x)(Df (x))T ) if n < m
g(x) =
(395)
det((Df (x))T Df (x)) else.
√
Lemmata K.2 and K.3 imply that g is real analytic. As · is real analytic on R+ by Lemma K.6, real
analyticity of Jf follows from Lemma K.4. Finally, as Jf is real analytic, it vanishes by Lemma K.5
either identically or on a set of Lebesgue measure zero.
We have the following important properties of real analytic mappings.
Lemma K.8. Let h : Rs → Rm be a real analytic mapping of s-dimensional Jacobian Jh 6≡ 0. Then, the
following properties hold.
i) The set O = {z ∈ Rs : Jh(z) > 0} is open and satisfies λs (Rs \O) = 0.
ii) For each set A ∈ B(Rs ) of positive Lebesgue measure there exists a set B ∈ B(Rs ) of positive
Lebesgue measure such that B ⊆ A and the mapping h|B is an embedding, i.e., Jh(z) > 0 for all
z ∈ B and h is one-to-one on B .
Proof. Openness of O follows from continuity of Jh, and Lemma K.7 together with Jh 6≡ 0 implies
λs (Rs \O) = 0. To prove ii), consider A ∈ B(Rs ) of positive Lebesgue measure. As λs (Rs \O) = 0,
with O from i), it follows that λs (A ∩ O) > 0. Thus, by Lemma H.7, there must exist a z0 ∈ A ∩ O
such that
λs (Bs (z0 , r) ∩ A ∩ O) > 0
for all r > 0.
(396)
Since Jh(z0 ) > 0, which follows from z0 ∈ O, Corollary K.4 implies that there must exist an r0 > 0
such that h is one-to-one on Bs (z0 , r0 ). Setting B = Bs (z0 , r0 ) ∩ A ∩ O concludes the proof.
Definition K.3. (Real analytic submanifold) [52, Definition 2.7.1] A subset M ⊆ Rn is an m-dimensional
real analytic submanifold of Rn if, for each y ∈ M, there exist an open set U ⊆ Rm and a real analytic
immersion f : U → Rn such that y ∈ f (U) and open subsets of U are mapped onto relatively open
subsets in M. Here, a subset W ⊆ M is called relatively open in M if there exists an open set V ⊆ Rn
such that W = V ∩ M.
March 20, 2018
DRAFT
70
Lemma K.9. Let M ⊆ Rs be a t-dimensional real analytic submanifold of Rs and z0 ∈ M. Then, there
exist a real analytic embedding ζ : Rt → Rs and an η > 0 such that
ζ(0) = z0 ,
(397)
Bs (z0 , η) ∩ M ⊆ ζ Rt ,
(398)
and ζ Rt is relatively open in M, i.e., there exists an open set V ⊆ Rs such that ζ Rt = V ∩ M.
Proof. Since M ⊆ Rs is a t-dimensional real analytic submanifold of Rs , Definition K.3 implies the
existence of an open set U ⊆ Rt and a real analytic immersion ξ : U → Rs that maps open subsets of U
onto relatively open subsets in M and satisfies z0 = ξ(u0 ) with u0 ∈ U . As ξ is an immersion,
rank(Dξ(v)) = t
for all v ∈ U.
(399)
Since U is open and u0 ∈ U , there exists a ρ > 0 such that Bt (u0 , ρ) ⊆ U , which implies in turn
ξ Bt (u0 , ρ) ⊆ ξ(U).
(400)
Using Corollary K.4 and Jξ(u0 ) > 0 (recall that ξ is an immersion) we may choose ρ sufficiently small
for ξ to be one-to-one on Bt (u0 , ρ). As ξ(Bt (u0 , ρ)) is relatively open in M, there must exist an open
set V ⊆ Rs such that
ξ Bt (u0 , ρ) = V ∩ M.
(401)
Now, as V is open and z0 = ξ(u0 ) ∈ V , we can find an η > 0 such that Bs (z0 , η) ⊆ V , which, together
with (401), yields
Bs (z0 , η) ∩ M ⊆ ξ Bt (u0 , ρ) .
(402)
Let κ : Rt → Bt (0, ρ) be the real analytic diffeomorphism constructed in Lemma K.10 below and set
ζ : Rt → Rs
(403)
v 7→ ξ(κ(v) + u0 ).
(404)
The mapping ζ is a real analytic mapping by Lemmata K.2 and K.4. Clearly, ζ is one-to-one on Rt as κ is
a diffeomorphism and ξ is one-to-one on Bt (u0 , ρ). Since ζ Rt = ξ(Bt (u0 , ρ)), (402) establishes (398)
and (401) proves that ζ Rt is relatively open in M. Finally, (397) follows from ζ(0) = ξ(κ(0) + u0 ) =
ξ(u0 ) = z0 , where in the second equality we used (408) in Lemma K.10 below. It remains to show that
ζ is an immersion, which is effected by proving that rank(Dζ(v)) = t for all v ∈ Rt . The chain rule
implies
Dζ(v) = (Dξ)(κ(v) + u0 )Dκ(v)
March 20, 2018
for all v ∈ Rt .
(405)
DRAFT
71
It now follows
i) from (410) in Lemma K.10 below that rank(Dκ(v)) = t for all v ∈ Rt , and
ii) from (399) and κ(v) + u0 ∈ Bt (u0 , ρ) ⊆ U that rank((Dξ)(κ(v) + u0 )) = t for all v ∈ Rt .
Applying Lemma K.12 to (Dξ)(κ(v) + u0 ) ∈ Rs×t and Dκ(v) ∈ Rt×t , and using i) and ii) above yields
rank(Dζ(v)) ≥ t for all v ∈ Rt , which in turn implies Jζ(v) > 0 for all v ∈ Rt , thereby concluding
the proof.
Lemma K.10. For ρ > 0, the mapping
κ : Rk → Bk (0, ρ)
(406)
x 7→ p
(407)
is a real analytic diffeomorphism on Rk satisfying
ρx
1 + kxk22
κ(0) = 0,
(408)
κ(Rk ) = Bk (0, ρ),
rank(Dκ(x)) = k
(409)
for all x ∈ Rk .
(410)
Proof. It follows immediately from the definition of κ that κ(0) = 0. The mapping κ is real analytic
thanks to Lemmata K.2, K.4, and K.6. Now, consider the mapping
σ : Bk (0, ρ) → Rk
y 7→ p
ρ2
(411)
y
− kyk22
.
(412)
Again, σ is real analytic thanks to Lemmata K.2, K.4, and K.6. Since (κ◦σ)(y) = y for all y ∈ Bk (0, ρ),
it follows that κ(Rk ) = Bk (0, ρ). Moreover, as (σ ◦ κ)(x) = x for all x ∈ Rk , σ is the inverse of κ,
which establishes that κ is a real analytic diffeomorphism on Rk . Finally, since (σ ◦ κ)(x) = x for all
x ∈ Rk , the chain rule implies Ik = D(σ ◦ κ)(x) = (Dσ)(κ(x))Dκ(x) for all x ∈ Rk , which yields
(410).
Proposition K.1. [52, Proposition 2.7.3] Let M ⊆ Rn . The following statements are equivalent:
i) M is a m-dimensional real analytic submanifold of Rn .
ii) For each z ∈ M there exist an open set U ⊆ Rn with z ∈ U , a real analytic diffeomorphism
ρ : U → Rn , and a m-dimensional linear subspace L ⊆ Rn , such that
ρ(M ∩ U) = ρ(U) ∩ L.
March 20, 2018
(413)
DRAFT
72
Proposition K.1 allows us to state the following sufficient condition on the inverse image of a point
under a real analytic function to result in a real analytic submanifold.
Lemma K.11. Let ψ : Rs → R be a real analytic function, y0 ∈ ψ(Rs ), and set My0 = ψ −1 ({y0 }).
Suppose that Jψ(z) > 0 for all z ∈ My0 . Then, My0 is a (s − 1)-dimensional real analytic submanifold
of Rs .
Proof. We may assume, w.l.o.g., that y0 = 0 (if y0 6= 0, we set ψ̃(z) = ψ(z) − y0 and prove the lemma
with ψ̃ in place of ψ , noting that My0 = ψ̃ −1 ({0}) and J ψ̃(z) = Jψ(z) for all z ∈ Rs .) The proof
is effected by verifying that M0 satisfies Statement ii) of Proposition K.1. Let z0 ∈ M0 be arbitrary
but fixed and set aT
1 = Dψ(z0 ). Since Jψ(z0 ) > 0 by assumption, we must have a1 6= 0. Choose
a2 , . . . , as ∈ Rs such that A = (a1 . . . as )T is a regular matrix and consider the mapping
ρ : Rs → Rs
T
z 7→ ψ(z) aT
2 z . . . as z
(414)
T
.
(415)
Note that Dρ(z0 ) = A. Since rank(A) = s by construction, Jρ(z0 ) > 0. It therefore follows from
Corollary K.3 that there exists an r > 0 such that ρ|Bs (z0 ,r) is a real analytic diffeomorphism. Finally,
we can write
ρ(M0 ∩ Bs (z0 , r)) = ρ({z ∈ Bs (z0 , r) : ψ(z) = 0})
(416)
= ρ({z ∈ Bs (z0 , r) : eT
1 ρ(z) = 0})
(417)
= ρ(Bs (z0 , r)) ∩ L,
(418)
where (416) follows from M0 = ψ −1 ({0}), (417) is by (415), and in (418) we set L = {(w1 . . . ws )T ∈
Rs : w1 = 0}.
Lemma K.12. [37, Property (c), Chapter 0.4.5] If A ∈ Rm×k and B ∈ Rk×n , then
rank(AB) ≥ rank(A) + rank(B) − k.
(419)
R EFERENCES
[1] G. Alberti, H. Bölcskei, C. De Lellis, G. Koliander, and E. Riegler, “Lossless linear analog compression,” in Proc. IEEE
Int. Symp. Inf. Theory, 2016, pp. 2789–2793.
[2] D. L. Donoho and P. B. Stark, “Uncertainty principles and signal recovery,” SIAM J. Appl. Math., vol. 49, no. 3, pp.
906–931, Jun. 1989.
[3] D. L. Donoho, “Compressed sensing,” IEEE Trans. Inf. Theory, vol. 52, no. 4, pp. 1289–1306, Apr. 2006.
March 20, 2018
DRAFT
73
[4] E. J. Candès, J. Romberg, and T. Tao, “Robust uncertainty principles: Exact signal reconstruction from highly incomplete
frequency information,” IEEE Trans. Inf. Theory, vol. 52, no. 2, pp. 489–509, Feb. 2006.
[5] E. J. Candès, Y. C. Eldar, D. Needell, and P. Randall, “Compressed sensing with coherent and redundant dictionaries,”
Appl. Comp. Harm. Anal., vol. 31, no. 1, pp. 59–73, Jul. 2011.
[6] P. Kuppinger, G. Durisi, and H. Bölcskei, “Uncertainty relations and sparse signal recovery for pairs of general signal
sets,” IEEE Trans. Inf. Theory, vol. 58, no. 1, pp. 263–277, Jan. 2012.
[7] R. Gribonval and M. Nielsen, “Sparse representations in unions of bases,” IEEE Trans. Inf. Theory, vol. 49, no. 12, pp.
3320–3325, Jan. 2003.
[8] M. Elad and A. M. Bruckstein, “A generalized uncertainty principle and sparse representation in pairs of bases,” IEEE
Trans. Inf. Theory, vol. 48, no. 9, pp. 2558–2567, Sep. 2002.
[9] D. L. Donoho and M. Elad, “Optimally sparse representation in general (nonorthogonal) dictionaries via `1 minimization,”
in Proc. Natl. Acad. Sci. USA, vol. 100, no. 5, 2003, pp. 2197–2002.
[10] J. A. Tropp, “Greed is good: Algorithmic results for sparse approximation,” IEEE Trans. Inf. Theory, vol. 50, no. 10, pp.
2231–2242, Oct. 2004.
[11] L. R. Welch, “Lower bounds on the maximum cross correlation of signals,” IEEE Trans. Inf. Theory, vol. 20, no. 3, pp.
397–399, May 1974.
[12] E. J. Candès and T. Tao, “Near-optimal signal recovery from random projections: Universal encoding strategies?” IEEE
Trans. Inf. Theory, vol. 52, no. 12, pp. 5406–5425, Dec. 2006.
[13] J. A. Tropp, “On the conditioning of random subdictionaries,” Appl. Comp. Harm. Anal., vol. 25, pp. 1–24, 2008.
[14] G. Pope, A. Bracher, and C. Studer, “Probabilistic recovery guarantees for sparsely corrupted signals,” IEEE Trans. Inf.
Theory, vol. 59, no. 5, pp. 3104–3116, May 2013.
[15] R. G. Baraniuk, M. Davenport, R. DeVore, and M. Wakin, “A simple proof of the restricted isometry property for random
matrices,” Constructive Approximation, vol. 28, no. 3, pp. 253–263, Dec. 2008.
[16] Y. Wu and S. Verdú, “Rényi information dimension: Fundamental limits of almost lossless analog compression,” IEEE
Trans. Inf. Theory, vol. 56, no. 8, pp. 3721–3748, Aug. 2010.
[17] ——, “Optimal phase transitions in compressed sensing,” IEEE Trans. Inf. Theory, vol. 58, no. 10, pp. 6241–6263, Oct.
2012.
[18] D. Stotz, E. Riegler, E. Agustsson, and H. Bölcskei, “Almost lossless analog signal separation and probabilistic uncertainty
relations,” IEEE Trans. Inf. Theory, vol. 63, no. 9, pp. 5445–5460, Sep. 2017.
[19] G. Koliander, G. Pichler, E. Riegler, and F. Hlawatsch, “Entropy and source coding for integer-dimensional singular random
variables,” IEEE Trans. Inf. Theory, vol. 62, no. 11, pp. 6124–6154, 2016.
[20] Y. C. Eldar and M. Mishali, “Robust recovery of signals from a structured union of subspaces,” IEEE Trans. Inf. Theory,
vol. 55, no. 11, pp. 5302–5316, Nov. 2009.
[21] R. G. Baraniuk, V. Cevher, M. F. Duarte, and C. Hegde, “Model-based compressive sensing,” IEEE Trans. Inf. Theory,
vol. 56, no. 4, pp. 1982–2001, Apr. 2010.
[22] P. Feng and Y. Bresler, “Spectrum-blind minimum-rate sampling and reconstruction of multiband signals,” in Proc. IEEE
Int. Conf. Acoust. Speech Signal Process., 1996, pp. 1688–1691.
[23] M. Mishali and Y. C. Eldar, “Blind multiband signal reconstruction: Compressed sensing for analog signals,” IEEE Trans.
Signal Process., vol. 57, no. 3, pp. 993–1009, Mar. 2009.
[24] R. G. Baraniuk and M. B. Wakin, “Random projections of smooth manifolds,” Found. Comput. Math., vol. 9, no. 1, pp.
51–77, Feb 2009.
March 20, 2018
DRAFT
74
[25] M. Stojnic, F. Parvaresh, and B. Hassibi, “On the reconstruction of block-sparse signals with an optimal number of
measurements,” IEEE Trans. Signal Process., vol. 57, no. 8, pp. 3075–3085, Aug. 2009.
[26] Y. C. Eldar, P. Kuppinger, and H. Bölcskei, “Block-sparse signals: Uncertainty relations and efficient recovery,” IEEE
Trans. Signal Process., vol. 58, no. 6, Jun. 2010.
[27] J. Huang and T. Zhang, “The benefit of group sparsity,” Ann. Stat., vol. 38, no. 4, pp. 1978–2004, 2010.
[28] E. J. Candès and B. Recht, “Exact matrix completion via convex optimization,” Found. Comput. Math., vol. 9, no. 6, pp.
717–772, 2009.
[29] E. J. Candès and Y. Plan, “Tight oracle inequalities for low-rank matrix recovery from a minimal number of noisy random
measurements,” IEEE Trans. Inf. Theory, vol. 4, no. 57, pp. 2342–2359, Apr. 2011.
[30] E. Riegler, D. Stotz, and H. Bölcskei, “Information-theoretic limits of matrix completion,” in Proc. IEEE Int. Symp. Inf.
Theory, Jun. 2015, pp. 1836–1840.
[31] J. C. Robinson, Dimensions, Embeddings, and Attractors.
[32] K. Falconer, Fractal Geometry, 1st ed.
Cambridge, NY: Cambridge University Press, 2011.
New York, NY: Wiley, 1990.
[33] T. Kawabata and A. Dembo, “The rate-distortion dimension of sets and measures,” IEEE Trans. Inf. Theory, vol. 40, no. 5,
pp. 1564–1572, Sep. 1994.
[34] H. Federer, Geometric Measure Theory.
New York, NY: Springer, 1969.
[35] P. Mattila, Geometry of Sets and Measures in Euclidean Space: Fractals and Rectifiability. Cambridge, UK: Cambridge
Univ. Press, 1999.
[36] S. G. Krantz and H. R. Parks, Geometric Integration Theory.
[37] R. A. Horn and C. R. Johnson, Matrix Analysis, 2nd ed.
Princeton, NJ: Birkhäuser, 2008.
Cambridge, U.K.: Cambridge Univ. Press, 2013.
[38] K. Stromberg, “An elementary proof of Steinhaus’s theorem,” in Proc. of the American Mathematical Society, vol. 36,
no. 1, 1976, p. 308.
[39] W. Rudin, Principles of Mathematical Analysis, 3rd ed.
[40] J. M. Lee, Introduction to Smooth Manifolds, 2nd ed.
New York, NY: McGraw-Hill, 1976.
New York, NY: Springer, 2013.
[41] G. Folland, Real Analysis: Modern Techniques and Their Applications, 2nd ed.
[42] H. H. Sohrab, Basic Real Analysis, 2nd ed.
New York, NY: Wiley, 2013.
New York, NY: Birkhäuser, 2014.
[43] L. Ambrosio, N. Fusco, and D. Pallara, Functions of Bounded Variation and Free Discontinuity Problems.
New York,
NY: Oxford Univ. Press, 2000.
[44] R. G. Bartle, The Elements of Integration and Lebesgue Measure.
[45] T. Tao, An Introduction to Measure Theory.
New York, NY: Wiley, 1995.
Providence RI: AMS, 2011.
[46] L. Gasiński and N. S. Papageorgiou, Exercises in Analysis-Part 2: Nonlinear Analysis, ser. Problem Books in Mathematics.
Cham, Switzerland: Springer, 2016.
[47] T.-W. Ma, Classical Analysis on Normed Spaces.
Singapore: World Scientific Pub. Co., 1995.
[48] F. W. Warner, Foundations of Differentiable Manifolds and Lie Groups.
[49] C. H. Edwards, Advanced Calculus of Several Variables.
New York, NY: Springer, 1971.
New York, NY: Academic Press, Inc., 1973.
[50] A. Sard, “The measure of the critical values of differentiable maps,” Bull. Amer. Math. Soc., vol. 48, no. 12, pp. 883–890,
1942.
[51] R. T. Rockafellar and R. J.-B. Wets, Variational Analysis, 3rd ed.
Heidelberg, Germany: Springer, 2009.
[52] S. G. Krantz and H. R. Parks, A Primer of Real Analytic Functions, 2nd ed.
March 20, 2018
Basel, Switzerland: Birkhäuser, 2002.
DRAFT
| 7cs.IT
|
Elite Bases Regression: A Real-time Algorithm for
Symbolic Regression
Chen Chen†‡ , Changtong Luo∗† , Zonglin Jiang†‡
arXiv:1704.07313v2 [cs.DS] 15 May 2017
† State
Key Laboratory of High Temperature Gas Dynamics, Institute of Mechanics, Chinese Academy of Sciences
Beijing 100190, China
‡ School of Engineering Sciences, University of Chinese Academy of Sciences, Beijing 100049, China
∗ Email: [email protected]
Abstract—Symbolic regression is an important but challenging
research topic in data mining. It can detect the underlying
mathematical models. Genetic programming (GP) is one of the
most popular methods for symbolic regression. However, its
convergence speed might be too slow for large scale problems
with a large number of variables. This drawback has become a
bottleneck in practical applications. In this paper, a new nonevolutionary real-time algorithm for symbolic regression, Elite
Bases Regression (EBR), is proposed. EBR generates a set of
candidate basis functions coded with parse-matrix in specific
mapping rules. Meanwhile, a certain number of elite bases are
preserved and updated iteratively according to the correlation
coefficients with respect to the target model. The regression
model is then spanned by the elite bases. A comparative study
between EBR and a recent proposed machine learning method
for symbolic regression, Fast Function eXtraction (FFX), are conducted. Numerical results indicate that EBR can solve symbolic
regression problems more effectively.
I. I NTRODUCTION
Symbolic regression aims to find a mathematical model that
can describe and predict a given system based on observed
input-response data. It plays an increasingly important role
in many engineering applications including signal processing
[23], system identification [20], industrial data analysis [13],
etc. Unlike conventional linear/nonlinear regression methods
that assume a linear/nonlinear trend, or require you to provide
a mathematical model of a given form, symbolic regression
searches an appropriate model from a space of all possible
expressions S defined by a set of given arithmetic operations
(e.g., +, −, ×, ÷, etc.) and mathematical functions (e.g., sin,
cos, exp, ln, etc.). Mathematically, symbolic regression finds
the best combination of these operations and functions, and
optimizes the model structure and coefficients simultaneously,
which can be described as follows:
X
f (x(i) ) − yi ,
(1)
f ∗ = arg min
f ∈S
i
where x(i) ∈ Rd , yi ∈ R are numeric input-response data,
and f is the model function. However, given that symbolic
regression is a kind of global optimization problem in data
mining, there are still several difficulties in dealing with both
structure optimization and coefficient optimization at the same
time [2]. Hence, how to use a appropriate method to solve a
symbolic regression problem is considered as a kaleidoscope
in this research field [5].
Genetic programming (GP) [12], as a evolutionary computing (EC) technique, is one of the most popular methods for
symbolic regression in recent years. Corresponding different
improved versions of basic GP have also been proposed
continually, for instance, linear genetic programming (LGP)
[9], gene expression programming (GEP) [6], parse-matrix
evolution (PME) [14], etc. The core idea of GP is to apply
Darwin’s theory of natural evolution to the artificial world of
computers and modeling. Theoretically, GP can get accurate
results provided that the computation time is long enough.
However, due to its stochasticity, GP is difficult to realize
the real-time calculation and hard to give repeated results. In
addition, the convergence speed of GP might be too slow for
large scale problems with a large number of variables. Hence,
GP’s practical applications are limited.
To overcome these difficulties, more recently, a number of
researchers have focused mainly on using non-evolutionary
optimization methods to solve symbolic regression problems.
McConaghy [16] presented the first non-evolutionary algorithm based on machine learning for symbolic regression,
which confined its search space to generalized linear space.
Icke & Bongard [10] proposed a hybrid algorithm which
combined deterministic machine learning method and conventional GP. Worm [22] introduced a deterministic machine
learning algorithm, Prioritized Grammar Enumeration (PGE),
in his thesis, which made a large reduction to the search
space. Deklel et al. [3] presented a new approach based on
artificial neural networks, which could solve problems with
large number of inputs and even more complex examples and
applications.
Among these non-evolutionary methods, FFX is the first
deterministic symbolic regression implementation. Based on
generalized linear model (GLM), FFX applies a kind of machine learning method, namely pathwise regularized learning,
to identify the best coefficients and bases in GLM, which is
given by
β ∗ = min ky − B(x) · βk2 + λ2 kβk2 + λ1 kβk1 ,
(2)
where B(x) = (φ1 (x), φ2 (x), ..., φN (x)) represents the vector
of N univariate and bivariate generated bases and β is the
regression parameter of GLM. λ1 and λ2 are set to λ1 = λ
and λ2 = (1 − ρ)λ respectively, where λ is the regularization
weight. It is reported that FFX can be an order of magnitude
faster than GP, and has been successfully applied to analog
circuit design and modeling [7], the reliability analysis for
analog circuit [15], etc.
However, note that pathwise regularized learning (2) needs
to solve a quadratic optimization problem, which is equal
to solve a large linear system. With the increase of basis
number N , the computation cost will increase quadratically.
This restricts the speed of FFX in further promote for large
scale problems.
In this paper, we propose a new non-evolutionary algorithm,
Elite Bases Regression (EBR), to solve symbolic regression
problems. Different from FFX, most of the generated bases are
discarded, and simultaneously, only elite bases are preserved
and updated iteratively according to the correlation coefficients
with respect to the target model. The regression model is then
spanned by the elite bases. This makes EBR do not need
to solve a large-scale system of linear equations. Hence, it
can save computation time with little memory overhead. The
performance of EBR is compared with FFX. Numerical results
indicate that EBR has lower normalized mean square error
(NMSE) and concise regression models than FFX’s.
The presentation of this paper is organized as follows:
related concepts used in EBR algorithm is introduced in
Section II, EBR algorithm is described in Section III, and
experiments and results are presented and discussed in Section
IV. The paper is concluded in Section V with remarking the
future work.
II. R ELATED CONCEPTS
USED IN
EBR
ALGORITHM
A. Generalized linear model
Generalized linear model (GLM) [18] is a generalization of
classical linear regression model. GLM aims to find f ∗ in a
finite dimensional space of functions spanned by a set of given
basis functions. In other words, GLM specifies a set of bases
φ0 , φ1 , ..., φN from RN to R and finds f ∗ in the form of a
linear combination of N basis functions φi , i = 1, 2, ..., N :
f ∗ = β0 +
N
X
βi φi (x),
(3)
i=1
where βi is the regression parameter. In Elite Basis Regression
(EBR) algorithm, the regression model is spanned by a set of
elite bases. The number of elite bases is denoted by npresv .
EBR algorithm is inspired from the expansion method in
mathematics. In theoretical analysis, two main methods of
linear expansion, namely Taylor series and Fourier series, are
very powerful tools that are widely applied to many research
fields [17], [4]. However, Taylor series can only be used in
local expansion (the neighborhood of a certain point) with
special functions, while Fourier series is utilized in periodic
functions exclusively. We hope to find a global and universal
expansion strategy in practical applications. This motivates
us to design such a kind of global and universal linear
approximation method for symbolic regression.
B. Correlation coefficient
Correlation coefficient aims to measure linear relationship
between two vectors, and has a wide application scope in
statistical analysis. Suppose a function with n continuous
variables
f (x) = f (x1 , x2 , ..., xn ) , xi ∈ [ai , bi ], i = 1, 2, ..., n.
(4)
For each variable xi , a column vector xi is defined after a
set of random sample points in [ai , bi ] are generated. That is
T
(m)
(2)
(1)
Xi = xi = xi , xi , ..., xi
, where m is the number
of sample points. Thus, we can get a column vector with m
components respect to the initial function (4)
(1)
(1) (1)
xn
x2
x1
(2)
(2) (2)
xn
x1 x2
F (X) = F
.. , .. , . . . , ..
.
. .
(m)
(m)
(m)
x2
x1
xn
(1)
(1)
(1)
(1)
f x1 , x2 , . . . , xn
f
(2)
(2)
(2)
(2)
f x1 , x2 , . . . , xn
f
= .
=
.
..
.
.
(m)
f
(m)
(m)
(m)
f x1 , x2 , . . . , xn
(5)
From the above discussion, for an n-dimensional problem
in EBR, the target model f (x1 , x2 , ..., xn ) and a certain
generated basis φ (x1 , x2 , ..., xn ) can be regarded as two
column vectors of m components after sampling, namely
T
Y = F (X) = f (1) , f (2) , · · · , f (m) and ξΦ = Φ (X) =
T
φ(1) , φ(2) , · · · , φ(m) . The correlation coefficient of these
two vectors, Y and ξΦ , can be expressed as
Cov (ξΦ , Y )
p
,
ρξΦ ,Y = p
D(ξΦ ) D(Y )
(6)
where the operator D represents variance.
Note that the basis functions might be nonlinear, but the
ensemble process of the GLM is still linear. Therefore, correlation test is effective in EBR. That is, if |ρξΦ ,Y | is close to 1,
Y and ξΦ are closely related. Particularly, when |ρξΦ ,Y | = 1,
Y and ξΦ are linearly correlated in probability one.
In fact, once we regard two nonlinear functions as two
column vectors after sampling, it still makes sense that correlation coefficient can be used for correlation analysis of
nonlinear function. To give an illustrative example, consider a
two-dimensional test function
f (x1 , x2 ) = x1 x2 + sin ((x1 − 1) (x2 − 1)) ,
(7)
and a certain basis function
φ(x1 , x2 ) = x1 x2 ,
(8)
where x ∈ [−3, 3]2 . To enhance the stability of the test, the
distribution of sample points in [−3, 3]2 should be as uniform
as possible by using controlled sampling method of Latin
hypercube design [1]. Then, a correlation coefficient of the
Fig. 2.
(a) The test function.
An example of parse-matrix encoding process of EBR.
We can see that the encoding of parse-matrix is a natural
and easy process. Note that the second and the third columns
a.2 , a.3 are used to control the dimension of a given problem.
The parse-matrix encoding scheme ensures the generated candidate basis functions can cover all possible bases, according
to the mapping rules in Table I. In the following section, we
will use this table to do our numerical experiments (see Section
IV).
III. E LITE BASIS R EGRESSION
ALGORITHM
two vector functions Y and ξΦ could be obtained, which is
|ρξΦ ,Y | = 0.9838. This means Y and ξΦ are closely related.
Note that, as is shown in Fig. 1, the simple basis function (8)
successfully ‘sketch out’ the landscape of test function (7).
The new algorithm is a kind of deterministic linear approximation method. It does not rely on other GP method. To
illustrate our proposed algorithm more clearly, in this section,
EBR is introduced by two main parts, which are generation
and preservation of the bases in Section III-A and ensemble
and evaluation of the model in Section III-B. Then, we discuss
the method of complexity control in EBR in Section III-C.
Finally, the whole procedure of EBR is introduced in Section
III-D.
C. Parse-matrix encoding scheme
A. Generation and preservation of bases
The generation of basis functions is a crucial step in EBR.
However, multiple nested-loop used to enumerate the possible
bases in FFX is not easy to extend. On the other hand, the
complexity of bases is implemented with the if-else statement,
which makes FFX difficult to control the complexity and limits
its ability of modeling highly nonlinear target function. Hence,
parse-matrix encoding scheme becomes a good candidate for
bases generation engine of EBR.
Parse-matrix encoding scheme was initially provided for
parse-matrix evolution (PME) [14], a special version of GP.
PME use a two-dimensional matrix with integer entries to
express a chromosome (individual), which can carry more information than conventional chromosome representations [8],
[19]. The matrix representation makes PME easy to control
the complexity and simple to program.
In EBR, basis functions are coded with parse-matrix encoding scheme. This process could be sketched in Fig. 2. Suppose
an example mapping table defined as Table I. Then, a given
basis function φ = sin(x + y) can be produced in the steps
listed in Table II. According to the mapping table (see Table
I) and the encoding steps (see Table II), the basis function
φ = sin(x + y) can be described by a parse-matrix of order
3 × 3 as follows:
1 1 2
A = 3 3 2
(9)
12 3 1
In EBR, the basis function is coded with parse-matrix
encoding scheme in specific mapping rules (refer to Table
I), which has been discussed in Section II-C. Enumeration
method is used to generate a set of candidate basis functions.
This process ensures the generated bases can cover all possibilities in given arithmetic operations (e.g., +, −, ×, ÷,
etc.) and mathematical functions (e.g., sin, cos, exp, ln, etc.).
Simultaneously, npresv elite bases are preserved and updated
iteratively according to the correlation coefficients with respect
to the given target model.
(b) The basis function.
Fig. 1. An example of correlation analysis applied to 2D nonlinear function.
Remark 1. The number of preserved elite bases npresv determines the most computation costs of EBR and the complexity
of the regression model. A more detailed discussion of this
control parameter will be conducted in Section IV-C.
Remark 2. If the the correlation coefficients of two generated
bases with respect to the target model, namely ρξΦ ,Y and
ρ∗ξΦ ,Y , are very close to zero (e.g., |ρξΦ ,Y | − |ρ∗ξΦ ,Y | <
10−7 ), one basis function of them will be discarded.
B. Ensemble and evaluation of the model
The regression model is established by GLM. Note that the
number of bases participated in computation is npresv , not
all generated candidate bases. This makes EBR can realize
real-time computation. We take normalized mean square error
(NMSE) as test error in the numerical study, which is used to
TABLE I
A N EXAMPLE MAPPING TABLE FOR A
a.1
T
a.2 , a.3
expr
1
s1
1
x1
2
s2
2
x2
3
+
3
f
4
−
5
∗
6
/
TABLE II
E NCODING STEPS OF THE BASIS FUNCTION sin(x + y).
T
x1
+
sin
Step
1
2
3
Fig. 3.
s1
x1
f
f
s2
x2
x2
x1
Update
f = x1
f = x1 + x2
f = sin(x1 + x2 )
Complexity control of EBR algorithm.
evaluate the regression model. The NMSE is defined by Eq.
(10):
NMSE (f, f ∗ ) =
kf − f ∗ k22
,
kf k22
(10)
∗
where the f and f are the target model and regression model,
respectively.
C. Complexity control
In EBR, complexity control mainly includes two parts,
namely the inner control and outer control.
Recall from the parse-matrix encoding scheme in Section
II-C that, the inner control aims to determine the complexity of
a basis function and the dimension of a given symbolic regression problem, by controlling the rows of the parse-matrix (9)
and the the second and the third columns a.2 , a.3 , respectively.
More precisely, for a muti-dimensional problem, the entries aij
are bounded integers according to the mapping rules in Table
I, namely the parse-matrix entries a.1 ∈ {1, 2, 3, ..., 13} and
a.j ∈ {1, 2, 3, ..., d + 1} (j = 2, 3), where d is the dimension
of the target model. The outer control in EBR focuses on the
overall complexity of the regression model, and is controlled
by the prespecified npresv . A detailed discussion of this control
parameter is in Section IV-C. The complexity control makes
EBR algorithm have a good flexibility. The relations between
the inner control and outer control are shown in Fig. 3.
D. Procedure
Up to this point in our discussion, we have a general
understanding of EBR algorithm. The main steps of EBR is
also given in the flow-chart of EBR algorithm in Fig. 4. The
procedure of EBR could be described as follows.
Procedure of EBR:
7
√
8
s21
BASIS FUNCTION .
9
1/s1
10
log
11
exp
12
sin
13
cos
Step 1. Initialize: Input the number of basis functions
needed to be preserved npresv , the sampling range
[a, b] and the test function f .
Step 2. Generate candidate basis: An enumeration
method is used to generate a candidate basis function
φi coded with parse-matrix encoding schemes.
Step 3. Evaluate and preserve:
(3.1) Evaluate: Evaluate each generated candidate
basis by its correlation with respect to target model.
(3.2) Preserve: Preserve the basis with higher correlation with respect to target model and update the
elite bases.
Step 4. Repeat step 2 and 3 until all possible bases
are evaluated. The preserved npresv functions form
a set of elite bases for GLM.
Step 5. Model: Solving the GLM (3) to get regression
model based on the set of elite bases. Output the
decoded model and its test error (NMSE).
IV. N UMERICAL
RESULTS AND DISCUSSION
In order to test the performance of our proposed algorithm,
several numerical experiments of classical symbolic regression
problems are conducted. These problems, given in Table III, V
and VI, are mostly taken from references [14], [11]. The results
are compared with machine learning algorithm Fast Function
eXtraction (FFX) [16]. NMSE is used as the test error, which
is governed by Eq. (10).
The test problems are partitioned into two groups: exact
fitting problems (Section IV-A) and linear approximation
fitting problems (Section IV-B), to test EBR’s capabilities
of structure optimization and coefficient optimization, respectively. Additionally, we give a discussion on control parameter
(npresv ) of EBR in section IV-C. To enhance the stability of
the EBR, the distribution of sample points in should be as
uniform as possible. Therefore, controlled sampling method,
Latin hypercube sampling [1] and orthogonal sampling [21]
are preferred to generate training points (sample points).
A. Exact fitting problems
In this test group (Case 1-16, see Table III), all of cases
have exact fitting models, which recover the target models.
In other words, these cases are chosen to test the ability of
function structure optimization for EBR, which is one of the
most concerned issues in symbolic regression.
The full names of the notations in Table III are the dimension of modeled system (Dim), the target model (Target
model), the domain of the target model (Domain), the number
of training points (No. samples), total bases generated by EBR
(Total bases of EBR), the number of bases of EBR in final
form of identities. Particularly, in Case 6 and 16, namely
f = sin x21 cos x1 − 1 and f = 6 sin x1 cos x2 , note that
EBR can reduce a product term to summation of trigonometric
function.
Moreover, almost all of the number of the bases in final
results of EBR is far less than FFX, while the results are
much better than FFX. This is because that FFX does not
cover a larger operation and function space. Some cases of test
errors of FFX are extremely large (namely Case 6, 11, 13 and
14), which shows that FFX is poor at providing a symbolic
regression model in highly nonlinear function. From all of
the results above, we can draw a conclusion that the EBR
has a good capability of structure optimization in symbolic
regression problems.
Start
Set up control parameters
Generate sample points
Generation
and
preservation
of the bases
Generate basis function based on
parse-matrix encoding scheme
Calculate the correlation coefficient
of basis and the target model
Preserve and update the
elite bases
B. Linear approximation fitting problems
Are all bases
generated?
No
Yes
Solve the GLM
Ensemble
and
evaluation
of the model
Is the NMSE
satisfied ?
No
Yes
Output the decoded model
and its NMSE
Stop
Fig. 4.
The flow chart of EBR algorithm.
regression model (No. bases of EBR), the number of bases
of FFX in final regression model (No. bases of FFX), the test
error of FFX in final regression model (Test error(%) of FFX).
1) Control parameter setting: Here, we set npresv = 35 for
1D and 2D cases, npresv = 200 for 3D cases, where npresv
is the prespecified parameter in step 1 of EBR (see Section
III-D). The region is chosen as [−3, 3],[−3, 3]2 and [−3, 3]3
for one-dimensional (1D), 2D and 3D problems, respectively.
If there is a square-root function or logarithmic function in
our target model, the left endpoint of the interval is replaced
to 1. The number of training point is set up to 30, 302 and
303 for 1D, 2D and 3D problems, respectively. The order of
the parse-matrix is fixed to 3.
2) Numerical results: Table III shows the test models
and performance of EBR and corresponding results of FFX.
Numerical results (regression models) are listed in Table IV.
3) Discussion: The computation results from Table III
show that EBR can recover the target models for all these test
problems (Case 1-16). In Case 5, 6, 9 and 13-16, although
EBR does not find the bases involved in the corresponding
target models, EBR might give the regression models in
As we know, practical engineering applications of symbolic
regression are generally complex, so whether an algorithm can
give an approximation fitting model becomes very important.
The purpose of this test group is to show EBR’s capability of
providing an linear approximation regression model. This can
be regraded as the ability of coefficient optimization.
1) Control parameter setting: Similar to the first test group,
we set npresv = 35 to all cases. The number of training point
is set up to 30 and 302 for 1D and 2D problems, respectively.
The order of the parse-matrix is fixed to 3.
2) Numerical results: An overview numerical results are
listed in Table V.
3) Discussion: In this test group, 8 cases show the performance of EBR for its linear approximation fitting, which
could be regarded as a capability of coefficient optimization.
To recap briefly for Section II-A, EBR is deemed to a linear
approximation method based on GLM. We hope to find a
global and universal expansion strategy, different from Taylor
series and Fourier series.
Note that the regression models of EBR are closer to the
target model. For most cases, EBR performs better than FFX
for its succinct regression models (less number of bases),
especially for highly nonlinear target models (Case 17, 21
and 24). Meanwhile, the comparison of NMSEs in Table
IV indicates that EBR has much lower NMSE at all cases.
EBR exhibits reasonable accuracy, which indicate that the
proposed algorithm EBR can fit the target functions in forms of
polynomial functions, trigonometric, logarithmic and bivariate
functions. Good performances for modeling target functions
show the potential of EBR to be applied in practical applications.
C. Study on control parameters
The paramount control parameter in EBR is the npresv . In
this part, the value of npresv to be set is different from the
previous test groups, in order to study its influence on the
regression model. The results of this part (Case 25-28) is given
in Table VI. Note that the target models given in Table VI are
all highly nonlinear functions in 1D and 2D.
TABLE III
T EST MODELS AND PERFORMANCE OF EBR AND FFX (C ASE 1-16).
No.
Dim
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
1
1
1
1
1
1
2
2
2
2
2
2
2
2
3
3
Target model
√
x
x2 − sin x
cos x2 − x
sin x + 2x
x4 + x3 + x2 + x
sin x21 ∗ cos x1 − 1
x
x1 2
ln(x1 + x2 )
x21 + x1 − x2
x1 + 2x2
sin(x21 − x2 )
x1 − ex1 +x2
(x1 + x2 )/x2
6 sin x1 cos x2
x1 + x2 + x3
x1 x2 + x2 x3
Domain
No.
samples
[1, 3]
[1, 3]
[−3, 3]
[−3, 3]
[−3, 3]
[−3, 3]
[1, 3]2
[1, 3]2
[−3, 3]2
[−3, 3]2
[−3, 3]2
[−3, 3]2
[−3, 3]2
[−3, 3]2
[−3, 3]3
[−3, 3]3
30
30
30
30
30
30
302
302
302
302
302
302
302
302
103
103
Total
bases
of EBR
7488
7493
5510
5481
5512
5511
7499
7489
5507
5505
5509
5489
5493
5515
17163
17139
No.
bases
of EBR
1
2
1
1
4
2
1
1
4
2
1
1
2
2
3
4
No.
bases
of FFX
5
5
8
4
8
9
8
8
8
2
9
10
2
9
11
2
Test
error(%)
of FFX
0.869
0.988
6.19
0.672
2.08
16.9
0.991
0.851
0.986
0.968
28.0
1.00
7.42
25.6
0.987
0.99
TABLE IV
E XACT FITTING RESULTS OF EBR.
No.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
Regression model√
f ∗ = 0.7071 ∗ x + x
f ∗ = (−1) ∗ (sin x − x) + (x2 − x)
f ∗ = cos x2 − x
f ∗ = sin x + x + x
f ∗ = −0.5 ∗ (xex − x) + 0.5 ∗ (x4 + x) + 0.5 ∗ (x2 + xex ) + 0.5 ∗ (x2 + x)2
f ∗ = (−1) + 0.5 ∗ sin(x21 + x1 ) + 0.5 ∗ sin(x21 − x1 )
f ∗ = ex2 ∗ln x1
f ∗ = 0.5 ∗ ln(x1 + x2 )2
f ∗ = 0.5 ∗ (x21 − x2 ) + 0.5 ∗ (x21 + x1 ) + 0.5 ∗ (ex1 − 2x2 ) − 0.5 ∗ (ex1 − x2 − x1 )
f ∗ = x1 + x2 + x2
f ∗ = sin(x21 − x2 )
f ∗ = (−1) ∗ (ex1 +x2 − x1 )
f ∗ = 0.2 ∗ (2x1 − x2 )/x2 + 0.6 ∗ (2x2 + x1 )/x2
f ∗ = 3 ∗ sin(x1 + x2 ) + 3 ∗ sin(x1 − x2 )
f ∗ = 0.5 ∗ (x1 + x2 ) + 0.25 ∗ (2x2 + 2x3 ) + 0.25 ∗ (2x1 + 2x3 )
f ∗ = (−0.25) ∗ (x21 − 2x1 x2 ) + 0.25 ∗ (x21 + 2x1 x2 ) + 0.25 ∗ (x22 ) + 0.25 ∗ (x23 + 2x2 x3 )
TABLE V
T EST MODELS AND PERFORMANCE OF EBR AND FFX (C ASE 17-24).
No.
Dim
Target model
Domain
17
18
19
20
21
22
23
24
1
1
1
1
2
2
2
2
0.3x sin(2πx)
ln(x + 1) + ln(x2 + 1)
x5 + x4 + x3 + x2 + x
x6 + x5 + x4 + x3 + x2 + x
ln(x1 + x2 ) + sin(x1 + x2 )
x41 − x31 + x22 /2 − x2
x31 /5 + x32 /2 − x2 − x1
x1 x2 + sin ((x1 − 1)(x2 − 1))
[−3, 3]
[1, 3]
[−3, 3]
[−3, 3]
[1, 3]2
[−3, 3]2
[−3, 3]2
[−3, 3]2
Using the given control parameter npresv = 35, EBR failed
to get the exact fitting models or the approximate models with
NMSE ≤ 5% in this test group (expect the case 25). However,
once we increase the npresv towards a large value, for instance,
npresv = 200, EBR might provide a approximate models in
a complex form. That is, the basis number of all regression
models is larger than 20.
In this test group, the performance of EBR is also compared
with the FFX’s. Although the bases number of regression
models of EBR is more than FFX’s, its NMSEs is still much
lower than FFX’s, as shown in Table VI. The increasing of
npresv will cause the increasing computation cost of EBR.
No.
bases
of EBR
7
4
7
11
10
5
7
4
Test
error(%)
of EBR
4.37
6.58e-10
1.45e-7
4.97e-4
1.19e-2
5.87e-2
0.26
3.69
No.
bases
of FFX
10
5
6
7
8
16
12
14
Test
error(%)
of FFX
21.09
0.967
1.91
2.08
16.25
3.47
0.991
4.18
So, in practical applications, we do not set npresv to a large
value. npresv < 40 is acceptable.
V. C ONCLUSION
A new deterministic algorithm, Elite Bases Regression
(EBR), for symbolic regression has been proposed in this
paper. It is a linear approximation method based on the generalized linear model (GLM). In EBR, all generated candidate
bases are coded with parse-matrices in specific mapping rules.
The correlation coefficients with respect to the target model are
used to evaluate the candidate bases, and only a certain number
of elite bases are preserved to form the regression model. This
TABLE VI
S TUDY ON CONTROL PARAMETER .
No.
Dim
Target model
Domain
25
26
27
28
1
1
1
2
0.3x sin (2πx)
sin x3 + x
sin x sin x + x2
2
sin x1 + sin x2
[−3, 3]
[−3, 3]
[−3, 3]
[−3, 3]2
makes EBR easy to realize real-time computation.
A comparative study between EBR and a recent proposed
deterministic machine learning method for symbolic regression, Fast Function eXtraction (FFX), have been conducted.
Numerical results indicate that EBR performs better for its
more concise linear approximation regression models and
lower normalized mean square error than FFX. Moreover,
EBR can provide exact fitting models with regard to the target
models, which shows the ability of structure optimization.
As a future work, it is planned to study on improving the
performance of EBR by introducing new modifications. For
example, nonlinear correlation detection is desired, so that
EBR can be applied to complicated real-world applications
more effectively.
ACKNOWLEDGMENT
This work has been supported by the National Natural
Science Foundation of China (Grant No. 11532014).
R EFERENCES
[1] B. K. Beachkofski, R. V. Grandhi, Improved distributed hypercube
sampling, in: Proceedings of the 43rd AIAA/ASME/ASCE/AHS/ASC
Structures, Structural Dynamics, and Materials Conference, Denver, Colorado, 2002.
[2] J.W. Davidson, D.A. Savic, G.A. Walters, Symbolic and numerical
regression: experiments and applications, Information Sciences 150 (1)
(2003) 95-117.
[3] A. K. Deklel, A. M. Hamdy, E. M. Saad, Multi-objective symbolic
regression using long-term artificial neural network memory (LTANNMEM) and neural symbolization algorithm (NSA), in: Neural Computing
and Applications, Springer, 2016, pp. 1-8.
[4] J. L. Deuzé, M. Herman, R. Santer, Fourier series expansion of the
transfer equation in the atmosphere-ocean system, Journal of Quantitative
Spectroscopy & Radiative Transfer 41 (6) (1989) 483-494.
[5] X. Dong, W. Dong, Y. Yi, Y. Wang, X. Xu, The Recent Developments and
Comparative Analysis of Neural Network and Evolutionary Algorithms
for Solving Symbolic Regression, in: Intelligent Computing, Springer,
2015, pp. 703-714.
[6] C. Ferreira, Gene expression programming in problem solving, Springer,
2002, pp. 635-653.
[7] G. Gielen, E. Maricau, Stochastic degradation modeling and simulation
for analog integrated circuits in nanometer CMOS, in: Proceedings of
Design, Automation & Test in Europe Conference & Exhibition (DATE),
IEEE, 2013, pp. 326-331.
[8] B. Goldberg, Functional programming languages, ACM Computing Surveys 28 (1) (1996) 249-251.
[9] P. Holmes, P. J. Barclay, Functional languages on linear chromosomes,
in: Proceedings of the 1st annual conference on genetic programming,
MIT Press, 1996, pp. 427-427.
[10] I. Icke, J. C. Bongard, Improving genetic programming based symbolic
regression using deterministic machine learning, in: Proceedings of the
IEEE Congress on Evolutionary Computation, IEEE-CEC, 2013, pp.
1763-1770.
[11] D. Karaboga, C. Ozturk, N. Karaboga, B. Gorkemli, Artificial bee colony
programming for symbolic regression, Information Sciences 209 (2012)
1-15.
No.
bases
of EBR
21
22
28
21
Test
error(%)
of EBR
1.70e-2
2.93e-5
2.23e-20
4.68e-8
No.
bases
of FFX
8
10
9
16
Test
error(%)
of FFX
19.74
29.61
13.29
7.771
[12] J.R. Koza, Genetic programming: on the programming of computers by
means of natural selection, MIT Press, Cambridge, MA, USA, 1992.
[13] C. Luo, Z, Hu, S. Zhang, Z. Jiang, Adaptive space transformation:
An invariant based method for predicting aerodynamic coefficients of
hypersonic vehicles, Engineering Applications of Artificial Intelligence
46 (2015) 93-103.
[14] C. Luo, S. Zhang, Z. Jiang, Parse-matrix evolution for symbolic regression, Engineering Applications of Artificial Intelligence 25 (6) (2012)
1182-1193.
[15] E. Maricau, D. D. Jonghe, G. Gielen, Hierarchical analog circuit reliability analysis using multivariate nonlinear regression and active learning
sample selection, in: Proceedings of Design, Automation & Test in Europe
Conference & Exhibition (DATE), IEEE, 2012, pp. 745-750.
[16] T. McConaghy, FFX: Fast, scalable, deterministic symbolic regression
technology, in: Genetic Programming Theory and Practice IX, Springer,
2011, pp. 235-260.
[17] T. Moller, R. Machiraju, K. Mueller, R. Yagel, Evaluation and design of
filters using a Taylor series expansion, IEEE Transactions on Visualization
and Computer Graphics 3 (2) (1997) 184-199.
[18] J. A. Nelder, R. W. M. Wedderburn, Generalized linear models, Journal
of the Royal Statistical Society, 135 (1972) 370-384.
[19] M. ONeil, C. Ryan, Grammatical evolution, IEEE Transactions on
Evolutionary Computation 5 (2001) 349-358.
[20] A. Patelli, L, Ferariu, Elite based multiobjective genetic programming
in nonlinear systems identification, Advances in Electrical and Computer
Engineering 10 (1) (2010) 94-99.
[21] D. M. Steinberg, D. K. J. Lin, A construction method for orthogonal
Latin hypercube designs, Biometrika 93 (2) (2006) 279-288.
[22] A. Worm, Prioritized Grammar Enumeration: A novel method for
symbolic regression, Binghamton University - State University of New
York, 2016 Ph.D. thesis.
[23] Y. W. Yang, C. Wang, C. K. Soh, Force identification of dynamic systems
using genetic programming, International journal for numerical methods
in engineering 63 (9) (2005) 1288-1312.
Preserve and update the
elite basis function
Are all bases
generated?
Use preserved bases to
generate regression model
Is the NMSE
satisfied ?
Decode the model based
on encoding scheme
Stop
| 8cs.DS
|
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
arXiv:1308.4060v3 [math.RT] 21 Jan 2018
STEVEN DUPLIJ
A BSTRACT. Polyadic systems and their representations are reviewed and a classification of general
polyadic systems is presented. A new multiplace generalization of associativity preserving homomorphisms, a ’heteromorphism’ which connects polyadic systems having unequal arities, is introduced
via an explicit formula, together with related definitions for multiplace representations and multiactions. Concrete examples of matrix representations for some ternary groups are then reviewed. Ternary
algebras and Hopf algebras are defined, and their properties are studied. At the end some ternary generalizations of quantum groups and the Yang-Baxter equation are presented.
C ONTENTS
1. I NTRODUCTION
2. P RELIMINARIES
3. S PECIAL ELEMENTS AND PROPERTIES OF POLYADIC SYSTEMS
4. H OMOMORPHISMS OF POLYADIC SYSTEMS
5. M ULTIPLACE MAPPINGS OF POLYADIC SYSTEMS AND HETEROMORPHISMS
6. A SSOCIATIVITY QUIVERS AND HETEROMORPHISMS
7. M ULTIPLACE REPRESENTATIONS OF POLYADIC SYSTEMS
8. M ULTIACTIONS AND G- SPACES
9. R EGULAR MULTIACTIONS
10. M ULTIPLACE REPRESENTATIONS OF TERNARY GROUPS
11. M ATRIX REPRESENTATIONS OF TERNARY GROUPS
12. T ERNARY ALGEBRAS AND H OPF ALGEBRAS
13. T ERNARY QUANTUM GROUPS
14. C ONCLUSIONS
R EFERENCES
2
3
6
11
13
18
24
29
31
33
37
40
44
47
47
Date: 20 May 2012.
2010 Mathematics Subject Classification. 16T05, 16T25, 17A42, 20N15, 20F29, 20G05, 20G42, 57T05.
Published: Journal of Kharkov National University, (ser. Nuclei, Particles and Fields), Vol. 1017, 3(55) (2012) 28-59.
1
2
STEVEN DUPLIJ
1. I NTRODUCTION
One of the most promising directions in generalizing physical theories is the consideration of higher
arity algebras K ERNER [2000], in other words ternary and n-ary algebras, in which the binary composition law is substituted by a ternary or n-ary one DE A ZCARRAGA AND I ZQUIERDO [2010].
Firstly, ternary algebraic operations (with the arity n = 3) were introduced already in the XIX-th
century by A. Cayley in 1845 and later by J. J. Sylvester in 1883. The notion of an n-ary group
was introduced in 1928 by D ÖRNTE [1929] (inspired by E. Nöther) and is a natural generalization
of the notion of a group. Even before this, in 1924, a particular case, that is, the ternary group
of idempotents, was used in P R ÜFER [1924] to study infinite abelian groups. The important coset
theorem of Post explained the connection between n-ary groups and their covering binary groups
P OST [1940]. The next step in study of n-ary groups was the Gluskin-Hosszú theorem H OSSZ Ú
[1963], G LUSKIN [1965]. Another definition of n-ary groups can be given as a universal algebra with
additional laws D UDEK ET AL . [1977] or identities containing special elements RUSAKOV [1979].
The representation theory of (binary) groups W EYL [1946], F ULTON AND H ARRIS [1991] plays
an important role in their physical applications C ORNWELL [1997]. It is initially based on a matrix
realization of the group elements with the abstract group action realized as the usual matrix multiplication C URTIS AND R EINER [1962], C OLLINS [1990]. The cubic and n-ary generalizations of
matrices and determinants were made in K APRANOV ET AL . [1994], S OKOLOV [1972], and their
physical application appeared in K AWAMURA [2003], R AUSCH DE T RAUBENBERG [2008]. In general, particular questions of n-ary group representations were considered, and matrix representations
derived, by the author B OROWIEC ET AL . [2006], and some general theorems connecting representations of binary and n-ary groups were presented in D UDEK AND S HAHRYARI [2012]. The intention
here is to generalize the above constructions of n-ary group representations to more complicated and
nontrivial cases.
In physics, the most applicable structures are the nonassociative Grassmann, Clifford and Lie algebras L ÕHMUS ET AL . [1994], L OUNESTO AND A BLAMOWICZ [2004], G EORGI [1999], and so
their higher arity generalizations play the key role in further applications. Indeed, the ternary analog of Clifford algebra was considered in A BRAMOV [1995], and the ternary analog of Grassmann
algebra A BRAMOV [1996] was exploited to construct various ternary extensions of supersymmetry
A BRAMOV ET AL . [1997].
The construction of realistic physical models is based on Lie algebras, such that the fields take their
values in a concrete binary Lie algebra G EORGI [1999]. In the higher arity studies, the standard Lie
bracket is replaced by a linear n-ary bracket, and the algebraic structure of the corresponding model
is defined by the additional characteristic identity for this generalized bracket, corresponding to the
Jacobi identity DE A ZCARRAGA AND I ZQUIERDO [2010]. There are two possibilities to construct
the generalized Jacobi identity: 1) The Lie bracket is a derivation by itself; 2) A double Lie bracket
vanishes, when antisymmetrized with respect to its entries. The first case leads to the so called
Filippov algebras F ILIPPOV [1985] (or n-Lie algebra) and second case corresponds to generalized
Lie algebras M ICHOR AND V INOGRADOV [1996] (or higher order Lie algebras).
The infinite-dimensional version of n-Lie algebras are the Nambu algebras NAMBU [1973],
TAKHTAJAN [1994], and their n-bracket is given by the Jacobian determinant of n functions,
the Nambu bracket, which in fact satisfies the Filippov identity F ILIPPOV [1985]. Recently,
the ternary Filippov algebras were successfully applied to a three-dimensional superconformal
gauge theory describing the effective worldvolume theory of coincident M2-branes of M-theory
BAGGER AND L AMBERT [2008a,b], G USTAVSSON [2009]. The infinite-dimensional Nambu
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
3
bracket realization H O ET AL . [2008] gave the possibility to describe a condensate of nearly coincident M2-branes L OW [2010].
From another side, Hopf algebras A BE [1980], S WEEDLER [1969], M ONTGOMERY [1993]
play a fundamental role in quantum group theory K ASSEL [1995], S HNIDER AND S TERNBERG
[1993]. Previously, their Von Neumann generalization was introduced in D UPLIJ AND L I [2001],
D UPLIJ AND S INEL’ SHCHIKOV [2009], L I AND D UPLIJ [2002], their actions on the quantum plane
were classified in D UPLIJ AND S INEL’ SHCHIKOV [2010], and ternary Hopf algebras were defined
and studied in D UPLIJ [2001], B OROWIEC ET AL . [2001].
The goal of this paper is to give a comprehensive review of polyadic systems and their representations. First, we classify general polyadic systems and introduce n-ary semigroups and groups. Then
we consider their homomorphisms and multiplace generalizations, paying attention to their associativity. We define multiplace representations and multiactions, and give examples of matrix representations for some ternary groups. We define and investigate ternary algebras and Hopf algebras, study
their properties and give some examples. At the end we consider some ternary generalizations of
quantum groups and the Yang-Baxter equation.
2. P RELIMINARIES
Let G be a non-empty set (underlying set, universe, carrier), its elements we denote by lower-case
Latin letters gi ∈ G. The n-tuple (or polyad) g1 , . . . , gn of elements from G is denoted by (g1 , . . . , gn ).
n
z
}|
{
1
The Cartesian product G × . . . × G = G×n consists of all n-tuples (g1 , . . . , gn ), such that gi ∈ G,
i = 1, . . . , n. For all equal elements g ∈ G, we denote n-tuple (polyad) by power (g n ). If the
number of elements in the n-tuple is clear from the context or is not
important, we denote it with one
bold letter (g), in other cases we use the power in brackets g (n) . We now introduce two important
constructions on sets.
(n)
Definition 2-1. The i-projection of the Cartesian product G×n on its i-th “axis” is the map Pri
G×n → G such that (g1 , . . . gi , . . . , gn ) 7−→ gi .
:
Definition 2-2. The i-diagonal Diagn : G → G×n sends one element to the equal element n-tuple
g 7−→ (g n ).
The one-point set {•} can be treated as a unit for the Cartesian product, since there are bijections
between G and G×{•}×n , where G can be on any place. On the Cartesian product G×n one can define
a polyadic (n-ary, n-adic, if it is necessary to specify n, its arity or rank) operation µn : G×n → G. For
operations we use small Greek letters and place arguments in square brackets µn [g]. The operations
with n = 1, 2, 3 are called unary, binary and ternary. The case n = 0 is special and corresponds to
(c)
fixing a distinguished element of G, a “constant” c ∈ G, and it is called a 0-ary operation µ0 , which
(c)
(c)
maps the one-point set {•} to G, such that µ0 : {•} → G, and formally has the value µ0 [{•}] =
c ∈ G. The 0-ary operation “kills” arity, which can be seen from the following B ERGMAN [1995]:
the composition of n-ary and m-ary operations µn ◦ µm gives (n + m − 1)-ary operation by
µn+m−1 [g, h] = µn [g, µm [h]] .
(2.1)
(c)
Then, if to compose µn with the 0-ary operation µ0 , we obtain
(c)
µn−1 [g] = µn [g, c] ,
(2.2)
1We place the sign for the Cartesian product (×) into the power, because the same abbreviation will also be used below
for other types of product.
4
STEVEN DUPLIJ
because g is a polyad of length (n − 1). So, it is necessary to make a clear distinction between the
(c)
0-ary operation µ0 and its value c in G, as will be seen and will become important below.
Definition 2-3. A polyadic system G is a set G which is closed under polyadic operations.
We will write G = hset|operationsi or G = hset|operations|relationsi, where “relations” are some
additional properties of operations (e.g., associativity conditions for semigroups or cancellation properties). In such a definition it is not necessary to list the images of 0-ary operations (e.g. the unit or
zero in groups), as is done in various other definitions. Here, we mostly consider concrete polyadic
systems with one “chief” (fundamental) n-ary operation µn , which is called polyadic multiplication
(or n-ary multiplication).
Definition 2-4. A n-ary system G n = hG | µn i is a set G closed under one n-ary operation µn
(without any other additional structure).
Note that a set with one closed binary operation without any other relations was called a groupoid
by Hausmann and Ore H AUSMANN AND O RE [1937] (see, also C LIFFORD AND P RESTON [1961]).
However, nowadays the term “groupoid” is widely used in category theory and homotopy theory for
a different construction with binary multiplication, the so-called Brandt groupoid B RANDT [1927]
(see, also, B RUCK [1966]). Alternatively, and much later on, Bourbaki B OURBAKI [1998] introduced the term “magma” for binary systems. Then, the above terms were extended to the case of
one fundamental n-ary operation as well. Nevertheless, we will use some neutral notations “polyadic
system” and “n-ary system” (when arity n is fixed/known/important), which adequately indicates all
of their main properties.
Let us consider the changing arity problem:
Definition 2-5. For a given n-ary system hG | µn i to construct another polyadic system hG | µ′n′ i
over the same set G, which has multiplication with a different arity n′ .
The formulas (2.1) and (2.2) give us the simplest examples of how to change the arity of a polyadic
system. In general, there are 3 ways:
(1) Iterating. Using composition of the operation µn with itself, one can increase the arity from n
to n′iter (as in (2.1)) without changing the signature of the system. We denote the number of
ℓ
iterating multiplications by ℓµ , and use the bold Greek letters µnµ for the resulting composition
of n-ary multiplications, such that
ℓµ
µ′n′ = µℓnµ
}|
z
{
def
×(n−1)
×(n−1)
,
. . . × id
= µn ◦ µn ◦ . . . µn × id
(2.3)
where
n′ = niter = ℓµ (n − 1) + 1,
ℓ
(2.4)
which gives the length of a polyad (g) in the notation µnµ [g]. Without assuming associativity
ℓ
there many variants for placing µn ’s among id’s in the r.h.s. of (2.3). The operation µnµ is
named a long product D ÖRNTE [1929] or derived D UDEK [2007].
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
5
(2) Reducing (Collapsing). Using nc distinguished elements or constants (or nc additional 0-ary
(c )
operations µ0 i , i = 1, . . . nc ), one can decrease arity from n to n′red (as in (2.2)), such that2
n
}|c
{
z
(c )
(c )
(c ...c ) def
(2.5)
µ′n′ = µn′1 nc = µn ◦ µ0 1 × . . . × µ0 nc × id×(n−nc ) ,
where
n′ = nred = n − nc ,
(2.6)
n′ = niter→red = ℓµ (n − 1) − nc + 1.
(2.7)
nmax
= ℓµ (n − 1) − 1
c
(2.8)
(c )
and the 0-ary operations µ0 i can be on any places.
(3) Mixing. Changing (increasing or decreasing) arity may be done by combining iterating and
reducing (maybe with additional operations of different arity). If we do not use additional
operations, the final arity can be presented in a general form using (2.4) and (2.6). It will
depend on the order of iterating and reducing, and so we have two subcases:
(a) Iterating →Reducing. We have
The maximal number of constants (when n′iter→red = 2) is equal to
and can be increased by increasing the number of multiplications ℓµ .
(b) Reducing →Iterating. We obtain
n′ = nred→iter = ℓµ (n − 1 − nc ) + 1.
(2.9)
Now the maximal number of constants is
nmax
=n−2
c
(2.10)
and this is achieved only when ℓµ = 1.
To give examples of the third (mixed) case we put n = 4, ℓµ = 3, nc = 2 for both subcases of
opposite ordering:
(1) Iterating →Reducing. We can put
(c ,c )′
µ8 1 2 g (8) = µ4 [g1 , g2 , g3, µ4 [g4 , g5 , g6 , µ4 [g7 , g8 , c1 , c2 ]]] ,
(2.11)
which corresponds to the following commutative diagram
ǫ
G×8 P✲ G×8 × {•}2
id×6 ×µ4
✲
id×3 ×µ4
✲ G×4
PP
PP
PP
µ4
PP
PP
PP
(c ,c )′
µ8 1 2
PP
PP
❄
Pq
P
G×7
(2.12)
G
(2) Reducing →Iterating. We can have
(c ,c )′
µ4 1 2 g (4) = µ4 [g1 , c1 , c2 , µ4 [g2 , c1 , c2 , µ4 [g3 , c1 , c2 , g4 ]]] ,
2In
(c ...cnc )
D UDEK AND M ICHALSKI [1984] µn 1
category theory for another construction).
(2.13)
is named a retract (which term is already busy and widely used in
6
STEVEN DUPLIJ
such that the diagram
ǫ
✲ G × {•}2
G×4
❳❳
×3
id×6 ×µ4
✲
id×3 ×µ4
✲ G×4
❳❳❳
❳❳❳
❳❳
µ4
❳❳❳
❳❳❳
(c ,c )′
❳❳❳
µ4 1 2
❳❳
❳❳❳ ❄
❳❳
③
×G
G×7
(2.14)
G
is commutative.
It is important to find conditions where iterating and reducing compensate each other, i.e. they
do not change arity overall. Indeed, let the number of the iterating multiplications ℓµ be fixed, then
(0)
we can find such a number of reducing constants nc , such that the final arity will coincide with the
initial arity n. The result will depend on the order of operations. There are two cases:
(0)
(1) Iterating →Reducing. For the number of reducing constants nc we obtain from (2.4) and
(2.6)
n(0)
(2.15)
c = (n − 1) (ℓµ − 1) ,
such that there is no restriction on ℓµ .
(0)
(2) Reducing →Iterating. For nc we get
n(0)
c =
(n − 1) (ℓµ − 1)
,
ℓµ
(2.16)
(0)
and now ℓµ ≤ n − 1. The requirement that nc should be an integer gives two further
possibilities
(
n−1
, ℓµ = 2,
(2.17)
n(0)
=
2
c
n − 2, ℓµ = n − 1.
The above relations can be useful in the study of various n-ary multiplication structures and their
presentation in special form is needed in concrete problems.
3. S PECIAL
ELEMENTS AND PROPERTIES OF POLYADIC SYSTEMS
Let us recall the definitions of some standard algebraic systems and their special elements, which
will be considered in this paper, using our notation.
Definition 3-1. A zero of a polyadic system is a distinguished element z (and the corresponding 0-ary
(z)
operation µ0 ) such that for any (n − 1)-tuple (polyad) g ∈ G×(n−1) we have
µn [g, z] = z,
(3.1)
where z can be on any place in the l.h.s. of (3.1).
There is only one zero (if its place is not fixed) which can be possible in a polyadic system. As in
the binary case, an analog of positive powers of an element P OST [1940] should coincide with the
number of multiplications ℓµ in the iterating (2.3).
Definition 3-2. A (positive) polyadic power of an element is
g hℓµ i = µℓnµ g ℓµ(n−1)+1 .
(3.2)
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
7
Definition 3-3. An element of a polyadic system g is called ℓµ -nilpotent (or simply nilpotent for
ℓµ = 1), if there exist such ℓµ that
g hℓµ i = z.
(3.3)
Definition 3-4. A polyadic system with zero z is called ℓµ -nilpotent, if there exists ℓµ such that for
any (ℓµ (n − 1) + 1)-tuple (polyad) g we have
µℓnµ [g] = z.
(3.4)
Therefore, the index of nilpotency (number of elements whose product is zero) of an ℓµ -nilpotent
n-ary system is (ℓµ (n − 1) + 1), while its polyadic power is ℓµ .
Definition 3-5. A polyadic (n-ary) identity (or neutral element) of a polyadic system is a distinguished
(e)
element e (and the corresponding 0-ary operation µ0 ) such that for any element g ∈ G we have
µn g, en−1 = g,
(3.5)
where g can be on any place in the l.h.s. of (3.5).
In binary groups the identity is the only neutral element, while in polyadic systems, there exist
neutral polyads n consisting of elements of G satisfying
µn [g, n] = g,
(3.6)
where g can be also on any place. The neutral polyads are not determined uniquely. It follows from
(3.5) that the sequence of polyadic identities en−1 is a neutral polyad.
Definition 3-6. An element of a polyadic system g is called ℓµ -idempotent (or simply idempotent for
ℓµ = 1), if there exist such ℓµ that
g hℓµ i = g.
(3.7)
Both zero and the identity are ℓµ -idempotents with arbitrary ℓµ . We define (total) associativity as
the invariance of the composition of two n-ary multiplications
µ2n [g, h, u] = µn [g, µn [h] , u] = invariant
(3.8)
under placement of the internal multiplication in r.h.s. with a fixed order of elements in the whole
polyad of (2n − 1) elements t(2n−1) = (g, h, u). Informally, “internal brackets/multiplication can be
moved on any place”, which gives n relations
µn ◦ µn × id×(n−1) = . . . = µn ◦ id×(n−1) ×µn .
(3.9)
There are many other particular kinds of associativity which were introduced in T HURSTON [1949]
and studied in B ELOUSOV [1972], S OKHATSKY [1997]. Here we will confine ourselves the most
general, total associativity (3.8). In this case, the iteration does not depend on the placement of
internal multiplications in the r.h.s. of (2.3).
Definition 3-7. A polyadic semigroup (n-ary semigroup) is a n-ary system in which the operation is
associative, or G semigrp
= hG | µn | associativityi.
n
In a polyadic system with zero (3.1) one can have trivial associativity, when all n terms are (3.8)
are equal to zero, i.e.
µ2n [g] = z
(3.10)
for any (2n − 1)-tuple g. Therefore, we state that
8
STEVEN DUPLIJ
Assertion 3-8. Any 2-nilpotent n-ary system (having index of nilpotency (2n − 1)) is a polyadic
semigroup.
In the case of changing arity one should use in (3.10) not the changed final arity n′ , but the “real”
arity which is n for the reducing case and ℓµ (n − 1)+1 for all other cases. Let us give some examples.
Example 3-9. In the mixed (interacting-reducing)
D case withEn = 2, ℓµ = 3, nc = 1, we have a ternary
(c)
system hG | µ3 i iterated from a binary system G | µ2 , µ0
with one distinguished element c (or an
3
additional 0-ary operation)
(c)
µ3 [g, h, u] = (g · (h · (u · c))) ,
(3.11)
E
(c)
where for binary multiplication we denote g · h = µ2 [g, h]. Thus, if the ternary system G | µ3
(c)
D
is nilpotent of index 7 (see 3-4), then it is a ternary semigroup (because µ3 is trivially associative)
independently of the associativity of µ2 (see, e.g. B OROWIEC ET AL . [2006]).
It is very important to find the associativity preserving conditions (constructions), where an associative initial operation µn leads to an associative final operation µ′n′ during the change of arity.
Example 3-10. An associativity preserving reduction can be given by the construction of a binary
associative operation using (n − 2)-tuple c consisting of nc = n − 2 different constants
(c)
µ2 [g, h] = µn [g, c, h] .
(3.12)
Associativity preserving mixing constructions with different arities and places were considered in
D UDEK AND M ICHALSKI [1984], M ICHALSKI [1981], S OKHATSKY [1997].
Definition 3-11. An associative polyadic system with identity (3.5) is called a polyadic monoid.
The structure of any polyadic monoid is fixed P OP AND P OP [2004]: it can be obtained by iterating a binary operation Č UPONA AND T RPENOVSKI [1961] (for polyadic groups this was shown in
D ÖRNTE [1929]).
In polyadic systems, there are several analogs of binary commutativity. The most straightforward
one comes from commutation of the multiplication with permutations.
Definition 3-12. A polyadic system is σ-commutative, if µn = µn ◦ σ, or
µn [g] = µn [σ ◦ g] ,
(3.13)
where σ◦g = gσ(1) , . . . , gσ(n) is a permutated polyad and σ is a fixed element of Sn , the permutation
group on n elements. If (3.13) holds for all σ ∈ Sn , then a polyadic system is commutative.
A special type of the σ-commutativity
µn [g, t, h] = µn [h, t, g] ,
(3.14)
where t is any fixed (n − 2)-polyad, is called semicommutativity. So for a n-ary semicommutative
system we have
µn g, hn−1 = µn hn−1 , g .
(3.15)
If a n-ary semigroup Gsemigrp is iterated from a commutative binary semigroup with identity, then
G semigrp is semicommutative.
3
This construction is named the b-derived groupoid in D UDEK
AND
M ICHALSKI [1984].
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
9
Example 3-13. Let G be the set of natural numbers N, and the 5-ary multiplication is defined by
µ5 [g] = g1 − g2 + g3 − g4 + g5 ,
(3.16)
then G N5 = hN, µ5 i is a semicommutative 5-ary monoid having the identity eg = µ5 [g 5 ] = g for each
g ∈ N. Therefore, G N5 is the idempotent monoid.
Another possibility is to generalize the binary mediality in semigroups
(g11 · g12 ) · (g21 · g22 ) = (g11 · g21 ) · (g12 · g22 ) ,
(3.17)
which, obviously, follows from binary commutativity. But for n-ary systems they are different. It is
seen that the mediality should contain (n + 1) multiplications, it is a relation between n × n elements,
and therefore can be presented in a matrix from. The latter can be achieved by placing the arguments
of the external multiplication in a column.
Definition 3-14. A polyadic system is medial (or entropic), if E VANS [1963], B ELOUSOV [1972]
µn [g11 , . . . , gn1]
µn [g11 , . . . , g1n ]
..
..
.
= µn
(3.18)
µn
.
.
µn [g1n , . . . , gnn ]
µn [gn1 , . . . , gnn ]
For polyadic semigroups we use the notation (2.3) and can present the mediality as follows
µnn [G] = µnn GT ,
(3.19)
where G = kgij k is the n × n matrix of elements and GT is its transpose. The semicommutative polyadic semigroups are medial, as in the binary case, but, in general (except n = 3) not
vice versa G ŁAZEK AND G LEICHGEWICHT [1982a]. A more general concept is σ-permutability
S TOJAKOVI Ć AND D UDEK [1986], such that the mediality is its particular case with σ = (1, n).
Definition 3-15. A polyadic system is cancellative, if
µn [g, t] = µn [h, t] =⇒ g = h,
(3.20)
where g, h can be on any place. This means that the mapping µn is one-to-one in each variable. If
g, h are on the same i-th place on both sides, the polyadic system is called i-cancellative.
The left and right cancellativity are 1-cancellativity and n-cancellativity respectively. A right and
left cancellative n-ary semigroup is cancellative (with respect to the same subset).
Definition 3-16. A polyadic system is called (uniquely) i-solvable, if for all polyads t, u and element
h, one can (uniquely) resolve the equation (with respect to h) for the fundamental operation
µn [u, h, t] = g
(3.21)
where h can be on any i-th place.
Definition 3-17. A polyadic system which is uniquely i-solvable for all places i is called a n-ary (or
polyadic) quasigroup.
It follows, that, if (3.21) uniquely i-solvable for all places, than
µℓnµ [u, h, t] = g
can be (uniquely) resolved with respect to h on any place.
Definition 3-18. An associative polyadic quasigroup is called a n-ary (or polyadic) group.
(3.22)
10
STEVEN DUPLIJ
The above definition is the most general one, but it is overdetermined. Much work on polyadic
groups was done RUSAKOV [1998] to minimize the set of axioms (solvability not in all places P OST
[1940], C ELAKOSKI [1977], decreasing or increasing the number of unknowns in determining equations G AL’ MAK [2003]) or construction in terms of additionally defined objects (various analogs of
the identity and sequences Ǔ SAN [2003]), as well as using not total associativity, but instead various
partial ones S OKOLOV [1976], S OKHATSKY [1997], Y UREVYCH [2001].
In a polyadic group the only solution of (3.21) is called a querelement of g and denoted by ḡ
D ÖRNTE [1929], such that
µn [h, ḡ] = g,
(3.23)
where ḡ can be on any place. So, any idempotent g coincides with its querelement ḡ = g. It follows
from (3.23) and (3.6), that the polyad
ng = g n−2ḡ
(3.24)
is neutral for any element of a polyadic group, where ḡ can be on any place. If this i-th place is
important, then we write ng;i . The number of relations in (3.23) can be reduced from n (the number
of possible places) to only 2 (when g is on the first and last places D ÖRNTE [1929], T IMM [1972],
or on some other 2 places). In a polyadic group the Dörnte relations
µn [g, nh;i] = µn [nh;j , g] = g
(3.25)
hold true for any allowable i, j. In the case of a binary group the relations (3.25) become g · h · h−1 =
h · h−1 · g = g.
The relation (3.23) can be treated as a definition of the unary queroperation
µ̄1 [g] = ḡ,
(3.26)
such that the diagram
id×(n−1) ×µ̄1
µn
✲
✑
✸
✻
✑
✑
✑ Prn
✑
G×n
G
(3.27)
G×n
commutes. Then, using the queroperation (3.26) one can give a diagrammatic definition of a polyadic
group (cf. G LEICHGEWICHT AND G ŁAZEK [1967]).
Definition 3-19. A polyadic group is a universal algebra
G grp
n = hG | µn , µ̄1 | associativity, Dörnte relationsi ,
(3.28)
where µn is a n-ary associative operation and µ̄1 is the queroperation, such that the following diagram
G×(n)
id×(n−1) ×µ̄1
✻
id ×Diag(n−1)
✲
G×n ✛
µ̄1 ×id×(n−1)
✻
Diag(n−1) ×id
µn
Pr1
❄
✲G✛
G×n
(3.29)
Pr2
G×G
G×G
commutes, where µ̄1 can be only on first and second places from the right (resp. left) on the left (resp.
right) part of the diagram.
A straightforward generalization of the queroperation concept and corresponding definitions can
be made by substituting in the above formulas (3.23)–(3.26) the n-ary multiplication µn by iterating
ℓ
the multiplication µnµ (2.3) (cf. D UDEK [1980] for ℓµ = 2).
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
11
Definition 3-20. Let us define the querpower k of g recursively
ḡ hhkii = (ḡ hhk−1ii ),
(3.30)
k
where ḡ
(3.26).
hh0ii
= g, ḡ
hh1ii
= ḡ, or as the k composition
µ̄◦k
1
z
}|
{
= µ̄1 ◦ µ̄1 ◦ . . . ◦ µ̄1 of the queroperation
n−3
◦2
For instance G AL’ MAK [2003], µ̄◦2
1 = µn , such that for any ternary group µ̄1 = id, i.e. one has
ḡ = g. Using the queroperation in polyadic groups we can define the negative polyadic power of an
element g by the following recursive relation
µn g hℓµ −1i , g n−2, g h−ℓµ i = g,
(3.31)
or (after use of (3.2)) as a solution of the equation
µℓnµ g ℓµ(n−1) , g h−ℓµ i = g.
(3.32)
It is known that the querpower and the polyadic power are mutually connected D UDEK [1993].
Here, we reformulate this connection using the so called Heine numbers H EINE [1878] or q-deformed
numbers K AC AND C HEUNG [2002]
qk − 1
,
q−1
which have the “nondeformed” limit q → 1 as [k]q → k. Then
[[k]]q =
ḡ hhkii = g h−[[k]]2−n i ,
(3.33)
(3.34)
which can be treated as follows: the querpower coincides with the negative polyadic deformed power
with a “deformation” parameter q which is equal to the “deviation” (2 − n) from the binary group.
4. H OMOMORPHISMS
OF POLYADIC SYSTEMS
Let G n = hG; µn i and G ′n′ = hG′ ; µ′n′ i be two polyadic systems of any kind (quasigroup, semigroup, group, etc.). If they have the multiplications of the same arity n = n′ , then one can define the
mappings from G n to G ′n . Usually such polyadic systems are similar, and we call mappings between
them the equiary mappings.
′
us take n + 1 mappings ϕGG
: G → G′ , i = 1, . . . , n + 1. An ordered system of mappings
i
Let
′
GG
ϕi
is called a homotopy from G n to G ′n , if B ELOUSOV [1972]
h
i
′
′
GG′
GG′
ϕGG
(µ
[g
,
.
.
.
,
g
])
=
µ
ϕ
(g
)
,
.
.
.
,
ϕ
(g
)
, gi ∈ G.
(4.1)
n 1
n
1
n
n+1
n
1
n
In general, one should add to this definition the “mapping” of the multiplications
(µµ′ )
ψnn′
µn 7→ µ′n′ .
(4.2)
n
o
′
′
(µµ )
In such a way, homotopy can be defined as the extended system of mappings ϕGG
; ψnn
. The
i
corresponding commutative (equiary) diagram is
′
G
ϕGG
n+1
✲
✻
µn
G′
(µ) ..............
✲
............. ψnn
G×n
′
′
ϕGG
×...×ϕGG
n
1
✲
✻
µ′n
(G′ )
×n
(4.3)
12
STEVEN DUPLIJ
(µµ′ )
The existence of the additional “mapping” ψnn acting on the second component of hG; µn i is
(µµ′ )
tacitly implied. We will write/mention the “mappings” ψnn′ manifestly, e.g.,
Gn
′
(µµ′ )
ϕGG
;ψnn
i
⇒
G ′n′ ,
(4.4)
GG′
only as needed. If all the components ϕi of a homotopy are bijections, it is called an isotopy. In
′
case of polyadic quasigroups B ELOUSOV [1972] all mappings ϕGG
are usually taken as permutations
i
′
of the same underlying set G = G . If the multiplications are also coincide µn = µ′n , then ϕGG
i ; id
is called an autotopy of the polyadic system G n . Various properties of homotopy in universal algebras
were studied, e.g. in P ETRESCU [1977], H ALA Š [1994].
The homotopy, isotopy and autotopy are widely used equiary mappings in the study of polyadic
′
quasigroups and loops, while their diagonal equiary counterparts (all ϕGG
coincide), the homomori
phism, isomorphism and automorphism, are more suitable in investigation of polyadic semigroups,
groups and rings and their wide applications in physics. Usually, it is written about the latter between
similar (equiary) polyadic systems: they “...are so well known that we shall not bother to define them
carefully” H OBBY AND M C K ENZIE [1988]. Nevertheless, we give a diagrammatic definition of the
standard homomorphism between similar polyadic systems in our notation, which will be convenient
to explain the clear way of its generalization.
′
A homomorphism from Gn to G ′n is given, if there exists a mapping ϕGG : G → G′ satisfying
h
i
′
′
′
ϕGG (µn [g1 , . . . , gn ]) = µ′n ϕGG (g1 ) , . . . , ϕGG (gn ) , gi ∈ G,
(4.5)
which means that the corresponding (equiary) diagram is commutative
′
G
ϕGG
✻
µn
✲
G′
(µµ′ ) .........
✲
....... ψnn
′ ×n
ϕGG
✻
µ′n
(4.6)
✲ (G′ )
G×n
′
Usually the homomorphism is denoted by the same onenletter ϕGG , while
o it would be more consis(µµ′ )
GG′
tent to use for its notation the extended pair of mappings ϕ ; ψnn
. We will use both notations
on a par.
We first mention a small subset of known generalizations of the homomorphism (for bibliography
till 1982 see, e.g., G ŁAZEK AND G LEICHGEWICHT [1982b]) and then introduce a concrete construction for an analogous mapping which can change the arity of the multiplication (fundamental
operation) without introducing additional (term) operations. A general approach to mappings between free algebraic systems was initiated in F UJIWARA [1959], where the so-called basic mapping
formulas for generators were introduced, and its generalization to many-sorted algebras was given in
V IDAL AND T UR [2010]. In N OVOTN Ý [2002] it was shown that the construction of all homomorphisms between similar polyadic systems can be reduced to some homomorphisms between corresponding mono-unary algebras N OVOTN Ý [1990]. The notion of n-ary homomorphism is realized as
a sequence of n consequent homomorphisms ϕi , i = 1, . . . , n, of n similar polyadic systems
×n
n
}|
{
z
ϕn−1
ϕ1
′ ϕ2
′′ ϕn
′′′
Gn → Gn → . . . → Gn → Gn
(4.7)
(generalizing Post’s n-adic substitutions P OST [1940]) was introduced in G AL’ MAK [1998], and
studied in G AL’ MAK [2001a, 2007].
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
13
The above constructions do not change the arity of polyadic systems, because they are based on the
corresponding diagram which gives a definition of an equiary mapping. To change arity one has to:
1) add another equiary diagram with additional operations using the same formula (4.5), where
both do not change arity;
2) use one modified (and not equiary) diagram and the underlying formula (4.5) by themselves,
which will allow us to change arity without introducing additional operations.
The first way leads to the concept of weak homomorphism which was introduced in
G OETZ [1966], M ARCZEWSKI [1966], G ŁAZEK AND M ICHALSKI [1974] for non-indexed algebras and in G ŁAZEK [1980] for indexed algebras, then developed in T RACZYK [1965] for
Boolean and Post algebras, in D ENECKE AND W ISMATH [2009] for coalgebras and F -algebras
D ENECKE AND S AENGSURA [2008] (see also C HUNG AND S MITH [2008]). To define the weak
homomorphism in our notation we should incorporate into the polyadic systems hG; µn i and hG′ ; µ′n′ i
′
the following additional term operations of opposite arity νn′ : G×n → G and νn′ : G′×n → G′ and
consider two equiary mappings between hG; µn , νn′ i and hG′ ; µ′n′ , νn′ i.
′
A weak homomorphism from hG; µn , νn′ i to hG′ , µ′n′ , νn′ i is given, if there exists a mapping ϕGG :
G → G′ satisfying two relations simultaneously
h
i
′
′
′
ϕGG (µn [g1 , . . . , gn ]) = νn′ ϕGG (g1 ) , . . . , ϕGG (gn ) ,
(4.8)
i
h
′
′
′
(4.9)
ϕGG (νn′ [g1 , . . . , gn′ ]) = µ′n′ ϕGG (g1 ) , . . . , ϕGG (gn′ ) ,
which means that two equiary diagrams commute
′
′
G
ϕGG
✲
✻
µn
G′
(µν ′ ) .........
✲
....... ψnn
G×n
′ ×n
ϕGG
✲
(G′ )
×n
✲
✻
✻
′
νn
ϕGG
G
νn′
G′
′)
✲
......... ψ(νµ
..........
′ ′
G×n
′
n ′n×n′
ϕGG
✲
✻
(G′ )
µ′n′
(4.10)
×n′
If only one of the relations (4.8) or (4.9) holds, such a mapping is called a semi-weak ho′
momorphism KOLIBIAR [1984]. If ϕGG is bijective, then it defines a weak isomorphism.
Any weak epimorphism can be decomposed into a homomorphism and a weak isomorphism
G ŁAZEK AND M ICHALSKI [1977], and therefore the study of weak homomorphisms reduces to
weak isomorphisms (see also C Z ÁK ÁNY [1962], M AL’ TCEV [1957], M AL’ TSEV [1958]).
5. M ULTIPLACE
MAPPINGS OF POLYADIC SYSTEMS AND HETEROMORPHISMS
Let us turn to the second way of changing the arity of the multiplication and use only one relation
which we then modify in some natural manner. First, recall that in any set G there always exists
the additional distinguished mapping, viz. the identity idG . We use the multiplication µn with its
combination of idG . We define an (ℓid -intact) id-product for the polyadic system hG; µn i as
id )
µ(ℓ
= µn × (idG )×ℓid ,
n
(5.1)
id )
µ(ℓ
: G×(n+ℓid ) → G×(1+ℓid ) .
n
(5.2)
(ℓ )
To indicate the exact i-th place of µn in the r.h.s. of (5.1), we write µn id (i), as needed. Here we
use the id-product to generalize the homomorphism and consider mappings between polyadic systems
of different arity. It follows from (5.2) that, if the image of the id-product is G alone, than ℓid = 0.
14
STEVEN DUPLIJ
(n,n′ )
Let us introduce a multiplace mapping Φk
acting as follows
(n,n′ )
Φk
: G×k → G′ .
(5.3)
(n,n′ )
While constructing the corresponding diagram, we are allowed to take only one upper Φk
,
′
because of one G in the upper right corner. Since we already know that the lower right corner
′
is exactly G′×n (as a pre-image of one multiplication µ′n′ ), the lower horizontal arrow should be a
(n,n′ )
product of n′ multiplace mappings Φk
. So we can write a definition of a multiplace analog of
homomorphisms which changes the arity of the multiplication using one relation.
Definition 5-1. A k-place heteromorphism from G n to G′n′ is given, if there exists a k-place mapping
(n,n′ )
Φk
(5.3) such that the following (arity changing or unequiary) diagram is commutative
Φk
G×k
(ℓ )
µn id
✲
G′
✻
✻
µ′n′
(Φk )
×n′
(5.4)
✲ (G′ )
G×kn
and the corresponding defining equation (a modification of (4.5)) depends on the place i of µn in
(5.1).
′
×n′
For i = 1 a heteromorphism is defined by the formula
µn [g1 , . . . , gn ]
g1
gk(n′ −1)
gn+1
(n,n′ )
..
= µ′ ′ Φ(n,n′ ) ... , . . . , Φ(n,n′ )
.
Φk
.
.
n
k
k
..
gk
gkn′
gn+ℓid
(5.5)
The notion “heteromorphism” is motivated by E LLERMAN [2006, 2007], where mappings between
objects from different categories were considered and called “chimera morphisms”. See, also, P ÉCSI
[2011].
In the particular case n = 3, n′ = 2, k = 2, ℓid = 1 we have
µ3 [g1 , g2 , g3]
g1
g3
(3,2)
(3,2)
(3,2)
′
Φ2
= µ2 Φ2
, Φ2
.
(5.6)
g4
g2
g4
This formula was used in the construction of the bi-element representations of ternary groups
B OROWIEC ET AL . [2006]. Consider the example.
Example 5-2. Let G = M2adiag (K), a set of antidiagonal 2 × 2 matrices over the field K and G′ = K,
where K = R, C, Q, H. The ternary multiplication µ3 is a product
Obviously, µ3 is
of 3 matrices.
0 ai
not derived from a binary multiplication. For the elements gi =
, i = 1, 2, we construct a
bi 0
2-place mapping G × G → G′ as
g1
(3,2)
Φ2
= a1 a2 b1 b2 ,
(5.7)
g2
which is a heteromorphism, because it satisfies (5.6). Let us introduce a standard 1-place mapping by
ϕ (gi ) = ai bi , i = 1, 2. It is important to note, that ϕ (gi ) is not a homomorphism, because the product
g1 g2 belongs to diagonal matrices. Consider the product of mappings
ϕ (g1 ) · ϕ (g2 ) = a1 b1 a2 b2 ,
(5.8)
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
15
where the product (·) in l.h.s. is taken in K. We observe that (5.7) and (5.8) coincide for the commutative field K only (= R, C) only, and in this case we can have the relation between the heteromorhism
(3,2)
Φ2 and the 1-place mapping ϕ
g1
(3,2)
Φ2
= ϕ (g1 ) · ϕ (g2 ) ,
(5.9)
g2
while for the noncommutative field K (= Q or H) there is no such relation.
A heteromorphism is called derived, if it can be expressed through a 1-place mapping (not necessary a homomorphism). So, in the above Example 5-2 the heteromorphism is derived (by formula
(5.9)) for the commutative field K and nonderived for the noncommutative K.
For arbitrary n a slightly modified construction (5.6) with still binary final arity, defined by n′ = 2,
k = n − 1, ℓid = n − 2,
µn [g1 , . . . , gn−1, h1 ]
g1
h1
h2
(n,2)
= µ′2 Φ(n,2)
... , Φ(n,2)
... .
Φn−1
(5.10)
.
n−1
n−1
..
gn−1
hn−1
hn−1
was used in D UDEK [2007] to study representations of n-ary groups. However, no new results compared with B OROWIEC ET AL . [2006] (other than changing 3 to n in some formulas) were obtained.
This reflects the fact that a major role is played by the final arity n′ and the number of n-ary multiplications in the l.h.s. of (5.6) and (5.10). In the above cases, the latter number was one, but can make it
arbitrary below n.
Definition 5-3. A heteromorphism is called a ℓµ -ple heteromorphism, if it contains ℓµ multiplications
(n,n′ )
in the argument of Φk
in its defining relation.
According this definition the mapping defined by (5.5) is the 1µ -ple heteromorphism. So by analogy
with (5.1)–(5.2) we define a ℓµ -ple ℓid -intact id-product for the polyadic system hG; µn i as
µ ,ℓid )
µ(ℓ
= (µn )×ℓµ × (idG )×ℓid ,
n
µ ,ℓid )
µ(ℓ
: G×(nℓµ +ℓid ) → G×(ℓµ +ℓid ) .
n
(5.11)
(5.12)
G ′n′
Definition 5-4. A ℓµ -ple k-place heteromorphism from Gn to
is given, if there exists a k-place
(n,n′ )
mapping Φk
(5.3) such that the following unequiary diagram is commutative
Φk
G×k
(ℓµ ,ℓid )
µn
✻
✲
G′
✻
µ′n′
×n′
(Φk )
✲ (G′ )
G×kn
The corresponding main heteromorphism equation is
µn [g1 , . . . , gn ] ,
..
ℓ
.
µ
(n,n′ ) µn gn(ℓµ −1) , . . . , gnℓµ
= µ′ ′ Φ(n,n′)
Φk
n
k
gnℓµ +1 ,
..
ℓid
.
gnℓµ +ℓid
′
(5.13)
×n′
g1
gk(n′ −1)
.. , . . . , Φ(n,n′ )
..
.
.
.
k
gk
gkn′
(5.14)
16
STEVEN DUPLIJ
F IGURE 1. Dependence of the final arity n′ through the number of heteromorphism places
k for the initial arity n = 9 with the fixed number of intact elements ℓid (left) and the fixed
number of multiplications ℓµ (right): =1 (solid curves), =2 (dash curves)
Obviously, we can consider various permutations of the multiplications on both sides, as further
additional demands (associativity, commutativity, etc.), are introduced, which will be considered below. The commutativity of the diagram (5.13) leads to the system of equation connecting initial and
final arities
kn′ = nℓµ + ℓid,
k = ℓµ + ℓid .
(5.15)
(5.16)
Excluding ℓµ or ℓid , we obtain two arity changing formulas, respectively
n−1
ℓid ,
k
n−1
n′ =
ℓµ + 1,
k
n′ = n −
(5.17)
(5.18)
ℓid ≥ 1 and n−1
ℓµ ≥ 1 are integer.
where n−1
k
k
As an example, the dependences n′ (k) for the fixed ℓµ = 1, 2 and ℓid = 1, 2 with n = 9 are
presented on Figure 1.
The following inequalities hold valid
1 ≤ ℓµ ≤ k,
0 ≤ ℓid ≤ k − 1,
ℓµ ≤ k ≤ (n − 1) ℓµ ,
′
2 ≤ n ≤ n,
(5.19)
(5.20)
(5.21)
(5.22)
which are important for the further classification of heteromorphisms. The main statement follows
from (5.22)
(n,n′ )
Proposition 5-5. The heteromorphism Φk
the arity of polyadic multiplication.
defined by the general relation (5.14) always decreases
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
17
Another important observation is the fact that only the id-product (5.11) with ℓid 6= 0 leads to
a change of the arity. In the extreme case, when k approaches its minimum, k = kmin = ℓµ , the
final arity approaches its maximum n′max = n, and the id-product becomes a product of ℓµ initial
multiplications µn without id’s, since now ℓid = 0 in (5.14). Therefore, we call a heteromorphism
defined by (5.14) with ℓid = 0 a k (= ℓµ )-place homomorphism. The ordinary homomorphism (4.1)
corresponds to k = ℓµ = 1, and so it is really a 1-place homomorphism. An opposite extreme case,
when the final arity approaches its minimum n′min = 2 (the final operation is binary), corresponds to
the maximal value of k, that is k = kmax = (n − 1) ℓµ . The number of id’s now is ℓid = (n − 2) ℓµ ≥
0, which vanishes, when the initial operation is binary as well. This is the case of the ordinary
homomorphism (4.1) for both binary operations n′ = n = 2 and k = ℓµ = 1. We conclude that:
Any polyadic system can be mapped into a binary system by means of the special k-place ℓµ -ple
(n,n′ )
heteromorphism Φk
, where k = (n − 1) ℓµ (we call it a binarizing heteromorphism) which is
defined by (5.14) with ℓid = (n − 2) ℓµ .
In relation to the Gluskin-Hosszú theorem G LUSKIN [1965] (any n-ary group can be constructed
from the special binary group and its homomorphism) our statement can be treated as:
Theorem 5-6. Any n-ary system can be mapped into a binary system, using a suitable binarizing
(n,2)
heteromorphism Φk (5.14).
The case of 1-ple binarizing heteromorphism (ℓµ = 1) corresponds to the formula (5.10). Further
requirements (associativity, commutativity, etc.) will give additional relations between multiplications
(n,n′ )
and Φk
, and fix the exact structure of (5.14). Thus, we arrive to the following
Proposition 5-7. Classification of ℓµ -ple heteromorphisms:
(n,n)
(1) n′ = n′max = n =⇒ Φk
is the ℓµ -place or multiplace homomorphism, i.e.,
k = kmin = ℓµ .
(n,n′ )
(2) 2 < n′ < n =⇒ Φk
(5.23)
is the intermediate heteromorphism with
n−1
ℓµ .
(5.24)
n′ − 1
In this case the number of intact elements is proportional to the number of multiplications
k=
ℓid =
(n,2)
(3) n′ = n′min = 2 =⇒ Φk
i.e.,
n − n′
ℓµ .
n′ − 1
(5.25)
is the (n − 1) ℓµ -place (multiplace) binarizing heteromorphism,
k = kmax = (n − 1) ℓµ .
(5.26)
In the extreme (first and third) cases there are no restrictions on the initial arity n, while in the
intermediate case n is “quantized” due to the fact that fractions in (5.17) and (5.18) should be integers.
Observe, that in the extreme (first and third) cases there are no restrictions on the initial arity n,
while in the intermediate case n is “quantized” due to the fact that fractions in (5.17) and (5.18) should
be integer. In this way, we obtain the TABLE 1 for the series of n and n′ (we list only first ones, just
for 2 ≤ k ≤ 4 and include the binarizing case n′ = 2 for completeness).
Thus, we have established a general structure and classification of heteromorphisms defined by
(5.14). The next important issue is the preservation of special properties (associativity, commutativity,
etc.), while passing from µn to µ′n′ , which will further restrict the concrete shape of the main relation
18
STEVEN DUPLIJ
TABLE 1. “Quantization” of heteromorphisms
n/n′
3, 5,
2, 3,
4, 7,
2, 3,
4, 7,
3, 5,
5, 9,
2, 3,
3, 5,
2, 3,
5, 9,
4, 7,
k ℓµ ℓid
2
1
1
3
1
2
3
2
1
4
1
3
4
2
2
4
3
1
n=
n′ =
n=
n′ =
n=
n′ =
n=
n′ =
n=
n′ =
n=
n′ =
7,
4,
10,
4,
10,
7,
13,
4,
7,
4,
13,
10,
...
...
...
...
...
...
...
...
...
...
...
...
(5.14) for each choice of the heteromorphism parameters: arities n, n′ , places k, number of intacts ℓid
and multiplications ℓµ .
6. A SSOCIATIVITY
QUIVERS AND HETEROMORPHISMS
The most important property of the heteromorphism, which is needed for its next applications to
representation theory, is the associativity of the final operation µ′n′ , when the initial operation µn is
associative. In other words, we consider here the concrete form of semigroup heteromorphisms. In
general, this is a complicated task, because it is not clear from (5.14), which permutation in the l.h.s.
should be taken to get an associative product in its r.h.s. for each set of the heteromorphism parameters. Straightforward checking of the associativity of the final operation µ′n′ for each permutation
in the l.h.s. of (5.14) is almost impossible, especially for higher n. To solve this difficulty we introduce the concept of the associative polyadic quiver and special rules to construct the associative final
operation µ′n′ .
Definition 6-1. A polyadic quiver of products is the set of elements from G n (presented as several
copies of some matrix of the elements glued together) and arrows, such that the elements along arrows
form n-ary products µn .
For instance, for the 4-ary multiplication µ4 [g1 , h2 , g2 , u1] (elements from G n are arbitrary here) a
corresponding 4-adic quiver will be denoted by {g1 → h2 → g2 → u1 }, and graphically this 4-adic
quiver is
g1
g2
❅❅
❅
h1
u1
h2
u2
g1
*
g2
KS
corr
h1
7 u1
♥♥♥
♥
♥
♥
♥♥♥
h2
u2
(6.1)
µ4 [g1 , h2 , g2 , u1 ] .
Next we define polyadic quivers which are related to the main heteromorphism equation (5.14) in
the following way:
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
19
1) the matrix of elements is the transposed matrix from the r.h.s. of (5.14), such that different letters
(n,n′ )
correspond to their place in Φk
and the low index of an element is related to its position in the µ′n′
product;
2) the number of polyadic quivers is ℓµ , which corresponds to ℓµ multiplications in the l.h.s. of
(5.14);
3) the heteromorphism parameters (n, n′ , k, ℓid and ℓµ ) are not arbitrary, but satisfy the arity
changing formulas (5.17)-(5.18);
4) the intact elements will be placed after a semicolon.
In this way, a polyadic quiver makes a clear visualization of the main heteromorphism equation
(5.14), and later on it will allow us to distinguish associativity preserving heteromorphisms by precise
graphical rules.
For example, the polyadic quiver {g1 → h2 → g2 → u1 ; h1 , u2 } corresponds to the unequiary heteromorphism with n = 4, n′ = 2, k = 3, ℓid = 2 and ℓµ = 1 is
g1
❅❅
❅
g2
(4,2)
Φ3
h1
u1
h2
u2
g1
*
g2
h1 ♥7 u1
♥
♥♥♥
♥
♥
♥♥
h2
u2
KS
corr
(6.2)
g2
g1
µ4 [g1 , h2 , g2 , u1 ]
(4,2)
(4,2)
′
h2 ,
h1 , Φ3
= µ2 Φ3
h1
u2
u1
u2
where the intact elements h1 , u2 are boxed in squares. As it is seen from (6.2), the product µ′2 is not
associative, if µ4 is associative. So, not all polyadic quivers preserve associativity.
Definition 6-2. An associative polyadic quiver is a polyadic quiver which ensures the final associativity of µ′n′ in the main heteromorphism equation (5.14), when the initial multiplication µn is
associative.
One of the associative polyadic quivers which corresponds to the same heteromorphism parameters,
as the non-associative quiver (6.2), is {g1 → h2 → u1 → g2 ; h1 , u2 } which corresponds to
g1
g2
❅❅
❅
h1
h2
u1
g1
⑥> ❅❅❅
⑥
⑥
❅
⑥
u2
g2
KS
corr
h1
u1
h2
u2
(6.3)
g1
g2
µ4 [g1 , h2 , u1 , g2 ]
(4,2)
h1 , Φ(4,2)
h2 .
= µ′2 Φ(4,2)
h1
Φ3
3
3
u1
u2
u2
Here we propose a classification of associative polyadic quivers and the rules of construction of
the corresponding heteromorphism equations, and then use the heteromorphism parameters for the
classification of ℓµ -ple heteromorphisms (5.24). In other words, we describe a consistent procedure
for building the semigroup heteromorphisms.
Let us consider the first class of heteromorphisms (without intact elements ℓid = 0 or intactless),
that is ℓµ -place (multiplace) homomorphisms. In the simplest case, associativity can be achieved,
when all elements in a product are taken from the same row. The number of places k is not fixed by
20
STEVEN DUPLIJ
the arity relation (5.17) and can be arbitrary, while the arrows can have various directions. There are
2k such combinations which preserve associativity. If the arrows have the same direction, this kind of
mapping is also called a homomorphism. As an example, for n = n′ = 3, k = 2, ℓµ = 2 we have
g1
h1
g2
h2
g3
h3
(6.4)
KS
(3,3)
Φ2
µ3 [g1 , g2 , g3 ]
µ3 [h1 , h2 , h3 ]
corr
(3,3)
= µ′3 Φ2
g1
h1
(3,3)
, Φ2
g2
h2
(3,3)
, Φ2
g3
h3
.
Note that the analogous quiver with opposite arrow directions is
hO 1
g1
g2
hO 2
g3
h3
(6.5)
KS
(3,3)
Φ2
µ3 [g1 , g2 , g3 ]
µ3 [h3 , h2 , h1 ]
corr
(3,3)
= µ′3 Φ2
g1
h1
(3,3)
, Φ2
g2
h2
(3,3)
, Φ2
g3
h3
,
The latter mapping and the corresponding vertical quiver were used in constructing the middle representations of ternary groups B OROWIEC ET AL . [2006].
For nonvertical quivers the main rule is the following: all arrows of an associative quiver should
have direction from left to right or vertical, and they should not instersect. Also, we start always from
the upper left corner, because of the permutation symmetry of (5.14), we can rearrange and rename
variables in the necessary way.
An important class of intactless heteromorphisms (with ℓid = 0) preserving associativity can be
constructed using an analogy with the Post substitutions P OST [1940], and therefore we call it the
Post-like associative quiver. The number of places k is now fixed by k = n − 1, while n′ = n and
ℓµ = k = n − 1. An example of the Post-like associative quiver with the same heteromorphisms
parameters as in (6.4)-(6.5) is
g1
❅❅
❅
g2
h1
❅❅ g1
❅❅
h1
h2 ❅ g2
h
❅❅ 2
❅
❅❅
g3
h3
❅
g3
h3
(6.6)
(3,3)
Φ2
µ3 [g1 , h2 , g3 ]
µ3 [h1 , g2 , h3 ]
=
µ′3
KS
corr
(3,3)
Φ2
g1
h1
(3,3)
, Φ2
g2
h2
(3,3)
, Φ2
g3
h3
.
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
21
This construction appeared in the study of ternary semigroups of morphisms C HRONOWSKI
[1994b,a], C HRONOWSKI AND N OVOTN Ý [1995]. Its n-ary generalization was used in the consideration of polyadic operations on Cartesian powers G AL’ MAK [2008], polyadic analogs of the
Cayley and Birkhoff theorems G AL’ MAK [2001b, 2007] and special representations of n-groups
G LEICHGEWICHT ET AL . [1983], WANKE -J AKUBOWSKA AND WANKE -J ERIE [1984] (where the
n-group with the multiplication µ′2 was called the diagonal n-group). Consider the following example.
Example 6-3. Let Λ be the Grassmann algebra consisting of even and odd parts Λ = Λ0̄ ⊕ Λ1̄ (see
(1̄)
e.g., B EREZIN [1987]). The odd part can be considered as a ternary semigroup G 3 = hΛ1̄ , µ3 i,
its multiplication µ3 : Λ1̄ × Λ1̄ × Λ1̄ → Λ1̄ is defined by µ3 [α, β, γ] = α · β · γ, where (·) is
(1̄)
multiplication in Λ and α, β, γ ∈ Λ1̄ , so G 3 is nonderived and contains no unity. The even part can
(0̄)
be treated as a ternary group G 3 = hΛ0̄ , µ′3 i with the multiplication µ′3 : Λ0̄ × Λ0̄ × Λ0̄ → Λ0̄ , defined
(0̄)
by µ3 [a, b, c] = a · b · c, where a, b, c ∈ Λ0̄ , thus G 3 is derived and contains unity. We introduce the
(3,3)
(0̄)
(1̄)
heteromorphism G 3 → G 3 as a mapping (2-place homomorphism) Φ2 : Λ1̄ × Λ1̄ → Λ0̄ by the
formula
(3,3)
Φ2
α
β
= α · β,
(6.7)
(3,3)
where α, β ∈ Λ1̄ . It is seen that Φ2 defined by (6.7) satisfies the Post-like heteromorphism equation
(6.6), but not the “vertical” one (6.4), due to the anticommutativity of odd elements from Λ1̄ . In other
(0̄)
(1̄)
words, G 3 can be treated as a nontrivial example of the “diagonal” semigroup of G 3 (according
to the notation of G LEICHGEWICHT ET AL . [1983], WANKE -J AKUBOWSKA AND WANKE -J ERIE
[1984]).
Note that for the number of places k ≥ 3 there exist additional (to the above) associative quivers
having the same heteromorphism parameters. For instance, when n′ = n = 4 and k = 3 we have the
Post-like associative quiver
g1
g2
❅❅
❅
h1 ❆ u1 ❅ g1
❆
❅
h1
u1
h2 ❆ u2 ❅ g2
h
❅❅ 2
❆
❅
u2
❆❆
❆❆
❅❅
❅❅
g3
h3
u3 ❅ g3
h
u
❅❅ 3 ❆❆ 3
❅
g4
h4
u4
❅❅
(4,4)
Φ3
❅
g4
KS
corr
❆❆
❅
h4
u4
g1
g2
µ4 [g1 , h2 , u3 , g4 ]
g1
g2
(4,4)
(4,4)
(4,4)
(4,4)
′
h1 , Φ3
h2 , Φ3
µ4 [h1 , u2 , g3 , h4 ] = µ4 Φ3
h1 , Φ3
h2 .
µ4 [u1 , g2 , h3 , u4 ]
u1
u2
u1
u2
(6.8)
22
STEVEN DUPLIJ
Also, we have one intermediate non-Post associative quiver
g1 PP h1 PP u1 PP g1
h1
u1
g1
h1
PPP PPP PPP
PPP PPP
P
P
P'
P'
P'
g2
h2
u2 PP g2 PP h2 PP u2
g2
h2
PPP PPP PPP
PPP PPPP PPPP
'
'
'
g3
h3
u3
g3
h3 PP u3 PP g3 PP h3
PPP PPP PPP
P
PPP
P' PPP' PPP'
g4
h4
u4
g4
h4
u4
g4
h4
(4,4)
Φ3
u1
u2
u3
u4
(6.9)
KS
corr
g2
g1
g2
g1
µ4 [g1 , u2 , h3 , g4 ]
(4,4)
(4,4)
(4,4)
(4,4)
′
h2 .
h1 , Φ3
h2 , Φ3
h1 , Φ3
µ4 [h1 , g2 , u3 , h4 ] = µ4 Φ3
u2
u1
u2
u1
µ4 [u1 , h2 , g3 , u4 ]
A general method of constructing associative quivers for n′ = n, ℓid = 0 and k = n − 1 can be
illustrated from the following more complicated example with n = 5. First, we draw k = n − 1
(= 4) copies of element matrices. Then we go from the first element in the first column g1 to the last
element in this column gn′ (= g5 ) by different k = n − 1 (= 4) ways: by the vertical quiver, by the
Post-like quiver (going to the second copy of the element matrix) and by the remaining k − 2 (= 2)
non-Post associative quivers, as
❱
g1 ❘
u1
❍❍❘❱❘❱❘❱❱h❱1❱
❍❍ ❘❘❘ ❱❱❱
g2
g3
g4
g5
vertical
❍$
v1
g1
h1
u1
v1
g1
h1
u1
v1
g1
❘❘❘ ❱❱❱❱
❱*
)
h2 ❆ u2 ◗◗ v2 ❱❱❱❱g2
h2
u2
v2
g2
h2
u2
v2
g2
◗◗◗
❱
❆❆
◗◗◗ ❱❱❱❱❱❱❱❱
❆
◗(
❱❱*
h3
u3 ❅ v3
g3 ◗◗◗ h3
u3 ❲❲❲v❲3❲
h3
u3
v3
g3
❲❲❲❲❲ g3
◗◗◗
❅❅
❲
❲
◗
❲
❲❲❲❲❲
❅
◗◗(
+
h4
u4
v4 ❉ g4
h4
u4 ❘❘❘v4
g4
g4
h4 ❱❱❱u
❱❱4 ❱❱ v4
❘❘❘
❉❉
❱❱❱❱
❘❘❘
❉❉
❱❱❱❱
❘❘)
!
❱+
h5
u5
v5
g5
Post
h5
u5
v5
g5
non-Post
h5
u5
v5
g5
h1
u1
v1
h2
u2
v2
h3
u3
v3
h4
u4
v4
h5
u5
v5
non-Post
(6.10)
Here we show, for short, only the quiver itself without the corresponding heteromorphism equation
and only the arrows corresponding to the first product, while other arrows (starting from h1 , u1 , v1 )
are parallel to it, as in (6.9).
The next type of heteromorphisms (intermediate) is described by the equations (5.15)-(5.25), it
contains intact elements (ℓid ≥ 1) and changes (decreases) arity n′ < n. For each fixed k the arities
are not arbitrary and presented in TABLE 1. The first general rule is: the associative quivers are nondecreasing in both,vertical (from up to down) and horizontal (from left to right), directions. Second,
if there are several multiplications (ℓµ ≥ 2), the corresponding associative quivers do not intersect.
Let us present some examples and start from the smallest number of heteromorphism places in
(n,n′ )
Φk
. For k = 2, the first (nonbinarizing n′ ≥ 3) case is n = 5, n′ = 3, ℓid = 1 (see the first row
.
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
23
and second n/n′ pair of TABLE 1). The corresponding associative quivers are
g1
/ h1
❅❅ g1
❅❅
g2
h2
g2
g3
h3
g3
(5,3)
Φ2
h1
g1
h1
g1
/ h2
❅❅ g2
❅❅
h2
h3
h3
g3
h1
g1
g2
h2
/ g2
g3
h3
g3
❅❅
❅
❅❅
❅
h1
g1
h1
h2
g2
h2
h3
/ g3
h3
KS
KS
corr
corr
µ5 [g1 , h1, g2 , h2, g3 ]
h3
(5,3)
Φ2
µ5 [g1 , h2, g2 , h3, g3 ]
h1
(6.11)
,
where we do not write the r.h.s. of the heteromorphism equation (5.14), because it is simply related
to the transposed quiver matrix of the size n′ × k.
More complicated examples can be given for k = 3, which corresponds to the second and third lines
of TABLE 1. That is we can obtain a ternary final product (n′ = 3) by using one or two multiplications
ℓµ = 1, 2. Examples of the corresponding associative quivers are (ℓµ = 1)
g1
/ h1
❆❆ u1
❆❆
g1
g2
h2
u2
/ g2
g3
h3
u3
g3
❅❅
❅
h1
u1
g1
h1
u1
h2
u2
g2
h2
u2
h3
/ u3
/ g3
h3
u3
(6.12)
KS
(7,3)
Φ3
corr
µ7 [g1 , h1, u2 , g2 , h3 , u3 , g3 ]
h2
u1
and (ℓµ = 2)
g1 PP h1
u1
PPP
PPP
P'
g2
h2
u2
g1
g3
g3
h3
u3
g2
h1
2 u1 ❅ g1
❅❅
❅
h1
u1
g2
h
4 h2 PPP u2
❅❅ 2
PPP
❅
PPP
'
u2
h3
u3
g3
KS
(4,3)
Φ3
h3
u3
(6.13)
corr
µ4 [g1 , u2 , h2 , g3 ]
µ4 [h1, u1 , g2 , h3 ]
u3
respectively. Finally, for the case k = 4 one can construct the associative quiver corresponding to
the first pair of the last line in TABLE 1. It has three multiplications and one intact element, and the
24
STEVEN DUPLIJ
corresponding quiver is, e.g.,
g1
g2
❅❅
❅
u1
v1
h2 ❆ u2
❆
- v2
h1
❆❆
g3
h3
u3
v3
g4
h4
u4
v4
g1
h1
u1
v1
3 u1 PP v1
PPP
PPP
'
g2
h2
g2
h
u2
v2
3 u2 ❅ v2
❅❅ 2
❅❅
❅
❅
v3
v3 PP g3
h3 PP u3
4 g3 ❲❲❲h❲❲3 ❲ u3
PPP
PPP
❲❲❲❲❲
PPP
❲❲❲❲❲
PPP
P'
❲❲❲+
'
g1
h1
g4
h4
u4
v4
g4
h4
u4
v4
(6.14)
KS
corr
µ5 [g1 , h2 , u3 , g3 , g4 ]
(5,4) µ5 [h1 , v2 , u2 , v3 , h4 ]
Φ4
µ5 [v1 , u1 , g2 , h3 , v4 ] .
u4
There are many other possibilities (using permutations and different variants of quivers) to obtain
an associative final product µ′n′ corresponding the same heteromorphism parameters, and therefore
we do list them all. The above examples are sufficient to understand the rules of the associative quiver
construction and obtain the polyadic semigroup heteromorphisms.
7. M ULTIPLACE
REPRESENTATIONS OF POLYADIC SYSTEMS
Representation theory K IRILLOV [1976] deals with mappings from abstract algebraic systems
into linear systems, such as, e.g. linear operators in vector spaces, or into general (semi)groups of
transformations of some set. In our notation, this means that in the mapping of polyadic systems
(4.4) the final multiplication µ′n′ is a linear map. This leads to some restrictions on the final polyadic
structure G ′n′ , which are considered below.
Let V be a vector space over a field K (usually algebraically closed) and End V be a set of linear
endomorphisms of V , which is in fact a binary group. In the standard way, a linear representation
of a binary semigroup G 2 = hG; µ2 i is a (1-place) map Π1 : G2 → End V , such that Π1 is a
homomorphism
Π1 (µ2 [g, h]) = Π1 (g) ∗ Π1 (h) ,
(7.1)
where g, h ∈ G and (∗) is the binary multiplication in End V (usually, it is a (semi)group with
multiplication as composition of operators or product of matrices, if a basis is chosen). If G 2 is a
binary group with the unity e, then we have the additional condition
Π1 (e) = idV .
(7.2)
We will generalize these known formulas to the corresponding polyadic systems along with the
heteromorphism concept introduced above. Our general idea is to use the heteromorphism equation
(5.14) instead of the standard homomorphism equation (7.1), such that the arity of the representation
will be different from the arity of the initial polyadic system n′ 6= n.
Consider the structure of the final n′ -ary multiplication µ′n′ in (5.14), taking into account that the
final polyadic system G ′n′ should be constructed from End V . The most natural and physically applicable way is to consider the binary End V and to put G ′n′ = dern′ (End V ), as it was proposed for the
ternary case in B OROWIEC ET AL . [2006]. In this way G′n′ becomes a derived n′ -ary (semi)group of
′
endomorphisms of V with the multiplication µ′n′ : (End V )×n → End V , where
µ′n′ [v1 , . . . , vn′ ] = v1 ∗ . . . ∗ vn′ , vi ∈ End V.
(7.3)
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
25
Because the multiplication µ′n′ (7.3) is derived and is therefore associative by definition, we may
consider the associative initial polyadic systems (semigroups and groups) and the associativity preserving mappings that are the special heteromorphisms constructed in the previous section.
Let G n = hG; µn i be an associative n-ary polyadic system. By analogy with (5.3), we introduce
the following k-place mapping
(n,n′ )
: G×k → End V.
Πk
(7.4)
A multiplace representation of an associative polyadic system G n in a vector space V is given,
if there exists a k-place mapping (7.4) which satisfies the (associativity preserving) heteromorphism
equation (5.14), that is
µn [g1 , . . . , gn ] ,
n′
z
..
}|
ℓ
{
µ
.
gk(n′ −1)
g1
(n,n′ ) µn gn(ℓµ −1) , . . . , gnℓµ
..
= Π(n,n′) ... ∗ . . . ∗ Π(n,n′)
, (7.5)
Πk
.
k
k
gnℓµ +1 ,
gkn′
gk
..
ℓid
.
gnℓµ +ℓid
and the following diagram commutes
Πk
G×k
(ℓµ ,ℓid )
µn
✲
✻
End V
✻
(7.6)
′
(∗)n
G×kn
′
×n′
(Πk )
✲
(End V )×n
′
(ℓ ,ℓ )
where µn µ id is given by (5.11), ℓµ and ℓid are the numbers of multiplications and intact elements in
the l.h.s. of (7.5), respectively.
The exact permutation in the l.h.s. of (7.5) is given by the associative quiver presented in the
previous section. The representation parameters (n, n′ , k, ℓµ and ℓid ) in (7.5) are the same as the
heteromorphism parameters, and they satisfy the same arity changing formulas (5.17) and (5.18).
Therefore, a general classification of multiplace representations can be done by analogy with that of
the heteromorphisms (5.23)–(5.26) as follows:
(1) The hom-like multiplace representation which is a multiplace homomorphism with n′ =
(min)
n′max = n, without intact elements lid = lid = 0, and minimal number of places
k = kmin = ℓµ .
(7.7)
(2) The intact element multiplace representation which is the intermediate heteromorphism with
2 < n′ < n and the number of intact elements is
n − n′
ℓµ .
lid = ′
n −1
(7.8)
k = kmax = (n − 1) ℓµ .
(7.9)
(3) The binary multiplace representation which is a binarizing heteromorphism (5.26) with n′ =
(max)
n′min = 2, the maximal number of intact elements lid
= (n − 2) ℓµ and maximal number
of places
26
STEVEN DUPLIJ
The multiplace representations for n-ary semigroups have no additional defining relations, as compared with (7.5). In case of n-ary groups, we need an analog of the “normalizing” relation (7.2). If
the n-ary group has the unity e, then one can put
e
(n,n′ ) .
..
Πk
k = idV .
e
(7.10)
If there is no unity at all, one can “normalize” the multiplace representation, using analogy with (7.2)
in the form
Π1 h−1 ∗ h = idV ,
(7.11)
as follows
h̄
...
ℓµ
(n,n′ ) h̄
Πk
h
.
..
ℓid
h
= idV ,
(7.12)
for all h ∈G n , where h̄ is the querelement of h. The latter ones can be placed on any places in the
l.h.s. of (7.12) due to the Dörnte identities. Also, the multiplications in the l.h.s. of (7.5) can change
their place due to the same reason.
A general form of multiplace representations can be found by applying the Dörnte identities to each
n-ary product in the l.h.s. of (7.5). Then, using (7.12) we have schematically
g1
′
′
(n,n ) .
(n,n )
..
Πk
= Πk
gk
t1
..
.
tℓµ
g
..
. ℓid
g
,
(7.13)
where g is an arbitrary fixed element of the n-ary group and
ta = µn [ga1 , . . . , gan−1 , ḡ] , a = 1, . . . , ℓµ .
(7.14)
This is the special shape of some multiplace representations, while the concrete formulas should be
obtained in each case separately. Nevertheless, some conclusions can be drawn from (7.13). Firstly,
(n,n′ )
the equivalence classes on which Πk
is constant are determined by fixing ℓµ + 1 elements, i.e. by
the surface ta = const, g = const. Secondly, some k-place representations of a n-ary group can be
reduced to ℓµ -place representations of its retract. In the case ℓµ = 1, multiplace representations of
a n-ary group derived from a binary group correspond to ordinary representations of the latter (see
B OROWIEC ET AL . [2006], D UDEK [2007]).
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
27
(n,2)
Example 7-1. Let us consider the case of a binary multiplace representation Π2n−2 of n-ary group
G n = hG; µn i with two multiplications ℓµ = 2 defined by the associativity preserving equation
(n,2)
Π2n−2
µn [g1 , u1 , . . . un−2 , g2 ]
u′1
..
.
u′n−2
µn [h1 , v1 , . . . vn−2 , h2 ]
v1′
..
.
′
vn−2
(n,2)
=
Π
2n−2
g1
u1
..
.
un−2
h1
v1
..
.
vn−2
(n,2)
∗
Π
2n−2
g2
u′1
..
.
u′n−2
h2
v1′
..
.
′
vn−2
(7.15)
and normalizing condition
(n,2)
Π2n−2
h
..
.
h
h̄
h
..
.
h
h̄
n−2
= idV ,
n−2
(7.16)
where h ∈G n is arbitrary. Using (7.15) and (7.16) for a general form of this (2n − 2)-place representation we have
(n,2)
Π2n−2
g1
u1
..
.
(n,2) un−2
= Π2n−2
h1
v1
..
.
vn−2
g1
u1
..
.
(n,2) un−2
= Π2n−2
h1
v1
..
.
vn−2
(n,2)
∗ Π2n−2
h
..
. n − 2
h
h̄
h
..
. n − 2
h
h̄
(n,2)
= Π2n−2
µn [g1 , u1 , . . . un−2 , h]
h
..
. n − 3
h
h̄
µn [h
1 , v1 , . . . vn−2 , h]
h
..
. n − 3
h
h̄
.
(7.17)
The Dörnte identity applied to the first elements of the products in the r.h.s. of (7.17) together with
associativity of µn gives
n−2
z
}|
{
µn g, . . . , g, tg , h
h
.
..
n−3
h
h̄
(n,2)
(n,2)
=
Π
= Π2n−2
2n−2
n−2
z }| {
µn g, . . . , g, th , h
h
.
.
. n − 3
h
h̄
(n,2)
Π2n−2
g
..
. n −2
g
tg
g
..
. n −2
g
th
(n,2)
∗Π2n−2
h
..
.
h
h̄
h
..
.
h
h̄
n−2
(n,2)
= Π2n−2
n−2
g
..
. n − 2
g
tg
g
..
. n − 2
g
th
(7.18)
,
28
STEVEN DUPLIJ
where
tg = µn [ḡ, g1 , u1 , . . . un−2] , th = µn [ḡ, h1 , v1 , . . . vn−2 ] .
Thus, the equivalent classes of the multiplace representation
(ℓµ + 1 = 3)-element surface
(n,2)
Π2n−2
(7.19)
(7.15) are determined by the
tg = const, th = const, g = const.
(7.20)
Note that the “ℓµ -place reduction” of a multiplace representation is possible not for all associativity
preserving heteromorphism equations. For instance, if to exchange vi ↔ vi′ in l.h.s. of (7.15), then
the associativity remains, but the “ℓµ -place reduction”, analogous to (7.18) will not be possible.
In case, when it is possible, the corresponding ℓµ -place representation can be realized on the binary
retract of G n (for some special n-ary groups and ℓµ = 1 see see B OROWIEC
AL . [2006], D UDEK
" ETn−2
#
z
}|
{
def
[2007]). Indeed, let Gret
2 = hG, ⊛i = retg hG; µn i, where g1 ⊛ g2 = µn g1 , g, . . . , g, g2 , g1 , g2 , g ∈
G and we define the reduced ℓµ -place representation (now ℓµ = 2) through (7.18) as follows
g
Πret
2
g1
g2
def
(n,2)
= Π2n−2
g
..
. n − 2
g
g1
g
..
. n − 2
g
g2
.
(7.21)
From (7.17) and (7.16) we obtain
n−2
z }| {
µn g, . . . , g, g1 , g
g
g
.
..
..
n−3
. n − 2
g
g
′
′
g1
g1
(n,2)
= Π2n−2
n−2
g
z }| {
..
µn g, . . . , g, g2 , g
n
−
2
.
g
g
′
g2
..
n
−
3
.
g
g2′
g
Πret
2
g1
g2
(n,2)
= Π2n−2
g
∗ Πret
2
g
..
.
g
g1′
g2′
(n,2)
= Π2n−2
n−2
n−2
z }| {
µn g1 , g, . . . , g, g1′
g
..
.
g
n−2
n−2
z }| { ′
µn g2 , g, . . . , g, g2
g
..
. n − 2
g
g1
g
..
. n − 2
g
g2
(n,2)
=Π
2n−2
(n,2)
∗ Π2n−2
g
..
. n − 2
g
g1
⊛ g1′
g
..
. n − 2
g
g2 ⊛ g2′
g1 ⊛ g1′
ret g
.
= Π2
g2 ⊛ g2′
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
In the framework of our classification
g
Πret
2
g1
g2
29
is a hom-like 2-place (binary) representation.
The above formulas describe various properties of multiplace representations, but they give no
idea of how to build representations for concrete polyadic systems. The most common method of
representation construction uses the concept of a group action on a set (see, e.g., K IRILLOV [1976]).
Below we extend this concept to the multiplace case and use corresponding heteromorphisms, as it
was done above for homomorphisms and representations.
8. M ULTIACTIONS
AND
G- SPACES
Let G n = hG; µn i be a polyadic system and X be a set. A (left) 1-place action of G n on X is the
(n)
external binary operation ρ1 : G × X → X such that it is consistent with the multiplication µn , i.e.
composition of the binary operations ρ1 {g|x} gives the n-ary product, that is,
n
o o
n
(n)
(n)
(n)
(n)
g2 | . . . |ρ1 {gn |x} . . . , g1 , . . . , gn ∈ G, x ∈ X. (8.1)
g1 |ρ1
ρ1 {µn [g1 , . . . gn ] |x} = ρ1
If the polyadic system is a n-ary group, then in addition to (8.1) it is implied the there exist such
(n)
ex ∈ G (which may or may not coincide with the unity of G n ) that ρ1 {ex |x} = x for all x ∈ X,
(n)
and the mapping x 7→ ρ1 {ex |x} is a bijection of X. The right 1-place actions of G n on X are
defined in a symmetric way, and therefore we will consider below only one of them. Obviously, we
(n)
(n′ )
cannot compose ρ1 and ρ1 with n 6= n′ . Usually X is called a G-set or G-space depending on its
properties (see, e.g., H USEM ÖLLER ET AL . [2008]).
The application of the 1-place action defined by (8.1) to the representation theory of n-ary groups
gave mostly repetitions of the ordinary (binary) group representation results (except for trivial bderived ternary groups) D UDEK AND S HAHRYARI [2012]. Also, it is obviously seen that the construction (8.1) with the binary external operation ρ1 cannot be applied for studying the most important
regular representations of polyadic systems, when the X coincides with Gn itself and the action arises
from translations.
Here we introduce the multiplace concept of action for polyadic systems, which is consistent with
heteromorphisms and multiplace representations. Then we will show how it naturally appears when
X = G n and apply it to construct examples of representations including the regular ones.
For a polyadic system G n = hG; µn i and a set X we introduce an external polyadic operation
ρk : G×k × X → X,
(8.2)
which is called a (left) k-place action or multiaction. To generalize the 1-action composition (8.1),
we use the analogy with multiplication laws of the heteromorphisms (5.14) and the multiplace representations (7.5) and propose (schematically)
µn [g1 , . . . , gn ] ,
n′
.
}|
z
.
ℓµ
{
.
′ −1)
g
g
k(n
1
µ g
n
n(ℓµ −1) , . . . , gnℓµ
(n)
(n)
(n)
..
..
.
.
.
x
. . . . (8.3)
ρ
ρk
x = ρk
.
k
.
gnℓµ+1 ,
gkn′
gk
..
ℓid
.
gnℓµ +ℓid
The connection between all the parameters here is the same as in the arity changing formulas
(5.17)–(5.18). Composition of mappings is associative, and therefore in concrete cases we can use
30
STEVEN DUPLIJ
the associative quiver technique, as it is described in the previous sections. If G n is n-ary group, then
we should add to (8.3) the “normalizing” relations analogous with (7.10) or (7.12). So, if there is a
unity e ∈G n , then
e
(n)
.. x = x, for all x ∈ X.
ρk
(8.4)
.
e
In terms of the querelement, the normalization has the form
h̄
..
ℓ
µ
.
h̄
(n)
ρk
(8.5)
x = x, for all x ∈ X and for all h ∈ G n .
h
..
ℓ
id
.
h
(n)
(n)
The multiaction ρk is transitive, if any two points x and y in X can be “connected” by ρk , i.e.
there exist g1 , . . . , gk ∈Gn such that
g1
(n)
.. x = y.
ρk
(8.6)
.
gk
(n)
If g1 , . . . , gk are unique, then ρk is sharply transitive. The subset of X, in which any points
are connected by (8.6) with fixed g1 , . . . , gk can be called the multiorbit of X. If there is only one
multiorbit, then we call X the heterogenous G-space (by analogy with the homogeneous one). By
analogy with the (ordinary) 1-place actions, we define a G-equivariant map Ψ between two G-sets X
and Y by (in our notation)
g1
g1
(n)
.. Ψ (x) ∈ Y,
.. x = ρ(n)
(8.7)
Ψ ρk
k
.
.
gk
gk
which makes G-space into a category (for details, see, e.g., H USEM ÖLLER ET AL . [2008]). In the
particular case, when X is a vector space over K, the multiaction (8.2) can be called a multi-G-module
which satisfies (8.4) and the additional (linearity) conditions
g1
g1
g1
(n)
.. ax + by = aρ(n)
.. x + bρ(n)
.. y ,
ρk
(8.8)
.
.
k
k
.
gk
gk
gk
where a, b ∈ K. Then, comparing (7.5) and (8.3) we can
multi-G-module by the following formula
g1
(n,n′ ) .
(n)
..
Πk
(x) = ρk
gk
define a multiplace representation as a
g1
.. x .
.
gk
(8.9)
In a similar way, one can generalize to polyadic systems many other notions from group action
theory K IRILLOV [1976].
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
9. R EGULAR
31
MULTIACTIONS
The most important role in the study of polyadic systems is played by the case, when X =G n ,
and the multiaction coincides with the n-ary analog of translations M AL’ TCEV [1954], so called
i-translations B ELOUSOV [1972]. In the binary case, ordinary translations lead to regular represenreg(n)
tations K IRILLOV [1976], and therefore we call such an action a regular multiaction ρk
. In this
connection, the analog of the Cayley theorem for n-ary groups was obtained in G AL’ MAK [1986,
2001b]. Now we will show in examples, how the regular multiactions can arise from i-translations.
Example 9-1. Let G 3 be a ternary semigroup, k = 2, and X =G 3 , then 2-place (left) action can be
defined as
def
g
reg(3)
ρ2
u = µ3 [g, h, u] .
(9.1)
h
This gives the following composition law for two regular multiactions
g2
g1
reg(3)
reg(3)
= µ3 [g1 , h1 , µ3 [g2 , h2 , u]]
u
ρ2
ρ2
h2
h1
= µ3 [µ3 [g1 , h1 , g2] , h2 , u] =
reg(3)
ρ2
µ3 [g1 , h1 , g2 ]
u .
h2
(9.2)
Thus, using the regular 2-action (9.1) we have, in fact, derived the associative quiver corresponding
to (5.6).
The formula (9.1) can be simultaneously treated as a 2-translation B ELOUSOV [1972]. In this way,
the following left regular multiaction
g1
reg(n)
.. h def
ρk
= µn [g1 , . . . , gk , h] ,
(9.3)
.
gk
corresponds to (5.10), where in the r.h.s. there is the i-translation with i = n. The right regular
multiaction corresponds to the i-translation with i = 1. The binary composition of the left regular
multiactions corresponds to (5.10). In general, the value of i fixes the minimal final arity n′reg , which
differs for even and odd values of the initial arity n.
It follows from (9.3) that for regular multiactions the number of places is fixed
kreg = n − 1,
(9.4)
and the arity changing formulas (5.17)–(5.18) become
n′reg = n − ℓid
n′reg
= ℓµ + 1.
(9.5)
(9.6)
From (9.5)–(9.6) we conclude that for any n a regular multiaction having one multiplication ℓµ = 1
is binarizing and has n − 2 intact elements. For n = 3 see (9.2). Also, it follows from (9.5) that for
regular multiactions the number of intact elements gives exactly the difference between initial and
final arities.
If the initial arity is odd, then there exists a special middle regular multiaction generated by the
i-translation with i = (n + 1) 2. For n = 3 the corresponding associative quiver is (6.5) and
such 2-actions were used in B OROWIEC ET AL . [2006] to construct middle representations of ternary
32
STEVEN DUPLIJ
groups, which did not change arity (n′ = n). Here we give a more complicated example of a middle
regular multiaction, which can contain intact elements and can therefore change arity.
Example 9-2. Let us consider 5-ary semigroup and the following middle 4-action
g
i=3
↓
h
reg(5)
s = µ5 g, h, s , u, v .
ρ4
u
v
(9.7)
Using (9.6) we observe that there are two possibilities for the number of multiplications ℓµ = 2, 4.
The last case ℓµ = 4 is similar to the vertical associative quiver (6.5), but with a more complicated
l.h.s., that is
µ
[g
,
h
,
g
,
h
g
]
5 1 1 2 2, 3
µ5 [h3 , g4 , h4 , g5, h5 ]
reg(5)
s =
ρ4
µ5 [u5 , v5 , u4, v4 , u3]
µ [v , u , v , u , v ]
5 3
2 2
1 1
g
g
g
g
g
5
4
3
2
1
h5
h4
h3
h2
h1
reg(5)
reg(5)
reg(5)
reg(5)
reg(5)
. (9.8)
s
ρ4
ρ4
ρ4
ρ4
ρ4
u5
u4
u3
u2
u1
v
v
v
v
v
2
1
5
4
3
Now we have an additional case with two intact elements ℓid and two multiplications ℓµ
g
g
g
µ
[g
,
h
,
g
,
h
g
]
2
1
5
1
1
2
2,
3
3
h
h
h
h
reg(5)
reg(5)
reg(5)
reg(5)
3
2
1
3
ρ4
ρ4
s = ρ4
ρ4
u
u
u
µ
[h
,
v
,
u
,
v
,
u
]
3
2
1
5
3
3
2
2
1
v
v
v
v
1
1
2
3
= 2 as
,
s
(9.9)
with arity changing from n = 5 to n′reg = 3. In addition to (9.9) we have 3 more possible regular
multiactions due to the associativity of µ5 , when the multiplication brackets in the sequences of 6
elements in the first two rows and the second two ones can be shifted independently.
For n > 3, in addition to left, right and middle multiactions, there exist intermediate cases. First,
observe that the i-translations with i = 2 and i = n − 1 immediately fix the final arity n′reg = n.
Therefore, the composition of multiactions will be similar to (9.8), but with some permutations in the
l.h.s.
Now we consider some multiplace analogs of regular representations of binary groups K IRILLOV
[1976]. The straightforward generalization is to consider the previously introduced regular multiactions (9.3) in the r.h.s. of (8.9). Let G n be a finite polyadic associative system and KG n
be a vector space spanned by G n (some properties of n-ary group rings were considered in
Z EKOVI Ć AND A RTAMONOV
P[1999, 2002]). This means that any element of KG n can be uniquely
presented in the form w = l al · hl , al ∈ K, hl ∈ G. Then, using (9.3) and (8.9) we define the
i-regular k-place representation by
g1
X
reg(i) .
..
Πk
(w) =
al · µk+1 [g1 . . . gi−1 hl gi+1 . . . gk ] .
(9.10)
l
gk
Comparing (9.3) and (9.10) one can conclude that all the general properties of multiplace regular
representations are similar to those of the regular multiactions. If i = 1 or i = k, the multiplace
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
33
representation is called a right or left regular representation respectively. If k is even, the representation with i = k 2 + 1 is called a middle regular representation. The case k = 2 was considered in
B OROWIEC ET AL . [2006] for ternary groups.
10. M ULTIPLACE
REPRESENTATIONS OF TERNARY GROUPS
Let us consider the case n = 3, k = 2 in more detail, paying attention to its special peculiarities,
which corresponds to the 2-place (bi-element) representations of ternary groups B OROWIEC ET AL .
[2006]. Let V be a vector space over K and End V be a set of linear endomorphisms of V . From now
on we denote the ternary multiplication
bysquare brackets only, as follows µ3 [g1 , g2 , g3 ] ≡ [g1 g2 g3 ],
g1
and use the “horizontal” notation Π
≡ Π (g1 , g2 ).
g2
Definition 10-1. A left representation of a ternary group G, [ ]) in V is a map ΠL : G × G → End V
such that
ΠL (g1 , g2 ) ◦ ΠL (g3 , g4 ) = ΠL ([g1 g2 g3 ] , g4 ) ,
ΠL (g, g) = idV ,
(10.1)
(10.2)
where g, g1, g2 , g3 , g4 ∈ G.
Replacing in (10.2) g by g we obtain ΠL (g, g) = idV , which means that in fact (10.2) has the form
Π (g, g) = ΠL (g, g) = idV , ∀g ∈ G. Note that the axioms considered in the above definition are
the natural ones satisfied by left multiplications g 7→ [abg]. For all g1 , g2, g3 , g4 ∈ G we have
L
ΠL ([g1 g2 g3 ] , g4 ) = ΠL (g1 , [g2 g3 g4 ]) .
For all g, h, u ∈ G we have
ΠL (g, h) = ΠL ([guu], h) = ΠL (g, u) ◦ ΠL (u, h)
(10.3)
and
ΠL (g, u) ◦ ΠL (u, g) = ΠL (u, g) ◦ ΠL (g, u) = idV ,
(10.4)
−1
and therefore every ΠL (g, u) is invertible and ΠL (g, u)
= ΠL (u, g). This means that any left
representation gives a representation of a ternary group by a binary group B OROWIEC ET AL . [2006].
If the ternary group is medial, then
ΠL (g1 , g2) ◦ ΠL (g3 , g4 ) = ΠL (g3 , g4 ) ◦ ΠL (g1 , g2 ) ,
i.e. the group so obtained is commutative. If the ternary group hG, [ ]i is commutative, then also
ΠL (g, h) = ΠL (h, g), because
ΠL (g, h) = ΠL (g, h) ◦ ΠL (g, g) = ΠL ([g h g] , g)
= ΠL ([h g g] , g) = ΠL (h, g) ◦ ΠL (g, g) = ΠL (h, g) .
In the case of a commutative and idempotent ternary group any of its left representations is idempotent
−1
and ΠL (g, h)
= ΠL (g, h), so that commutative and idempotent ternary groups are represented
by Boolean groups.
Assertion 10-2. Let hG, [ ]i = der (G, ⊙) be a ternary group derived from a binary group hG, ⊙i,
then there is one-to-one correspondence between representations of (G, ⊙) and left representations
of (G, [ ]).
34
STEVEN DUPLIJ
Indeed, because (G, [ ]) = der (G, ⊙), then g ⊙ h = [geh] and e = e, where e is unity of the binary
group (G, ⊙). If π ∈ Rep (G, ⊙), then (as it is not difficult to see) ΠL (g, h) = π (g) ◦ π (h) is a left
representation of hG, [ ]i. Conversely, if ΠL is a left representation of hG, [ ]i then π (g) = ΠL (g, e)
is a representation of (G, ⊙). Moreover, in this case ΠL (g, h) = π (g) ◦ π (h), because we have
ΠL (g, h) = ΠL (g, [ehe]) = ΠL ([geh], e) = ΠL (g, e) ◦ ΠL (h, e) = π (g) ◦ π (h) .
Let (G, [ ]) be a ternary group and (G × G, ∗) be a semigroup used in the construction of left
representations. According to Post P OST [1940] one says that two pairs (a, b), (c, d) of elements of
G are equivalent, if there exists an element g ∈ G such that [abg] = [cdg]. Using a covering group we
can see that if this equation holds for some g ∈ G, then it holds also for all g ∈ G. This means that
ΠL (a, b) = ΠL (c, d) ⇐⇒ (a, b) ∼ (c, d),
i.e.
ΠL (a, b) = ΠL (c, d) ⇐⇒ [abg] = [cdg]
for some g ∈ G. Indeed, if [abg] = [cdg] holds for some g ∈ G, then
ΠL (a, b) = ΠL (a, b) ◦ ΠL (g, g) = ΠL ([abg], g)
By analogy we can define
= ΠL ([cdg], g) = ΠL (c, d) ◦ ΠL (g, g) = ΠL (c, d).
Definition 10-3. A right representation of a ternary group (G, [ ]) in V is a map ΠR : G×G → End V
such that
ΠR (g3 , g4 ) ◦ ΠR (g1 , g2 ) = ΠR (g1 , [g2 g3 g4 ]) ,
ΠR (g, g) = idV ,
(10.5)
(10.6)
where g, g1, g2 , g3 , g4 ∈ G.
From (10.5)-(10.6) it follows that
ΠR (g, h) = ΠR (g, [u u h]) = ΠR (u, h) ◦ ΠR (g, u) .
(10.7)
−1
It is easy to check that ΠR (g, h) = ΠL h, g = ΠL (g, h) . So it is sufficient to consider only
left representations (as in the binary case). Consider the following example of a group algebra ternary
generalization B OROWIEC ET AL . [2006].
Example 10-4. Let G be a ternary group and KG be a vector space
P spanned by G, which means that
any element of KG can be uniquely presented in the form t = ni=1 ki hi , ki ∈ K, hi ∈ G, n ∈ N (we
do not assume that G has finite rank). Then left and right regular representations are defined by
n
X
ΠLreg (g1 , g2 ) t =
ki [g1 g2 hi ] ,
(10.8)
i=1
ΠR
reg (g1 , g2 ) t =
n
X
ki [hi g1 g2 ] .
(10.9)
i=1
Let us construct the middle representations as follows.
Definition 10-5. A middle representation of a ternary group hG, [ ]i in V is a map ΠM : G × G →
End V such that
ΠM (g3 , h3 ) ◦ ΠM (g2 , h2 ) ◦ ΠM (g1 , h1 ) = ΠM ([g3 g2 g1 ] , [h1 h2 h3 ]) ,
ΠM (g, h) ◦ ΠM g, h = ΠM g, h ◦ ΠM (g, h) = idV
(10.10)
(10.11)
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
35
It can be seen that a middle representation is a ternary group
homomorphism ΠM : G × Gop →
der End V. Note that instead of (10.11) one can use ΠM g, h ◦ ΠM (g, h) = idV after changing g to
g and taking into account that g = g. In the case of idempotent elements g and h we have ΠM (g, h) ◦
ΠM (g, h) = idV , which means that the matrices ΠM are Boolean. Thus all middle representation
matrices of idempotent ternary groups are Boolean. The composition ΠM (g1 , h1 ) ◦ ΠM (g2 , h2 ) is not
a middle representation, but the following proposition nevertheless holds.
Let ΠM be a middle representation of a ternary group hG, [ ]i, then, if ΠLu (g, h) = ΠM (g, u) ◦
M
Π (h, u) is a left representation of hG, [ ]i, then ΠLu (g, h) ◦ ΠLu′ (g ′ , h′ ) = ΠLu′ ([ghu′] , h′ ), and,
M
M
′
′
R
R
if ΠR
u (g, h) = Π (u, h) ◦ Π (u, g) is a right representation of hG, [ ]i, then Πu (g, h) ◦ Πu′ (g , h ) =
R
′ ′
L
R
Πu (g, [hg h ]). In particular, Πu (Πu ) is a family of left (right) representations.
If a middle representation ΠM of a ternary group hG, [ ]i satisfies ΠM (g, g) = idV for all g ∈ G,
then it is a left and a right representation and ΠM (g, h) = ΠM (h, g) for all g, h ∈ G. Note that in
general ΠM
reg (g, g) 6= id. For regular representations we have the following commutation relations
R
L
ΠLreg (g1 , h1 ) ◦ ΠR
reg (g2 , h2 ) = Πreg (g2 , h2 ) ◦ Πreg (g1 , h1 ) .
Let hG, [ ]i be a ternary group and let G × G, [ ]′ be a ternary group used in the construction of
the middle representation. In hG, [ ]i, and consequently in G × G, [ ]′ , we define the relation
(a, b) ∼ (c, d) ⇐⇒ [aub] = [cud]
for all u ∈ G. It is not difficult to see that this relation is a congruence in the ternary group
M
G × G, [ ]′ . For regular representations ΠM
reg (a, b) = Πreg (c, d) if (a, b) ∼ (c, d). We have the
following relation
a h a′ ⇐⇒ a = [ga′ g] for some g ∈ G
or equivalently
a h a′ ⇐⇒ a′ = [gag] for some g ∈ G.
It is not difficult to see that it is an equivalence relation on hG, [ ]i, moreover, if hG, [ ]i is medial,
then this relation is a congruence.
Let G × G, [ ]′ be a ternary group used in a construction of middle representations, then
(a, b) h (a′ , b) ⇐⇒ a′ = [gag] and
b′ = [hbb ] for some (g, h) ∈ G × G
is an equivalence relation on G × G, [ ]′ . Moreover, if (G, [ ]) is medial, then this relation is a
congruence. Unfortunately, however it is a weak relation. In a ternary group Z3 , where [ghu] =
(g − h + u) (mod 3) we have only one class, i.e. all elements are equivalent. In Z4 with the operation
[ghu] = (g + h + u + 1) (mod 4) we have a h a′ ⇐⇒ a = a′ . However, for this relation the
following statement holds. If (a, b) h (a′ , b′ ), then
tr ΠM (a, b) = tr ΠM (a′ , b′ ).
We have tr(AB) = tr(BA) for all A, B ∈ EndV , and
tr ΠM (a, b) = tr ΠM ([ga′ g], [hb′ h] ) = tr ΠM (g, h) ◦ ΠM (a′ , b′ ) ◦ ΠM (g, h)
= tr ΠM (g, h) ◦ ΠM (g, h) ◦ ΠM (a′ , b′ ) = tr idV ◦ ΠM (a′ b′ ) = tr ΠM (a′ , b′ )
In our derived case the connection with standard group representations is given by the following.
Let (G, ⊙) be a binary group, and the ternary derived group as hG, [ ]i = der (G, ⊙). There is
one-to-one correspondence between a pair of commuting binary group representations and a middle
36
STEVEN DUPLIJ
ternary derived group representation. Indeed, let π, ρ ∈ Rep (G, ⊙), π (g) ◦ ρ (h) = ρ (h) ◦ π (g) and
ΠL ∈ Rep (G, [ ]). We take
ΠM (g, h) = π (g) ◦ ρ h−1 , π (g) = ΠM (g, e) , ρ (g) = ΠM (e, g) .
Then using (10.10) we prove the needed representation laws.
Let hG, [ ]i be a fixed ternary group, G × G, [ ]′ a corresponding ternary group used in the construction of middle representations, ((G × G)∗ , ⊛) a covering group of G × G, [ ]′ , (G × G, ⋄) =
ret(a,b) (G × G, h i). If ΠM (a, b) is a middle representation of hG, [ ]i, then π defined by
π(g, h, 0) = ΠM (g, h), π(g, h, 1) = ΠM (g, h) ◦ ΠM (a, b)
is a representation of the covering group P OST [1940]. Moreover
ρ(g, h) = ΠM (g, h) ◦ ΠM (a, b) = π(g, h, 1)
is a representation of the above retract induced by (a, b). Indeed, (a, b) is the identity of this retract
and ρ(a, b) = ΠM (a, b) ◦ ΠM (a, b) = idV . Similarly
ρ ((g, h) ⋄ (u, u)) = ρ (h(g, h), (a, b), (u, u)i) = ρ ([gau], [ubh]) = ΠM ([gau], [ubh])) ◦ ΠM (a, b)
= ΠM (g, h) ◦ ΠM (a, b) ◦ ΠM (u, u) ◦ ΠM (a, b) = ρ(g, h) ◦ ρ(u, u)
But τ (g) = (g, g) is an embedding of (G, [ ]) into G × G, [ ]′ . Hence µ defined by µ(g, 0) =
ΠM (g, g) and µ(g, 1) = ΠM (g, g) ◦ ΠM (a, a) is a representation of a covering group G∗ for (G, [ ])
(see the Post theorem P OST [1940] for a = c). On the other hand, β(g) = ΠM (g, g) ◦ ΠM (a, a) is a
representation of a binary retract (G, · ) = reta (G, [ ]). Thus β can induce some middle representation
of (G, [ ]) (by the Gluskin-Hosszú theorem G LUSKIN [1965]).
Note that in the ternary group of quaternions hK, [ ]i (with norm 1), where [ghu] = ghu(−1) =
−ghu and gh is the multiplication of quaternions (−1 is a central element) we have 1 = −1, −1 = 1
and g = g for others. In K × K, [ ]′ we have (a, b) ∼ (−a, −b) and (a, −b) ∼ (−a, b), which gives
32 two-element equivalence classes. The embedding τ (g) = (g, g) suggest that ΠM (i, i) = π(i) 6=
π(−i) = ΠM (−i, −i). Generally ΠM (a, b) 6= ΠM (−a, −b) and ΠM (a, −b) 6= ΠM (−a, b).
The relation (a, b) ∼ (c, d) ⇐⇒ [abg] = [cdg] for all g ∈ G is a congruence on (G × G, ∗). Note
that this relation can be defined as ”for some g”. Indeed, using a covering group we can see that if
[abg] = [cdg] holds for some g then it holds also for all g. Thus π L (a, b) = ΠL (c, d) ⇐⇒ (a, b) ∼
(c, d). Indeed
ΠL (a, b) = ΠL (a, b) ◦ ΠL (g, g) = ΠL ([a b g], g)
= ΠL ([c d g], g) = ΠL (c, d) ◦ ΠL (g, g) = ΠL (c, d).
We conclude, that every left representation of a commutative group hG, [ ]i is a middle representation. Indeed,
ΠL (g, h) ◦ ΠL (g, h) = ΠL ([g h g], h) = ΠL ([g g h], h) = ΠL (h, h) = idV
and
ΠL (g1 , g2 ) ◦ ΠL (g3 , g4 ) ◦ ΠL (g5 , g6) = ΠL ([[g1 g2 g3 ]g4 g5 ], g6 ) = ΠL ([[g1 g3 g2 ]g4 g5 ], g6 )
= ΠL ([g1 g3 [g2 g4 g5 ]], g6 ) = ΠL ([g1 g3 [g5 g4 g2 ]], g6 ) = ΠL ([g1 g3 g5 ], [g4 g2 g6 ]) = ΠL ([g1 g3 g5 ], [g6 g4 g2 ]).
Note that the converse holds only for the special kind of middle representations such that
Π (g, g) = idV . Therefore,
M
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
37
Assertion 10-6. There is one-one correspondence between left representations of hG, [ ]i and binary
representations of the retract reta (G, [ ]).
Indeed, let ΠL (g, a) be given, then ρ(g) = ΠL (g, a) is such representation of the retract, as can be
directly shown. Conversely, assume that ρ(g) is a representation of the retract reta (G, [ ]). Define
ΠL (g, h) = ρ(g) ◦ ρ(h)−1 , then ΠL (g, h) ◦ ΠL (u, u) = ρ(g) ◦ ρ(h)−1 ◦ ρ(u) ◦ ρ(u)−1 = ρ(g ⊛ (h)−1 ◦
⊛u) ◦ ρ(u)−1 = ρ([ [ g a [ a h a ] ] a u ]) ◦ ρ(u)−1 = ρ([ g h g ]) ◦ ρ(u)−1 = ΠL ([ g h u ], u),
11. M ATRIX
REPRESENTATIONS OF TERNARY GROUPS
Here we give several examples of matrix representations for concrete ternary groups. Let G =
Z3 ∋ {0, 1, 2} and the ternary multiplication be [ghu] = g − h + u. Then [ghu] = [uhg] and 0 = 0,
1 = 1, 2 = 2, therefore (G, [ ]) is an idempotent medial ternary group. Thus ΠL (g, h) = ΠR (h, g)
and
ΠL (a, b) = ΠL (c, d) ⇐⇒ (a − b) = (c − d)mod 3.
(11.1)
The calculations give the left regular representation in the manifest matrix form
ΠLreg (0, 0) = ΠLreg (2, 2) = ΠLreg (1, 1) = ΠR
reg (0, 0)
1 0 0
R
0 1 0 = [1] ⊕ [1] ⊕ [1],
= ΠR
reg (2, 2) = Πreg (1, 1) =
0 0 1
(11.2)
0 1 0
R
R
0 0 1
ΠLreg (2, 0) = ΠLreg (1, 2) = ΠLreg (0, 1) = ΠR
reg (2, 1) = Πreg (1, 0) = Πreg (0, 2) =
1 0 0
√
3
1
−
−
√
√
1
1
1
1
√2
2 = [1] ⊕ − + i 3 ⊕ − − i 3 ,
(11.3)
= [1] ⊕
3
1
2 2
2 2
−
2
2
0 0 1
R
R
1 0 0
ΠLreg (2, 1) = ΠLreg (1, 0) = ΠLreg (0, 2) = ΠR
reg (2, 0) = Πreg (1, 2) = Πreg (0, 1) =
0 1 0
√
1
3
√
√
1
1
1
1
−√2
2 = [1] ⊕ − − i 3 ⊕ − + i 3 .
= [1] ⊕
(11.4)
3
1
2 2
2 2
−
−
2
2
Consider next the middle representation construction. The middle regular representation is defined
by
n
X
ΠM
(g
,
g
)
t
=
ki [g1 hi g2 ] .
1 2
reg
i=1
For regular representations we have
R
R
M
ΠM
reg (g1 , h1 ) ◦ Πreg (g2 , h2 ) = Πreg (h2 , h1 ) ◦ Πreg (g1 , g2 ) ,
ΠM
reg
(g1 , h1 ) ◦
ΠLreg
(g2 , h2 ) =
ΠLreg
(g1 , g2 ) ◦
ΠM
reg
(h2 , h1 ) .
(11.5)
(11.6)
38
STEVEN DUPLIJ
For the middle regular representation matrices we obtain
1 0 0
M
M
0 0 1
ΠM
(0,
0)
=
Π
(1,
2)
=
Π
(2,
1)
=
reg
reg
reg
0 1 0
0 1 0
M
M
1 0 0
(1,
0)
=
Π
(2,
2)
=
ΠM
(0,
1)
=
Π
reg
reg
reg
0 0 1
0 0 1
M
M
0 1 0
ΠM
(0,
2)
=
Π
(2,
0)
=
Π
(1,
1)
=
reg
reg
reg
1 0 0
,
,
.
The above representation ΠM
reg of hZ3 , [ ]i is equivalent to the orthogonal direct sum of two irreducible representations
−1 0
M
M
M
Πreg (0, 0) = Πreg (1, 2) = Πreg (2, 1) = [1] ⊕
,
0 1
√
1
3
−
M
M
2
2
√
ΠM
,
reg (0, 1) = Πreg (1, 0) = Πreg (2, 2) = [1] ⊕
3
1
−
−
2 √ 2
1
3
√2
M
M
2
ΠM
,
reg (0, 2) = Πreg (2, 0) = Πreg (1, 1) = [1] ⊕
3
1
−
2
2
i.e. one-dimensional trivial [1] and two-dimensional irreducible. Note, that in this example
ΠM (g, g) = ΠM (g, g) 6= idV , but ΠM (g, h) ◦ ΠM (g, h) = idV , and so ΠM are of the second degree.
Consider a more complicated example of left representations. Let G = Z4 ∋ {0, 1, 2, 3} and the
ternary multiplication be
[ghu] = (g + h + u + 1) mod 4.
(11.7)
We have the multiplication table
1
2
[g, h, 0] =
3
0
3
0
[g, h, 2] =
1
2
2
3
0
1
3
0
1
2
0
1
2
3
1
2
3
0
0
1
2
3
2
3
0
1
2
3
[g, h, 1] =
0
1
0
1
[g, h, 3] =
2
3
3
0
1
2
0
1
2
3
1
2
3
0
2
3
0
1
1
2
3
0
3
0
1
2
Then the skew elements are 0 = 3, 1 = 2, 2 = 1, 3 = 0, and therefore (G, [ ]) is a (non-idempotent)
commutative
ternary group. The left representation is defined by the expansion ΠLreg (g1 , g2 ) t =
Pn
i=1 ki [g1 g2 hi ], which means that (see the general formula (9.10))
ΠLreg (g, h) |u >= | [ghu] > .
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
39
Analogously, for right and middle representations
M
ΠR
reg (g, h) |u >= | [ugh] >, Πreg (g, h) |u >= | [guh] > .
Therefore | [ghu] >= | [ugh] >= | [guh] > and
M
ΠLreg (g, h) = ΠR
reg (g, h) |u >= Πreg (g, h) |u >,
M
so ΠLreg (g, h) = ΠR
reg (g, h) = Πreg (g, h). Thus it is sufficient to consider the left representation only.
In this case the equivalence is ΠL (a, b) = ΠL (c, d) ⇐⇒ (a + b) = (c + d)mod 4, and we obtain
the following classes
0 0 0 1
1 0 0 0
ΠLreg (0, 0) = ΠLreg (1, 3) = ΠLreg (2, 2) = ΠLreg (3, 1) =
0 1 0 0 = [1] ⊕ [−1] ⊕ [−i] ⊕ [i] ,
0 0 1 0
0 0 1 0
0 0 0 1
ΠLreg (0, 1) = ΠLreg (1, 0) = ΠLreg (2, 3) = ΠLreg (3, 2) =
1 0 0 0 = [1] ⊕ [−1] ⊕ [−1] ⊕ [−1] ,
0 1 0 0
0 1 0 0
0 0 1 0
ΠLreg (0, 2) = ΠLreg (1, 1) = ΠLreg (2, 0) = ΠLreg (3, 3) =
0 0 0 1 = [1] ⊕ [−1] ⊕ [i] ⊕ [−i] ,
1 0 0 0
1 0 0 0
0 1 0 0
ΠLreg (0, 3) = ΠLreg (1, 2) = ΠLreg (2, 1) = ΠLreg (3, 0) =
0 0 1 0 = [1] ⊕ [−1] ⊕ [1] ⊕ [1] .
0 0 0 1
It is seen that, due to the fact that the ternary operation (11.7) is commutative, there are only onedimensional irreducible left representations.
Let us “algebralize” the above regular representations in the following way. From (10.1) we have,
for the left representation
ΠLreg (i, j) ◦ ΠLreg (k, l) = ΠLreg (i, [jkl]) ,
(11.8)
where [jkl] = j − k + l, i, j, k, l ∈ Z3 . Denote γiL = ΠLreg (0, i), i ∈ Z3 , then we obtain the algebra
with the relations
L
γiL γjL = γi+j
.
(11.9)
Conversely, any matrix representation of γi γj = γi+j leads to the left representation by ΠL (i, j) =
M
γj−i . In the case of the middle regular representation we introduce γk+l
= ΠM
reg (k, l), k, l ∈ Z3 , then
we obtain
M
γiM γjM γkM = γ[ijk]
,
i, j, k ∈ Z3 .
(11.10)
In some sense (11.10) can be treated as a ternary analog of the Clifford algebra. As before, any
matrix representation of (11.10) gives the middle representation ΠM (k, l) = γk+l .
40
STEVEN DUPLIJ
12. T ERNARY
ALGEBRAS AND
H OPF
ALGEBRAS
Let us consider associative ternary algebras C ARLSSON [1980], DE A ZCARRAGA AND I ZQUIERDO
[2010]. One can introduce an autodistributivity property [[xyz] ab] = [[xab] [yab] [zab]] (see D UDEK
[1993]). If we take 2 ternary operations { , , } and [ , , ], then distributivity is given by {[xyz] ab} =
[{xab} {yab} {zab}]. If (+) is a binary operation (addition), then left linearity is
[(x + z) , a, b] = [xab] + [zab] .
(12.1)
By analogy one can define central (middle) and right linearity. Linearity is defined, when left, middle
and right linearity hold simultaneously.
Definition 12-1. An associative ternary algebra is a triple A, µ3 , η (3) , where A is a linear space over
a field K, µ3 is a linear map A ⊗ A ⊗ A → A called ternary multiplication µ3 (a ⊗ b ⊗ c) = [abc]
which is ternary associative [[abc] de] = [a [bcd] e] = [ab [cde]] or
µ3 ◦ (µ3 ⊗ id ⊗ id) = µ3 ◦ (id ⊗µ3 ⊗ id) = µ3 ◦ (id ⊗ id ⊗µ3 ) .
There are two types D UPLIJ [2001] of ternary unit maps η (3) : K → A:
1) One strong unit map
µ3 ◦ η (3) ⊗ η (3) ⊗ id = µ3 ◦ η (3) ⊗ id ⊗η (3) = µ3 ◦ id ⊗η (3) ⊗ η (3) = id;
(3)
(12.2)
(12.3)
(3)
2) Two sequential units η1 and η2 satisfying
(3)
(3)
(3)
(3)
(3)
(3)
µ3 ◦ η1 ⊗ η2 ⊗ id = µ3 ◦ η1 ⊗ id ⊗η2
= µ3 ◦ id ⊗η1 ⊗ η2
= id;
(12.4)
In first case the ternary analog of the binary relation η (2) (x) = x1, where x ∈ K, 1 ∈ A, is
η (3) (x) = [x, 1, 1] = [1, 1, x] = [1, x, 1] .
(12.5)
Let (A, µA , ηA ), (B, µB , ηB ) and (C, µC , ηC ) be ternary algebras, then the ternary tensor product
space A ⊗ B ⊗ C is naturally endowed with the structure of an algebra. The multiplication µA⊗B⊗C
on A ⊗ B ⊗ C reads
[(a1 ⊗ b1 ⊗ c1 )(a2 ⊗ b2 ⊗ c2 )(a3 ⊗ b3 ⊗ c3 )] = [a1 a2 a3 ] ⊗ [b1 b2 b3 ] ⊗ [c1 c2 c3 ] ,
(12.6)
and so the set of ternary algebras is closed under taking ternary tensor products. A ternary algebra
map (homomorphism) is a linear map between ternary algebras f : A → B which respects the ternary
algebra structure
f ([xyz]) = [f (x) , f (y) , f (z)] ,
(12.7)
f (1A ) = 1B .
(12.8)
Let C be a linear space over a field K.
Definition 12-2. A ternary comultiplication ∆3 is a linear map over a field K such that
∆3 : C → C ⊗ C ⊗ C.
(12.9)
Pn
In the standard Sweedler notations S WEEDLER [1969] ∆3 (a) = i=1 a′i ⊗a′′i ⊗a′′′
i = a(1) ⊗a(2) ⊗
a(3) . Consider different possible types of ternary coassociativity D UPLIJ [2001], B OROWIEC ET AL .
[2001].
(1) A standard ternary coassociativity
(∆3 ⊗ id ⊗ id) ◦ ∆3 = (id ⊗∆3 ⊗ id) ◦ ∆3 = (id ⊗ id ⊗∆3 ) ◦ ∆3 ,
(12.10)
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
41
(2) A nonstandard ternary Σ-coassociativity (Gluskin-type positional operatives)
(∆3 ⊗ id ⊗ id) ◦ ∆3 = (id ⊗ (σ ◦ ∆3 ) ⊗ id) ◦ ∆3 ,
where σ ◦ ∆3 (a) = ∆3 (a) = a(σ(1)) ⊗ a(σ(2)) ⊗ a(σ(3)) and σ ∈ Σ ⊂ S3 .
(3) A permutational ternary coassociativity
(∆3 ⊗ id ⊗ id) ◦ ∆3 = π ◦ (id ⊗∆3 ⊗ id) ◦ ∆3 ,
where π ∈ Π ⊂ S5 .
A ternary comediality is
(∆3 ⊗ ∆3 ⊗ ∆3 ) ◦ ∆3 = σmedial ◦ (∆3 ⊗ ∆3 ⊗ ∆3 ) ◦ ∆3 ,
∈ S9 . A ternary counit is defined as a map ε(3) : C → K. In general,
where σmedial = 123456789
147258369
ε(3) 6= ε(2) satisfying one of the conditions below. If ∆3 is derived, then maybe ε(3) = ε(2) , but another
counits may exist. There are two types of ternary counits:
(1) Standard (strong) ternary counit
(ε(3) ⊗ ε(3) ⊗ id) ◦ ∆3 = (ε(3) ⊗ id ⊗ε(3) ) ◦ ∆3 = (id ⊗ε(3) ⊗ ε(3) ) ◦ ∆3 = id,
(3)
(12.11)
(3)
(2) Two sequential (polyadic) counits ε1 and ε2
(3)
(3)
(3)
(3)
(3)
(3)
(ε1 ⊗ ε2 ⊗ id) ◦ ∆ = (ε1 ⊗ id ⊗ε2 ) ◦ ∆ = (id ⊗ε1 ⊗ ε2 ) ◦ ∆ = id,
(12.12)
Below we will consider only the first standard type of associativity (12.10). The σ-cocommutativity
is defined as σ ◦ ∆3 = ∆3 .
Definition 12-3. A ternary coalgebra is a triple C, ∆3 , ε(3) , where C is a linear space and ∆3 is a
ternary comultiplication (12.9) which is coassociative in one of the above senses and ε(3) is one of the
above counits.
Let (A, µ3 ) be a ternary algebra and (C, ∆3 ) be a ternary coalgebra and f, g, h ∈ HomK (C, A).
Ternary convolution product is
[f, g, h]∗ = µ3 ◦ (f ⊗ g ⊗ h) ◦ ∆3
or in the Sweedler notation [f, g, h]∗ (a) = f a(1) g a(2) h a(3) .
(12.13)
Definition 12-4. A ternary coalgebra is called derived, if there exists a binary (usual, see e.g.
S WEEDLER [1969]) coalgebra ∆2 : C → C ⊗ C such that
∆3,der = (id ⊗∆2 ) ⊗ ∆2 .
(3)
Definition 12-5. A ternary
bialgebra B is B, µ3 , η (3) , ∆3 , ε
for which B, µ3 , η
(3)
algebra and B, ∆3 , ε
is a ternary coalgebra and they are compatible
where τ12
= 12
.
21
∆3 ◦ µ3 = µ3 ◦ (id ⊗τ12 ⊗ id) ◦ ∆3 ,
(12.14)
(3)
is a ternary
(12.15)
One can distinguish four kinds of ternary bialgebras with respect to a “being derived” property:
(1) A ∆-derived ternary bialgebra
∆3 = ∆3,der = (id ⊗∆2 ) ◦ ∆2
(12.16)
µ3,der = µ2 ◦ (µ2 ⊗ id)
(12.17)
(2) A µ-derived ternary bialgebra
42
STEVEN DUPLIJ
(3) A derived ternary bialgebra is simultaneously µ-derived and ∆-derived ternary bialgebra.
(4) A non-derived ternary bialgebra which does not satisfy (12.16) and (12.17).
Possible types of ternary antipodes can be defined by analogy with binary coalgebras.
Definition 12-6. A skew ternary antipode is
(3)
(3)
(3)
µ3 ◦ (Sskew ⊗ id ⊗ id) ◦ ∆3 = µ3 ◦ (id ⊗Sskew ⊗ id) ◦ ∆3 = µ3 ◦ (id ⊗ id ⊗Sskew ) ◦ ∆3 = id . (12.18)
If only one equality from (12.18) is satisfied, the corresponding skew antipode is called left, middle
or right.
Definition 12-7. Strong ternary antipode is
(3)
(3)
(µ2 ⊗ id) ◦ (id ⊗Sstrong ⊗ id) ◦ ∆3 = 1 ⊗ id, (id ⊗µ2 ) ◦ (id ⊗ id ⊗Sstrong ) ◦ ∆3 = id ⊗1,
where 1 is a unit of algebra.
If in a ternary coalgebra the relation
holds true, where τ13 =
123
321
∆3 ◦ S = τ13 ◦ (S ⊗ S ⊗ S) ◦ ∆3
(12.19)
, then it is called skew-involutive.
Definition 12-8. A ternary Hopf algebra H, µ3 , η (3) , ∆3 , ε(3) , S (3) is a ternary bialgebra with a
ternary antipode S (3) of the corresponding above type .
Let us consider concrete constructions of ternary comultiplications, bialgebras and Hopf algebras.
A ternary group-like element can be defined by ∆3 (g) = g ⊗ g ⊗ g, and for 3 such elements we have
∆3 ([g1 g2 g3 ]) = ∆3 (g1 ) ∆3 (g2 ) ∆3 (g3 ) .
(12.20)
But an analog of the binary primitive element (satisfying ∆2 (x) = x ⊗ 1 + 1 ⊗ x) cannot be chosen
simply as ∆3 (x) = x ⊗ e ⊗ e + e ⊗ x ⊗ e + e ⊗ e ⊗ x, since the algebra structure is not preserved.
Nevertheless, if we introduce two idempotent units e1 , e2 satisfying “semiorthogonality” [e1 e1 e2 ] = 0,
[e2 e2 e1 ] = 0, then
∆3 (x) = x ⊗ e1 ⊗ e2 + e2 ⊗ x ⊗ e1 + e1 ⊗ e2 ⊗ x
(12.21)
and now ∆3 ([x1 x2 x3 ]) = [∆3 (x1 ) ∆3 (x2 ) ∆3 (x3 )]. Using (12.21) ε (x) = 0, ε (e1,2 ) = 1, and
S (3) (x) = −x, S (3) (e1,2 ) = e1,2 , one can construct a ternary universal enveloping algebra in full
analogy with the binary case (see e.g. K ASSEL [1995]).
One of the most important examples of noncocommutative Hopf algebras is the well known
Sweedler Hopf algebra S WEEDLER [1969] which in the binary case has two generators x and y
satisfying
µ2 (x, x) = 1,
(12.22)
µ2 (y, y) = 0,
(12.23)
(2)
(2)
σ+ (xy) = −σ− (xy) .
(12.24)
∆2 (x) = x ⊗ x,
(12.25)
It has the following comultiplication
∆2 (y) = y ⊗ x + 1 ⊗ y,
(12.26)
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
43
counit ε(2) (x) = 1, ε(2) (y) = 0, and antipode S (2) (x) = x, S (2) (y) = −y, which respect the algebra
structure. In the derived case a ternary Sweedler algebra is generated also by two generators x and y
obeying D UPLIJ [2001]
µ3 (x, e, x) = µ3 (e, x, x) = µ3 (x, x, e) = e,
(3)
σ+ ([yey]) = 0,
(3)
σ+
([xey]) =
(3)
−σ−
(12.27)
(12.28)
([xey]) .
(12.29)
The derived Hopf algebra structure is given by
∆3 (x) = x ⊗ x ⊗ x,
(12.30)
∆3 (y) = y ⊗ x ⊗ x + e ⊗ y ⊗ x + e ⊗ e ⊗ y,
(12.31)
ε(3) (x) = ε(2) (x) = 1,
(12.32)
ε(3) (y) = ε(2) (y) = 0,
(12.33)
S (3) (x) = S (2) (x) = x,
(12.34)
S
(3)
(y) = S
(2)
(y) = −y,
(12.35)
and it can be checked that (12.30)-(12.32) are algebra maps, while (12.34) are antialgebra maps. To
obtain a non-derived ternary Sweedler example we have the following possibilities: 1) one “even”
generator x, two “odd” generators y1,2 and one ternary unit e; 2) two “even” generators x1,2 , one
“odd” generator y and two ternary units e1,2 . In the first case the ternary algebra structure is (no
summation, i = 1, 2)
[xxx] = e,
(12.36)
[yi yi yi ] = 0,
(12.37)
(3)
(3)
σ+ ([yi xyi ]) = σ+ ([xyi x]) = 0,
[xeyi ] = − [xyi e] ,
[exyi ] = − [yi xe] ,
(12.38)
(12.39)
[eyi x] = − [yi ex] ,
(12.40)
σ+ ([y1 xy2 ]) = −σ− ([y1 xy2 ]) .
(12.41)
(3)
(3)
The corresponding ternary Hopf algebra structure is
∆3 (x) = x ⊗ x ⊗ x, ∆3 (y1,2 ) = y1,2 ⊗ x ⊗ x + e1,2 ⊗ y2,1 ⊗ x + e1,2 ⊗ e2,1 ⊗ y2,1 ,
(12.42)
ε(3) (x) = 1, ε(3) (yi ) = 0,
(12.43)
S (3) (x) = x,
(12.44)
S (3) (yi ) = −yi .
In the second case we have for the algebra structure
[xi xj xk ] = δij δik δjk ei , [yyy] = 0,
(12.45)
σ+ ([yxi y]) = 0, σ+ ([xi yxi ]) = 0,
(3)
(12.46)
(3)
(12.47)
(3)
σ+ ([y1 xy2 ]) = 0,
(3)
σ− ([y1 xy2 ]) = 0,
44
STEVEN DUPLIJ
and the ternary Hopf algebra structure is D UPLIJ [2001]
∆3 (xi ) = xi ⊗ xi ⊗ xi ,
∆3 (y) = y ⊗ x1 ⊗ x1 + e1 ⊗ y ⊗ x2 + e1 ⊗ e2 ⊗ y,
(12.48)
ε(3) (xi ) = 1,
(12.49)
ε(3) (y) = 0,
(12.50)
S (3) (xi ) = xi ,
(12.51)
S (3) (y) = −y.
13. T ERNARY
(12.52)
QUANTUM GROUPS
A ternary commutator can be obtained in different ways B REMNER AND H ENTZEL [2000].
We will consider the simplest version called a Nambu bracket (see e.g. TAKHTAJAN [1994],
(3)
DE A ZCARRAGA AND I ZQUIERDO [2010]). Let us introduce two maps ω± : A⊗A⊗A → A⊗A⊗A
by
ω+ (a ⊗ b ⊗ c) = a ⊗ b ⊗ c + b ⊗ c ⊗ a + c ⊗ a ⊗ b,
(3)
(13.1)
ω− (a ⊗ b ⊗ c) = b ⊗ a ⊗ c + c ⊗ b ⊗ a + a ⊗ c ⊗ b.
(3)
(13.2)
(3)
(3)
(3)
Thus, obviously µ3 ◦ ω± = σ± ◦ µ3 , where σ± ∈ S3 denotes a sum of terms having even and
(2)
(2)
odd permutations respectively. In the binary case ω+ = id ⊗ id and ω− = τ is the twist operator
(2)
(2)
τ : a ⊗ b → b ⊗ a, while µ2 ◦ ω− is permutation σ− (ab) = ba. So the Nambu product is
(3)
(3)
(3)
(3)
(3)
(3)
ωN = ω+ − ω− , and the ternary commutator is [, , ]N = σN = σ+ − σ− , or TAKHTAJAN [1994]
[a, b, c]N = [abc] + [bca] + [cab] − [cba] − [acb] − [bac]
(13.3)
An abelian ternary algebra is defined by the vanishing of the Nambu bracket [a, b, c]N = 0 or
(3)
(3)
ternary commutation relation σ+ = σ− . By analogy with the binary case a deformed ternary
algebra can be defined by
(3)
(3)
σ+ = qσ− or [abc] + [bca] + [cab] = q ([cba] + [acb] + [bac]) ,
(13.4)
where multiplication by q is treated as an external operation.
Let us consider a ternary analog of the Woronowicz example of a bialgebra construction, which has
(2)
(2)
two generators satisfying xy = qyx (or σ+ (xy) = qσ− (xy)), then the following coproducts
∆2 (x) = x ⊗ x
∆2 (y) = y ⊗ x + 1 ⊗ y
(13.5)
(13.6)
are the algebra maps. In the derived ternary case using (13.4) we have
(3)
(3)
σ+ ([xey]) = qσ− ([xey]) ,
(13.7)
where e is the ternary unit and ternary coproducts are
∆3 (e) = e ⊗ e ⊗ e,
∆3 (x) = x ⊗ x ⊗ x,
∆3 (y) = y ⊗ x ⊗ x + e ⊗ y ⊗ x + e ⊗ e ⊗ y,
(13.8)
(13.9)
(13.10)
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
45
which are ternary algebra maps, i.e. they satisfy
(3)
(3)
σ+ ([∆3 (x) ∆3 (e) ∆3 (y)]) = qσ− ([∆3 (x) ∆3 (e) ∆3 (y)]) .
(13.11)
Let us consider the group G = SL (n, K). Then the algebra generated by aij ∈ SL (n, K) can be
endowed with the structure of a ternary Hopf algebra (see, e.g., M ADORE [1995] for the binary case)
by choosing the ternary coproduct, counit and antipode as (here summation is implied)
i
∆3 aij = aik ⊗ akl ⊗ alj , ε aij = δji , S (3) aij = a−1 j .
(13.12)
This antipode is a skew one since from (12.18) it follows that
i
µ3 ◦ (S (3) ⊗ id ⊗ id) ◦ ∆3 aij = S (3) aik akl alj = a−1 k akl alj = δli alj = aij .
This ternary Hopf algebra is derived since for ∆2 = aij ⊗ ajk we have
∆3 = (id ⊗∆2 ) ⊗ ∆2 aij = (id ⊗∆2 ) aik ⊗ akj = aik ⊗ ∆2 akj = aik ⊗ akl ⊗ alj .
(13.13)
(13.14)
In the most important case n = 2 we can obtain the manifest action of the ternary coproduct ∆3 on
components. Possible non-derived matrix representations of the ternary product
ij can be done only by
four-rank n × n × n × n twice covariant and twice contravariant tensors akl . Among all products
ij ol ko
jl ko
the non-derived ones are only the following: aoi
jk boo cil and aok bio cil (where o is any index). So using
e.g. the first choice we can define the non-derived Hopf algebra structure by
iµ
vσ
ρj
∆3 aij
(13.15)
kl = avρ ⊗ akl ⊗ aµσ ,
1 i j
i j
(13.16)
δ
δ
+
δ
δ
=
ε aij
k
l
l
k ,
kl
2
iµ vσ
i µ σ
(3)
and the skew antipode sij
aij
kl = S
kl which is a solution of the equation svρ akl = δρ δk δl .
Next consider ternary dual pair kG (push-forward) and F (G) (pull-back) which are related by
(kG)∗ ∼
= F (G) (see, e.g., KOGORODSKI AND S OIBELMAN [1998]). Here kG = span (G) is a
(3)
ternary group algebra (G has a ternary product [ , , ]G or µG ) over a field k.
If u ∈ kG (u = ui xi , xi ∈ G), then
[uvw]k = ui v j w l [xi xj xl ]G
(13.17)
is associative, and so (kG, [ , , ]k ) becomes a ternary algebra. Define a ternary coproduct ∆3 : kG →
kG ⊗ kG ⊗ kG by
∆3 (u) = ui xi ⊗ xi ⊗ xi
(13.18)
(derived and associative), then ∆3 ([uvw]k ) = [∆3 (u) ∆3 (v) ∆3 (w)]k , and kG is a ternary bialgebra.
(3)
If we define a ternary antipode by Sk = ui x̄i , where x̄i is a skew element of xi , then kG becomes a
ternary Hopf algebra.
In the dual case of functions F (G) : {ϕ : G → k} a ternary product [ , , ]F or µ3,F (derived and
associative) acts on ψ (x, y, z) as
(µ3,F ψ) (x) = ψ (x, x, x) ,
(13.19)
and so F (G) is a ternary algebra. Let F (G) ⊗ F (G) ⊗ F (G) ∼
= F (G × G × G), then we define a
ternary coproduct ∆3 : F (G) → F (G) ⊗ F (G) ⊗ F (G) as
(∆3 ϕ) (x, y, z) = ϕ ([xyz]F ) ,
(13.20)
which is derive and associative. Thus we can obtain ∆3 ([ϕ1 ϕ2 ϕ3 ]F ) = [∆3 (ϕ1 ) ∆3 (ϕ2 ) ∆3 (ϕ3 )]F ,
and therefore F (G) is a ternary bialgebra. If we define a ternary antipode by
(3)
SF (ϕ) = ϕ (x̄) ,
(13.21)
46
STEVEN DUPLIJ
where x̄ is a skew element of x, then F (G) becomes a ternary Hopf algebra.
Let us introduce a ternary analog of the R-matrix D UPLIJ [2001]. For a ternary Hopf algebra H
we consider a linear map R(3) : H ⊗ H ⊗ H → H ⊗ H ⊗ H.
Definition 13-1. A ternary Hopf algebra H, µ3 , η (3) , ∆3 , ε(3) , S (3) is called quasifiveangular4 if it
satisfies
(3)
(3)
(3)
(13.22)
(id ⊗∆3 ⊗ id) = R125 R145 R135 ,
(3)
(3)
(3)
(13.23)
(id ⊗ id ⊗∆3 ) = R125 R124 R123 ,
(3)
(3)
(3)
(13.24)
(∆3 ⊗ id ⊗ id) = R145 R245 R345 ,
where as usual the index of R denotes action component positions.
Using the standard procedure (see, e.g., K ASSEL [1995], C HARI AND P RESSLEY [1996], M AJID
[1995]) we obtain a set of abstract ternary quantum Yang-Baxter equations, one of which has the form
D UPLIJ [2001]
(3)
(3)
(3)
(3)
(3)
(3)
(3)
(3)
(3)
(3)
R243 R342 R125 R145 R135 = R123 R132 R145 R245 R345 ,
(13.25)
and others can be obtained by corresponding permutations. The classical ternary Yang-Baxter equations form a one parameter family of solutions R (t) can be obtained by the expansion
R(3) (t) = e ⊗ e ⊗ e + rt + O t2 ,
(13.26)
where r is a ternary classical R-matrix, then e.g. for (13.25) we have
r342 r125 r145 r135 + r243 r125 r145 r135 + r243 r342 r145 r135 + r243 r342 r125 r135 + r243 r342 r125 r145
= r132 r145 r245 r345 + r123 r145 r245 r345 + r123 r132 r245 r345 + r123 r132 r145 r345 + r123 r132 r145 r245 .
(3)
(3)
(3)
(3)
(3)
For three ternary Hopf algebras HA,B,C , µA,B,C , ηA,B,C , ∆A,B,C , εA,B,C , SA,B,C we can introduce
a non-degenerate ternary “pairing” (see, e.g., C HARI AND P RESSLEY [1996] for the binary case)
h , , i(3) : HA × HB × HC → K, trilinear over K, satisfying D UPLIJ [2001]
D
E(3) D
E(3) D
E(3) D
E(3)
(3)
(3)
(3)
(3)
ηA (a) , b, c
= a, εB (b) , c
, a, ηB (b) , c
= εA (a) , b, c
,
D
E(3) D
E(3) D
E(3) D
E(3)
(3)
(3)
(3)
(3)
b, ηB (b) , c
= a, b, εC (c)
, a, b, ηC (c)
= a, εB (b) , c
,
D
E(3) D
E(3) D
E(3) D
E(3)
(3)
(3)
(3)
(3)
a, b, ηC (c)
= εA (a) , b, c
, ηA (a) , b, c
= a, b, εC (c)
,
4
The reason for such notation is clear from (13.25).
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
D
(3)
µA (a1 ⊗ a2 ⊗ a3 ) , b, c
D
E(3)
E(3)
(3)
∆A (a) , b1 ⊗ b2 ⊗ b3 , c
D
E(3)
(3)
a, µB (b1 ⊗ b2 ⊗ b3 ) , c
E(3)
D
(3)
a, ∆B (b) , c1 ⊗ c2 ⊗ c3
D
E(3)
(3)
a, b, µC (c1 ⊗ c2 ⊗ c3 )
D
E(3)
(3)
a1 ⊗ a2 ⊗ a3 , b, ∆C (c)
D
(3)
SA
(a) , b, c
E(3)
=
D
47
D
E(3)
(3)
= a1 ⊗ a2 ⊗ a3 , ∆B (b) , c
,
D
E(3)
(3)
= a, µB (b1 ⊗ b2 ⊗ b3 ) , c
,
D
E(3)
(3)
= a, b1 ⊗ b2 ⊗ b3 , ∆C (c)
,
E(3)
D
(3)
,
= a, b, µC (c1 ⊗ c2 ⊗ c3 )
D
E(3)
(3)
= ∆A (a) , b, c1 ⊗ c2 ⊗ c3
,
D
E(3)
(3)
= µA (a1 ⊗ a2 ⊗ a3 ) , b, c
,
(3)
a, SB
(b) , c
E(3)
=
D
(3)
a, b, SC
(c)
E(3)
,
where a, ai ∈ HA , b, bi ∈ HB . The ternary “paring” between HA ⊗ HA ⊗ HA and HB ⊗ HB ⊗ HB
is given by ha1 ⊗ a2 ⊗ a3 , b1 ⊗ b2 ⊗ b3 i(3) = ha1 , b1 i(3) ha2 , b2 i(3) ha3 , b3 i(3) . These constructions can
naturally lead to ternary generalizations of the duality concept and the quantum double which are key
ingredients in the theory of quantum groups D RINFELD [1987], K ASSEL [1995], M AJID [1995].
14. C ONCLUSIONS
In this paper we presented a review of polyadic systems and their representations, ternary algebras
and Hopf algebras. We have classified general polyadic systems and considered their homomorphisms
and their multiplace generalizations, paying attention to their associativity. Then, we defined multiplace representations and multiactions and have given examples of matrix representations for some
ternary groups. We defined and investigated ternary algebras and Hopf algebras, and have given some
examples. We then considered some ternary generalizations of quantum groups and the Yang-Baxter
equation.
R EFERENCES
A BE , E. (1980). Hopf Algebras. Cambridge: Cambridge Univ. Press.
A BRAMOV, V. (1995). Z3 -graded analogues of Clifford algebras and algebra of Z3 -graded symmetries. Algebras Groups Geom. 12 (3), 201–221.
A BRAMOV, V. (1996). Ternary generalizations of Grassmann algebra. Proc. Est. Acad. Sci., Phys. Math. 45
(2-3), 152–160.
A BRAMOV, V., R. K ERNER , AND B. L E ROY (1997). Hypersymmetry: a Z3 -graded generalization of supersymmetry. J. Math. Phys. 38, 1650–1669.
BAGGER , J. AND N. L AMBERT (2008a). Comments on multiple M2-branes. J. High Energy Phys. 2, 105.
BAGGER , J. AND N. L AMBERT (2008b). Gauge symmetry and supersymmetry of multiple M2-branes. Phys.
Rev. D77, 065008.
B ELOUSOV, V. D. (1972). n-ary Quasigroups. Kishinev: Shtintsa.
B EREZIN , F. A. (1987). Introduction to Superanalysis. Dordrecht: Reidel.
B ERGMAN , G. M. (1995). An invitation to general algebra and universal constructions. Berkeley: University
of California.
B OROWIEC , A., W. D UDEK , AND S. D UPLIJ (2001). Basic concepts of ternary Hopf algebras. J. Kharkov
National Univ., ser. Nuclei, Particles and Fields 529 (3(15)), 21–29. math.QA/0306208.
48
STEVEN DUPLIJ
B OROWIEC , A., W. D UDEK , AND S. D UPLIJ (2006). Bi-element representations of ternary groups. Comm.
Algebra 34 (5), 1651–1670.
B OURBAKI , N. (1998). Elements of Mathematics: Algebra 1. Springer.
B RANDT, H. (1927). Über eine Verallgemeinerung des Gruppenbegriffes. Math. Annalen 96, 360–367.
B REMNER , M. AND I. H ENTZEL (2000). Identities for generalized lie and jordan products on totally associative triple systems. J. Algebra 231 (1), 387–405.
B RUCK , R. H. (1966). A Survey on Binary Systems. New York: Springer-Verlag.
C ARLSSON , R. (1980). N -ary algebras. Nagoya Math. J. 78 (1), 45–56.
C ELAKOSKI , N. (1977). On some axiom systems for n-groups. Mat. Bilt. 1, 5–14.
C HARI , V. AND A. P RESSLEY (1996). A Guide to Quantum Groups. Cambridge: Cambridge University Press.
C HRONOWSKI , A. (1994a). On ternary semigroups of homomorphisms of ordered sets. Arch. Math. (Brno) 30
(2), 85–95.
C HRONOWSKI , A. (1994b). A ternary semigroup of mappings. Demonstr. Math. 27 (3-4), 781–791.
C HRONOWSKI , A. AND M. N OVOTN Ý (1995). Ternary semigroups of morphisms of objects in categories.
Archivum Math. (Brno) 31, 147–153.
C HUNG , K. O. AND J. D. H. S MITH (2008). Weak homomorphisms and graph coalgebras. Arab. J. Sci. Eng.,
Sect. C, Theme Issues 33 (2), 107–121.
C LIFFORD , A. H. AND G. B. P RESTON (1961). The Algebraic Theory of Semigroups, Volume 1. Providence:
Amer. Math. Soc.
C OLLINS , M. J. (1990). Representations And Characters Of Finite Groups. Cambridge: Cambridge University
Press.
C ORNWELL , J. F. (1997). Group Theory in Physics: An Introduction. London: Academic Press.
Č UPONA , G. AND B. T RPENOVSKI (1961). Finitary associative operations with neutral elements. Bull. Soc.
Math. Phys. Macedoine 12, 15–24.
C URTIS , C. W. AND I. R EINER (1962). Representation theory of finite groups and associative algebras.
Providence: AMS.
C Z ÁK ÁNY, B. (1962). On the equivalence of certain classes of algebraic systems. Acta Sci. Math. (Szeged) 23,
46–57.
DE A ZCARRAGA , J. A. AND J. M. I ZQUIERDO (2010). n-Ary algebras: A review with applications. J.
Phys. A43, 293001.
D ENECKE , K. AND K. S AENGSURA (2008). State-based systems and (F1 , F2 )-coalgebras. East-West J.
Math., Spec. Vol., 1–31.
D ENECKE , K. AND S. L. W ISMATH (2009). Universal Algebra and Coalgebra. Singapore: World Scientific.
D ÖRNTE , W. (1929). Unterschungen über einen verallgemeinerten Gruppenbegriff. Math. Z. 29, 1–19.
D RINFELD , V. G. (1987). Quantum groups. In: A. G LEASON (Ed.), Proceedings of the ICM, Berkeley, Phode
Island: AMS, pp. 798–820.
D UDEK , W. A. (1980). Remarks on n-groups. Demonstratio Math. 13, 165–181.
D UDEK , W. A. (1993). Autodistributive n-groups. Annales Sci. Math. Polonae, Commentationes Math. 23,
1–11.
D UDEK , W. A. (2007). Remarks to Głazek’s results on n-ary groups. Discuss. Math., Gen. Algebra Appl. 27
(2), 199–233.
D UDEK , W. A., K. G ŁAZEK , AND B. G LEICHGEWICHT (1977). A note on the axioms of n-groups. In: Coll.
Math. Soc. J. Bolyai. 29. Universal Algebra, Esztergom (Hungary), pp. 195–202.
D UDEK , W. A. AND J. M ICHALSKI (1984). On retracts of polyadic groups. Demonstratio Math. 17, 281–301.
D UDEK , W. A. AND M. S HAHRYARI (2012). Representation theory of polyadic groups. Algebr. Represent.
Theory 15 (1), 29–51.
D UPLIJ , S. (2001). Ternary Hopf algebras. In: A. G. N IKITIN AND V. M. B OYKO (Eds.), Symmetry in
Nonlinear Mathematical Physics, Kiev: Institute of Mathematics, pp. 25–34.
D UPLIJ , S. AND F. L I (2001). Regular solutions of quantum Yang-Baxter equation from weak Hopf algebras.
Czech. J. Phys. 51 (12), 1306–1311.
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
49
D UPLIJ , S. AND S. S INEL’ SHCHIKOV (2009). Quantum enveloping algebras with von Neumann regular
Cartan-like generators and the Pierce decomposition. Commun. Math. Phys. 287 (1), 769–785.
D UPLIJ , S. AND S. S INEL’ SHCHIKOV (2010). Classification of Uq ( SL 2 )-module algebra structures on the
quantum plane. J. Math. Physics, Analysis, Geometry 6 (6), 21–46.
E LLERMAN , D. (2006). A theory of adjoint functors – with some thoughts about their philosophical significance. In: G. S ICA (Ed.), What is category theory?, Monza (Milan): Polimetrica, pp. 127–183.
E LLERMAN , D. (2007). Adjoints and emergence: applications of a new theory of adjoint functors. Axiomathes 17 (1), 19–39.
E VANS , T. (1963). Abstract mean values. Duke Math J. 30, 331–347.
F ILIPPOV, V. T. (1985). n-Lie algebras. Sib. Math. J. 26, 879–891.
F UJIWARA , T. (1959). On mappings between algebraic systems. Osaka Math. J. 11, 153–172.
F ULTON , W. AND J. H ARRIS (1991). Representation Theory: a First Course. N. Y.: Springer.
G AL’ MAK , A. M. (1986). Translations of n-ary groups. Dokl. Akad. Nauk BSSR 30, 677–680.
G AL’ MAK , A. M. (1998). Generalized morphisms of algebraic systems. Vopr. Algebry 12, 36–46.
G AL’ MAK , A. M. (2001a). Generalized morphisms of Abelian m-ary groups. Discuss. Math. 21 (1), 47–55.
G AL’ MAK , A. M. (2001b). Polyadic analogs of the Cayley and Birkhoff theorems. Russ. Math. 45 (2), 10–15.
G AL’ MAK , A. M. (2003). n-Ary Groups, Part 1. Gomel: Gomel University.
G AL’ MAK , A. M. (2007). n-Ary Groups, Part 2. Minsk: Belarus State University.
G AL’ MAK , A. M. (2008). Polyadic associative operations on Cartesian powers. Proc. of the Natl. Academy of
Sciences of Belarus, Ser. Phys.-Math. Sci. 3, 28–34.
G EORGI , H. (1999). Lie Algebras in Particle Physics. New York: Perseus Books.
G ŁAZEK , K. (1980). Weak homomorphisms of general algebras and related topics. Math. Semin. Notes, Kobe
Univ. 8, 1–36.
G ŁAZEK , K. AND B. G LEICHGEWICHT (1982a). Abelian n-groups. In: Colloq. Math. Soc. J. Bolyai. Universal Algebra (Esztergom, 1977), Amsterdam: North-Holland, pp. 321–329.
G ŁAZEK , K. AND B. G LEICHGEWICHT (1982b). Bibliography of n-groups (polyadic groups) and some group
like n-ary systems. In: Proceedings of the Symposium on n-ary structures, Skopje: Macedonian Academy
of Sciences and Arts, pp. 253–289.
G ŁAZEK , K. AND J. M ICHALSKI (1974). On weak homomorphisms of general non-indexed algebras. Bull.
Acad. Pol. Sci., Sér. Sci. Math. Astron. Phys. 22, 651–656.
G ŁAZEK , K. AND J. M ICHALSKI (1977). Weak homomorphisms of general algebras. Commentat. Math. 19
(1976), 211–228.
G LEICHGEWICHT, B. AND K. G ŁAZEK (1967). Remarks on n-groups as abstract algebras. Colloq. Math. 17,
209–219.
G LEICHGEWICHT, B., M. B. WANKE -J AKUBOWSKA , AND M. E. WANKE -J ERIE (1983). On representations
of cyclic n-groups. Demonstr. Math. 16, 357–365.
G LUSKIN , L. M. (1965). Position operatives. Math. Sbornik 68(110) (3), 444–472.
G OETZ , A. (1966). On weak isomorphisms and weak homomorphisms of abstract algebras. Colloq. Math. 14,
163–167.
G USTAVSSON , A. (2009). One-loop corrections to Bagger-Lambert theory. Nucl. Phys. B807, 315–333.
H ALA Š , R. (1994). A note on homotopy in universal algebra. Acta Univ. Palacki. Olomuc., Fac. Rerum Nat.,
Math. 33 (114), 39–42.
H AUSMANN , B. A. AND Ø. O RE (1937). Theory of quasigroups. Amer. J. Math. 59, 983–1004.
H EINE , E. (1878). Handbuch der Kugelfunktionan. Berlin: Reimer.
H O , P.-M., R.-C. H OU , Y. M ATSUO , AND S. S HIBA (2008). M5-brane in three-form flux and multiple
M2-branes. J. High Energy Phys. 08, 014.
H OBBY, D. AND R. M C K ENZIE (1988). The Structure of Finite Algebras. Providence: AMS.
H OSSZ Ú , L. M. (1963). On the explicit form of n-group operation. Publ. Math. Debrecen 10, 88–92.
H USEM ÖLLER , D., M. J OACHIM , B. J UR ČO , AND M. S CHOTTENLOHER (2008). G-spaces, G-bundles,
50
STEVEN DUPLIJ
and G-vector bundles. In: D. H USEM ÖLLER AND S. E CHTERHOFF (Eds.), Basic Bundle Theory and KCohomology Invariants, Berlin-Heidelberg: Springer, pp. 149–161.
K AC , V. AND P. C HEUNG (2002). Quantum calculus. New York: Springer.
K APRANOV, M., I. M. G ELFAND , AND A. Z ELEVINSKII (1994). Discriminants, Resultants and Multidimensional Determinants. Berlin: Birkhäuser.
K ASSEL , C. (1995). Quantum Groups. New York: Springer-Verlag.
K AWAMURA , Y. (2003). Cubic matrices, generalized spin algebra and uncertainty relation. Progr. Theor.
Phys. 110, 579–587.
K ERNER , R. (2000). Ternary algebraic structures and their applications in physics, preprint Univ. P. & M.
Curie, Paris, 15 p.
K IRILLOV, A. A. (1976). Elements of the Theory of Representations. Berlin: Springer-Verlag.
KOGORODSKI , L. I. AND Y. S. S OIBELMAN (1998). Algebras of Functions on Quantum Groups. Providence:
AMS.
KOLIBIAR , M. (1984). Weak homomorphisms in some classes of algebras. Stud. Sci. Math. Hung. 19, 413–
420.
L I , F. AND S. D UPLIJ (2002). Weak Hopf algebras and singular solutions of quantum Yang-Baxter equation.
Commun. Math. Phys. 225 (1), 191–217.
L ÕHMUS , J., E. PAAL , AND L. S ORGSEPP (1994). Nonassociative Algebras in Physics. Palm Harbor:
Hadronic Press.
L OUNESTO , P. AND R. A BLAMOWICZ (2004). Clifford Algebras: Applications To Mathematics, Physics, And
Engineering. Birkhäuser.
L OW, A. M. (2010). Worldvolume superalgebra of BLG theory with Nambu-Poisson structure. J. High Energy
Phys. 04, 089.
M ADORE , J. (1995). Introduction to Noncommutative Geometry and its Applications. Cambridge: Cambridge
University Press.
M AJID , S. (1995). Foundations of Quantum Group Theory. Cambridge: Cambridge University Press.
M AL’ TCEV, A. I. (1954). On general theory of algebraic systems. Mat. Sb. 35 (1), 3–20.
M AL’ TCEV, A. I. (1957). Free topological algebras. Izv. Akad. Nauk SSSR. Ser. Mat. 21, 171–198.
M AL’ TSEV, A. I. (1958). The structural characteristic of some classes of algebras. Dokl. Akad. Nauk SSSR 120,
29–32.
M ARCZEWSKI , E. (1966). Independence in abstract algebras. Results and problems. Colloq. Math. 14, 169–
188.
M ICHALSKI , J. (1981). Covering k-groups of n-groups. Arch. Math. (Brno) 17 (4), 207–226.
M ICHOR , P. W. AND A. M. V INOGRADOV (1996). n-ary Lie and associative algebras. Rend. Sem. Mat. Univ.
Pol. Torino 54 (4), 373–392.
M ONTGOMERY, S. (1993). Hopf algebras and their actions on rings. Providence: AMS.
NAMBU , Y. (1973). Generalized Hamiltonian dynamics. Phys. Rev. 7, 2405–2412.
N OVOTN Ý , M. (1990). Mono-unary algebras in the work of Czechoslovak mathematicians. Arch. Math.,
Brno 26 (2-3), 155–164.
N OVOTN Ý , M. (2002). Homomorphisms of algebras. Czech. Math. J. 52 (2), 345–364.
P ÉCSI , B. (2011). On Morita contexts in bicategories. Applied Categorical Structures, 1–18.
P ETRESCU , A. (1977). On the homotopy of universal algebras. I. Rev. Roum. Math. Pures Appl. 22, 541–551.
P OP, M. S. AND A. P OP (2004). On some relations on n-monoids. Carpathian J. Math. 20 (1), 87–94.
P OST, E. L. (1940). Polyadic groups. Trans. Amer. Math. Soc. 48, 208–350.
P R ÜFER , H. (1924). Theorie der abelshen Gruppen I. Grundeigenschaften. Math. Z. 20, 165–187.
R AUSCH DE T RAUBENBERG , M. (2008). Cubic extentions of the poincare algebra. Phys. Atom. Nucl. 71,
1102–1108.
RUSAKOV, S. A. (1979). A definition of an n-ary group. Dokl. Akad. Nauk BSSR 23, 965–967.
RUSAKOV, S. A. (1998). Some Applications of n-ary Group Theory. Minsk: Belaruskaya navuka.
S HNIDER , S. AND S. S TERNBERG (1993). Quantum Groups. Boston: International Press.
POLYADIC SYSTEMS, REPRESENTATIONS AND QUANTUM GROUPS
51
S OKHATSKY, F. M. (1997). On the associativity of multiplace operations. Quasigroups Relat. Syst. 4, 51–66.
S OKOLOV, E. I. (1976). On the theorem of Gluskin-Hosszú on Dörnte groups. Mat. Issled. 39, 187–189.
S OKOLOV, N. P. (1972). Introduction to the Theory of Multidimensional Matrices. Kiev: Naukova Dumka.
S TOJAKOVI Ć , Z. AND W. A. D UDEK (1986). On σ-permutable n-groups. Publ. Inst. Math., Nouv. Sér. 40(54),
49–55.
S WEEDLER , M. E. (1969). Hopf Algebras. New York: Benjamin.
TAKHTAJAN , L. (1994). On foundation of the generalized Nambu mechanics. Commun. Math. Phys. 160,
295–315.
T HURSTON , H. A. (1949). Partly associative operations. J. London Math. Soc. 24, 260–271.
T IMM , J. (1972). Verbandstheoretische Behandlung n-stelliger Gruppen. Abh. Math. Semin. Univ. Hamb. 37,
218–224.
T RACZYK , T. (1965). Weak isomorphisms of Boolean and Post algebras. Colloq. Math. 13, 159–164.
Ǔ SAN , J. (2003). n-groups in the light of the neutral operations. Math. Moravica Special vol.
V IDAL , J. C. AND J. S. T UR (2010). A 2-categorial generalization of the concept of institution. Stud. Log. 95
(3), 301–344.
WANKE -J AKUBOWSKA , M. B. AND M. E. WANKE -J ERIE (1984). On representations of n-groups. Annales
Sci. Math. Polonae, Commentationes Math. 24, 335–341.
W EYL , H. (1946). Classical Groups, Their Invariants and Representations. Princeton: Princeton Univ. Press.
Y UREVYCH , O. V. (2001). Criteria for invertibility of elements in associates. Ukr. Math. J. 53 (11), 1895–1905.
Z EKOVI Ć , B. AND V. A. A RTAMONOV (1999). A connection between some properties of n-group rings and
group rings. Math. Montisnigri 15, 151–158.
Z EKOVI Ć , B. AND V. A. A RTAMONOV (2002). On two problems for n-group rings. Math. Montisnigri 15,
79–85.
C ENTER FOR M ATHEMATICS , S CIENCE AND E DUCATION RUTGERS U NIVERSITY, 118 F RELINGHUYSEN R D .,
P ISCATAWAY, NJ 08854-8019
Current address: V.N. Karazin Kharkov National University, Svoboda Sq. 4, Kharkov 61022, Ukraine
E-mail address: [email protected], [email protected]
URL: http://homepages.spa.umn.edu/~duplij
| 4math.GR
|
Behavior-based Navigation of Mobile Robot in Unknown
Environments Using Fuzzy Logic and Multi-Objective
Optimization
Thi Thanh Van Nguyen, Manh Duong Phung and Quang Vinh Tran
University of Engineering and Technology - Vietnam National University, Hanoi
[email protected]
Abstract
This study proposes behavior-based navigation architecture, named BBFM, to deal
with the problem of navigating the mobile robot in unknown environments in the presence
of obstacles and local minimum regions. In the architecture, the complex navigation task
is split into principal sub-tasks or behaviors. Each behavior is implemented by a fuzzy
controller and executed independently to deal with a specific problem of navigation. The
fuzzy controller is modified to contain only the fuzzification and inference procedures so
that its output is a membership function representing the behavior’s objective. The
membership functions of all controllers are then used as the objective functions for a
multi-objective optimization process to coordinate all behaviors. The result of this
process is an overall control signal, which is Pareto-optimal, used to control the robot. A
number of simulations, comparisons, and experiments were conducted. The results show
that the proposed architecture outperforms some popular behaviorbased architectures in
term of accuracy, smoothness, traveled distance, and time response.
Keywords: Behavior-based navigation, fuzzy logic, multi-objective optimization,
mobile robot
1. Introduction
Mobile robot navigation is one of the most challenging problems in robotics. To
complete a navigation task, the robot must be capable of perceiving its surrounding
environment, interpreting data from sensors, planning the path to be tracked and
controlling the actuators to reach the target [1]. Navigation, on the other hand, is
fundamental for mobile robot applications. In order to complete any given task, the robot
first needs to have the capability of reaching the target safely. Navigation of mobile robot
thus has been receiving much research attention and the approaches can be classified into
two main categories: hierarchical architectures and reactive or behavior-based
architectures [2].
The hierarchical architecture operates through sequent steps of sensing, planning and
acting based on a known model of the environment. This architecture is appropriate for
static and structured environments. For unknown environments, the behavior-based
architecture is often used. This approach splits a complex navigation task into sub-tasks or
behaviors as shown in Figure 1. Each behavior is an independent control module dealing
with a specific problem of navigation. It takes input data from sensors and generates an
output control signal specifying for its objective. The output control signals of all
behaviors are then combined in accordance with the global navigating objective to
generate an overall control signal. As the combination only uses the local data, the
behavior-based architecture does not need to have a global map of the environment.
Besides, the use of behaviors enables the modularization and scalability of the
architecture.
Task
Control signal 1
Behavior 2
Control signal 2
...
Behavior N
Control signal N
Command
fusion
Overall control signal
Sensors
Behavior 1
Figure 1. General Scheme of the Behavior-based Navigation Architecture
The main challenge with the behavior-based architecture is how to combine behaviors,
alled command fusion, to achieve the navigation objective efficiently. The superposition
techniques deal with this problem by using a linear combination of behaviors generated
from potential fields [3] or motor schemata [4]. The potential fields select an action based
on vector sums of the potential fields produced by an attractive force from goal and
repulsive forces from obstacles. The motor schemata use the potential field to define the
output of each schema and then combine them to generate a motor action basing on
predetermined weighting factors. Those techniques are simple to implement but difficult
to adjust gains.
Another command fusion approach is voting techniques in which each behavior votes
for or against a set of actions. The actions are then combined to generate the best one. In
[5], the combination is simply the sum of votes. In DAMN [6], the combination is based
on weighting factors assigned by a mode manager. Due to compromise without priority,
these techniques may present poor performance in conflict situations, for example, if the
”obstacle avoidance” behavior votes for turning right to avoid an obstacle in front of it,
while the ”goal reaching” behavior votes for turning left since the goal is on the left of it,
the compromised action may direct the robot forward resulting a collision with the
obstacle.
A command fusion approach commonly used in recent mobile robot navigation
systems [7] - [14] is the fuzzy technique. This technique presents each behavior by a fuzzy
controller. The output fuzzy sets of all controllers are then combined and defuzzified to
generate the overall control signal. This technique is simple to implement and quite
efficient in navigation. The fusion, however, is not optimal as each defuzzification
method often results in a different control value [15], [16].
In order to deal with the optimization problem in command fusion, a technique based
on multi-objective optimization theory, called MOASM, was proposed [17]. This
technique represents each behavior by an objective function that assigns to each control
signal a value reflecting the grade of behavior’s objective. A multiobjective optimization
process is then applied to find the solution which best maximizes all the objective
functions. The main advantage of this technique is its theoretical approach to ensure the
optimality of the found solutions. However, the lack of a framework for designing
objective functions, which are usually complicated, prevented it from practical use.
In this study, we propose an approach to integrating the advantages of fuzzy logic and
multi-objective optimization into a single behavior-based navigation architecture called
BBFM. In this architecture, each behavior is represented by a fuzzy controller which only
contains the fuzzification and fuzzy inference procedures. Consequently, the output of
each fuzzy controller is a function of input variables whose value represents the grade of
behavior’s objective. They are then used as the input for a multi-objective optimization
process to find the optimal value of the overall control signal. The results from a number
of simulations, comparisons, and experiments confirm the efficiency of the proposed
architecture in navigating the mobile robot in complex and unknown environments.
350
The structure of paper includes five sections. Section 2 presents the BBFM
architecture. Section 3 simulates and compares the BBFM with two other popular
architectures. Section 4 presents experimental results. The paper finishes with conclusions
in Section 5.
2. Proposed Architecture
In this section, the overall structure of the BBFM architecture is first described. After
that, detailed implementation of each behavior is introduced. Finally, the command fusion
using multi-objective optimization is presented.
2.1. Overall Structure of the BBFM
The robot used to evaluate the proposed architecture is the type of differential drive
wheeled mobile robot with non-holonomic constraints. Its parameters are shown in Figure
2.
Figure 2. The Differential Drive Wheeled Mobile Robot and Its Parametes
where R is the wheel diameter, L is the distance between two wheels, and (x, y, θ)
represents the position and orientation of the robot. Let (xd, yd, θd) be the position and
orientation of the target. We define three additional variables for navigation: ρ defined by
Equation (1) is the distance from the center of the robot to the target; α defined by
Equation (2) is the angle between the robot heading and the vector connecting the robot
center with the target; and ed defined by Equation (3) represents the movement of the
robot with the target. For the sake of simplicity, whenever we refer to the position of the
robot in this paper, we mean its position and orientation.
( xd x)2 ( yd y)2
arctan( yd y, xd x) , [ , ]
ed i i 1
(1)
(2)
(3)
In the system, the motion of robot is controlled by adjusting its linear velocity, u, and
angular velocity, ω. The position of robot is determined via optical quadrature encoders.
To sense the environment, the robot is equipped with eight ultrasonic sensors clustered
into three groups of left, right, and front as shown in Figure 3. The measuring value of
each group is the minimum value of all sensors in that group:
d r min(d1 , d 2, d3 )
d f min(d 4 , d 5 )
(4)
dl min(d 6 , d 7, d8 )
where di is the distance to obstacles measured by sensor i. The mission of the robot is
to navigate in an unknown environment from an initial position to a desired target without
colliding with obstacles and getting trapped in any trap area.
351
Figure 3. Arrangement of Ultrasonic Sensors on the Robot
In order to complete this task, it is natural to subdivide it into small and easy-tomanage behaviors that focus on execution of specific subtasks. The subtasks would be: (i)
reaching the target from an arbitrary position, (ii) avoiding obstacles, and (iii) escaping
local minimum (trapped) regions. As a result, this approach simplifies the navigation
solution while offering a possibility to add new behaviors to the system without causing
any major increase in complexity. Individual behavior, however, needs to cope with
uncertainties and incompleteness of sensory information as well as with the fact that the
operating environment contains elements of dynamics and variability. Fuzzy logic is
known to be an organized method for dealing with those problems. Using linguistic rules,
fuzzy logic requires neither mathematical models of the environment nor the robot
dynamics to design motion controllers. Instead, it fuzzifies the inputs and takes advantage
of expert knowledge to discover and represent data relationships and to improve
uncertainty reasoning. Therefore, fuzzy logic provides a framework for designing
individual behavior.
The command fusion using fuzzy logic, however, does not give reliable solutions.
Figure 4 shows that the two ways of fusion using fuzzy logic: defuzzifying first and then
combining individual decisions, and combining individual decisions first and then
defuzzifying, generate different results. The reason is due to the concurrent activation of
multiple behaviors with possibility of conflict between them. In this context, it may not
exist a globally optimal solution that is simultaneously optimal with respect to all
behaviors. The optimization of one behavior’s objective might be associated with a
simultaneous deterioration of other behavior’s objective. Thus, it is more appropriate to
search for a ”good enough” solution that can guarantee a suitable trade-off between a
multitude of conflicting behaviors’ objectives. Multi-objective optimization provides a
theoretical approach to find this solution. It quantifies the problem through objective
functions that assign to each input a value reflecting their objectives’ desirability and
provide methods to optimize those functions simultaneously. Interestingly, if we remove
the defuzzification procedure in the design process of fuzzy controllers, the output of each
behavior will be a fuzzy membership function that its value represents the grade of
behavior’s objective. This function, therefore, encrypts the semantics meaning of
objective functions. This way, the fuzzy technique and multi-objective optimization can
be combined.
Defuzzification
0
-10
0+10
∑
4
-10
0
-10
8
Defuzzification
-10
0+10
(a)
Defuzzification
7.5
+10
-10
0
0
+10
+10
(b)
Figue 4. Two Fuzzy-based Approaches to Command Fusion: (a)
Defuzzifying First and Then Combining, (b) Combining First and Then
Defuzzifying
352
Figure 5 shows the structure of our proposed architecture. It has three behaviors
including the local minimum avoidance, obstacle avoidance, and goal reaching. Each
behavior is implemented by one customized fuzzy controller consisting of fuzzification
and inference modules. The inputs include variables defined in equations (1) - (4). The
output of each fuzzy controller includes two membership functions that map the input
space to the interval of [0, 1] representing the grade of behavior objective. They are then
used as the objective functions for a multi-objective optimization process to generate the
overall control signals, u* and ω*, which are the Pareto-optimal solutions.
Task
Local minimum avoidance Objective function
dl
df
dr
ed
Fuzzification
Inference
Obstacle avoidance
Fuzzification
Inference
Goal reaching
Fuzzification
Inference
LM u
LM
OA u
OA
GR u
GR
Command fusion
Multi-objective Optimization
uˆ max { LM (u), OA (u), GR (u)}
ˆ max { LM (), OA (), GR ()}
u*
*
Figure 5. The Overall Structure of the BBFM
2.2. Design of Individual Behavior
Each behavior is implemented by one customized fuzzy controller which only contains
the fuzzification and inference procedures, ignoring the defuzzification. The fuzzification
procedure maps the crisp input values to the fuzzy linguistic terms. Each linguistic term is
determined by a fuzzy set that is characterized by its membership function. In this study,
the Gaussian and Sigmoid membership functions are chosen to represent the fuzzy sets as
follows:
( xi cij )2
A e
2( ij ) 2
ij
B e
lk
( yl clk )2
2( lk )2
(5)
(6)
for the input and output Gaussian membership functions, respectively, and
1
Aij
(7)
aij ( xi bij )
1 e
1
Blk
(8)
alk ( xl blk )
1 e
for the input and output Sigmoid membership functions, respectively, where xi is the ith
input variable, yl is the lth output variables, Aij is the jth fuzzy set of the ith input variable,
Blk is the kth fuzzy set of the lth output variable, {cij, σij} and {clk, σlk} are the parameters
of input and output Gaussian membership functions, respectively, and {aij, bij} and {alk,
blk} are the parameters of input and output Sigmoid membership functions, respectively.
The inference procedure is responsible for formulating the relationship between the
inputs and the outputs. It is based on rules of the form” if...then...”, for example,” if x1 =
A1j and x2 = A2j and . . . xm = Amj then y1 = B1k and y2 = B2k and . . . yn = Bnk”. The result of
each rule for each output variable is then given by:
353
R ( yl ) min( H , B ( yl )),
k
lk
(9)
H min{ A1 j ( x1 ), A2 j ( x2 ),...., Amj ( xm )}
where Rk denotes the kth rule. For M control rules, the implication result of each output
variable according to the max-min method is an aggregated output fuzzy set with the
membership function determined by:
R ( yl ) max(R1 ( yl ), R2 ( yl ),...., RM ( yl )) .
(10)
The membership function (10) represents the degree of membership of each element in
the set of output variable yl. Detailed implementations of the fuzzification and inference
procedures for each behavior are described as follows.
2.2.1. Obstacle Avoidance
The obstacle avoidance behavior includes four input variables dr, df, dl, and α, and two
output variables u and ω as shown in Figure 5. Their linguistic terms and membership
functions are defined as shown in Figure 6.
1 N
F
M
0.5
0
1
1.5
2
2.5
1 LN
3
3.5 d (m)
-3
-2
N
Z
P
-1
0
1
(a)
S
0
2
3 (rad)
(b)
1
LP
L
M
0.2
0.4
0.6
0.8
1 LNO
1 u (m/s)
-4
-3
-2
(c)
PO
ZO
NO
-1
0
1
2
PNO
3 (rad/s)
(d)
Figure 6. The Linguistic Terms and Membership Function of Input and
Output Variables of the Obstacle Avoidance Behavior: (a) dl, df, dr; (b) α; (c):
u; (d): ω
Table 1 presents 28 control rules defined for the behavior. Let ROA,k (u ) and
R
OA ,k
( ) be respectively the results of the kth rule in this table for u and ω by using
Equation (9). The implication results according to the max-min method are then given by:
ROA (u ) max( ROA ,1 (u ), ROA ,2 (u ),..., ROA ,28 (u ))
(11)
ROA ( ) max( ROA ,1 ( ), ROA ,2 ( ),..., ROA ,28 ( )).
Table 1. Rules for Obstacle Avoidance
Collisions
354
Rule
1
dl
N
Input
df
dr
N
F
2
3
4
5
F
M
F
N
N
N
N
N
N
N
M
F
α
Output
u
ω
S
Po
M
M
N
M
Po
Po
LPo
LNo
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
N
F
M
F
N
N
M
N
N
N
N
N
N
N
N
M
F
M
F
F
F
F
F
N
N
N
N
M
N
N
M
M
F
F
F
F
N
F
M
M
F
F
F
F
F
F
M
F
F
M
N
N
M
M
F
M
F
F
F
N
F
N
N
N
N
N
N
N
N
M
M
M
M
M
S
S
M
M
M
S
S
L
L
L
M
M
M
L
L
L
S
S
LN
N
Z
LP
P
LN
N
Z
LP
P
No
Lpo
No
Po
Po
Po
Po
No
No
No
Lno
No
Zo
Zo
Zo
Po
Po
Po
Zo
Zo
Zo
Lpo
LPo
2.2.2. Goal Reaching
The objective of goal-reaching behavior is to control the robot to reach the target in the
shortest time. In order to complete this task, the behavior continuously adjusts the robot
direction to match the target direction while driving the robot at the fastest possible speed.
The behavior uses two input and two output variables as shown in Figure 5. Three of them
including the deflection angle α and the velocities u and ω have the same definition of
linguistic terms and membership functions as in the obstacle avoidance behavior. The
fourth variable, ρ, has the linguistic terms and membership functions defined as shown in
Figure 7.
1N M
0
2
F
4
6
8
10
12 14 16 18
20
(m)
Figure 7. The Linguistic Terms and Membership Function of ρ
The behavior has 15 rules defined as in Table 2. Let R
GR ,k
(u ) and RGR ,k ( ) be
respectively the results of the kth rule in the table for output variables u and ω by using
Equation (9). The implication results for u and ω according to the max-min method are
then given by:
RGR (u ) max( RGR ,1 (u ), RGR ,2 (u ),..., RGR ,15 (u))
(12)
RGR ( ) max( RGR ,1 ( ), RGR ,2 ( ),..., RGR ,15 ( )).
355
Table 2. Rules for Goal Reaching
Rule
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
Input
ρ
α
N
Z
N
N
N LN
N
P
N LP
M
Z
M
N
M LN
M
P
M LP
F
Z
F
N
F LN
F
P
F
LP
Output
u
ω
S
Zo
S
No
S LNo
S
Po
S LPo
M Zo
M No
M LNo
M Po
M LPo
L
Zo
L
No
L LNo
L
Po
L LPo
2.2.3. Local Minimun Avoidance
During navigation, the robot may be trapped in U-shape regions called local minimum
problem. It happens when the robot repeatedly conducts opposite turning commands
during avoiding obstacles. Figure 8(a) illustrates such a situation. The robot starts from A,
moves toward the target according to rules 6 and 11 in Table 1; at B, it turns left
according to rules 11 and 12 to move to C or it can turn right to move to D depending on
the priority; at C, it turns left to go toward the target according to rules 27 and 28 and
back to B. The process is repeated causing the robot to be trapped.
Target
Target
k 1
B
D
C
C
k
B
C1
Cn
(a)
(b)
Figure 8. The Fuzzy-based Method for Dealing with the Local Minimum
Problem: (a) the Robot is trapped in a Local Minimum Region; (b) the Robot
Scapes the Local Minimum Region Using the Fuzzy-based Method
In order to deal with this problem, we carefully analyze various trapped situations and
realize that it is critical to detect the trapped points which are points C and D in the
example shown above. We detect them by checking if the distance ed is increasing, the
target is behind the robot (corresponding to the positive value of α), and the obstacle is on
the right (or left) side of robot. Figure 8(b) shows that if the robot goes straight instead of
turning left at the trapped point C, it will reach Cn as the conditions do not change from C
to Cn. At Cn, the robot will turn right to exit the local minimum region. In order to
implement the proposed solution, the behavior of local minimum avoidance has five input
variables including dl, df, dr, α, and ed and two output variables including u and ω. The
linguistic terms and membership functions of ed are defined as in Figure 9.
356
1
ZT
NT
-1 -0.8 -0.6 -0.4 -0.2
PT
0 0.2 0.4 0.6 0.8
1 ed (m)
Figure 9. The Linguistic Terms and Membership Function of ed
The linguistic terms and membership functions of the remained variables are defined as
in the obstacle avoidance behavior. Table 3 presents 7 control rules for detecting trapped
points and escaping local minimum regions. Let R (u ) and R ( ) be respectively
LM ,k
LM ,k
the results of the kth rule in the table for output variables u and ω by using Equation (9).
The implication results for u and ω according to the max-min method are then given by:
RLM (u ) max( RLM ,1 (u ), RLM ,2 (u ),..., RLM ,7 (u ))
(13)
RLM ( ) max( RLM ,1 ( ), RLM ,2 ( ),..., RLM ,7 ( )).
Table 3. Rules for Local Minimum Avoidance
Rule
1
2
3
4
5
6
7
dl
N
F
F
F
F
F
F
df
N
N
N
F
F
F
F
Input
dr
N
N
N
N
N
F
F
ed
PT
PT
PT
PT
α
Z
P
LP
P
LP
P
LP
Output
u
ω
S
Po
M
Po
S
LPo
M
Zo
M
Zo
M
Zo
M LNo
2.3. Command Fusion
The command fusion is carried out by using multiobjective optimization. As shown in
Figure 5, the membership functions of output variables are used as the objective
functions. Let μi(y) be the ith objective function, y be an output control signal (y = u or y
= ω), Y be the domain of y, and N be the number of objective functions. The optimal
value of each output control signal is then the solution of the following equation:
yˆ arg max[1 ( y), 2 ( y),..., N ( y)]
(14)
According to the theory of multi-objective optimization, there might not exist the
optimal solution, ŷ , of Equation (14), but only the ”good enough” solution, y*, which is
the best fit for all objectives. This solution is called the Pareto-optimal solution or nondominated solution defined as follows: y* is the Pareto-optimal solution of Equation (14)
if there does not exist any y Y such that μi(y) > μi(y*) for at least one i and μj(y) ≥ μj(y*)
for all j. In other words, the Pareto optimal solution is the one in which there is not other
solution that improves an objective without resulting in the deterioration of at least
another objective.
In order to find the Pareto-optimal solution, there exist a number of methods [17] such
as weighting, lexicographic, and goal programming. In the context of command fusion,
we choose to use the lexicographic method due to its efficiency. This method first
requires ranking the order of importance of all objectives, for instance, µ1 is the most
important and µN is the least important. The Pareto-optimal solutions are then obtained by
357
solving the following sequence of problems until either a unique solution is found or all
the problems are solved:
P1 : max[ 1 ( y )]
yY
P2 : max[ 2 ( y )]
yY1
...
PN : max[ N ( y )]
(15)
yYN 1
Yi { y | y solves Pi }, i 1,..., N 1
In our system, the optimal values of the overall control signal are determined by the
output membership functions of (11), (12), and (13) as follows:
uˆ arg max[ ROA (u ), RGR (u ), RLM (u )],
(16)
ˆ arg max[ ROA ( ), RGR ( ), RLM ( )]
The lexicographic method used to find the Pareto optimal solutions of (16) are carried
out as follows:
Sorting all behaviors in descending order of importance: local minimum
avoidance, obstacle avoidance, and goal reaching.
Sequentially solving equations Pi by using discrete values of u and ω on their
domains U and W until a unique solution is obtained, or all equations are solved:
P1 : max[ RLM (u )],
P1 : max[ RLM ( )],
uU
W
P
:
max[
(
u
)],
P
:
max[
2 uU
2 W ROA ( )],
ROA
1
1
*
*
u : U1 {u | u solves P1},
: W1 { | solves P1}, (17)
P : max[ (u )],
P : max[ ( )],
3 uU 2 RGR
3 W2 RGR
U 2 {u | u solves P2 }
W2 { | solves P2 }
If more than one Pareto-optimal solution is obtained, the one with the greatest
value of u and the smallest value of ω is chosen.
3. Simulations
Simulations have been conducted to evaluate the efficiency of the BBFM compared to
two other popular architectures including the MOASM [17] and CDB [10]. The MOASM
uses multi-objective optimization for command fusion. It is implemented with three
behaviors including avoiding obstacles, maintaining the target heading and moving fast
forward. The objective functions are defined as in the origin [17]. The overall control
signal is determined by using the lexicographic method. The CDB is implemented with
three behaviors as in the MOASM. Each behavior is a fuzzy controller. The overall
control signal is determined by using fuzzy meta rules and defuzzification. In order to
ensure the fairness in comparison, the BBFM only uses the obstacle avoidance and goal
reaching behaviors.
All architectures are simulated in Matlab and use the same robot configuration. Its
mechanical parameters are set as follows: R = 0.085 m, L = 0.265 m, u [0, 1.3] m/s,
and ω [-4.3, 4.3] rad/s. The ultrasonic sensors have the sensing range from 0 m to 4 m
and the radiation cone of 15°. They are arranged in front of the robot as shown in Figure 3
to cover the range of 160°. The universe of discourse of ρ is in the range of [0, 20] and
that of ed is [-1, 1]. In simulations, three scenarios were chosen for navigating the robot. In
each scenario, 15 runs were carried out to evaluate the performance of each architecture.
Details and results of each scenario are presented as follows.
358
3.1. Popular Operating Environment
In scenario 1, the operating environment is chosen to be the same as in the original
paper of MOASM [17]. The start position is (-2, -1.8, 180°) and the target position is (-6, 4.8, 0°). Figure 10 shows the path of robot generated by three architectures MOASM,
BBFM, and CDB. It can be seen that all architectures successfully navigate the robot to
avoid obstacles and reach the target. Table 4 shows the average performance of each
architecture, where the index smoothness is the average absolute value of the difference
between the current and the previous orientation, thus showing how smooth the
maneuvers is; the index target error is the distance between the final position of the robot
and the target position, thus evaluating the reachable ability to the target at the steady
state; and the indexes traveled distance and elapsed time are respectively the total distance
of the robot’s traveled path and the time taken to go through that path. As can be inferred
from Table 4, the BBFM is more efficient than the remaining architectures in almost all
criteria.
(a)
(b)
(c)
Figure 10. Travelling Paths of the Robot Navigated by Three Architectures
in Scenario 1: (a) BBFM, (b) MOASM, (c) CDB
Table 4. Navigation Results in Scenario 1
Index
Traveled distance (m)
Elapsed time (second)
Smoothness (degree)
Target error (m)
BBFM
10.36
28.26
0.88
0.05
MOASM
11.02
41.45
1.29
0.2
CDB
11.02
36.43
6.1
0.05
3.2. Office-like Operating Environment
In scenario 2, the operating environment is chosen to be more like an office with wall
and bulkhead obstacles. The start position is (-7, -6, 0°) and the target is (-2.5, -1.5, 0°).
Figure 11 and Table 5 show the navigation results. It can be seen that the MOASM does
not complete the navigation task as its objective functions are built based on the principle
of Instantaneous Center of Curvature (ICC) of differential drive wheeled mobile robot
which does not efficient in escaping corners. On the other hand, both the BBFM and CDB
can safely navigate the robot to reach the target due to the efficiency of fuzzy control.
However, the BBFM is more efficient than the CDB.
359
(a)
(b)
(c)
Figure 11. Travelling Paths of the Robot Navigated by Three Architectures
in Scenario 2: (a) BBFM, (b) MOASM, (c) CDB
Table 5. Navigation Results in Scenario 2
Index
Traveled distance (m)
Elapsed time (second)
Smoothness (degree)
Target error (m)
BBFM
9.35
12.09
2.04
0.05
CDB
15.66
24.45
3.16
0.05
3.3. Operating Environment with Local Minimun Regions
In the third scenario, the operating environment has local minimum regions. The robot
starts at position (-6, -6, 0°) and desires to reach position (-2.5, -1.5, 0°). The results in
Figure 12 show that all three architectures are not able to complete the navigation task
because of local minimum problem. This problem can be solved if adding a local
minimum avoidance behavior to the BBFM as shown in Figure 13(a). Even in a more
complicated local minimum situation, the BBFM still completes the navigation task as
shown in Figure 13(b).
(a)
(b)
(c)
Figure 12. Travelling Paths of the Robot Navigated by Three Architectures
in Scenario 3: (a) BBFM, (b) MOASM, (c) CDB
Figure 13. Travelling Paths of the Robot Navigated by the BBFM in
Operating Environment with Local Minimum Regions
360
4. Experiments
In order to evaluate the performance of the BBFM in real environments, we carried out
a number of experiments described as follows.
4.1. Experimental Setup
The robot used in experiments is a Sputnik robot [18] as shown in Figure 14. It is
equipped with three ultrasonic sensors DUR5200 at left, front and right directions creating
the scanning range from −60° to 60°. To extend the scanning range from −90° to 90°, we
added two additional ultrasonic sensors SRF05 to the left and right sides of the robot.
Each employs a microcontroller PIC12F1572 to synchronize its data with the main board
of Sputnik robot. The linear and angular velocities of the robot are respectively set to [0,
0.5] m/s and [-3.7, 3.7] rad/s. The position of the robot is determined via optical encoder
sensors. The robot has a wireless module connecting it to a Wifi router (Figure 14). The
BBFM is written in Matlab and installed on a PC which communicates with the robot
through a Wifi router. The sampling time Ts is 300 ms and the experimental environment
is an indoor office with the size of 4 m x 3 m and changeable obstacles.
Figure 14. The Sputnik Robot and Its Configuration to Communicate
with the Control Computer
4.2. Experimental Results
Experiments were carried out with different configurations of the environment and
target position. Figure 15 shows results of the traveled paths, velocity responses and
images of the robot in three of such configurations. In configuration 1 (Figure 15(a)), the
robot starts at A (0, 0, 90°) and then moves forward along two walls to B. At B, it turns
right two times to C and then goes straight to D. At D, the robot continuously adjusts its
direction to avoid the bulkhead corners while on average maintaining the target heading to
reach E and finally goes straight to reach the target F (1.9, 2.5, 0°). Figure 15(b) shows the
correspondence of linear and angular velocities of the robot with those movements, for
instance, between D and E near the target, the angular velocity switches between the left
and right directions to avoid the bulkhead corners while the linear velocity gradually
decreases to prepare for stopping at the target.
In configuration 2 as shown in Figure 15(d), the environment structure changed with
more potential local minimum regions. The robot starts at position (-0.1, -1, 90°) and the
target is set to (-0.1, 2.5, 0°). The path from B to D shows that the robot successfully
escapes the local minimum region near the starting position. From D to F, it succeeds in
avoiding obstacles located at the center and near the target. Figure 15(e) shows the
correspondence of the robot’s velocities with its movements. The operating environment
in Configuration 3 is similar to the configuration 2. However, the start and target positions
are changed to (0.1, -0.2, 0°) and (1.8, 2.3, 0°), respectively. As can be inferred from
Figure 15(g)-(i), the robot can reach the target while avoiding obstacles and potential
local minimum regions.
Table 6 shows the navigation performance in each configuration. As can be inferred
from the table, configuration 1 introduces the best performance in smoothness and elapsed
361
time because it is the simplest case. On the contrary, configuration 2 introduces the worst
performance due to obstacles and local minimum regions. Nevertheless, the closeness
between values of average linear velocities in all configurations implies that the operation
of robot is stable and suitable for the indoor environment.
Table 6. Navigation Results in Three Configurations
Configuration
1
2
3
uaverage
(m/s)
0.18
0.15
0.16
Smoothness
( degree)
3.72
6.54
5.75
Elapsed
time (s)
22.5
33
24
Traveled
distance (m)
4.06
4.86
3.77
(a)
(b)
(c)
(d)
(e)
(f)
(g)
(h)
(i)
Target
error (m)
0.05
0.08
0.07
Figure 15. Travelling Paths, Velocity Responses, and Photos of the
Robot Operating in Three Different Navigation Configurations: (a) – (c):
Configuration 1, (d) – (f): Configuration 2, (g) – (i): Configuration 3
5. Conclusions
In this paper, we have proposed new behaviorbased navigation architecture for
navigating the mobile robot in unknown environments. We modified the procedures of
designing fuzzy controllers for behaviors so that their outputs can be used as inputs for a
multiobjective optimization process to coordinate the behaviors. Consequently, the
architecture inherits advantages of fuzzy logic in dealing with uncertainties of sensory
information while providing a framework for designing objective functions. It also takes
advantage of multiobjective optimization to generate Pareto-optimal solutions for
command fusion. During the development, we proposed a very simple yet effective fuzzybased approach for dealing with the local minimum problem which often happens in
362
mobile robot navigation. The results show that the proposed architecture is practical in
implementation and possibly navigates the robot to reach the target along a smooth and
efficient trajectory in environments with unpredictable obstacles, topographies and local
minimum regions.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
S. Roland and N. I. R, “Introduction to autonomous mobile robots”, The MIT Press Cambridge, (2004),
pp. 10–11.
D. Nakhaeinia, S. H. Tang, S. B. M. Noor and O. Motlagh, “A review of control architectures for
autonomous navigation of mobile robot”, International Journal of the Physical Sciences, vol. 6, no. 2,
(2011), pp. 169–174.
O. Khatib, “Real-time obstacle avoidance for manipulators and mobile robots”, The International
Journal of Robotics Research, vol. 1, no. 5, (1986), pp. 90–98.
R. Arkin, “Motor-schema based mobile robot navigation”, Int. J. Robot, vol. 8, no. 4, (1989).
J. Hoff and G. Bekey, “An architecture for behavior coordination learning”, In IEEE International
Conference on Neural Networks, (1995).
J. Rosenblatt, “DAMN: A distributed architecture for mobile navigation”, In Proceedings of the AAAI
Spring Symposium on Software Architecture for Physical Agents, (1995).
Saffiotti, “The uses of fuzzy logic in autonomous robot navigation”, Soft Computing, Springer Verlag,
(1997), pp. 180–197.
M. S.-F. Eduardo Freire, Teodiano Bastos-Filho and R. Carelli, “A new mobile robot control approach
via fusion of control signals”, IEEE transactions on system, mam and cybernetics, vol. 34, no. 1, (2004).
S. A. Yahmedi, A. B. El-Tahir El-Dirdiri and T. Pervez, “Behavior based control of a robotic based
navigation aid for the blind”, Control and Applications Conference, (2009).
S. A. Yahmedi and M. A. Fatmi, “Fuzzy logic based navigation of mobile robot”, Recent Advances in
Mobile Robotics, (2011).
G. Z. Ye and D.-K. Kang, “Fusion of hierarchical behavior-based actions in mobile robot using fuzzy
logic”, Journal of Information and Communication Convergence Engineering, vol. 10, no. 2, (2012), pp.
149–155.
Q. T. Hongwei Mo and L. Meng, “Behavior - based fuzzy control for mobile robot navigation”,
Mathematical Problems in Engineering, no. Article ID 561451, (2013).
M. Faisal, K. Al-Mutib, R. Hedjar, H. Mathkour, M. Alsulaiman and E. Mattar, “Behavior based mobile
for mobile robots navigation and obstacle avoidance”, International Journal of computers and
communications, vol. 8, (2014).
R. V. D. Sousa, R. A. Tabile, R. Y. Inamasu and A. J. V. Porto, “A row crop following behavior based
on primitive fuzzy behaviors for navigation system of agricultural robots”, 4th IFAC Conference on
Modelling and Control in Agriculture, Horticulture and Post Harvest Industry, vol. 46, (2013), pp. 91–96.
D. Driankov, H. Hellendoorn and M. Reinfrank, “An introduction to fuzzy control,” Spinger, (2010), pp.
132–141.
S. Naaz, A. Alam and R. Biswas, “Effect of different defuzzification methods in a fuzzy based load
balancing application”, International Journal of Computer Science Issues, vol. 8, no. 1, (2011), pp. 261–
267.
P. Pirjanian, “Multiple objective behavior-based control”, Robotics and Autonomous Systems, Elsevier,
no. 31, (2000), pp. 53–60.
D. R. Manual, http://www.drrobot.com/products.
363
| 3cs.SY
|
A New Simulation Algorithm for Absorbing
Receiver in Molecular Communication
Yiran Wang† , Adam Noel‡ , and Nan Yang†
arXiv:1803.04638v1 [cs.ET] 13 Mar 2018
†
Research School of Engineering, Australian National University, Canberra, ACT 2601, Australia
‡
School of Engineering, University of Warwick, Coventry, CV4 7AL, UK
Email: [email protected], [email protected], [email protected]
Abstract—The simulation of diffusion-based molecular communication systems with absorbing receivers often requires a
high computational complexity to produce accurate results. In
this work, a new a priori Monte Carlo (APMC) algorithm
is proposed to precisely simulate the molecules absorbed at
a spherical receiver when the simulation time step length is
relatively large. This algorithm addresses the limitations of the
current refined Monte Carlo (RMC) algorithm, since the RMC
algorithm provides accurate simulation only for a relatively
small time step length. The APMC algorithm is demonstrated to
achieve a higher simulation efficiency than the existing algorithms
by finding that the APMC algorithm, for a relatively large
time step length, absorbs the fraction of molecules expected by
analysis, while other algorithms do not.
I. I NTRODUCTION
Molecular communication (MC) has recently emerged as
an underpinning paradigm of exchanging information in nanoscale environments such as organs, tissues, soil, and water [1].
This information exchange is conducted among nanomachines
which are the devices with nano-scale functional components.
Within the research community, diffusion-based MC is a
simple but commonly adopted MC system. In such a system,
information molecules propagate using kinetic energy only,
which preserves a high energy efficiency. One of the major
challenges in designing and analyzing a diffusion-based MC
system is receiver modeling. The majority of the existing
MC studies have considered two types of receivers: passive
receivers and active receivers. Passive receivers do not impose
any impact on molecule propagation, while active receivers
can absorb molecules when they hit the receiver’s surface.
In nature, most receiver nanomachines react with or remove
selected information molecules from the environment when
they hit the receiver’s surface. Thus, active receivers are
generally more realistic than passive ones in describing the
chemical detection mechanism in MC. This motivates us to
investigate the properties of absorbing receivers.
The notion of diffusion with absorption is a long existing
phenomenon that has been described in the literature, e.g., [2].
Considering a diffusion-based MC system with a single fully
absorbing receiver within an unbounded three-dimensional
environment, [3] presented the hitting rate of molecules at
different times and the fraction of molecules absorbed until
a given time. Recently, [4] and [5] evaluated the impact of
receiver with reversible absorption on the performance of
MC systems, where a molecule can be released back to the
environment at some time after being captured by the receiver.
Apart from theoretical analysis such as [3]–[5], the simulation of MC systems is also an effective means for performance
evaluation. One of the most common MC simulation methods
is particle-based microscopic simulation. For diffusion-based
MC, molecules are moved by adding Gaussian random variables (RVs) to their x-, y-, and z-coordinates at the end of
every simulation time step. In practice, molecules may actually
diffuse into an absorbing receiver between two sampling times.
To determine this absorption, some existing simulation algorithms simply compared the observed coordinates of molecules
with those of the receiver [4], [6], [7]. As a consequence,
the molecules that hit a receiver between sampling times
cannot be considered as “absorbed” by using these simulation
algorithms. Recently, [8] and [9] investigated the possibility of
molecule absorption between sampling times. Specifically, [8]
declared a molecule as “absorbed” if its straight-line trajectory
within a time step crossed an absorbing surface. Differently,
[9] approximated the intra-step absorption probability for
spherical receiver boundaries using the equation for planar
receiver boundaries given by [10]. This approximation was
referred to as the refined Monte Carlo (RMC) algorithm.
The key contribution in this paper is that we propose a new
algorithm, i.e., the a priori Monte Carlo (APMC) algorithm,
for simulating an MC system with an absorbing receiver,
considering a relatively large time step length. We show that
using our APMC algorithm, the fraction of molecules absorbed
at the receiver precisely matches the corresponding analytical
result when the time step length is relatively large, or equivalently, the receiver’s radius is small. This demonstrates that
our APMC algorithm achieves a high simulation efficiency.
II. S YSTEM M ODEL
We consider a diffusion-based MC system within a threedimensional space, as depicted in Fig. 1. In this system, a
point transmitter (TX) is located at the origin of the space and
a single fully absorbing spherical receiver (RX) is centered
at location (r0 , 0, 0). We denote rr as the RX’s radius and
Ωr as the RX’s boundary. At the beginning of a transmission
process, the TX instantaneously releases N molecules. We
assume that the molecules are small enough to be considered
as points. Once released, molecules diffuse in the environment
according to Brownian motion until hitting the RX’s boundary.
z-axis
Molecules
Algorithm 1 The common structure of simulation algorithms
for a single absorbing RX
Ωr
1:
Receiver
(r0,0,0)
Transmitter
(0,0,0)
y-axis
2:
3:
4:
x-a
xis
rr
5:
6:
7:
Fig. 1: Illustration of our system model. The TX is a point transmitter located
at (0, 0, 0), the RX is a spherical receiver located at (r0 , 0, 0) with rr
being the radius and Ωr being the RX’s fully absorbing boundary. Molecules
propagate in the environment according to Brownian motion.
Molecule’s initial position
before propagation
Absorbing
receiver
8:
9:
10:
11:
12:
13:
Determine the end time of simulation.
for all simulation time steps do
if t = 0 then
Add N molecules to environment.
end if
Scan all not-yet-absorbed molecules, i.e., molecules
which are not absorbed by the RX.
for all not-yet-absorbed molecules do
Propagate each molecule for one step according to
Brownian motion.
Determine if the molecule is absorbed. molecules
absorbed
Update the location and status of the molecule.
end for
Record the number of molecules absorbed by the RX.
end for
Molecule’s final location
after propagation
Fig. 2: Illustration of the intra-step molecule movement. There is a possibility
that a molecule crossed an absorbing boundary within one time step, even
if its initial and final positions during that time step are both outside the
absorbing receiver.
We denote Nhit (Ωr , t|r0 ) as the number of molecules released
from the origin at time t0 = 0 s and absorbed by the RX at
time t. As per [11, Eq. (3.116)], we express Nhit (Ωr , t|r0 ) as
N rr
r0 − rr
,
(1)
erfc √
Nhit (Ωr , t|r0 ) =
r0
4Dt
where D is the diffusion coefficient and erfc (·) is the complementary error function. Here, D describes the proportionality
constant between the flux due to molecular diffusion and the
gradient in the concentration of molecules.
As mentioned in Section I, the majority of the existing
simulation algorithms for absorbing RXs did not consider the
possibility of intra-step molecule absorption. This absorption
is depicted in Fig. 2, which shows that the actual trajectory
of a molecule may cross the absorbing boundary during one
simulation time step, even if its initial position at the beginning
of the time step and its final position at the end of the same
time step are both outside the absorbing RX. If this crossing
occurs, the molecule is absorbed by the RX in practice. The
ignorance of this absorption leads to an underestimation of the
number of molecules absorbed, thus deteriorating the accuracy
of simulation. In this work, we refer to the probability that
a molecule is absorbed during a simulation time step as the
intra-step absorption probability. We note that the intra-step
absorbing probability was only considered in [8], [9]. In [9],
the intra-step absorption probability of a fully absorbing RX
with the spherical boundary was approximated as that of a
fully absorbing RX with an infinite planar boundary, given
by [10, Eq. (10)]
li lf
,
(2)
PrßRMC = exp −
D∆t
where li is the initial distance of a molecule from the absorbing
boundary at the beginning of a time step, lf is the final distance
of a molecule from the absorbing boundary at the end of the
same time step, and ∆t is the simulation time step length.
III. S IMULATION A LGORITHMS
In this section, we first clearly present the common structure
of the existing simulation algorithms for absorbing receivers.
Then we present our proposed APMC algorithm.
A. Common Structure of Existing Simulation Algorithms
Built upon the observation of the existing simulation algorithms for microscopic molecule absorption (such as those
in [6], [8], [9]), we find that they follow a common structure.
This common structure is presented in Algorithm 1.
We clarify that each algorithm has its own criterion for
determining whether or not a molecule is absorbed by the RX,
as given in Line 9 of Algorithm 1. The determination criteria
for the existing algorithms in [6], [8], [9] are summarized as
follows:
•
•
•
As per the determination criterion in [6], the molecules
being observed inside the RX at the end of a time step
are absorbed. This criterion is referred to as the simplistic
Monte Carlo (SMC) algorithm.
As per the determination criterion in [8], a molecule is
absorbed if the line segment from its initial position to
its final position crosses the RX’s boundary. However, we
note that the line segment crossing the RX’s surface is
neither sufficient nor necessary to correctly detect intrastep absorption.
As per the determination criterion in [9], referred to as
the RMC algorithm, (2) is used to calculate the intra-step
absorption probability of a fully absorbing RX with the
spherical boundary.
which is obtained by scaling (1) by N , replacing the total
simulation time t with ∆t, and replacing r0 with dj . Then the
molecule absorption is determined by generating a uniform RV
u, where 0 ≤ u ≤ 1, and comparing its value with the probability obtained by (3). A molecule is marked as “absorbed”
if u ≤ PrßAPMC . After determining the molecules absorbed,
each not-yet-absorbed molecule is propagated according to
Brownian motion. If any not-yet-absorbed molecule is inside
the RX’s boundary at the end of the current time step, we
revert the movement of this molecule and let it propagate
again, until this molecule diffuses to a location outside the
RX. The APMC algorithm is detailed in Algorithm 2.
IV. N UMERICAL R ESULTS
In this section, we use Monte Carlo simulations to compare
the time-varying results produced by the SMC algorithm, the
RMC algorithm, and the APMC algorithm for a single receiver.
The algorithms are implemented in MATLAB. Throughout this
section, we set the diffusion coefficient as D = 10−9 m2 /s and
assume that other parameters vary. In the figures, M denotes
the number of time steps. If not otherwise noted, the TX-RX
distance is r0 = 50 µm and the number of molecules released
is N = 106 . The fraction of molecules absorbed is defined
as the ratio between the number of molecules absorbed at the
RX and the total number of molecules.
Fig. 3 plots the simulated fraction of molecules absorbed
versus time t for M = 100, ∆t = 0.1 s, and rr = 20 µm
or 0.5 µm. In this figure, the analytical result obtained from
(1) is also plotted for evaluating the accuracy of the algorithms. From this figure, we observe that our APMC algorithm
achieves a higher accuracy when rr decreases while ∆t,
D, and r0 remain unchanged. Specifically, Fig. 3(a) shows
that the RMC algorithm matches the fraction of
√ molecules
absorbed expected by the analytical result when D∆t/rr is
small, which meets our expectation. Indeed, the performance
√
of the √
RMC algorithm depends on the value of D∆t/rr .
When D∆t/rr approaches 0 and r0 is larger than rr , the
surface area of the RX can be approximated by an infinite
plane. Therefore, the probability of a molecule entering the
RX between sampling times is comparable to the probability
of a molecule crossing a planar boundary between sampling
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
Determine the end time of simulation.
for all simulation time steps do
if t = 0 then
Release N molecules into environment.
end if
Scan all not-yet-absorbed molecules.
for all not-yet-absorbed molecules do
Calculate the distance between the jth molecule to
(r0 , 0, 0), denote by dj .
Calculate the absorbed probability PrßAPMC for
each not-yet-absorbed molecule using (3) with rr , dj , D,
and ∆t.
if PrßAPMC ≥ u then
The molecule is absorbed.
end if
end for
for all not-yet-absorbed molecules do
Propagate the molecule for one step.
end for
for all free molecules in environment do
while the molecule’s distance to (r0 , 0, 0) ≤ rr do
Revert the movement of this molecule.
Propagate this molecule again.
end while
end for
Record the number of molecules absorbed at the RX
after this time step.
end for
0.6
0.5
0.4
0.04
RMC
SMC
APMC
Analytical
Fraction of Molecules Absorbed
In this subsection, we propose a new simulation algorithm
for approximating the√fraction of molecules absorbed at a fully
absorbing RX when D∆t/rr is larger than that considered
for the RMC algorithm in [9]. We refer to the newly proposed
algorithm as the APMC algorithm.
The procedure of the APMC algorithm is to first calculate,
before the jth molecule diffuses, the probability that this
molecule will be absorbed in the current time step. This
probability depends on the distance between this molecule
and the center of the RX, dj , and the time step length, ∆t.
Specifically, this probability is calculated as
dj − rr
rr
,
(3)
PrßAPMC = erfc √
dj
4D∆t
Algorithm 2 The APMC algorithm for molecule absorption
Fraction of Molecules Absorbed
B. A New A Priori Monte Carlo Algorithm
0.3
0.2
0.1
00
2
4
6
Time [s]
(a) rr = 20 µm
8
0.03
RMC
SMC
APMC
Analytical
0.02
0.01
0
0
2
4
6
8
Time [s]
(b) rr = 0.5 µm
Fig. 3: Comparison of the time-varying results produced by the SMC
algorithm, the RMC algorithm, and the APMC algorithm for M = 100 and
∆t = 0.1 s.
√
times. We also observe from Fig. 3(a) that when D∆t/rr is
small, the SMC algorithm and the APMC algorithm underestimates and overestimates the fraction of molecules absorbed,
respectively. Unlike Fig. 3(a), Fig. 3(b) shows that the APMC
algorithm matches the fraction of molecules absorbed expected
by the analytical result whereas the other two algorithms do
not. Particularly, the fractions of molecules absorbed produced
by the RMC and SMC algorithms are very far away from the
0.4
0.15
0.1
RMC
SMC
APMC
Analytical
0.05
0
0
1
2
3
Fraction of Molecules Absorbed
Fraction of Molecules Absorbed
0.2
0.3
0.2
0
4
RMC
SMC
APMC
Analytical
0.1
0
10
20
30
40
Time [s]
Time [s]
(a) ∆t = 0.5 s
(b) ∆t = 5 s
80
0.5
Distribution of
absorbed molecules
Analytical
60
0.4
0.3
40
0.2
20
0.1
0
1
2
3
4
5
0
Number of Molecules Newly Absorbed
Number of Molecules Newly Absorbed
Fig. 4: Comparison of the time-varying results produced by the SMC
algorithm, the RMC algorithm, and the APMC algorithm for M = 100 and
rr = 10 µm.
0.5
Distribution of
absorbed molecules
Analytical
60
0.4
0.3
40
0.2
20
0.1
0
1
2
3
4
Time [s]
Time [s]
(a) RMC Algorithm
(b) APMC Algorithm
5
0
Fig. 5: Distribution of the number of molecules newly absorbed during each
time step for the RMC and APMC algorithms with ∆t = 0.5 s, M = 10, and
rr = 10 µm. Simulations are repeated 103 times and N = 103 molecules
are released each time. The spectrum bars represent observation probabilities.
√
analytical result when t > 0 s. Indeed, when D∆t/rr is
large, the spherical RX’s boundary cannot be approximated
by a plane. Therefore, the RMC algorithm overestimates the
fraction of molecules absorbed.
Fig. 4 plots the simulated fraction of molecules absorbed,
together with the analytical result from (1), versus time t for
M = 100, rr = 10 µm and ∆t = 0.5 s or 5 s. We observe
from this figure that when ∆t increases from 0.5 s to 5 s,
our APMC algorithm achieves a higher accuracy. Specifically,
Fig. 4(a) shows that the gap between the fraction of molecules
absorbed produced by the APMC algorithm and that produced
by the RMC algorithm is very small. Also, Fig. 4(a) shows
that both the APMC and the RMC algorithms overestimate
the fraction of molecules absorbed. Unlike Fig. 4(a), Fig. 4(b)
shows that the APMC algorithm matches the fraction of
molecules absorbed expected by the analytical result. Fig. 4(b)
shows that the fraction of molecules absorbed produced by the
RMC algorithm is approximately twice of that produced by
the APMC algorithm when t ≥ 10 s. This
√ demonstrates the
accuracy of our APMC algorithm when D∆t/rr is large.
Fig. 5 depicts the distribution of molecules newly absorbed
during each time step (e.g., from t = 1 s to t = 1.5 s) for
both the RMC and APMC algorithms for M = 10, rr =
10µm, and ∆t = 0.5 s. The analytical result in this figure
is obtained based on (1). We observe from this figure that
the RMC algorithm significantly overestimates the number of
molecules newly absorbed when t = 1 s, while our APMC
algorithm gives an improved estimation accuracy when t = 1 s
by absorbing molecules according to (3). When t increases,
the overestimation of the RMC algorithm is slightly less severe
than that of the APMC algorithm. We further observe that
the distribution of molecules newly absorbed in Fig. 5(b) is
very similar to that in Fig. 5(a), which demonstrates that our
APMC algorithm does not noticeably affect this distribution.
Importantly, the average of the simulated distribution using
our APMC algorithm is not far from the analytical result.
V. C ONCLUSION
We proposed a new a priori Monte Carlo (APMC) algorithm to simulate the probability that a molecule is absorbed
by a spherical receiver in MC systems. Based on numerical
results, we confirmed that our APMC algorithm produces a
more accurate result for the fraction of molecules√absorbed
than the existing RMC and SMC algorithms when D∆t/rr
is large. This demonstrates that our algorithm is suitable to
enable an efficient simulation with a relatively large diffusion
time step length.
R EFERENCES
[1] T. Nakano, A. W. Eckford, and T. Haraguchi, Molecular Communication. Cambridge University Press, 2013.
[2] H. C. Berg, Random Walks in Biology. Princeton University Press,
1993.
[3] H. B. Yilmaz, A. C. Heren, T. Tugcu, and C. B. Chae, “Threedimensional channel characteristics for molecular communications with
an absorbing receiver,” IEEE Commun. Lett., vol. 18, no. 6, pp. 929–932,
June 2014.
[4] Y. Deng, A. Noel, M. Elkashlan, A. Nallanathan, and K. C. Cheung,
“Modeling and simulation of molecular communication systems with
a reversible adsorption receiver,” IEEE Trans. Mol. Biol. Multi-Scale
Commun., vol. 1, no. 4, pp. 347–362, Dec. 2015.
[5] A. Ahmadzadeh, H. Arjmandi, A. Burkovski, and R. Schober, “Comprehensive reactive receiver modeling for diffusive molecular communication systems: Reversible binding, molecule degradation, and finite
number of receptors,” IEEE Trans. Nanobiosci., vol. 15, no. 7, pp. 713–
727, Dec. 2016.
[6] H. B. Yilmaz and C.-B. Chae, “Simulation study of molecular communication systems with an absorbing receiver: Modulation and ISI
mitigation techniques,” Simulation Model. Practice. Th., vol. 49, pp.
136–150, Dec. 2014.
[7] I. Llatser, D. Demiray, A. Cabellos-Aparicio, D. T. Altilar, and
E. Alarcón, “N3sim: Simulation framework for diffusion-based molecular communication nanonetworks,” Simulation Model. Practice. Th.,
vol. 42, pp. 210–222, Mar. 2014.
[8] A. Noel, K. C. Cheung, R. Schober, D. Makrakis, and A. Hafid,
“Simulating with AcCoRD: Actor-based communication via reaction–
diffusion,” Nano Commun. Netw., vol. 11, pp. 44–75, Mar. 2017.
[9] D. Arifler and D. Arifler, “Monte Carlo analysis of molecule absorption
probabilities in diffusion-based nanoscale communication systems with
multiple receivers,” IEEE Trans. Nanobiosci., vol. 16, no. 3, pp. 157–
165, Apr. 2017.
[10] S. S. Andrews and D. Bray, “Stochastic simulation of chemical reactions
with spatial resolution and single molecule detail,” Phys. Biol., vol. 1,
no. 3, p. 137, Aug. 2004.
[11] K. Schulten and I. Kosztin, “Lectures in theoretical biophysics,” University of Illinois, vol. 117, 2000.
| 7cs.IT
|
1
Threshold Phenomena for Interference with
Randomly Placed Sensors
arXiv:1611.06329v2 [cs.DS] 4 May 2017
Rafał Kapelko
Abstract—This paper investigates the problem of efficient displacement of random sensors where good communication within the
network is provided, there is no interference between sensors and the movement cost is minimized in expectation.
Fix d ∈ N \ {0}. Assume n sensors are initially placed in the hyperoctant [0, ∞)d according to d identical and independent Poisson
processes each with arrival rate n1/d . Let 0 < s ≤ v be given real numbers. We are allowed to move the sensors, so that every two
consecutive sensors are placed at distance greater than or equal to s and less than or equal to v. When a sensor is displaced a
distance equal to m, the cost of movement is proportional to some (fixed) power a > 0 of the m distance travelled. We consider a−total
movement (see Definition 2) for one dimension and d−dimensional a−total movement (see Definition 3) for the higher dimension.
Motivation for using this extended cost metric arises from the fact that the parameter a in the exponents represents various conditions
on the barrier: lubrication, obstacles, etc. The main contribution of this paper can be summarized as follows.
1)
2)
If the sensors are displaced in the half-infinite interval [0, ∞), the tradeoffs between the interference distances s, v and the
expected minimum a−total movement are derived. We discover a sharp decline and a sharp increase in the expected minimum
1
a−total movement around the interference distance s, v equal to n
.
For the sensors in the hyperoctant [0, ∞)d (d > 1), the tradeoffs between the interference distances s, v and the expected
minimum d−dimensional a−total movement are obtained. We also discover a sharp decline and a sharp increase in the
1
expected minimum d−dimensional a−total movement around the interference distance s, v equal to 1/d
.
n
Index Terms—Interference, Analysis of algorithms, Random, Sensors, Poisson process
✦
1
I NTRODUCTION
Mobile sensors are being deployed for detecting and monitoring events which occur in many instances of every day life.
However, it is often their case that monitoring may not be as
effective due to external factors such as harsh environmental
conditions, sensor faults, geographic obstacles, etc. In such cases
sensor realignments may be required, e.g., sensors must be relocated from their initial positions to new positions so as to attain
the desired communication characteristics. The resulting problem
is assigning final positions to the sensors in order to minimize the
reallocation cost.
Fix d ∈ N \ {0}. The present paper is concerned with random
realignments of sensors in the hyperoctant. Assume that n sensors
are initially placed in the hyperoctant [0, ∞)d according to d
identical and independent Poisson processes each with arrival rate
n1/d . Given that the sensors are initially placed on the domain
at random according to some well-defined random process, we
are interested to ensure that by moving the sensors the following
scheduling requirement is satisfied.
Definition 1 ((s, v) − IP ). The (s, v)−interference problem
requires that every two consecutive sensors1 are placed at distance
greater than or equal to s and less than or equal to v for some
s, v such that 0 < s ≤ v.
It is very well known that proximity between sensors affects
•
R. Kapelko is with the Department of Computer Science, Faculty of
Fundamental Problems of Technology, Wrocław University of Science and
Technology, Poland.
E-mail: [email protected].
Manuscript received ???? ??, ????; revised ???? ??, ????.
1. The precise meaning of consecutive sensors in the higher dimension
(d > 1) will be explain further in Section 4.
transmission and reception signals and causes degradation of performance (see [6]). The closer the distance between neighbouirng
sensors, the higher the resulting interference. Additionally, to
ensure a good communication or connection of the whole network
the sensors can not be too far from each other.
Thus, in (s, v) − IP problem the goal is to ensure a good
communication of the whole network while at the same time the
neighbouring sensors are not too close.
The initial placement of the sensors does not guarantee (s, v)−
IP interference requirement since the sensors have been placed
randomly according to the arrival times of Poisson processes. To
attain the requirements that any pair of sensors are separated by a
distance of at least s and no two consecutive sensors are placed at
distance greater than v, the sensors have to move from their initial
location to new position.
For the case of the half-infinite interval [0, ∞) we define the
cost measure a−total movement as follows.
Definition 2 (a−total movement). Let a > 0 be a constant.
Suppose that the i−th sensor’s displacement is equal
P |m(i)|. The
a−total movement is defined as the sum Ma := ni=1 |m(i)|a .
Motivation for this extended cost metric arises from the fact
that the cost of individual sensor displacement may not be linear
in this displacement, but rather be dependent on some power
of the distance traversed. Moreover, the parameter a may well
represent various conditions on the barrier: obstacles, lubrication,
etc. Therefore, the a−total movement is a more realistic metric
than the one previously considered a = 1.
Let d be a natural number greater than 1. For the case
of the hyperoctant [0, ∞)d we define below the concept of
d−dimensional a−total movement which refers to a movement
2
of sensors only along the axes.
Definition 3 (d−dimensional a−total movement). Let a > 0
be a constant. Consider a sensor Si located in the position
(y1 (i), y2 (i), . . . , yd (i)), where y1 (i), y2 (i), . . . , yd (i) ∈ R+ .
We move the sensor Si to the position (y1 (i) + m1 (i), y2 (i) +
m2 (i), . . . , yd (i) + md (i)). Then the d−dimensional a−total
movement is defined as the double sum Md,a
:=
Pn Pd
a
|m
(i)|
.
j
j=1
i=1
The main question we study in this paper is the following. Assume that n sensors are initially placed in the hyperoctant [0, ∞)d
according to d identical and independent Poisson processes each
with arrival rate n1/d . What is the expected minimum a−total
movement on the line and d−dimensional a−total movement in
the higher dimension so as to solve the problem (s, v) − IP as a
function of the parameter s, v, n, a, d? Our goal is to investigate
tradeoffs arising among the parameters s, v, n, a, d.
Let us fix ǫ, τ > 0 arbitrary small constants independent
on the number of sensors. For the case of the half-infinitive
interval [0, ∞) we explain the threshold phenomena around the
interference distances s = v = n1 . We discover
•
•
a sharp decline in the expected minimal a−total movement
1
as s decreases from n1 to 1−ǫ
n and v increases from n to
1+τ
n ,
a sharp increase in the expected minimal a−total movement as s increases from n1 to 1+ǫ
n and v increases from
1+τ
1
n to n .
exact asymptotics) which explain the tradeoffs arising among the
parameters s, v, n, a, d.
1.2 Preliminaries and Model
In this subsection we recall some useful properties of the Poisson
process. Consider n sensors initially placed in the half-infinite
interval [0, ∞) according to Poisson process with arrival rate
n. Assume that, the i−th event represents the location of the
i−th sensor, for i = 1, 2, . . . , n. Let Xi be the arrival time of
the i−th event in this Poisson process, i.e., the position of the
i−th sensor in the interval [0, ∞). We know that the random
variable Xi obeys the Gamma distribution with parameters i, n. Its
(nt)i−1
probability density function is given by fi,n (t) = ne−nt (i−1)!
R∞
(nt)i−1
and Pr [Xi ≥ t] = t ne−nt (i−1)! . Moreover, the probability
density function of random variable Xj+l − Xj = Xl is given by
the formula
(nt)l−1
fl,n (t) = ne−nt
(1)
(l − 1)!
(see [9], [10], [16] for additional details on the Poisson process).
Notice that,
Z ∞
fl,n (t)dt = 1,
(2)
0
Z
d
For the case of the hyperoctant [0, ∞) the threshold phenomena
1
around the interference distances s = v = n1/d
is discovered for
the expected minimal d−dimensional a−total movement.
1.1 Related Work
Interference has been the subject of extensive interest in research
community in the last decade. Some papers study interference in
relation to network performance degradation [6], [8]. Moscibroda
et al. [14] consider the average interference problem while maintaining the desired network properties such as connectivity, multicast trees or point-to-point connections, while in [1] the authors
propose connectivity preserving and spanner constructions which
are interference optimal. The interference minimization in wireless
ad-hoc networks in a plane was studied in [7]. Further, [12],
[13] investigates the problem of optimally determining sourcedestination connectivity in random networks. The asymptotic
analysis of interference problem on the line using queueing theory
was provided in [11].
More importantly, our work is related to the paper [2] where
the authors consider the expected minimum total displacement
required so that every pair of sensors are in their final positions
at distance greater or equal to s for n sensors placed uniformly
according to Poisson process.
Compared to the minimum distance between two sensors, the
(s, v) − IP scheduling requirement not only avoids interference,
but also ensures good connectivity and is more reasonable when
considering interference. Our analysis also generalizes the result
of the paper [2] from a = 1 to all exponents a > 0 for more
realistic expected minimum a−total movement on the line and
d−dimensional a−total movement in the higher dimension. It is
worth mentioning that some asymptotic bounds in [2] are onesided. We give full asymptotic results (lower and upper bound,
∞
tfl,n (t)dt =
0
l
,
n
(3)
where l, n are positive integers ( see [4, Chapter 15]).
Finally, we derive the following definite integral
Z
∞
z
∞
l−1+b
X (nt)j
(l
−
1
+
b)!
e−nt
tb fl,n (t)dt = −
(l − 1)!nb
j!
j=0
z
=
(l − 1 + b)! −nz
e
(l − 1)!nb
l−1+b
X
j=0
j
(nz)
.
j!
(4)
where l, n, b are positive integers.
1.3 Outline and Results of the Paper
Throughout this paper ǫ, τ > 0 are arbitrary small constants
independent on n.
Fix d ∈ N \ {0}. Assume n sensors are initially placed in
the hyperoctant [0, ∞)d according to d identical and independent
Poisson processes each with arrival rate n1/d . We want to have the
sensors moving from their current random locations to positions
to ensure (s, v) − IP interference requirement.
For the case of one dimension we derive tradeoffs between
the expected minimum a−total movement and the interferences
distances s, v. Table 1 summarizes the results proved in Section
3.2
2. We recall the following asymptotic notation: (i) f (n) = O(g(n)) if
there exists a constant c1 > 0 and integer N such that |f (n)| ≤ C1 |g(n)| for
all n > N, (ii) f (n) = Ω(g(n)) if there exists a constant c2 > 0 and integer
N such that |f (n)| ≥ C2 |g(n)| for all n > N, (iii) f (n) = Θ(g(n)) if and
only if f (n) = O(g(n)) and f (n) = Ω(g(n)).
3
TABLE 1
The expected minimum a−total movement of n sensors in the
half-infinite interval [0, ∞) as a function of the interference distances
s, v.
Interference
distances
s, v
s = 1−ǫ
n
v = 1+τ
n
0 < ǫ, τ
1
n
1
n
1+ǫ
n
1+τ
n
s=
v=
s=
v=
0<ǫ≤τ
Expected minimum
a−total movement
O n1−a
a
Θ n1− 2 , a ≥ 1,
a
O n1− 2 , a ∈ (0, 1)
Θ (n) , a ≥ 1,
O (n) , a ∈ (0, 1)
Algorithm
I1 n,
1−ǫ 1+τ
, n
n
M V1 n,
M V1 n,
1
n
1+ǫ
n
Let us consider case a = 2. We prove the following results.
•
•
•
For interference distances s = v = n1 the expected
minimal 2−total movement is in Θ(1).
1
1+τ
1
If s = 1−ǫ
n is below n and v =
n is above n , the
expected
minimal 2−total movement declines sharply to
O n1 .
1+τ
If both interference distances s = 1+ǫ
n and v =
n
are above n1 , the expected minimal 2−total movement
increses sharply to Θ(n).
Similar sharp decline and increase hold for all exponent a > 0.
Thus, our investigations explain the threshold phenomena around
the interference distances equal to n1 as this affects the expected
minimum a−total movement of the sensors to fulfill (s, v) − IP
interference requirement.
For the sensors placed in the higher dimension (d > 1) we
obtain tradeoffs between d−dimensional a−total movement and
the interferences distances s, v. Table 2 summarizes the results
obtained in Section 4.
TABLE 2
The expected minimum d−dimensional a−total movement of n
sensors in the hyperoctant [0, ∞)d as a function of the interference
distances s, v.
•
1+ǫ
√
n
1+τ
√
n
If both interference distances s =
and v =
are
1
above √n , the expected minimal 2−dimensional 2−total
movement increases sharply to Θ(n).
Similar sharp decrease and increase hold in all dimensions
(d ∈ N \ {1}) and for all exponents a > 0. Hence, our
investigations explain the threshold phenomena around the in1
terference distances equal to n1/d
as this affects the expected
minimum d−dimensional a−total movement of the sensors to
fulfill (s, v) − IP interference requirement.
Here is an outline of the paper. In Section 2 we provide several
combinatorical facts that will be used in the sequel. In Section 3
we investigate sensors on the line. We show
that the expected
a−total movement of algorithm M V1 n, n1 is
a
a!
a
n1− 2 + O n− 2 ,
a
a
22 2 + 1 !
when a is an even natural number. Further, we analyze the expected minimal a−total movement near the interference distances
s = v = n1 . In Section 4 we investigate sensors in the higher
dimensions.
2
BASIC FACTS AND N OTATIONS
In this section we recall some known facts about special functions
and special numbers which will be useful in the analysis in the
next sections. We also prove Lemma 4 which will be helpful in
the analysis of Algorithm 1.
We will use the following notations for the rising factorial [5]
(
1
for k = 0
k
n =
n(n + 1) . . . (n + k − 1) for k ≥ 1.
n
Let k be the Stirling numbers of the first, which are defined for
all integer numbers such that 0 ≤ k ≤ n.
The Stirling numbers of the first kind arise as coefficients of
the rising factorial (see [5, Identity 6.11])
" #
X m
m
x =
(5)
xl2 .
l
2
l
2
n
k
Interference
distances
s, v
s = 1−ǫ
n1/d
v = 1+τ
n
0 < ǫ, τ
s=
v=
1
n1/d
1
n1/d
1+ǫ
n1/d
1+τ
n1/d
s=
v=
0<ǫ≤τ
Expected minimum
d−dimensional
a−total movement
a
O n1− d
Algorithm
Id n,
1−ǫ
, 1+τ
n1/d n1/d
a
Θ n1− 2d , a ≥ 1,
a
O n1− 2d , a ∈ (0, 1)
M Vd n,
1
n1/d
Θ (n) , a ≥ 1,
O (n) , a ∈ (0, 1)
M Vd n,
1+ǫ
n1/d
•
Let us recall the definition of a finite difference of a function f
∆f (x) = f (x + 1) − f (x).
Let us consider cases d = 2 and a = 2. We prove the
following results.
•
be the Eulerian numbers of the second kind, which
Let
are defined for all integer numbers such that 0 ≤ k ≤ n. The
following two identities for Eulerian numbers of the second kind
are known (see Identities (6.42), and (6.44) in [5]):
X DD m EE
(2m)! 1
=
,
(6)
l
(m)! 2m
l
"
#
!
X DD p EE m + l
m
=
.
(7)
m−p
l
2p
l
√1
n
the expected
For interference distances s = v =
√
minimal 2−dimensional 2−total movement is in Θ ( n) .
√
√
If s = 1−ǫ
is below √1n and v = 1+τ
is above √1n ,
n
n
the expected minimal 2−dimensional 2−total movement
declines sharply to O (1) .
Then, high-order differences are defined by iteration
∆a f (x) = ∆∆a−1 f (x).
It is easy to prove by induction the following formula (see also [5,
Identity 5.40])
!
X a
a
(−1)a−j f (x + j).
(8)
∆ f (x) =
j
j
4
A crucial observation is the following identity, which will be
useful in the asymptotic analysis of Algorithm 1 when interference
distances s, v are equal to n1 .
Lemma 4. Assume that a is an even positive number. Then
!
"
#
0
X a
if 2l1 < a
j
(−1)j
=
a!
if
2l1 = a.
a
a !2 2
j
j − l1
(2)
j
x
Proof. Choosing f (x) = x−l
in (8) we see that
1
"
#
!
"
#
X a
x
x+j
a
a−j
∆
=
(−1)
x − l1
j
x + j − l1
j
x=0
x=0
!
"
#
X a
j
=
(−1)j
.
j
j − l1
j
Let f be non-negative integer. Then
n
X
i=2
3
This is enough to prove Lemma 4.
We will also use many times Jensen’s inequality for expectactions. If f is a convex function, then
(9)
provided the expectations exists (see [16, Proposition 3.1.2]).
The following inequality for the general random variable will
be useful in the threshold tight bounds when interference distances
are greater or equal to n1 .
If Z is the random variable such that E[|Z|] < ∞ and q ∈ R,
then
|E[Z] − q| ≤ E[|Z − q|] ≤ E[|Z|] + |q|.
(10)
Notice that the left side of Inequality (10) is the special case of
Jensen’s inequality for X := Z −q and f (x) = |x|. The right side
of Inequality (10) follows from the triangle inequality |Z − q| ≤
|Z| + |q| and the monotonicity of the expected value.
We will also use the following notation
|x|+ = max{x, 0}
for positive parts of x ∈ R.
The Euler Gamma function (see [15])
Z ∞
Γ(z) =
tz−1 e−t dt
(11)
(12)
0
is defined for z > 0. Notice that Γ(z) satisfies the following
functional equations
Γ(z + 1) = Γ(z)z.
Moreover, for n natural number we have
Γ(n + 1) = n!.
(i − 1)f =
f
X
1
cl n l ,
nf +1 +
f +1
l=0
(15)
where cl are some constants independent on n (see [5, Formula
(6.78)]).
Applying equations
(6), (7) and the following identity
(
0
if
2l1 < a
x+l
∆a 2l1
=
, we easily derive
x=0
1 if 2l1 = a
!
"
#
X l1
x
a x+l
a
∆
∆
=
l
2l1
x − l1
l
x=0
x=0
0
if
2l1 < a
DD a EE
= P
a!
2
= a a2 if 2l1 = a.
l
l
( 2 )!2
f (E[X]) ≤ E [f (X)]
We will also use the following forms of Stirling’s formula (see
[3, page 54])
√
√
1
1
1
1
2πN N + 2 e−N + 12N +1 < N ! < 2πN N + 2 e−N + 12N . (14)
(13)
S ENSORS ON THE L INE
Let us recall that ǫ, τ > 0 are arbitrary small constants independent on n. In this section we analyze (s, v) − IP interference
problem when the sensors are placed in the half-infinite interval
[0, ∞) according to Poisson process with arrival rate n.
3.1 Analysis of Algorithm 1
In this subsection we present algorithm M V1 (n, s) (see Algorithm
1). We show
that the expected a−total movement of algorithm
M V1 n, n1 is
2
a
2
a
a!
a
n1− 2 + O n− 2 ,
+1 !
a
2
when a is an even positive number.
We prove that the expected a−total movement of algorithm
M V1 n, 1+ǫ
is Θ(n), when a is an even positive number or
n
a = 1 and ǫ > 0.
Algorithm 1 M V1 (n, s) Moving sensors in the [0, ∞); s > 0.
Require: The initial location X1 ≤ X2 ≤ · · · ≤ Xn of the n
sensors in the [0, ∞) according to Poisson process with arrival
rate n.
Ensure: The final positions of the sensors such that each pair of
consecutive sensors is separated by the distance equal to s.
1: for i = 2 to n do
2:
move the sensor Xi at the position X1 + (i − 1)s;
3: end for
We prove the following theorem.
Theorem 5. Fix ǫ1 ≥ 0 independent on n. Let a be an even
natural number.
The expected a−total movement of algorithm
1
is respectively
M V1 n, 1+ǫ
n
a
a
a a!
when ǫ1 = 0,
n1− 2 + O n− 2
a
2
2 ( 2 +1)!
Θ (n)
when ǫ1 > 0.
Proof. Let Xi be the arrival time of the ith event in a Poisson
process with arrival rate n. We know that the random variable
Xi − X1 = Xi−1 obeys the Gamma distribution with density
fi−1,n (t) = ne−nt
(nt)i−2
(i − 2)!
for i = 2, 3, . . . , n. (see equation (1) for j = 1, l = i − 1).
(a)
Assume that, a is even natural number. Let Di be the expected
5
(1)
Di
distance to the power a between Xi − X1 and the i−th sensor
position, ti−1 = (1 + ǫ1 ) i−1
n , hence given by
Z ∞
(a)
Di =
|t − ti−1 |a fi−1,n (t)dt
0
Z ∞
(ti−1 − t)a fi−1,n (t)dt.
=
be the expected distance between Xi − X1 and
Proof. Let
the i−th sensor position, ti−1 = (1 + ǫ) i−1
n for i = 2, 3, . . . , n,
hence given by
Z ∞
(1)
|t − ti−1 |fi−1,n (t)dt.
Di =
Let j ∈ {0, . . . , a}. Applying Identity (5) we deduce that
This completes the proof of Theorem 6.
0
Let us recall that E [Xi − X1 ] = i−1
n (see (3) for l = i − 1).
Applying Inequality (10) for Z = Xi − X1 , q = ti−1 and ǫ > 0
Observe that
!
Z ∞
we have
X a
i − 1 a−j
i−1
i−1
(a)
j
(1)
(−1)
tj fi−1,n (t)dt.
(1 + ǫ1 )
Di =
ǫ
≤ Di ≤ (2 + ǫ)
.
n
j
n
n
0
j
Pn
(n−1)n
, we have
Since i=2 (i − 1) =
Using the integral Identities (2) and (3) we see that
2
!
n
X (1)
1 X a
(i + j − 2)!
(a)
(16)
Di = Θ(n)
Di = a
(1 + ǫ1 )a−j (−1)j (i − 1)a−j
.
n j
(i − 2)!
j
i=2
0
(i + j − 2)!
(i − 1)a−j
= (i − 1)a−j (i − 1)j
(i − 2)!
"
#
X
j
(i − 1)a−l1 .
=
j
−
l
1
l
1
Hence
(a)
Di
"
!
#
1 XX a
j
= a
(1 + ǫ1 )a−j (−1)j (i − 1)a−l1
.
j − l1
j
n j l
1
Changing the summation we get
(a)
Di
!
"
#
X a
j
1 X
a−j
j
a−l1
(1 + ǫ1 ) (−1)
(i − 1)
.
= a
j
j − l1
n l
j
1
Now, we will estimate separately when ǫ1 = 0 and when ǫ1 > 0.
Case ǫ1 = 0.
Applying Lemma 4 we get
a
1 X
a!
1
(a)
C1,a−l1 · (i − 1)a−l1 ,
Di = a a a · (i − 1) 2 + a
n 2 !2 2
n 2l >a
3.2 Expected Minimum a−total Movement for s = v =
1
n
In this subsection we look at the expected minimum a−total
movement when the interference distances
s and v are equal to
1
1− a
2
.
We
prove
the
upper
bound
O
n
on the expected a−total
n
movement, when a > 0 (see Theorem 7). Our
Theorem 8 and
a
Theorem 9 give the lower bound Ω n1− 2 on the expected
a−total movement, when a ≥ 1.
We begin with a theorem which indicates how to apply the
results of Theorem 5 to the upper bound on the expected a−total
movement, when a > 0.
Theorem 7. Let a >
movement of
0. The expected a−total
a
algorithm M V1 n, n1 is respectively O n1− 2 .
(a)
Proof. Assume that a > 0. Let Di be the expected distance to
the power a between Xi − X1 and the ith sensor position. Let
b be the even natural number such that b − a > 0. Then we use
b
discrete Hölder inequality with parameters ab and b−a
and get
1
where C1,a−l1 depends only on a and l1 . Using Identity (15) we
conclude that the expected
sum of displacements to the power a
of algorithm M V1 n, n1 is
n
X
i=2
(a)
Di
=
2
a
2
a
a!
n1− 2 + O n− 2 .
a
2 +1 !
a
This is enough to prove the case
ǫ1 = 0.
Pwhen
Case ǫ1 > 0. Observe that j aj (−1)j (1+ǫ1 )a−j jj = ǫa1 .
Therefore
1 X
ǫa
(a)
C2,a−l1 · (i − 1)a−l1 ,
Di = 1a · (i − 1)a + a
n
n l >0
1
where C2,a−l1 depends only on a and l1 . Again, applying Identity
(15) we conclude that the expected sum
of displacements to the
1
power a of algorithm M V1 n, 1+ǫ
is Θ(n). This is enough
n
to prove the case when ǫ1 > 0 which completes the proof of
Theorem 5.
n
X
i=2
(a)
Di
≤
=
n
b
X
(a) a
Di
i=2
n
b
X
(a) a
Di
i=2
! ab
! ab
n
X
i=2
1
! b−a
b
(n − 1)
b−a
b
.
(17)
b
Next we use Jensen’s inequality (see (9)) for f (x) = x a and
(a)
E[X] = Di and get
b
(a) a
(b)
Di
≤ Di .
(18)
Combining together (17), (18) and Theorem 5 we deduce that
ab
n
X
b−a
a
1
(a)
Di ≤ Θ
(n − 1) b = Θ n1− 2 .
b
n 2 −1
i=2
This is enough to prove the upper bound which finishes the proof
of Theorem 7.
We also prove the following tight bound for 1−total movement
of Algorithm 1 when s = 1+ǫ
n and ǫ > 0.
We now prove the desired lower bound for expected 1−total
movement.
Theorem 6. Let ǫ > 0 be a constant independent on n. Then
is
the expected 1−total movement of algorithm M V1 n, 1+ǫ
n
respectively Θ(n).
Theorem
8. Any sensor’s displacement algorithm which solves
1 1
,
n n −
√IP problem requires expected 1−total movement of at
least Ω ( n) .
6
Proof. Before providing the proof of the theorem we make two
important observations.
Let X1 < X2 < · · · < Xn be the initial positions of the
sensors. Recall that by the monotonicity lemma no sensor Xi is
ever placed before sensor Xj , for all i < j.
We assume that the final position of the first sensor is the
nonnegative random variable Z with E[Z] < ∞.
We are now ready to prove the theorem. Let Xi be the arrival
times of the i−th event in Poisson process with arrival rate. Let
1
ai = ni − 2n
, for i = 1, 2, . . . , n be the anchor points. The
following result was proved in [2, Lemma 1]
n
X
i=1
E [|Xi − ai |] = C1
√
√
n +o n ,
(19)
where C1 is some constant independent on n. There are two cases
to consider.
Case 1. The algorithm moves the sensor Xi to the position
Z + bi , where bi = n1 (i − 1), for i = 1, 2, . . . n and E[Z] >
C1
1√
2 n.
Combining together Inequality (10) for Z := Xi − Z , q = bi
Equation (3) for l = i and the triangle inequality we get
n
X
i=1
E [|Xi − Z − bi |] ≥
n
X
i=1
E[Z] −
This is enough to prove the first case.
Case 2. The algorithm moves the sensor Xi to the position
Z + bi , where bi = n1 (i − 1), for i = 1, 2, . . . n and E[Z] ≤
C1
1√
2 n.
i
n
be the expected distance
Proof. Assume that a > 1. Let
to the power a of i−th sensor for i = 1, 2, . . . , n. Then we use
a
discrete Hölder inequality with parameters a and a−1
and get
n
X
(1)
Ei
n
X
(1) a
Ei
≤
i=1
i=1
n
X
(1) a
Ei
=
i=1
! a1
! a1
n
X
1
i=1
n
a−1
a
! a−1
a
.
(20)
Next we use Jensen’s inequality (see (9)) for f (x) = xa and
(1)
E[X] = Ei and get
(1) a
(a)
Ei
≤ Ei .
(21)
Combining together (20), (21) and Theorem 8 we deduce that
!a
n
n
X
X
√ a −a+1
(a)
(1)
n
Ei ≥
Ei
n−a+1 = Ω n
i=1
i=1
a
= Ω n1− 2 .
This is enough to prove the lower bound and completes the proof
of Theorem 9.
3.3 Expected Minimum a−total Movement for
v
1
n
n
X
√
1
1 C1
√ −
≥
=Θ n .
2
n
n
i=1
Let us recall that ai =
triangle inequality
(a)
Ei
1
− 2n
, for i = 1, 2, . . . , n. Using the
|Xi − ai | ≤ |Xi − (Z + bi )| + |(Z + bi ) − ai |
1
n
<s≤
In this subsection we study the expected minimum a−total
movement when the interference distances s and v are greater
than n1 . We give the upper bound O(n) on the expected a−total
movement, when a > 0 (see Theorem 10) and the lower bound
Ω(n) on the expected a−total movement, when a ≥ 1 (see
Theorem 11 and Theorem 12).
We begin with a theorem which indicates how to apply the
results of Theorem 5 to the upper bound on the expected a−total
movement, when a > 0.
Theorem 10. Let ǫ > 0 be a constant independent on n. Let a >
0. The expected a−total movement of algorithm M V1 n, 1+ǫ
is
n
in O(n).
(a)
we get
n
X
i=1
E [|Xi − (Z + bi )|] ≥
n
X
i=1
E [|Xi − ai |]−
n
X
i=1
E Z−
1
.
2n
C1
and
Combining together Equation (19), assumption E[Z] ≤ 12 √
n
1
1
the triangle inequality E Z − 2n ≤ E[Z] + 2n we have
n
X
i=1
E [|Xi − (Z + bi )|] ∈ Ω(n).
This is enough to prove the second case and sufficient to complete
the proof of Theorem 8.
Proof. Let Di be the expected distance to the power a between
Xi − X1 and the ith sensor position. Let b be the even natural
number such that b − a > 0. Then we proceed as in the upper
bound treatment from the proof of Theorem 7 and get
n
X
(a)
Di
i=2
≤
n
b
X
(a) a
Di
i=2
(a)
Di
ab
! ab
(n − 1)
b−a
b
(b)
≤ Di
(22)
(23)
Combining together (22), (23) and Theorem 5 we deduce that
n
X
i=2
(a)
Di
a
≤ (Θ(n)) b (n − 1)
b−a
b
= Θ(n).
This is sufficient to complete the proof of Theorem 10.
We now apply Theorem 8 in order to derive the lower bound
on the expected a−total movement when a > 1.
We can now prove the desired lower bound for expected
1−total movement.
Theorem 9. Let a > 1. Then any sensor’s displacement algorithm
which solves n1 , n1 − IP problem
requires expected a−total
a
movement of at least Ω n1− 2 .
Theorem 11. Fix τ ≥ ǫ > 0 independent on n. Then
any sensor’s
1+τ
displacement algorithm which solves 1+ǫ
− IP problem
,
n
n
requires expected 1−total movement of at least Ω(n).
7
Proof. The proof of the theorem is analogous to the proof of
Theorem 8.
Fix τ ≥ ǫ > 0 independent on n. Let Xi be the arrival
times of the i−th event in Poisson process with arrival rate. We
assume that the algorithm moves the sensor Xi to the position
i
bi = Z + 1+∆
n (i − 1) provided ǫ ≤ ∆i ≤ τ for i = 1, 2, . . . , n,
where Z is the nonnegative random variable with E[Z] < ∞.
It is sufficient to show that
n
X
i=1
E [|Xi − bi |] ∈ Ω(n).
Applying Inequality (10) for Z := Xi −Z −∆i i−1
n , q =
and Equation (3) for l = i we get
n
X
1 + ∆i
E Xi − Z +
(i − 1)
n
i=1
i−1
n
n
X
E[∆i ]
1 + E[∆i ]
i + E[Z] −
n
n
i=1
n
X E[∆i ]
1 + E[∆i ]
≥
i + E[Z] −
n
n
i=1
n
X ǫ
1+τ
= Θ(n)
i + E[Z] −
≥
n
n
i=1
≥
This completes the proof of Theorem 11.
We now apply Theorem 11 in order to derive the lower bound
on the expected a−total movement when a > 1.
Theorem 12. Fix τ ≥ ǫ > 0 independent on n. Let a > 1. Then
1+τ
−
any sensor’s displacement algorithm which solves 1+ǫ
n , n
IP problem requires expected a−total movement of at least Ω(n).
Algorithm 2 I1 (n, s, v) Moving sensors in the [0, ∞); 0 < s <
v.
Require: The initial location X1 ≤ X2 ≤ · · · ≤ Xn of the n
sensors in the [0, ∞) according to Poisson process with arrival
rate n.
Ensure: The final positions of the sensors such that
∀i=2,3...,n v ≥ Xi − Xi−1 ≥ s (so as that the distance
between consecutive sensors is less or equal v and greater or
equal to s).
1: for i = 2 to n do
2:
if Xi − Xi−1 < s then
3:
move left-to-right the sensor Xi at the new position s +
Xi−1 ;
4:
else if Xi − Xi−1 > v then
5:
move right-to-left the sensor Xi at the new position v +
Xi−1 ;
6:
else
7:
do nothing;
8:
end if
9: end for
Lemma 13. Fix ǫ > 0 independent on n. Let a > 0 and let
s = 1−ǫ
n . Assume that random variable Xl obeys the Gamma
distribution with parameters l, n. Then
n
X
n
E (|sl − Xl |+ )a = O(n1−a ).
l
l=1
Proof. See Appendix for the proof.
Lemma 14. Fix τ > 0 independent on n. Let a > 0 and let
v = 1+τ
n . Assume that random variable Xl obeys the Gamma
distribution with parameters l, n. Then
n
X
n
(a)
Proof. Assume that a > 1. Let Ei be the expected distance to
the power a of i−th sensor for i = 1, 2, . . . , n. As in the proof of
Theorem 9 we get two inequalities:
n
X
(1)
Ei
i=1
≤
n
X
(1) a
Ei
i=1
! a1
(1) a
(a)
Ei
≤ Ei .
n
a−1
a
,
(24)
(25)
Combining together (24), (25) and Theorem 11 we deduce that
!a
n
n
X
X
(a)
(1)
a
n−a+1 = (Ω(n)) n−a+1 = Ω(n).
Ei ≥
Ei
i=1
i=1
This is enough to prove the lower bound and completes the proof
of Theorem 12.
3.4 Expected Minimum a−total Movement for s <
and v > n1
1
n
In this subsection we give algorithm I1 (n, s, v) (see Algorithm
2) to solve (s, v) − IP interference problem. Let ǫ, τ > 0
be constants independent on n. Let a > 0. We show
that
1+τ
is in
,
expected a−total movement of algorithm I1 n, 1−ǫ
n
n
O n1−a .
We begin with the following two lemmas which will be helpful
in the proof of Theorem 15.
l=1
l
E (|Xl − vl|+ )a = O(n1−a ).
Proof. See Appendix for the proof.
We can now prove the desired result.
Theorem 15. Fix ǫ, τ > 0 independent on n. Let a > 0.
The
1+τ
is in
,
expected a−total movement of algorithm I1 n, 1−ǫ
n
n
O n1−a .
1+τ
Proof. Let s = 1−ǫ
n , v = n , where ǫ and τ are arbitrary small
constants independent on n.
Firstly, we consider the following scenario.
•
•
•
Algorithm 2 leaves the sensors X1 , X2 , . . . , Xi at the
same positions.
Algorithm 2 moves the sensors Xi+1 , Xi+2 , . . . Xi+q at
the new locations.
Algorithm 2 leaves the sensors Xi+q+1 , Xi+q+2 , . . . , Xn
at the same positions.
Let q = q1 + q2 + · · · + qk for some q1 , q2 , . . . , qk ∈ N+ . Define
(
0 if i = 0
Si =
q1 + q2 + · · · + qi if i ∈ {1, 2, . . . , k − 1}.
Consider an integer configuration (q1 , q2 , . . . , qk ) as specified
above.
•
For each i ∈ {0, 1, . . . , k − 1} Algorithm 2 moves the
sensors XSi +1 , XSi +2 , . . . , XSi +qi+1 in the one chosen
direction left to right or right to left.
8
•
For each i ∈ {0, 1, . . . , k − 2} the movement direction of sensors XSi +1 , XSi +2 , . . . , XSi +qi+1
is opposite to the movement direction of sensors
XSi+1 +1 , XSi+1 +2 , . . . , XSi+1 +qi+2 .
Let T (q1 , q2 , . . . , qk ) be the movement to the power a of the sensors Xi+1 , Xi+2 , . . . Xi+q and let T (qk ) be the displacement to
the power a of the sensors XSk−1 +1 , XSk−1 +2 , . . . , XSk−1 +qk .
There are two cases to consider.
Case 1. The sensor XSk−1 moves left to right to the new
position Y and the sensors XSk−1 +1 , XSk−1 +2 , . . . , XSk−1 +qk
move right to left.
Observe that
qk
X
+ a
XSk−1 +l − (Y + vl)
.
T (qk ) =
l=1
Since XSk−1 < Y we upper bound the displacement T (qk ) as
follows
qk
X
+ a
XSk−1 +l − XSk−1 + vl
.
T (qk ) ≤
l=1
Using Identity Xi+l − Xi = Xl for i := Sk−1 (see (1)) we get
T (qk ) ≤
qk
X
l=1
(|Xl − vl|+ )a .
(26)
Case 2. The sensor XSk−1 moves right to left to the new
position Z and the sensors XSk−1 +1 , XSk−1 +2 , . . . , XSk−1 +qk
move left to right.
Observe that
qk
X
+ a
.
Z + sl − XSk−1 +l
T (qk ) =
Observe that the displacements E [(|sl − Xl |+ )a ] and |Xl −
vl|+ )a can appear in the Algorithm 2 at most nl times. Therefore
Ta ≤
n
X
n
l=1
l
(|sl − Xl |+ )a +
n
X
n
l=1
l
(|Xl − vl|+ )a
Passing to the expectations we have
E [Ta ] ≤
n
X
n
E (|sl − Xl |+ )a +
E (|Xl − vl|+ )a
l
l
l=1
n
X
n
l=1
1+τ
Finally, using Lemma 13 for s = 1−ǫ
n and
Lemma 14 for v = n
1−a
we conclude that E [Ta ] = O n
. This is enough to prove
Theorem 15.
4
S ENSORS IN THE H IGHER D IMENSION .
Fix d ∈ N \ {0}. Let us recall that ǫ, τ > 0 are arbitrary small
constants independent on n.
We consider n sensors that are placed in the hyperoctant
[0, ∞)d according to d identical and independent Poisson pro(d)
(2)
(1)
cesses Xi , Xi , . . . , Xi , for i = 1, 2, . . . , n1/d each with
1/d
arrival rate n . The position of a sensor in the R+ is de(d)
(2)
(1)
termined by the d coordinates (Xi1 , Xi2 , . . . , Xid ), where
1 ≤ i1 , i2 , . . . , id ≤ n1/d .
Hence, we have initially n(d−1)/d rows and n(d−1)/d columns
such that each column and each row has n1/d sensors. Figure 4
illustrates our random displacement of sensors in two dimensions.
(2)
s
s
s
s
(2)
s
s
s
s
(2)
s
s
s
s
(2)
s
s
s
s
X4
X3
l=1
Since Z < XSk−1 we upper bound the displacement T (qk ) as
follows
qk
X
+ a
.
XSk−1 + sl − XSk−1 +l
T (qk ) =
X2
l=1
X1
Using Identity Xi+l − Xi = Xl for i := Sk−1 (see (1)) we get
T (qk ) ≤
qk
X
l=1
(|sl − Xl |+ )a .
(27)
Combining together inequalities (26) and (27) we have
T (q1 , q2 , . . . , qk ) ≤T (q1 , q2 , . . . , qk−1 ) +
+
qk
X
l=1
qk
X
(|Xl − vl|+ )a
l=1
(|sl − Xl |+ )a .
Hence, by induction we get
(1)
≤
j=1 l=1
(|Xl − vl| ) +
qj
k X
X
j=1 l=1
+ a
(|sl − Xl | ) .
Next we make an important observation that extends our estimation to general scenario of Algorithm 2. Let Ta be the a−total
displacement of Algorithm 2.
(1)
X4
We reallocate the sensors so as:
(a)
(b)
(c)
+ a
(1)
X3
Fig. 1. The mobile sensors located in the quadrant [0, ∞)2 according to
2 identical and independent Poisson processes.
T (q1 ,q2 , . . . , qk )
qj
k X
X
(1)
X1 X2
the sensors move only along the axes,
we have finally n(d−1)/d rows and n(d−1)/d columns
such that each column and each row has n1/d sensors,
in each column and in each row the sensors satisfy
(s, v) − IP interference requirement.
In order to fulfill the requirements (a), (b), (c) two algorithms
are presented. Namely,
•
1
1
for the case of s = n1/d
, v = n1/d
we show that the
expected d−dimensional
a−
total
movement
of algorithm
a
1
is in O n1− 2d (see proof of Theorem
M Vd n, n1/d
17),
9
•
1−ǫ
,
n1/d
1+τ
n1/d
v =
we prove that the
for the case of s =
expected d−dimensional
a−
total
movement
of algorithm
1+τ
1− a
d
Id n, n1−ǫ
(see
Theorem
19).
,
is
in
O
n
1/d n1/d
Algorithm 3 M Vd (n, s) Moving sensors in the [0, ∞)d , d ≥ 2,
s > 0.
(1)
(2)
(d)
Require: The initial location (Xi1 , Xi2 , . . . , Xid ) of the n
sensors in the [0, ∞)d , 1 ≤ i1 , i2 , . . . , id ≤ n1/d .
Ensure: The final positions of the sensors such that in each
column and in each row the consecutive sensors are separated
by the distance equal to s.
move the sensor at
1: ∀1 ≤ i ,i ,...,i ≤ n1/d
1 2
d
(1)
(d)
(2)
the location (Xi1 , Xi2 , . . . , Xid ) to the
(d)
(2)
(1)
(X1+(i1 −1)s , X1+(i2 −1)s , . . . , X1+(id −1)s );
position
Algorithm 4 Id (n, s, v) Moving sensors in the [0, ∞)d , d ≥ 2,
0 < s < v.
(1)
(2)
(d)
Require: The initial location (Xi1 , Xi2 , . . . , Xid ) of the n
sensors in the [0, ∞)d , 1 ≤ i1 , i2 , . . . , id ≤ n1/d .
Ensure: The final positions of the sensors such that in each
column and in each row the sensors satisfy (s, v) − IP
interference requirement.
1: For each column and row in the [0, ∞)d apply algorithm
I1 (n1/d , s, v);
We call a move of a sensor a sliding move if the final position
of the sensor is either in the same row or column as its initial
position.
In this section, we restrict the movement of sensors to a sliding
movement. The claim is justified in Lemma 16 whose simple proof
is omitted. Such a reduction of the movement is indeed crucial and
reduces the displacement of sensors in the higher dimension to the
displacement of sensors in the half-infinite interval [0, ∞).
Lemma 16. The optimal reallocation of sensors which ensures
the requirements (a), (b), (c) is a sliding movement.
We now embark to extend the results from Section 3 to the high
dimensions. We can prove the following sequences of Theorem.
The next theorem clarifies how the interference distances s =
1
1
, v = n1/d
affect the expected minimum d−dimensional
n1/d
a−total movement.
Theorem 17. Let a > 0 be a constant. Assume that n sensors
are placed in the [0, ∞)d according to d independent identical
Poisson processes, each with arrival rate n1/d and the reallocation of sensors ensures the requirements (a), (b), (c). If the
1
1
interference distances s = n1/d
, v = n1/d
then the expected
minimal d−dimensional a−total movement is in
(
a
Θ n1− 2d
if a ≥ 1,
a
1− 2d
O n
if a ∈ (0, 1).
Proof. First of all, we discuss the proof of the upper bound. By
Theorem 7 applied to n := n1/d and for n(d−1)/d columns and
n(d−1)/d rows we have that the expected
d−dimensional a−total
1
movement of algorithm M Vd n, n1/d
is
1− a2
a
2n(d−1)/d O n1/d
= O n1− 2d , when a > 0.
Next we prove the lower bound. Since the movement of sensors
along the axes is a sliding move to attain the interference distances
1
1
s = n1/d
, v = n1/d
in the [0, ∞)d the sensors have to attain the
1
1
interference distances s = n1/d
, v = n1/d
in each column and
each row. By Theorem 8 and Theorem 9 applied to n := n1/d and
for n(d−1)/d columns and n(d−1)/d rows we have the following
lower bound
1− a2
a
2n(d−1)/d Ω n1/d
= Ω n1− 2d , when a ≥ 1.
This is sufficient to complete the proof of Theorem 17.
We now analyze the expected minimum d−dimensional
a−total movement when the interference distances s and v are
1
.
greater than n1/d
Theorem 18. Fix τ ≥ ǫ > 0 independent on n. Let a > 0
be a constant. Assume that n sensors are placed in the [0, ∞)d
according to d independent identical Poisson processes, each with
arrival rate n1/d and the reallocation of sensors ensures the
requirements (a), (b), (c). If the interference distances are equal to
1+τ
s = n1+ǫ
1/d , v = n1/d , then the expected d−dimensional a−total
movement is in
(
Θ(n) if a ≥ 1,
O(n) if a ∈ (0, 1).
Proof. Fix τ ≥ ǫ > 0 independent on n. The proof is analogous
to the proof of Theorem 17. By Theorem 10 applied to n :=
n1/d and for n(d−1)/d columns and n(d−1)/d rows and have the
expected a−total movement of algorithm M Vd n, n1+ǫ
is
1/d
2n(d−1)/d O n1/d = O(n), when a > 0.
This completes the prove of upper bound.
By Theorem 11 and Theorem 12 applied to n := n1/d and for
(d−1)/d
n
columns and n(d−1)/d rows we have that the following
lower bound
2n(d−1)/d Ω n1/d = Ω (n) , when a ≥ 1.
This is enough to prove Theorem 18.
The next theorem provides the expected minimum
1
1
d−dimensional a−total movement for s < n1/d
and v > n1/d
.
Theorem 19. Fix ǫ, τ > 0 independent on n. Let a > 0 be
a constant. Assume that n sensors are placed in the [0, ∞)d
according to d independent identical Poisson processes, each with
arrival rate n1/d and the reallocation of sensors ensures the
requirements (a), (b), (c). The expected d−dimensional a−total
a
1+τ
movement of algorithm Id n, n1−ǫ
is in O n1− d .
1/d , n1/d
Proof. Fix ǫ, τ > 0 independent on n. By Theorem 15 applied to
n := n1/d and for n(d−1)/d columns and n(d−1)/d rows we have
the following upper bound
1−a
a
2n(d−1)/d O n1/d
= O n1− d , when a > 0
which proves the theorem.
5
C ONCLUSION
In this paper we obtained tradeoffs between interference distances s, v and the expected minimum (a−total
movement)/(d−dimensional a−total movement) of n random
sensors.
10
R EFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
Wattenhofer R. Burkhart, M. and A. Zollinger. Does topology control
reduce interference? In Proceedings of the 5th ACM International
Symposium on Mobile Ad Hoc Networking and Computing, pages 9–19.
ACM, 2004.
Kranakis E. and Shaikhet G. Displacing random sensors to avoid
interference. In COCOON, volume 8591 of LNCS, pages 501–512.
Springer, 2014.
W. Feller. An Introduction to Probability Theory and its Applications,
volume 1. John Wiley, NY, 1968.
Evans M. Hastings N. Peacock B. Forbes, C. Statistical Distributions.
John Wiley & Sons, 2011.
R. Graham, D. Knuth, and O. Patashnik. Concrete Mathematics A
Foundation for Computer Science. Addison-Wesley, Reading, MA, 1994.
P. Gupta and P.R. Kumar. The capacity of wireless networks. IEEE
Transactions on Information Theory, 46(2), 2000.
M. Halldórsson and T. Tokuyama. Minimizing interference of a wireless
ad-hoc network in a plane. Theoretical Computer Science, 402(1), 2008.
Padhye J. Padmanabhan V.N. Jain, K. and L. Qiu. Impact of interference
on multi-hop wireless network performance. Wireless Networks, 11(4),
2005.
J.F.C. Kingman. Poisson Process, volume 3. Oxford University Press,
1992.
E. Kranakis. On the event distance of poisson processes with applications
to sensors. Discrete Applied Mathematics, 179:152 – 162, 2014.
E. Kranakis and G. Shaikhet. Sensor allocation problems on the real line.
Journal of Applied Probability, 53(3):667–687, 2016.
Wang X. L. Fu, L. and P.R. Kumar. Optimal determination of sourcedestination connectivity in random graphs. In Proceedings of the 15th
ACM International Symposium on Mobile Ad Hoc Networking and
Computing, pages 205–214. ACM, 2014.
Wang X. L. Fu, L. and P.R. Kumar. Are we connected? optimal
determination of source-destination connectivity in random networks.
IEEE/ACM Transactions on Networking, PP(99):1–14, 2016.
T. Moscibroda and R. Wattenhofer. Minimizing interference in ad hoc
and sensor networks. In Proceedings of the 2005 Joint Workshop on
Foundations of Mobile Computing, pages 24–33. ACM, 2005.
NIST Digital Library of Mathematical Functions. http://dlmf.nist.gov/5.
S. M. Ross. Probability Models for Computer Science. Academic press,
2002.
11
A PPENDIX
Proof. (Lemma 13) First of all observe that
h
i Z sl
+ a
E (|sl − Xl | ) =
(sl − t)a fl,n (t)dt
0
Z sl
≤ (sl)a
fl,n (t)dt,
Applying Identity (4) for z = rl and b = ⌈a⌉ we have
Z
(28)
∞
t
⌈a⌉
rl
(l − 1 + ⌈a⌉)! −nrl
fl,n (t)dt =
e
(l − 1)!n⌈a⌉
0
0
(l−1)!
e
(30)
Applying Stirling’s formula (14) for N = l we get
1
ll
≤ el l 2 .
(l − 1)!
(31)
Using (31) with assumption s = 1−ǫ
n in Inequality (30) we imply
(
Z sln
sne l 21
l
if l ≥ 1ǫ
xl−1
esn
(32)
dx ≤ snl
e−x
(l−1)l−1
(l − 1)!
if l < 1ǫ .
0
el−1 (l−1)!
Putting together (28), (29) and (32) we have
n
X
n
l=1
l
E (|sl − Xl |+ )a
≤n
1−a
a
(ns)
n
X
nse l
ens
l≥ 1ǫ
+ n1−a (ns)a+1
l
X
(nrl)j
(nrl)l
≤ (l + 1)
.
j!
(l)!
j=0
l
X la (l − 1)l−1
.
el−1 (l − 1)!
1
l< ǫ
l≥ ǫ
l< ǫ
l
j=l+1
l⌈a⌉−1 (nrl)l
(nrl)j
≤ (⌈a⌉ − 1) (nr)⌈a⌉−1
.
j!
l + 1 (l)!
(38)
From Stirling’s formula (14) for N = l we get
ll
el
(39)
≤ 1 ≤ el .
l!
l2
Putting all together (35–39) we have
h
⌈a⌉ i
1 nre l
E |Xl − vl|+
≤ f (l) ⌈a⌉
,
(40)
enr
n
where
!
⌈a⌉−1
(l − 1 + ⌈a⌉)!
⌈a⌉−1 l
.
f (l) = (l + 1) + (⌈a⌉ − 1)(nr)
l+1
(l − 1)!
Since rn = 1 + τ is some constant independent on n we derive
f (l) = Θ lmax(⌈a⌉+1,2⌈a⌉−2) .
(41)
⌈a⌉
a
and
(42)
Together (40) and (42) imply that
E
(34)
E (|sl − Xl |+ )a = O n1−a .
This is enough to prove Lemma 13.
Proof. (Lemma 14) Let ⌈a⌉ be the smallest integer greater or
equal to a. Observe that
⌈a⌉ Z ∞
+
E |Xl − vl|
(t − rl)⌈a⌉ fl,n (t)dt
=
rl
Z ∞
≤
t⌈a⌉ fl,n (t)dt.
(35)
rl
X
(33)
Finally, putting together (33) and (34) we get
n
X
n
l+⌈a⌉−1
a
h
a i
⌈a⌉ ⌈a⌉
+
+
.
E |Xl − vl|
≤ E |Xl − vl|
a− 12
(37)
Observe that
Next we apply Jensen’s inequality (see (9)) for f (x) = x
⌈a⌉
+
X = |Xl − vl|
and get
Combining assumption sn < 1 with the elementary inequality
xe < ex when x < 1 we deduce that nse
ens < 1. Hence
n
X nse l
X la (l − 1)l−1
a− 12
a+1
= O(1).
l
(ns)a
+
(ns)
ens
el−1 (l − 1)!
1
1
l=1
Hence
−x l−1
Observe that, the function g(x) = e x
is monotonically
increasing over the interval [0, l−1] and monotonically decreasing
over the interval [l − 1, ∞]. Therefore, we easily derive the
following inequality
Z sln
xl−1
xl−1
dx ≤ sln max e−x
e−x
(l − 1)!
x∈[0,sln]
(l − 1)!
0
l
l
l
sn
if sln ≤ l − 1
esn
(l−1)!
= snl
l−1
(l−1)
l−1
if sln > l − 1.
j=0
(nrl)j
.
j!
(36)
(nrl)j
(nrl)j+1
≤
, when j ≤ l − 1.
j!
(j + 1)!
l−1
. The substitution tn = x yields
Z sln
xl−1
fl,n (t)dt =
e−x
dx.
(29)
(l − 1)!
0
X
Using assumption rn > 1 we can easily derive
ne−nt (nt)
(l−1)!
where fl,n (t) =
Z sl
l−1+⌈a⌉
a l
h
a i
a
1 nre ⌈a⌉
+
|Xl − vl|
.
≤ f (l) ⌈a⌉ a
n
enr
(43)
Combining assumption rn > 1 with the elementary inequality
xe < ex , when x > 1 we deduce that
nre
< 1.
(44)
enr
From (41) and (44) we have
a
n
a l
X
f (l) ⌈a⌉ nre ⌈a⌉
= O(1).
(45)
l
enr
l=1
Finally, putting together (43) and (45) we get
n
X
n
l=1
l
E (|Xl − vl|+ )a = O n1−a .
This is enough to prove Lemma 14.
| 8cs.DS
|
arXiv:cs/0208008v1 [cs.PL] 6 Aug 2002
Soft Concurrent Constraint Programming
Stefano Bistarelli
Istituto di Informatica e Telematica, C.N.R., Pisa, Italy
and
Ugo Montanari
Dipartimento di Informatica, Università di Pisa, Italy
and
Francesca Rossi
Dipartimento di Matematica Pura ed Applicata, Università di Padova, Italy.
Soft constraints extend classical constraints to represent multiple consistency levels, and thus
provide a way to express preferences, fuzziness, and uncertainty. While there are many soft
constraint solving formalisms, even distributed ones, by now there seems to be no concurrent
programming framework where soft constraints can be handled. In this paper we show how the
classical concurrent constraint (cc) programming framework can work with soft constraints, and
we also propose an extension of cc languages which can use soft constraints to prune and direct
the search for a solution. We believe that this new programming paradigm, called soft cc (scc),
can be also very useful in many web-related scenarios. In fact, the language level allows web
agents to express their interaction and negotiation protocols, and also to post their requests in
terms of preferences, and the underlying soft constraint solver can find an agreement among the
agents even if their requests are incompatible.
Categories and Subject Descriptors: D.1.3 [Programming Techniques]: Concurrent Programming—Distributed programming; D.3.1 [Programming Languages]: Formal Definitions and
Theory—Semantics; Syntax; D.3.2 [Programming Languages]: Language Classifications—
Concurrent, distributed, and parallel languages; Constraint and logic languages; D.3.3 [Programming Languages]: Language Constructs and Features—Concurrent programming structures; Constraints; F.3.2 [Logics and Meanings of Programs]: Semantics of Programming
Languages—Operational semantics
General Terms: Languages
Additional Key Words and Phrases: constraints, soft constraints, concurrent constraint programming
1. INTRODUCTION
The concurrent constraint (cc) paradigm [Saraswat 1993] is a very interesting computational framework which merges together constraint solving and concurrency.
The main idea is to choose a constraint system and use constraints to model comResearch supported in part by the the the Italian MIUR Projects cometa and napoli and by ASI
project ARISCOM.
Permission to make digital/hard copy of all or part of this material without fee for personal
or classroom use provided that the copies are not made or distributed for profit or commercial
advantage, the ACM copyright/server notice, the title of the publication, and its date appear, and
notice is given that copying is by permission of the ACM, Inc. To copy otherwise, to republish,
to post on servers, or to redistribute to lists requires prior specific permission and/or a fee.
c 2018 ACM 1529-3785/2018/0700-0001 $5.00
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018, Pages 1–25.
2
·
Stefano Bistarelli et al.
munication and synchronization among concurrent agents.
Until now, constraints in cc were crisp, in the sense that they could only be satisfied or violated. Recently, the classical idea of crisp constraints has been shown
to be too weak to represent real problems and a big effort has been done toward
the use of soft constraints [Freuder and Wallace 1992; Dubois et al. 1993; Ruttkay 1994; Fargier and Lang 1993; Schiex et al. 1995; Bistarelli et al. 1997; 2001;
Bistarelli 2001], which can have more than one level of consistency. Many real-life
situations are, in fact, easily described via constraints able to state the necessary
requirements of the problems. However, usually such requirements are not hard,
and could be more faithfully represented as preferences, which should preferably be
satisfied but not necessarily. Also, in real life, we are often challenged with overconstrained problems, which do not have any solution, and this also leads to the
use of preferences or in general of soft constraints rather than classical constraints.
Generally speaking, a soft constraint is just a classical constraint plus a way to
associate, either to the entire constraint or to each assignment of its variables, a
certain element, which is usually interpreted as a level of preference or importance.
Such levels are usually ordered, and the order reflects the idea that some levels are
better than others. Moreover, one has also to say, via suitable combination operators, how to obtain the level of preference of a global solution from the preferences
in the constraints.
Many formalisms have been developed to describe one or more classes of soft
constraints. For instance consider fuzzy CSPs [Dubois et al. 1993; Ruttkay 1994],
where crisp constraints are extended with a level of preference represented by a
real number between 0 and 1, or probabilistic CSPs [Fargier and Lang 1993], where
the probability to be in the real problem is assigned to each constraint. Some
other examples are partial [Freuder and Wallace 1992] or valued CSPs [Schiex et al.
1995], where a preference is assigned to each constraint, in order to satisfy as many
constraints as possible, and thus handle also overconstrained problems.
We think that many network-related problem could be represented and solved
by using soft constraints. Moreover, the possibility to use a concurrent language
on top of a soft constraint system, could lead to the birth of new protocols with an
embedded constraint satisfaction and optimization framework.
In particular, the constraints could be related to a quantity to be minimized/maximized but they could also satisfy policy requirements given for performance or administrative reasons. This leads to change the idea of QoS in routing
and to speak of constraint-based routing [Awduche et al. 1999; Clark 1989; Jain
and Sun 2000; Calisti and Faltings 2000]. Constraints are in fact able to represent
in a declarative fashion the needs and the requirements of agents interacting over
the web.
The features of soft constraints could also be useful in representing routing problems where an imprecise state information is given [Chen and Nahrstedt 1998].
Moreover, since QoS is only a specific application of a more general notion of Service Level Agreement (SLA), many applications could be enhanced by using such
a framework. As an example consider E-commerce: here we are always looking for
establishing an agreement between a merchant, a client and possibly a bank. Also,
all auction-based transactions need an agreement protocol. Moreover, also secuACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
Soft Concurrent Constraint Programming
·
3
rity protocol analysis have shown to be enhanced by using security levels instead
of a simple notion of secure/insecure level [Bella and Bistarelli 2001]. All these
considerations advocate for the need of a soft constraint framework where optimal
answers are extracted.
In this paper, we use one of the frameworks able to deal with soft constraints
[Bistarelli et al. 1995; 1997]. The framework is based on a semiring structure that
is equipped with the operations needed to combine the constraints present in the
problem and to choose the best solutions. According to the choice of the semiring,
this framework is able to model all the specific soft constraint notions mentioned
above. We compare the semiring-based framework with constraint systems “a la
Saraswat” and then we show how use it inside the cc framework. The next step is
the extension of the syntax and operational semantics of the language to deal with
the semiring levels. Here, the main novelty with respect to cc is that tell and ask
agents are equipped with a preference (or consistency) threshold which is used to
prune the search.
After a short summary of concurrent constraint programming (§2.1) and of
semiring-based SCSPs (§2.2), we show how the concurrent constraint framework
can be used to handle also soft constraints (§3). Then we integrate semirings inside
the syntax of the language and we change its semantics to deal with soft levels (§4).
Some notions of observables able to deal with a notion of optimization and with
success (§6.1), fail (§6.2) and hang computations (§6.3) are then defined. Some
examples (§5) and an application scenario (§7) conclude our presentation showing
the expressivity of the new language. Finally, conclusions (§8) are added to point
out the main results and possible directions for future work.
2. BACKGROUND
2.1 Concurrent Constraint Programming
The concurrent constraint (cc) programming paradigm [Saraswat 1993] concerns
the behaviour of a set of concurrent agents with a shared store, which is a conjunction of constraints. Each computation step possibly adds new constraints to the
store. Thus information is monotonically added to the store until all agents have
evolved. The final store is a refinement of the initial one and it is the result of the
computation. The concurrent agents do not communicate directly with each other,
but only through the shared store, by either checking if it entails a given constraint
(ask operation) or adding a new constraint to it (tell operation).
2.1.1 Constraint Systems. A constraint is a relation among a specified set of
variables. That is, a constraint gives some information on the set of possible values
that these variables may assume. Such information is usually not complete since
a constraint may be satisfied by several assignments of values of the variables (in
contrast to the situation that we have when we consider a valuation, which tells
us the only possible assignment for a variable). Therefore it is natural to describe
constraint systems as systems of partial information [Saraswat 1993].
The basic ingredients of a constraint system (defined following the information
systems idea) are a set D of primitive constraints or tokens, each expressing some
partial information, and an entailment relation ⊢ defined on ℘(D) × D (or its
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
·
4
Stefano Bistarelli et al.
extension defined on ℘(D) × ℘(D))1 satisfying:
—u ⊢ P for all P ∈ u (reflexivity) and
—if u ⊢ v and v ⊢ z, then u ⊢ z (transitivity).
We also define u ≈ v if u ⊢ v and v ⊢ u.
As an example of entailment relation, consider D as the set of equations over
the integers; then ⊢ could include the pair h{x = 3, x = y}, y = 3i, which means
that the constraint y = 3 is entailed by the constraints x = 3 and x = y. Given
X ∈ ℘(D), let X be the set X closed under entailment. Then, a constraint in an
information system h℘(D), ⊢i is simply an element of ℘(D).
As it is well known, h℘(D), ⊆i is a complete algebraic lattice, the compactness
of ⊢ gives the algebraic structure for ℘(D), with least element true = {P | ∅ ⊢ P },
greatest element D (which we will mnemonically denote f alse), glbs (denoted by
⊓) given by the closure of the intersection and lubs (denoted by ⊔) given by the
closure of the union. The lub of chains is, however, just the union of the members
in the chain. We use a, b, c, d and e to stand for elements of ℘(D); c ⊆ d means
c ⊢ d.
2.1.2 The hiding operator: Cylindric Algebras. In order to treat the hiding operator of the language (see later), a general notion of existential quantifier for
variables in constraints is introduced, which is formalized in terms of cylindric algebras. This leads to the concept of cylindric constraint system over an infinite set
of variables V such that for each variable x ∈ V , ∃x : ℘(D) → ℘(D) is an operation
satisfying:
(1)
(2)
(3)
(4)
u ⊢ ∃x u;
u ⊢ v implies (∃x u) ⊢ (∃x v);
∃x (u ⊔ ∃x v) ≈ (∃x u) ⊔ (∃x v);
∃x ∃y u ≈ ∃y ∃x u.
2.1.3 Procedure calls. In order to model parameter passing, diagonal elements
are added to the primitive constraints. We assume that, for x, y ranging in V , ℘(D)
contains a constraint dxy which satisfies the following axioms:
(1) dxx = true,
(2) if z 6= x, y then dxy = ∃z (dxz ⊔ dzy ),
(3) if x =
6 y then dxy ⊔ ∃x (c ⊔ dxy ) ⊢ c.
Note that the in the previous definition we assume the cardinality of the domain
for x, y and z greater than 1. Note also that, if ⊢ models the equality theory, then
the elements dxy can be thought of as the formulas x = y.
2.1.4 The language. The syntax of a cc program is show in Table I: P is the
class of programs, F is the class of sequences of procedure declarations (or clauses),
A is the class of agents, c ranges over constraints, and x is a tuple of variables.
Each procedure is defined (at most) once, thus nondeterminism is expressed via the
+ combinator only. We also assume that, in p(x) :: A, we have vars(A) ⊆ x, where
1 The
extension is s.t. u ⊢ v iff u ⊢ P for every P ∈ v.
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
Soft Concurrent Constraint Programming
Table I.
·
5
cc syntax
P ::= F.A
F ::= p(x) :: A | F.F
A ::= success | f ail | tell(c) → A | E | AkA | ∃x A | p(x)
E ::= ask(c) → A | E + E
vars(A) is the set of all variables occurring free in agent A. In a program P = F.A,
A is the initial agent, to be executed in the context of the set of declarations F .
This corresponds to the language considered in [Saraswat 1993], which allows only
guarded nondeterminism.
In order to better understand the extension of the language that we will introduce
later, let us remind here the operational semantics of the agents.
—agent “success” succeeds in one step,
—agent “f ail” fails in one step,
—agent “ask(c) → A” checks whether constraint c is entailed by the current store
and then, if so, behaves like agent A. If c is inconsistent with the current store,
it fails, and otherwise it suspends, until c is either entailed by the current store
or is inconsistent with it;
—agent “ask(c1 ) → A1 + ask(c2 ) → A2 ” may behave either like A1 or like A2 if
both c1 and c2 are entailed by the current store, it behaves like Ai if ci only is
entailed, it suspends if both c1 and c2 are consistent with but not entailed by
the current store, and it behaves like “ask(c1 ) → A1 ” whenever “ask(c2 ) → A2 ”
fails (and vice versa);
—agent “tell(c) → A” adds constraint c to the current store and then, if the
resulting store is consistent, behaves like A, otherwise it fails.
—agent A1 kA2 behaves like A1 and A2 executing in parallel;
—agent ∃x A behaves like agent A, except that the variables in x are local to A;
—p(x) is a call of procedure p.
A formal treatment of the cc semantics can be found in [Saraswat 1993; de Boer and
Palamidessi 1991]. Also, a denotational semantics of deterministic cc programs,
based on closure operators, can be found in [Saraswat 1993]. A more complete
survey on several concurrent paradigms is given also in [de Boer and Palamidessi
1994].
2.2 Soft Constraints
Several formalization of the concept of soft constraints are currently available.
In the following, we refer to the one based on c-semirings [Bistarelli et al. 1997;
Bistarelli 2001], which can be shown to generalize and express many of the others.
A soft constraint may be seen as a constraint where each instantiations of its
variables has an associated value from a partially ordered set which can be interpreted as a set of preference values. Combining constraints will then have to take
into account such additional values, and thus the formalism has also to provide
suitable operations for combination (×) and comparison (+) of tuples of values and
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
6
·
Stefano Bistarelli et al.
constraints. This is why this formalization is based on the concept of c-semiring,
which is just a set plus two operations.
2.2.1
C-semirings. A semiring is a tuple hA, +, ×, 0, 1i such that:
(1) A is a set and 0, 1 ∈ A;
(2) + is commutative, associative and 0 is its unit element;
(3) × is associative, distributes over +, 1 is its unit element and 0 is its absorbing
element.
A c-semiring is a semiring hA, +, ×, 0, 1i such that + is idempotent, 1 is its
absorbing element and × is commutative. Let us consider the relation ≤S over A
such that a ≤S b iff a + b = b. Then it is possible to prove that (see [Bistarelli et al.
1997]):
(1) ≤S is a partial order;
(2) + and × are monotone on ≤S ;
(3) 0 is its minimum and 1 its maximum;
(4) hA, ≤S i is a complete lattice and, for all a, b ∈ A, a + b = lub(a, b).
Moreover, if × is idempotent, then: + distribute over ×; hA, ≤S i is a complete
distributive lattice and × its glb. Informally, the relation ≤S gives us a way to
compare semiring values and constraints. In fact, when we have a ≤S b, we will
say that b is better than a. In the following, when the semiring will be clear from
the context, a ≤S b will be often indicated by a ≤ b.
2.2.2 Soft Constraints and Problems. Given a semiring S = hA, +, ×, 0, 1i, a
finite set D (the domain of the variables) and an ordered set of variables V , a
constraint is a pair hdef , coni where con ⊆ V and def : D|con| → A. Therefore, a
constraint specifies a set of variables (the ones in con), and assigns to each tuple
of values of these variables an element of the semiring. Consider two constraints
c1 = hdef1 , coni and c2 = hdef2 , coni, with |con| = k. Then c1 ⊑S c2 if for all
k-tuples t, def1 (t) ≤S def2 (t). The relation ⊑S is a partial order.
A soft constraint problem is a pair hC, coni where con ⊆ V and C is a set of
constraints: con is the set of variables of interest for the constraint set C, which
however may concern also variables not in con. Note that a classical CSP is a SCSP
where the chosen c-semiring is: SCSP = h{f alse, true}, ∨, ∧, f alse, truei. Fuzzy
CSPs [Schiex 1992] can instead be modeled in the SCSP framework by choosing the
c-semiring SF CSP = h[0, 1], max, min, 0, 1i. Many other “soft” CSPs (Probabilistic,
weighted, . . . ) can be modeled by using a suitable semiring structure (for example,
Sprob = h[0, 1], max, ×, 0, 1i, Sweight = hR, min, +, 0, +∞i, . . . ).
Figure 1 shows the graph representation of a fuzzy CSP. Variables and constraints
are represented respectively by nodes and by undirected arcs (unary for c1 and c3
and binary for c2 ), and semiring values are written to the right of the corresponding
tuples. The variables of interest (that is the set con) are represented with a double
circle. Here we assume that the domain D of the variables contains only elements
a and b.
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
Soft Concurrent Constraint Programming
<a> --> 0.9
<b> --> 0.1
c1
X
<a, a> --> 0.8
<a, b> --> 0.2
<b, a> --> 0
<b, b> --> 0
Fig. 1.
·
7
<a> --> 0.9
<b> --> 0.5
c3
Y
c2
A fuzzy CSP
2.2.3 Combining and projecting soft constraints. Given two constraints c1 =
hdef 1 , con1 i and c2 = hdef 2 , con2 i, their combination c1 ⊗ c2 is the constraint
con
hdef , coni defined by con = con1 ∪ con2 and def (t) = def 1 (t ↓con
con1 ) × def 2 (t ↓con2 ),
X
where t ↓Y denotes the tuple of values over the variables in Y , obtained by projecting tuple t from X to Y . In words, combining two constraints means building
a new constraint involving all the variables of the original ones, and which associates to each tuple of domain values for such variables a semiring element which is
obtained by multiplying the elements associated by the original constraints to the
appropriate subtuples.
Given a constraint c = hdef , coni and a subset I of V , the projection of c over
′
′ ′
′
′
I,
P written c ⇓I is the constraint hdef , con i where con = con ∩ I and def (t ) =
′ def (t). Informally, projecting means eliminating some variables. This
t/t↓con
I∩con =t
is done by associating to each tuple over the remaining variables a semiring element
which is the sum of the elements associated by the original constraint to all the
extensions of this tuple over the eliminated variables. In short, combination is
performed via the multiplicative operation of the semiring, and projection via the
additive one.
2.2.4 Solutions.
The solution of an SCSP problem P = hC, coni is the constraint
N
Sol(P ) = ( C) ⇓con . That is, we combine all constraints, and then project over
the variables in con. In this way we get the constraint over con which is “induced”
by the entire SCSP.
For example, the solution of the fuzzy CSP of Figure 1 associates a semiring
element to every domain value of variable x. Such an element is obtained by first
combining all the constraints together. For instance, for the tuple ha, ai (that
is, x = y = a), we have to compute the minimum between 0.9 (which is the value
assigned to x = a in constraint c1 ), 0.8 (which is the value assigned to hx = a, y = ai
in c2 ) and 0.9 (which is the value for y = a in c3 ). Hence, the resulting value for
this tuple is 0.3. We can do the same work for tuple ha, bi → 0.2, hb, ai → 0 and
hb, bi → 0. The obtained tuples are then projected over variable x, obtaining the
solution hai → 0.8 and hbi → 0.
Sometimes it may be useful to find only a semiring value representing the least
upper bound among the values yielded by the solutions. This is called the best level
of consistency of an SCSP problem P and it is defined by blevel(P ) = Sol(P ) ⇓∅
(for instance, the fuzzy CSP of Figure 1 has best level of consistency 0.8). We also
say that: P is α-consistent if blevel(P ) = α; P is consistent iff there exists α > 0
such that P is α-consistent; P is inconsistent if it is not consistent.
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
8
·
Stefano Bistarelli et al.
3. CONCURRENT CONSTRAINT PROGRAMMING OVER SOFT CONSTRAINTS
Given a semiring S = hA, +, ×, 0, 1i and an ordered set of variables V over a finite
domain D, we will now show how soft constraints over S with a suitable pair of
operators form a semiring, and then, we highlight the properties needed to map
soft constraints over constraint systems “a la Saraswat” (as recalled in Section 2.1.
We start by giving the definition of the carrier set of the semiring.
Definition 3.1 (functional constraints). We define C = (V → D) → A as the set
of all possible constraints that can be built starting from S = hA, +, ×, 0, 1i, D and
V.
A generic function describing the assignment of domain elements to variables will
be denoted in the following by η : V → D. Thus a constraint is a function which,
given an assignment η of the variables, returns a value of the semiring.
Note that in this functional formulation, each constraint is a function and not a
pair representing the variable involved and its definition. Such a function involves
all the variables in V , but it depends on the assignment of only a finite subset of
them. We call this subset the support of the constraint. For computational reasons
we require each support to be finite.
Definition 3.2 (constraint support). Consider a constraint c ∈ C. We define his
support as supp(c) = {v ∈ V | ∃η, d1 , d2 .cη[v := d1 ] 6= cη[v := d2 ]}, where
(
d
if v = v ′ ,
′
η[v := d]v =
ηv ′ otherwise.
Note that cη[v := d1 ] means cη ′ where η ′ is η modified with the association v := d1
(that is the operator [ ] has precedence over application).
Definition 3.3 (functional mapping). Given
any
soft
constraint
hdef, {v1 , . . . , vn }i ∈ C, we can define its corresponding function c ∈ C s.t.
cη[v1 := d1 ] . . . [vn := dn ] = def (d1 , . . . , dn ). Clearly supp(c) ⊆ {v1 , . . . , vn }.
Definition 3.4 (Combination and Sum). Given the set C, we can define the combination and sum functions ⊗, ⊕ : C × C → C as follows:
(c1 ⊗ c2 )η = c1 η ×S c2 η
and
(c1 ⊕ c2 )η = c1 η +S c2 η.
Notice that function ⊗ has the same meaning of the already defined ⊗ operator
(see Section 2.2) while function ⊕ models a sort of disjunction.
By using the ⊕S operator we can easily extend the partial order ≤S over C by
defining c1 ⊑S c2 ⇐⇒ c1 ⊕S c2 = c2 . In the following, when the semiring will be
clear from the context, we will use ⊑.
We can also define a unary operator that will be useful to represent the unit
elements of the two operations ⊕ and ⊗. To do that, we need the definition of
constant functions over a given set of variables.
Definition 3.5 (constant function). We define function ā as the function that
returns the semiring value a for all assignments η, that is, āη = a. We will usually
write ā simply as a.
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
Soft Concurrent Constraint Programming
·
9
An example of constants that will be useful later are 0̄ and 1̄ that represent respectively the constraint associating 0 and 1 to all the assignment of domain values.
It is easy to verify that each constant has an empty support. More generally we
can prove the following:
Proposition 3.6. The support of a constraint c ⇓I is always a subset of I(that
is supp(c ⇓I ) ⊆ I).
Proof. By definition of ⇓I , for any variable x 6∈ I we have c ⇓I η[x = a] = c ⇓I
η[x = b] for any a and b. So, by definition of support x 6∈ supp(c ⇓I ).
Theorem 3.7 (Higher order semiring). The structure SC = hC, ⊕, ⊗, 0, 1i
where
—C : (V → D) → A is the set of all the possible constraints that can be built
starting from S, D and V as defined in Definition 3.1,
—⊗ and ⊕ are the functions defined in Definition 3.4, and
—0 and 1 are constant functions defined following Definition 3.5,
is a c-semiring.
Proof. To prove the theorem it is enough to check all the properties with the
fact that the same properties hold for semiring S. We give here only a hint, by
showing the commutativity of the ⊗ operator:
c1 ⊗ c2 η = (by definition of ⊗)
c1 η × c2 η = (by commutativity of ×)
c2 η × c1 η = (by definition of ⊗)
c2 ⊗ c1 η.
All the other properties can be proved similarly.
The next step is to look for a notion of token and of entailment relation. We define
as tokens the functional constraints in C and we introduce a relation ⊢ that is an
entailment relation when the multiplicative operator of the semiring is idempotent.
Definition 3.8 (⊢ relation). Consider the high order semiring carrier set C and
the partial order ⊑. We define
N the relation ⊢⊆ ℘(C) × C s.t. for each C ∈ ℘(C) and
c ∈ C, we have C ⊢ c ⇐⇒
C ⊑ c.
The next theorem shows that, when the multiplicative operator of the semiring
is idempotent, the ⊢ relation satisfies all the properties needed by an entailment.
Theorem 3.9 (⊢ with idempotent × is an entailment relation).
Consider the higher order semiring carrier set C and the partial order ⊑. Consider
also the relation ⊢ of Definition 3.8. Then, if the multiplicative operation of the
semiring is idempotent, ⊢ is an entailment relation.
Proof. Is enough to check that for any c ∈ C, and for any C1 , C2 and C3 subsets
of C we have
N
(1) C ⊢ c when c ∈ C: We need to show that
C ⊑ c when c ∈ C. This follows
from the extensivity of ×.
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
10
·
Stefano Bistarelli et al.
(2) if C1 ⊢ C2 and C2 ⊢ C3 then C1 ⊢ C3 : To prove this we use the extended
version of the relationN⊢ able to deal with subsets of C : ℘(C) × ℘(C) s.t.
C1 ⊢ C2 ⇐⇒ C1 ⊢
C2 . Note
N that when × is idempotent we have that,
∀c2 ∈ C2 , C1 ⊢ N
c2 ⇐⇒ N
C1 ⊢
CN
2 . In this
Ncase to prove
N the item
N we have
to prove that if
C1 ⊑
C2 and
C2 ⊑
C3 , then
C1 ⊑
C3 . This
comes from the transitivity of ⊑.
Note that in this setting the notion of token (constraint) and of set of tokens (set
of constraints) closed under entailment is used indifferently. In fact, given a set
of constraint functions C1 , its closure
w.r.t. entailment is a set C¯1 that contains
N
all the constraints greater
than
C
.
1 This set is univocally representable by the
N
constraint function
C1 .
The definition of the entailment operator ⊢ on top of the higher order semiring
SC = hC, ⊕, ⊗, 0, 1i and of the ⊑ relation leads to the notion of soft constraint
system. It is also important to notice that in [Saraswat 1993] it is claimed that a
constraint system is a complete algebraic lattice. Here we do not ask for this, since
the algebraic nature of the structure C strictly depends on the properties of the
semiring.
3.1 Non-idempotent ×
If the constraint system is defined on top of a non-idempotent multiplicative operator, we cannot obtain a ⊢ relation satisfying all the properties of an entailment.
Nevertheless, we can give a denotational semantics to the constraint store, as described in Section 4, using the operations of the higher order semiring.
To treat the hiding operator of the language, a general notion of existential
quantifier has to be introduced by using notions similar to those used in cylindric
algebras. Note however that cylindric algebras are first of all boolean algebras.
This could be possible in our framework only when the × operator is idempotent.
Definition 3.10 (hiding). Consider a set of variables V with domain D and the
corresponding softPconstraint system C. We define for each x ∈ V the hiding
function (∃x c)η = di ∈D cη[x := di ].
Notice that x does not belong to the support of ∃x c.
By using the hiding function we can represent the ⇓ operator defined in Section 2.2.
Proposition 3.11. Consider a semiring S = hA, +, ×, 0, 1i, a domain of the
variables D, an ordered set of variables V , the corresponding structure C and the
class of hiding functions ∃x : C → C as defined in Definition 3.10. Then, for any
constraint c and any variable x ⊆ V , c ⇓V −x = ∃x c.
Proof. IsP
enough to apply the definition of ⇓V −x and ∃x and check that both
are equal to di ∈D cη[x := di ].
We now show how the hiding function so defined satisfies the properties of cylindric algebras.
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
Soft Concurrent Constraint Programming
·
11
Theorem 3.12. Consider a semiring S = hA, +, ×, 0, 1i, a domain of the variables D, an ordered set of variables V , the corresponding structure C and the class
of hiding functions ∃x : C → C as defined in Definition 3.10. Then C is a cylindric
algebra satisfying:
(1 ) c ⊢ ∃x c
(2 ) c1 ⊢ c2 implies ∃x c1 ⊢ ∃x c2
(3 ) ∃x (c1 ⊗ ∃x c2 ) ≈ ∃x c1 ⊗ ∃x c2 ,
(4 ) ∃x ∃y c ≈ ∃y ∃x c
Proof. Let us consider all the items:
(1) It follows from the intensivity of +;
(2) It follows from the monotonicity of +;
(3) It follows from theorems about distributivity and idempotence, proven in
[Bistarelli et al. 1997];
(4) It follows from commutativity and associativity of +.
To model parameter passing we need also to define what diagonal elements are.
Definition 3.13 (diagonal elements). Consider an ordered set of variables V and
the corresponding soft constraint system C. Let us define for each x, y ∈ V a
constraint dxy ∈ C s.t., dxy η[x := a, y := b] = 1 if a = b and dxy η[x := a, y := b] = 0
if a 6= b. Notice that supp(dxy ) = {x, y}.
We can prove that the constraints just defined are diagonal elements.
Theorem 3.14. Consider a semiring S = hA, +, ×, 0, 1i, a domain of the variables D, an ordered set of variables V , and the corresponding structure C. The
constraints dxy defined in Definition 3.13 represent diagonal elements, that is
(1 ) dxx = 1,
(2 ) if z 6= x, y then dxy = ∃z (dxz ⊗ dzy ),
(3 ) if x 6= y then dxy ⊗ ∃x (c ⊗ dxy ) ⊢ c.
Proof. (1) It follows from the definition of the 1 constant and of the diagonal
constraint;
(2) The constraint dxz ⊗ dzy is equal to 1 when x = y = z, and is equal to 0 in all
the other cases. If we project this constraint over z, we obtain the constraint
∃z (dxz ⊗ dzy ) that is equal to 1 only when x = y;
(3) The constraint (c ⊗ dxy )η has value 0 whenever η(x) 6= η(y) and cη elsewhere.
Now, (∃x (c⊗dxy ))η is by definition equal to cη[x := y]. Thus (dxy ⊗∃x (c⊗dxy ))η
is equal to cη when η(x) = η(y) and 0 elsewhere. By the last assumption, we
have the relation of entailment with c.
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
12
·
Stefano Bistarelli et al.
Table II.
scc syntax
P :: = F.A
F :: = p(X) :: A | F.F
A :: = stop | tell(c) →φ A | tell(c) →a A | E | AkA | ∃X.A | p(X)
E :: = ask(c) →φ A | ask(c) →a A | E + E
3.2 Using cc on top of a soft constraint system
The only problem in using a soft constraint system in a cc language is the interpretation of the consistency notion necessary to deal with the ask and tell operations.
Usually SCSPs with best level of consistency equal to 0 are interpreted as inconsistent, and those with level greater than 0 as consistent, but we can be more
general. In fact, we can define a suitable function α that, given the best level
of the actual store, will map such a level over the classical notion of consistency/inconsistency. More precisely, given a semiring S = hA, +, ×, 0, 1i, we can
define a function α : A → {f alse, true}. Function α has to be at least monotone,
but functions with a richer set of properties could be used. It is worth to notice
that in a different environment some of the authors use a similar function to map
elements from a semiring to another, by using abstract interpretation techniques
[Bistarelli et al. 2000a; 2000b]
Whenever we need to check the consistency of the store, we will first compute
the best level and then we will map such a value by using function α over true or
f alse.
It is important to notice that changing the α function (that is, by mapping in
a different way the set of values A over the boolean elements true and f alse), the
same cc agent yields different results: by using a high cut level, the cc agent will
either finish with a failure or succeed with a high final best level of consistency of
the store. On the other hand, by using a low level, more programs will end in a
success state.
4. SOFT CONCURRENT CONSTRAINT PROGRAMMING
The next step in our work is now to extend the syntax of the language in order to
directly handle the cut level. This means that the syntax and semantics of the tell
and ask agents have to be enriched with a threshold to specify when tell/ask agents
have to fail, succeed or suspend.
Given a soft constraint system hS, D, V i and the corresponding structure C, and
any constraint φ ∈ C, the syntax of agents in soft concurrent constraint programming is given in Table II.
The main difference w.r.t. the original cc syntax is
the presence of a semiring element a and of a constraint φ to be checked whenever
an ask or tell operation is performed. More precisely, the level a (resp., φ) will be
used as a cut level to prune computations that are not good enough.
We present here a structured operational semantics for scc programs, in the SOS
style, which consists of defining the semantic of the programming language by specifying a set of configurations Γ, which define the states during execution, a relation
→ ⊆ Γ×Γ which describes the transition relation between the configurations, and a
set T of terminal configurations. To give an operational semantics to our language,
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
Soft Concurrent Constraint Programming
Table III.
·
13
Transition rules for scc
hstop, σi −→ hsuccess, σi
(Stop)
(σ ⊗ c) ⇓∅ 6< a
(Valued-tell)
htell(c) →a A, σi −→ hA, σ ⊗ ci
σ ⊗ c 6⊏ φ
htell(c) →φ A, σi −→ hA, σ ⊗ ci
σ ⊢ c, σ ⇓∅ 6< a
hask(c) →a A, σi −→ hA, σi
σ ⊢ c, σ 6⊏ φ
hask(c) →φ A, σi −→ hA, σi
hA1 , σi −→ hA′1 , σ′ i
(Ask)
hA1 kA2 , σi −→ hA2 , σ′ i
hA2 kA1 , σi −→ hA2 , σ′ i
(Parallelism)
hE1 , σi −→ hA1 , σ′ i
hE1 + E2 , σi −→ hA1 , σ′ i
hE2 + E1 , σi −→ hA1 , σ′ i
h∃x A, σi −→ hA′ , σ′ i
(Valued-ask)
hA1 , σi −→ hsuccess, σ′ i
hA1 kA2 , σi −→ hA′1 kA2 , σ′ i
hA2 kA1 , σi −→ hA2 kA′1 , σ′ i
hA[y/x], σi −→ hA′ , σ′ i
(Tell)
with y fresh
hp(y), σi −→ hA[y/x], σi when p(x) :: A
(Nondeterminism)
(Hidden variables)
(Procedure call)
we need to describe an appropriate transition system.
Definition 4.1 (transition system). A transition system is a triple hΓ, T, →i
where Γ is a set of possible configurations, T ⊆ Γ is the set of terminal configurations and →⊆ Γ × Γ is a binary relation between configurations.
The set of configurations represent the evolutions of the agents and the modifications in the constraint store. We define the transition system of soft cc as
follows:
Definition 4.2 (configurations). The set of configurations for a soft cc system is
the set Γ = {hA, σi} ∪ {hsuccess, σi}, where σ ∈ C. The set of terminal configurations is the set T = {hsuccess, σi} and the transition rule for the scc language are
defined in Table III.
Here is a brief description of the transition rules:
Stop. The stop agent succeeds in one step by transforming itself into terminal
configuration success.
Valued-tell. The valued-tell rule checks for the α-consistency of the SCSP defined
by the store σ ⊗c. The rule can be applied only if the store σ ⊗c is b-consistent with
b 6< a. In this case the agent evolves to the new agent A over the store σ ⊗ c. Note
that different choices of the cut level a could possibly lead to different computations.
Tell. The tell action is a finer check of the store. In this case, a pointwise comparison between the store σ ⊗ c and the constraint φ is performed. The idea is to
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
14
·
Stefano Bistarelli et al.
perform an overall check of the store and to continue the computation only if there
is the possibility to compute a solution not worse than φ.
Valued-ask. The semantics of the valued-ask is extended in a way similar to what
we have done for the valued-tell action. This means that, to apply the rule, we need
to check if the store σ entails the constraint c and also if the store is “consistent
enough” w.r.t. the threshold a set by the programmer.
Ask. Similar to the tell rule, here a finer (pointwise) threshold φ is compared to
the store σ.
Nondeterminism and parallelism. The composition operators + and k are not
modified w.r.t. the classical ones: a parallel agent will succeed if all the agents
succeeds; a nondeterministic rule chooses any agent whose guard succeeds.
Hidden variables. The semantics of the existential quantifier is similar to that
described in [Saraswat 1993] by using the notion of freshness of the new variable
added to the store.
Procedure calls. The semantics of the procedure call is not modified w.r.t. the
classical one. The only difference is the different use of the diagonal constraint to
represent parameter passing.
4.1 Eventual Tell/Ask
We recall that both ask and tell operations in cc could be either atomic (that is,
if the corresponding check is not satisfied, the agent does not evolve) or eventual
(that is, the agent evolves regardless of the result of the check). It is interesting to
notice that the transition rules defined in Table III could be used to provide both
interpretations of the ask and tell operations. In fact, while the generic tell/ask
rule represents an atomic behaviour, by setting φ = 0 or a = 0 we obtain their
eventual version:
htell(c) → A, σi −→ hA, σ ⊗ ci
σ⊢c
hask(c) → A, σi −→ hA, σi
(Eventual tell)
(Eventual ask)
Notice that, by using an eventual interpretation, the transition rules of the scc
become the same as those of cc (with an eventual interpretation too). This happens
since, in the eventual version, the tell/ask agent never checks for consistency and
so the soft notion of α-consistency does not play any role.
5. A SIMPLE EXAMPLE
In this section we will show the behaviour of some of the rules of our transition system. We consider in this example a soft constraint system over the fuzzy semiring.
Consider the fuzzy constraints
c : {x, y} → R2 → [0, 1]
c′ : {x} → R → [0, 1]
1
s.t. c(x, y) =
1 + |x − y|
(
1 if x ≤ 10,
s.t. c′ (x) =
0 otherwise.
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
and
Soft Concurrent Constraint Programming
·
15
Notice that the domain of both variables x and y is in this example any integer (or
real) number. As any fuzzy CSP, the definition of the constraints is instead in the
interval [0, 1].
Let’s now evaluate the agent
htell(c) →0.4 ask(c′ ) →0.8 stop, 1i
in the empty starting store 1.
By applying the Valued-tell rule we need to check (1 ⊗ c) ⇓∅ 6< 0.4. Since 1 ⊗ c = c
and c ⇓∅ = 1, the agent can perform the step, and it reaches the state
hask(c′ ) →0.8 stop, ci.
Now we need to check (by following the rule of Valued-ask) if c ⊢ c′ and c ⇓∅ 6< 0.8.
While the second relation easily holds, the first one does not hold (in fact, for x = 11
and y = 10 we have c′ (x) = 0 and c(x, y) = 0.5).
1
in place of c′ , then we
If instead we consider the constraint c′′ (x, y) = 1+2×|x−y|
have
hask(c′′ ) →0.8 stop, ci.
Here the condition c ⊢ c” easily holds and the agent ask(c′′ ) →0.8 stop can perform
its last step, reaching the stop and success states:
hstop, c ⊗ c′′ i → hsuccess, c ⊗ c′′ i.
6. OBSERVABLES AND CUTS
Sometimes one could desire to see an agent, and a corresponding program, execute
with a cut level which is different from the one originally given. We will therefore
define cutψ (A) the agent A where all the occurrences of any cut level, say φ, in any
subagent of A or in any clause of the program, are replaced by ψ if φ ⊑ ψ. This
means that the cut level of each subagent and clause becomes at least ψ, or is left
to the original level.
In this paper, for simplicity and generality reasons, this cut level change applies
only to those programs with cut levels which are constraints (φ), and not single
semiring levels (a).
Definition 6.1 (cut function). Consider an scc agent A; we define the function
cutψ : A → A that transforms ask and tell subagents as follows:
(
ask/tell(c) →ψ if φ ⊏ ψ,
cutψ (ask/tell(c) →φ ) =
ask/tell(c) →φ otherwise.
By definition of cutψ , it is easy to see that cut0 (A) = A.
We can then prove the following Lemma (that will be useful later):
Lemma 6.2 (tell and ask cut). Consider the Tell and Ask rules of Table III,
and the constraints σ and c as defined in such rules. Then:
—If the Tell rule can be applied to agent A, then the rule can be applied also to
cutψ (A) when when ψ ⊑ σ ⊗ c.
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
16
·
Stefano Bistarelli et al.
—If the Ask rule can be applied to agent A, then the rule can be applied also to
cutψ (A) when ψ ⊑ σ.
Proof. We will prove only the first item; the second can be easily proved by
using the same ideas. By the definition of the tell transition rules of Table III, if
we can apply the rule it means that A ::= tell(c) →φ A′ and if σ is the store we
have σ ⊗ c 6⊏ φ. Now, by definition of cutψ , we can have
—cutψ (A) ::= tell(c) →ψ cutψ (A′ ) when φ ⊏ ψ.
—cutψ (A) ::= tell(c) →φ cutψ (A′ ) when φ 6⊏ ψ,
In the first case, the statement holds by initial hypothesis over A. In the second
case, since by hypothesis we have σ ⊗ c 6⊏ ψ, again the statement holds by the
definition of the tell transition rules of Table III.
It is now interesting to notice that the thresholds appearing in the program are
related to the final computed stores:
Theorem 6.3 (thresholds). Consider an scc computation
hA, 1i → hA1 , σ1 i → . . . hAn , σn i → hsuccess, σi
for a program P . Then, also
hcutσ (A), 1i → hcutσ (A1 ), σ1 i → . . . hcutσ (An ), σn i → hsuccess, σi
is an scc computation for program P .
Proof. First of all, notice that during the computation an agent can only add
constraints to the store. So, since × is extensive, the store can only monotonically
decrease starting from the initial store 1 and ending in the final store σ. So we
have
1 ⊒ σ1 . . . ⊒ σn ⊒ σ.
Now, the statement follows by applying at each step the results of Lemma 6.2. In
fact, at each step the hypothesis of the lemma hold:
—the cut σ is always lower than the current store (σ ⊑ σi ⊗ c);
—the ask and tell operations can be applied (moving from agent Ai to agent Ai +1).
6.1 Capturing Success Computations.
Given the transition system as defined in the previous section, we now define what
we want to observe of the program behaviour as described by the transitions. To
do this, we define for each agent A the set of constraints
SA = {σ ⇓var(A) | hA, 1i →∗ hsuccess, σi}
that collects the results of the successful computations that the agent can perform.
Notice that the computed store σ is projected over the variables of the agent A to
discard any fresh variable introduced in the store by the ∃ operator.
The observable SA could be refined by considering, instead of the set of successful computations starting from hA, 1i, only a subset of them. For example,
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
Soft Concurrent Constraint Programming
·
17
one could be interested in considering only the best computations: in this case,
all the computations leading to a store worse than one already collected are disregarded. With a pessimistic view, the representative subset could instead collect
all the worst computations (that is, all the computations better than others are
disregarded). Finally, also a set containing both the best and the worst computations could be considered. These options are reminiscent of Hoare, Smith and
Egli-Milner powerdomains respectively [Plotkin 1981].
At this stage, the difference between don’t know and don’t care nondeterminism
arises only in the way the observables are interpreted: in a don’t care approach,
agent A can commit to one of the final stores σ ⇓var(A) , while, in a don’t know
approach, in classical cc programming it is enough that one of the final stores is
consistent. Since existential quantification corresponds to the sum in our semiringbased approach, for us a don’t know approach leads to the sum (that is, the lub)
of all final stores:
M
σ.
Sdk
A =
σ∈SA
It is now interesting to notice that the thresholds appearing in the program are
related also to the observable sets:
Proposition 6.4 (Thresholds and SA (1)). For each ψ, we have SA ⊇
Scutψ (A) .
Proof. By definition of cuts (Definition 6.1), we can modify the agents only by
changing the thresholds with a new level, greater than the previous one. So, easily,
we can only cut away some computations.
dk
Corollary 6.5 (Thresholds and Sdk
⊇
A (1)). For each ψ, we have SA
Sdk
cutψ (A) .
Proof. It follows from the definition of Sdk
A and from Proposition 6.4.
Theorem 6.6 (Thresholds and SA (2)). Let ψ ⊑ glb{σ ∈ SA }. Then SA =
Scutψ (A) .
Proof. By Proposition 6.4, we have SA ⊆ Scutψ (A) . Moreover, since ψ is lower
than all σ in SA , by Theorem 6.3 we have that all the computations are also in
Scutψ (A) . So, the statement follows.
Notice that, thanks to Theorem 6.6 and to Proposition 6.4, whenever we have
a lower bound ψ of the glb of the final solutions, we can use ψ as a threshold to
eliminate some computations. Moreover, we can prove the following theorem:
Theorem 6.7. Let σ ∈ SA and σ 6∈ Scutψ (A) . Then we have σ ⊏ ψ.
Proof. If σ ∈ SA and σ 6∈ Scutψ (A) , it means that the cut eliminates some
computations. So, at some step we have changed the threshold of some tell or ask
agent. In particular, since we know by Theorem 6.3 that when ψ ⊑ σ we do not
modify the computation, we need ψ 6⊑ σ. Moreover, since the tell and ask rules fail
only if σ ⊏ ψ, we easily have the statement of the theorem.
The following theorem relates thresholds and Sdk
A .
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
18
·
Stefano Bistarelli et al.
′
Theorem 6.8 (Thresholds and Sdk
∈
A (2)). Let ΨA = {σ ∈ SA |6 ∃σ
SA with σ ′ ⊒ σ} (that is, ΨA is the set of “greatest” elements of SA ). Let also
dk
ψ ⊑ glb{σ ∈ ΨA }. Then Sdk
A = Scutψ (A) .
L
σ∈SA σ =
LProof. Since we have a + b = b ⇐⇒ a ≤ b, we easily have
σ.
Now,
by
following
a
reasoning
similar
to
Theorem
6.6,
by
applying
a
σ∈ΨA
cut with a threshold
ψ
⊑
glb{σ
∈
Ψ
}
we
do
not
eliminate
any
computation.
So
A
L
L
dk
we obtain Sdk
A =
σ∈ΨA σ = Scutψ (A)
σ∈SA σ =
Lemma 6.9. Given any constraint ψ, we have:
dk
Sdk
A ⊑ ψ + Scutψ (A) .
dk
Proof. Let S be the set of all solutions; then Sdk
A = lub(S) and Scutψ (A) =
lub(S1 ) where S1 ⊆ S. The solutions that have been eliminated by the cut ψ (that
is all the σ ∈ S − S1 ) are all lower than ψ by Theorem 6.7. So, it easily follows that
dk
Sdk
A ⊑ ψ + Scutψ (A) .
Theorem 6.10. Given any constraint ψ, we have:
dk
dk
Sdk
cutψ (A) ⊑ SA ⊑ ψ + Scutψ (A) .
dk
Proof. From Corollary 6.5, we have Sdk
cutψ (A) ⊑ SA . From Lemma 6.9 we have
dk
dk
instead SA ⊑ ψ + Scutψ (A) .
This theorem suggests a way to cut useless computations while generating the
observable Sdk
A of an scc program P starting from agent A. A very naive way to
obtain such an observable would be to first generate all final states, of the form
hsuccess, σi i, and then compute their lub. An alternative, smarter way to compute
this same observable would be to do the following. First we start executing the
program as it is, and find a first solution, say σ1 . Then we restart the execution
applying the cut level σ1 .
By Theorem 6.8, this new cut level cannot eliminate solutions which influence
the computation of the observable: the only solutions it will cut are those that are
lower than the one we already found, thus useless in terms of the computation of
Sdk
A .
In general, after having found solutions σ1 , . . . , σk , we restart execution with cut
level ψ = σ1 + . . .+ σk . Again, this will not cut crucial solutions but only some that
are lower than the sum of those already found. When the execution of the program
terminates with no solution we can be sure that the cut level just used (which is
the sum of all solutions found) is the desired observable (in fact, by Theorem 6.10
dk
dk
when Sdk
cutψ (A) = ψ we necessarily have Scutψ (A) = SA = ψ).
In a way, such an execution method resembles a branch & bound strategy, where
the cut levels have the role of the bounds.
The following corollary is important to show the correctness of this approach.
Corollary 6.11. Given any constraint ψ ⊑ Sdk
A , we have:
dk
Sdk
A = ψ + Scutψ (A) .
Proof. It easily comes from Theorem 6.10.
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
Soft Concurrent Constraint Programming
Table IV.
·
19
Failure in the scc language
σ⊗c ⊏φ
(Tell1 )
htell(c) →φ A, σi −→ f ail
(σ ⊗ c) ⇓∅ < a
htell(c) →a A, σi −→ f ail
σ ⊏φ
(Valued-tell1 )
(Ask1 )
hask(c) →φ A, σi −→ f ail
σ ⇓∅ < a
hask(c) →a A, σi −→ f ail
hE1 , σi −→ f ail, hE2 , σi −→ f ail
hE1 + E2 , σi −→ f ail
hE2 + E1 , σi −→ f ail
hA1 , σi −→ f ail
hA1 kA2 , σi −→ f ail
hA2 kA1 , σi −→ f ail
(Valued-ask1 )
(Nondeterminism1 )
(Parallelism1 )
Let us now use this corollary to prove the correctness of the whole procedure
above.
Let σ1 be the first final state reached by agent A. By stopping the algorithm
dk
after one step, what we have to prove is Sdk
A = σ1 + Scutσ (A) . Since σ1 is for sure
1
lower than Sdk
A , this is true by Corollary 6.11.
By applying this procedure iteratively, we will collect a superset Ψ′A of ΨA =
{σ ∈ SA |6 ∃σ ′ ∈ SA with σ ′ ⊒ σ} (Ψ′A is a superset of ΨA because we could collect
a final state σi before computing a final state σj ⊒ σi ; in this
L will be in
L case both
Ψ′A ). Even if Ψ′A contains more elements then ΨA , we have σ∈Ψ′ = σ∈ΨA (for
A
the extensivity and idempotence properties of +).
The only difference with the procedure we have tested correct w.r.t. the algorithm
is that, at each step, it performs a cut by using the sum of all the previously
computed final state. This means that the algorithm can at each step eliminate
more computations, but by the results of Theorem 6.7 the eliminated computations
does not change the final result.
6.2 Failure
The transition system we have defined considers only successful computations. If
this could be a reasonable choice in a don’t know interpretation of the language it
will lead to an insufficient analysis of the behaviour in a pessimistic interpretation
of the indeterminism. To capture agents’ failure, we add the terminal fail to the
configurations and the transition rules of Table IV to those of Table III.
(Valued)tell1 /ask1 . The failing rule for ask and tell simply checks if the
added/checked constraint c is inconsistent with the store σ and in this case stops
the computation and gives fail as a result. Note that since we use soft constraints
we enriched this operator with a threshold (a or φ). This is used also to compute
failure. If the level of consistency of the resulting store is lower than the threshold
level, then this is considered a failure.
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
20
·
Stefano Bistarelli et al.
Table V.
Hanging computations in the scc language
σ 6⊢ c, σ 6⊏ φ
(Ask2 )
hask(c) →φ A, σi −→ hang
σ 6⊢ c, σ ⇓∅ 6< a
(Valued-ask2 )
hask(c) →a A, σi −→ hang
hE1 , σi −→ f ail/hang, hE2 , σi −→ hang
hE1 + E2 , σi −→ hang
hE2 + E1 , σi −→ hang
(Nondeterminism2 )
Nondeterminism1 . Since the failure of a branch arises only from the failure of
a guard, and since we use angelic non-determinism (that is, we check the guards
before choosing one path), we fail only when all the branches fail.
Parallelism1 . In this case the computation fails as soon as one of the branches
ails.
The observables of each agent can now be enlarged by using the function
FA = {f ail | hA, 1V i →∗ f ail}
that computes a failure if at least a computation of agent A fails.
By considering also the failing computations, the difference between don’t know
and don’t care becomes finer. In fact, in situations where we have SA = Sdk
A ,
the failing computations could make the difference: in the don’t care approach the
notion of failure is existential and in the don’t know one becomes universal [de Boer
and Palamidessi 1994]:
dk
FA
= {f ail | all computations for A which lead to f ail}.
This means that in the don’t know nondeterminism we are interested in observing
a failure only if all the branches fail. In this way, given an agent A with an empty
dk
Sdk
A and a non-empty FA , we cannot say for sure that the semantic of this agent is
f ail. In fact, the transition rules we have defined do not consider hang and infinite
computations. Similar semidecibility results for soft constraint logic programming
are proven in [Bistarelli et al. 2001].
6.3 Hanging and infinite computations
To complete the possible observables of a goal, we need also to observe the hanged
states or those representing infinite computation. To this extent, we extend the
configurations with the terminals hang and ⊥, and we add some transition rules
(those in Table V) to handle hanging computations.
Nondeterminism. The only case that can lead the system to a hanging state is
when all the branches are stuck. In this case, we can assume no future change of
the state will happen that give the possibility to the agents to evolve.
To deal with hang states and infinite computations, we enlarged the observables
with the functions
HA = {hang | hA, 1i →∗ hang}
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
Soft Concurrent Constraint Programming
·
21
that collects all the hang computations of the agent A and
DA = {⊥ | hA, 1i diverges}
to represent infinite computations.
7. AN EXAMPLE FROM THE NETWORK SCENARIO
We consider in this section a simple network problem, involving a set of processes
running on distinct locations and sharing some variables, over which they need to
synchronize, and we show how to model and solve such a problem in scc.
Each process is connected to a set of variables, shared with other processes, and
it can perform several moves. Each of such moves involves performing an action
over some or all the variables connected to the process. An action over a variable
consists of giving a certain value to that variable. A special value “idle” models the
fact that a process does not perfom any action over a variable. Each process has
also the possibility of not moving at all: in this case, all its variables are given the
idle value.
The desired behavior of a network of such processes is that, at each move of the
entire network:
(1) processes sharing a variable perform the same action over it;
(2) as few processes as possible remain idle.
To describe a network of processes with these features, we use an SCSP where
each variable models a shared variable, and each constraint models a process and
connects the variables corresponding to the shared variables of that process. The
domain of each variable in this SCSP is the set of all possible actions, including the
idle one. Each way of satisfying a constraint is therefore a tuple of actions that a
process can perform on the corresponding shared variables.
In this scenario, softness can be introduced both in the domains and in the
constraints. In particular, since we prefer to have as many moving processes as
possible, we can associate a penalty to both the idle element in the domains, and
to tuples containing the idle action in the constraints. As for the other domain
elements and constraint tuples, we can assign them suitable preference values to
model how much we like that action or that process move.
For example, we can use the semiring S = h[−∞, 0], max, +, −∞, 0i, where 0 is
the best preference level (or, said dually, the weakest penalty), −∞ is the worst
level, and preferences (or penalties) are combined by summing them. According
to this semiring, we can assign value −∞ to the idle action or move, and suitable
other preference levels to the other values and moves. Figure 2 gives the details
of a part of a network and it shows eight processes (that is, c1 , . . . , c8 ) sharing a
total of six variables. In this example, we assume that processes c1 , c2 and c3 are
located on site a, processes c5 and c6 are located on site b, and c4 is located on
site c. Processes c7 and c8 are located on site d. Site e connects this part of the
network to the rest. Therefore, for example, variables xd , yd and zd are shared
between processes located in distinct locations.
As desired, finding the best solution for the SCSP representing the current state
of the process network means finding a move for all the processes such that they
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
22
·
Stefano Bistarelli et al.
u_a
e
x_d
c_1 c_3
a
c_2
c_7
y_d
c
d
c_8
c_4
c_6
w_c
c_5
z_d
v_b
b
Fig. 2.
The SCSP describing part of a process network
perform the same action on the shared variables and there is a minimum number
of idle processes. However, since the problem is inherently distributed, it does not
make sense, and it might not even be possible, to centralize all the information and
give it to a single soft constraint solver.
On the contrary, it may be more reasonable to use several soft constraint solvers,
one for each network location, which will take care of handling only the constraints
present in that location. Then, the interaction between processes in different locations, and the necessary agreement to solve the entire problem, will be modelled via
the scc framework, where each agent will represent the behaviour of the processes
in one location.
More precisely, each scc agent (and underlying soft constraint solver) will be in
charge of receiving the necessary information from the other agents (via suitable
asks) and using it to achieve the synchronization of the processes in its location.
For this protocol to work, that is, for obtaining a global optimal solution without
a centralization of the work, the SCSP describing the network of processes has
to have a tree-like shape, where each node of the tree contains all the processes
in a location, and the agents have to communicate from the bottom of the tree
to its root. In fact, the proposed protocol uses a sort of Dynamic Programming
technique to distribute the computation between the locations. In this case the
use of a tree shape allows us to work, at each step of the algorithm, only locally
to one of the locations. In fact, a non tree shape would lead to the construction
of non-local constraints and thus require computations which involve more than
one location at a time. In our example, the tree structure we will use is the one
shown in Figure 3(a), which also shows the direction of the child-parent relation
links (via arrows). Figure 3(b) describes instead the partition of the SCSP over the
four involved locations. The gray connections represent the synchronization to be
assured between distinct locations. Notice that, w.r.t. Figure 2, we have duplicated
the variables representing variables shared between distinct locations, because of
our desire to first perform a local work and then to communicate the results to the
other locations.
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
·
Soft Concurrent Constraint Programming
23
5
u_a
a
4
c_1
1
1
x_a
x_e
c_2
c_3
y_a
3
a
1
x_d
2
d
1
x_c
d
y_b
c_8
1
c_6
c_4
c
c
z_d
w_c
b
(a) A possible tree structure for our network.
Fig. 3.
c_7
y_d
1
b
z_c
z_b
c_5
v_b
(b) The SCSP partitioned over the four locations.
The ordered process network
The scc agents (one for each location plus the parallel composition of all of them)
are therefore defined as follows:
Aa : ∃ua (tell(c1 (xa , ua ) ∧ c2 (ua , ya ) ∧ c3 (xa , ya )) → tell(enda = true) → stop)
Ab : ∃vb (tell(c5 (yb , vb ) ∧ c6 (zb , vb )) → tell(endb = true) → stop)
Ac : ∃wc (tell(c4 (xc , wc , zc )) → tell(endc = true) → stop)
Ad : ask(enda = true ∧ endb = true ∧ endc = true ∧ endd = true) →
tell(c7 (xd , yd ) ∧ c8 (xd , yd , zd ) ∧ xa = xd = xc ∧ ya = yd = yb ∧ zb = zd = zc )
→ tell(endd = true) → stop
A : Aa | Ab | Ac | Ad
Agents Aa ,Ab ,Ac and Ad represent the processes running respectively in the
location a, b, c and d. Note that, at each ask or tell, the underlying soft constraint
solver will only check (for consistency or entailment) a part of the current set of
constraints: those local to one location. Due to the tree structure chosen for this
example, where agents Aa , Ab , and Ac correspond to leaf locations, only agent
Ad shows all the actions of a generic process: first it needs to collect the results
computed separately by the other agents (via the ask); then it performs its own
constraint solving (via a tell), and finally it can set its end flag, that will be used by
a parent agent (in this case the agent corresponding to location e, which we have
not modelled here).
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
24
·
Stefano Bistarelli et al.
8. CONCLUSIONS AND FUTURE WORK
We have shown that cc languages can deal with soft constraints. Moreover, we have
extended their syntax to use soft constraints also to direct and prune the search
process at the language level. We believe that such a new programming paradigm
could be very useful for web and internet programming.
In fact, in several network-related areas, constraints are already being used [Bella
and Bistarelli 2001; Awduche et al. 1999; Clark 1989; Jain and Sun 2000; Calisti
and Faltings 2000]. The soft constraint framework has the advantage over the classical one of selecting a “best” solution also in overconstrained or underconstrained
systems. Moreover, the need to express preferences and to search for optimal solutions shows that soft constraints can improve the modelling of web interaction
scenarios.
ACKNOWLEDGMENTS
We are indebted to Paolo Baldan for valuable suggestions.
REFERENCES
Awduche, D., Malcolm, J., Agogbua, J., O’Dell, M., and McManus, J. 1999. RFC2702:
Requirements for traffic engineering over mpls. Tech. rep., Network Working Group. Sept.
Bella, G. and Bistarelli, S. 2001. Soft constraints for security protocol analysis: Confidentiality. In Proceedings of the 3rd International Symposium on Practical Aspects of Declarative
Languages (PADL ’01), I. Ramakrishnan, Ed. LNCS, vol. 1990. Springer-Verlag, Heidelberg,
Germany, 108–122.
Bistarelli, S. 2001. Soft constraint solving and programming: a general framework. Ph.D. thesis,
Dipartimento di Informatica, Università di Pisa, Italy. TD-2/01.
Bistarelli, S., Codognet, P., and Rossi, F. 2000b. Abstracting soft constraints. In Proceedings
of the 1999 ERCIM/Compulog Net workshop on Constraints, K. Apt, E. Monfroy, T. Kakas,
and F. Rossi, Eds. LNCS, vol. 1865. Springer, Heidelberg, Germany.
Bistarelli, S., Codognet, P., and Rossi, F. 2000a. An abstraction framework for soft constraints and its relationship with constraint propagation. In Proceedings of the Symposium on
Abstraction, Reformulation and Approximation (SARA2000), B. Y. Chouery and T. Walsh,
Eds. LNAI, vol. 1864. Springer, Heidelberg, Germany.
Bistarelli, S., Montanari, U., and Rossi, F. 1995. Constraint Solving over Semirings. In
Proceedings of the 14th International Joint Conference on Artificial Intelligence (IJCAI’95).
Morgan Kaufman, San Francisco, CA, USA.
Bistarelli, S., Montanari, U., and Rossi, F. 1997. Semiring-based Constraint Solving and
Optimization. Journal of the ACM 44, 2 (Mar), 201–236.
Bistarelli, S., Montanari, U., and Rossi, F. 2001. Semiring-based Constraint Logic Programming: Syntax and Semantics. ACM Trans. Program. Lang. Syst. 23, 1–29.
Calisti, M. and Faltings, B. 2000. Distributed constrained agents for allocating service demands
in multi-provider networks,. Journal of the Italian Operational Research Society XXIX, 91.
Special Issue on Constraint-Based Problem Solving.
Chen, S. and Nahrstedt, K. 1998. Distributed QoS routing with imprecise state information.
In ICCCCN98.
Clark, D. 1989. RFC1102: Policy routing in internet protocols. Tech. rep., Network Working
Group. May.
de Boer, F. and Palamidessi, C. 1991. A fully abstract model for concurrent constraint
programming. In Proceeding of the 16th Colloquium on Trees in Algebra and Programming
(CAAP1991), S. Abramsky and T. Maibaum, Eds. Vol. 493. Springer-Verlag, Heidelberg, Germany.
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
Soft Concurrent Constraint Programming
·
25
de Boer, F. and Palamidessi, C. 1994. From Concurrent Logic Programming to Concurrent
Constraint Programming. In Advances in Logic Programming Theory, G. Levi, Ed. Oxford
University Press, 55–113.
Dubois, D., Fargier, H., and Prade, H. 1993. The calculus of fuzzy restrictions as a basis for
flexible constraint satisfaction. In Proceedings of the 2nd IEEE International Conference on
Fuzzy Systems (FUZZ-IEEE 1993). IEEE, Piscataway, NJ, U.S.A., 1131–1136.
Fargier, H. and Lang, J. 1993. Uncertainty in constraint satisfaction problems: a probabilistic
approach. In Proceeding of the 2nd European Conference on Symbolic and Quantitative Approaches to Reasoning with Uncertainty (ECSQARU1993). LNCS, vol. 747. Springer-Verlag,
Heidelberg, Germany, 97–104.
Freuder, E. and Wallace, R. 1992. Partial constraint satisfaction. Artificial Intelligence
Journal 58.
Jain, R. and Sun, W. 2000. QoS/Policy/Constraint-based routing. In Carrier IP Telephony 2000
Comprehensive Report. International Engineering Consortium, Heidelberg, Germany. ISBN: 0933217-75-7.
Plotkin, G. 1981. Post-graduate lecture notes in advanced domain theory (incorporating the
pisa lecture notes). Technical report, Dept. of Computer Science, Univ. of Edinburgh.
Ruttkay, Z. 1994. Fuzzy constraint satisfaction. In Proceedings of the 3rd IEEE International
Conference on Fuzzy Systems (FUZZ-IEEE 1994). 1263–1268.
Saraswat, V. 1993. Concurrent Constraint Programming. MIT Press.
Schiex, T. 1992. Possibilistic constraint satisfaction problems, or “how to handle soft constraints?”. In Proceeding of the 8th Conference on Uncertainty in Artificial Intelligence
(UAI1992). 269–275.
Schiex, T., Fargier, H., and Verfaille, G. 1995. Valued Constraint Satisfaction Problems: Hard
and Easy Problems. In Proceedings of the 14th International Joint Conference on Artificial
Intelligence (IJCAI’95). Morgan Kaufmann, San Francisco, CA, USA, 631–637.
...
ACM Transactions on Computational Logic, Vol. V, No. N, March 2018.
| 6cs.PL
|
Invariant Clusters for Hybrid Systems
Hui Kong∗, Sergiy Bogomolov∗, Christian Schilling† , Yu Jiang‡, Thomas A. Henzinger∗
∗ IST
Austria, Klosterneuburg, Austria
of Freiburg, Freiburg, Germany
‡ University of Illinois at Urbana-Champaign, Illinois, USA
arXiv:1605.01450v1 [math.OC] 4 May 2016
† University
Abstract—In this paper, we propose an approach to automatically compute invariant clusters for semialgebraic hybrid systems.
An invariant cluster for an ordinary differential equation (ODE)
is a multivariate polynomial invariant g(~u, ~x) = 0, parametric
in ~u, which can yield an infinite number of concrete invariants
by assigning different values to ~u so that every trajectory of the
system can be overapproximated precisely by a union of concrete
invariants. For semialgebraic systems, which involve ODEs with
multivariate polynomial vector flow, invariant clusters can be
obtained by first computing the remainder of the Lie derivative
of a template multivariate polynomial w.r.t. its Gröbner basis and
then solving the system of polynomial equations obtained from
the coefficients of the remainder. Based on invariant clusters and
sum-of-squares (SOS) programming, we present a new method
for the safety verification of hybrid systems. Experiments on
nonlinear benchmark systems from biology and control theory
show that our approach is effective and efficient.
Index Terms—ybrid system, nonlinear system, semialgebraic
system, invariant, safety verification, SOS programmingybrid
system, nonlinear system, semialgebraic system, invariant, safety
verification, SOS programmingh
I. I NTRODUCTION
Hybrid systems [1] are models for systems with interacting
discrete and continuous dynamics. Safety verification is among
the most challenging problems in verifying hybrid systems,
asking whether a set of bad states can be reached from a set
of initial states. The safety verification problem for systems
described by nonlinear differential equations is particularly
complicated because computing the exact reachable set is
usually intractable. Existing approaches are mainly based
on approximate reachable set computations [2], [3], [4] and
abstraction [5], [6], [7], [8], [9].
An invariant is a special kind of overapproximation for the
reachable set of a system. Since invariants do not involve
direct computation of the reachable set, they are especially
suitable for dealing with nonlinear hybrid systems. However,
automatically and efficiently generating sufficiently strong
invariants is challenging on its own.
In this work, we propose an approach to automatically compute invariant clusters for a class of nonlinear semialgebraic
systems whose trajectories are algebraic, i.e., every trajectory
of the system is essentially a common zero set (algebraic
variety) of a set of multivariate polynomial equations. An
invariant cluster for a semialgebraic system is a parameterized
multivariate polynomial invariant g(~u, ~x) = 0, with parameter
~u, which can yield an infinite number of concrete invariants
by assigning different values to ~u so that every trajectory of
the system can be overapproximated precisely by a union of
concrete invariants.
The basic idea of computing invariant clusters is as follows.
A sufficient condition for a trajectory of a semialgebraic
system to start from and to always stay in the solution set
of a multivariate polynomial equation g(~x) = 0 is that the
Lie derivative of g(~x) w.r.t. f~ belongs to the ideal generated
by g(~x), i.e., Lf~g ∈ hg(~x)i, where f~ is the vector flow of
the system. Therefore, if some g(~x) satisfies this condition,
g(~x) = 0 is an invariant of the system. Then, according to
Gröbner basis theory, Lf~g ∈ hg(~x)i implies that the remainder
of Lf~g w.r.t. hg(~x)i must be identical to 0. Based on this
theory, we first set up a template polynomial g(~u, ~x) and then
compute the remainder r(~u, ~x) of Lf~g w.r.t. hg(~u, ~x)i. Since
r(~u, ~x) ≡ 0 implies that all coefficients ai (~u) of ~x in r(~u, ~x)
are equal to zero, we can set up a system P of polynomial
equations on ~u from the coefficients ai (~u). By solving P we
get a set C of constraints on ~u. For those elements in C that
are linear in ~u, the corresponding parameterized polynomial
equations g(~u, ~x) = 0 form invariant clusters. Based on
invariant clusters and SOS programming, we propose a new
method for the safety verification of hybrid systems.
The main contributions of this paper are as follows: 1)
We propose to generate invariant clusters for semialgebraic
systems based on computing the remainder of the Lie derivative of a template polynomial w.r.t. its Gröbner basis and
solving the system of polynomial equations obtained from the
coefficients of the remainder. Our approach avoids Gröbner
basis computation and first-order quantifier elimination. 2) We
present a method to overapproximate trajectories precisely by
using invariant clusters. 3) We apply invariant clusters to the
safety verification of semialgebraic hybrid systems based on
SOS programming. 4) We implemented a prototype tool to
perform the aforementioned steps automatically. Experiments
show that our approach is effective and efficient.
The paper is organized as follows. Section II is devoted to
the preliminaries. In Section III, we introduce the approach
to computing invariant clusters and using them to characterize
trajectories. In Section IV, we present a method to verify safety
properties for semialgebraic continuous and hybrid systems
based on invariant clusters. In Section V, we present our
experimental results. In Section VI, we introduce some related
works. Finally, we conclude our paper in Section VII.
II. P RELIMINARIES
In this section, we recall some backgrounds we need
throughout the paper. We first clarify some notation conventions. If not specified otherwise, we decorate vectors ~·, we
use the symbol K for a field, R for the real number field, C
for the complex number field (which is algebraically closed)
and N for the set of natural numbers, and all the polynomials
involved are multivariate polynomials. In addition, for all the
polynomials g(~u, ~x), we denote by ~u the vector composed of
all the ui and denote by ~x the vector composed of all the
remaining variables that occur in the polynomial.
Definition 1 (Ideal): [10] A subset I of K[~x], is called an
ideal if 1) 0 ∈ I; 2) if p, q ∈ I, then p + q ∈ I; and 3) if
p ∈ I and q ∈ K[~x], then pq ∈ I.
Definition 2 (Generated ideal): [10] Let g1 , . . . , gs be
polynomials in K[~x]. The ideal generated by {g1 , . . . , gs }
is
s
X
def
hg1 , . . . , gs i =
hi gi | h1 , . . . , hs ∈ K[~x] .
i=1
Definition 3 (Algebraic variety): Let K be an algebraically
closed field and I ⊂ K[~x] be an ideal. We define the algebraic
variety of I as
def
V(I) = {~x ∈ Kn | f (~x) = 0 for f ∈ I}.
Next, we present the notation of the Lie derivative, which
is widely used in the discipline of differential geometry.
Definition 4 (Lie derivative): For a given polynomial p ∈
K[~x] over ~x = (x1 , · · · , xn ) and a continuous system ~x˙ = f~,
where f~ = (f1 , . . . , fn ), the Lie derivative of p ∈ K[~x] along
f of order k is defined as follows.
(
p,
k=0
k def
(1)
Lf~p =
p
Pn ∂Lfk−1
~
· fi , k ≥ 1
i=1
∂xi
Essentially, the k-th order Lie derivative of p is the k-th
derivative of p w.r.t. time, i.e., reflects the change of p over
time. We write Lf~p for L1f~p.
We furthermore use the following theorem for deciding
the existence of a real solution of a system of polynomial
constraints.
Theorem 1 (Real Nullstellensatz): [11] The system of multivariate polynomial equations and inequalities p1 (~x) = 0, . . . ,
pm1 (~x) = 0, q1 (~x) ≥ 0, . . . , qm2 (~x) ≥ 0, r1 (~x) > 0, . . . ,
rm3 (~x) > 0 either has a solution in Rn , or there exists the
following polynomial identity
m1
X
βi p i +
v∈{0,1}m2
i=1
+
X
X
v∈{0,1}m3
X
t
ctv
X
t
2
·
m3
m2
Y
2 Y
vj
qj +
rkdk
btv ·
j=1
m3
Y
k=1
rkvk
+
k=1
X
s2w
(2)
=0
w
where dk ∈ N and pi , qj , rk , βj , btv , ctv , sw are polynomials
in R[~x].
Remark 1: Theorem 1 enables us to efficiently decide if
a system of polynomial equations
P and inequalities has a real
solution. By moving the term w s2w in equation (2) to the
right-hand side and denoting
P 2the remaining terms by R(~x, ~y),
x, ~y )
we have −R(~x, ~y) =
w sw , which means that −R(~
is a sum-of-squares. Therefore, finding a set of polynomials
βj , rk , btv , ctv , sw as well as some dk ’s that make −R(~x, ~y )
a sum-of-squares is sufficient to prove that the system has
no real solution, which can be done efficiently by SOS
programming [12].
In this paper, we focus on semialgebraic continuous and hybrid systems which are defined in the following, respectively.
Definition 5 (Semialgebraic system): A semialgebraic sysdef
tem is a tuple M = hX, f~, X0 i, where
1) X is the state space of the system M ,
2) f~ is a Lipschitz continuous vector flow function, and
3) X0 is the initial set described by a semialgebraic set.
The Lipschitz continuity guarantees existence and uniqueness of the differential equation ~x˙ = f~. A trajectory of a
semialgebraic system is defined as follows.
Definition 6 (Trajectory): Let M be a semialgebraic system. A trajectory originating from a point ~x0 ∈ X0 to
time T > 0 is a continuous and differentiable function
~x(t) : [0, T ) → Rn such that i) ~x(0) = ~x0 , and ii)
x
~ x(τ )).
∀τ ∈ [0, T ) : d~
dt t=τ = f (~
Definition 7 (Safety): Given an unsafe set XU , a semialgebraic system M = hX, f~, X0 i is said to be safe if no trajectory
~x(t) of M satisfies both ~x(0) ∈ X0 and ∃τ ∈ R≥0 : ~x(τ ) ∈
XU .
Definition 8 (Hybrid System): A hybrid system is dedef
scribed by a tuple H = hL, X, E, R, G, I, F i, where
• L is a finite set of locations (or modes);
n
• X ⊆ R is the continuous state space. The hybrid state
space of the system is denoted by X = L × X and a state
is denoted by (l, ~x) ∈ X ;
• E ⊆ L × L is a set of discrete transitions;
X
• G : E → 2 is a guard mapping over discrete transitions;
X
• R : E × 2
→ 2X is a reset mapping over discrete
transitions;
X
• I :L →2
is an invariant mapping;
• F : L → (X → X) is a vector field mapping which
assigns to each location l a vector field f~.
The transition and dynamic structure of the hybrid system
defines a set of trajectories. A trajectory is a sequence originating from a state (l0 , ~x0 ) ∈ X0 , where X0 ⊆ X is an initial set,
and consisting of a series of interleaved continuous flows and
discrete transitions. During the continuous flows, the system
evolves following the vector field F (l) at some location l ∈ L
as long as the invariant condition I(l) is not violated. At some
state (l, ~x), if there is a discrete transition (l, l′ ) ∈ E such
that (l, ~x) ∈ G(l, l′ ) (we write G(l, l′ ) for G((l, l′ ))), then
the discrete transition can be taken and the system state can
be reset to R(l, l′ , ~x). The problem of safety verification of a
hybrid system is to prove that the hybrid system cannot reach
an unsafe set XU from an initial set X0 .
III. C OMPUTATION
OF I NVARIANT
C LUSTERS
In this section, we first introduce the notion of invariant
cluster and then show how to compute a set of invariant
clusters and how to use it to represent every trajectory of a
semialgebraic system.
A. Foundations of invariants and invariant clusters
Given a semialgebraic system M , for any trajectory ~x(t)
of M , if we can find a multivariate polynomial g(~x) ∈ R[~x]
such that g(~x(0)) ∼ 0 implies g(~x(t)) ∼ 0 for all t > 0,
where ∼ ∈ {<, ≤, =, ≥, >}, then g(~x) ∼ 0 is an invariant of
the system. For the convenience of presentation, we call g(~x)
an invariant polynomial of M . In addition, a trajectory ~x(t) is
said to be algebraic if there exists an invariant g(~x) = 0 for
~x(t) and g(~x) 6≡ 0. In the following, we present a sufficient
condition for a polynomial g(~x) to be an invariant polynomial.
Proposition 1: Let M = hX, f~, X0 i be a semialgebraic
system and g(~x) ∈ R[~x]. Then g ∼ 0 is an invariant of M for
every ∼ ∈ {<, ≤, =, ≥, >} if g(~x) satisfies
Lf~g ∈ hgi
Fig. 1: Example 2. Curve defined by invariant cluster of
Class(C ∗ , ~x0 ) for ~x0 = (4, 2).
(3)
Proof: See Appendix A.
Proposition 1 states that all the polynomial equations and
inequalities g ∼ 0 are invariants of M if the Lie derivative of
g belongs to the ideal hgi. Note that every invariant satisfying
condition (3) defines a closed set for trajectories, that is, no
trajectory can enter or leave the set defined by the invariant.
For a semialgebraic system whose trajectories are algebraic,
the trajectories can usually be divided into several groups and
in each group all trajectories show similar curves. Essentially,
these similar curves can be described identically by a unique
parameterized polynomial equation, which we characterize
as an invariant cluster. The computation method of invariant
clusters is presented in Subsection III-B.
Definition 9 (Invariant cluster): An invariant cluster C of
a semialgebraic system is a set of invariants which can be uniK
~
formly described
PM as C =i {g(~u, ~x) = 0 | ~u ∈ R \{0}}, where
g(~u, ~x) = i=1 ci (~u)X satisfies Lf~g ∈ hgi and ci (~u) ∈ R[~u]
are fixed linear polynomials on ~u = (u1 , · · · , uK ), X i are
monomials on ~x = (x1 , · · · , xn ), and M, K ∈ N.
Note that by requiring ~u 6= ~0 in Definition 9 and the
following related definitions, we exclude the trivial invariant
0 = 0. Given an invariant cluster, by varying the parameter
~u, we may obtain an infinite set of concrete invariants for
the system. To be intuitive, we present a running example to
demonstrate the related concepts throughout the paper.
Example 1 (running example): Consider
the semialgebraic
system M1 described by [ẋ, ẏ] = y 2 , xy . The set C ∗ =
{u1 − u3 (x2 − y 2 ) = 0 | (u1 , u3 ) ∈ R2 \ {~0}} is an invariant
cluster. It is easy to verify that the polynomial u1 −u3 (x2 −y 2 )
satisfies condition (3) for all (u1 , u3 ) ∈ R2 .
Next we introduce the notion of an invariant class.
Definition 10 (Invariant class): Given a semialgebraic system M with some initial point ~x0 and an invariant cluster
C = {g(~u, ~x) = 0 | ~u ∈ RK \ {~0}} of M , where K ∈ N, an
invariant class of C at ~x0 , denoted by Class(C, ~x0 ), is the
set {g(~u, ~x) = 0 | g(~u, ~x0 ) = 0, ~u ∈ RK \ {~0}}.
Given an invariant cluster C, by substituting a specific point
~x0 for ~x in g(~u, ~x) = 0, we obtain a constraint g(~u, ~x0 ) = 0 on
~u, which yields a subset Class(C, ~x0 ) of C. Apparently, every
Fig. 2: Example 6. X0 : (x + 15)2 + (y − 17)2 ≤ 1, XU :
(x1 − 11)2 + (y1 − 16.5)2 ≤ 1.
member of Class(C, ~x0 ) is an invariant for the trajectory
originating from ~x0 .
Example 2 (running example): For the given invariant cluster C ∗ in Example 1 and a given initial point ~x0 = (4, 2),
we get the invariant class Class(C ∗ , ~x0 ) = {u1 − u3 (x2 −
y 2 ) = 0 | u1 − 12u3 = 0, (u1 , u3 ) ∈ R2 \ {~0}}. Every
member of Class(C ∗ , ~x0 ) is an invariant for the trajectory
of M1 originating from ~x0 . The algebraic variety defined by
Class(C ∗ , ~x0 ) is shown in Figure 1.
An invariant class has the following important properties.
Theorem 2: Given an n-dimensional semialgebraic system
M and an invariant class D = {g(~u, ~x) = 0 | g(~u, ~x0 ) = 0,
~u ∈ RK \ {~0}} of M at a specified point ~x0 , let Dg be the
set of all invariant polynomials occurring in D and π~x0 be the
trajectory of M originating from ~x0 . Then,
1) π~x0 ⊆ V(Dg );
2) there must exist a subset B of Dg consisting of m members such that hDg i = hBi, where m is the dimension
of the hyperplane g(~u, ~x0 ) = 0 in terms of ~u in RK and
hDg i and hBi denote the ideal generated by the members
of Dg and B, respectively.
Proof: See Appendix B.
Remark 2: The first property in Theorem 2 reveals that a
trajectory π~x0 is always contained in the intersection of all the
invariants in the invariant class D of ~x0 . The second property
concludes that the invariant class can be generated by a finite
subset B of D if it consists of an infinite number of invariants.
The algebraic variety V(B) (which is equivalent to V(Dg ))
forms an overapproximation for π~x0 and the quality of the
overapproximation depends largely on the dimension m of
V(B). More specifically, the lower the dimension is, the better
the overapproximation is. The ideal case is that, when m = 1,
V(B) shrinks to an algebraic curve and hence some part of
the algebraic curve matches the trajectory precisely. In the
case of m > 1, V(B) is usually a hypersurface. To make the
overapproximation less conservative, we may take the union
of multiple invariant classes from different invariant clusters
to reduce the dimension of the algebraic variety.
Algorithm 1: Computation of invariant clusters
input : f : n-dimensional polynomial vector field;
N : upper bound for the degree of invariant
polynomials
output: CFamily: a set of invariant clusters
1
2
3
4
5
6
B. Computing invariant clusters
7
8
In the following, we propose a computation method for
invariant clusters. The main idea arises from the following
proposition.
Proposition 2: Let M = hX, f~, X0 i be a semialgebraic
system and suppose for a given multivariate polynomial g ∈
R[~x] that the polynomial Lf~g can be written as Lf~g = pg + r.
Then Lf~g ∈ hgi if r ≡ 0.
According to Proposition 2, if we can find a polynomial
g(~x) such that the remainder of its Lie derivative w.r.t. g(~x)
is identical to 0, then g(~x) = 0 is an invariant of M . The idea
is as follows. We first establish a template g(~u, ~x) for g(~x)
with parameter ~u and then compute the remainder r(~u, ~x) of
Lf~g w.r.t. g(~u, ~x). According to the computing procedure of
PK
a remainder [10], r(~u, ~x) must be of the form i=1 biu(~du) X i ,
j
where d ∈ N, bi (~u) are homogeneous polynomials of degree
d+1 over ~u, uj is the coefficient of the leading term of g(~u, ~x)
by some specified monomial order of ~x, and X i are monomials
on ~x. Since r(~u, ~x) ≡ 0 implies uj 6= 0 and bi (~u) = 0
for all i = 1, . . . , K, we obtain a system C of homogeneous
polynomial equations on ~u plus uj 6= 0 from the coefficients
of r(~u, ~x). Solving C can provide a set of invariant clusters of
M if it exists. Note that all the aforementioned steps can be
performed automatically in a mathematical software such as
Maple. Pseudocode for computing invariant clusters is shown
in Algorithm 1. The principle for the loop in line 5 is that the
remainder may vary from the monomial order of ~x and hence
produce different solutions. Using multiple orders helps to get
more solutions.
Remark 3 (Complexity): In Algorithm 1, the key steps are
computing the remainder in line 4 and solving the system of
equations on ~u in line 8. The former takes only linear time
and hence is very efficient. The latter involves solving a system
of homogeneous polynomial equations, which is known to be
NP-complete [13]. In our implementation in Maple, we use the
command solve for solving the system of equations. In our
experiments on nonlinear (parametric) systems of dimension
ranging from 2 to 8, the solver can return quickly in most cases
whether they have solutions or not. Among the solutions we
obtained, we found some complex solutions, however, we have
not seen nonlinear solutions.
Example 3 (running example): According to Algorithm 1,
9
10
11
CFamily ← ∅;
for i ← 1 to N do
gu~ ,~x ← generate parameterized polynomial over ~x of
degree i;
Lf g ← compute the Lie derivative of gu~ ,~x ;
foreach monomial order O of ~x do
Ru~ ,~x ← compute remainder of Lf g w.r.t. g by O;
Coeffs ← collect coefficients of ~x in Ru~ ,~x ;
Solution ← solve system Coeffs on ~u;
CFamily ← CFamily ∪ Solution;
end
end
the steps for computing the invariant clusters of degree 2 are
as follows:
1) Generate the template polynomial of degree 2:
g2 (~u, ~x) = u6 x2 + u5 xy + u4 x + u3 y 2 + u2 y + u1
2) Compute the Lie derivative Lf~g2 using Definition 4:
Lf~g2 =
∂g2
∂g2
ẋ +
ẏ = u5 x2 y + (2 u3 + 2 u6 ) xy 2
∂x
∂y
+ u2 xy + y 3 u5 + u4 y 2
3) Compute the remainder of Lf~g2 w.r.t. g2 by graded
reverse lexicographic (grevlex) order of (x, y). Using this
order, the leading term of Lf~g2 and g2 is u5 x2 y and
u6 x2 , respectively. Then:
2 u3 u6 − u25 + 2 u26 xy 2
u5 y
r(~u, ~x) = Lf~g2 −
g2 =
u6
u6
(u2 u6 − u4 u5 ) xy
(−u3 u5 + u5 u6 ) y 3
+
+
u6
u6
u1 u5 y
(−u2 u5 + u4 u6 ) y 2
−
+
u6
u6
4) Collect the coefficients of r(~u, ~x):
u2 u6 − u4 u5 −u3 u5 + u5 u6 −u2 u5 + u4 u6
S :=
,
,
,
u6
u6
u6
u1 u5
2 u3 u6 − u25 + 2 u26
,−
u6
u6
5) Solve the system formed by S. To save space, we just
present one of the six solutions we obtained:
C6 = {u6 = −u3 , u2 = u4 = u5 = 0, u3 = u3 , u1 = u1 }
6) Substitute the above solution C6 for ~u in g2 (~u, ~x). We
get the following parameterized invariant polynomial:
g2 (~u, ~x) = −u3 x2 + u3 y 2 + u1
Note that the other five solutions obtained in step 5) are in
fact the products of the invariant polynomials {u2 y, u1 (x +
y), u1 (x − y)}, which have been obtained when initially
computing the invariants of degree 1. Hence they cannot
increase the expressive power of the set of invariant clusters
and should be dropped. The solution presented above is the
one we have given in Example 1.
Algorithm 2: Computation of invariant classes
input : CFamily: set of invariant clusters;
~x0 : an initial point
output: ICls: list of invariant classes
1
2
3
4
C. Overapproximating trajectories by invariant classes
In this section, we address how to overapproximate trajectories precisely by using invariant classes.
Invariant clusters can be divided into two categories according to the number of invariant classes that they can yield by
varying the parameter ~u. 1) finite invariant cluster: This kind
of invariant cluster can yield only one invariant class no matter
how ~u changes. For example, {u1 (x − y) = 0 | u1 ∈ R \ {0}}
is such an invariant cluster for the running example. In this
case, the trajectories that can be covered by the invariant
class is very limited. Moreover, the overapproximation is also
conservative due to the high dimension of the algebraic variety
defined by the invariant class. 2) infinite invariant cluster:
One such invariant cluster C can yield an infinite number of
invariant classes Class(C, ~x0 ) as the initial point ~x0 varies,
e.g., the invariant cluster C ∗ in Example 1. For the trajectory π~x0 , the overapproximating precision of Class(C, ~x0 )
depends largely on the dimension m of the algebraic variety
defined by Class(C, ~x0 ). The best case is m = 1 and then it
gives a curve-to-curve match in part for the trajectory.
Now, we introduce how to identify the invariant classes for
a given point ~x0 from a set of invariant clusters and how to get
a finite representation for it. To be intuitive, we first present a
3-dimensional system and a set of invariant clusters for it.
Example 4: Consider the following semialgebraic system
M2 : [ẋ, ẏ, ż] = [yz, xz, xy]. We obtain a set of invariant
clusters consisting of 7 elements. Here we only present the
infinite invariant cluster (for other clusters see Appendix I).
C7 = {g7 (~u, ~x) = (−u5 − u6 )x2 + u5 y 2 + u6 z 2 + u0
= 0 | ~u ∈ R3 \ {~0}}
The invariant clusters are capable of overapproximating all the
trajectories of the system M2 . For any given initial state, how
can we identify the invariant classes from the set of invariant
clusters? Suppose we want to find the invariant classes that can
overapproximate the trajectory from the state ~x0 = (1, 2, 3).
According to Theorem 2, we have Algorithm 2 for the purpose.
Remark 4: In Algorithm 2, we enumerate the invariant
clusters to find out which one can provide a non-empty
invariant class Class(C, ~x0 ) for ~x0 . For a Class(C, ~x0 ) to
be nonempty, the corresponding hyperplane g(~u, ~x0 ) = 0 must
have at least one solution to ~u ∈ RK \{~0}, which is equivalent
to that its dimension must be at least 1. For a hyperplane in
RK , its dimension is equal to K − 1. Therefore, Class(C, ~x0 )
must be nonempty if K > 1 and the basis of the hyperplane
can be obtained through a basic linear algebraic computation
5
6
7
8
9
10
11
12
ICls ← ∅;
foreach C ∈ CFamily do
D ← Class(C, ~x0 );
if D 6= ∅ then
m ← the dimension of the hyperplane
g(~u, ~x0 ) = 0 defining D;
if m ≥ 1 then
Basis ← basis {u1 , . . . , um } of the
hyperplane g(~u, ~x0 ) = 0;
D ← the polynomials
{g(~u1 , ~x), . . . , g(~um , ~x)}, ui ∈ Basis;
end
ICls ← ICls ∪ D;
end
end
(which will be illustrated in what follows). However, in the
case of g(~u, ~x0 ) being identical to 0, the hyperplane degenerates to the space RK \ {~0} and the dimension will be K.
Therefore, an invariant class with K = 1 is nonempty iff
g(~u, ~x0 ) is identical to 0. For example, given an invariant
cluster C 0 = {u1 (x − y) = 0 | u1 ∈ R \ {0}} and a point
~x0 = (x0 , y0 ), Class(C 0 , ~x0 ) is empty if x0 6= y0 , however,
Class(C 0 , ~x0 ) is equal to C 0 if x0 = y0 .
Example 5: We continue from Example 4. For the given
point ~x0 = (1, 2, 3), according to Algorithm 2, we find that
only Class(C7 , ~x0 ) = {g7 (~u, x, y, z) = 0 | 3u5 + 8u6 + u0 =
0, ~u ∈ R3 \{~0}} is nonempty. The dimension of the hyperplane
H : 3u5 + 8u6 + u0 = 0 is 2. Since u0 = −3u5 − 8u6 , to
get the basis of H, we can write (u0 , u5 , u6 ) = (−3u5 −
8u6 , u5 , u6 ) = u5 (−3, 1, 0) + u6 (−8, 0, 1). Hence, we have
the following basis for H: {(−3, 1, 0), (−8, 0, 1)}. As a result,
we get the finite representation B = {y 2 − x2 − 3 = 0,
z 2 − x2 − 8 = 0} for Class(C7 , ~x0 ). It is easy to check by
the Maple function HilbertDimension that dim(B) = 1.
Therefore, we finally get an algebraic variety V(B) which
provides in part a curve-to-curve match to the trajectory π~x0 .
The 3-D vector field and the algebraic curve V(B) is shown
in Figure 3a and Figure 3b, respectively.
IV. S AFETY V ERIFICATION BASED
C LUSTERS
ON I NVARIANT
A. Safety Verification of Continuous Systems
In this section, we show how to verify a safety property for
a nonlinear system based on invariant clusters.
In Section III, we have demonstrated that a trajectory can
be overapproximated by an invariant class. Since an invariant
class is determined uniquely by a single hyperplane g(~u, ~x0 ) =
0 in RK for an initial point ~x0 , and a hyperplane without
j2 = m1 + 1, . . . , m1 + m2 , k2 = n1 + 1, . . . , n1 + n2 }.
Then we have the following theorem for deciding the safety
of a semialgebraic system.
Theorem 4: Given a semialgebraic system M = hX, f~, X0 i
and invariant cluster C = {g(~u, ~x) = 0 | ~u ∈ RK \ {~0}} of
M with K ≥ 2. Suppose the normal vector of the hyperplane
g(~u, ~x) = 0 over ~u is (1, ψ1 (~x), . . . , ψK (~x)). Then the system
M is safe if there exists the following polynomial identity
(a)
(b)
Fig. 3: (a) 3D vector field of Example 4. (b) The intersection
of the invariants y 2 − x2 − 3 = 0 (blue) and z 2 − x2 − 8 =
0 (orange) overapproximates the trajectory originating from
~x0 = (1, 2, 3) (green ball).
K
X
k=1
γk · (ψk (~x1 ) − ψk (~x2 )) +
+
v∈{0,1}m1 +m2
+
constant term (this is the case for g(~u, ~x0 ) = 0) is uniquely
determined by its normal vector, we can verify if two states
lie on the same trajectory based on the following theorem.
Theorem 3: Given a semialgebraic system M = hX, f~, X0 i
and an invariant cluster C = {g(~u, ~x) = 0 | ~u =
(u1 , · · · , uKP
) ∈ RK \ {~0}} of M with K > 1, where
K
x)ui and ψi (~x) ∈ R[~x], let X0 be the
g(~u, ~x) =
i=1 ψi (~
initial set and XU be the unsafe set. Then, if there exists a
pair of states (~x1 , ~x2 ) ∈ X0 × XU such that ~x1 and ~x2 lie on
the same trajectory, one of the following two formulae must
hold:
(i)
(ii)
∃k ∈ R\{0} : kψi (~x1 ) = ψi (~x2 ), i = 1, . . . , K
ψi (~x1 ) = ψi (~x2 ) = 0, i = 1, . . . , K
(4)
(5)
Moreover, if some ψi (~x) ≡ 1, i.e. g(~u, ~x) contains a
constant term ui , then formula (4) simplifies to
ψi (~x1 )) = ψi (~x2 ), i = 1, . . . , K
(6)
Proof: See Appendix D.
Remark 5: Instead of computing the invariants explicitly,
Theorem 3 provides an alternative way to verify whether two
states ~x1 , ~x2 lie on the same trajectory by checking that the
difference between the normal vectors of g(~u, ~x1 ) = 0 and
g(~u, ~x2 ) = 0. Let us take Example 4 for illustration. We think
of C7 as a hyperplane over ~u ∈ R3 : u0 + (y 2 − x2 )u5 +
(z 2 − x2 )u6 = 0, hence the corresponding normal vector is
~ (~x) = (1, y 2 − x2 , z 2 − x2 ). Given two random points ~x1 =
N
√ √
(1, 2, 3) and ~x2 = (5, 27, 34), it is easy to verify that
~ (~x1 ) 6= N
~ (~x2 ), which means that ~x1 and ~x2 are not on
N
the same trajectory. This can be verified in another way, as
we know that the invariant class of ~x1 is {y 2 − x2 − 3 =
0, z 2 − x2 − 8 = 0} and ~x2 does not belong to its solution set.
Now we demonstrate how to verify a safety property of
semialgebraic systems. Assume X0 and XU can be written
as semialgebraic sets, i.e., X0 = {~x1 ∈ Rn | pi1 (~x1 ) =
0, qj1 (~x1 ) ≥ 0, rk1 (~x1 ) > 0, i1 = 1, . . . , l1 , j1 = 1 . . . m1 ,
k1 = 1, . . . , n1 } and XU = {~x2 ∈ Rn | pi2 (~x2 ) =
0, qj2 (~x2 ) ≥ 0, rk2 (~x2 ) > 0, i2 = l1 + 1, . . . , l1 + l2 ,
X
X
v∈{0,1}n1 +n2
i=1
m1Y
+m2
X
(
btv )2 ·
t
lX
1 +l2
βi p i
v
qj j
(7)
j=1
n1Y
+n2
n1Y
+n2
X
X
(
ctv )2 ·
rkvk +
s2w +
rkdk = 0
t
k=1
w
k=1
where dk ∈ N and βj , γk , btv , ctv , sw are polynomials in
R[~x1 , ~x2 ].
Proof: See Appendix E.
Remark 6: Theorem 4 transforms the safety verification
problem into a decision problem about the existence of a real
solution of a system of polynomial equations and inequalities.
As noted in Remark 1, this decision problem can be solved
by SOS programming. Our implementation uses the efficient
tool SOSTOOLS [12]. Appendix C contains more information
on SOS programming.
In Theorem 4, we deal with a general semialgebraic system
where the initial set and the unsafe set are represented by a
set of polynomial equations and inequalities. However, if the
system is described by much simpler set representations such
as a single polynomial equation or inequality, the programming
problem can be simplified correspondingly. For example, if
both sets can be represented or overapproximated by a single
polynomial equation I(~x1 ) = 0 and U (~x2 ) = 0, respectively,
then the programming problem is simplified to (see [11])
K
X
αj (ψj (~x1 ) − ψj (~x2 )) + βI + θU − 1 is an SOS (8)
j=1
where (ψ1 (~x), . . . , ψK (~x)) is the same as in Theorem 4 and
αj , β, θ ∈ R[~x1 , ~x2 ]. For how to overapproximate a compact
set we refer the reader to Appendix J. The algorithm for
safety verification based on the condition (8) is shown in
Algorithm 3.
The next example demonstrates the application of the verification method.
Example 6 (running example
2): Given the semialgebraic
system M3 by [ẋ, ẏ] = y 2 , xy and the initial set be
X0 = {(x, y) ∈ R2 | I(x, y) = (x+15)2 +(y −17)2 −1 ≤ 0},
verify that if the unsafe set XU = {(x, y) ∈ R2 | U (x, y) =
(x−11)2 +(y−16.5)2 −1 ≤ 0} can be reached. The parameter
space of C ∗ = {g(~u, ~x) = u1 − u3 (x2 − y 2 ) = 0 | (u1 , u3 ) ∈
R2 \ {~0}} has dimension 1 and hence can provide an invariant
class for every state in X0 and XU . The normal vector of the
hyperplane g(~u, ~x) = 0 is (1, ψ1 (x, y)) = (1, y 2 − x2 ). Let
Algorithm 3: Safety verification
~ the K-dimensional normal vector of an
input : ψ:
invariant cluster;
I(~x1 ): the initial set;
U(~x2 ): the unsafe set;
N : the maximum degree of programming
polynomials α
~ , β, θ
output: IsSafe: whether the system is safe
1
2
3
4
5
6
7
8
9
10
11
12
IsSafe ← False;
for i ← 1 to N do
α
~ ← generate a vector of polynomials of degree i for
ψ;
β ← generate a polynomial of degree i for I(~x1 );
θ ← generate a polynomial of degree i for U(~x2 );
PK
~j (~x1 ) − ψ
~j (~x2 )) + βI+θU −1;
P ← j=1 αj (ψ
Solution ← perform SOS programming on P ;
if Solution is found then
IsSafe ← True;
break;
end
end
ϕ(x1 , y1 , x2 , y2 ) = ψ1 (x1 , y1 ) − ψ1 (x2 , y2 ). By Theorem 4,
to verify the safety property, we only need to verify that the
following system of equations has no real solution.
I(x1 , y1 ) = (x1 + 15)2 + (y1 − 17)2 − 1 = 0
U (x2 , y2 ) = (x2 − 11)2 + (y2 − 16.5)2 − 1 = 0
ϕ(x1 , y1 , x2 , y2 ) = y12 − x21 − (y22 − x22 ) = 0
Note that we substitute (x1 , y1 ), (x2 , y2 ) for (x, y) in I(x, y)
and U (x, y), respectively, to denote the different points in X0
and XU . To prove that the system is safe, we need to find
αi ∈ R[x1 , y1 , x2 , y2 ], i = 1, 2, 3 such that Prog = α1 I +
α2 U + α3 ϕ − 1 is a sum-of-squares. Finally, we found three
polynomials of degree 2 for αi , respectively (see Appendix K
for the expressions of αi and Prog), hence the system is safe.
As shown in Figure 2, although the relative position of XU
to the reachable set of X0 is very close, we can still verify
the safety property using an invariant cluster. However, we
failed to find a barrier certificate for this system by using the
methods in [14], [15].
In Theorem 4, we present a sufficient condition for deciding
if a semialgebraic system is safe. The theory originates from
the fact that the system is safe if there is no invariant class
intersecting both the initial and the unsafe set, which is equivalent to that the formula (7) holds. To verify the latter, we need
to find a set of witness polynomials by SOS programming.
However, as the dimension of the system increases, the number
of parametric polynomials involved increases correspondingly,
which also leads to an increase in computational complexity. In
what follows, we present a new method for safety verification,
which avoids the aforementioned problem. The new method
is based on Proposition 1, that is, for any polynomial g(~x)
satisfying Lf~g ∈ hgi, g(x) ∼ 0 is an invariant for any
∼ ∈ {<, ≤, =, ≥, >}.
Proposition 3: Given a semialgebraic system M =
hX, f~, X0 i and an invariant cluster C = {g(~u, ~x) = 0 | ~u ∈
RK \ {~0}} of M , let X0 and XU be the initial set and the
unsafe set, respectively. Then, the system is safe if there exists
a ~u∗ ∈ RK \ {~0} such that
∀~x ∈ X0 : g(~u∗ , ~x) ≥ 0
∗
∀~x ∈ XU : g(~u , ~x) < 0
(9)
(10)
Proof: See Appendix F.
According to Proposition 3, to verify the safety property,
it suffices to find a ~u∗ ∈ RK \ {~0} which satisfies the
constraints (9) and (10). There are some constraint solving
methods available, e.g, SMT solvers. However, the high complexity of SMT theory limits the applicability of the method.
In the following, we transform the above constraint-solving
problem into an SOS programming problem, which can be
solved efficiently. We write P~ (~x) ~0 to denote pi (~x) ≥ 0, i =
1, . . . , m for a polynomial vector P~ (x) = (p1 (~x), . . . , pm (~x)).
Proposition 4: Given a semialgebraic system M =
hX, f~, X0 i and an invariant cluster C = {g(~u, ~x) = 0 |,
~u ∈ RK \ {~0}} of M and a constant ǫ ∈ R>0 , let
X0 = {~x ∈ Rn | I~ ~0, I~ ∈ R[~x]m1 } and XU = {~x ∈ Rn |
~ ~0, U
~ ∈ R[~x]m2 }. Then, the system is safe if there exist a
U
∗
K
~u ∈ R \ {~0} and two SOS polynomial vectors ~µ1 ∈ R[~x]m1
and ~µ2 ∈ R[~x]m2 such that the following polynomials are SOS
polynomials.
g(~u∗ , ~x) − µ
~ 1 · I~
~ −ǫ
− g(~u∗ , ~x) − ~µ2 · U
(11)
(12)
Proof: See Appendix G.
Similar to Theorem 4, Proposition 4 also reduces to an
SOS programming problem. However, the ideas behind these
two theories are different in that by Theorem 4 we attempt
to prove no invariant class which overapproximates trajectory
can intersect both X0 and XU , while by Proposition 4 we
mean to find a hypersurface which is able to separate the
reachable set of X0 from the unsafe set XU . Apparently,
there must exist no invariant class intersecting both X0 and
XU if there exists such a hypersurface, but not vice versa.
Hence the latter is more conservative than the former, but it
is also more efficient in theory because it usually involves
less unknown polynomials. For example, for an n-dimensional
system with X0 and XU defined by a single polynomial
inequality, respectively, we usually need n + 1 unknown
polynomials for the former method, however, we need only
2 for the latter. See Algorithm 4 for the pseudocode of the
method based on Proposition 4.
Let us use the following example to demonstrate the application of Algorithm 4.
Example 7: Given the semialgebraic system S4 by
2
ẋ
y − 2y
= 2
,
ẏ
x + 2x
Algorithm 4: Algorithm for safety verification based on
SOS programming
input : g(~u, ~x): the invariant polynomial defining an
invariant cluster;
~ the polynomial vector describing initial set;
I:
~ : the polynomial vector describing unsafe set;
U
N : the maximum degree of the polynomial
vectors ~µ1 , ~µ2 ;
ǫ: a positive real number
output: IsSafe: whether the system is safe;
g(~u∗ , ~x), µ
~ 1, µ
~ 2 : the feasible solution if IsSafe;
1
2
3
4
5
6
7
8
9
10
11
12
13
IsSafe ← False;
for i ← 0 to N do
~µ1 ← generate a parametric polynomial vector of
~
degree 2i for I;
~µ2 ← generate a parametric polynomial vector of
~;
degree 2i for U
~
p1 ← g(~u, ~x) − ~
µ1 · I;
~ − ǫ;
p2 ← −g(~u, ~x) − ~
µ2 · U
Solution ← perform SOS programming on
{~µ1 , ~µ2 , p1 , p2 };
if Solution is found then
IsSafe ← True;
[g(~u∗ , ~x), ~
µ1 , ~
µ2 ] ← get the feasible solution from
Solution;
break;
end
end
Fig. 4: Example 7. Invariant g(~u∗ , ~x) = 0 (blue curve)
separating reachable set of X0 (green patch) from XU (red
patch).
all locations. The idea is to pick out a polynomial gl (~u∗l , ~x)
from the respective invariant cluster Cl for each location l such
that gl (~u∗l , ~x) ≥ 0 is an invariant for the location l and all the
invariants coupled together by the constraints at the discrete
transitions form a hybrid invariant for the hybrid system. The
aforementioned idea is formalized in Proposition 5.
Proposition 5: Given an n-dimensional hybrid system H =
hL, X, E, R, G, I, F i and a set of invariant clusters {Cl , l =
1, . . . , n}, where Cl = {gl (~ul , ~x) = 0 | ~ul ∈ RKl \ {~0}} with
Kl > 1 is an invariant cluster for location l, the system H is
safe if there exists a set Su~ = {~u∗l ∈ RKl \ {~0}, l = 1, . . . , n}
such that, for all l ∈ L and (l, l′ ) ∈ E, the following formulae
hold:
∀~x ∈ Init(l) : gl (~u∗l , x) ≥ 0
∀~x ∈ G(l, l′ ), ∀~x′ ∈ R((l, l′ ), ~x) : gl (~u∗l , ~x) ≥ 0
2
let the initial set be X0 = {(x, y) ∈ R | I(x, y) = 1 − (x +
6.0)2 − (y + 6.0)2 ≥ 0}, decide whether the unsafe set XU =
{(x, y) ∈ R2 | U (x, y) = 1 − (x − 8.2)2 − (y − 4.0)2 ≥ 0}
can be reached.
By Algorithm 1, we first get an invariant cluster C4 =
{g(~u, ~x) = 13 u3 x3 − 31 u3 y 3 + u3 x2 + u3 y 2 + u1 = 0 | ~u ∈ R2 }
for the system S4 , then by Algorithm 4, we find an invariant
g(~u∗ , ~x) = 0 with ~u∗ = (−3081.9, 7.1798) from the invariant
cluster C4 : g(~u∗ , ~x) = 7.1798x3 − 7.1798y 3 + 21.539x2 +
21.539y 2 − 3081.9 and two polynomials µ1 (~x), µ2 (~x) of
degree 2. The stream plot of S4 and the plot of g(~u∗ , ~x) = 0
are shown in Figure 4. Note that we failed to find a barrier
certificate by using the method in [15] and [14] for this system.
B. Safety Verification of Hybrid Systems
In this section, we extend the safety verification method for
continuous systems based on invariant clusters to semialgebraic hybrid systems.
A hybrid system consists of a set of locations and a set
of discrete transitions between locations. In general, different
locations have different continuous dynamics and hence correspond to different invariant clusters. An invariant for the hybrid
system can be synthesized from the set of invariant clusters of
=⇒ gl′ (~u∗l′ , ~x′ ) ≥ 0
∀~x ∈ I(l) ∩ Uns(l) : gl (~u∗l , ~x) < 0
(13)
(14)
(15)
where Init(l) and Uns(l) denote respectively the initial set
and the unsafe set at location l.
Proof: See Appendix H.
Similar to Proposition 3, we further transform the problem
into an SOS programming problem. Consider a semialgebraic
hybrid system H = hL, X, E, R, G, I, F i, where the mappings
R, G, and I are defined in terms of polynomial inequalities as
follows:
′
~ ll′ 0, G
~ ll′ ∈ R[~x]mll′ }
• G : (l, l ) 7→ {~
x ∈ Rn | G
′
~ ll′ ~x 0, R
~ ll′ ~x ∈ R[~x]nll′ }
• R : (l, l , ~
x) 7→ {~x ∈ Rn | R
• I : l 7→ {~
x ∈ Rn | I~l 0, I~l ∈ R[~x]pl }
and the mappings of the initial and the unsafe set are defined
as follows:
~ l 0, Init
~ l ∈ R[~x]rl }
• Init : l 7→ {~
x ∈ Rn | Init
n
~
~ l ∈ R[~x]sl }
• Uns : l 7→ {~
x ∈ R | Unsl 0, Uns
where mll′ , nll′ , rl , pl and sl are the dimensions of the polynomial vector spaces. Then we have the following proposition
for safety verification of the semialgebraic hybrid system H.
Proposition 6: Let the hybrid system H, the initial set
mapping Init, and the unsafe set mapping Uns be defined
as above. Given a set of invariant clusters {Cl , l = 1, . . . , n}
of H, where Cl = {gl (~ul , ~x) = 0 | ~ul ∈ RKl \ {~0}} with
Kl > 1 is an invariant cluster for location l, a set Sγ =
{γll′ ∈ R≥0 , (l, l′ ) ∈ E} of constants, and a constant vector
~ǫ ∈ Rn>0 , the system is safe if there exists a set Su = {~u∗l ∈
RKl \ {~0}, l = 1, . . . , n} and five sets of SOS polynomial
vectors {~θl ∈ R[~x]sl , l ∈ L}, {~κll′ ∈ R[~x]pll′ , (l, l′ ) ∈ E},
ηl ∈ R[~x]tl , l ∈ L}, and
{~σll′ ∈ R[~x]qll′ , (l, l′ ) ∈ E}, {~
wl
{~νl ∈ R[~x] , l ∈ L} such that the following polynomials
are SOS for all l ∈ L and (l, l′ ) ∈ E:
~ l
gl (~u∗l , ~x) − θ~l · Init
gl′ (~u∗l′ , ~x′ )
(16)
γll′ gl (~u∗l , ~x)
− ~κll′ · Gll′ − ~σll′ · Rll′ ~x
~
− ~νl · Il − ~ηl · Unsl − gl (~u∗l , ~x) − ǫl
−
(17)
(18)
Proof: Similar to the proof of Proposition 4, we can
easily derive the formulae (13)–(15) from the SOS’s (16)–(18),
respectively. Then, by Proposition 5, the system is safe.
The algorithm for computing invariants for semialgebraic
hybrid systems based on Proposition 6 is very similar to
Algorithm 4 for semialgebraic continuous systems except that
it involves more SOS constraints on continuous and discrete
transitions.
V. I MPLEMENTATION AND E XPERIMENTS
Based on the approach presented in this paper, we implemented a prototype tool in Maple and Matlab, respectively.
In Maple, we implemented the tool for computing invariant
clusters and identifying invariant classes based on the remainder computation of the Lie derivative of a polynomial w.r.t. its
Gröbner basis and solving the system of polynomial equations
obtained from the coefficients of the remainder. In Matlab,
we implemented the tool for safety verification based on the
SOS programming tool package SOSTOOLS. Currently, we
manually transfer the invariant clusters computed in Maple to
Matlab for safety verification. In the future, we will integrate
the two packages into a single tool.
Now, we present the experimental results on nonlinear
benchmark systems, run on a laptop with an 3.1GHz Intel
Core i7 CPU and 8GB memory.
A. Longitudinal Motion of an Airplane
In this experiment, we study the 6th order longitudinal equations of motion that captures the vertical motion (climbing,
descending) of an airplane ([16], Chapter 5). Let g denote the
gravity acceleration, m the total mass of an airplane, M the
aerodynamic and thrust moment w.r.t. the y axis, (X, Z) the
aerodynamics and thrust forces w.r.t. axis x and z, and Iyy the
second diagonal element of its inertia matrix. Then the motion
of the airplane is described as follows.
X
− g sin(θ) − qw,
m
ẋ = w sin(θ) + v cos(θ),
v̇ =
θ̇ = q,
Z
+ g cos(θ) + qv,
m
ż = −v sin(θ) + w cos(θ),
M
,
q̇ =
Iyy
ẇ =
where the meanings of the variables are as follows: v: axial
velocity, w: vertical velocity, x: range, z: altitude, q: pitch rate,
θ: pitch angle.
To transform the above system into a semialgebraic system,
we first introduce two additional variables d1 , d2 such that
d1 = sin(θ), d2 = cos(θ) and then substitute d1 and d2
respectively for sin(θ) and cos(θ) in the model. In addition, we
get two more constraints d˙1 = qd2 and d˙2 = −qd1 . As a result,
the dimension of the system rises to 8. For this system, using
the method in [17], Ghorbal et al. spent 1 hour finding three
invariant polynomials of degree 3 on a laptop with a 1.7GHz
Intel Core i5 CPU and 4GB memory. Using our method,
we spent only 0.406 seconds obtaining an invariant cluster
g9 (~u, ~x) = 0 of degree 3, which is presented in Appendix L.
By the constraint d21 + d22 = 1, we reduce the normal vector
of the hyperplane g9 (~u, ~x) = 0 in ~u to (1, ψ1 , ψ2 , ψ3 ), where
ψ1 , ψ2 , ψ3 are defined as follows.
M mz gmθ mqv
ψ1 =
+
+
+ 1 sin (θ)
Iyy Z
Z
Z
X
mqw
+
−
cos (θ)
Z
Z
Xqv qw
gIyy Xθ
Xz
sin(θ)
+x−
− Iyy
+
ψ2 = −
Z
ZM
ZM
M
qv
X2 + Z2
Xqw
cos(θ)
−
−
+ Iyy
ZM
M
ZM m
Mθ
ψ3 = q 2 − 2
Iyy
Given a symbolic initial point ~x0 = (v0 , w0 , x0 , z0 , θ0 , q0 ,
d01 , d02 ), we have verified that our invariant cluster defines the
same algebraic variety as defined by the invariants in [17]
by comparing their Gröbner bases. However, our method is
much more efficient. Moreover, we also obtained the invariant
clusters of higher degrees (4 − 6) quickly. The experimental
result is shown in Table I. The first column is the degree of the
invariants, the second column is the variables to be decided,
the third column is the computing time in seconds, and the
last column is the number of invariant clusters generated. As
can be seen, in the most complicated case, where the number
of the indeterminates reaches up to 3003, we spent only 200.9
seconds to discover an invariant cluster of degree 6. However,
we found that these higher order invariant clusters have the
same expressive power as the invariant cluster of degree 3 in
terms of algebraic variety.
B. Looping particle
Consider a heavy particle on a circular path of radius r. The
motion of the particle is described by the following differential
equation.
˙
rcos(θ)
−r sin(θ)θ̇
ẋ
−yω
˙
ẏ =
rsin(θ)
= r cos(θ)θ̇ = xω
g cos(θ)
ω̇
− gx
− g(r cos(θ))
−
2
r2
r
r
Note that the above is a parameterized system with gravity
acceleration g and radius r as parameters. Our tool finds the
TABLE I: Benchmark results for the Longitudinal Motion of an Airplane (B1), the Looping Particle system (B2), the Coupled
Spring-Mass system (B3), and the 3D-Lotka-Volterra System (B4).
Degree of
invariants
1
2
3
4
5
6
B1
9
45
165
495
1287
3003
No. of variables
B2
B3
4
6
10
21
20
56
35
126
56
252
84
462
B4
4
10
20
35
56
84
B1
0.016
0.031
0.484
3.844
25.172
200.903
Running time (sec)
B2
B3
0.015
0.047
0.047
0.078
0.049
0.250
0.156
1.109
0.703
6.641
3.000
32.109
following invariant cluster consisting of a parametric polynomial of degree 2: {g(~u, ~x) = 0 | g(~u, ~x) = u5 x2 + u5 y 2 +
u2 ω 2 + 2ur22 g y + u0 , ~u ∈ R3 \ {~0}}. Given an arbitrary point
(x0 , y0 , ω0 ) = (2, 0, ω0 ), we get the invariant class {g(~u, ~x) =
0 | (x20 + y02 )u5 + (ω02 + 2g
u ∈ R3 \ {~0}}.
r 2 y0 )u2 + u0 = 0, ~
According to Algorithm 2, the algebraic variety representing
the trajectory originating from (x0 , y0 , ω0 ) is {(x, y, ω) ∈ R3 |
2g
2
x2 + y 2 − x20 − y02 = 0, ω 2 + 2g
r 2 y − ω0 − r 2 y0 = 0}. The results
in [19] and [18] are special cases of our result when setting
(r, g) = (2, 10) and (r, g, x0 , y0 ) = (2, 10, 2, 0), respectively.
Therefore, our method is more powerful in finding parameterized invariants for parameterized systems. See Table I for
detailed experimental result.
C. Coupled spring-mass system
Consider the system
v1
x˙1
k
k
v˙1 − 1 x1 − 2 (x1 − x2 )
m1
= m1
x˙2
v2
k
2
v˙2
− m2 (x2 − x1 )
The model consists of two springs and two weights w1 , w2 .
One spring, having spring constant k1 , is attached to the ceiling
and the weight w1 of mass m1 is attached to the lower end
of this spring. To the weight w1 , a second spring is attached
having spring constant k2 . To the bottom of this second spring,
a weight w2 of mass m2 is attached. x1 and x2 denote the
displacements of the center of masses of the weights w1 and
w2 from equilibrium, respectively.
In this benchmark experiment, we first tried an instantiated
version of the system by using the same parameters as in [20]:
k2
k1
m1 = m2 = k and m1 = 5m2 . The experimental result
is presented in Table I. We found that the expressive power
of the invariant clusters does not increases any more as the
degree is greater than 3 and it took only 0.250 seconds to
compute the invariant cluster of degree 3. Finally, we perform
the computation directly on the fully parameterized system
and we get the following parameterized invariant cluster.
k2 x1 x2 (m1 u8 − 2m2 u10 )
+ u10 v12 + u1
g(~u, ~x) = u8 v1 v2 +
m1 m2
1 v22 (k1 m2 u8 − k2 m1 u8 + k2 m2 u8 + 2k2 m2 u10 )
+
2
k2 m1
1 (2k1 m2 u10 − k2 m1 u8 + 2k2 m2 u10 )x21
+
2
m1 m2
1 (k1 m2 u8 − k2 m1 u8 + 2k2 m2 u10 )x22
+
2
m1 m2
B4
<0.001
0.031
0.109
0.312
0.750
1.641
B1
0
1
1
1
1
1
No. of invariant clusters
B2
B3
0
0
1
0
0
1
1
1
0
1
1
1
B4
2
3
7
6
6
16
This invariant cluster enables us to analyze the system
properties under different parameter settings.
D. 3D-Lotka-Volterra system
Consider the system [ẋ, ẏ, ż] = [xy − xz, yz − yx, xz − yz].
The experimental result is presented in Table I. Here we
present only two invariant clusters to show their expressive
power: one of degree 1 and one of degree 3. C1 = {g1 (~u, ~x) =
u3 x + u3 y + u3 z + u4 = 0 | ~u ∈ R2 \ {~0}}, C3 =
{g3 (~u, ~x) = 0 | ~u ∈ R4 \ {~0}}, where g3 (~u, ~x) = u5 xyz +
u10 x3 + 3u10 x2 y + 3u10 x2 z + 3xy 2 u10 + 3xz 2 u10 + u10 y 3 +
3u10 y 2 z + 3u10 yz 2 + u10 z 3 + u16 x2 + 2u16 xy + 2u16 xz +
u16 y 2 + 2u16 yz + u16 z 2 + u19 x + u19 y + u19z + u20 . Invariant
cluster C1 shows that every trajectory of the system must
lie in some plane. C3 can overapproximate the trajectories
with a much higher precision because it can provide algebraic
varieties of dimension 1. For example, in [20], the initial
set is given as X0 = {(x, y, z) ∈ R3 | x2 − 1 = 0,
y 2 − 1 = 0, z 2 − 1 = 0} and they discovered four invariants.
In our experiment, we spend 0.109 seconds discovering an
invariant cluster. This invariant cluster can overapproximate
all the trajectories precisely, say, for x0 = (1, −1, 1) ∈ X0 ,
we get the invariant class Class(C3 , x0 ) = {g3 (~u, ~x) = 0 |
−u5 +7u10 +u16 +u19 +u20 = 0, ~u ∈ R5 \{~0}}. By taking the
intersection of its basis xyz + 1 = 0 and x + y + z − 1 = 0, we
immediately obtain a one-dimensional algebraic variety which
overapproximates the trajectory precisely.
E. Hybrid controller
Consider a hybrid controller consisting of two control
modes. The discrete transition diagram of the system is shown
in Fig. 5a and the vector fields describing the continuous
behaviors are as follows:
y 2 + 10y + 25
f1 (~x) =
,
2xy + 10x − 40y − 200
−y 2 − 10y − 25
f2 (~x) =
8xy + 40x − 160y − 800
The system starts from some point in X0 = {(x, y) ∈ R2 |
(x−9)2 +(y−20)2 ≤ 4} and then evolves following the vector
field f1 (~x) at location l1 (Switch-On). The value of x keeps
increasing until it reaches 35, then the system switches immediately to location l2 (Switch-Off) without performing any
reset operation. At location l2 , the system operates following
the vector field f2 (~x) and the value of x keeps decreasing. As
the value of x drops to 5, the system switches immediately
x
d x d
x f x
d x d
x f x
l
l
x
(a)
(b)
Fig. 5: (a) Hybrid automaton from Subsection V-E. x = 5
and x = 35 are guards for discrete transitions and no reset
operation is performed. (b) Hybrid invariant for the system in
Subsection V-E. Solid patch in green: initial set, curve in blue:
invariant for l1 , curve in purple: invariant for l2 , red shadow
region on the top: unsafe region.
back to location l1 again. Our objective is to verify that the
value of y will never exceed 48 in both locations.
For the convenience of SOS programming, we define the
unsafe set as Uns(l1 ) = Uns(l2 ) = {(x, y) ∈ R2 | 48 < y <
60}, which is sufficient to prove y ≤ 48 in locations l1 and
l2 . According to the theory proposed in Section IV-B, we first
find an invariant cluster for each location, which is composed
of a parameterized polynomial, respectively: g1 (~u1 , ~x) =
1
u12 y 2 + 8u12 x + u12 y + u11 and g2 (~u2 , ~x) =
− 51 u12 x2 + 10
1
4
2
2
5 u22 x + 10 u22 y −32u22 x+u22 y +u21 . In the second phase,
we make use of the constraint condition in Proposition 6 to
compute a pair of vectors ~u∗1 and ~u∗2 . By setting γ12 = γ21 = 1,
our tool found a pair of ~u∗1 = (u11 , u12 ) = (2.9747, 382.14)
and ~u∗2 = (u21 , u22 ) = (2.9747, 138.44). As shown in Fig. 5b,
the curves of g1 (~u∗1 , ~x) = 0 and g2 (~u∗2 , ~x) = 0 form an upper
bound for the reachable set in location l1 and l2 , respectively,
which lie below the unsafe region y ≥ 48. Therefore, the
system is safe.
VI. R ELATED W ORK
In recent years, many efforts have been made toward generating invariants for hybrid systems. Matringe et al. reduce
the invariant generation problem to the computation of the
associated eigenspaces by encoding the invariant constraints
as symbolic matrices [21]. Ghorbal et al. use the invariant
algebraic set formed by a polynomial and a finite set of its successive Lie derivatives to overapproximate vector flows [17].
Both of the aforementioned methods involve minimizing the
rank of a symbolic matrix, which is inefficient in dealing
with parametric systems. In addition, none of these methods
involve how to verify safety properties based on the invariants.
Sankaranarayanan discovers invariants based on invariant ideal
and pseudo ideal iteration [20], but this method is limited to
algebraic systems. Tiwari et al. compute invariants for special
types of linear and nonlinear systems based on Syzygy computation and Gröbner basis theory as well as linear constraint
solving [22]. Platzer et al. use quantifier elimination to find
differential invariants [23]. The methods based on Gröbner
basis and first-order quantifier elimination suffer from the high
complexity significantly. Another approach considers barrier
certificates based on different inductive conditions [15], [14]
which can be solved by SOS programming efficiently but is
limited by the conservative inductive condition. Carbonell et
al. generate invariants for linear systems [24]. Some other
approaches focusing on different features of systems have also
been proposed for constructing inductive invariants [25], [26],
[18], [27], [28].
VII. C ONCLUSION
In this paper, we proposed an approach to automatically
generate invariant clusters for semialgebraic hybrid systems.
The benefit of invariant clusters is that they can overapproximate trajectories of the system precisely. The invariant clusters
can be obtained efficiently by computing the remainder of
the Lie derivative of a template polynomial w.r.t. its Gröbner
basis and then solving a system of homogeneous polynomial
equations obtained from the remainder. Moreover, based on
invariant clusters and SOS programming, we propose a new
method for safety verification of hybrid systems. Experiments
show that our approach is effective and efficient for a large
class of biological and control systems.
ACKNOWLEDGEMENT
This research was supported in part by the European Research Council (ERC) under grant 267989 (QUAREM) and by
the Austrian Science Fund (FWF) under grants S11402-N23
(RiSE) and Z211-N23 (Wittgenstein Award).
R EFERENCES
[1] T. Henzinger, “The theory of hybrid automata,” in Proc. IEEE Symp.
Logic in Computer Science, 1996, pp. 278–292.
[2] E. Asarin, T. Dang, and A. Girard, “Reachability analysis of nonlinear
systems using conservative approximation,” in Hybrid Systems: Computation and Control. Springer, 2003, pp. 20–35.
[3] S. Bogomolov, A. Donzé, G. Frehse, R. Grosu, T. T. Johnson, H. Ladan,
A. Podelski, and M. Wehrle, “Guided search for hybrid systems based
on coarse-grained space abstractions,” International Journal on Software
Tools for Technology Transfer, pp. 1–19, 2015.
[4] S. Bogomolov, G. Frehse, R. Grosu, H. Ladan, A. Podelski, and
M. Wehrle, “A box-based distance between regions for guiding the
reachability analysis of SpaceEx,” in Computer Aided Verification (CAV
2012), ser. LNCS, vol. 7358. Springer, 2012, pp. 479–494.
[5] A. Tiwari, “Abstractions for hybrid systems,” Formal Methods in System
Design, vol. 32, no. 1, pp. 57–83, 2008.
[6] R. Alur, T. Dang, and F. Ivančić, “Progress on reachability analysis
of hybrid systems using predicate abstraction,” in Hybrid Systems:
Computation and Control. Springer, 2003, pp. 4–19.
[7] S. Bogomolov, G. Frehse, M. Greitschus, R. Grosu, C. S. Pasareanu,
A. Podelski, and T. Strump, “Assume-guarantee abstraction refinement
meets hybrid systems,” in HVC 2014. Springer, pp. 116–131.
[8] S. Bogomolov, C. Schilling, E. Bartocci, G. Batt, H. Kong, and R. Grosu,
“Abstraction-based parameter synthesis for multiaffine systems,” in 11th
International Haifa Verification Conference (HVC 2015), ser. LNCS, vol.
9434. Springer, 2015, pp. 19–35.
[9] S. Bogomolov, C. Herrera, M. Muñiz, B. Westphal, and A. Podelski,
“Quasi-dependent variables in hybrid automata,” in Hybrid Systems:
Computation and Control. ACM, 2014, pp. 93–102.
[10] D. A. Cox, J. Little, and D. O’Shea, Ideals, Varieties, and Algorithms:
An Introduction to Computational Algebraic Geometry and Commutative
Algebra. Springer, 2007.
[11] G. Stengle, “A Nullstellensatz and a Positivstellensatz in semialgebraic
geometry,” Mathematische Annalen, vol. 207, no. 2, pp. 87–97, 1974.
[12] S. Prajna, A. Papachristodoulou, P. Seiler, and P. Parrilo, “SOSTOOLS
and its control applications,” Positive polynomials in control, pp. 580–
580, 2005.
[13] A. Ayad, “A survey on the complexity of solving algebraic systems,” in
International Mathematical Forum, vol. 5, no. 7, 2010, pp. 333–353.
[14] H. Kong, F. He, X. Song, W. N. Hung, and M. Gu, “Exponentialcondition-based barrier certificate generation for safety verification of
hybrid systems,” in CAV. Springer, 2013, pp. 242–257.
[15] S. Prajna and A. Jadbabaie, “Safety verification of hybrid systems using
barrier certificates,” HSCC, pp. 271–274, 2004.
[16] R. F. Stengel, “Flight dynamics,” Fluid Dynamics, vol. 1, 2004.
[17] K. Ghorbal and A. Platzer, “Characterizing algebraic invariants by differential radical invariants,” in Tools and Algorithms for the Construction
and Analysis of Systems. Springer, 2014, pp. 279–294.
[18] S. Sankaranarayanan, H. Sipma, and Z. Manna, “Constructing invariants
for hybrid systems,” HSCC, pp. 69–77, 2004.
[19] R. Rebiha, A. V. Moura, and N. Matringe, “Generating invariants for
non-linear hybrid systems,” TCS, vol. 594, pp. 180–200, 2015.
[20] S. Sankaranarayanan, “Automatic invariant generation for hybrid systems using ideal fixed points,” in International Conference on Hybrid
Systems: Computation and Control. ACM, 2010, pp. 221–230.
[21] N. Matringe, A. V. Moura, and R. Rebiha, “Generating invariants
for non-linear hybrid systems by linear algebraic methods,” in Static
Analysis. Springer, 2010, pp. 373–389.
[22] A. Tiwari and G. Khanna, “Nonlinear systems: Approximating reach
sets,” Hybrid Systems: Computation and Control, pp. 171–174, 2004.
[23] A. Platzer and E. Clarke, “Computing differential invariants of hybrid
systems as fixedpoints,” in CAV. Springer, 2008, pp. 176–189.
[24] E. Rodríguez-Carbonell and A. Tiwari, “Generating polynomial invariants for hybrid systems,” HSCC, pp. 590–605, 2005.
[25] T. T. Johnson and S. Mitra, “Invariant synthesis for verification of
parameterized cyber-physical systems with applications to aerospace
systems,” in AIAA Infotech at Aerospace Conference, 2013.
[26] S. Gulwani and A. Tiwari, “Constraint-based approach for analysis of
hybrid systems,” in CAV. Springer, 2008, pp. 190–203.
[27] B. Sassi, M. Amin, A. Girard, and S. Sankaranarayanan, “Iterative
computation of polyhedral invariants sets for polynomial dynamical
systems,” in CDC. IEEE, 2014, pp. 6348–6353.
[28] J. Liu, N. Zhan, and H. Zhao, “Computing semi-algebraic invariants
for polynomial dynamical systems,” in International Conference on
Embedded Software. ACM, 2011, pp. 97–106.
[29] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan, Linear matrix
inequalities in system and control theory.
Society for Industrial
Mathematics, 1994, vol. 15.
[30] P. Parrilo, “Semidefinite programming relaxations for semialgebraic
problems,” Mathematical Programming, vol. 96, no. 2, pp. 293–320,
2003.
A PPENDIX
A. Proof of Proposition 1
Proof:
Let g(x) ∈ R[~x] be a polynomial satisfying Lf g ∈ hgi, we
first prove that
∀k ≥ 1 : Lkf g ∈ hgi
(19)
The proof is by induction. Since g(x) satisfies Lf g ∈ hgi, we
must have Lf g = p1 g for some p1 ∈ R[x]. Assume ∀k ≤ m :
Lkf g ∈ hgi holds, we only need to prove Lm+1
g ∈ hgi holds.
f
By assumption, we have Lm
g
∈
hgi,
that
is,
Lm
f
f g = pm g
m+1
m
for some pm ∈ R[x]. Then, Lf g = Lf Lf g = Lf pm g =
gLf pm + pm Lf g = gLf pm + pm p1 g = g(Lf pm + pm p1 ).
Therefore, the formula (19) holds.
Second, we prove that g(x) = 0 is an invariant, i.e. for any
trajectory x(t),
g(x(0)) = 0 =⇒ ∀t ≥ 0 : g(x(t)) = 0
(20)
m
g(x(t))
By the formula (19), we have d dt
|t=τ = Lm
m
f g|t=τ =
pm
g(x(τ
))
for
all
m
≥
1,
then
g(x(0))
=
0
=⇒
dm g(x(t))
|
=
0
for
all
m
≥
1.
Consider
the
Taylor
t=0
dtm
expansion of g(x(t)) at t = τ ,
g(x(t)) = g(x(τ )) +
∞
X
1 dm g(x(t))
|t=τ (t − τ )m (21)
m
m!
dt
m=1
Hence, the formula (20) holds.
Next, we prove that for any trajectory x(t),
g(x(0)) 6= 0 =⇒ ∀t ≥ 0 : g(x(t)) 6= 0
(22)
then, according to the continuity of g(x(t)), we can assert that
g(x) ∼ 0 is an invariant for each ∼∈ {<, ≤, =, ≥, >}.
The proof is by contradiction. Given an arbitrary trajectory
x(t), assume g(x(τ )) = 0 for some τ > 0 and
g(x(t)) > 0
m
g(x(t))
(or g(x(t)) < 0) for all t ∈ [0, τ ). Since d dt
|t=τ =
m
Lm
g|
=
p
g(x(τ
))
=
0
for
all
m
≥
1,
then
according
to
t=τ
m
f
the Taylor expansion (21), there must exist a δ ∈ R>0 such
that g(x(t)) = 0 for all t ∈ [τ −δ, τ +δ], which contradicts the
assumption that g(x(t)) > 0 (or g(x(t)) < 0) for all t ∈ [0, τ ).
Thus, the proposition holds.
B. Proof of Theorem 2
Proof: 1. By Definition 10, for all g(~u, ~x) ∈ Dg ,
g(~u, ~x) = 0 is an invariant for the trajectory originating from
~x0 , which means π~x0 ⊆ {~x ∈ Rn | g(~u, ~x) = 0}. Hence,
π~x0 ⊆ V(Dg ) holds.
2. Since the dimension of the hyperplane g(~u, ~x0 ) = 0 is
m, there must exist m vectors {~u1 , . . . , ~um } such that for
all ~u satisfying g(~u, ~x0 ) = 0, P
there exist scalars c1 , . . . , cm ,
m
ui . Therefore,
for evnot all zero, such that ~u =
i=1 ci ~
Pm
ery
g(~
u
,
~
x
)
∈
D
,
we
have
g(~
u
,
~
x
)
=
g(
c
~
u
x) =
g
i=1 i i , ~
Pm
c
g(~
u
,
~
x
),
hence,
g(~
u
,
~
x
)
∈
hg(~
u
,
~
x
),
.
.
.
,
g(~
u
x)i.
i
1
m, ~
i=1 i
For B = {g(~u1 , ~x), . . . , g(~um , ~x)}, we have hDg i ⊆ hBi.
The converse is trivial. Therefore, hDg i = hBi.
C. Principle of SOS programming
The principle of SOS programming is based on the fact that
a polynomial of P
degree 2k can be written as a sum-of-squares
(SOS) P (~x) =
qi (~x)2 for some polynomials qi (~x) of degree k if and only if P (~x) has a positive semidefinite quadratic
form, i.e. P (~x) = ~v (~x)M~v (~x)T , where ~v (~x) is a vector of
monomials with respect to x of degree k or less and M is a real
symmetric positive semidefinite matrix with the coefficients of
P (~x) as its entries. Therefore, the problem of finding a SOS
polynomial P (~x) can be converted to the problem of solving
a linear matrix inequality (LMI) M 0 [29], which can
be solved by semidefinite programming [30]. Currently, there
exists an efficient implementation, named SOSTOOLS [12],
for SOS programming and our implementation is based on
this tool. For details on SOS programming we refer the reader
to the literature.
D. Proof of Theorem 3
Proof: When considering g(~u, ~x) = 0 as a parameterized
hyperplane over ~u, since there is no constant term, the normal
vector of the hyperplane is (ψ1 (~xi ), . . . , ψK (~xi )). Therefore,
if ~x1 and ~x2 are in the same trajectory, they must correspond
to the same hyperplane, namely, the normal vectors of g(~u, ~x1 )
and g(~u, ~x2 ) are either all equal to ~0, which indicates the
formula (5) holds, or are proportional to one another, which
means the formula (4) holds. Moveover, if ψi (~x) ≡ 1, then
formula (6) follows from (4) immediately.
E. Proof of Theorem 4
Proof: By Theorem 1, the identity (7) holds if and only
if there exists no real solution for the following system P of
polynomial equations and inequalities
ψk (~x1 ) − ψk (~x2 ) = 0, k = 1 . . . K
G. Proof of Proposition 4
Proof: Since the polynomial (11) is an SOS, then ∀~x ∈
~
Rn : −g(~u∗ , ~x)−~µ1 · I~ ≥ 0, i.e. ∀~x ∈ Rn : −g(~u∗ , ~x) ≥ ~µ1 · I.
Since ∀~x ∈ X0 : I~ ~0 and µ
~ 1 is an SOS polynomial
vector, which implies ∀~x ∈ Rn : ~µ1 ~0, then ∀~x ∈ X0 :
−g(~u∗ , ~x) ≥ ~µ1 · I~ ≥ 0. Similarly, we can derive from the
SOS polynomial 12 that ∀~x ∈ XU : −g(~u∗ , ~x) < 0. By
Proposition 3, we can conclude that the system is safe.
H. Proof of Proposition 5
Proof: Assume Su is the set that satisfies the formulae (13), (14) and (15), by Proposition 1, every gl (~u∗l , ~x) ≥ 0
is an invariant at location l. To prove this proposition, it
is sufficient to prove that given any trajectory, say π, of
the system H, it cannot reach an unsafe state. Suppose the
infinite time interval R≥0 associated with π is divided into an
infinite
sequence of continuous time subintervals, i.e., R≥0 =
S∞
I
, where In = {t ∈ R≥0 |tn ≤ t ≤ tn+1,tn <tn+1 } is
n
n=0
the time interval that the system spent at location ρ(In ) (where
ρ(In ) returns the location corresponding to In associated to π),
we define the trajectory as π = {(ρ(In ), ~x(t))|t ∈ In , n ∈ N},
where ~x(t0 ) ∈ Init(ρ(I0 )). Then, our objective is to prove the
following assertion:
∀n ∈ N : ∀t ∈ In : gρ(In ) (~u∗ρ(In ) , ~x(t)) ≥ 0.
The basic proof idea is by induction.
Basis: n = 0. Since gρ(I0 ) (~u∗ρ(I0 ) , ~x) ≥ 0 is an invariant, it is
obvious that
∀t ∈ I0 : gρ(I0 ) (~u∗ρ(I0 ) , ~x(t)) ≥ 0.
Induction: n = k. Assume for some k,
pi1 (~x1 ) = 0, pi2 (~x2 ) = 0,
i1 = 1, . . . , l1 , i2 = l1 + 1, . . . l1 + l2
∀j ∈ [0, k] : ∀t ∈ Ij : gρ(Ij ) (~u∗ρ(Ij ) , ~x(t)) ≥ 0.
we mean to prove that
qj1 (~x1 ) ≥ 0, qj2 (~x2 ) ≥ 0,
j1 = 1, . . . , m1 , j2 = m1 + 1 . . . m1 + m2
rk1 (~x1 ) > 0, rk2 (~x2 ) > 0,
k1 = 1, . . . , n1 , k2 = n1 + 1 . . . n1 + n2 .
(24)
∀t ∈ Ik+1 : gρ(Ik+1 ) (~u∗ρ(Ik+1 ) , ~x(t)) ≥ 0.
(23)
That formula (23) has no solution indicates that there exists
no (~x1 , ~x2 ) ∈ X0 × XU such that ψi (~x1 ) = ψi (~x2 ) for all i =
1 . . . K. By the assumption K ≥ 2, we have Class(C, xi ) 6=
∅, i = 1, 2 for any (~x1 , ~x2 ) ∈ X0 × XU . By Theorem 3, the
system is safe.
F. Proof of Proposition 3
Proof: By Proposition 1, g(~u, ~x) ≥ 0 is an invariant of M
for any ~u ∈ RK . Suppose g(~u∗ , ~x) satisfies the constraints (9)
and (10) for some ~u∗ ∈ RK and ~x(t) is an arbitrary trajectory
of M such that ~x(0) ∈ X0 , we must have g(~u∗ , ~x(t)) ≥ 0 for
all t ≥ 0, which means ~x(t) cannot reach XU . Therefore, the
system is safe.
Case 1. (Discrete Transition) By the inductive assumption, we
know that
∀t ∈ Ik : gρ(Ik ) (~u∗ρ(Ik ) , ~x(t)) ≥ 0.
hence
∀t ∈ Ik : ~x(t) ∈ G(ρ(Ik ), ρ(Ik+1 )) =⇒ gρ(Ik ) (~u∗ρ(Ik ) , ~x(t)) ≥ 0
According to condition (14), ∀~x′ (t) ∈ R((l, l′ ), ~x) :
gρ(Ik+1 ) (~u∗ρ(Ik+1 ) , ~x′ (t)) ≥ 0 holds.
Case 2. (Continuous Transition)
Based on Case 1 and the fact that gρ(Ik+1 ) (~u∗ρ(Ik+1 ) , ~x) ≥ 0
is an invariant at ρ(Ik+1 ), we can conclude that ∀t ∈ Ik+1 :
gρ(Ik+1 ) (~u∗ρ(Ik+1 ) , ~x(t)) ≥ 0.
By induction, we know that the assertion (24) holds. Therefore, the system is safe.
I. Invariant clusters of Example 4
C1 = {u2 (x − y) = 0 | u2 ∈ R}, C2 = {u2 (x + y) = 0 | u2 ∈ R},
C3 = {u3 (x + z) = 0 | u3 ∈ R}, C4 = {u3 (x − z) = 0 | u3 ∈ R},
C5 = {u3 (y + z) = 0 | u3 ∈ R}, C6 = {u3 (y − z) = 0 | u3 ∈ R}.
J. Overapproximating a convex compact set
In fact, overapproximating a compact semialgebraic set
by one or multiple ellipsoids is reasonable. If a compact
semialgebraic set is overapproximated by an ellipsoid (~x −
~v )T A(~x−~v ) ≤ 1, where A is a positive definite matrix and x, v
are vectors, then we can use (x−v)T A(x−v) = 1 equivalently
instead of (~x − ~v )T A(~x − ~v ) ≤ 1 for the verification, the
reason is that any trajectory originating from the interior of
the ellipsoid must pass through some point on the surface of
the ellipsoid.
K. Additional information for Example 6
α1 = − 0.027854x21 − 0.048663y12 − 0.014007x22 − 0.031091y22
− 0.019017
α2 = − 0.0227x21 − 0.027808y12 − 0.034068x22 − 0.030096y22
− 0.00016071
α3 = − 1.8255x21 + 1.8251y12 + 1.8211x22 − 1.8282y22 + 0.014587
Prog = 1.8533x41 − 3.5741x21 y12 − 3.6098x21x22 + 3.7075x21 y22
+ 1.8738y14 + 3.688y12x22 − 3.5944y12y22 + 1.8551x42
− 3.5851x22y22 + 1.8583y24 + 0.83561x31 − 0.94703x21y1
− 0.49941x21x2 − 0.74911x21y2 + 1.4599x1y12 + 0.4202x1x22
+ 0.93272x1y22 − 1.6545y13 − 0.61178y12x2 − 0.91767y12y2
− 0.47623y1x22 − 1.0571y1y22 − 0.7495x32 − 1.1242x22y2
− 0.66212x2y22 − 0.99318y23 + 23.1976x21 + 35.9056y12
+ 20.5634x22 + 27.7403y22 + 0.5705x1 − 0.64657y1
− 0.0035356x2 − 0.0053034y2 + 9.8186
L. Invariant Cluster for Subsection V-A
mqwd2
mM z Xd2
mgθ d21
g9 (~u, ~x) = −
+
+
+ d1 +
Z
Iyy Z
Z
Z
2
mgθ d2
Iyy Xqwd2
mqvd1
Iyy qwd1
+
u5 +
+
−
Z
Z
ZM
M
2
2
(Iyy X + Iyy Z )d2
gIyy Xθ d21
Xz
+x−
−
−
Z
ZM m
ZM
gIyy Xθ d22
Iyy qvd2
Iyy Xqvd1
−
u85
−
−
ZM
M
ZM
Mθ
+ d21 + d22 u8 + q 2 − 2
u17 + u1
Iyy
| 3cs.SY
|
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
arXiv:1505.05497v4 [math.GR] 17 Mar 2017
STÉPHANE LAMY
Abstract. We study the group Tame(A3 ) of tame automorphisms of the 3dimensional affine space, over a field of characteristic zero. We recover, in a
unified way, previous results of Kuroda, Shestakov, Umirbaev and Wright, about
the theory of reduction and the relations in Tame(A3 ). The novelty in our presentation is the emphasis on a simply connected 2-dimensional simplicial complex
on which Tame(A3 ) acts by isometries.
Contents
Introduction
1. Simplicial complex
1.A. General construction
1.B. Degrees
1.C. The complex in dimension 3
2. Parachute Inequality and Principle of Two Maxima
2.A. Degree of polynomials and forms
2.B. Parachute Inequality
2.C. Consequences
2.D. Principle of Two Maxima
3. Geometric theory of reduction
3.A. Degree of automorphisms and vertices
3.B. Elementary reductions
3.C. K-reductions
3.D. Elementary K-reductions
3.E. Proper K-reductions
3.F. Stability of K-reductions
4. Reducibility Theorem
4.A. Reduction paths
4.B. Reduction of a strongly pivotal simplex
4.C. Vertex with two low degree neighbors
4.D. Proof of the Reducibility Theorem
5. Simple connectedness
5.A. Consequences of the Reducibility Theorem
5.B. Local homotopies
6. Examples
References
Date: March 20, 2017.
This research was partially supported by ANR Grant “BirPol” ANR-11-JS01-004-01.
1
2
5
5
6
7
8
8
9
11
13
15
15
16
18
21
22
26
29
30
31
38
41
43
43
43
46
48
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
2
Introduction
Let k be a field, and let An = Ank be the affine space over k. We are interested in
the group Aut(An ) of algebraic automorphisms of the affine space. Concretely, we
choose once and for all a coordinate system (x1 , . . . , xn ) for An . Then any element
f ∈ Aut(An ) is a map of the form
f ∶ (x1 , . . . , xn ) ↦ (f1 (x1 , . . . , xn ), . . . , fn (x1 , . . . , xn )),
where the fi are polynomials in n variables, such that there exists a map g of
the same form satisfying f ○ g = id. We shall abbreviate this situation by writing
f = (f1 , . . . , fn ), and g = f −1 . Observe a slight abuse of notation here, since we
are really speaking about polynomials, and not about the associated polynomial
functions. For instance over a finite base field, the group Aut(An ) is infinite (for
n ⩾ 2) even if there is only a finite number of induced bijections on the finite number
of k-points of An .
The group Aut(An ) contains the following natural subgroups. First we have the
affine group An = GLn (k) ⋉ kn . Secondly we have the group En of elementary
automorphisms, which have the form
f ∶ (x1 , . . . , xn ) ↦ (x1 + P (x2 , . . . , xn ), x2 , . . . , xn ),
for some choice of polynomial P in n − 1 variables. The subgroup
Tame(An ) = ⟨An , En ⟩
generated by the affine and elementary automorphisms is called the subgroup of
tame automorphisms.
A natural question is whether the inclusion Tame(An ) ⊆ Aut(An ) is in fact an
equality. It is a well-known result, which goes back to Jung (see e.g. [Lam02] for a
review of some of the many proofs available in the literature), that the answer is yes
for n = 2 (over any base field), and it is a result by Shestakov and Umirbaev [SU04b]
that the answer is no for n = 3, at least when k is a field of characteristic zero.
The main purpose of the present paper is to give a self-contained reworked proof
of this last result: see Theorem 4.1 and Corollary 4.2. We follow closely the line of
argument by Kuroda [Kur10]. However, the novelty in our approach is the emphasis
on a 2-dimensional simplicial complex C on which Tame(A3 ) acts by isometries. In
fact, this construction is not particular to the 3-dimensional case: In §1 we introduce,
for any n ⩾ 2 and over any base field, a (n − 1)-dimensional simplicial complex on
which Tame(An ) naturally acts.
We now give an outline of the main notions and results of the paper. Since the
paper is quite long and technical, we hope that this informal outline will serve as
a guide for the reader, even if we cannot avoid being somewhat imprecise at this
point.
● The 2-dimensional complex C contains three kinds of vertices, corresponding to
three orbits under the action of Tame(A3 ). In this introduction we shall focus on
the so-called type 3 vertices, which correspond to a tame automorphism (f1 , f2 , f3 )
up to post-composition by an affine automorphism. Such an equivalence class is
denoted v3 = [f1 , f2 , f3 ] (see §1.C and Figure 1).
● Given a vertex v3 one can always choose a representative (f1 , f2 , f3 ) such that
the top monomials of the fi are pairwise distinct. Such a good representative is
not unique, but the top monomials are. This allows to define a degree function
(with values in N3 ) on vertices, with the identity automorphism corresponding to
the unique vertex of minimal degree (see §3.A).
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
3
By construction of the complex C, two type 3 vertices v3 and v3′ are at distance
2 in C if and only if they admit representatives of the form v3 = [f1 , f2 , f3 ] and
v3′ = [f1 + P (f2 , f3 ), f2 , f3 ], that is, representatives that differ by an elementary
automorphism. If moreover deg v3 > deg v3′ , we say that v3′ is an elementary reduction
of v3 . The whole idea is that it is ‘almost’ true that any vertex admits a sequence
of elementary reductions to the identity. However a lot of complications lie in this
‘almost’, as we now discuss.
● Some particular tame automorphisms are triangular automorphisms, of the form
(up to a permutation of the variables) (x1 + P (x2 , x3 ), x2 + Q(x3 ), x3 ). There are
essentially two ways to decompose such an automorphism as a product of elementary
automorphisms, namely
●
(x1 + P (x2 , x3 ), x2 + Q(x3 ), x3 )
= (x1 , x2 + Q(x3 ), x3 ) ○ (x1 + P (x2 , x3 ), x2 , x3 )
= (x1 + P (x2 − Q(x3 ), x3 ), x2 , x3 ) ○ (x1 , x2 + Q(x3 ), x3 ).
This leads to the presence of squares in the complex C, see Figure 3. Conversely,
the fact that a given vertex admits two distinct elementary reductions often leads
to the presence of such a square, in particular if one of the reductions corresponds
to a polynomial that depends only on one variable instead of two, like Q(x3 ) above.
We call ‘simple’ such a particular elementary reduction.
● When v ′ = [f1 + P (f2 , f3 ), f2 , f3 ] is an elementary reduction of v3 = [f1 , f2 , f3 ],
3
we shall encounter the following situations (see Corollary 4.28):
(i) The most common situation is when the top monomial of f1 is the dominant
one, that is, the largest among the top monomials of the fi .
(ii) Another (non-exclusive) situation is when P depends only on one variable.
As mentioned before this is typically the case when v3 admits several elementary
reductions. This corresponds to the fact that the top monomial of a component is
a power of another one, and we name ‘resonance’ such a coincidence (see definition
in §3.A).
(iii) Finally another situation is when f1 does not realize the dominant monomial,
but f2 , f3 nevertheless satisfy a kind of minimality condition via looking at the degree
of the 2-form df2 ∧ df3 . We call this last case an elementary K-reduction (see §3.C
for the definition, Corollary 3.13 for the characterization in terms of the minimality
of deg df2 ∧ df3 , and §6 for examples). This case is quite rigid (see Proposition 3.35),
and at posteriori it forbids the existence of any other elementary reduction from v3
(see Proposition 5.1).
● Finally we define (see §3.C again) an exceptional case, under the terminology
‘normal proper K-reduction’, that corresponds to moving from a vertex v3 to a
neighbor vertex w3 of the same degree, and then realizing an elementary K-reduction
from w3 to another vertex u3 . The fact that v3 and w3 share the same degree is not
part of the technical definition, but again is true only at posteriori (see Corollary
5.2).
● Then the main result (Reducibility Theorem 4.1) is that we can go from any
vertex to the vertex corresponding to the identify by a finite sequence of elementary
reductions or normal proper K-reductions.
● The proof proceeds by a double induction on degrees which is quite involved.
The Induction Hypothesis is precisely stated on page 36. The analysis is divided
into two main branches:
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
4
(i) §4.B where the slogan is “a vertex that admits an elementary K-reduction
does not admit any other reduction” (Proposition 4.27 is an intermediate technical
statement, and Proposition 5.1 the final one);
(ii) §4.C where the slogan is “a vertex that admits several elementary reductions
must admit some resonance” (see in particular Lemmas 4.32 and 4.34).
For readers familiar with previous works on the subject, we now give a few word
about terminology. In the work of Kuroda [Kur10], as in the original work of Shestakov and Umirbaev [SU04b], elementary reductions are defined with respect to one
of the three coordinates of a fixed coordinate system. In contrast, as explained
above, we always work up to an affine change of coordinates. Indeed our simplicial
complex C is designed so that two tame automorphisms correspond to two vertices
at distance 2 in the complex if and only if they differ by the left composition of an
automorphism of the form a1 ea2 , where a1 , a2 are affine and e is elementary. This
slight generalization of the definition of reduction allows us to absorb the so-called
“type I” and “type II” reductions of Shestakov and Umirbaev in the class of elementary reductions: In our terminology they become “elementary K-reductions” (see
§3.C). On the other hand, the “type III” reductions, which are technically difficult
to handle, are still lurking around. One can suspect that such reductions do not
exist (as the most intricate “type IV” reductions which were excluded by Kuroda
[Kur10]), and an ideal proof would settle this issue. Unfortunately we were not able
to do so, and these hypothetical reductions still appear in our text under the name
of “normal proper K-reduction”. See Example 6.5 for more comments on this issue.
One could say that the theory of Shestakov, Umirbaev and Kuroda consists in
understanding the relations inside the tame group Tame(A3 ). This was made explicit
by Umirbaev [Umi06], and then it was proved by Wright [Wri15] that this can be
rephrased in terms of an amalgamated product structure over three subgroups (see
Corollary 5.8). In turn, it is known that such a structure is equivalent to the action of
the group on a 2-dimensional simply connected simplicial complex, with fundamental
domain a simplex. Our approach allows to recover a more transparent description
of the relations in Tame(A3 ). After stating and proving the Reducibility Theorem
4.1 in §4, we directly show in §5 that the natural complex on which Tame(A3 ) acts
is simply connected (see Proposition 5.7), by observing that the reduction process
of [Kur10, SU04b] corresponds to local homotopies.
We should stress once more that this paper contains no original result, and consists
only in a new presentation of previous works by the above cited authors. In fact,
for the sake of completeness we also include in Section 2 some preliminary results
where we only slightly differ from the original articles [Kur08, SU04a].
Our motivation for reworking this material is to prepare the way for new results
about Tame(A3 ), such as the linearizability of finite subgroups, the Tits alternative or the acylindrical hyperbolicity. From our experience in related settings (see
[BFL14, CL13, Lon16, Mar15]), such results should follow from some non-positive
curvature properties of the simplicial complex. We plan to explore these questions
in some follow-up papers (see [LP16] for a first step).
Acknowledgement
In August 2010 I was assigned the paper [Kur10] as a reviewer for MathSciNet.
I took this opportunity to understand in full detail his beautiful proof, and in November of the same year, at the invitation of Adrien Dubouloz, I gave some lectures
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
5
in Dijon on the subject. I thank all the participants of this study group “Automorphismes des Espaces Affines”, and particularly Jérémy Blanc, Eric Edo, Pierre-Marie
Poloni and Stéphane Vénéreau who gave me useful feedback on some early version
of these notes.
A few years later I embarked on the project to rewrite this proof focusing on the
related action on a simplicial complex. At the final stage of the present version,
when I was somehow loosing faith in my endurance, I greatly benefited from the
encouragements and numerous suggestions of Jean-Philippe Furter.
Finally, I am of course very much indebted to Shigeru Kuroda. Indeed without
his work the initial proof of [SU04b] would have remained a mysterious black box
to me.
1. Simplicial complex
We define a (n − 1)-dimensional simplicial complex on which the tame automorphism group of An acts. This construction makes sense in any dimension n ⩾ 2, over
any base field k.
1.A. General construction. For any 1 ⩽ r ⩽ n, we call r-tuple of components
a morphism
f ∶ An → Ar
x = (x1 , . . . , xn ) ↦ (f1 (x), . . . , fr (x))
that can be extended as a tame automorphism f = (f1 , . . . , fn ) of An . One defines n
distinct types of vertices, by considering r-tuple of components modulo composition
by an affine automorphism on the range, r = 1, . . . , n. We use a bracket notation to
denote such an equivalence class:
[f1 , . . . , fr ] ∶= Ar (f1 , . . . , fr ) = {a ○ (f1 , . . . , fr ); a ∈ Ar }
where Ar = GLr (k) ⋉ kr is the r-dimensional affine group. We say that vr =
[f1 , . . . , fr ] is a vertex of type r, and that (f1 , . . . , fr ) is a representative of vr .
We shall always stick to the convention that the index corresponds to the type of a
vertex: for instance vr , vr′ , ur , wr , mr will all be possible notation for a vertex of type
r.
Now given n vertices v1 , . . . , vn of type 1, . . . , n, we attach a standard Euclidean
(n − 1)-simplex on these vertices if there exists a tame automorphism (f1 , . . . , fn ) ∈
Tame(An ) such that for all i ∈ {1, . . . , n}:
vi = [f1 , . . . , fi ].
We obtain a (n − 1)-dimensional simplicial complex Cn on which the tame group acts
by isometries, by the formulas
g ⋅ [f1 , . . . , fr ] ∶= [f1 ○ g−1 , . . . , fr ○ g−1 ].
Lemma 1.1. The group Tame(An ) acts on Cn with fundamental domain the simplex
[x1 ], [x1 , x2 ], . . . , [x1 , . . . , xn ].
In particular the action is transitive on vertices of a given type.
Proof. Let v1 , . . . , vn be the vertices of a simplex (recall that the index corresponds
to the type). By definition there exists f = (f1 , . . . fn ) ∈ Tame(An ) such that vi =
[f1 , . . . , fi ] for each i. Then
[x1 , . . . , xi ] = [(f1 , . . . , fi ) ○ f −1 ] = f ⋅ vi .
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
6
Remark 1.2.
(1) One could make a similar construction by working with the
full automorphism group Aut(An ) instead of the tame group. The complex Cn we
consider is the gallery connected component of the standard simplex [x1 ], [x1 , x2 ],
. . . , [x1 , . . . , xn ] in this bigger complex. See [BFL14, §6.2.1] for more details.
(2) When n = 2, the previous construction yields a graph C2 . It is not difficult to
show (see [BFL14, §2.5.2]) that C2 is isomorphic to the classical Bass-Serre tree of
Aut(A2 ) = Tame(A2 ).
1.B. Degrees. We shall compare polynomials in k[x1 , . . . , xn ] by using the graded
lexicographic order on monomials. We find it more convenient to work with an
additive notation, so we introduce the degree function, with value in Nn ∪ {−∞},
by taking
deg xa11 xa22 . . . xann = (a1 , a2 , . . . , an )
and by convention deg 0 = −∞. We extend this order to Qn ∪ {−∞}, since sometimes
it is convenient to consider difference of degrees, or degrees multiplied by a rational
number. The top term of g ∈ k[x1 , . . . , xn ] is the uniquely defined ḡ = cxd11 . . . xdnn
such that
(d1 , . . . , dn ) = deg g > deg(g − ḡ).
Observe that two polynomials f, g ∈ k[x1 , . . . , xn ] have the same degree if and only if
their top terms f¯, ḡ are proportional. If f = (f1 , . . . , fr ) is a r-tuple of components,
we call top degree of f the maximum of the degree of the fi :
topdeg f ∶= max deg fi ∈ Nn .
Lemma 1.3. Let f = (f1 , . . . , fr ) be a r-tuple of components, and consider V ⊂
k[x1 , . . . , xn ] the vector space generated by the fi . Then
(1) The set H of elements g ∈ V satisfying topdeg f > deg g is a hyperplane in
V;
(2) There exist a sequence of degrees δr > ⋅ ⋅ ⋅ > δ1 and a flag of subspaces V1 ⊂
⋅ ⋅ ⋅ ⊂ Vr = V such that dim Vi = i and deg g = δi for any g ∈ Vi ∖ Vi−1 .
Proof.
(1) Up to permuting the fi we can assume topdeg f = deg fr . Then for
each i = 1, . . . , r − 1 there exists a unique ci ∈ k such that deg fr > deg(fi + ci fr ). The
conclusion follows from the observation that an element of V is in H if and only if
it is a linear combination of the fi + ci fr , i = 1, . . . , r − 1.
(2) Immediate, by induction on dimension.
Using the notation of the lemma, we call r-deg f = (δ1 , . . . , δr ) the r-degree of f ,
and deg f = ∑ri=1 δi ∈ Nn the degree of f . Observe that for any affine automorphism
a ∈ Ar we have r- deg f = r- deg(a ○ f ), so we get a well-defined notion of r-degree
and degree for any vertex of type r.
If vr = [f1 , . . . , fr ] ∈ Cn with the deg fi pairwise distinct, we say that f is a good
representative of vr (we do not ask deg fr > ⋅ ⋅ ⋅ > deg f2 > deg f1 ). We use a double
bracket notation such as v2 = Jf1 , f2 K or v3 = Jf1 , f2 , f3 K, to indicate that we are using
a good representative.
Lemma 1.4. Let v1 , . . . , vn be a (n − 1)-simplex in Cn . Then there exists f =
(f1 , . . . , fn ) ∈ Tame(An ) such that vi = Jf1 , . . . , fi K for each n ⩾ i ⩾ 1.
Proof. We pick f1 such that v1 = Jf1 K, and we define the other fi by induction as
follows. If the i-degree of vi = Jf1 , . . . , fi K is (δ1 , . . . , δi ) (recall that by definition the
δj are equal to the degrees of the fj only up to a permutation), then there exist δ ∈ N3
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
[x1 +x3 ]
7
[x1 ,x3 ]
●
[x3 ]
[x1 ]
[x1 ,x2 ,x3 ]
[x2 ,x3 ]
[x1 ,x2 ]
●
●
[x1 +Q(x2 ),x2 ,x3 ]
[x1 ,x2 ,x3 +P (x1 ,x2 )]
[x2 ]
[x1 ,x2 ,x3 +Q(x2 )]
●
[x1 +Q(x2 ),x2 ]
●
[x1 +x3 +Q(x2 ),x2 ]
●
[x2 ,x3 +P (x1 ,x2 )]
[x1 +Q(x2 ),x2 ,x3 +P (x1 ,x2 )]
Figure 1. A few simplexes of the complex C.
and i + 1 ⩾ s ⩾ 1 such that the (i + 1)-degree of vi+1 is (δ1 , . . . , δs−1 , δ, δs , . . . , δi ). That
exactly means that there exists fi+1 of degree δ such that vi+1 = Jf1 , . . . , fi+1 K.
1.C. The complex in dimension 3. Now we specialize the general construction to
the dimension n = 3, which is our main interest in this paper. We drop the index and
simply denote by C the 2-dimensional simplicial complex associated to Tame(A3 ).
To get a first feeling of the complex one can draw pictures such as Figure 1, where
we use the following convention for vertices: , ● or corresponds respectively to
a vertex of type 1, 2 and 3. However one should keep in mind, as the following
formal discussion makes it clear, that the complex is not locally finite. A first step
in understanding the geometry of the complex C is to understand the link of each
type of vertex. In fact, we will now see that if the base field k is uncountable, then
the link of any vertex or any edge also has uncountably many vertices.
Consider first the link L(v3 ) of a vertex of type 3. By transitivity of the action of
Tame(A3 ), it is sufficient to describe the link L([id]). A vertex of type 1 at distance
1 from [id] has the form [a1 x1 +a2 x2 +a3 x3 ] where the ai ∈ k are uniquely defined up
to a common multiplicative constant. In other words, vertices of type 1 in L([id])
are parametrized by P2 . We denote by P2 (v3 ) this projective plane of vertices of
type 1 in the link of v3 . Similarly, vertices of type 2 in L(v3 ) correspond to lines in
P2 (v3 ), that is, to points in the dual projective space P̂2 (v3 ). The edges in L(v3 )
correspond to incidence relations (“a point belongs to a line”). We will often refer
to a vertex of type 2 as a “line in P2 (v3 )”. In the same vein, we will sometimes refer
to a vertex of type 1 as being “the intersection of two lines in P2 (v3 )”, or we will
express the fact that v1 and v2 are joined by an edge in C by saying “the line v2
passes through v1 ”.
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
8
Now we turn to the description of the link of a vertex v2 of type 2. By transitivity
we can assume v2 = [x1 , x2 ], and one checks that vertices of type 1 in L(v2 ) are
parametrized by P1 and are of the form
[a1 x1 + a2 x2 ], (a1 ∶ a2 ) ∈ P1 .
On the other hand vertices of type 3 in L(v2 ) are of the form
[x1 , x2 , x3 + P (x1 , x2 )], P ∈ k[y, z].
Precisely by taking the P without constant or linear part we obtain a complete set
of representatives for such vertices of type 3 in L(v2 ). Using the transitivity of the
action of Tame(A3 ) on vertices of type 2, the following lemma and its corollary are
then immediate:
Lemma 1.5. The link L(v2 ) of a vertex of type 2 is the complete bipartite graph
between vertices of type 1 and 3 in the link.
Corollary 1.6. Let v2 = [f1 , f2 ] and v3 = [f1 , f2 , f3 ] be vertices of type 2 and 3.
Then any vertex u3 distinct from v3 such that v2 ∈ P̂2 (u3 ) has the form
u3 = [f1 , f2 , f3 + P (f1 , f2 )]
where P ∈ k[y, z] is a non-affine polynomial in two variables (that is, not of the form
P (y, z) = ay + bz + c). In particular, v2 is the unique type 2 vertex in P̂2 (v3 ) ∩ P̂2 (u3 ).
The link of a vertex of type 1 is more complicated. Let us simply mention without
proof, since we won’t need it in this paper (but see Lemma 5.6 for a partial result,
and also [LP16, §3]), that in contrast with the case of vertices of type 2 or 3, the link
of a vertex of type 1 is a connected unbounded graph, which admits a projection to
an unbounded tree.
2. Parachute Inequality and Principle of Two Maxima
We recall here two results from [Kur08] (in turn they were adaptations from
[SU04a, SU04b]). The Parachute Inequality is the most important; we also recall
some direct consequences. From now on k denotes a field of characteristic zero.
2.A. Degree of polynomials and forms. Recall that we define a degree function
on k[x1 , x2 , x3 ] with value in N3 ∪{−∞} by taking deg xa11 xa22 xa33 = (a1 , a2 , a3 ) and by
convention deg 0 = −∞. We compare degrees using the graded lexicographic order.
We now introduce the notion of virtual degree in two distinct situations, which
should be clear by context.
Let g ∈ k[x1 , x2 , x3 ], and ϕ = ∑i∈I Pi y i ∈ k[x1 , x2 , x3 ][y] where Pi ≠ 0 for all i ∈ I,
that is, I is the support of ϕ. We define the virtual degree of ϕ with respect to g as
degvirt ϕ(g) ∶= max(deg Pi gi ) = max(deg Pi + i deg g).
i∈I
i∈I
Denoting by I¯ ⊆ I the subset of indexes i that realize the maximum, we also define
the top component of ϕ with respect to g as
ϕg ∶= ∑ P̄i y i .
i∈I¯
Similarly if g, h ∈ k[x1 , x2 , x3 ], and ϕ = ∑(i,j)∈S ci,j y i z j ∈ k[y, z] with support S,
we define the virtual degree of ϕ with respect to g and h as
degvirt ϕ(g, h) ∶= max deg gi hj = max (i deg g + j deg h).
(i,j)∈S
(i,j)∈S
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
9
Observe that ϕ(g, h) can be seen either as an element coming from ϕh (y) ∶=
ϕ(y, h) ∈ k[h][y] ⊂ k[x1 , x2 , x3 ][y] or from ϕ(y, z) ∈ k[y, z], and that the two possible notions of virtual degree coincide:
degvirt ϕh (g) = degvirt ϕ(g, h).
Example 2.1. In general we have degvirt ϕ(g) ⩾ deg ϕ(g) and degvirt ϕ(g, h) ⩾
deg ϕ(g, h). We now give two simple examples where these inequalities are strict.
(1) Let ϕ = x23 y − x3 y 2 , and g = x3 . Then ϕ(g) = 0, but
degvirt ϕ(g) = deg x33 = (0, 0, 3).
(2) Let ϕ = y 2 − z 3 , and g = x31 , h = x21 . Then ϕ(g, h) = 0, but
degvirt ϕ(g, h) = deg x61 = (6, 0, 0).
We extend the notion of degree to algebraic differential forms. Given
ω = ∑ fi1 ,⋯,ik dxi1 ∧ ⋯ ∧ dxik
where k = 1, 2 or 3 and fi1 ,⋯,ik ∈ k[x1 , x2 , x3 ], we define
deg ω ∶= max{deg fi1 ,⋯,ik xi1 ⋯xik } ∈ N3 ∪ {−∞}.
We gather some immediate remarks for future reference (observe that here we use
the assumption char k = 0).
Lemma 2.2. If ω, ω ′ are forms, and g is a non constant polynomial, we have
deg ω + deg ω ′ ⩾ deg ω ∧ ω ′ ;
deg g = deg dg;
deg gω = deg g + deg ω.
2.B. Parachute Inequality. If ϕ ∈ k[x1 , x2 , x3 ][y], we denote by ϕ(n) ∈ k[x1 , x2 , x3 ][y]
the nth derivative of ϕ with respect to y. We simply write ϕ′ instead of ϕ(1) .
Lemma 2.3. Let ϕ ∈ k[x1 , x2 , x3 ][y] and g ∈ k[x1 , x2 , x3 ]. Then:
(1) If degy ϕg ⩾ 1, then degvirt ϕ′ (g) = degvirt ϕ(g) − deg g.
g
(2) If degy ϕg ⩾ j ⩾ 1, then ϕ(j) = (ϕg )(j) .
Proof. We note as before
ϕ = ∑ Pi y i ,
i∈I
ϕg = ∑ P̄i y i
and
ϕ′ =
i∈I¯
i−1
∑ iPi y ,
i∈I∖{0}
where I is the support of ϕ, and I¯ ⊆ I is the subset of indexes i that realize the
maximum maxi∈I (deg Pi + i deg g).
Now if degy ϕg ⩾ 1, that is, if I¯ ≠ {0}, then the indexes in I¯∖{0} are precisely those
that realize the maximum maxi∈I∖{0} (deg Pi + (i − 1) deg g). Thus we get assertion
g
(1), and ϕ′ = (ϕg )′ . Assertion (2) for j ⩾ 2 follows by induction.
Lemma 2.4. Let ϕ ∈ k[x1 , x2 , x3 ][y] and g ∈ k[x1 , x2 , x3 ]. Then, for m ⩾ 0, the
following two assertions are equivalent:
(1) For j = 0, . . . , m − 1 we have degvirt ϕ(j) (g) > deg ϕ(j) (g), but
degvirt ϕ(m) (g) = deg ϕ(m) (g).
(2) There exists ψ ∈ k[x1 , x2 , x3 ][y] such that ψ(ḡ) ≠ 0 and
ϕg = (y − ḡ)m ⋅ ψ.
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
10
Proof. Observe that we have the equivalences
degvirt ϕ(g) > deg ϕ(g) ⇐⇒ ϕg (ḡ) = 0 ⇐⇒ y − ḡ divides ϕg .
(2.5)
g
First we prove (2) ⇒ (1). Assuming (2), by Lemma 2.3(2) we have ϕ(j) = (ϕg )(j)
for j = 0, . . . , m − 1. The second equivalence in (2.5) yields (ϕg )(j) (g) = 0 for j =
0, . . . , m − 1, and (ϕg )(m) (g) ≠ 0, and then the first equivalence gives the result.
To prove (1) ⇒ (2), it is sufficient to show that if degvirt ϕ(j) (g) > deg ϕ(j) (g) for
j = 0, . . . , k −1, then ϕg = (y − ḡ)k ⋅ψk for some ψk ∈ k[x1 , x2 , x3 ][y]. The remark (2.5)
g
gives it for k = 1. Moreover, by Lemma 2.3(2), if ϕg depends on y then ϕ′ = (ϕg )′ ,
hence the result by induction.
In the situation of Lemma 2.4, we call the integer m the multiplicity of ϕ with
respect to g, and we denote it by m(ϕ, g). In other words, the top term ḡ is a
multiple root of ϕg of order m(ϕ, g).
Following Vénéreau [Vén11], where a similar inequality is proved, we call the next
result a “Parachute Inequality”. Indeed its significance is that the real degree cannot
drop too much with respect to the virtual degree. However we follow Kuroda for
the proof.
Recall that (over a field k of characteristic zero) some polynomials f1 , ⋯, fr ∈
k[x1 , . . . , xn ] are algebraically independent if and only if df1 ∧ ⋯ ∧ dfr ≠ 0. Indeed
this is equivalent to asking that the map f = (f1 , . . . , fr ) from An to Ar is dominant,
which in turn is equivalent to saying that the differential of this map has maximal
rank on an open set of An (for details see for instance [HP94, Theorem III p.135]).
Proposition 2.6 (Parachute Inequality, see [Kur08, Theorem 2.1]). Let r = 2 or 3,
and let f1 , ⋯, fr ∈ k[x1 , x2 , x3 ] be algebraically independent. Let ϕ ∈ k[f2 , ⋯, fr ][y] ∖
{0}. Then
deg ϕ(f1 ) ⩾ degvirt ϕ(f1 ) − m(ϕ, f1 )(deg ω + deg f1 − deg df1 ∧ ω).
where ω = df2 if r = 2, or ω = df2 ∧ df3 if r = 3.
Proof. Denoting as before ϕ′ the derivative of ϕ with respect to y, we have
d(ϕ(f1 )) = ϕ′ (f1 )df1 + other terms involving df2 or df3 .
So we obtain d(ϕ(f1 )) ∧ ω = ϕ′ (f1 )df1 ∧ ω. Using Lemma 2.2 this yields
deg ϕ(f1 ) + deg ω = deg d(ϕ(f1 )) + deg ω ⩾ deg d(ϕ(f1 )) ∧ ω
= deg ϕ′ (f1 )df1 ∧ ω = deg ϕ′ (f1 ) + deg df1 ∧ ω,
which we can write as
− deg df1 ∧ ω + deg ω + deg ϕ(f1 ) ⩾ deg ϕ′ (f1 ).
(2.7)
Now we are ready to prove the inequality of the statement, by induction on
m(ϕ, f1 ).
If m(ϕ, f1 ) = 0, that is, if deg ϕ(f1 ) = degvirt ϕ(f1 ), there is nothing to do.
If m(ϕ, f1 ) ⩾ 1, it follows from Lemma 2.4 that m(ϕ′ , f1 ) = m(ϕ, f1 )−1. Moreover,
the condition m(ϕ, f1 ) ⩾ 1 implies that ϕf1 does depend on y, hence by Lemma 2.3(1)
we have
degvirt ϕ′ (f1 ) = degvirt ϕ(f1 ) − deg f1 .
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
11
By induction hypothesis, we have
deg ϕ′ (f1 ) ⩾ degvirt ϕ′ (f1 ) − m(ϕ′ , f1 )(deg ω + deg f1 − deg df1 ∧ ω)
= degvirt ϕ(f1 ) − deg f1 − (m(ϕ, f1 ) − 1)(deg ω + deg f1 − deg df1 ∧ ω)
= degvirt ϕ(f1 ) − m(ϕ, f1 )(deg ω + deg f1 − deg df1 ∧ ω)
− deg df1 ∧ ω + deg ω.
Combining with (2.7), and canceling the terms − deg df1 ∧ ω + deg ω on each side, one
obtains the expected inequality.
2.C. Consequences. We shall use the Parachute Inequality 2.6 mostly when r = 2,
and when we have a strict inequality degvirt ϕ(f1 , f2 ) > deg ϕ(f1 , f2 ). In this context
the following easy lemma is crucial. Ultimately this is here that lies the difficulty
when one tries to extend the theory in dimension 4 (or more!).
Lemma 2.8. Let f1 , f2 ∈ k[x1 , x2 , x3 ] be algebraically independent, and ϕ ∈ k[y, z]
such that degvirt ϕ(f1 , f2 ) > deg ϕ(f1 , f2 ). Then:
(1) There exist coprime p, q ∈ N∗ such that
p deg f1 = q deg f2 .
In particular, there exists δ ∈ N3 such that deg f1 = qδ, deg f2 = pδ, so that
the top terms of f1p and f2q are equal up to a constant: there exists c ∈ k such
that f¯1p = cf¯2q .
(2) Considering ϕ(f1 , f2 ) as coming from ϕ(y, f2 ) ∈ k[x1 , x2 , x3 ][y], we have
ϕf1 = (y p − f¯1p )m(ϕ,f1 ) ⋅ ψ = (y p − cf¯2q )m(ϕ,f1 ) ⋅ ψ
for some ψ ∈ k[x1 , x2 , x3 ][y].
Proof.
(1) We write ϕ(f1 , f2 ) = ∑ ci,j f1i f2j . Since degvirt ϕ(f1 , f2 ) > deg ϕ(f1 , f2 ),
there exist distinct (a, b) and (a′ , b′ ) such that
deg f1a f2b = deg f1a f2b = degvirt ϕ(f1 , f2 ).
′
′
Moreover we can assume that a, a′ are respectively maximal and minimal for this
property. We obtain
(a − a′ ) deg f1 = (b′ − b) deg f2 .
Dividing by m, the GCD of a − a′ and b′ − b, we get the expected relation.
(2) With the same notation, we have a = a′ + pm where m ⩾ 1, and in particular degy ϕ(y, f2 ) ⩾ p. So if p > degy P (y, f2 ) for some P ∈ k[f2 ][y], we have
degvirt P (f1 , f2 ) = deg P (f1 , f2 ). By the first assertion, there exists c ∈ k such that
deg f1p > deg (f1p − cf2q ). By successive Euclidean divisions in k[f2 ][y] we can write:
ϕ(y, f2 ) = ∑ Ri (y) (y p − cf2q )
i
with p > degy Ri for all i. Denote by I the subset of indexes such that
f1
i
ϕf1 = ∑ Ri (y p − cf¯2q ) .
(2.9)
i∈I
Let i0 be the minimal index in I. We want to prove that i0 ⩾ m(ϕ, f1 ). By contradiction, assume that m(ϕ, f1 ) > i0 . Since y− f¯1 is a simple factor of (y p −cf¯2q ) = (y p − f¯1p),
f1
and is not a factor of any Ri , we obtain that (y − f¯1 )i0 +1 divides all summands of
f1
(2.9) except Ri (y p − cf¯q )i0 . In particular (y − f¯1 )i0 +1 , hence also (y − f¯1 )m(ϕ,f1 ) ,
0
2
do not divide ϕf1 : This is a contradiction with Lemma 2.4.
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
12
We now list some consequences of the Parachute Inequality 2.6.
Corollary 2.10. Let f1 , f2 ∈ k[x1 , x2 , x3 ] be algebraically independent with deg f1 ⩾
deg f2 , and ϕ ∈ k[y, z] such that degvirt ϕ(f1 , f2 ) > deg ϕ(f1 , f2 ). Following Lemma
2.8, we write p deg f1 = q deg f2 where p, q ∈ N∗ are coprime. Then:
(1) deg ϕ(f1 , f2 ) ⩾ p deg f1 − deg f1 − deg f2 + deg df1 ∧ df2 ;
(2) If deg f1 ∈/ N deg f2 , then deg ϕ(f1 , f2 ) > deg df1 ∧ df2 ;
(3) Assume deg f1 ∈/ N deg f2 and deg f1 ⩾ deg ϕ(f1 , f2 ). Then p = 2, q ⩾ 3 is odd,
and
deg ϕ(f1 , f2 ) ⩾ deg f1 − deg f2 + deg df1 ∧ df2 .
If moreover deg f2 ⩾ deg ϕ(f1 , f2 ), then q = 3.
(4) Assume deg f1 ∈/ N deg f2 and deg f1 ⩾ deg ϕ(f1 , f2 ). Then
deg d(ϕ(f1 , f2 )) ∧ df2 ⩾ deg f1 + deg df1 ∧ df2 .
Proof.
(1) By Lemma 2.8(2), we have
degvirt ϕ(f1 , f2 ) ⩾ m(ϕ, f1 )p deg f1 .
(2.11)
On the other hand the Parachute Inequality 2.6 applied to ϕ(y, f2 ) ∈ k[f2 ][y] yields
deg ϕ(f1 , f2 ) ⩾ degvirt ϕ(f1 , f2 ) − m(ϕ, f1 )(deg f1 + deg f2 − deg df1 ∧ df2 ).
Combining with (2.11), and remembering that m(ϕ, f1 ) ⩾ 1, we obtain
deg ϕ(f1 , f2 ) ⩾
deg ϕ(f1 , f2 )
⩾ p deg f1 − deg f1 − deg f2 + deg df1 ∧ df2 .
m(ϕ, f1 )
(2) From Lemma 2.8 we have deg f1 = qδ and deg f2 = pδ for some δ ∈ N3 . The
inequality (1) gives
deg ϕ(f1 , f2 ) ⩾ p deg f1 − deg f1 − deg f2 + deg df1 ∧ df2 =
(pq − p − q)δ + deg df1 ∧ df2 . (2.12)
The assumption deg f1 /∈ N deg f2 implies q > p ⩾ 2. Thus pq − p − q > 0, and finally
deg ϕ(f1 , f2 ) > deg df1 ∧ df2 .
(3) Again the assumptions imply q > p ⩾ 2. Since qδ = deg f1 ⩾ deg ϕ(f1 , f2 ), we
get from (2.12) that q > pq − p − q. This is only possible if p = 2, and so q ⩾ 3 is odd.
Replacing p by 2 in (2.12), we get the inequality.
If deg f2 ⩾ deg ϕ(f1 , f2 ), we obtain 2δ > (q − 2)δ, hence q = 3.
(4) Denote ϕ(y, z) = ∑ ci,j y i z j , and consider the partial derivatives
ϕ′y (y, z) = ∑ ici,j y i−1 z j ;
ϕ′z (y, z) = ∑ jci,j y i z j−1 .
We have d(ϕ(f1 , f2 )) = ϕ′y (f1 , f2 )df1 +ϕ′z (f1 , f2 )df2 . In particular d(ϕ(f1 , f2 ))∧df2 =
ϕ′y (f1 , f2 )df1 ∧ df2 , and
deg d(ϕ(f1 , f2 )) ∧ df2 = deg ϕ′y (f1 , f2 ) + deg df1 ∧ df2 .
Now we consider ϕ′y (f1 , f2 ) as coming from ϕ′y (y, f2 ) ∈ k[f2 ][y], and we simply
write ϕ′ (f1 ) instead of ϕ′y (f1 , f2 ), in accordance with the convention for derivatives
introduced at the beginning of §2.B. We want to show deg ϕ′ (f1 ) ⩾ deg f1 . Recall
that by (2.11), degvirt ϕ(f1 ) ⩾ 2m(ϕ, f1 ) deg f1 , and so, using also Lemma 2.3(1):
degvirt ϕ′ (f1 ) = degvirt ϕ(f1 ) − deg f1 ⩾ 2(m(ϕ, f1 ) − 1) deg f1 + deg f1 .
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
13
The Parachute Inequality 2.6 then gives (for the last inequality recall that deg f1 ⩾
deg f2 by assumption):
deg ϕ′ (f1 ) ⩾ degvirt ϕ′ (f1 ) − m(ϕ′ , f1 )(deg f1 + deg f2 − deg df1 ∧ df2 )
⩾ 2(m(ϕ, f1 ) − 1) deg f1 + deg f1
− (m(ϕ, f1 ) − 1)(deg f1 + deg f2 − deg df1 ∧ df2 )
= (m(ϕ, f1 ) − 1)(deg f1 − deg f2 + deg df1 ∧ df2 ) + deg f1
⩾ deg f1 .
Corollary 2.13. Let f1 , f2 , f3 ∈ k[x1 , x2 , x3 ] be algebraically independent, and ϕ ∈
k[y, z] such that
degvirt ϕ(f1 , f2 ) > deg ϕ(f1 , f2 ),
degvirt ϕ(f1 , f2 ) > deg f3 .
Following Lemma 2.8, we write p deg f1 = q deg f2 where p, q ∈ N∗ are coprime. Then
deg(f3 + ϕ(f1 , f2 )) > p deg f1 − deg df2 ∧ df3 − deg f1 ,
Proof. The Parachute Inequality 2.6 applied to ψ = f3 + ϕ(y, f2 ) ∈ k[f2 , f3 ][y] gives
deg(f3 + ϕ(f1 , f2 )) ⩾ degvirt ψ(f1 )
− m(ψ, f1 )(deg df2 ∧ df3 + deg f1 − deg df1 ∧ df2 ∧ df3 ). (2.14)
By assumption degvirt ϕ(f1 , f2 ) > deg f3 . Thus not only degvirt ψ(f1 ) = degvirt ϕ(f1 ),
f1
but also ψ = ϕf1 , hence m(ψ, f1 ) = m(ϕ, f1 ) ⩾ 1. By Lemma 2.8(2), we obtain
degvirt ψ(f1 ) ⩾ m(ϕ, f1 )p deg f1 = m(ψ, f1 )p deg f1
Replacing in (2.14), and dividing by m(ψ, f1 ), we get the result.
2.D. Principle of Two Maxima. The proof of the next result, which we call the
“Principle of Two Maxima”, is one of the few places where the formalism of Poisson
brackets used by Shestakov and Umirbaev seems to be more transparent (at least
for us) than the formalism of differential forms used by Kuroda. In this section we
propose a definition that encompasses the two points of view, and then we recall the
proof following [SU04a, Lemma 5].
Proposition 2.15 (Principle of Two Maxima, [Kur08, Theorem 5.2] and [SU04a,
Lemma 5]). Let (f1 , f2 , f3 ) be an automorphism of A3 . Then the maximum between
the following three degrees is realized at least twice:
deg f1 + deg df2 ∧ df3 ,
deg f2 + deg df1 ∧ df3 ,
deg f3 + deg df1 ∧ df2 .
Let Ω be the space of algebraic 1-forms ∑ fi dgi where fi , gi ∈ k[x1 , x2 , x3 ]. We
consider Ω as a free module of rank three over k[x1 , x2 , x3 ], with basis dx1 , dx2 , dx3 ,
and we denote by
∞
T = ⊕ Ω⊗p
p=0
the associative algebra of tensorial powers of Ω, where as usual Ω⊗0 = k[x1 , x2 , x3 ].
The degree function on Ω extends naturally to a degree function on T. Recall that
T has a natural structure of Lie algebra: For any ω, µ ∈ T, we define their bracket as
[ω, µ] ∶= ω ⊗ µ − µ ⊗ ω.
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
14
In particular, if df, dg ∈ Ω are 1-forms, we have
[df, dg] = df ⊗ dg − dg ⊗ df = df ∧ dg.
It is easy to check that the bracket satisfies the Jacobi identity: For any α, β, γ ∈ T,
we have
[[α, β], γ] + [[β, γ], α] + [[γ, α], β]
= α⊗β ⊗γ −β ⊗α⊗γ −γ ⊗α⊗β +γ ⊗β ⊗α
+β ⊗γ ⊗α−γ ⊗β ⊗α−α⊗β ⊗γ +α⊗γ ⊗β
+γ ⊗α⊗β −α⊗γ ⊗β −β ⊗γ ⊗α+β ⊗α⊗γ
=0
since each one of the six possible permutations appears twice, with different signs.
Lemma 2.16. The nine elements [dxi ∧ dxj , dxk ], for 1 ⩽ i < j ⩽ 3 and 1 ⩽ k ⩽ 3
generate a 8-dimensional free submodule in T, the only relation between them being
the Jacobi identity:
[dx1 ∧ dx2 , dx3 ] + [dx2 ∧ dx3 , dx1 ] − [dx1 ∧ dx3 , dx2 ] = 0.
Proof. We work inside the 27-dimensional free sub-module of T generated by the
dxi ⊗ dxj ⊗ dxk for 1 ⩽ i, j, k ⩽ 3. We compute, for i < j:
[dxi ∧ dxj , dxi ] = [dxi ⊗ dxj − dxj ⊗ dxi , dxi ]
= 2dxi ⊗ dxj ⊗ dxi − dxj ⊗ dxi ⊗ dxi − dxi ⊗ dxi ⊗ dxj ,
[dxi ∧ dxj , dxj ] = [dxi ⊗ dxj − dxj ⊗ dxi , dxj ]
= −2dxj ⊗ dxi ⊗ dxj + dxi ⊗ dxj ⊗ dxj + dxj ⊗ dxj ⊗ dxi .
This shows that the elements [dxi ∧ dxj , dxi ] and [dxi ∧ dxj , dxj ], for i < j, generate
a 6-dimensional free submodule. On the other hand, for {i, j, k} = {1, 2, 3}:
[dxi ∧ dxj , dxk ] = [dxi ⊗ dxj − dxj ⊗ dxi , dxk ]
= dxi ⊗ dxj ⊗ dxk − dxj ⊗ dxi ⊗ dxk − dxk ⊗ dxi ⊗ dxj + dxk ⊗ dxj ⊗ dxi ,
so that [dx1 ∧ dx2 , dx3 ] and [dx2 ∧ dx3 , dx1 ] are independent, and together with the
above family they generate a 8-dimensional free submodule.
The proof of the Principle of Two Maxima 2.15 now follows from the observation:
Lemma 2.17. Let f, g, h ∈ k[x1 , x2 , x3 ], with h non-constant. Then
deg [[df, dg] , dh] = deg h + deg df ∧ dg.
Proof. We have
dh = ∑
1⩽k⩽3
and
[df, dg] = df ∧ dg =
∂h
∂xk dxk
∂f ∂g
∑ ( ∂xi ∂xj −
1⩽i<j⩽3
∂f ∂g
∂xj ∂xi ) dxi
∧ dxj .
Thus
[[df, dg], dh] = ∑
∑
1⩽k⩽3 1⩽i<j⩽3
∂h
∂xk
∂f ∂g
∂f ∂g
( ∂x
− ∂x
) [dxi ∧ dxj , dxk ] .
i ∂xj
j ∂xi
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
15
∂h
dxk , k = 1, 2, 3, then
If the degree of dh is realized by at most two of the terms ∂x
k
by Lemma 2.16 the terms realizing the maximum of the degrees
∂f ∂g
∂h
deg ( ∂x
( ∂x
−
i ∂xj
k
∂f ∂g
∂xj ∂xi ) [dxi
∧ dxj , dxk ])
(2.18)
are independent (because at most two of them occur in the Jacobi relation), hence
the result since deg [dxi ∧ dxj , dxk ] = deg xi xj xk .
∂h
dxk have the same degree, then among
On the other hand if the three terms ∂x
k
the indexes (i, j, k) that realize the maximum of the degrees in (2.18), we must find
some with k = i or k = j, hence again we get the conclusion since by Lemma 2.16
such terms cannot cancel each other.
Proof of the Principle of Two Maxima 2.15. Since by the Jacobi identity
[[df1 , df2 ], df3 ] + [[df2 , df3 ], df1 ] + [[df3 , df1 ], df2 ] = 0,
the dominant terms must cancel each other. In particular the maximum of the
degrees, which are computed in Lemma 2.17, is realized at least twice: This is the
five lines proof of the Principle of Two Maxima by Shestakov and Umirbaev!
3. Geometric theory of reduction
In this section we mostly follow Kuroda [Kur10], but we reinterpret his theory of
reduction in a combinatorial way, using the complex C. Recall that k is a field of
characteristic zero.
3.A. Degree of automorphisms and vertices. Recall that in §3.A we defined
a notion of degree for an automorphism f = (f1 , f2 , f3 ) ∈ Tame(A3 ). The point is
that we want a degree that is adapted to the theory of reduction of Kuroda, so for
instance taking the maximal degree of the three components of an automorphism is
not good, because we would not detect a reduction of the degree on one of the two
lower components (such reductions do exist, see §6). We also want a definition that
is adapted to working on the complex C, so directly taking the sum of the degree of
the three components is no good either, since it would not give a degree function on
vertices of C.
Recall that the 3-degree of f ∈ Tame(A3 ), or of the vertex v3 = [f ], is the triple
(δ1 , δ2 , δ3 ) given by Lemma 1.3, where in particular δ3 > δ2 > δ1 . By definition the
top degree of v3 is δ3 ∈ N3 , and the degree of v3 is the sum
deg v3 ∶= δ1 + δ2 + δ3 ∈ N3 .
Similarly we have a 2-degree (ν1 , ν2 ) associated with any vertex v2 of type 2, a top
degree equal to ν2 and a degree deg v2 ∶= ν1 + ν2 . Finally for a vertex of type 1 the
notions of 1-degree, top degree and degree coincide.
Lemma 3.1. Let v3 be a vertex of type 3. Then
deg v3 ⩾ (1, 1, 1)
with equality if and only if v3 = [id].
Proof. If v3 = [f ] with deg v3 = (1, 1, 1), then the 3-degree (δ1 , δ2 , δ3 ) of v3 must be
equal to ((0, 0, 1), (0, 1, 0), (0, 0, 1)), hence the result.
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
16
Let v3 be a vertex with 3-degree (δ1 , δ2 , δ3 ). The unique m1 ∈ P2 (v3 ) such that
deg m1 = δ1 is called the minimal vertex in P2 (v3 ), and the unique m2 ∈ P̂2 (v3 )
such that deg m2 = (δ1 , δ2 ) is called the minimal line in P2 (v3 ). If v2 ∈ P̂2 (v3 )
has 2-degree (ν1 , ν2 ), there is a unique degree δ such that v3 has degree ν1 + ν2 + δ.
We denote this situation by δ ∶= deg(v3 ∖ v2 ). Observe that, by definition, deg(v3 ∖
v2 ) = deg v3 − deg v2 . There is also a unique v1 such that v2 passes through v1 and
deg v1 = ν1 : we call v1 the minimal vertex of v2 . Observe that if v2 = m2 ∈ P̂2 (v3 ),
then the minimal vertex of v2 coincides with the minimal vertex of v3 , and if v2 ≠ m2 ,
then v1 is the intersection of v2 with m2 .
We call triangle T in P2 (v3 ) the data of three non-concurrent lines. A good
triangle T in P2 (v3 ) is the data of three distinct lines m2 , v2 , u2 , such that m2 is
the minimal line, v2 passes through the minimal vertex m1 , and u2 does not pass
through m1 . Equivalently, a good triangle corresponds to a good representative
v3 = Jf1 , f2 , f3 K with deg f1 > deg f2 > deg f3 , by putting m2 = Jf2 , f3 K, v2 = Jf1 , f3 K,
u2 = Jf1 , f2 K. If v1 , v2 , v3 is a simplex in C, we say that a good triangle T in P2 (v3 ) is
compatible with this simplex if v2 is one of the lines of T , and v1 is the intersection
of v2 with another line of T . Each simplex v1 , v2 , v3 admits such a compatible
good triangle (not unique in general): Indeed it corresponds to a choice of good
representatives as given by Lemma 1.4.
Let v2 ∈ P̂2 (v3 ) be a vertex with 2-degree (δ1 , δ2 ). We say that v2 has inner
resonance if δ2 ∈ Nδ1 . We say that v2 has outer resonance in v3 if deg(v3 ∖ v2 ) ∈
Nδ1 + Nδ2 .
3.B. Elementary reductions. Let v3 , v3′ be vertices of type 3. We say that v3′
is a neighbor of v3 if v3′ ≠ v3 and there exists a vertex v2 of type 2 such that
v2 ∈ P̂2 (v3 ) ∩ P̂2 (v3′ ). Equivalently, this means that v3 and v3′ are at distance 2 in C.
We denote this situation by v3′ ≬ v3 , or if we want to make v2 explicit, by v3′ ≬v2 v3
We also say that v2 is the center of v3 ≬ v3′ . Recall that the center v2 is uniquely
defined, and that we can choose representatives as in Corollary 1.6.
We say that v3′ is an elementary reduction (resp. a weak elementary reduction)
of v3 with center v2 , if deg v3 > deg v3′ (resp. deg v3 ⩾ deg v3′ ) and v3′ ≬v2 v3 . Let v1
be the minimal vertex in the line v2 . We say that v1 , v2 , v3 is the pivotal simplex
of the reduction, and that v1 is the pivot of the reduction. Moreover we say that
the reduction is optimal if v3′ has minimal degree among all neighbors of v3 with
center v2 . We say that v3′ is a simple elementary reduction (resp. a weak simple
elementary reduction) of v3 if there exist good representatives v3 = Jf1 , f2 , f3 K and
v3′ = Jf1 + P (f2 ) + af3 , f2 , f3 K for some a ∈ k and non-affine polynomial P , satisfying
deg f1 > max{f1 + P (f2 ), f1 + P (f2 ) + af3 };
(resp. f1 ⩾ max{f1 + P (f2 ), f1 + P (f2 ) + af3 }).
In this situation we say that v2 = Jf2 , f3 K, v1 = Jf2 K is the simple center of the
reduction, and when drawing pictures we represent this relation by adding an arrow
on the edge from v2 to v1 (see Figure 2). Beware that this representation is imperfect,
since the arrow does not depend only on the edge from v2 to v1 but indicates a
relation between the two vertices v3 and v3′ .
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
17
v1 = [f2 ]
v3 = Jf1 ,f2 ,f3 K
●
v2 = Jf2 ,f3 K
v3′ = Jf1 +P (f2 )+af3 ,f2 ,f3 K
Figure 2. Simple reduction with simple center v2 , v1 .
Remark 3.2. In the definition of a (weak) simple elementary reduction, if deg f3 =
topdeg v3 , then we must have a = 0. For instance in the following example:
v3 = Jx1 + x22 , x2 , x3 + x22 + x32 K = Jf1 , f2 , f3 K,
v3′ = Jx1 − x3 , x2 , x3 + x22 + x32 K = Jf1 + f23 − f3 , f2 , f3 K,
we do not want to call v3′ a simple reduction of v3 because deg(f1 + f23 ) > deg f1 .
On the other hand, consider the following example:
v3 = Jx1 + x22 + x32 , x2 , x3 + x22 K = Jf1 , f2 , f3 K,
v3′ = Jx1 − x3 , x2 , x3 + x22 K = Jf1 − f23 − f3 , f2 , f3 K.
Here v3′ is a simple elementary reduction of v3 , and the coefficient a = −1 is necessary
to get a good representative.
Lemma 3.3. Let v3′ be a neighbor of v3 = [f1 , f2 , f3 ] with center v2′ = Jf1 , f2 K. Then
there exists a non-affine polynomial P ∈ k[y, z] such that v3′ = Jf1 , f2 , f3 + P (f1 , f2 )K.
Moreover:
(1) If v3′ is a weak elementary reduction of v3 , then deg f3 ⩾ deg P (f1 , f2 );
(2) If v3′ is an elementary reduction of v3 , then deg f3 = deg P (f1 , f2 ).
Proof. From Corollary 1.6 we know that v3′ has the form v3′ = [f1 , f2 , f3 + P (f1 , f2 )].
Since by assumption deg f1 ≠ deg f2 , there exist a, b ∈ k such that (f1 , f2 , f3 +
P (f1 , f2 ) + af1 + bf2 ) is a good representative for v3′ . So up to changing P by a
linear combination of f1 and f2 we can assume v3′ = Jf1 , f2 , f3 + P (f1 , f2 )K.
If v3′ is a weak elementary reduction of v3 , then we have
deg f1 + deg f2 + deg f3 ⩾ deg v3 ⩾ deg v3′ = deg f1 + deg f2 + deg(f3 + P (f1 , f2 )).
So deg f3 ⩾ deg(f3 + P (f1 , f2 )), which implies deg f3 ⩾ deg P (f1 , f2 ).
Finally if v3′ is an elementary reduction of v3 , that is, deg v3 > deg v3′ , then the
same computation gives deg f3 > deg(f3 + P (f1 , f2 )), which implies that deg f3 =
deg P (f1 , f2 ).
Lemma 3.4 (Square Lemma). Let v3 , v3′ , v3′′ be three vertices such that:
● v ′ ≬v′ v3 and v ′′ ≬v′′ v3 for some v ′ ≠ v ′′ that are part of a good triangle of v3
3
3
2
2
2
2
(this is automatic if v2′ or v2′′ is the minimal line of v3 );
● Denoting v1 the common vertex of v ′ and v ′′ , v ′′ is a (possibly weak) simple
2
2
3
elementary reduction of v3 with simple center v2′′ , v1 ;
● deg v3 ⩾ deg v ′ , deg v3 ⩾ deg v ′′ , with at least one of the inequalities being strict.
3
3
Then there exists u3 such that u3 ≬ v3′ , u3 ≬ v3′′ and deg v3 > deg u3 .
Proof. We pick f2 such that v1 = [f2 ], and then we take good representatives v2′ =
Jf1 , f2 K, v2′′ = Jf2 , f3 K. Since v2′ and v2′′ are part of a good triangle, we have v3 =
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
18
v3 =Jf1 ,f2 ,f3 K
v2′′ =Jf2 ,f3 K
●
●
v2′ =Jf1 ,f2 K
v1 =[f2 ]
v3′′ =Jf1 +af3 +Q(f2 ),f2 ,f3 K
●
v3′ =Jf1 ,f2 ,f3 +P (f1 ,f2 )K
●
u′2 =[f1 +Q(f2 ),f2 ]
u′′
2 =Jf2 ,f3 +P (f1 ,f2 )K
u3 =[f1 +Q(f2 ),f2 ,f3 +P (f1 ,f2 )]
Figure 3. Square Lemma 3.4.
Jf1 , f2 , f3 K. By Lemma 3.3, and since v3′′ is a (possibly weak) simple reduction of v3 ,
there exist a ∈ k, Q ∈ k[f2 ] and P ∈ k[f1 , f2 ] such that
v3′ = Jf1 , f2 , f3 + P (f1 , f2 )K;
v3′′ = Jf1 + af3 + Q(f2 ), f2 , f3 K.
We have
deg f3 ⩾ deg(f3 + P (f1 , f2 )),
deg f1 ⩾ max{deg(f1 + af3 + Q(f2 )), deg(f1 + Q(f2 ))},
with one of the two inequalities being strict. We define
u3 ∶= [f1 + Q(f2 ), f2 , f3 + P (f1 , f2 )].
Observe that u3 is a neighbor of both v3′ , with center Jf2 , f3 + P (f1 , f2 )K, and v3′′ ,
with center [f1 + Q(f2 ), f2 ] (see Figure 3). The inequality on degrees follows from:
deg v3 = deg f3 + deg f2 + deg f1
> deg(f3 + P (f1 , f2 )) + deg f2 + deg(f1 + Q(f2 ))
⩾ deg u3 .
3.C. K-reductions. If f1 , f2 ∈ k[x1 , x2 , x3 ] are two algebraically independent polynomials with deg f1 > deg f2 , we introduce the degree
∆(f1 , f2 ) ∶= deg f1 − deg f2 + deg df1 ∧ df2 ∈ Z3 .
Assuming that v2 = Jf1 , f2 K is a vertex of type 2, we define
d(v2 ) ∶= deg df1 ∧ df2
and
∆(v2 ) ∶= ∆(f1 , f2 ).
We call d(v2 ) and ∆(v2 ) respectively the differential degree and the delta degree of v2 . It is easy to check that these definitions do not depend on a choice of
representative. In fact, for any ( αγ βδ ) ∈ GL2 (k) we have
d(αf1 + βf2 ) ∧ d(γf1 + δf2 ) = (αδ − βγ)df1 ∧ df2 ,
so in the definition of d(v2 ) we could use any representative. On the other hand in
the definition of ∆(v2 ), because of the term deg f1 − deg f2 , we really need to work
with a good representative. Observe also that, by definition, for any vertex v2 we
have
∆(v2 ) > d(v2 ).
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
v3 =Jf1 ,f2 ,f3 K
↺●
v1 =[f2 ]
19
w3 =Jg1 ,f2 ,f3 K
v2 =Jf1 ,f2 K
u3 =Jf1 ,f2 ,g3 K
m2 =Jf2 ,f3 K
w2 =Jg1 ,f2 K
↺●
●
v3 =Jf1 ,f2 ,f3 K
u3 =Jg1 ,f2 ,g3 K
v1 =[f2 ]
Figure 4. Elementary and proper K-reductions, with good representatives as in Set-Up 3.5.
We now introduce the key concept of K-reduction, where we let the reader decide
for himself whether the K should stand for “Kuroda” or for “Kazakh”.
More precisely by a K-reduction we shall mean either an elementary K-reduction,
or a proper K-reduction, two notions that we now define. Let v3 and u3 be vertices
of type 3.
We say that u3 is an elementary K-reduction of v3 if u3 ≬ v3 and:
(K0)
(K1)
(K2)
(K3)
(K4)
deg v3 > deg u3 ;
the center v2 of v3 ≬ u3 has no inner resonance;
v2 has no outer resonance in v3 ;
v2 is not the minimal line in P2 (v3 );
∆(v2 ) > deg(u3 ∖ v2 ).
Denoting by v1 the minimal point in v2 , as before (see definition from page 16) we
call v1 the pivot, and v1 , v2 , v3 the pivotal simplex of the elementary K-reduction
(denoted by ↺ on Figure 4).
We say that u3 is a proper K-reduction of v3 via the auxiliary vertex w3 if
v3 is a weak elementary reduction of w3 with center the minimal line of w3 , and
u3 is an elementary K-reduction of w3 . Formally, this corresponds to the following
conditions:
(K0′ )
(K1′ )
(K2′ )
(K3′ )
(K4′ )
(K5′ )
(K6′ )
deg w3 > deg u3 ;
the center w2 of w3 ≬ u3 has no inner resonance;
w2 has no outer resonance in w3 ;
w2 is not the minimal line in P2 (w3 );
∆(w2 ) > deg(u3 ∖ w2 ).
deg w3 ⩾ deg v3 ;
the center m2 of w3 ≬ v3 is the minimal line in P2 (w3 ).
Observe that the pivot v1 of the elementary K-reduction from w3 to u3 is the common
vertex of the distinct lines m2 and w2 in P2 (w3 ). The simplex v1 , w2 , w3 is still called
the pivotal simplex of the proper K-reduction. It will be proved in Proposition
3.26 that the above conditions (K0′ ) to (K6′ ) imply deg v3 > deg u3 , so that the
terminology of “reduction” is not misleading, even if by no means obvious at this
point.
Let v1 , v2 , v3 be a simplex, and let s ⩾ 3 be an odd integer. We say that the
simplex v1 , v2 , v3 has Strong Pivotal Form ↺(s) if
(↺1)
(↺2)
(↺3)
(↺4)
deg v1 = 2δ and 2-deg v2 = (2δ, sδ) for some δ ∈ N3 ;
v2 has no outer resonance in v3 ;
v2 is not the minimal line in P2 (v3 );
deg(v3 ∖ v2 ) ⩾ ∆(v2 ).
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
20
In all the previous definitions we took care of working with vertices, and not
with particular representatives. However for writing proofs it will often be useful to
choose representatives.
Set-Up 3.5.
(1) Let u3 be an elementary K-reduction of v3 , with pivotal simplex
v1 , v2 , v3 . Then there exist representatives
v1 = [f2 ]
v2 = Jf1 , f2 K
v3 = Jf1 , f2 , f3 K
u3 = Jf1 , f2 , g3 K
such that g3 = f3 + ϕ3 (f1 , f2 ), where ϕ3 ∈ k[x, y]. Observe that, by definition,
deg f1 > deg f2 and deg f1 > deg f3 > deg g3 .
(2) Let u3 be a proper K-reduction of v3 , with pivotal simplex v1 , w2 , w3 . Then
there exist representatives
v1 = [f2 ]
w3 = Jg1 , f2 , f3 K
m2 = Jf2 , f3 K
v3 = Jf1 , f2 , f3 K
w2 = Jg1 , f2 K
u3 = Jg1 , f2 , g3 K
such that g1 = f1 + ϕ1 (f2 , f3 ) and g3 = f3 + ϕ3 (g1 , f2 ), where ϕ1 , ϕ3 ∈ k[x, y].
Proof.
(1) Pick any good representatives v1 = [f2 ], v2 = Jf1 , f2 K, v3 = Jf1 , f2 , f3 K
as given by Lemma 1.4, and apply Lemma 3.3 to get g3 .
(2) Pick any representative v1 = [f2 ], and then pick f3 , g1 such that v2 = Jf2 , f3 K
and w2 = Jg1 , f2 K. Since m2 is the minimal line in P̂2 (w3 ), we have deg g1 > deg f2
and deg g1 > deg f3 , hence (g1 , f2 , f3 ) is a good representative for w3 . Now apply
Lemma 3.3 twice to get f1 and g3 .
We establish a first property of a simplex with Strong Pivotal Form.
Lemma 3.6. Let v1 , v2 , v3 be a simplex with Strong Pivotal Form ↺(s) for some
odd s ⩾ 3. Then the minimal line m2 in P2 (v3 ) has no inner resonance.
Proof. We pick representatives v1 = [f2 ], v2 = Jf1 , f2 K, v3 = Jf1 , f2 , f3 K as given by
Lemma 1.4. By (↺1) we have deg f1 = sδ > 2δ = deg f2 . Since v2 is not the
minimal line by (↺3), the minimal line must be m2 = Jf2 , f3 K. Then by (↺4)
we have deg f3 > (s − 2)δ ⩾ δ, so that deg f2 /∈ N deg f3 . Since by (↺2) we also
have deg f3 ∈/ N deg f2 , we conclude that the minimal line m2 = Jf2 , f3 K has no inner
resonance.
We can rephrase results from Corollary 2.10 with the previous definitions (see
also Example 6.3 for some complements):
Proposition 3.7. Let v3 be a vertex that admits an elementary reduction with pivotal simplex v1 , v2 , v3 .
(1) Assume v2 has no inner resonance, and no outer resonance in v3 . Then
deg(v3 ∖ v2 ) > d(v2 ).
(2) If moreover v2 is not the minimal line in P2 (v3 ), then v1 , v2 , v3 has Strong
Pivotal Form ↺(s) for some odd s ⩾ 3.
Proof.
(1) We pick good representatives v1 = [f2 ], v2 = Jf1 , f2 K, v3 = Jf1 , f2 , f3 K as
given by Lemma 1.4. By Lemma 3.3, the elementary reduction has the form u3 =
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
21
Jf1 , f2 , f3 + P (f1 , f2 )K with deg f3 = deg P (f1 , f2 ). Since v2 has no outer resonance
in v3 , we have deg f3 /∈ N deg f1 + N deg f2 , hence
degvirt P (f1 , f2 ) > deg P (f1 , f2 ).
Since moreover v2 has no inner resonance, we can apply Corollary 2.10(2) to get the
inequality deg(v3 ∖ v2 ) > d(v2 ).
(2) By assumption the simplex v1 , v2 , v3 already satisfies conditions (↺2) and
(↺3). The condition that v2 is not the minimal line in P2 (v3 ) is equivalent to
max{deg f1 , deg f2 } > deg f3 = deg P (f1 , f2 ),
hence we can apply Corollary 2.10(3), which yields conditions (↺1) and (↺4).
3.D. Elementary K-reductions. Here we list some properties of an elementary
K-reduction. First we have the following corollary from Proposition 3.7.
Corollary 3.8. The pivotal simplex of a K-reduction has Strong Pivotal Form ↺(s)
for some odd s ⩾ 3.
Proof. First, let v1 , v2 , v3 be the pivotal simplex of an elementary K-reduction. We
know that, by (K1), v2 has no inner resonance, by (K2), v2 has no outer resonance
in v3 , and by (K3), v2 is not the minimal line in v3 , so we can apply Proposition
3.7(2).
Now the pivotal simplex of proper K-reduction is by definition the pivotal simplex
of an elementary K-reduction from the auxiliary vertex, so that the above argument
applies.
Lemma 3.9. Let u3 be an elementary K-reduction of v3 with center v2 , m2 the
minimal line in P2 (v3 ), and u2 ∈ P2 (v3 ) a line not passing through the pivot v1 of
the reduction. Then
(1) d(m2 ) ⩾ deg(v3 ∖ m2 ) + d(v2 );
(2) d(u2 ) > d(m2 ) > d(v2 );
(3) The function
t2 ∈ P̂2 (v3 ) ↦ d(t2 ) ∈ N3
only takes the three distinct values d(u2 ) > d(m2 ) > d(v2 ), and it takes its
minimal value only at the point v2 ;
(4) deg(v1 ) + d(u2 ) > 2 deg(v3 ∖ m2 ).
Proof. The assumption means that v2 , m2 , u2 form a (not necessarily good) triangle.
We use the notation from Set-Up 3.5, we therefore have m2 = Jf2 , f3 K, v2 = Jf1 , f2 K
and v1 = [f2 ].
(1) On the one hand:
df2 ∧ df3 = df2 ∧ dg3 − df2 ∧ d(ϕ3 (f1 , f2 )).
On the other hand, the following sequence of inequalities holds, where the first one
comes from Corollary 2.10(4), the second one from (K4), and the third one from
Lemma 2.2:
deg df2 ∧ d(ϕ3 (f1 , f2 )) ⩾ deg f1 + deg df1 ∧ df2 > deg f2 + deg g3 ⩾ deg df2 ∧ dg3 . (3.10)
So deg df2 ∧ df3 = deg df2 ∧ d(ϕ3 (f1 , f2 )), and replacing in (3.10) we obtain the expected inequality:
d(m2 ) = deg df2 ∧ df3 ⩾ deg f1 + deg df1 ∧ df2 = deg(v3 ∖ m2 ) + d(v2 ).
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
22
(2) By the previous point we have
deg f1 + deg df2 ∧ df3 > deg f3 + deg df1 ∧ df2 ,
hence by the Principle of Two Maxima 2.15 we get
deg f2 + deg df1 ∧ df3 = deg f1 + deg df2 ∧ df3 .
(3.11)
Since deg f1 > deg f2 we get deg df1 ∧ df3 > deg df2 ∧ df3 , and finally
deg df1 ∧ df3 > deg df2 ∧ df3 > deg df1 ∧ df2 .
(3.12)
The general form of u2 being u2 = Jf1 + αf2 , f3 + βf2 K, we have
d(u2 ) = deg(df1 + α df2 ) ∧ (df3 + β df2 ) = deg df1 ∧ df3 ,
so that the expected result is exactly (3.12).
(3) We just saw that if t2 is any line not passing through [f2 ], then d(t2 ) =
deg df1 ∧ df3 = d(u2 ).
Now consider t2 passing through [f2 ] but not equal to v2 . Then t2 = [f2 , f3 + αf1 ]
for some α ∈ k and, using (3.12):
d(t2 ) = deg (df2 ∧ df3 − α df1 ∧ df2 ) = deg df2 ∧ df3 = deg(m2 ).
We obtain that v2 is the unique minimum of the function
t2 ∈ P̂2 (v3 ) ↦ d(t2 ).
(4) By (3.11) we have
deg(v1 ) + d(u2 ) = deg(v3 ∖ m2 ) + d(m2 ).
Since by assertion (1) we have d(m2 ) > deg(v3 ∖ m2 ), the result follows.
As an immediate consequence of Lemma 3.9(3) we get:
Corollary 3.13. Assume that v3 admits an elementary K-reduction, and that one
of the following holds:
(1) There exists a (non necessarily good) triangle u2 , v2 , w2 ∈ P̂2 (v3 ) such that
d(u2 ) > d(w2 ) > d(v2 );
(2) m2 is the minimal line in P2 (v3 ), and v2 is another line in P2 (v3 ) such that
d(m2 ) > d(v2 ).
Then v2 is the center of the K-reduction.
3.E. Proper K-reductions. In this section we list some properties of proper Kreductions, and introduce the concept of a normal K-reduction.
Proposition 3.14. Let u3 be a proper K-reduction of v3 , via w3 . Then (using
notation from Set-Up 3.5):
(1) g1 = f1 + ϕ1 (f2 , f3 ) with degvirt ϕ1 (f2 , f3 ) = deg ϕ1 (f2 , f3 ).
(2) If the pivotal simplex has Strong Pivotal Form ↺(s) with s ⩾ 5, then v3 is
a weak simple elementary reduction of w3 , with simple center m2 , v1 , and
deg v3 = deg w3 .
Proof. First observe that by Corollary 3.8 we know that the pivotal simplex v1 , w2 , w3
has Strong Pivotal Form ↺(s). In particular by Lemma 3.6 the minimal line m2 =
Jf2 , f3 K of w3 has no inner resonance.
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
23
(1) Assume by contradiction that degvirt ϕ1 (f2 , f3 ) > deg ϕ1 (f2 , f3 ).
By non resonance of Jf2 , f3 K we can apply Corollary 2.10(2) and Lemma 3.9(1) to
get the contradiction
deg ϕ1 (f2 , f3 ) > deg df2 ∧ df3 = d(m2 ) > deg g1 .
(2) We just established
deg g1 ⩾ deg ϕ1 (f2 , f3 ) = degvirt ϕ1 (f2 , f3 ).
Since v1 , w2 , w3 has Strong Pivotal Form ↺(s), we have deg g1 = sδ, deg f2 = 2δ
and deg f3 > (s − 2)δ, so that deg(f2 f3 ) > deg g1 . As soon as s ⩾ 5 we also have
2(s − 2) > s, hence deg f32 > deg g1 . This implies that if s ⩾ 5 then ϕ1 has the form
ϕ1 (f2 , f3 ) = af3 + Q(f2 ), as expected. Moreover since deg Q(f2 ) = 2rδ for some
r ⩾ 2 and since s is odd, we have deg g1 > degvirt ϕ1 (f2 , f3 ) = deg ϕ(f2 , f3 ) hence
deg f1 = deg g1 and deg v3 = deg w3 .
In view of the previous proposition, we introduce the following definition. We say
that u3 is a normal K-reduction of v3 in any of the two following situations:
● either u3 is an elementary K-reduction of v3 ;
● or u3 is a proper K-reduction of v3 via an auxiliary vertex w3 , and, denoting
by v1 the pivot of the reduction and m2 the minimal line in w3 , the vertex
v3 is not a weak simple elementary reduction of w3 with center m2 , v1 .
Given a proper K-reduction, Corollary 3.8 and Proposition 3.14(2) say that if the
reduction is normal then the pivotal simplex has Strong Pivotal Form ↺(3). We
now prove the converse, and give some estimations on the degrees involved.
Lemma 3.15. Assume that u3 is a proper K-reduction of v3 , via w3 , and that the
pivotal simplex has Strong Pivotal Form ↺(3). Then the reduction is normal, and
using representatives as from Set-Up 3.5, we have:
deg g1 = 3δ,
deg f2 = 2δ,
3
2δ
⩾ deg f3 > δ,
(3.16)
deg df1 ∧ df3 = δ + deg df2 ∧ df3 ⩾ 4δ + deg dg1 ∧ df2 ,
(3.17)
deg df1 ∧ df2 = deg f3 + deg df2 ∧ df3 .
(3.18)
Moreover we have the implications:
deg w3 > deg v3 Ô⇒ deg f3 = 32 δ,
deg f1 > 52 δ.
deg w3 = deg v3 Ô⇒ deg f1 = deg g1 = 3δ.
(3.19)
(3.20)
In any case we have
deg f1 > deg f2 > deg f3 ,
(3.21)
deg df1 ∧ df2 > deg df1 ∧ df3 > deg df2 ∧ df3 ,
(3.22)
2
m2 = Jf2 , f3 K is the minimal line of v3 , and for any other line ℓ2 ∈ P̂ (v3 ), we have
d(ℓ2 ) > d(m2 ).
(3.23)
Proof. The equalities deg g1 = 3δ and deg f2 = 2δ come from the fact that the pivotal
simplex has Strong Pivotal Form ↺(3). Property (↺4) gives deg f3 > δ, and
Proposition 3.14(1) says that degvirt ϕ1 (f2 , f3 ) = deg ϕ1 (f2 , f3 ). Hence apart from
f2 , f3 and f32 , any monomial in f2 , f3 has degree strictly bigger than g1 . So there
exists a, b, c ∈ k such that
w3 = Jg1 = f1 + af32 + bf2 + cf3 , f2 , f3 K.
Moreover, since w3 ≠ v3 , we have a ≠ 0, which implies 3δ = deg g1 ⩾ 2 deg f3 . This
proves (3.16), and the fact that the K-reduction is normal.
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
24
By Lemma 3.9(1) we have
deg df2 ∧ df3 = d(m2 ) ⩾ deg(w3 ∖ m2 ) + d(w2 ) = deg g1 + deg dg1 ∧ df2 .
(3.24)
This implies
deg g1 + deg df2 ∧ df3 > deg f3 + deg dg1 ∧ df2 .
By the Principle of Two Maxima 2.15, we get
deg g1 + deg df2 ∧ df3 = deg f2 + deg dg1 ∧ df3 .
(3.25)
Now dg1 ∧ df3 = df1 ∧ df3 + bdf2 ∧ df3 , and the previous equality implies deg dg1 ∧ df3 >
deg df2 ∧ df3 , so that
deg dg1 ∧ df3 = deg df1 ∧ df3 .
Now combining (3.24) and (3.25) we get the expected inequality (3.17):
deg df1 ∧ df3 = δ + deg df2 ∧ df3 ⩾ 4δ + deg dg1 ∧ df2 .
Observe that deg w3 > deg v3 is equivalent to deg g1 = deg f32 > deg f1 . So in this
situation deg f3 = 32 δ, and since by Lemma 2.2 we have deg f1 + deg f3 ⩾ deg df1 ∧ df3 ,
from (3.17) we also get deg f1 > 4δ − 32 δ = 25 δ. This proves (3.19), and (3.20) is
immediate. These two assertions imply that we always have deg f1 > 25 δ, so that we
get (3.21), and the minimal line m2 = Jf2 , f3 K of w3 also is the minimal line of v3 .
For the equality (3.18) we start again from g1 = f1 + af32 + bf2 + cf3 , which gives
dg1 ∧ df2 = df1 ∧ df2 − 2af3 df2 ∧ df3 − c df2 ∧ df3 .
Since (3.24) implies deg(f3 df2 ∧ df3 ) > deg dg1 ∧ df2 , the first two terms on the righthand side must have the same degree, which is the expected equality.
Now (3.17), (3.18) and the inequality deg f3 > δ from (3.16) immediately implies
(3.22).
Finally, for any line ℓ2 distinct from m2 = Jf2 , f3 K, we have
d(ℓ2 ) = deg(α df1 ∧ df2 + β df1 ∧ df3 + γ df2 ∧ df3 )
with (α, β) ≠ (0, 0), from which we obtain (3.23).
Now we can justify the terminology of “reduction”, as announced when we gave
the definition of a proper K-reduction:
Proposition 3.26. Let u3 be a proper K-reduction of a vertex v3 . Then
deg v3 > deg u3 .
Proof. We use the notation from Set-Up 3.5. Observe that if deg w3 = deg v3 , then
the proposition is obvious from (K0′ ).
Assume first that the reduction is not normal, that is, g1 = f1 + af3 + Q(f2 ). We
know by Corollary 3.8 that deg g1 = sδ, deg f2 = 2δ and sδ > deg f3 > (s − 2)δ, where
s ⩾ 3 is odd. An inequality deg g1 > deg f1 would imply sδ = deg(g1 ) = deg Q(f2 ) =
2rδ for some integer r (the degree of Q), a contradiction with s odd. Thus we obtain
deg g1 = deg f1 > deg(af3 + Q(f2 )), hence deg w3 = deg v3 and we are done.
Now assume we have a normal proper K-reduction, and that deg w3 > deg v3 . By
Proposition 3.14(2) (see also the discussion just before Lemme 3.15), we are in the
setting of Lemma 3.15. By condition (K4′ ) we have
δ + deg dg1 ∧ df2 > deg g3 .
Adding 3δ = deg g1 , and using (3.17) from Lemma 3.15, we get
deg df1 ∧ df3 ⩾ 4δ + deg dg1 ∧ df2 > deg g1 + deg g3 .
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
25
Finally adding deg f2 we get
deg v3 = deg f1 + deg f2 + deg f3
⩾ deg df1 ∧ df3 + deg f2
> deg g1 + deg g3 + deg f2
= deg u3 .
by Lemma 2.2
In the following result we prove that if a vertex admits a non-normal proper Kreduction, then it already admits an elementary (and therefore normal) K-reduction.
It follows that any vertex admitting a K-reduction admits a normal K-reduction.
Lemma 3.27 (Normalization of a K-reduction). Let u3 be a non-normal proper
K-reduction of v3 , via w3 . Then there exists u′3 such that
(1) u′3 ≬ v3 and u′3 ≬ u3 ;
(2) u′3 is an elementary (hence by definition normal) K-reduction of v3 .
w3 =Jg1 ,f2 ,f3 K
m2 =Jf2 ,f3 K
●
↺ ●
w2 =Jg1 ,f2 K
v3 =Jf1 =g1 −af3 −Q(f2 ),f2 ,f3 K
u3 =Jg1 ,f2 ,f3 +P (g1 ,f2 )K
↺
●
v2 =Jg1 −Q(f2 ),f2 K
v1 =[f2 ]
●
u2 =Jf2 ,f3 +P (g1 ,f2 )K
u′3 =Jg1 −Q(f2 ),f2 ,f3 +P (g1 ,f2 )K
Figure 5. Normalization of a K-reduction.
Proof. By Corollary 3.8 the pivotal simplex of the reduction has Strong Pivotal
Form ↺(s) for some odd s, and by Lemma 3.15 we have s ⩾ 5. Note also that
by Proposition 3.14 we have deg v3 = deg w3 , and v3 is a weak simple elementary
reduction of w3 with simple center m2 , v1 : see the upper-half of Figure 5, where we
use the notation of Set-Up 3.5.
By the Square Lemma 3.4, we get the existence of u′3 with u′3 ≬ v3 , u′3 ≬ u3 and
deg w3 > deg u′3 : see Figure 5. In particular deg v3 > deg u′3 , which is (K0).
Since v3 and w3 have the same 3-degrees, the vertices v2 and w2 also have the
same 2-degrees. So Properties (K1′ ) and (K2′ ) for the initial proper K-reduction
from v3 to u3 imply (K1) and (K2) for the elementary reduction from v3 to u′3 .
Finally m2 is the minimal line of v3 , and is distinct from v2 , which gives (K3),
and v2 = Jg1 − Q(f2 ), f2 K so that d(v2 ) = d(w2 ), which gives (K4).
Corollary 3.28. If v3 = Jf1 , f2 , f3 K admits a K-reduction, then one of the following
holds:
(1) Any line in P2 (v3 ) has no inner resonance;
(2) Up to permuting the fi , we have deg f1 = 2 deg f3 and 2 deg f1 = 3 deg f2 . In
particular with this numbering [f3 ] is the minimal vertex of v3 .
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
26
Proof. Any line in P2 (v3 ) has 2-degree (δ1 , δ2 ) with δi ∈ {deg f1 , deg f2 , deg f3 } for
i = 1, 2. So there exists a line in P2 (v3 ) with inner resonance if and only if there
exist two indexes i ≠ j in {1, 2, 3} such that deg fi ∈ N deg fj .
If the K-reduction is elementary, using the notation from Set-Up 3.5(1), we have
deg f1 = sδ, deg f2 = 2δ and sδ > deg f3 > (s − 2)δ for some odd s ⩾ 3. Moreover
an inner resonance in Jf2 , f3 K would be of the form deg f3 = (s − 1)δ = s−1
2 deg f2 for
s ⩾ 5, but this is impossible by Property (K2). So the only possible resonance is
between deg f1 and deg f3 , in the case s = 3, as stated in (2).
If the K-reduction is proper, we use the notation from Set-Up 3.5(2). Either
deg g1 = deg f1 and we are reduced to the previous case; or deg g1 > deg f1 and by
Lemma 3.27 we can assume that the K-reduction is normal (and proper, otherwise
again we are reduced to the previous case). Then by Lemma 3.15 we have 3δ >
deg f1 > 25 δ, deg f2 = 2δ, deg f3 = 32 δ, hence there is no relation of the form deg fi ∈
N deg fj for any i ≠ j and we are in case (1).
Remark 3.29. We shall see later in Corollary 5.3 that in fact Case (2) in the
previous corollary never happens.
3.F. Stability of K-reductions. Consider v3 a vertex that admits a normal Kreduction. In this section we want to show that most elementary reductions of v3
still admit a K-reduction. First we prove two lemmas that give some constraint on
the (weak) elementary reductions that such a vertex v3 can admit.
Lemma 3.30. Let u3 be a normal K-reduction of v3 , with pivot v1 . Let u2 be any
line in P2 (v3 ) not passing through v1 . Then v3 does not admit a weak elementary
reduction with center u2 .
Proof. We start with the notation v3 = Jf1 , f2 , f3 K from Set-Up 3.5(1) when u3 is an
elementary K-reduction of v3 , and with the Set-Up 3.5(2) to which we apply Lemma
3.15 when u3 is a normal proper K-reduction of v3 . It follows that m2 = Jf2 , f3 K
is in both cases the minimal line of v3 . We have u2 = Jh1 , h3 K where h1 = f1 + af2 ,
h3 = f3 + bf2 for some a, b ∈ k. Then [f2 , h3 ] is the minimal line m2 in P2 (v3 ),
however it is possible that [f2 , h3 ] is not a good representative of m2 . There are two
possibilities:
● either (h1 , f2 , h3 ) is still a good representative for v3 , and we have deg h3 =
deg f3 , deg f2 = deg(v3 ∖ u2 );
● or deg f2 = deg h3 > deg m1 where m1 = [f3 ] is the minimal point of v3 , and
deg f2 > deg(v3 ∖ u2 ) = deg f3 .
In both cases we have
deg h1 = topdeg v3 ,
deg f2 ⩾ deg(v3 ∖ u2 ) and
deg h3 ⩾ deg f3 .
Assume by contradiction that v3 admits a weak elementary reduction with center
u2 . By Lemma 3.3 there exists a non-affine polynomial P ∈ k[y, z] such that
deg(v3 ∖ u2 ) ⩾ deg P (h1 , h3 ).
On the other hand we know from Corollary 3.8 that the pivotal simplex of the
K-reduction has Strong Pivotal Form ↺(s) for some odd s ⩾ 3, hence
deg h3 ⩾ deg f3 > (s − 2)δ ⩾ δ which implies 2 deg h3 > 2δ = deg f2 .
In consequence, since deg h1 > deg h3 , we have
degvirt P (h1 , h3 ) ⩾ 2 deg h3 > deg f2 ⩾ deg(v3 ∖ u2 ),
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
so that
27
degvirt P (h1 , h3 ) > deg P (h1 , h3 ).
If u2 has no inner resonance, then we get a contradiction as follows, in both cases
of an elementary or a normal proper K-reduction:
deg(v3 ∖ u2 ) ⩾ deg P (h1 , h3 )
> deg dh1 ∧ dh3 = d(u2 )
> d(m2 )
> deg(v3 ∖ m2 )
by Corollary 2.10(2),
by Lemma 3.9(2) or (3.23),
by Lemma 3.9(1).
More precisely, in the case of a normal proper K-reduction the last inequality comes
from d(m2 ) > deg(w3 ∖ m2 ) ⩾ deg(v3 ∖ m2 ) by Lemma 3.9(1) and by (K5′ ).
Now consider the case where u2 = Jh1 , h3 K has inner resonance. By Corollary
3.28(2) we have deg h3 = min{deg f2 , deg f3 }, and since by assumption deg h3 ⩾ deg f3
we get deg h3 = deg f3 . Then Corollary 3.28(2) gives the two relations
1
2
2
3
deg(v3 ∖ m2 ) = 21 deg h1 = deg h3 ,
deg(v3 ∖ m2 ) = 32 deg h1 = deg f2 ⩾ deg(v3 ∖ u2 ).
(3.31)
In particular we have deg f1 > deg f2 > deg f3 , and b = 0, that is, h3 = f3 . We apply
Corollary 2.10(1) which gives
deg(v3 ∖ u2 ) ⩾ deg P (h1 , h3 ) ⩾ d(u2 ) − deg h3 ,
which we rewrite as
deg h3 + 2 deg(v3 ∖ u2 ) ⩾ deg(v3 ∖ u2 ) + d(u2 ).
(3.32)
deg(v3 ∖ u2 ) + d(u2 ) > 2 deg(v3 ∖ m2 ).
(3.33)
If u3 is an elementary K-reduction of v3 , and since in our situation deg(v3 ∖ u2 ) =
deg[f2 ], by Lemma 3.9(4) we have
If on the other hand u3 is a proper K-reduction of v3 via w3 , let us prove that
(3.33) still holds, by using Lemma 3.15. First note that deg(v3 ∖ u2 ) = deg f2 ,
deg(v3 ∖ m2 ) = deg f1 and
d(u2 ) = deg dh1 ∧ dh3 = deg d(f1 + af2 ) ∧ df3 = deg df1 ∧ df3 by (3.17).
Then, using (3.16) and (3.17) from Lemma 3.15, we get
deg f2 + deg df1 ∧ df3 > 2δ + 4δ = 2 deg g1 ⩾ 2 deg f1
as expected.
Adding the first equality of (3.31) to twice the second one, and combining with
(3.32) and (3.33), we get the contradiction
( 12 + 2.2
3 ) deg(v3 ∖ m2 ) > 2 deg(v3 ∖ m2 ).
Lemma 3.34. Let u3 be a normal proper K-reduction of v3 , with pivot v1 . Let
v2′ ≠ m2 be a line in P2 (v3 ) passing through v1 . If v3′ is a weak elementary reduction
of v3 with center v2′ , then this reduction is simple with center v2′ , v1 .
Proof. We use the notation from Set-Up 3.5(2), and set v2′ = Jh1 , f2 K with h1 =
f1 + af3 . By Lemma 3.15, h1 realizes the top degree of v3 . Then v3 = Jf1 , f2 , f3 K =
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
28
Jh1 , f2 , f3 K, and by Lemma 3.3 we have v3′ = Jh1 , f2 , f3 + P (h1 , f2 )K for some nonaffine polynomial P . We want to prove that P (h1 , f2 ) ∈ k[f2 ]. It is sufficient to
prove deg h1 > degvirt P (h1 , f2 ). Assume the contrary. Then
degvirt P (h1 , f2 ) ⩾ deg h1 > deg f3 ⩾ deg P (h1 , f2 ).
By Lemma 3.15, we have deg f1 > deg f2 > deg f3 , so that by Corollary 3.28 we have
deg h1 = deg f1 ∈/ N deg f2 . Thus we can apply Corollary 2.10(2) to get
deg f1 > deg P (h1 , f2 ) > deg dh1 ∧ df2 .
By (3.18) of Lemma 3.15 we get
deg dh1 ∧ df2 = deg(df1 ∧ df2 − adf2 ∧ df3 ) = deg df1 ∧ df2 > deg df2 ∧ df3 .
Then by (3.17) and (3.16) of Lemma 3.15 we have
deg df2 ∧ df3 ⩾ 3δ = deg g1 ⩾ deg f1 ,
hence the contradiction deg f1 > deg dh1 ∧ df2 ⩾ deg f1 .
Proposition 3.35 (Stability of a K-reduction). Let u3 be a normal K-reduction
of v3 , and v3′ a weak elementary reduction of v3 with center v2′ . Denote by m2 the
minimal line of v3 . If u3 is an elementary K-reduction, assume moreover that the
centers of v3′ ≬ v3 and u3 ≬ v3 are distinct. Then u3 is a K-reduction of v3′ , and
more precisely, we are in one of the following cases:
(1) u3 is an elementary K-reduction of v3 , and v2′ = m2 : then u3 is a (possibly
non-normal) proper K-reduction of v3′ , via v3 ;
(2) u3 is an elementary K-reduction of v3 , and v2′ ≠ m2 : then u3 is a (possibly non-normal) proper K-reduction of v3′ , via an auxiliary vertex w3′ that
satisfies deg w3′ = deg v3 ;
(3) u3 is a normal proper K-reduction of v3 via w3 , and v3′ = w3 : then u3 is an
elementary K-reduction of v3′ ;
(4) u3 is a normal proper K-reduction of v3 via w3 , and v2′ = m2 : then u3 also
is a normal proper K-reduction of v3′ via w3 .
Proof. First assume that u3 is an elementary K-reduction of v3 . We denote by
v1 = [f2 ], v2 = Jf1 , f2 K, v3 = Jf1 , f2 , f3 K the pivotal simplex of the K-reduction u3
(following Set-Up 3.5(1)). By Lemma 3.30, the line v2′ passes through v1 .
(1). If v2′ = m2 , since by assumption deg v3 ⩾ deg v3′ , we directly get that u3 is a
proper K-reduction of v3′ , via v3 .
(2). Now assume that v2′ ≠ m2 , so that v2′ = Jf1 + af3 , f2 K for some a ∈ k, and a ≠ 0
since we assume v2′ ≠ v2 . Then by Lemma 3.3 we can write
v3′ = Jf1 + af3 , f2 , f3 + P (f1 + af3 , f2 )K with deg f3 ⩾ deg P (f1 + af3 , f2 ).
If we can show that P depends only on f2 we are done: indeed then deg f3 ≠
deg P (f2 ), because m2 = Jf2 , f3 K has no inner resonance by Corollary 3.28, hence we
have deg f3 = deg(f3 + P (f2 )). It follows that u3 is a proper K-reduction of v3′ via
w3′ = Jf1 , f2 , f3 + P (f2 )K, where m′2 = Jf2 , f3 + P (f2 )K is the minimal line of w3′ (see
Figure 6, Case (2)).
To show that P depends only on f2 it is sufficient to show that deg f3 ⩾ degvirt P (f1 +
af3 , f2 ). By contradiction, assume that this is not the case. Then
degvirt P (f1 + af3 , f2 ) > deg f3 ⩾ deg P (f1 + af3 , f2 ).
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
29
v3 =Jf1 ,f2 ,f3 K
v3 =Jf1 ,f2 ,f3 K
v2 =Jf1 ,f2 K
u3
v2′ =Jf1 +af3 ,f2 K ●
v2′ =m2 =Jf2 ,f3 K
↺ ●
v2 =Jf1 ,f2 K
↺●
●
v1
v3′
u3
↺
v3′ =Jf1 +af3 ,f2 ,f3 +P (f2 )K
w3′ =Jf1 ,f2 ,f3 +P (f2 )K
v1 =[f2 ]
m′2 =Jf2 ,f3 +P (f2
Case (1)
●
)K
Case (2)
v3′ =w3
w3
v3′
v2′ =m2
v2′ =m2
w2
↺●
●
v3
↺●
●
u3
v3
u3
v1
v1
Case (3)
Case (4)
Figure 6. Stability of a K-reduction.
Since v2′ has the same 2-degree as v2 , it has no inner resonance by (K1), and by
Corollary 2.10(2) we get
deg P (f1 + af3 , f2 ) > deg(df1 ∧ df2 − adf2 ∧ df3 ).
By Lemma 3.9(1) we have deg df2 ∧ df3 > df1 ∧ df2 and deg df2 ∧ df3 > deg f3 , so finally
we obtain the contradiction
deg P (f1 + af3 , f2 ) > deg df2 ∧ df3 > deg f3 .
Now assume that u3 is a normal proper K-reduction of v3 , via an auxiliary vertex
w3 . Recall that the minimal line m2 of v3 is also the minimal line of the intermediate
vertex w3 (last assertion of Lemma 3.15).
(3). If v3′ = w3 , then by definition u3 is an elementary K-reduction of v3′ .
(4). If v3′ ≠ w3 , but v2′ = m2 , then the conclusion is also direct, because deg w3 ⩾
deg v3 ⩾ deg v3′ .
Finally we prove that the situation where v2′ ≠ m2 leads to a contradiction. By
Lemma 3.34 the reduction from v3 to v3′ is simple with center v2′ , v1 . But then
by Remark 3.2 (or directly from the proof of Lemma 3.34) we should have v2′ =
Jf1 + af3 , f2 K and v3′ = Jf1 + af3 , f2 , f3 + P (f2 )K. By Lemma 3.15 we have
deg f2 = 2δ > 23 δ ⩾ deg f3 ,
so deg P (f2 ) > deg f3 and we get a contradiction with deg v3 ⩾ deg v3′ .
4. Reducibility Theorem
In this section we state and prove the main result of this paper, that is, the
Reducibility Theorem 4.1.
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
30
4.A. Reduction paths. Given a vertex v3 with a choice of good triangle T , we call
elementary T -reduction any elementary reduction with center one of the three lines
of T .
We now define the notion of a reducible vertex in a recursive manner as follows:
We declare that the vertex [id] is reducible, where by Lemma 3.1 [id] is the
unique type 3 vertex realizing the minimal degree (1, 1, 1).
● Let µ > ν be two consecutive degrees, and assume that we have already defined
the subset of reducible vertices among type 3 vertices of degree at most ν. Then
we say that a vertex v3 with deg v3 = µ is reducible if for any good triangle T in
P2 (v3 ), there exists either a T -elementary reduction or a (proper or elementary)
K-reduction from v3 to u3 , with u3 reducible.
●
Let v3 , v3′ be vertices of type 3. A reduction path of length n ⩾ 0 from v3 to v3′
is a sequence of type 3 vertices v3 (0), v3 (1), . . . , v3 (n) such that:
●
●
●
v3 (0) = v3 and v3 (n) = v3′ ;
v3 (i) is reducible for all i = 0, . . . , n;
For all i = 0, . . . , n − 1, v3 (i + 1) is either an elementary reduction, or a
K-reduction, of v3 (i).
Observe that, by definition, a reducible vertex v3 admits a reduction path from
v3 to the vertex [id].
In the following sections we shall prove the main result:
Theorem 4.1 (Reducibility Theorem). Any vertex of type 3 in the complex C is
reducible.
Directly from the definition, this theorem has the following consequence: For any
vertex v3 ≠ [id] of type 3 in the complex C, and for any good triangle T in P2 (v3 ),
the vertex v3 admits either a T -elementary reduction or a (proper or elementary)
K-reduction.
We remark that this result immediately implies that Tame(A3 ) is a proper subgroup of Aut(A3 ):
Corollary 4.2. The Nagata’s automorphism
f = (x1 + 2x2 (x22 − x1 x3 ) + x3 (x22 − x1 x3 )2 , x2 + x3 (x22 − x1 x3 ), x3 )
is not tame.
Proof. Denote f = (f1 , f2 , f3 ) the components of f . Assume that f is tame. Let v3 =
Jf1 , f2 , f3 K be the associated vertex in C, and let T be the good triangle associated
with this representative. We have deg f1 = (2, 0, 3), deg f2 = (1, 0, 2) and deg f3 =
(0, 0, 1).
On the one hand, if f admits a K-reduction, by Corollary 3.8 one of the fi (the
pivot of the reduction) should have a degree of the form 2δ: this is not the case.
On the other hand, the degrees of the fi are pairwise Z-independent, so for any
distinct i, j ∈ {1, 2, 3} and any polynomial P we have degvirt P (fi , fj ) = deg P (fi , fj ).
This implies that if f admits an elementary T -reduction, then one of the deg fi should
be a N-combination of the other two. Again this is not the case.
Thus v3 is not reducible, a contradiction.
We shall prove Theorem 4.1 in §4.D. In the next two sections we establish preliminaries technical results.
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
31
4.B. Reduction of a strongly pivotal simplex. First we describe the set-up
that we shall use in this section.
Set-Up 4.3. Let v1 , v2 , v3 be a simplex in C with Strong Pivotal Form ↺(s) for
some odd s ⩾ 3. We choose some good representatives v1 = [f2 ], v2 = Jf1 , f2 K and
v3 = Jf1 , f2 , f3 K. Condition (↺1) means that
deg f1 = sδ,
deg f2 = 2δ.
By (↺3) we have deg f1 > deg f3 . we have
deg f3 ⩾ (s − 2)δ + deg df1 ∧ df2 .
Condition (↺2) is equivalent to the condition
deg f3 /∈ N deg f2 .
Since deg f3 > (s − 2)δ ⩾ δ, we also obtain
deg f32 > deg f2 and deg f2 ∈/ N deg f3 .
(4.4)
deg f1 /∈ N deg f3 except if s = 3 and deg f1 = 2 deg f3 .
(4.5)
In particular, as already noticed in Lemma 3.6, the minimal line m2 = Jf2 , f3 K of v3
has no inner resonance. Observe also that
Lemma 4.6. Assume Set-Up 4.3. Then v3 does not admit a normal proper Kreduction.
Proof. Assume v3 admits a normal proper K-reduction, via w3 . Then we get a
contradiction as follows:
d(m2 ) > deg(w3 ∖ m2 )
⩾ deg(v3 ∖ m2 )
> deg(v3 ∖ v2 ) ⩾ ∆(v2 )
> d(v2 ) > d(m2 )
by Lemma 3.9(1)
by (K5′ )
by (↺3) and (↺4)
by (3.23) in Lemma 3.15.
Lemma 4.7. Assume Set-Up 4.3, and deg f1 ≠ 2 deg f3 . Assume that v3 admits a
weak elementary reduction v3′ with center v2′ . Then v2′ passes through v1 .
Proof. By contradiction, assume that v2′ does not pass through v1 . Recall that
v1 = [f2 ] and v2 = Jf1 , f2 K with deg f1 > deg f2 . Up to replacing f1 by f1 + af2 for
some a ∈ k, we can assume v2 ∩ v2′ = [f1 ] while keeping all the properties stated in
Set-Up 4.3. Then let [h3 ] = [f3 + af2 ] be the intersection of v2′ with the minimal line
Jf2 , f3 K of P2 (v3 ). Then we have v2′ = Jf1 , h3 K, v3 = [f1 , f2 , h3 ], and by Lemma 3.3
v3′ = Jf1 , f2 + ϕ2 (f1 , h3 ), h3 K for some non-affine polynomial ϕ2 , with deg f1 > deg f2 ⩾
deg ϕ2 (f1 , h3 ). We have either deg f2 = deg h3 or deg f3 = deg h3 , hence in any case
by (4.4)
degvirt ϕ2 (f1 , h3 ) > deg ϕ2 (f1 , h3 ).
In particular ϕ2 (f1 , h3 ) /∈ k[h3 ], and then the inequality deg f1 > deg f2 implies
degvirt ϕ2 (f1 , h3 ) > deg f2 .
By Lemma 2.8 there exist coprime q > p such that
q deg h3 = p deg f1 = psδ.
Moreover if p = 1, we would have deg h3 = deg f3 and deg f1 = 2 deg f3 , in contradiction with our assumption. Hence we have q > p ⩾ 2.
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
32
Observe that even if (f2 , h3 ) is not a good representative of the minimal line in
P2 (v3 ), in any case we have deg h3 ⩾ deg(v3 ∖ v2 ) = deg f3 , and Property (↺4) gives
deg h3 ⩾ (s − 2)δ + deg df1 ∧ df2 .
Then Corollary 2.13 yields:
2δ = deg f2 ⩾ deg(f2 + ϕ2 (f1 , h3 )) ⩾ q deg h3 − deg df1 ∧ df2 − deg h3
⩾ q deg h3 − deg h3 + (s − 2)δ − deg h3 .
Multiplying by q and replacing q deg h3 = psδ we get:
0 ⩾ (pqs − 2ps + sq − 4q)δ,
hence
0 ⩾ ps(q − 2) + q(s − 4).
This implies s = 3, and we get the contradiction:
0 ⩾ 3pq − 6p − q = (3p − 1)(q − 2) − 2 ⩾ 5 − 2.
Lemma 4.8. Assume Set-Up 4.3. Assume that v3 admits an elementary reduction
v3′ with center m2 , the minimal line of v3 . Assume moreover that v3′ is reducible.
Then v3′ also admits an elementary reduction with center m2 .
Proof. By Lemma 3.3 we can write v3′ = Jf1′ , f2 , f3 K, where f1′ has the form
f1′ = f1 + ϕ1 (f2 , f3 )
for some non-affine polynomial ϕ1 , with deg f1 = deg ϕ1 (f2 , f3 ) > deg f1′ . Without
loss in generality we can assume that ϕ1 has no constant term. Moreover we can
also assume
(4.9)
deg f1′ ∈/ N deg f2 + N deg f3 ,
otherwise the result is immediate.
Working with the good triangle associated with the representative (f1′ , f2 , f3 ), we
want to prove that v3′ does not admit a K-reduction, nor an elementary reduction
with center Jf1′ , f2 K or Jf1′ , f3 K: Indeed since v3′ is reducible by assumption, the only
remaining possibility will be that v3′ admits an elementary reduction with center
m2 = Jf2 , f3 K, as expected. The proof is quite long, so we prove several facts along
the way. The first one is:
Fact 4.10. If degvirt ϕ1 (f2 , f3 ) > deg ϕ1 (f2 , f3 ) then Lemma 4.8 holds.
Proof. If degvirt ϕ1 (f2 , f3 ) > deg ϕ1 (f2 , f3 ) then by Lemma 2.8, there exist coprime
p, q such that q deg f2 = p deg f3 . Observe that (↺2) and (4.4) imply p, q ≠ 1. We
have
sδ = deg f1 > deg f1′ = deg(f1 + ϕ1 (f2 , f3 ))
> p deg f3 − deg df1 ∧ df2 − deg f3
⩾ p deg f3 − (deg f3 − (s − 2)δ)) − deg f3
= (p − 2) deg f3 − 2δ + sδ.
by Lemma 3.3
by Corollary 2.13
by (↺4)
Multiplying by p, recalling that deg f2 = 2δ, p deg f3 = q deg f2 and putting δ in factor
we get:
0 > 2q(p − 2) − 2p = (2p − 4)(q − 1) − 4 ⩾ 2p − 8.
(4.11)
It follows that 3 ⩾ p. Now we deduce p = 3. If p = 2, then deg f3 = qδ. Condition
(↺4) gives
sδ > deg f3 > (s − 2)δ,
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
33
so q = s − 1, which contradicts q coprime with 2.
Replacing p = 3 in the first inequality of (4.11) we get 6 > 2q, hence q = 2. We
obtain deg f3 = 34 δ, and the condition deg f3 > (s − 2)δ yields s = 3. Finally
deg f1 = 3δ,
deg f2 = 2δ,
deg f3 = 34 δ,
3δ > deg f1′ > 37 δ.
(4.12)
v3′
First we observe that these values are not compatible with
admitting a Kreduction. Indeed, by Corollary 3.8 an elementary K-reduction would imply 2 deg f1′ =
s′ deg fj for some odd integer s′ ⩾ 3 and j ∈ {2, 3}, and one checks from (4.12) that
′
there is no such relation. Indeed if s′ = 3, we have 2 deg f1 > 14
3 δ > 4δ = s deg f3 ,
′
′
and if s′ ⩾ 5 we have s′ deg f3 ⩾ 20
3 δ > 6δ > 2 deg f1 . Finally, for any s ⩾ 3 we have
′
′
s deg f2 ⩾ 6δ > 2 deg f1 .
Now if v3′ admits a normal proper K-reduction, then, noting that deg f1′ > deg f2 >
deg f3 , there should exist δ′ ∈ N3 such that all the conclusions of Lemma 3.15 hold,
with f1′ , f2 , f3 , δ′ instead of f1 , f2 , f3 , δ. In particular (3.19) gives deg f2 = 2δ′ , hence
δ′ = δ. Now since 3δ > deg f1′ , by (3.19) and (3.20) from Lemma 3.15 would imply
deg f3 = 23 δ, incompatible with deg f3 = 34 δ.
On the other hand, if v3′ admits an elementary reduction with center Jf1′ , fj K with
j = 2 or 3, then, denoting by k the integer such that {j, k} = {2, 3}, there would exist
a non-affine polynomial ϕ such that deg fk > deg(fk + ϕ(f1′ , fj )), and in particular
deg f1′ > deg fk = deg ϕ(f1′ , fj ). Since Jf2 , f3 K has no inner resonance this implies
ϕ(f1′ , fj ) ∈/ k[fj ], so that degvirt ϕ(f1′ , fj ) ⩾ deg f1′ , and finally degvirt ϕ(f1′ , fj ) >
deg ϕ(f1′ , fj ). Now by (4.9) we can apply Corollary 2.10(3) to get an odd integer
q ′ ⩾ 3 such that 2 deg f1′ = q ′ deg fj : again this is not compatible with (4.12).
From now on we assume degvirt ϕ1 (f2 , f3 ) = deg ϕ1 (f2 , f3 ).
Fact 4.13. deg f1 = 2 deg f3 .
Proof. By contradiction, assume deg f1 ≠ 2 deg f3 . Then deg f1 ∈/ N deg f3 by (4.5).
Moreover we know that deg f1 ∈/ N deg f2 and deg f2 + deg f3 > deg f1 . This is not
compatible with the equalities
deg f1 = deg ϕ1 (f2 , f3 ) = degvirt ϕ1 (f2 , f3 ).
We deduce from (4.5) and Fact 4.13 that s = 3, so that
deg f1 = 3δ,
deg f2 = 2δ,
deg f3 = 32 δ,
and there exist a, c, e ∈ k such that (recall that ϕ1 has no constant term):
ϕ1 (f2 , f3 ) = af32 + cf3 + ef2 with a ≠ 0.
Now come some technical facts.
Fact 4.15. deg df1 ∧ df3 = deg df1′ ∧ df3 = δ + deg df2 ∧ df3 .
Proof. Recall from Set-Up 4.3 that we have 32 δ = deg f3 > deg df1 ∧ df2 , so
3δ > deg f3 + deg df1 ∧ df2 .
Since deg f1 = 3δ we get
deg f1 + deg df2 ∧ df3 > deg f3 + deg df1 ∧ df2 .
By the Principle of Two Maxima 2.15 we have
deg f2 + deg df1 ∧ df3 = deg f1 + deg df2 ∧ df3 .
(4.14)
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
34
Passing deg f2 to the right-hand side we get one of the expected equalities
deg df1 ∧ df3 = deg f1 − deg f2 + deg df2 ∧ df3 = δ + deg df2 ∧ df3 .
From (4.14) we get
df1′ ∧ df3 = df1 ∧ df3 + e df2 ∧ df3 .
By the previous equality we obtain deg df1 ∧ df3 = deg df1′ ∧ df3 .
Fact 4.16. deg df1′ ∧ df2 = 32 δ + deg df2 ∧ df3 .
Proof. From (4.14) we get
df1′ ∧ df2 = df1 ∧ df2 + 2af3 df3 ∧ df2 + c df3 ∧ df2 .
By (↺4) we have deg f3 > deg df1 ∧ df2 , so 2af3 df3 ∧ df2 has strictly larger degree
than the two other terms of the right-hand side. Finally,
deg df1′ ∧ df2 = deg f3 df3 ∧ df2 = 23 δ + deg df2 ∧ df3 .
Fact 4.17. deg f1′ > δ.
Proof. Consider P = f1 + ay 2 + cy + ef2 ∈ k[f1 , f2 ][y]. We have
degvirt P (f3 ) = deg f1 > deg f1′ = deg P (f3 ).
On the other hand P ′ = 2ay + c, so that degvirt P ′ (f3 ) = deg P ′ (f3 ) = deg f3 . Thus
m(P, f3 ) = 1, and the Parachute Inequality 2.6 yields
deg f1′ = deg P (f3 ) ⩾ degvirt P (f3 ) + deg df1 ∧ df2 ∧ df3 − deg df1 ∧ df2 − deg f3
> deg f1 − deg df1 ∧ df2 − deg f3
= deg f3 − deg df1 ∧ df2 .
Recall that by (↺4) we have
deg f3 ⩾ deg f1 − deg f2 + deg df1 ∧ df2 = δ + deg df1 ∧ df2 .
Replacing in the previous inequality we get the result.
Fact 4.18. The vertices Jf1′ , f2 K and Jf1′ , f3 K do not have outer resonance in v3′ =
Jf1′ , f2 , f3 K.
Proof. We have (the last inequality is Fact 4.17):
deg f2 = 2δ,
deg f3 = 23 δ,
deg f1′ > δ.
Moreover these degrees are pairwise distinct, because (f1′ , f2 , f3 ) is a good representative of v3′ . This implies that deg f3 is not a N-combination of deg f1′ and deg f2 ,
and deg f2 is not a N-combination of deg f1′ and deg f3 .
Now we are ready to finish the proof of Lemma 4.8. Observe that by Lemma 3.27,
to prove that v3′ does not admit a K-reduction it is sufficient to exclude normal Kreductions. Recall also that from Facts 4.15 and 4.16 we have
deg f1′ ∧ f2 > deg f1′ ∧ f3 > deg f2 ∧ f3 .
(4.19)
Fact 4.20. v3′ does not admit an elementary K-reduction.
Proof. By contradiction, assume that v3′ admits an elementary K-reduction. By
(4.19) and Corollary 3.13(1), it follows that the center of the reduction is Jf2 , f3 K.
But then by Corollary 3.8 we should have 2 deg f2 = s′ deg f3 for some odd integer
s′ ⩾ 3, and this is not compatible with deg f2 = 2δ and deg f3 = 23 δ.
Fact 4.21. v3′ does not admit a normal proper K-reduction.
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
35
Proof. By contradiction, assume that v3′ admits a normal proper K-reduction u3 via
w3 . By comparing (4.19) with (3.22) from Lemma 3.15, we obtain that the center
v2′ of v3′ ≬ w3 is equal to v2′ = Jf2 , f3 K, that v2′ is the minimal line of v3′ , and that
deg w3 = 3δ + deg f2 + deg f3 = deg v3 . Since Jf2 , f3 K is also the minimal line of v3 , we
get that u3 is also a proper K-reduction of v3 via w3 . This is a contradiction with
Lemma 4.6.
Now we are left with the task of proving that v3′ does not admit an elementary
reduction with center Jf1′ , f2 K or Jf1′ , f3 K.
Fact 4.22. v3′ does not admit an elementary reduction with center Jf1′ , f2 K.
Proof. By contradiction, assume there exists ϕ3 (f1′ , f2 ) such that
deg ϕ3 (f1′ , f2 ) = deg f3 .
By Fact 4.18 we have degvirt ϕ3 (f1′ , f2 ) > deg ϕ3 (f1′ , f2 ), and on the other hand
Proposition 3.7(1) gives
3
2δ
= deg f3 = deg ϕ3 (f1′ , f2 ) > deg df1′ ∧ df2 .
This is a contradiction with Fact 4.16.
Fact 4.23. v3′ does not admit an elementary reduction with center Jf1′ , f3 K.
Proof. By contradiction, assume there exists ϕ2 (f1′ , f3 ) such that deg f2 > deg(f2 +
ϕ2 (f1′ , f3 )), which implies
deg ϕ2 (f1′ , f3 ) = deg f2 .
By Fact 4.18 we have degvirt ϕ2 (f1′ , f3 ) > deg ϕ2 (f1′ , f3 ). By Lemma 2.8 there exist
coprime p, q ∈ N∗ and γ ∈ N3 such that
deg f1′ = pγ,
deg f3 = qγ.
Corollary 2.10(1) then yields
2δ = deg f2 = deg ϕ2 (f1′ , f3 ) ⩾ pqγ + deg df1′ ∧ df3 − pγ − qγ.
By Fact 4.15 we know that deg df1′ ∧ df3 = δ + deg df2 ∧ df3 , so that
δ ⩾ deg df2 ∧ df3 + (pq − p − q)γ.
(4.24)
We know that qγ = deg f3 = 23 δ > δ, and by Fact 4.17 we have pγ = deg f1′ > δ, so
min{p, q} > pq − p − q.
By Fact 4.18 we know that p ≠ 1 and q ≠ 1. The only possibilities for the pair (p, q)
are then (2, 3) or (3, 2).
If (p, q) = (2, 3) then Fact 4.17 gives the contradiction
3δ = 2 deg f3 = 3 deg f1′ > 3δ.
If (p, q) = (3, 2), the equalities deg f1′ = 3γ and deg f3 = 2γ = 32 δ yield
γ = 43 δ,
deg f1′ = 94 δ,
pq − p − q = 1.
The inequality (4.24) becomes
δ
4
⩾ deg df2 ∧ df3 .
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
36
Corollary 2.13 then yields
deg(f2 + ϕ2 (f1′ , f3 )) > 2 deg f1′ − deg df2 ∧ df3 − deg f1′
⩾ ( 94 − 41 ) δ
= 2δ.
This is a contradiction with 2δ = deg f2 > deg(f2 + ϕ2 (f1′ , f3 )).
This finishes the proof of Lemma 4.8.
We now introduce an induction hypothesis that will be the corner-stone for the
proof of the Reducibility Theorem 4.1.
Induction Hypothesis 4.25 (for degrees ν, µ in N3 ). Let v3 be a reducible vertex
such that ν ⩾ deg v3 . Then any neighbor u3 of v3 with µ ⩾ deg u3 also is reducible.
We shall refer to this situation by saying that we can apply the (ν, µ)-Induction
Hypothesis 4.25 to v3 ≬ u3 .
Lemma 4.26. Let µ > ν ∈ N3 be two consecutive degrees, and assume Induction
Hypothesis 4.25 for degrees ν, ν. Let v3 be a reducible vertex distinct from [id], with
µ ⩾ deg v3 , and let T be a good triangle for v3 . Then there exists a reducible vertex
u3 such that
● either u3 is an optimal elementary T -reduction or an optimal elementary
K-reduction of v3 ;
● or u3 is a normal proper K-reduction of v3 .
Proof. Let v3 (1) be the first step of a reduction path from v3 , with respect to the
triangle T . By this we mean that v3 (1) is a reducible vertex that is either a T elementary reduction of v3 , or a K-reduction of v3 .
First assume that v3 (1) is an elementary reduction of v3 . If this reduction is
optimal, we can take u3 = v3 (1) and we are done. If the reduction is non-optimal,
denote by v2 the center of v3 ≬ v3 (1). Then let u3 be an optimal elementary
reduction of v3 with the same center v2 . Since µ > ν are two consecutive degrees,
the inequalities
µ ⩾ deg v3 ,
deg v3 > deg v3 (1)
and
deg v3 > deg u3
imply
ν ⩾ deg v3 (1) and ν ⩾ deg u3 .
By the (ν, ν)-Induction Hypothesis 4.25 applied to v3 (1) ≬ u3 we get that u3 is
reducible, as expected. Moreover since u3 is an elementary reduction with center v2 ,
by construction u3 is an optimal elementary T -reduction or an optimal elementary
K-reduction of v3 .
Now assume that v3 (1) is a proper K-reduction of v3 . If this reduction is normal,
we are done. Otherwise, by Lemma 3.27, there exists u′3 an elementary K-reduction
of v3 , such that u′3 ≬ v3 (1). By the (ν, ν)-Induction Hypothesis 4.25 applied to
v3 (1) ≬ u′3 we get that u′3 is reducible. So we can take u′3 as the first step of a
reduction path from v3 , and we are reduced to the first case of the proof.
Proposition 4.27. Let µ > ν ∈ N3 be two consecutive degrees, and assume Induction
Hypothesis 4.25 for degrees ν, ν. Let v3 be a vertex that is part of a simplex v1 , v2 , v3
with Strong Pivotal Form ↺(s) for some odd s ⩾ 3. Let T be any good triangle
compatible with the simplex v1 , v2 , v3 . Assume that v3 is reducible, and that µ ⩾
deg v3 . Then
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
37
(1) There exists a reduction path from v3 to [id] that starts with an elementary
T -reduction or an elementary K-reduction;
(2) Any elementary T -reduction or elementary K-reduction of v3 admits v2 as
a center;
(3) Any optimal elementary T -reduction of v3 is an elementary K-reduction;
(4) There exists an elementary K-reduction u3 of v3 with center v2 , such that
u3 is reducible.
Proof. Let v3 (1) be the first step of a reduction path from v3 to [id], with respect to
the choice of good triangle T . By Lemma 4.26 we can assume that if this first step
is a K-reduction, then it is a normal K-reduction. Moreover we know from Lemma
4.6 that v3 does not admit any normal proper K-reduction. This gives (1).
We write v3 = Jf1 , f2 , f3 K as in Set-Up 4.3, such that the triangle T corresponds
to the choice of representative (f1 , f2 , f3 ). The remaining possibilities for the first
step v3 (1) of the reduction path are:
(i) An elementary T -reduction with center Jf2 , f3 K;
(ii) An elementary T -reduction with center Jf1 , f3 K;
(iii) An elementary K-reduction;
(iv) An elementary T -reduction with center v2 = Jf1 , f2 K.
In case (i), by Lemma 4.26 we can moreover assume that the elementary reduction
is optimal. Then Lemma 4.8 gives a contradiction.
In case (ii), Lemma 4.7 implies that deg f1 = 2 deg f3 , so there exists a ∈ k such
that w3 ∶= [f1 + af32 , f2 , f3 ] satisfies deg v3 > deg w3 . By the Square Lemma 3.4, there
exists u3 such that u3 ≬ w3 , u3 ≬ v3 (1) and deg v3 > deg u3 . By the (ν, ν)-Induction
Hypothesis 4.25 applied successively to v3 (1) ≬ u3 and u3 ≬ w3 , we obtain that w3
is reducible, and so could be chosen as the first step of a reduction path. We are
reduced to case (i), which leads to a contradiction.
In case (iii), by Lemma 3.9(1) we have
d(m2 ) > topdeg v3 .
Moreover by (↺4) we have
topdeg v3 > deg(v3 ∖ v2 ) ⩾ ∆(v2 ) > d(v2 ).
By Corollary 3.13(2) we conclude that v2 is the center of the K-reduction, hence we
also are in case (iv), which gives (2).
Now assume that u3 is an optimal elementary T -reduction of v3 . By (2), we
know that the center of this reduction is v2 , and by the (ν, ν)-Induction Hypothesis
4.25 applied to v3 (1) ≬v2 u3 , we get that u3 is reducible. We want to prove that
∆(v2 ) > deg(u3 ∖ v2 ), that is, Property (K4), which will imply that u3 is a Kreduction of v3 . By contradiction, assume deg(u3 ∖ v2 ) ⩾ ∆(v2 ), which is condition
(↺4) for the simplex v1 , v2 , u3 . Moreover conditions (↺1) and (↺3) for the simplex
v1 , v2 , u3 directly follow from the analogous conditions for the simplex v1 , v2 , v3 , and
condition (↺2) follows from the optimality of the reduction u3 . Thus v1 , v2 , u3 has
Strong Pivotal Form ↺(s), and from assertions (1) and (2) we conclude that u3
admits an elementary reduction with center v2 . This contradicts the optimality of
the reduction from v3 to u3 , and so we obtain (3).
Finally to prove (4), consider a first step of a reduction path from v3 to [id],
with respect to the good triangle T , as given by Lemma 4.26. By Lemma 4.6 this
first step is not a normal proper K-reduction, so that it is an optimal elementary
reduction, hence by (3) this is an elementary K-reduction, as expected.
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
38
Corollary 4.28. Let µ > ν ∈ N3 be two consecutive degrees, and assume Induction
Hypothesis 4.25 for degrees ν, ν. Let v3 be a reducible vertex with µ ⩾ deg v3 , and assume that u3 is an optimal elementary reduction of v3 with pivotal simplex v1 , v2 , v3 .
If v2 has no inner resonance, and no outer resonance in v3 , then either v2 is the
minimal line of v3 , or u3 is an elementary K-reduction of v3 .
Proof. Assume v2 is not the minimal line of v3 . By Proposition 3.7(2), the simplex
v1 , v2 , v3 has Strong Pivotal Form. Let T be a good triangle compatible with the
simplex v1 , v2 , v3 . In particular u3 is an optimal elementary T -reduction of v3 , so
we can conclude by Proposition 4.27(3).
4.C. Vertex with two low degree neighbors.
Set-Up 4.29. Let µ > ν ∈ N3 be two consecutive degrees, and assume the Induction
Hypothesis 4.25 for degrees ν, µ. Let v3 , v3′ , v3′′ be vertices such that
●
●
●
●
µ ⩾ deg v3 , deg v3 > deg v3′ (hence ν ⩾ deg v3′ ), deg v3 ⩾ deg v3′′ ;
v3′ ≬v2′ v3 and v3′′ ≬v2′′ v3 with v2′ ≠ v2′′ ;
v3′ is reducible (hence v3 also is by the (ν, µ)-Induction Hypothesis);
v2′ is minimal, in the sense that if u3 is an elementary reduction of v3 with
center u2 , which is the first step of a reduction path, then deg u2 ⩾ deg v2′ .
We denote by v1 = [f2 ] the intersection point of the lines v2′ and v2′′ . We fix choices
of f1 , f3 such that v2′ = Jf1 , f2 K and v2′′ = Jf2 , f3 K. Observe that it is possible that
deg f1 = deg f3 , and in this case (f1 , f2 , f3 ) is not a good representative of v3 . In any
case by Lemma 3.3 there exist some non-affine polynomials in two variables P1 , P3
such that (see Figure 7):
v3 = [f1 , f2 , f3 ];
v3′ = Jf1 , f2 , f3 + P3 (f1 , f2 )K;
v3′′ = Jf1 + P1 (f2 , f3 ), f2 , f3 K.
v3 =[f1 ,f2 ,f3 ]
v2′′ =Jf2 ,f3 K
●
●
v3′′ =Jf1 +P1 (f2 ,f3 ),f2 ,f3 K
v2′ =Jf1 ,f2 K
v3′ =Jf1 ,f2 ,f3 +P3 (f1 ,f2 )K
v1 =[f2 ]
Figure 7. Set-Up 4.29.
In this section we shall prove:
Proposition 4.30. Assume Set-Up 4.29. Then v3′′ is reducible.
We divide the proof in several lemmas. The proposition will be a direct consequence of Lemmas 4.32, 4.33, 4.34 and 4.35. We start with a consequence from the
minimality of v2′ .
Lemma 4.31. Assume Set-Up 4.29. If v2′ has inner resonance, then v2′ is the minimal line of v3 .
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
39
Proof. Assume first that there exists a ∈ k and r ⩾ 2 such that deg f2 > deg(f2 +af1r ).
Then consider the vertex w3 = [f1 , f2 + af1r , f3 ], which satisfies deg v3 > deg w3 .
Assume by contradiction that v2′ is not the minimal line of v3 . Then we have
deg f2 > deg f3 , hence m2 = [f1 , f3 ] is the minimal line of v3 . By the Square Lemma
3.4, we find u3 such that u3 ≬ w3 , u3 ≬ v3′ and deg v3 > deg u3 . By applying the
(ν, µ)-Induction Hypothesis 4.25 successively to v3′ ≬ u3 and u3 ≬ w3 , we find that
w3 is reducible. So we can take w3 as the first step of a reduction path from v3 ,
which contradicts the minimality of v2′ .
Now assume that there exist a ∈ k∗ and r ⩾ 2 such that deg f1 > deg(f1 + af2r ). If
′
v2 is not the minimal line, we have deg f1 ⩾ deg f3 .
If deg f1 = deg f3 , there exists b ∈ k∗ such that deg f1 > deg(f1 + bf3 ). Then
[f2 , f1 + bf3 ] is the minimal line of v3 , and we consider w3′ = [f1 + af2r , f2 , f1 + bf3 ],
which satisfies deg v3 > deg w3′ . As above, by the Square Lemma 3.4 and the (ν, µ)Induction Hypothesis 4.25, we obtain that w3′ can be chosen to be the first step of
a reduction path from v3 . This contradicts the minimality of v2′ .
If deg f1 > deg f3 , v2′′ = Jf2 , f3 K is the minimal line of v3 . We consider w3′′ =
[f1 + af2r , f2 , f3 ] which satisfies deg v3 > deg w3′′ . As before w3′′ can be chosen to be
the first step of a reduction path from v3 , with center v2′′ : again this contradicts the
minimality of v2′ .
Lemma 4.32. Assume Set-Up 4.29. If v2′ is not the minimal line in v3 , and has
outer resonance in v3 , then v3′′ is reducible.
Proof. Since v2′ = Jf1 , f2 K is not the minimal line in v3 , one of the degrees deg f1 or
deg f2 must realize the top degree of v3 .
First consider the case deg f2 = topdeg v3 . In this case, we have deg f1 ≠ deg f3 .
Indeed, assuming by contradiction that deg f1 = deg f3 , there would exist α ∈ k∗ such
that f3′ = f3 − αf1 satisfies deg f3 > deg f3′ . Then, we would have deg f2 > deg f1 >
deg f3′ , and the line v2′ = Jf1 , f2 K could not have outer resonance in v3 . Therefore,
(f1 , f2 , f3 ) is a good representative of v3 , and the assumption on outer resonance
means that deg f3 = deg f1r for some r ⩾ 2. Recall that by Lemma 3.3, the nonaffine polynomial P1 ∈ k[y, z] associated with the weak elementary reduction v3′′
of v3 satisfies deg f1 ⩾ P1 (f2 , f3 ). We have deg f2 /∈ N deg f3 , otherwise we would
have deg f2 ∈ N deg f1 , that is, v2′ would have inner resonance, and by Lemma 4.31
this would contradict v2′ ≠ m2 . Corollary 2.10(3) then gives 2 deg f2 = 3 deg f3 ,
and deg(v3 ∖ v2′′ ) ⩾ ∆(v2′′ ). But then the existence of δ ∈ N such that deg f2 =
3δ, deg f3 = 2δ and deg f1 > δ is not compatible with the existence of r ⩾ 2 such that
deg f3 = r deg f1 : contradiction.
Now consider the case deg f1 = topdeg v3 . In particular deg f1 > deg f2 and v1 =
[f2 ] is on the minimal line of v3 . Now we distinguish two subcases (see Figure 8):
● Case deg f1 > deg f3 . Then there exist a ∈ k and r ⩾ 2 such that deg f3 >
deg(f3 + af2r ), and v2′′ is the minimal line of v3 . We apply the Square Lemma 3.4 to
get u3 a common neighbor of v3′′ and w3′ = [f1 , f2 , f3 + af2r ] satisfying deg v3 > deg u3 ,
and then we conclude by the (ν, µ)-Induction Hypothesis 4.25 applied successively
to w3′ ≬ u3 and u3 ≬ v3′′ .
● Case deg f1 = deg f3 . Then there exists b ∈ k∗ such that deg f1 > deg(bf1 + f3 ).
As in the first case of the proof, the assumption on outer resonance implies that
(f1 , f2 , f3 + bf1 ) is a good representative of v3 , and there exists r ⩾ 2 such that
deg(bf1 +f3 ) = deg f2r . Then there exists a ∈ k such that u′3 = [f1 , f2 , bf1 +f3 +af2r ] and
u′′3 = [bf1 + f3 + af2r , f2 , f3 ] are (simple) elementary reductions of v3 with respective
centers v2′ and v2′′ . Moreover u′3 and u′′3 are neighbors, with center [bf1 + f3 + af2r , f2 ].
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
v3
40
v3
v3′
v3′′
●
●
●
●
●
v3′′
w3′
●
u′′
3
v3′
u′3
●
u3
deg f1 > deg f3
deg f1 = deg f3
Figure 8. Lemma 4.32, case deg f1 = topdeg v3 .
Again we conclude by applying the (ν, µ)-Induction Hypothesis 4.25 to v3′ ≬ u′3 ,
u′3 ≬ u′′3 and u′′3 ≬ v3′′ successively.
We conclude the case where v2′ is not equal to the minimal line m2 of v3 with the
following:
Lemma 4.33. Assume Set-Up 4.29, v2′ ≠ m2 (where m2 is the minimal line in v3 ),
and has no outer resonance in v3 . Then there exists u3 an elementary K-reduction
of v3 with center v2′ . Moreover, v3′′ is reducible, and there exists a reduction path
starting with a proper K-reduction from v3′′ to u3 .
Proof. By Lemma 4.31, we know that v2′ has no inner resonance. Then Corollary 4.28
says that any optimal elementary reduction u3 of v3 with center v2′ is an elementary
K-reduction. By the (ν, µ)-Induction Hypothesis 4.25 applied to v3′ ≬ u3 , we get
that u3 is reducible. The last assertion follows by the Stability of K-reductions 3.35,
Case (1) or (2).
Now we treat the situation where v2′ = m2 is the minimal line of v3 , and first we
identify some cases that we can handle with the Square Lemma 3.4.
Lemma 4.34. Assume Set-Up 4.29, and v2′ = m2 . In the following cases, v3′′ is
reducible:
(1) v3 admits a simple elementary reduction with simple center v2′ , v1 ;
(2) v3 admits a simple elementary reduction with simple center v2′′ , v1 ;
(3) v3′′ is a simple weak elementary reduction of v3 with simple center v2′′ , v1 .
v3
v3′′
●
v2′
●
v3′
v3′′
w3′
w3′′
v3
v2′′
●
●
v3′
v3′′
●
v1
v1
●
v3
v2′′
●
u3
Case (1)
●
●
u3
Case (2)
●
v3′
v1
●
●
u3
Case (3)
Figure 9. Lemma 4.34.
Proof. Since m2 = Jf1 , f2 K, we have v3 = Jf1 , f2 , f3 K with deg f3 = topdeg v3 .
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
41
(1) Denote by w3′ a simple elementary reduction of v3 with simple center v2′ , v1 .
By the Square Lemma 3.4, there exists u3 a neighbor of both w3′ and v3′′ such that
deg v3 > deg u3 (see Figure 9). Then we conclude by applying the (ν, µ)-Induction
Hypothesis 4.25 successively to v3′ ≬ w3′ , w3′ ≬ u3 and u3 ≬ v3′′ .
(2) Denote by w3′′ a simple elementary reduction of v3 with simple center v2′′ , v1 .
By the Square Lemma 3.4, there exists u3 a neighbor of both w3′′ and v3′ such that
deg v3 > deg u3 . Then we conclude by applying the (ν, µ)-Induction Hypothesis 4.25
successively to v3′ ≬ u3 , u3 ≬ w3′′ and w3′′ ≬ v3′′ .
(3) By the Square Lemma 3.4, there exists u3 a neighbor of both v3′′ and v3′ such
that deg v3 > deg u3 . Again the (ν, µ)-Induction Hypothesis 4.25 applied successively
to v3′ ≬ u3 and u3 ≬ v3′′ yields that v3′′ is reducible.
The last logical step to finish the proof of Proposition 4.30 is the following lemma.
However, we should mention that we will be able to prove a posteriori that this case
never happens: see Corollary 5.4.
Lemma 4.35. Assume Set-Up 4.29, v2′ = m2 , and that we are not in one of the
cases covered by Lemma 4.34. Then there exists u′′3 an elementary K-reduction of
v3 with center v2′′ such that u′′3 is reducible. In particular, by the (ν, µ)-Induction
Hypothesis 4.25, v3′′ is reducible.
Proof. It is sufficient to check that the simplex v1 , v2′′ , v3 has Strong Pivotal Form
↺(s) for some odd s ⩾ 3: Indeed then one can apply Proposition 4.27(4) to get the
result.
On the one hand deg f3 > deg f1 ⩾ deg P1 (f2 , f3 ), and on the other hand since we
are not in the situation of Lemma 4.34(3) we have P1 (f2 , f3 ) ∈/ k[f2 ], so that
degvirt P1 (f2 , f3 ) > deg P1 (f2 , f3 ).
We also have deg f3 ∈/ N deg f2 since otherwise we could apply Lemma 4.34(1). So
we are in the hypotheses of Corollary 2.10(3), and there exist a degree δ ∈ N3 and
an odd integer s ⩾ 3 such that deg f2 = 2δ, deg f3 = sδ and
sδ > deg P1 (f2 , f3 ) ⩾ ∆(f3 , f2 ).
It remains to check (↺2): if v2′′ = Jf2 , f3 K had outer resonance in v3 then we would
have deg f1 ∈ N deg f2 , and we could apply Lemma 4.34(2), contrary to our assumption.
4.D. Proof of the Reducibility Theorem. Clearly Theorem 4.1 is a corollary of
Proposition 4.36. If a vertex v3 of type 3 in the complex C is reducible, then any
neighbor of v3 also is reducible.
Proof. We plan to prove Proposition 4.36 by induction on degree: we need to prove
that for any ν ∈ N3 , the Induction Hypothesis 4.25 holds for degrees ν, ν. Clearly
when ν = (1, 1, 1) this is true (because empty!).
Let µ > ν be two consecutive degrees in N3 . It is sufficient to prove the two
following facts.
Fact 4.37. Assume the Induction Hypothesis 4.25 for degrees ν, ν. Then it also
holds for degrees ν, µ.
Fact 4.38. Assume the Induction Hypothesis 4.25 for degrees ν, µ. Then it also
holds for degrees µ, µ.
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
42
To prove Fact 4.37, consider v3′ a reducible vertex with ν ⩾ deg v3′ , and let v3 be
a neighbor of v3′ , with center v2′ , and with deg v3 = µ (otherwise there is nothing to
prove). We want to prove that v3 is reducible.
If v2′ is the minimal line of v3 , then v3′ is a T -reduction of v3 for any good triangle
T and we are done.
If v2′ is not the minimal line of v3 , and has no inner or outer resonance in v3 , then
by Corollary 4.28 any optimal reduction u3 of v3 with respect to the center v2′ is
an elementary K-reduction of v3 . Since u3 is a neighbor of v3′ and ν ⩾ deg u3 , we
conclude by the (ν, ν)-Induction Hypothesis 4.25 that u3 is reducible. Hence v3 also
is reducible, with first step of a reduction path the K-reduction to u3 .
Finally assume that v2′ has resonance, and that v2′ is not the minimal line of v3 .
By Lemma 3.3 we can write
v3 = Jf1 , f2 , f3 K,
v3′ = Jf1 , f2 , g3 K,
v2′ = Jf1 , f2 K,
with deg f1 > max{deg f3 , deg f2 } and g3 = f3 + P (f1 , f2 ) for some polynomial P .
If v2′ has inner resonance, then deg f1 = r deg f2 for some r ⩾ 2. There exists a ∈ k
such that v3′′ = [f1 + af2r , f2 , f3 ] is a simple elementary reduction of v3 with center
Jf2 , f3 K, which is the minimal line of v3 , hence belongs to any good triangle T . Then
we can apply the Square Lemma 3.4 to get u3 with ν ⩾ deg u3 and u3 ≬ v3′ , u3 ≬ v3′′ .
We conclude by the (ν, ν)-Induction Hypothesis 4.25, applied successively to v3′ ≬ u3
and u3 ≬ v3′′ , that v3′′ is reducible, hence v3 also is.
If v2′ has no inner resonance, but has outer resonance in v3 , then deg f1 > deg f3 >
deg f2 and deg f3 = r deg f2 for some r ⩾ 2. There exists Q(f2 ) such that deg f3 >
deg(f3 + Q(f2 )) and deg(f3 + Q(f2 )) ∈/ N deg f2 . Let T be a good triangle of v3 . One
of the lines in T has the form u2 = Jf1 + af3 , f2 K. Set u3 = Jf1 + af3 , f2 , f3 + Q(f2 )K,
which is an elementary T -reduction of v3 with center u2 , and a neighbor of w3 =
Jf1 , f2 , f3 + Q(f2 )K. By the (ν, ν)-Induction Hypothesis 4.25 applied successively to
v3′ ≬ w3 and w3 ≬ u3 , we obtain that u3 is reducible, hence v3 also is.
To prove Fact 4.38, consider v3 a reducible vertex with deg v3 = µ, and let v3′′ be
a neighbor of v3 , with center v2′′ , such that
deg v3 ⩾ deg v3′′ .
We want to prove that v3′′ is reducible.
First assume that v3 admits a reduction path such that the first step v3′ is an
elementary reduction, with center v2′ . Moreover we assume that v2′ has minimal
degree between all possible such first center of a reduction path. If v2′ = v2′′ , then v3′′
is a neighbor of v3′ and by the (ν, µ)-Induction Hypothesis 4.25 we are done. If on
the contrary v2′ and v2′′ are two different lines in P2 (v3 ), we are in the situation of
Set-Up 4.29. Then we conclude by Proposition 4.30.
Finally assume that v3 admits a reduction path such that the first step is a
proper K-reduction v3′ . If this proper K-reduction is normal, then by Stability of
K-reduction 3.35, Case (3) or (4), we obtain that v3′ is also a K-reduction of v3′′ ,
and we are done. On the other hand if v3′ is a non-normal proper K-reduction of v3 ,
then by Lemma 3.27 we get the existence of a vertex u3 that is both an elementary
K-reduction of v3 and a neighbor of v3′ . By applying the (ν, µ)-Induction Hypothesis
4.25 to v3′ ≬ u3 we get that u3 is reducible, and we are reduced to the previous case
of the proof.
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
43
5. Simple connectedness
In this section we prove that the complex C is simply connected, which amounts
to saying that the group Tame(A3 ) is the amalgamated product of three subgroups
along their pairwise intersections.
5.A. Consequences of the Reducibility Theorem. Now that the Reducibility
Theorem 4.1 is proved, all previous results that were dependent of a reducibility
assumption become stronger. This is the case in particular for:
● The Induction Hypothesis 4.25, which is always true;
● Lemma 4.8, which now implies that if v3 is part of a simplex with Strong
Pivotal Form, then v3 does not admit an elementary reduction with center
m2 the minimal line of v3 : see the proof of Proposition 5.1 below;
● Proposition 4.27 and Corollary 4.28;
● Set-Up 4.29, hence also all results in §4.C.
In particular we single out the following striking consequences of the Reducibility
Theorem 4.1.
Proposition 5.1. Let u3 be an elementary K-reduction of v3 , with center v2 . Then
v3 does not admit any elementary reduction with center distinct from v2 .
Proof. By contradiction, assume that v3′ is an elementary reduction of v3 , with center
v2′ distinct from v2 . Then by Stability of K-reduction 3.35, Case (1) or (2), there
exists w3′ with deg w3′ = deg v3 such that u3 is a proper K-reduction of v3′ via w3′ . In
particular, v3′ is an elementary reduction of w3′ with center m2 , the minimal line of
w3′ . Moreover, by Corollary 3.8, the pivotal simplex of this proper K-reduction has
Strong Pivotal Form. Now consider v3′′ an optimal elementary reduction of w3′ with
center m2 : Lemma 4.8 gives a contradiction.
Corollary 5.2. Let u3 be a proper K-reduction of v3 , via w3 . Then deg w3 = deg v3 .
Proof. By Proposition 5.1 we cannot have deg w3 > deg v3 .
Corollary 5.3. Let v3 admitting a K-reduction, then any line u2 in P (v3 ) has no
inner resonance. In other words, Case (2) in Corollary 3.28 never happens.
2
Proof. We use the notation v3 = Jf1 , f2 , f3 K from Corollary 3.28, which says that if
there exists a line in v3 with inner resonance, then deg f1 = 2 deg f3 and deg f1 =
3
2
2 deg f2 . But in this case we would have an elementary reduction [f1 + af3 , f2 , f3 ]
of v3 with center the minimal line Jf2 , f3 K ≠ v2 , in contradiction with Proposition
5.1
We also obtain that the situation of Lemma 4.35 never happens:
Corollary 5.4. Assume Set-Up 4.29, and v2′ = m2 . Then we are in one of the cases
covered by Lemma 4.34.
Proof. Otherwise by Lemma 4.35 there would exist an elementary K-reduction of
v3 with center v2′′ , in addition to the elementary reduction with center v2′ : This is
not compatible with Proposition 5.1.
5.B. Local homotopies. To prove the simple connectedness of C, the following
terminology will be convenient.
We call combinatorial path a sequence of vertices v3 (i), i = 0, . . . , n, such
that for all i = 0, . . . , n − 1, v3 (i) ≬ v3 (i + 1). Denoting by v2 (i) the center of
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
44
v3 (i) ≬ v3 (i + 1), we think of such a sequence as equivalent to a path γ∶ [0, 2n] → C
where for each i = 0, . . . , n − 1, the interval [2i, 2i + 2] is mapped isometrically onto
the union of the two edges v3 (i), v2 (i) and v2 (i), v3 (i + 1). In particular we have
these parameterizations in mind when we say that two such combinatorial paths are
homotopic in C. We say that a combinatorial path is locally geodesic if for all i,
v2 (i) ≠ v2 (i+1) (and v3 (i) ≠ v3 (i+1), but this is already contained in the definition of
v3 (i) ≬ v3 (i + 1)). Observe that starting from any combinatorial path, by removing
some vertices we can always obtain a locally geodesic one. If v3 (0) = v3 (n) we say
that the path is a combinatorial loop with base point v3 (0). If v3 (0) = [id], the
maximal vertex of such a loop is defined as the vertex v3 (i0 ) that realizes the
maximum max deg v3 (i), with i0 maximal. In particular, we have
deg v3 (i0 ) > deg v3 (i0 + 1)
and
deg v3 (i0 ) ⩾ deg v3 (i0 − 1).
Lemma 5.5. Let v3 (i), i = 0, . . . , n, be a locally geodesic loop with base point at [id],
and let v3 (i0 ) be the maximal vertex. Then the combinatorial path v3 (i0 − 1), v3 (i0 ),
v3 (i0 + 1) is homotopic in C to a combinatorial path from v3 (i0 − 1) to v3 (i0 + 1)
where all type 3 intermediate vertices have degree strictly less than deg v3 (i0 ).
Proof. Since deg v3 (i0 ) > deg v3 (i0 + 1), we know that v3 (i0 ) admits an elementary
reduction, so it makes sense to choose v2′ a vertex of minimal degree such that
v3 = v3 (i0 ) admits an elementary reduction v3′ with center v2′ . Then we are going to
apply Set-Up 4.29 twice, taking v3′′ to be successively v3 (i0 − 1) and v3 (i0 + 1). It is
sufficient to prove that in both cases the combinatorial path v3′′ , v3 , v3′ is homotopic
to a combinatorial path from v3′′ to v3′ where all type 3 intermediate vertices have
degree strictly less than deg v3 . Observe that there are degenerate cases which are
easy to handle. First if v3′ = v3′′ , we just take the combinatorial path with one single
vertex v3′ . Second, if v3′ and v3′′ share the same center with v3 , we can just discard
v3 to obtain a combinatorial path from v3′′ to v3′ without any intermediate vertex, so
we can indeed assume that the centers v2′ and v2′′ are distinct as required in Set-Up
4.29.
If v2′ = m2 then by Corollary 5.4 we are in one of the cases covered by Lemma
4.34, and the homotopy is clear in all cases (see Figure 9). For instance in Case (1)
we replace the path v3′′ , v3 , v3′ by v3′′ , u3 , w3′ , v3′ , and the other cases are similar.
If v2′ ≠ m2 , and v2′ has outer resonance in v3 , then we are in one of the two cases
covered by Lemma 4.32, and again the homotopy is clear (see Figure 8).
If v2′ ≠ m2 , and v2′ has no outer resonance in v3 , then by Lemma 4.33 there
exists u3 an elementary K-reduction of v3 with center v2′ . Therefore, in that case,
up to replacing v3′ by u3 we may assume without loss in generality that v3′ is an
elementary K-reduction of v3 . By Proposition 5.1, this can only happen in the case
v3′′ = v3 (i0 − 1), and deg v3 (i0 ) = deg v3 (i0 − 1). By Stability of K-reduction 3.35,
Case (1) or (2), v3′ is a proper K-reduction of v3 (i0 − 1), and up to a local homotopy
(see Figure 6, Case (2)) we can assume that the intermediate vertex is v3 = v3 (i0 ).
If this proper K-reduction is not normal, we obtain the expected homotopy from
the normalization process of Lemma 3.27 (see Figure 5). Otherwise, By Stability of
K-reduction 3.35, Case (3) or (4), we get that v3′ is a normal proper K-reduction of
v3 (i0 − 2), with v2 (i0 − 2) = v2 (i0 − 1) the minimal line of v3 (i0 ): This contradicts
our assumption that we started with a locally geodesic loop.
We need one last ingredient before proving the simple connectedness of the complex C.
Lemma 5.6. The link L(v1 ) of a vertex of type 1 is a connected graph.
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
45
Proof. By transitivity of the action of Tame(A3 ) on vertices of type 1, it is sufficient
to work with the link L([x3 ]). Let v3 = Jf1 , f2 , x3 K be a vertex of type 3 in L([x3 ]),
where we choose our representative such that deg f1 > deg f2 .
First observe that v3 does not admit any elementary K-reduction. Indeed by
Corollary 3.8 the pivotal simplex of such a reduction should have Strong Pivotal
Form ↺(s) for some odd s ⩾ 3. In particular there exist δ ∈ N3 and a reordering
{g2 , g3 } = {f2 , x3 } such that
deg f1 = sδ,
deg g2 = 2δ
and
sδ > deg g3 > (s − 2)δ.
But since deg x3 = (0, 0, 1) is the minimal possible degree of a component of an
automorphism, both cases g2 = x3 or g3 = x3 are impossible.
It follows that v3 also does not admit a proper K-reduction: such a reduction
would be via w3 = Jg1 , f2 , x3 K, but we just proved that such a w3 cannot admit an
elementary K-reduction.
By Theorem 4.1, we conclude that v3 admits an elementary reduction v3′ , which
clearly must admit [x3 ] as pivot, since there is no non-constant polynomial P ∈
k[x1 , x2 , x3 ] with deg x3 > deg P . In particular, v3′ also is in L([x3 ]), and by induction on degree, we obtain the existence of a reduction path from v3 to [id] that stays
in L([x3 ]).
Now we recover a result of [Umi06] and [Wri15], about relations in Tame(A3 ).
Precisely, Umirbaev gives an algebraic description of the relations, and Wright shows
that this result can be rephrased in terms of an amalgamated product structure over
three subgroups. Our proof follows the same strategy as in [BFL14, Proposition
3.10].
Proposition 5.7. The complex C is simply connected.
Proof. Let γ be a loop in C. We want to show that it is homotopic to a trivial loop.
Without loss in generality, we can assume that the image of γ is contained in the
1-skeleton of the square complex, and that γ(0) = Jx1 , x2 , x3 K is the vertex of type 3
associated with the identity.
A priori (the image of) γ is a sequence of edges of arbitrary type. By Lemma 5.6,
we can perform a homotopy to avoid each vertex of type 1. So now we assume that
vertices in γ are alternatively of type 2 and 3. Precisely, up to a reparametrization
we can assume that for each i, γ(2i) has type 3 and γ(2i + 1) has type 2, so that γ
defines a combinatorial path by setting v3 (i) = γ(2i) for each i. By removing some
of these vertices we can also assume that γ is a locally geodesic loop.
Let v3 (i0 ) be the maximal vertex of the loop, and δ0 its degree. Then by Lemma
5.5 we can conclude by induction on the couple (δ0 , i0 ), ordered with lexicographic
order.
Since Tame(A3 ) acts on a simply connected 2-dimensional simplicial complex with
fundamental domain a simplex, we can recover the group from the data of the stabilizers of each type of vertex. This is a simple instance of the theory of developable
complexes of groups in the sense of Haefliger (see [BH99, III.C]). Following Wright
we can phrase this fact as follows:
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
46
Corollary 5.8 ([Wri15, Theorem 2]). The group Tame(A3 ) is the amalgamated
product of the three following subgroups along their pairwise intersections:
Stab([x1 , x2 , x3 ]) = A3 ;
Stab([x1 , x2 ]) = {(ax1 + bx2 + c, a′ x1 + b′ x2 + c′ , αx3 + P (x1 , x2 ))};
Stab([x1 ]) = {(ax1 + b, f2 (x1 , x2 , x3 ), f3 (x1 , x2 , x3 )}.
6. Examples
We gather in this last section a few examples of interesting reductions.
Example 6.1 (Elementary K-reduction with s = 3). Let
g = (x1 , x2 , x3 + x21 − x32 ),
t1 = (x1 + αx2 x3 + x33 , x2 + x23 , x3 ).
Clearly in the composition g ○ t1 the terms of degree 6 cancel each other. Moreover,
if we choose α = 32 this is also the case for terms of degree 5:
g ○ t1 = (x1 + 32 x2 x3 + x33 , x2 + x23 , x3 + x21 − x32 + 3x1 x2 x3 +
x23
4 (8x1 x3
− 3x22 )) .
Consider now a triangular automorphism preserving the quadratic form 8x1 x3 − 3x22
that appears as a factor:
t2 = (x1 , x2 + x21 , x3 + 34 x1 x2 + 38 x31 ).
A direct computation shows that the components of f = g ○t1 ○t2 = (f1 , f2 , f3 ) admits
the following degrees:
(9, 0, 0), (6, 0, 0), (7, 0, 1).
Finally, u3 = [t1 ○ t2 ], whose degrees of components are
(9, 0, 0), (6, 0, 0), (3, 0, 0),
is an elementary K-reduction of v3 = [f1 , f2 , f3 ]. Following notation from the definition of Strong Pivotal Form, we have s = 3 and δ = (3, 0, 0). Moreover
2
9 2
3
df1 ∧ df2 = − 23 (x21 − x2 )dx2 ∧ dx3 + ( 27
16 x1 x2 + 8 x2 + 2 x1 x3 + 1)dx1 ∧ dx2
+ (− 49 x31 − 32 x1 x2 + 2x3 )dx1 ∧ dx3 .
so that deg df1 ∧ df2 = (4, 0, 1), from the contribution of the factor x31 dx1 ∧ dx3 .
Example 6.2 (Elementary K-reduction with s = 5, see also [Kur09]). One can
apply the same strategy to produce examples of K-reduction with s ⩾ 3 an arbitrary
odd number. We give the construction for s = 5, and leave the generalization to the
reader. Let
g = (x1 , x2 , x3 + x21 − x52 ),
t1 = (x1 + αx22 x3 + βx2 x33 + x53 , x2 + x23 , x3 ).
Observe that αx22 x3 + βx2 x33 + x53 is homogeneous of degree 5, by putting weight 1
5
on x3 and weight 2 on x2 . By choosing α = 15
8 , β = 2 , we minimize the degree of the
composition:
g ○ t1 = (x1 +
15 2
8 x2 x3
+ 52 x2 x33 + x53 , x2 + x23 , x3 + 81 x43 (16x1 x3 − 5x32 ) + ⋯) .
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
47
Now take the following triangular automorphism, which preserves the polynomial
16x1 x3 − 5x32 :
5
t2 = (x1 , x2 + x21 , x3 + 16
(3x1 x22 + 3x31 x2 + x51 )) .
We compute the degrees of the components of f = g ○ t1 ○ t2 = (f1 , f2 , f3 ):
(25, 0, 0), (10, 0, 0), (20, 3, 0).
Finally, u3 = [t1 ○ t2 ], whose degrees of components are
(25, 0, 0), (10, 0, 0), (5, 0, 0),
is an elementary K reduction of v3 = [f1 , f2 , f3 ]. Here we have s = 5 and δ = (5, 0, 0).
Moreover
(x21 + x2 ) dx2 ∧ dx3 + (2x3 − 85 (5x51 − 9x31 x2 − 3x1 x22 )) dx1 ∧ dx3
df1 ∧ df2 = − 15
8
2
75
(5x41 x22 + 9x21 x32 + 3x42 ) +
+ ( 128
3
15
8 (x1 x3
+ 2x1 x2 x3 ) + 1) dx1 ∧ dx2 ,
so that deg df1 ∧ df2 = (5, 3, 0), from the contribution of the factor x41 x22 dx1 ∧ dx2 .
Example 6.3 (Elementary reduction without Strong Pivotal Form). We give examples of elementary reduction that show that the three assumptions in Proposition
3.7 are necessary to get Strong Pivotal Form.
(1) Let f = (f1 , f2 , f3 ) ∈ Tame(A3 ) and r ⩾ 2 such that
deg f1 > deg f3 = r deg f2 .
In particular there exists a ∈ k such that deg f3 > deg(f3 + af2r ). For instance
f = (x1 + x33 , x2 , x3 + x22 ) is such an automorphism, for r = 2 and a = −1. Then
u3 = Jf1 , f2 , f3 + af2r K is an elementary reduction of v3 = Jf1 , f2 , f3 K, and the pivotal
simplex does not have Strong Pivotal Form. Observe that v2 = Jf1 , f2 K has outer
resonance in v3 .
(2) Let u3 = Jf1 , f2 , f3 +P (f1 , f2 )K be an elementary K-reduction of w3 = Jf1 , f2 , f3 K,
with 2 deg f1 = s deg f2 for some odd s ⩾ 3. For instance we can start with one of the
′
r
r
examples 6.1 or 6.2. Pick any integer r ⩾ s+1
2 . Then v3 = Jf1 +f2 , f2 , f3 +P (f1 −f2 , f2 )K
r
is an elementary reduction of v3 = Jf1 + f2 , f2 , f3 K, and the pivotal simplex does not
have Strong Pivotal Form. Observe that the center v2 = Jf1 + f2r , f2 K has inner
resonance.
(3) With the notation of Example 6.1 or 6.2, the elementary reduction from Jg○t1 K
to Jt1 K gives an example of an elementary reduction where the center m2 , which is
the minimal line, does not have inner or outer resonance, and again the pivotal
simplex does not have Strong Pivotal Form.
Example 6.4 (Non-normal proper K-reduction). Pick the elementary K-reduction
from Example 6.2, and set v3′ = [f1 +f22 , f2 , f3 ], which is a weak elementary reduction
of v3 . Then u3 is a non-normal proper K-reduction of v3′ , via v3 . This corresponds
to Case (1) of Stability of a K-reduction 3.35. We mention again that it is an open
question whether there exists any normal proper K-reduction.
Non Example 6.5 (Hypothetical type II and type III reductions). From Corollary
3.8 we know that if v3 admits an elementary K-reduction, then the pivotal simplex
has Strong Pivotal Form ↺(s) for some odd s ⩾ 3. In particular if s = 3, then
v3 = Jf1 , f2 , f3 K with deg f1 = 3δ, deg f2 = 2δ and deg f3 > δ for some δ ∈ N3 . It is
not clear if there exists an example of such a reduction with 32 δ > deg f3 , or even
2δ > deg f3 . Observe that such an example would be the key for the existence of
COMBINATORICS OF THE TAME AUTOMORPHISM GROUP
48
the following reductions (for the definition of a reduction of type II or III, see the
original paper [SU04b], or [Kur10, §7]):
(1) If 32 δ > deg f3 , then v3′ = Jf1 + f32 , f2 , f3 K would admit a normal proper Kreduction, via v3 : This would correspond to a type III reduction.
(2) If 2δ > deg f3 > 23 δ, then v3 would admit an elementary K-reduction such
that the pivot [f2 ] is distinct from the minimal vertex [f3 ]: This would
correspond to a type II reduction.
(3) If 32 δ > deg f3 , then v3′′ = Jf1 , f2 + f32 , f3 K would be an example of a vertex
that admits a reduction along a center with outer resonance in v3′′ , but that
does not admit a reduction with center the minimal line of v3′′ (see Lemma
4.32).
References
[BFL14] C. Bisi, J.-P. Furter & S. Lamy. The tame automorphism group of an affine quadric
threefold acting on a square complex. Journal de l’École Polytechnique - Mathématiques,
1:161–223, 2014.
[BH99] M. R. Bridson & A. Haefliger. Metric spaces of non-positive curvature, volume 319 of
Grundlehren der Mathematischen Wissenschaften. Springer-Verlag, Berlin, 1999.
[CL13] S. Cantat & S. Lamy. Normal subgroups in the Cremona group. Acta Math., 210(1):31–94,
2013. With an appendix by Yves de Cornulier.
[HP94] W. Hodge & D. Pedoe. Methods of algebraic geometry. Vol. I. Cambridge Mathematical
Library. Cambridge University Press, Cambridge, 1994. Book I: Algebraic preliminaries,
Book II: Projective space, Reprint of the 1947 original.
[Kur08] S. Kuroda. A generalization of the Shestakov-Umirbaev inequality. J. Math. Soc. Japan,
60(2):495–510, 2008.
[Kur09] S. Kuroda. Automorphisms of a polynomial ring which admit reductions of type I. Publ.
Res. Inst. Math. Sci., 45(3):907–917, 2009.
[Kur10] S. Kuroda. Shestakov-Umirbaev reductions and Nagata’s conjecture on a polynomial automorphism. Tohoku Math. J. (2), 62(1):75–115, 2010.
[Lam02] S. Lamy. Une preuve géométrique du théorème de Jung. Enseign. Math. (2), 48(3-4):291–
315, 2002.
[Lon16] A. Lonjou. Non simplicité du groupe de Cremona sur tout corps. Ann. Inst. Fourier,
66(5):2021–2046, 2016.
[LP16] S. Lamy & P. Przytycki. Acylindrical hyperbolicity of the three-dimensional tame automorphism group. arXiv:1610.05457, 2016.
[Mar15] A. Martin. On the acylindrical hyperbolicity of the tame automorphism group of SL2 (C).
arXiv:1512.07526, 2015.
[SU04a] I. P. Shestakov & U. U. Umirbaev. Poisson brackets and two-generated subalgebras of
rings of polynomials. J. Amer. Math. Soc., 17(1):181–196, 2004.
[SU04b] I. P. Shestakov & U. U. Umirbaev. The tame and the wild automorphisms of polynomial
rings in three variables. J. Amer. Math. Soc., 17(1):197–227, 2004.
[Umi06] U. U. Umirbaev. Defining relations of the tame automorphism group of polynomial algebras
in three variables. J. Reine Angew. Math., 600:203–235, 2006.
[Vén11] S. Vénéreau. A parachute for the degree of a polynomial in algebraically independent ones.
Math. Ann., 349(3):589–597, 2011.
[Wri15] D. Wright. The generalized amalgamated product structure of the tame automorphism
group in dimension three. Transform. Groups, 20(1):291–304, 2015.
Institut de Mathématiques de Toulouse, Université Paul Sabatier, 118 route de
Narbonne, 31062 Toulouse Cedex 9, France
E-mail address: [email protected]
| 4math.GR
|
arXiv:1609.04735v1 [cs.NI] 15 Sep 2016
RALL — Routing-Aware Of Path Length, Link
Quality, And Traffic Load For Wireless Sensor
Networks
Vinícius N. Medeiros, Douglas V. Santana,
Bruno Silvestre, Vinicius da C. M. Borges
{viniciusnunesmedeiros, douglassantana}@inf.ufg.br
{brunoos, vinicius}@inf.ufg.br
Instituto de Informática – Universidade Federal de Goiás
Goiânia, GO – Brasil
15 September 2016
Abstract
Due to the enormous variety of application scenarios and ubiquity, Internet of Things (IoT) brought a new perspective of applications for the
current and future Internet. The Wireless Sensor Networks provide key
devices for developing the IoT communication paradigm, such as the sensors collecting various kind of information and the routing and MAC protocols. However, this type of network has strong power consumption and
transmission capacity restrictions (low speed wireless links and subject to
interference). In this context, it is necessary to develop solutions that enable a more efficient communication based on the optimized utilization of
the network resources. This papers aims to present a multi-objective routing algorithm, named Routing-Aware of path Length, Link quality, and
traffic Load (RALL), that seeks to balance three objectives: to minimize
bottlenecks, to minimize path length, and to avoid links with low quality. RALL results in good performance when taking into consideration
delivery rate, overhead, delay, and power consumption.
1
Introduction
The Internet has been employed more and more, and becoming an essential tool
for humans. Nowadays, not only people use this information and communication
technology, as well as machines employ it to communicate with each other, making measurement of various types of information. The Internet of Things (IoT)
paradigm allows these machines to communicate and feel the environment [3].
1
Wireless Sensor Networks (WSNs) play an important role within the IoT and
have gained increasing prominence as part of ubiquitous computing in various
environments, such as industry, smart cities, smart spaces, smart grids, health
monitoring environmental, real-time multimedia applications [3, 15, 2]. WSNs
consist of various nodes that generally have strong processing, battery, and
memory constraints. Typically, each node has only one radio interface (generally
802.15.4 CSMA/CA) with a fixed transmission rate (250 kbps) [20].
Usually, a flat model is employed to organize the network, where all the
nodes play the same role in sensing, processing, and (re)broadcast packets. The
information is forwarded (routed) node by node until it reaches the sink node,
where it is processed [1, 14, 17]. This paper will treat the routing optimization
problem based on three objectives: (i) minimize the path length (number of
hops), (ii) minimize the bottleneck, and (iii) minimize the interference in the
path. At the routing, each hop could increase the delay.
This article is organized as follows: In Section 2, we present a model for
the routing problem of multi-objective optimization. In Section 3, we describe
the proposed algorithm to determine a solution for the model. In Section 4,
we present the simulation results. In Section 5, we analyze the most relevant
related work on routing approaches for WSNs. The conclusion and future work
is carried out in Section 6.
2
System Model
WSNs consist of various sensors, distributed in an area, that enable data collection. Most of these sensors monitor or interact with the environment in which
they are. A node in particular, called sink, is designed to gather all the data
collected by the sensors. The sink node is attached to a machine with more
processing capacity (e.g. servers in cloud computing environment) so that collected data can be transformed into information and knowledge, to be used by
various applications.
WSNs are modeled as a graph G = (V, E), where V is the set of vertices
that represents the sensors, E is the set of communication links between two
network devices. The link esd ∈ E between s, d ∈ V exists only if the device s
accomplish a data transmission to the node d, i.e. the node d must be in the
transmission range of s. Some metrics can be employed for determining if a link
between two sensors can actually be used. Every sensor s ∈ V − {i}, where i
is the sink, is responsible for the origin of a data stream named fs ∈ F , where
F is defined as the set of all flows generated in the network. All flows of WSNs
have the sink node i as destination node.
The establishment of a path is demanded on the graph G so that the data
flow can reach the sink node. The selection of each flow path must follow some
objectives to ensure certain network characteristics, such as packet delivery rate,
energy consumption, delay, etc. A larger number of hops require the activation
of more links, that generates more use of the transmission medium and may
lead to more contention or interference as well as end-to-end delay [6].
2
Load balancing aims at distributing the flows over the network uniformly.
This prevents a small set of nodes to be used for the majority of the routes,
leading to a greater burden on those nodes. It is important to point out that the
nodes have limited resources, so that load can cause, for example, the reduction
of the network lifetime, due to a higher consumption of battery to forward
packets, as well as for packet loss, given the small amount of memory in the
nodes for packet storage [8].
The interference degrades the link quality [6]. It can be caused by the number of neighbors that are in communication range of each node and/or other
wireless networks that generally adopt communication standards using same
spectrum band. As nodes in a WSNs use this shared band, and they employ
only a radio interface and a single communication channel, more nodes in the
network result in higher level of interference. This implies a greater chance of
collision (requiring packet retransmission) because there is a greater probability
that the channel is busy. As a result, interference strongly impacts the effective transmission capacity. Thus, the routes that pass through nodes with a
greater degree of interference may suffer more delay in delivery and/or more
retransmission of packets, as well as may result in larger battery consumption.
It can be noticed interrelationships between these three objectives. The
increase in the number of hops in the path is a common consequence when a
routing solution seeks to minimize overload [6]. However, increasing path length
causes more nodes activation, in consequence more interference. It is important
to notice that not always shorter paths generate less interference in the WSN
as a whole. For instance, the shortest paths may suffer interference due to the
existence of neighbors or other wireless networks/devices that are in the same
range area of these paths.
Following we present the construction of an optimization model for the route
establishment on WSN with three objectives: minimizing the number of hops,
minimizing the network bottleneck on the links, and minimizing the use of low
quality links. These three objectives are described and modeled in the following
paragraphs.
Number of Hops Minimizing the number of hops in a required path can
reduce the end-to-end delay. A linear programming model is used to find the
shortest paths from f flows in F :
minimize
X
asd
(1)
esd ∈E
Subject to :
X
avd −
d∈V
X
X
adv = 1, ∀v ∈ V, v 6= i
(2)
d∈V
asi = |F |, i = sink
(3)
s∈V
asd ≥ 0
(4)
3
Variable asd is defined as the sum of the flows using the edge esd on your
route. Constraint (2) ensures that every vertex can only generate a single flow
and constraint (3) ensures that the sink node can receive a number of flows equal
to the number of sensors in the network. The combination of restrictions (2) and
(3) ensures that loops are not created. The constraint (4) ensures non-negative
values for sum of flows on links in asd . The objective function (1) minimizes
the sum of the amount of all flows that pass through each edge esd ∈ E. Thus,
each flow is allocated to the shortest path and therefore the objective function
value is minimized.
Wireless Links with Low Quality The shortest paths may be more susceptible to high level of interference that increases the packet loss rate, since
the interference decreases the link quality. The quality of the wireless communication affects the overall network capacity. Wireless links that do not have
their quality degraded by factors, such as interference and noise, become a good
choice for use on routes to the sink node, while links that have poor data transmission capacity (i.e., high level of interference) should be avoided to improve
package delivery rate.
The LQI (Link Quality Indication) is a measurement provided by the IEEE
802.15.4 physical layer that can be used as a quality metric of the wireless
transmission between two sensors. LQI values are represented in a range from 0
to 255 [7]. In this paper, a link is considered weak or with low quality if the LQI
value is smaller than a certain threshold (T HLQI ), according to the following
condition:
(
0
, LQIsd ≥ T HLQI
(5)
lsd =
sd
)
, LQIsd < T HLQI
1 − ( TLQI
HLQI
LQIsd is the value of LQI regarding the wireless link between nodes s, d ∈ V .
The variable lsd is used in the proposed model to represent link quality values
that should be avoided (higher ones) or selected (lower ones). When lsd is zero,
it indicates that the link between the devices has an acceptable LQI, i.e. it has
a good quality. Otherwise, the value is normalized to the interval (1, 0) in order
to distinguish the links with lower quality, but that have a good chance of being
used, since they have a higher LQI.
The variable lsd is fundamental to minimize the use of links with low quality
on the paths. The objective function for this approach is constructed combining
the objective function (1) and taking into the number of hops (variable lsd ):
(
)
X
minimize
lsd · asd
(6)
esd ∈E
Network Bottleneck The distribution of flows in a WSNs, in a balanced
way, increases the network lifetime. On the other hand, an unfair distribution of
paths for flows creates agglomeration flows on the same node (i.e. bottleneck),
4
increasing the power consumption due to the increase of forwarded packets.
Since the sensors usually have strong restrictions relating to energy consumption, the question of minimizing bottlenecks is a very important feature for WSN
scenarios. Furthermore, smoothing out network bottleneck can improve traffic
performance.
The variable asd enables the objective function that is employed to minimize
the network bottleneck. Therefore, the objective function is a combination of
paths for each flow on the graph G, where the highest sum of the amount of
flows passing through a node s is minimal:
(
(
max
minimize
!)
)
X
asd
, ∀s ∈ V
(7)
d∈V
Multi-objective Model The model of the multi-objective optimization problem for the routing problem in WSNs uses three objectives (number of hops,
quality links, and network bottleneck):
(
)
X
minimize
asd
esd ∈E
(
)
X
minimize
lsd · asd
esd ∈E
(
minimize
(
max
)
X
asd
!)
, ∀s ∈ V
d∈V
Subject to:
X
avd −
d∈V
X
X
adv = 1, ∀v ∈ V, v 6= i
d∈V
asi = |F | − 1, i = sink
s∈V
asd ≥ 0
The objective functions (1), (6), and (7), when combined in a single model,
create a problem with conflicting solutions by having a set of optimal solutions
that are not dominated by each other, called the set of Pareto optimal. In other
words, the modeled objectives are conflicting each other because there is no
solution that is able to meet optimally all three objectives.
The conflict between the objectives that minimizes the amount of hops (1),
minimizes the amount of low-quality wireless links (6), and minimizes the network bottleneck (7) is evident since the objectives (1) and (6) group the paths
in order to select shortest paths or a link with better quality, objective function
(7) try to distribute the flows equally over the paths on the entire network,
therefore increasing the average path length and the use of links with low qual-
5
ity. The objective functions (1) and (6) are conflicting when the establishment
of shortest route demands links with lower quality.
3
RALL — Routing-Aware of path
Length, Link quality, and
traffic Load
In this section we describe the multi-objective algorithm, called Routing-Aware
of path Length, Link quality, and traffic Load (RALL) for the routing problem in
WSNs. RALL algorithm employs some techniques to simplify the complexity of
selecting the set of Pareto optimal solutions. For this reason, it uses an algorithm
for determining paths with lower cost (Dijkstra’s algorithm) so that each path
is calculated for a flow generated on an ordinary node. The values related to the
objective function are changed according to the nodes that had already their
specific route for the flows in order to achieve the balancing. Initially, it is carried
out the transformation of the model so that the objective functions of path
length minimization (1) and low quality links minimization (6) are combined in
a single objective.
When a solution of an objective function is not dominated for the other objective, it is necessary to establish a trade-off between the functions to determine
which solution can satisfy partially or completely the objectives. The weighted
sum method of the objective functions is usually applied to perform the combination of objectives. Therefore, it is necessary to establish a weight for each
objective function and perform a weighted sum of these values. This method
was chosen because it is well known for the problems of linear programming [10]:
(
!
!)
X
X
minimize wp ·
asd + wl ·
lsd · asd
=
esd ∈E
esd ∈E
(
)
X
(wp + wl · lsd ) · asd
(8)
esd ∈E
wp and wl are the weights for the objective function related to the number
of hops (1) and number of low quality links (6), respectively. As a requirement
to apply this approach, the decision variables of the objective functions must
be at the same scale or magnitude, it is not the case of the function (7). In the
presented model, the decision variable for the objectives is the number of flows
passing through the links (edges).
RALL algorithm begins by performing a combination of objective functions
path length minimization and low quality links minimization in a single objective
function, using the weighted sum (8). Next, RALL algorithm performs the
minimization of network bottleneck by updating the values of the objective
6
Algorithm 1: RALL (Routing-Aware of path Length, Link quality, and
traffic Load.)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
Algorithm RALL()
input :
G − (V, E)
Of − set of ordered flows
i − sink, i ∈ V
wp , wl − weights for objective functions
begin
E ← SumWeights (E,wp ,wl )
A←E
P ←∅
foreach fs ∈ Of do
ps ← M CP ath(s, i, G < V, A >)
A ← UpdateEdges(ps , E, A)
P ← P ∪ {ps }
end
end
return P
Procedure SumWeights()
input :
E − set of edges
wp , wl − weights for objective functions
begin
Enew ← ∅
foreach esd ∈ E do
enew
← pconst · wp + wl · lsd
sd
Enew ← Enew ∪ {enew
sd }
end
return Enew
end
Procedure UpdateEdges()
input :
E − set of edges
Ps − router of s to sink
A − set of edges
begin
foreach bjh ∈ Ps do
rjh ← E ∩ {b
Pjh }
amount ←
asd , ∀esd ∈ E, s = j
rjh ← ejh + amount, ejh ∈ E
A ← A ∪ {rjh }
end
return A
end
function. The algorithm 1 shows the RALL’s pseudo-code. It has a set of flows
as input that are used to generate the paths with lower cost.
On line 3, the weighted sum of objective functions is called. In this procedure, path length and low quality links are combined in a single vector cost.
The variable lsd is the normalized value of LQI for edge esd , using the constraint
shown by the equation (5). pconst is a constant used to establish the value for
each hop. The value of each link is updated according to this constant value. If
the chosen value for pconst is small in some interactions, the objective function
to determine the lowest number of hops is almost negligible, and the trade-off
will be impaired. This happens because the values that represent the bottleneck
in the objective function, on each iteration, are incremented. It is used the value
7
of pconst = |V |, and the values of variable lsd are normalized to have pconst as
the threshold value.
After the weighted sum is generated minimal cost path using M CP ath(sender,
receiver, GRAP H) for each flow fs , which is assigned to the variable ps (on line
7). The links that make up the minimum cost path PS will have their values
updated by adding the amount of flows that use the device s ∈ V — line 27 of
U pdateEdges(ps , E, A) procedure. Finally, the path is inserted into the solution
set P that will contain all routes to every flow.
4
Simulation Results
We conducted a simulation study to analyze the performance of proposed approach. A network with a single sink node and many sensor nodes are taken
into consideration in the simulated network. The sink node receives data from
the sensor nodes and it serves as a connection point to an external network. The
sensor nodes send the collected data and forward/relay packets on a single route
to the sink node. All nodes are static and they use asymmetric links. This section is organized as follows: the scenario configuration is outlined in sub-section
4.1. The evaluation and the simulation results are discussed in sub-section 4.2.
4.1
Scenario Configuration
We employ the Castalia module [5] of OMNeT++ simulator [18], which are
widely used to evaluate WSN. Table 1 shows the general parameters used in the
simulation and Table 2 presents node configuration based on [13].
The simulations were performed in scenarios varying the number of nodes,
i.e. 10 up to 50 nodes (Table 1). For every set of nodes, it was generated ten
topologies and four random seeds, resulting in a total of 40 runs for each set of
nodes. The positions of the nodes in a topologies were generated using a normal
distribution, limited in area, that guarantees none of the nodes would remain
isolated from the network (out of range from any other node).
Table 1: Simulation Parameters
Packet Generation Rates
5 packets/minute
Traffic generation model
Poisson
Topology Area
50x50
Number of nodes
10, 20, 30, 40, and 50
Interference Model
Additive 1
Weight values for path
50%
length function (wp )
Weight values for link
50%
quality function (wl )
1 Additive interference: it use the SINR values of the package, the receiver receives the
strongest signal of the two transmissions (if it is strong enough).
8
Table 2: Node Configuration
Initial Energy
100 J
TX Energy
20 mJ/pkt
RX Energy
10 mJ/pkt
MAC layer
T-MAC
Radio Model
CC2420
Transmission Power
-10dBm
Packet Size
127 bytes
Link Communication Asymmetric
The simulation was carried out in two phases. In the first phase, the neighborhood discovery was done by disseminating broadcast messages. These control
messages are also used to calculate the average LQI value of the link for each
neighbor so that the sink node has the complete network graph information.
Due to the fact that the processing capacity and energy of the sensors devices
are very restricted, the routing algorithm is performed in a centralized way in
the sink node.
Once in possession of the network global view, the proposed routing algorithm is executed to select the path to the sink for each sensor node. The second
simulation phase uses the routing table information to forward the packets. Each
node sends messages to the sink, which will extract some of the performance
metrics for analysis.
4.2
Results
Other routing algorithms were chosen to assist the performance assessment of
RALL, as following:
• BALAN CED − LQI: It is based on similar criteria of RALL, however, it
only employs the load balancing and link quality objective functions, and
it does not perform the combination of functions using the weighted sum.
• BP R: It is a heuristic with two objective functions, load balancing and
path length [11]. For each flow generated on the network, BPR seeks to
select the shorter route belonging to set of candidate paths that does not
increase the network bottleneck.
• P AT H: Traditional approach to determine the shortest path length (e.g.
Dijkstra’s algorithm).
• LQI: It aims to select routes that use fewer links with low quality. A link
is considered low quality when it does not satisfied the described condition
in (5).
The comparison of these approaches aims to provide better impact of the
different criteria in the traffic and network performance through the route selection.
9
Figure 1 shows the packet loss rate. We figure out that the BALAN CED −
LQI and RALL algorithms achieve lower levels of losses, especially in high
density network (i.e. 50 nodes). Therefore, the amount of data increases. This
can be justified by the Jain fairness index (well-known metric to evaluate load
balancing) shown in Figure 2, and by the fact that the two algorithms take
into consideration the link quality. Despite the fact that BALAN CED − LQI
slightly results in lower packet loss, the confidence interval (Figure 1) shows
that the performance of both approaches is very similar.
Figure 1: Packet loss rate.
Figure 2: Jain fairness index.
Figure 3 shows the average latency in data delivery of the simulated network
scenarios. It is worth noting that in the scenario with 10 nodes, most of packets
10
were delivered within the range of [0, 600)ms. The other intervals follow the
behavior of the first three ones. In fact, it occurs in all scenarios, thus we show
only the first four intervals: 20, 30, 40, and 50.
(a) 10 nodes
(b) 20 nodes
(c) 30 nodes
(d) 40 nodes
(e) 50 nodes
Figure 3: Latency for 10, 20, 30, 40, and 50 nodes.
The greatest difference in latency between approaches can be noticed when
the number of nodes increases. Again, the BALAN CED − LQI and RALL
approaches are featured, however, RALL results in an significant improvement
of the latency, followed by P AT H. These results can be justified by analyzing
the average path length generated by each approach (Figure 4), where RALL
and P AT H reached the lowest values.
We also analyzed the average network lifetime, which was measured as from
the beginning of the second phase up to the first node of the network run out
the battery energy. In this analysis we used only BALAN CED − LQI and
RALL approaches, because they showed better performance for packet loss and
latency. Figure 5 illustrates the lifetime.
RALL approach achieves performance improvement when increasing the
number of nodes (i.e. 30, 40, and 50 nodes). In order to better understand this
behavior, we analyzed the packets exchanged by each algorithm. In Figure 6,
data packets are generated by the application layer and accounted in packet loss
by the sink. Control packets are accounted of the MAC layer, which uses the
T-MAC protocol. This protocol employs an adaptive method to avoid collision
11
Figure 4: Average path length.
Figure 5: Average network lifetime.
and to control the radio duty-cycle, in order to reduce the energy consumption.
It is important to stress out that BALAN CED − LQI has a slightly larger
number of data packets delivery to all scenarios. Moreover, BALAN CED−LQI
12
Figure 6: Average control overhead in MAC layer.
also requires a larger amount of control packets in the MAC layer. Each packet
has an energy cost for sending and receiving. This explains why the network
lifetime achieves the highest times by getting shortest paths (activating fewer
links).
5
Related Work
There are many works that seek to combine these objectives, some of them to
optimize routing in wireless networks, especially in WSNs. The most of these
works deal with one or two objectives described above. Wang and Zhang [19]
presented a protocol, called Interference Aware Multipath Routing (IAMR),
that selects paths spatially disjoint, adopting an interference model in which
interference range is twice greater than the transmission range. This model
may overestimate the interference level.
Radi et al. [16] proposed Low-Interference Energy-Efficient Multipath Routing Protocol (LIEMRO), which takes into account a load balancing multipath
approach and which also seeks to minimize interference through the use of ETX
routing metric. As it was described in the literature, ETX has limitations to
capture the interference in a more accurate way [4].
Minhas et al. [12] recommended an approach based on fuzzy logic for multiobjective optimization problem in routing that seeks to balance two objectives:
to maximize the network lifetime (analyzing the current and residual energy
levels of the nodes) and the delay between the ordinary node and the sink node
13
(taking into account of the number of hops).
Load balancing routing can provide both packet loss minimization and network lifetime maximization based on a uniform distribution of flows in the network. For this reason, routing approaches are proposed in wireless networks
to minimize bottlenecks and the average path length, such as Bottleneck, Path
Length and Routing heuristic overhead (BPR) [11].
Moghadam et al. [13] presented a heuristic, named Heuristic Load Distribution (HeLD), which aims to minimize the overhead communication and maximize the network lifetime by a load balancing routing based on a linear programming approach, and the use of a set of braided paths. Simulation results
showed that HeLD achieved less overhead and an increase in network lifetime,
however, packet delivery rate was lower and latency was not presented. This
can be partially justified by braided paths that can generate a high level of
interference.
Machado et al. [9] presented a heuristic for IoT environments called Routing
by Energy and Link quality (REL), which seeks to select paths that minimize
the number of hops, maximize network lifetime through the residual energy of
the nodes, and maximize the link quality of the in the selected paths based
on Received Signal Strength Indicator (RSSI) metric. However, REL is quite
simple and it is limited to establish a balance among the three objectives.
None of the related work handles the three mentioned objectives. Most
of them use heuristics for routing solution. This paper propose a new multiobjective routing algorithm for WSNs. The algorithm seeks to achieve the tradeoff among load balancing, path length, and interference, i.e. aggregate flows in
the shortest paths to avoid overloaded links and with high level of interference.
6
Conclusions and Future Work
We propose a multi-objective routing algorithm (RALL) for IoT environments
in this paper. RALL seeks to minimize three objectives: bottlenecks, links with
high levels of interference, and long paths.
The results showed that RALL algorithm results in good levels of packet loss
rate (very similar to the main comparative approach), the lowest latency, and
longer average network lifetime, mainly in dense scenarios.
Although RALL does not reach the best results in all tests, it shows the
importance of the all employed criteria in order to offer an overall routing approach for IoT. The weakness of one criterion is smoothed out by the others as
well as the equilibrium of the criteria to provide a more efficient performance
for WSNs. We believe that the proposed algorithm is an interesting approach
to IoT environments that have different types of traffic and require routing approaches that seeks to balance conflicting objectives to improve delivery rate,
delay, and power consumption.
As future work, we intend to start further studies about the weights used in
the objective functions, for example, trying to make a best correlation between
the weights and types of topologies. Moreover, the RALL algorithm will be
14
implemented in testbeds for performing experiments in a real environment.
References
[1] Ian F. Akyildiz, Tommaso Melodia, and Kaushik R. Chowdhury. A survey
on wireless multimedia sensor networks. Comput. Netw., 51(4):921–960,
March 2007.
[2] Adwan Alanazi and Khaled Elleithy. Real-time QoS routing protocols
in wireless multimedia sensor networks: Study and analysis. Sensors,
15(9):22209, 2015.
[3] Luigi Atzori, Antonio Iera, and Giacomo Morabito. The internet of things:
A survey. Computer Networks, 54(15):2787 – 2805, 2010.
[4] Vinicius C. M. Borges, Marilia Curado, and Edmundo Monteiro. CrossLayer Routing Metrics for Mesh Networks: Current Status and Research
Directions. Computer Communications, 34(6):681 – 703, 2011.
[5] Athanassios Boulis. Castalia: A simulator for wireless sensor networks and
body area networks. NICTA: National ICT Australia, 2011.
[6] Eduardo Feo Flushing and Gianni A. Di Caro. A flow-based optimization
model for throughput-oriented relay node placement in wireless sensor networks. In Proceedings of the 28th Annual ACM Symposium on Applied
Computing, SAC ’13, pages 632–639, New York, NY, USA, 2013. ACM.
[7] Carles Gomez, Antoni Boix, and Josep Paradells. Impact of LQI-Based
routing metrics on the performance of a one-to-one routing protocol for
IEEE 802.15.4 multihop networks. EURASIP Journal on Wireless Communications and Networking, 2010(1):1–20, 2010.
[8] Jaewon Kang, Yanyong Zhang, and Badri Nath. End-to-end channel capacity measurement for congestion control in sensor networks. In Proceedings
of the 2nd International Workshop on Sensor and Actor Network Protocols
and Applications (SANPA’04), 2004.
[9] Kássio Machado, Denis Rosário, Eduardo Cerqueira, Antonio A. F.
Loureiro, Augusto Neto, and José Neuman de Souza. A routing protocol based on energy and link quality for internet of things applications.
Sensors, 13(2):1942, 2013.
[10] R. Timothy Marler and Jasbir S. Arora. The weighted sum method for
multi-objective optimization: new insights. Structural and Multidisciplinary Optimization, 41(6):853–862, 2009.
[11] Micael O. M. C. Mello, Vinicius C. M. Borges, Leizer de L. Pinto, and Kleber Vieira Cardoso. Load balancing routing for path length and overhead
15
controlling in wireless mesh networks. In IEEE Symposium on Computers and Communications, ISCC 2014, Funchal, Madeira, Portugal, June
23-26, pages 1–6, 2014.
[12] Mahmood R. Minhas, Sathish Gopalakrishnan, and Victor C. M. Leung.
Multiobjective routing for simultaneously optimizing system lifetime and
source-to-sink delay in wireless sensor networks. In Proceedings of the 2009
29th IEEE International Conference on Distributed Computing Systems
Workshops, ICDCSW ’09, pages 123–129, Washington, DC, USA, 2009.
IEEE Computer Society.
[13] Meisam Moghadam, Hassan Taheri, and Mehdi Karrari. Minimum cost
load balanced multipath routing protocol for low power and lossy networks.
Wireless Networks, 20(8):2469–2479, 2014.
[14] D. Niculescu. Communication paradigms for sensor networks. IEEE Communications Magazine, 43(3):116–122, March 2005.
[15] Luís Oliveira and Joel Rodrigues. Wireless sensor networks: a survey on
environmental monitoring. Journal of Communications, 6(2), 2011.
[16] M. Radi, B. Dezfouli, S. A. Razak, and K. A. Bakar. Liemro: A lowinterference energy-efficient multipath routing protocol for improving QoS
in event-based wireless sensor networks. In Sensor Technologies and Applications (SENSORCOMM), pages 551–557, July 2010.
[17] Marjan Radi, Behnam Dezfouli, Kamalrulnizam Abu Bakar, Shukor Abd
Razak, and Mohammad Ali Nematbakhsh. Interference-aware multipath
routing protocol for QoS improvement in event-driven wireless sensor networks. Tsinghua Science & Technology, 16(5):475 – 490, 2011.
[18] András Varga. The OMNeT++ discrete event simulation system. In Proceedings of the European simulation multiconference (ESM’2001), volume 9,
page 65, 2001.
[19] Z. Wang and J. Zhang. Interference aware multipath routing protocol for
wireless sensor networks. In IEEE GLOBECOM Workshops, pages 1696–
1700, Dec 2010.
[20] Jennifer Yick, Biswanath Mukherjee, and Dipak Ghosal. Wireless sensor
network survey. Computer Networks, 52(12):2292 – 2330, 2008.
16
| 8cs.DS
|
On Variational Expressions for Quantum Relative Entropies
Mario Berta,1, 2 Omar Fawzi,3 and Marco Tomamichel4
1
Department of Computing, Imperial College London
Institute for Quantum Information and Matter, California Institute of Technology
3
Laboratoire de l’Informatique du Parallélisme, École Normale Supérieure de Lyon
4
Centre for Quantum Software and Information, University of Technology Sydney
arXiv:1512.02615v2 [quant-ph] 8 Nov 2017
2
Distance measures between quantum states like the trace distance and the fidelity can
naturally be defined by optimizing a classical distance measure over all measurement
statistics that can be obtained from the respective quantum states. In contrast, Petz showed
that the measured relative entropy, defined as a maximization of the Kullback-Leibler
divergence over projective measurement statistics, is strictly smaller than Umegaki’s
quantum relative entropy whenever the states do not commute. We extend this result in
two ways. First, we show that Petz’ conclusion remains true if we allow general positive
operator valued measures. Second, we extend the result to Rényi relative entropies and
show that for non-commuting states the sandwiched Rényi relative entropy is strictly larger
than the measured Rényi relative entropy for α ∈ ( 12 , ∞), and strictly smaller for α ∈ [0, 12 ).
The latter statement provides counterexamples for the data-processing inequality of the
sandwiched Rényi relative entropy for α < 21 . Our main tool is a new variational expression
for the measured Rényi relative entropy, which we further exploit to show that certain lower
bounds on quantum conditional mutual information are superadditive.
Keywords: Quantum entropy, measured relative entropy, relative entropy of recovery,
additivity in quantum information theory, operator Jensen inequality, convex optimization.
Mathematics Subject Classifications (2010): 94A17, 81Q99, 15A45.
I.
MEASURED RELATIVE ENTROPY
The relative entropy is the basic concept underlying various information measures like entropy,
conditional entropy and mutual information. A thorough understanding of its quantum generalization is thus of preeminent importance in quantum information theory. We start by considering
measured relative entropy, which is defined as a maximization of the Kullback-Leibler divergence
over all measurement statistics that are attainable from two quantum states.
For a positive measure Q on a finite set X and a probability measure P on X that is absolutely continuous with respect to Q, denoted P ≪ Q, the relative entropy or Kullback-Leibler
divergence [26] is defined as
X
P (x)
D(P kQ) :=
P (x) log
,
(1)
Q(x)
x∈X
P (x)
Q(x)
= 0 whenever P (x) = 0. By continuity we define it as +∞ if
where we understand P (x) log
P 6≪ Q. (We use log to denote the natural logarithm.)
To extend this concept to quantum systems, Donald [12] as well as Hiai and Petz [21] studied
measured relative entropy. In the following we restrict ourselves to a d-dimensional Hilbert space
for some d ∈ N. Let us denote the set of positive semidefinite operators acting on this space by
P and the subset of density operators (with unit trace) by S. For a density operator ρ ∈ S and
σ ∈ P, we define two variants of measured relative entropy. The general measured relative entropy
is defined as
(2)
D M (ρkσ) := sup D Pρ,M Pσ,M ,
(X ,M )
2
where the optimization is over finite sets X and positive operator valued measures (POVMs)
M on X . (More formally, M is a map from X to positive semidefinite operators and satisfies
P
x∈X M (x) = 1, whereas Pρ,M is a measure on X defined via the relation Pρ,M (x) = tr[M (x)ρ]
for any x ∈ X .) At first sight this definition seems cumbersome because we cannot restrict the size
of X that we optimize over. Therefore, following the works [12, 21, 35], let us also consider the
following projectively measured relative entropy, which is defined as
D P (ρkσ) := sup
{Pi }di=1
(
d
X
i=1
tr[Pi ρ]
tr[Pi ρ] log
tr[Pi σ]
)
,
(3)
where the maximization is over all sets of mutually orthogonal projectors {Pi }di=1 and we spelled
out the Kullback-Leibler divergence for discrete measures. Note that without loss of generality we
can assume that these projectors are rank-1 as any course graining of the measurement outcomes
can only reduce the relative entropy due to its data-processing inequality [29, 49]. Furthermore,
the quantity is finite and the supremum is achieved whenever ρ ≪ σ, which here denotes that the
support of ρ is contained in the support of σ. (To verify this, recall that the rank-1 projectors form
a compact set and the divergence is lower semi-continuous.)
The first of the following two variational expressions for the (projectively) measured relative
entropy is due to Petz [37]. Note that the second objective function is concave in ω so that the
optimization problem has a particularly appealing form.
Lemma 1. For ρ ∈ S and σ ∈ P non-zero, the following identities hold:
D P (ρkσ) = sup tr[ρ log ω] − log tr[σω] = sup tr[ρ log ω] + 1 − tr[σω] .
ω>0
(4)
ω>0
Moreover, the suprema are achieved when ρ and σ are positive definite operators, i.e. ρ > 0 and
σ > 0.
Proof. If ρ 6≪ σ then the two expressions in the suprema of (4) are unbounded, as expected. We
now assume that ρ ≪ σ. Let us consider the second expression in (4). We write the supremum
over ω > 0 as two suprema over {Pi }di=1 and {ωi }di=1 , where ωi > 0 are the eigenvalues of ω
corresponding to the eigenvectors given by rank-1 projectors Pi . Using the fact that tr[ρ] = 1, we
find
sup tr[ρ log ω] + 1 − tr[σω] = sup
ω>0
sup
d
X
{Pi }di=1 {ωi }di=1 i=1
tr[Pi ρ](log ωi + 1) − tr[Pi σ]ωi .
(5)
For i ∈ [d] such that tr[Pi σ] = 0, we also have tr[Pi ρ] = 0, and thus the corresponding term is
zero. When tr[Pi σ] > 0, let us first consider tr[Pi ρ] = 0. In this case, the supremum of i-th term
is supωi >0 − tr[Pi σ]ωi = 0 achieved in the limit ωi → 0. Now in the case tr[Pi ρ] > 0 (which is the
only possible case when ρ > 0), observe that the expression is concave in ωi , the inner supremum
is achieved by the local maxima at ωi = tr[Pi ρ]/ tr[Pi σ]. Plugging this into (5), we find
sup
d
X
{Pi }di=1 i=1
d
X
tr[Pi ρ]
tr[Pi ρ]
tr[Pi ρ] log
tr[Pi ρ] log
+ tr[Pi ρ] − tr[Pi ρ] = sup
.
tr[Pi σ]
tr[Pi σ]
{Pi }d
i=1
(6)
i=1
This is the expression for the measured relative entropy in (3). The remaining supremum is achieved
because the set of rank-1 projectors is compact and the divergence is lower semi-continuous.
3
It remains to show that the two variational expressions in (4) are equivalent. We have log(1 +
x) ≤ x for all x > −1 and, thus, − log tr[σω] ≥ 1 − tr[σω] for all ω > 0. This yields
sup tr[ρ log ω] − log tr[σω] ≥ sup tr[ρ log ω] + 1 − tr[σω] .
ω>0
(7)
ω>0
Now note that the expression on the left hand side is invariant under the substitution ω → γω for
γ > 0. Hence, as tr[σω] > 0 for ω > 0 and non-zero σ, we can add the normalization constraint
tr[σω] = 1 and we have
sup tr[ρ log ω] − log tr[σω] =
ω>0
sup tr[ρ log ω] − log tr[σω]
(8)
ω>0
tr[σω]=1
≤ sup tr[ρ log ω] + 1 − tr[σω] ,
(9)
ω>0
where we used that log tr[σω] = 1 − tr[σω] when tr[σω] = 1.
Using this variational expression, we are able to answer a question left open by Donald [12,
Sec. 3] as well as Hiai and Petz [21, Sec. 1], namely whether the two definitions of measured
relative entropy are equal.
Theorem 2. For ρ ∈ S and σ ∈ P, we have D P (ρkσ) = D M (ρkσ).
Proof. The direction ‘≤’ holds trivially. Moreover, if ρ 6≪ σ, we can choose P1 to be a rank-1
projector such that tr[P1 ρ] > 0 and tr[P1 σ] = 0 and thus D P (ρkσ) = D M (ρkσ) = +∞.
It remains to show the direction ‘≥’ when ρ ≪ σ holds. Let (X , M ) be a POVM. Recall that the
distribution Pρ,M is defined by Pρ,M (x) = tr[M (x)ρ]. Introducing X̄ = {x ∈ X : Pρ,M (x)Pσ,M (x) >
0}, we can write
X
Pρ,M (x)
(10)
D(Pρ,M kPσ,M ) =
Pρ,M (x) log
Pσ,M (x)
x∈X̄
X
Pρ,M (x)
(11)
M (x) log
= trρ
Pσ,M (x)
x∈X̄
Xp
p
Pρ,M (x)
(12)
M (x) log
M (x) .
·1
= trρ
Pσ,M (x)
x∈X̄
P
(x)
ρ,M
Now observe that for any x ∈ X̄ , the spectrum of the operator Pσ,M
(x) · 1 is included in (0, ∞).
As a result, we can apply the operator Jensen inequality for the function log, which is operator
concave on (0, ∞) and get
X
Pρ,M (x)
M (x)
D(Pρ,M kPσ,M ) ≤ trρ log
.
(13)
Pσ,M (x)
x∈X̄
Now we simply choose
X
Pρ,M (x)
ω=
M (x)
Pσ,M (x)
x∈X̄
so that
tr[σω] =
X
x∈X̄
Pσ,M (x)
X
Pρ,M (x)
=
Pρ,M (x) = 1 ,
Pσ,M (x)
(14)
x∈X
and which allows to further bound (13) by tr[ρ log ω] + 1 − tr[σω]. Comparing this with the
variational expression for the measured relative entropy in Lemma 1 yields the desired inequality.
Hence, the measured relative entropy, D M (ρkσ), achieves its maximum for projective rank-1
measurements and can be evaluated using the concave optimization problem in Lemma 1.
4
II.
MEASURED RÉNYI RELATIVE ENTROPY
Here we extend the results of the previous section to Rényi divergence. Using the same notation
as in the previous section, for α ∈ (1, ∞) we define the Rényi divergence [40] as
X
P (x) α−1
1
Dα (P kQ) :=
log
P x
(15)
α−1
Q(x)
x∈X
if P ≪ Q and as +∞ if P 6≪ Q. For α ∈ (0, 1) we rewrite the sum as
X
P (x)α Q(x)1−α .
(16)
x∈X
Hence we see that absolute continuity P ≪ Q is not necessary to keep Dα finite for α < 1. However,
the Rényi divergence instead diverges to +∞ when P and Q are orthogonal.1 It is well known that
the Rényi divergence converges to the Kullback-Leibler divergence when α → 1 and we thus set
P (x)
.
D ≡ D1 . Moreover, in the limit α → ∞ we find the max-divergence D∞ (P kQ) = supx∈X log Q(x)
Let us now define the measured Rényi relative entropy as before, namely
(17)
DαM (ρkσ) := sup Dα Pρ,M Pσ,M .
(X ,M )
We will later show that this is equivalent to the following projectively measured Rényi relative
entropy, which we define here for α ∈ (1, ∞) as
)
(
X
1
(18)
tr[Pi ρ]α tr[Pi σ]1−α ,
DαP (ρkσ) :=
log QPα (ρkσ) , with QPα (ρkσ) := sup
α−1
d
{Pi }
i=1
i
and analogously for α ∈ (0, 1) with
QPα (ρkσ) := min
{Pi }di=1
(
X
tr[Pi ρ]α tr[Pi σ]1−α
i
)
.
(19)
Note that the supremum in (18) is achieved and DαP (ρkσ) is finite whenever ρ ≪ σ. Similarly, the
minimum in (19) is non-zero and DαP (ρkσ) is finite whenever ρ 6⊥ σ, i.e. when the two states are
not orthogonal.
Next we give variational expressions for QPα (ρkσ) similar to the variational characterization of
the measured relative entropy in Lemma 1.
Lemma 3. For ρ ∈ S and σ ∈ P, the following identities hold:
h
i
α
α−1
inf
α
tr
ρω
+
(1
−
α)
tr
σω
for α ∈ (0, 12 )
i
h
ω>0
1
inf α tr ρω 1− α + (1 − α) tr[σω] for α ∈ [ 12 , 1)
QPα (ρkσ) = ω>0
i
h
sup α tr ρω 1− α1 + (1 − α) tr[σω] for α ∈ (1, ∞)
ω>0
h
i1−α
α
inf tr [ρω]α tr σω α−1
for α ∈ (0, 1)
= ω>0 h 1− 1 iα
.
sup tr ρω α tr[σω]1−α for α ∈ (1, ∞)
ω>0
The infima and suprema are achieved when ρ > 0 and σ > 0.
1
P and Q are orthogonal, denoted P ⊥ Q, if there exists an A ⊆ X such that P (A) = 1 and Q(A) = 0.
(20)
(21)
5
The expressions (21) can be seen as a generalization of Alberti’s theorem [1] for the fidelity
(which corresponds to α = 1/2) to general α ∈ (0, 1) ∪ (1, ∞).
Proof. We first show the identity (20). Let us discuss the case α ∈ (0, 1) in detail. Note that the
α
two expressions given for α ∈ (0, 12 ) and α ∈ [ 21 , 1) are equivalent by the transformation ω 7→ ω α−1
(the reason for the different ways of writing is to see that the expressions are convex in ω, which
will be useful later, in particular in Theorem 4). We first write
i
h
1
inf α tr ρω 1− α + (1 − α) tr[σω] =
ω>0
inf
inf α
{Pi }di=1 {ωi }di=1
X
1
1− α
ωi
i
tr[Pi ρ] + (1 − α)
X
ωi tr[Pi σ] . (22)
i
Let i ∈ [d] be such that tr[Pi ρ] and tr[Pi σ] are both strictly positive (which is the case when ρ > 0
and σ > 0). Then a local (and thus global) minimum for ωi is easily found at the point where
1 − α1
tr[Pi ρ] α
ω
α 1−
.
(23)
tr[Pi ρ] + (1 − α) tr[Pi σ] = 0
=⇒
ωi =
α i
tr[Pi σ]
If both tr[Pi ρ] = tr[Pi σ] = 0 we can chose ωi arbitrarily. If only tr[Pi ρ] = 0 the infimum is achieved
in the limit ωi → 0, and if only tr[Pi σ] = 0 in the limit ωi → ∞. In all these cases the infimum of
the i-th term is zero. Furthermore, it is achieved for a finite, non-zero ωi when ρ > 0 and σ > 0.
Plugging this solution into the above expression yields
X
tr[Pi ρ]α tr[Pi σ]1−α .
(24)
inf
{Pi }di=1
i∈[d]
This infimum is always achieved since the set we optimize over is compact. Comparing this with
the definition of QPα (ρkσ) yields the first equality.
For the case α ∈ (1, ∞), when ρ ≪ σ, the proof is analogous to the previous argument.
Otherwise, it is simple to see that the supremum is +∞.
Now we show (21). Using (20) and the weighted arithmetic-geometric mean inequality we have
h
i
1 α
QPα (ρkσ) ≥ inf tr ρω 1− α tr [σω]1−α for α ∈ (0, 1) .
(25)
ω>0
However, for any feasible ω > 0 in (20) and λ > 0, λω > 0 is also feasible and choosing λ =
1 α
tr ρω 1− α tr[σω]−α shows that (20) cannot exceed (25). Similarly, by Bernoulli’s inequality,
h
i
1 α
QPα (ρkσ) ≤ sup tr ρω 1− α tr [σω]1−α
ω>0
for α ∈ (1, ∞) .
(26)
And the same argument as above yields the equality.
As for the measured relative entropy, the restriction to rank-1 projective measurements is in
fact not restrictive at all.
Theorem 4. For ρ ∈ S and σ ∈ P, we have DαP (ρkσ) = DαM (ρkσ).
Proof. For α > 1 we follow the steps of the proof of Theorem 2. Consider any finite set X and
POVM M with induced measures Pρ,M and Pσ,M . We can write
Dα (Pρ,M kPσ,M ) =
X
1
log
Pρ,M (x)α Pσ,M (x)1−α ,
1−α
x∈X̄
(27)
6
where we can restrict the sum over X̄ = {x ∈ X : Pρ,M (x)Pσ,M (x) > 0}. We then find that the
sum satisfies
1
α−1
α 1− α
X
X
Pρ,M (x)
Pρ,M (x)
(28)
M (x)
≤ trρ
Pρ,M (x)
,
Pσ,M (x)
Pσ,M (x)
x∈X̄
x∈X̄
where the inequality again follows by the operator Jensen inequality and the operator concavity of
1
the function t 7→ t1− α on [0, ∞). Now we set
X
X
Pρ,M (x) α
Pρ,M (x) α−1
ω=
M (x)
so that tr[σω] =
Pρ,M (x)
.
(29)
Pσ,M (x)
Pσ,M (x)
x∈X
x∈X̄
Thus, we can bound
X
Pρ,M (x)
x∈X
Pρ,M (x)
Pσ,M (x)
α−1
1
≤ α tr ρω 1− α + (1 − α) tr[σω] .
(30)
Comparing this with the variational expression in Lemma 3 yields the desired inequality.
For α < 1, we use the same notation as in (27). We further distinguish the cases α ∈ (0, 12 ) and
α ∈ [ 21 , 1). For α ∈ (0, 12 ), we define
X
Pρ,M (x) α−1
.
(31)
ω=
M (x)
Pσ,M (x)
x∈X̄
We can then evaluate
α
α−1
h
i
α
Pρ,M (x) α−1
X
α−1
tr σω
M (x)
= trσ
Pσ,M (x)
(32)
x∈X̄
≤ trσ
X
x∈X̄
M (x)
Pρ,M (x)
Pσ,M (x)
α
α−1
α
=
X
Pρ,M (x)α Pσ,M (x)1−α .
(33)
x∈X̄
where we used the operator convexity of t 7→ t
on (0, ∞) and the operator Jensen inequality.
Moreover,
X
tr[ρω] =
Pρ,M (x)α Pσ,M (x)1−α .
(34)
x∈X̄
As a result
X
x∈X̄
α
Pρ,M (x)α Pσ,M (x)1−α ≥ α tr[ρω] + (1 − α) tr[σω α−1 ] .
(35)
Comparing this with the variational expression
3 yields the desired inequality.
in Lemma
P
Pρ,M (x) α
1
For α ∈ [ 2 , 1) we choose ω = x∈X̄ M (x) Pσ,M (x) , so that
h
i X
1
Pρ,M (x) α−1
tr ρω 1− α ≤
Pρ,M (x)
Pσ,M (x)
x∈X̄
X
tr [σω] =
Pρ,M (x)α Pσ,M (x)1−α ,
x∈X̄
and once again conclude using the variational expression in Lemma 3.
(36)
(37)
7
III.
ACHIEVABILITY OF RELATIVE ENTROPY
A.
Umegaki’s Relative Entropy
Here we compare the measured relative entropy to other notions of quantum relative entropy
that have been investigated in the literature and have found operational significance in quantum
information theory. Umegaki’s quantum relative entropy [50] has found operational significance as
the threshold rate for asymmetric binary quantum hypothesis testing [21]. For ρ ∈ S and σ ∈ P,
it is defined as
D(ρkσ) := tr ρ(log ρ − log σ) if σ ≫ ρ and as +∞ if σ 6≫ ρ.
(38)
We recall the following variational expression by Petz [36] (see also [25] for another variational
expression):
D(ρkσ) = sup tr[ρ log ω] − log tr[exp(log σ + log ω)]
(39)
ω>0
= sup tr[ρ log ω] + 1 − tr[exp(log σ + log ω)] .
(40)
ω>0
By the data-processing inequality for the relative entropy [29, 49] and Theorem 2 we always have
D P (ρkσ) = D M (ρkσ) ≤ D(ρkσ) ,
(41)
and moreover Petz [35] showed the inequality D P (ρkσ) ≤ D(ρkσ) is strict if ρ and σ do not commute
(for ρ > 0 and σ > 0). Theorem 2 strengthens this to show that the strict inequality persists even
when we take a supremum over POVMs. In the following we give an alternative proof of Petz’
result and then extend the argument to Rényi relative entropy in Section III B.
Proposition 5. Let ρ ∈ S with ρ > 0 and σ ∈ P with σ > 0. Then, we have
D M (ρkσ) ≤ D(ρkσ)
with equality if and only if
[ρ, σ] = 0 .
(42)
Our proof relies on the Golden–Thompson inequality [19, 46]. It states that for two Hermitian
matrices X and Y , it holds that
tr[exp(X + Y )] ≤ tr[exp(X) exp(Y )]
(43)
with equality if an only if [X, Y ] = 0 as shown in [43].
Proof. First, it is evident that equality holds if [ρ, σ] = 0 since then there exists a projective
measurement that commutes with ρ and σ and thus does not effect the states. For the following,
it is worth writing the variational expressions for the two quantities side by side. Namely, writing
H = log ω, we have
D(ρkσ) = 1 + sup tr[ρH] − tr[exp(log σ + H)]
(44)
H
D M (ρkσ) = 1 + max tr[ρH] − tr[σ exp(H)] ,
H
(45)
where we optimize over all Hermitian operators H. Note that, according to Lemma 1, we can write
a max for (45) because we are assuming ρ > 0 and σ > 0. The inequality in (42) can now be seen
as a direct consequence of the Golden–Thompson inequality.
8
It remains to show that D M (ρkσ) = D(ρkσ) implies [ρ, σ] = 0. Let H ∗ be any maximizer of the
variational problem in (45). Observe now that the equality D M (ρkσ) = D(ρkσ) necessitates
tr[exp(log σ + H ∗ )] = tr[σ exp(H ∗ )] ,
(46)
which holds only if and only if [σ, H ∗ ] = 0 by the equality condition in (43). Now define the
function
f (H) = tr[ρH] − tr[σ exp(H)] ,
(47)
and since H ∗ maximizes f (H), we must have for all Hermitian ∆,
0 = Df (H ∗ )[∆] = tr[ρ∆] − tr[σ exp(H ∗ )∆] .
(48)
To evaluate the second summand of this Fréchet derivative we used that σ and H ∗ commute. Since
this holds for all ∆ we must in fact have ρ = σ exp(H ∗ ), which means that [ρ, σ] = 0, as desired.
In some sense this result tells us that some quantum correlations, as measured by the relative
entropy, do not survive the measurement process. This fact appears in quantum information theory in various different guises, for example in the form of locking classical correlations in quantum
states [11]. (We also point to [38] for the use of measured relative entropy measures in quantum
information theory.) Moreover, since Umegaki’s relative entropy is the smallest quantum generalization of the Kullback-Leibler divergence that is both additive and satisfies data-processing (see,
e.g., [47, Sec. 4.2.2]), the same conclusion can be drawn for any sensible quantum relative entropy.
(An example being the quantum relative entropy introduced by Belavkin and Staszewski [4].)
B.
Sandwiched Rényi Relative Entropy
Next we consider a family of quantum Rényi relative entropies [33, 55] that are commonly called
sandwiched Rényi relative entropies and have found operational significance since they determine
the strong converse exponent in asymmetric binary quantum hypothesis testing [31]. They are of
particular interest here because they are, for α ≥ 21 , the smallest quantum generalization of the
Rényi divergence that is both additive and satisfies data-processing [47, Sec. 4.2.2]. (Examples
for other sensible quantum generalizations are the quantum Rényi relative entropy first studied by
Petz [34] and the quantum divergences introduced by Matsumoto [30].)
For ρ ∈ S and σ ∈ P, the sandwiched Rényi relative entropy of order α ∈ (0, 1) ∪ (1, ∞) is
defined as
h 1−α 1−α α i
1
Dα (ρkσ) :=
log Qα (ρkσ) with Qα (ρkσ) := tr σ 2α ρσ 2α
,
(49)
α−1
where the same considerations about finiteness as for the measured Rényi relative entropy apply.
We also consider the limits α → ∞ and α → 1 of the above expression for which we have [33],
D∞ (ρkσ) = inf {λ ∈ R|ρ ≤ exp(λ)σ}
and
D1 (ρkσ) = D(ρkσ) ,
respectively. We recall the following variational expression by Frank and Lieb [17]:
h 1 α−1 1 α i
α−1
inf α tr[ρω] + (1 − α) tr ω 2 σ α ω 2
for α ∈ (0, 1)
ω>0
h 1 α−1 1 α i
Qα (ρkσ) =
sup α tr[ρω] + (1 − α) tr ω 2 σ α ω 2 α−1
for α ∈ (1, ∞) .
ω>0
(50)
(51)
9
Alternatively, we can also write
h 1 α−1 1 α i1−α
α−1
inf tr[ρω]α tr ω 2 σ α ω 2
ω>0
h 1 α−1 1 α i1−α
Qα (ρkσ) =
sup tr[ρω]α tr ω 2 σ α ω 2 α−1
ω>0
for α ∈ (0, 1)
(52)
for α ∈ (1, ∞) ,
where we have used the same arguments as in the proof of the second part of Lemma 3. By the
data-processing inequality for the sandwiched Rényi relative entropy [3, 17, 33] we always have
1
.
(53)
2
In the following we give an alternative proof of this fact and show that
1
M
[ρ, σ] 6= 0
=⇒
Dα (ρkσ) < Dα (ρkσ) for α ∈
,∞ .
(54)
2
In contrast, at the boundaries α ∈ 21 , ∞ it is known that Dα (ρkσ) = DαM (ρkσ) [18, 24, 31]. (We
refer to [31, App. A] for a detailed discussion.)
Theorem 6. Let ρ ∈ S with ρ > 0 and σ ∈ P with σ > 0. For α ∈ 21 , ∞ , we have
DαM (ρkσ) ≤ Dα (ρkσ)
DαM (ρkσ) ≤ Dα (ρkσ)
for α ≥
with equality if and only if
[ρ, σ] = 0 .
(55)
The argument is similar to the proof of Proposition 5 but with the Golden–Thompson inequality
replaced by the Araki–Lieb–Thirring inequality [2, 28]. It states that for X, Y ≥ 0 we have
tr [Y XY ]r ≤ tr [Y r X r Y r ] for r ≥ 1
r
r
r
r
tr [Y XY ] ≥ tr [Y X Y ] for r ∈ [0, 1] ,
(56)
(57)
with equality if and only if [X, Y ] = 0 except for r = 1 as shown in [20].
Proof. We give the proof for α ∈ 21 , 1 and note that the argument for α ∈ (1, ∞) is analogous.
We have the following variational expressions from Lemma 3 and (51):
h 1 1−α
α i
1
1−α
Qα (ρkσ) = inf α tr[ρω] + (1 − α) tr ω − 2 σ α ω − 2
(58)
ω>0
i
h
α
(59)
QPα (ρkσ) = min α tr[ρω] + (1 − α) tr σω α−1 ,
ω>0
where the existence of the minima relies on the fact that both operators have full support. (Note
α
≥ 1, the inequality
also that these two expressions are in fact equivalent for α = 21 .) Since 1−α
then follows immediately by the Araki–Lieb–Thirring inequality (56):
QPα (ρkσ) ≥ Qα (ρkσ)
=⇒
DαM (ρkσ) ≤ Dα (ρkσ) .
(60)
Furthermore, if [ρ, σ] = 0 we have equality. To show that QPα (ρkσ) = Qα (ρkσ) implies [ρ, σ] = 0,
we define the function
i
h
α
(61)
fα (ω) = α tr[ρω] + (1 − α) tr σω α−1 .
For ωα∗ any minimizer of the variational problem in (59), we have
i
h
1
0 = Dfα(ωα∗ )[∆] = α tr[ρ∆] − α tr σ (ωα∗ ) α−1 ∆ ,
(62)
for all Hermitian ∆. To evaluate the second summand of this Fréchet derivative we used that σ and
ωα∗ commute, which holds by the equality condition for Araki–Lieb–Thirring. We thus conclude
1
that ρ = σ (ωα∗ ) α−1 which implies that [ρ, σ] = 0.
10
IV.
VIOLATION OF DATA-PROCESSING FOR α <
1
2
As a further application of the variational characterization of measured Rényi relative entropy,
we can show that the data-processing for the sandwiched Rényi relative entropy Dα fails for α < 12 .
(Numerical evidence pointed to the fact that data-processing does not hold in this regime [32].)
Theorem 7. Let ρ ∈ S with ρ > 0 and σ ∈ P with σ > 0, and [ρ, σ] 6= 0. For α ∈ 0, 12 , we have
DαP (ρkσ) > Dα (ρkσ).
In particular, there exists a rank-1 measurement that achieves DαP (ρkσ) and thus violates the
data-processing inequality.
Proof. First note that [ρ, σ] 6= 0 implies that the two states are not orthogonal and thus both
quantities are finite. For α ∈ (0, 21 ) the formulas (58) and (59) still hold. However, in contrast
α
to the proof of Theorem 6 we have 1−α
∈ [0, 1]. Hence, we find by the Araki–Lieb–Thirring
inequality (57) that
QPα (ρkσ) ≤ Qα (ρkσ)
=⇒
DαP (ρkσ) ≥ Dα (ρkσ) .
(63)
As in the proof of Theorem 6 we have equality if and only if [ρ, σ] = 0. This implies the claim.
V.
A.
EXPLOITING VARIATIONAL FORMULAS
Some Optimization Problems in Quantum Information
The variational characterizations of the relative entropy (39)–(40), the sandwiched Rényi relative entropy (51)–(52), and their measured counterparts (Lemma 1 and Lemma 3), can be used
to derive properties of various entropic quantities that appear in quantum information theory. We
are interested in operational quantities of the form
M(ρ) := min D(ρkσ) ,
σ∈C
(64)
where D(·k·) stands for any relative entropy, measured relative entropy, or Rényi variant thereof,
and C denotes some convex, compact set of states. For Umegaki’s relative entropy D ≡ D, prominent examples for C include the set of
• separable states, giving rise to the relative entropy of entanglement [53].
• positive partial transpose states, leading to the Rains bound on entanglement distillation [39].
• non-distillable states, leading to bounds on entanglement distillation [52].
• quantum Markov states, leading to insights about the robustness properties of these
states [22].
• locally recoverable states, leading to bounds on the quantum conditional mutual information [8, 16, 41].
• k-extendible states, leading to bounds on squashed entanglement [27].
Other examples are conditional Rényi entropies which are defined by optimizing the sandwiched
Rényi relative entropy over a convex set of product states with a fixed marginal, see, e.g., [48].
11
A central question is what properties of the underlying relative entropy D translate to properties
of the induced measure M(·). For example, all the relative entropies discussed in this paper are
superadditive on tensor product states in the sense that
D(ρ1 ⊗ ρ2 kσ1 ⊗ σ2 ) ≥ D(ρ1 kσ1 ) + D(ρ2 kσ2 ) .
(65)
We might then ask if we also have
?
min D(ρ1 ⊗ ρ2 kσ12 ) = M(ρ1 ⊗ ρ2 ) ≥ M(ρ1 ) + M(ρ2 ) = min D(ρ1 kσ1 ) + min D(ρ2 kσ2 ) . (66)
σ12 ∈C12
σ1 ∈C1
σ2 ∈C2
To study questions like this we propose to make use of the variational characterizations of the form
D(ρkσ) = sup f (ρ, σ, ω)
ω>0
in order to write M(ρ) = min sup f (ρ, σ, ω) = sup min f (ρ, σ, ω) , (67)
σ∈C ω>0
ω>0 σ∈C
where we made use of Sion’s minimax theorem [42] for the last equality. We note that the conditions
of the minimax theorem are often fulfilled. The minimization over σ ∈ C then typically simplifies
and is a convex or even semidefinite optimization. (As an example, for the measured relative
entropies the objective function becomes linear in σ.) We can then use strong duality of convex
optimization to rewrite this minimization as a maximization problem [7]:
min f (ρ, σ, ω) = max f¯(ρ, σ̄, ω) .
(68)
M(ρ) = sup max f¯(ρ, σ̄, ω) ,
(69)
σ∈C
σ̄∈C̄
This leads to the expression
ω>0 σ̄∈C¯
which, in contrast to the definition of M(ρ) in (64), only involves maximizations. This often gives
useful insights about M(ρ). As an example, let us come back to the question of superadditivity
raised in (66). We want to argue that the following two conditions on f¯ and C¯ imply superadditivity.
First, we need that the function f¯ is superadditive itself, i.e. we require that
f¯(ρ1 ⊗ ρ2 , σ̄1 ⊗ σ̄2 , ω1 ⊗ ω2 ) ≥ f¯(ρ1 , σ̄1 , ω1 ) + f¯(ρ2 , σ̄2 , ω2 ) .
(70)
And second, we require that the sets C¯ are closed under tensor products in the sense that σ̄1 ∈ C¯1
and σ̄2 ∈ C¯2 imply that σ̄1 ⊗ σ̄2 ∈ C¯12 . Using these two properties, we deduce that
M(ρ1 ⊗ ρ2 ) ≥ f¯(ρ1 ⊗ ρ2 , σ̄1 ⊗ σ̄2 , ω1 ⊗ ω2 ) ≥ f¯(ρ1 , σ̄1 , ω1 ) + f¯(ρ2 , σ̄2 , ω2 )
(71)
for any ω1 , ω2 > 0 and any σ̄1 ∈ C¯1 , σ̄2 ∈ C¯2 . Hence, the inequalities also hold true if we maximize
over these variables, implying superadditivity.
B.
Relative Entropy of Recovery
In the following we denote multipartite quantum systems using capital letters, e.g., A, B, C.
The set of density operators on A and B is then denoted S(AB), for example. We also use these
letters as subscripts to indicate which systems operators act on.
As a concrete application, we study the additivity properties of the relative entropy of recovery
defined as [8, 10, 41]
D rec (ρAD kσAE ) := inf D(ρAD k(IA ⊗ ΓE→D )(σAE )) ,
ΓE→D
(72)
12
where the infimum is taken over all trace-preserving completely positive maps ΓE→D . (We restrict
to σA > 0 such that the quantity is surely finite and the infimum becomes a minimum.) One
motivation for studying the additivity properties of the relative entropy of recovery is the study
of lower bounds on the quantum conditional mutual information and strengthenings of the dataprocessing inequality for the relative entropy [6, 8, 16, 23, 27, 41, 44, 45, 54]. In particular, [8,
Prop. 3] shows that
1 rec ⊗n
,
D
ρABC kρ⊗n
AB
n→∞ n
I(A : C|B) ≥ lim
(73)
where the systems are understood as D = BC and E = B. To obtain a lower bound that does not
involve limits, the authors of [8] use a Finetti-type theorem to show that
lim
n→∞
1 rec ⊗n
M,rec
(ρABC kρAB ) .
D (ρABC kρ⊗n
AB ) ≥ D
n
(74)
with the measured relative entropy of recovery defined as [8] (see also [41, Rmk. 6]),
D M,rec(ρAD kσAE ) := inf D M (ρAD k(IA ⊗ ΓE→D )(σAE )) .
ΓE→D
(75)
(We restrict to σA > 0 such that the quantity is surely finite and the infimum becomes a minimum.)
This gives an interpretation for the conditional mutual information in terms of recoverability.
The measured relative entropy is superadditive and if this property would translate to D M,rec we
would get an alternative proof for the step (74). Using the variational expression for the measured
relative entropy (Lemma 1) together with strong duality for semidefinite programming, we find the
following alternative characterization for D M,rec .
Lemma 8. Let ρAD ∈ S(AD) and σAE ∈ S(AE) with σA > 0, and let σAEF be a purification of
σAE . Then, we have
D M,rec (ρAD kσAE ) = maximize
tr[ρAD log RAD ]
subject to SAF > 0, RAD > 0
1D ⊗ SAF ≥ RAD ⊗ 1F
(76)
tr[SAF σAF ] = 1 .
Proof. We write
h
i
√
√
ΓE→D (σAE ) = trF ΓE→D (σAEF ) = trF ΓE→D σAF ΨAF :E σAF ,
(77)
where we denote by ΨAF :E the (unnormalized) maximally entangled state between AF : E in the
basis of the Schmidt decomposition of σAEF with respect to AF : E. With this we define the
Choi-Jamiolkowski state (unnormalized) of the map ΓE→D as
(78)
τADF = ΓE→D ΨAF :E , τAF = ΠσAF ,
where ΠσAF denotes the projector onto the support of σAF . Hence, we can write
√
√
σAF τADF σAF ,
ΓE→D (σAE ) = trF
(79)
and thus we can express the optimization problem for ΓE→D in terms of the Choi-Jamiolkowski
state in (78). Together with the variational characterization of the measured relative entropy
(Lemma 1) we find
√
√
D M,rec (ρAD kσAE ) =
min
sup tr[ρAD log RAD ] + 1 − tr [τADF σAF RAD σAF ] , (80)
τADF ∈Rec(σAE ) RAD >0
13
where Rec(σAE ) := {τADF ≥ 0, τAF = ΠσAF }. We now apply Sion’s minimax theorem [42] to
exchange the minimum with the supremum. The theorem is applicable as Rec(σAE ) is convex and
compact and {ωAD > 0} is convex. Moreover, as the logarithm is operator concave, the function
√
√
(81)
RAD 7→ tr[ρAD log RAD ] + 1 − tr [τADF σAF RAD σAF ]
is concave for any fixed τADF . Finally, for any fixed RAD , the function being optimized is linear
on τADF . As a result,
√
√
D M,rec (ρAD kσAE ) = sup tr[ρAD log RAD ] + 1 −
max
tr [τADF σAF RAD σAF ] . (82)
τADF ∈Rec(σAE )
RAD >0
Now for a fixed RAD > 0, the inner maximization is a semidefinite program for which we can write
the following programs:
√
√
minimize:
tr[SAF σAF ]
maximize:
tr τADF σAF RAD σAF
subject to:
τADF ≥ 0
subject to:
1D ⊗ SAF ≥ RAD ⊗ 1F .
(83)
σ
τAF = ΠAF
As the primal problem is strictly feasible, Slater’s condition (see, e.g., [7]) is satisfied and we have
strong duality. This leads to the expression:
D M,rec (ρAD kσAE ) = maximize
tr[ρAD log RAD ] + 1 − tr[SAF σAF ]
subject to SAF > 0, RAD > 0
(84)
1D ⊗ SAF ≥ RAD ⊗ 1F .
Observe now that we can restrict the optimization to tr[SAF σAF ] = 1. In fact, for arbitrary feasible
RAD and SAF , we can define
R̃AD =
RAD
tr[SAF σAF ]
and S̃AF =
SAF
.
tr[SAF σAF ]
(85)
The constraint 1D ⊗ SAF ≥ RAD ⊗ 1F is still satisfied and the value of the objective function can
only increase,
tr[ρAD log R̃AD ] = tr[ρAD log RAD ] − tr[ρAD ] log tr[SAF σAF ]
≥ tr[ρAD log RAD ] − tr[SAF σAF ] + 1 ,
(86)
(87)
where we used the inequality log x ≤ x − 1 and that tr[ρAD ] = 1. This concludes the proof.
This readily implies that D M,rec is indeed superadditive.
Proposition 9. Let ρAD ∈ S(AD), τA′ D′ ∈ S(A′ D ′ ), σAE ∈ S(AE), and ωA′ E ′ ∈ S(A′ E ′ ) with
σA , ωA′ > 0. Then, we have
D M,rec(ρAD ⊗ τA′ D′ kσAE ⊗ ωA′ E ′ ) ≥ D M,rec (ρAD kσAE ) + D M,rec (τA′ D′ kωA′ E ′ ) .
(88)
Proof. Given feasible operators RAD , SAF for the quantity D M,rec (ρAD kσAE ) and feasible operators
RA′ D′ , SA′ F ′ for the quantity D M,rec(τA′ D′ kωA′ E ′ ), we have
1D ⊗ SAF ≥ RAD ⊗ 1F ∧ 1D′ ⊗ SA′ F ′ ≥ RA′ D′ ⊗ 1F ′
=⇒
1DD′ ⊗ SAF ⊗ SA′ F ′ ≥ RAD ⊗ RA′ D′ ⊗ 1F F ′ .
(89)
14
Here we used that A ≥ B =⇒ M ⊗ A ≥ M ⊗ B for M ≥0, which holds since taking the
tensor
product with M is a positive map. Moreover, we have tr (SAF ⊗ SA′ F ′ )(σAF ⊗ σA′ F ′ ) = 1. In
other words, (RAD ⊗ RA′ D′ , SAF ⊗ SA′ F ′ ) is a feasible pair in the expression (76) for D M,rec (ρAD ⊗
τA′ D′ kσAE ⊗ ωA′ E ′ ) and we get
D M,rec (ρAD ⊗ τA′ D′ kσAE ⊗ ωA′ E ′ ) ≥ tr [(ρAD ⊗ τA′ D′ ) log (RAD ⊗ RA′ D′ )]
= tr [ρAD log RAD ] + tr [τA′ D′ log RA′ D′ ] .
(90)
(91)
Taking the supremum over feasible (RAD , SAF ) and (RA′ D′ , SA′ F ′ ), we get the claimed superadditivity.
Now, if the relative entropy of recovery would be superadditive we could strengthen (74) to
?
1 rec ⊗n
) ≥ D rec (ρABC kρAB ) .
(92)
D (ρABC kρ⊗n
AB
n→∞ n
This together with (73) would lead to a stronger lower bound on the conditional mutual information.
Using strong duality for convex programming we can write D rec in a dual form similar to Lemma 8.
lim
Lemma 10. Let ρAD ∈ S(AD) and σAE ∈ S(AE) with σA > 0, and let σAEF be a purification of
σAE . Then, we have
D rec (ρAD kσAE ) = maximize
tr[ρAD log ρAD ] − D M (ρAD kRAD )
subject to SAF > 0, RAD > 0
1D ⊗ SAF ≥ RAD ⊗ 1F
(93)
tr[SAF σAF ] = 1 ,
where RAD is not normalized and the measured relative entropy term is in general negative.
We provide a proof in Appendix A. This is to be compared to our expression for the measured
relative entropy of recovery from Lemma 8, which can be written as
D M,rec(ρAD kσAE ) = maximize
tr[ρAD log ρAD ] − D(ρAD kRAD )
subject to SAF > 0, RAD > 0
1D ⊗ SAF ≥ RAD ⊗ 1F
(94)
tr[SAF σAF ] = 1 ,
where RAD is not normalized and the relative entropy term is in general negative. Unfortunately, we
can not solve the open additivity question for the relative entropy of recovery using the variational
expression from Lemma 10. However, we note that the argument flipped relative entropy of recovery
D̄ rec (σAE kρAD ) := inf D((IA ⊗ ΓE→D )(σAE )kρAD )
ΓE→D
(95)
becomes additive (restricted to ρA > 0 to assure finiteness and making the infimum a minimum).
Proposition 11. Let ρAD ∈ S(AD) and σAE ∈ S(AE) with ρA > 0, and let σAEF be a purification
of σAE . Then, we have
D̄ rec (σAE kρAD ) = maximize
− tr[SAF σAF ]
subject to RAD > 0
1D ⊗ SAF ≥ (log ρAD − log RAD ) ⊗ 1F
(96)
tr[RAD ] = 1 .
Moreover, for τA′ D′ ∈ S(A′ D ′ ) with τA′ > 0 and ωA′ E ′ ∈ S(A′ E ′ ), we have
D̄ rec (σAE ⊗ ωA′ E ′ kρAD ⊗ τA′ D′ ) = D̄ rec (σAE kρAD ) + D̄ rec (ωA′ E ′ kτA′ D′ ) .
We provide a proof based on strong duality for convex programming in Appendix A.
(97)
15
C.
Rényi Relative Entropy of Recovery
As a generalization of the measured relative entropy of recovery (75) we define the measured
Rényi relative entropy of recovery for α ∈ (1, ∞) as (see also [41, Rmk. 6]),
1
log QM,rec
(ρAD kσAE ) ,
α
α−1
QM,rec
(ρAD kσAE ) := inf QM
α
α (ρAD k(IA ⊗ ΓE→D )(σAE )) ,
DαM,rec(ρAD kσAE ) :=
with
(98)
(99)
ΓE→D
and analogously for α ∈ (0, 1) with
QM,rec
(ρAD kσAE ) := sup QM
α (ρAD k(IA ⊗ ΓE→D )(σAE )) ,
α
(100)
ΓE→D
where we restrict to σA > 0 such that the quantity is surely finite and the infimum/supremum
is achieved. Using the variational expression for the measured Rényi relative entropy (Lemma 3)
together with strong duality for semidefinite programming, we find the following alternative characterization for QM,rec
.
α
Lemma 12. Let ρAD ∈ S(AD) and σAE ∈ S(AE) with σA > 0, and let σAEF be a purification of
σAE . For α ∈ [ 12 , 1), we have
QM,rec
(ρAD kσAE )
α
= minimize
tr
1− 1
ρAD RADα
α
tr[SAF σAF ]1−α
(101)
subject to SAF > 0, RAD > 0
1D ⊗ SAF ≥ RAD ⊗ 1F .
Similar dual formulas hold for α ∈ (0, 21 ) ∪ (1, ∞).
Proof. As in the proof of Lemma 8 and using the first variational formula from Lemma 3, we get
from Sion’s minimax theorem [42] that
QM,rec
(ρAD kσAE ) =
α
i
h
√
√
1− 1
inf α tr ρAD RADα + (1 − α) tr [τADF σAF RAD σAF ]
max
τADF ∈Rec(σAE ) RAD >0
(102)
=
inf
max
RAD >0 τADF ∈Rec(σAE )
h
1
1− α
α tr ρAD RAD
i
√
√
+ (1 − α) tr [τADF σAF RAD σAF ] .
(103)
The set Rec(σAE ) is convex and compact and {RAD > 0} is convex. Moreover, the function
h
i
√
√
1− 1
RAD 7→ α tr ρAD RADα + (1 − α) tr [τADF σAF RAD σAF ]
(104)
is convex for any fixed τADF because of the operator convexity of t 7→ tβ with β = 1 − α1 for
α ∈ [ 12 , 1) (with β ∈ [−1, 0)). Finally, for a fixed RAD , the function being optimized is linear on
τADF . As in (83) we then look at the convex dual of
max
τADF ∈Rec(σAE )
√
√
tr [τADF σAF RAD σAF ] ,
(105)
16
and find
QM,rec
(ρAD kσAE )
α
= minimize
α tr
1− 1
ρAD RADα
+ (1 − α) tr[SAF σAF ]
subject to SAF > 0, RAD > 0
(106)
1D ⊗ SAF ≥ RAD ⊗ 1F .
Invoking the arithmetic-geometric mean inequality as in the proof of Lemma 3 establishes the
claim.
We find that the measured Rényi entropy of recovery is superadditive for all Rényi parameters.
Proposition 13. Let ρAD ∈ S(AD), τA′ D′ ∈ S(A′ D ′ ), σAE ∈ S(AE), and ωA′ E ′ ∈ S(A′ E ′ ) with
σA , ωA′ > 0. For α ∈ (0, 1) ∪ (1, ∞), we have
DαM,rec (ρAD ⊗ τA′ D′ kσAE ⊗ ωA′ E ′ ) ≥ DαM,rec(ρAD kσAE ) + DαM,rec(τA′ D′ kωA′ E ′ ) .
(107)
Proof. We give the argument for α ∈ [ 21 , 1) based on Lemma 12 and note that the proof for
α ∈ (0, 12 ) ∪ (1, ∞) is similar. Given feasible operators RAD , SAF for the quantity QM,rec
(ρAD kσAE )
α
M,rec
and feasible operators RA′ D′ , SA′ F ′ for the quantity Qα (τA′ D′ kωA′ E ′ ), we apply exactly the same
argument as in (89) and find
(ρAD kσAE ) · QM,rec
(τA′ D′ kωA′ E ′ ) .
QM,rec
(ρAD ⊗ τA′ D′ kσAE ⊗ ωA′ E ′ ) ≤ QM,rec
α
α
α
(108)
This concludes the proof.
Analogously, the relative entropy of recovery (72) is generalized to the Rényi relative entropy
of recovery as (see also [41, Rmk. 6]),
Dαrec (ρAD kσAE ) :=
with
Qrec
α (ρAD kσAE ) :=
1
log Qrec
α (ρAD kσAE )
α−1
inf Qα (ρAD k(IA ⊗ ΓE→D )(σAE ))
ΓE→D
sup Qα (ρAD k(IA ⊗ ΓE→D )(σAE ))
ΓE→D
(109)
for α ∈ (1, ∞)
for α ∈ (0, 1) ,
(110)
where we restrict to σA > 0 such that the quantity is surely finite and the infimum/supremum is
achieved. Now, for the special case of α = 1/2, the sandwiched Rényi relative entropy is equal to
the negative logarithm of the fidelity and as the measured fidelity is the same as the fidelity (see,
e.g., [18, Sec. 3.3]), we have
QM
1/2 = Q1/2
as well as
M,rec
Qrec
1/2 = Q1/2 .
(111)
The same holds true at α → ∞ [31]. As such Proposition 13 can be seen as a generalization
of the multiplicativity of the fidelity of recovery [41] that was derived by two of the authors [6].
For the general sandwiched Rényi relative entropy of recovery, we only have preliminary numerics
indicating that they are not additive. Using a non-commutative extension of the techniques from [5,
Sec. 3.3.1], we can write the sandwiched Rényi relative entropy with fractional Rényi parameter
as a semidefinite program. Together with the Choi state as in (79), we then get a semidefinite
rec . Using this, we found weak numerical evidence for additivity violations for
program for, e.g., D2/3
small dimensional systems. This also challenges the possible additivity of the relative entropy of
recovery D rec (corresponding to the case α = 1).2
2
A very recent preprint [15] has demonstrated by numerical examples that the relative entropy of recovery is indeed
non-additive.
17
VI.
CONCLUSION
We presented variational characterizations of the measured relative entropy and the measured
Rényi relative entropy. Using these formulas we were able to show that these quantities can be
achieved by rank-1 projective measurements instead of general POVMs. We also showed that the
measured Rényi relative entropy is equal to the corresponding sandwiched Rényi relative entropy if
and only if the two quantum states commute (except for the special cases with Rényi parameters 1/2
and ∞ where they are always equal), and gave analytical counterexamples for the data-processing
inequality for the sandwiched Rényi relative entropy with Rényi parameters smaller than 1/2.
Finally, we applied our variational characterizations to analyze the additivity properties of various
relative entropies of recovery.
As an extension of our work it would be desirable to weaken the support conditions in Proposition 5 and Theorem 6 concerning the equality conditions for measured relative entropy and
relative entropy. (We note that this is non-trivial because we make use of the equality conditions for Golden-Thompson and Araki–Lieb–Thirring). Moreover, it would be neat to extend the
concept of measured relative entropy to general, continuous POVMs described by measure spaces
(see, e.g., [51] for the definition of Kullback-Leibler divergence and Rényi divergence on measure
spaces). Many of the steps still go through and we would like to point to a Jensen inequality for
operator-valued measures [14]. It seems, however, that an extension of this inequality would be
needed [13]. Finally, it would also be interesting to use the variational characterizations presented
in this work for studying the operational entropic quantities mentioned in Section V A. This might
lead to new applications of measured relative entropy in quantum information theory.
Acknowledgments: We acknowledge discussions with Fernando Brandão, Douglas Farenick and
Hamza Fawzi. MB acknowledges funding by the SNSF through a fellowship, funding by the
Institute for Quantum Information and Matter (IQIM), an NSF Physics Frontiers Center (NFS
Grant PHY-1125565) with support of the Gordon and Betty Moore Foundation (GBMF-12500028),
and funding support form the ARO grant for Research on Quantum Algorithms at the IQIM
(W911NF-12-1-0521). Most of this work was done while OF was also with the Department of
Computing and Mathematical Sciences, California Institute of Technology. MT would like to
thank the IQIM at CalTech and John Preskill for his hospitality during the time most of the
technical aspects of this project were completed. He is funded by an ARC Discovery Early Career
Researcher Award fellowship (Grant No. DE160100821).
Appendix A: Proofs for Results in Section V
Proof of Lemma 10. We start by writing out the relative entropy of recovery as a convex optimization program,
D rec (ρAD kσAE ) = tr[ρAD log ρAD ]+ minimize
− tr[ρAD log γAD ]
√
√
subject to γAD = trF [ σAF τADF σAF ]
τADF ≥ 0
(A1)
τAF = ΠσAF ,
where the notation is as
of Lemma 8. Clearly, the first and last constraint can be
√in the proof
√
σAF τADF σAF and τAF ≤ ΠσAF without changing the solution. The
relaxed to γAD ≤ trF
18
Lagrangian can then be written as
√
√
L(γ, τ, R, S) = − tr[ρAD log γAD ] + tr [RAD (γAD − trF [ σAF τADF σAF ])]
+ tr [SAF (τAF − ΠσAF )]
= − tr[ρAD log γAD ] + tr[RAD γAD ] − tr [τADF
− tr [SAF ΠσAF ] ,
√
√
( σAF RAD σAF − SAF )]
(A2)
(A3)
where we introduced the variables RAD ≥ 0 and SAF ≥ 0. In order to compute the dual objective
function, we should take the infimum of this quantity over γAD > 0 and τADF ≥ 0. Using the
variational expression for the measured relative entropy (Lemma 1) we find
inf
γAD >0, τADF ≥0
L(γ, θ, R, S) = −D M (ρAD kRAD ) + 1 − tr[SAF ΠσAF ] ,
(A4)
√
√
when SAF ≥ σAF RAD σAF . With Slater’s strong duality and a change of variable S̄AF :=
−1/2
−1/2
σAF SAF σAF (but not including the bar in the following), we get
D rec (ρAD kσAE ) = tr[ρAD log ρAD ]+ maximize
− D M (ρAD kRAD ) + 1 − tr[SAF σAF ]
subject to SAF > 0, RAD > 0
(A5)
1D ⊗ SAF ≥ RAD ⊗ 1F .
Adding the constraint tr[SAF σAF ] = 1 as in the proof of Lemma 8 concludes the proof.
Proof of Proposition 11. We start by writing out the argument flipped relative entropy of recovery
as a convex optimization program,
D̄ rec (σAE kρAD ) = minimize
tr[γAD log γAD ] − tr[γAD log ρAD ]
√
√
subject to γAD = trF [ σAF τADF σAF ]
τADF ≥ 0
(A6)
τAF = ΠσAF ,
where the notation is as in the proof of Lemma 8. The Lagrangian can be written as
√
√
L(γ, τ, T, S) = tr[γAD log γAD ] − tr[γAD log ρAD ] + tr [TAD (γAD − trF [ σAF τADF σAF ])]
+ tr [SAF (τAF − ΠσAF )]
= tr[γAD log γAD ] + tr[γAD (TAD − log ρAD )]
√
√
− tr [SAF ΠσAF ] − tr [τADF ( σAF TAD σAF − SAF )] ,
(A7)
(A8)
where TAD and SAF are Hermitian operators. In order to compute the dual objective function, we
should take the infimum of this quantity over γAD > 0 and τADF ≥ 0. From the last expression
√
√
we get SAF ≥ σAF TAD σAF and we also know how to optimize the first expression as it is an
entropy maximization question:
inf tr[γAD log γAD ] + tr[γAD (TAD − log ρAD )] .
γAD >0
(A9)
It is optimized when γAD = exp(log ρAD − TAD + α1AD ) for some α [9]. This means that this
infimum is given by
inf tr[exp(log ρAD − TAD + α1AD ) (log ρAD − TAD + α1AD )]
α
+ tr[exp(log ρAD − TAD + α1AD ) (TAD − log ρAD )]
= inf α tr[exp(log ρAD − TAD + α1AD )]
(A10)
= inf αeα tr[exp(log ρAD − TAD )]
(A11)
= − tr[exp(log ρAD − TAD − 1AD )] .
(A12)
α
α
19
−1/2
−1/2
With Slater’s strong duality and a change of variable S̄AF := σAF SAF σAF
the bar in the following), we get
D̄ rec (σAE kρAD ) = maximize
(but not including
− tr[SAF σAF ] − tr[exp(log ρAD − TAD − 1AD )]
subject to 1D ⊗ SAF ≥ TAD ⊗ 1F .
(A13)
We now do another change of variable (but not including the bar in the following)
RAD := exp(log ρAD − TAD − 1AD ) as well as
1D ⊗ S̄AF := 1D ⊗ SAF + 1ADF ,
(A14)
and the program becomes
D̄ rec (σAE kρAD ) = maximize
− tr[SAF σAF ] + 1 − tr[RAD ]
subject to RAD > 0
(A15)
1D ⊗ SAF ≥ (log ρAD − log RAD ) ⊗ 1F .
Observe now that we can add the constraint tr[RAD ] = 1. In fact, let
R̄AD :=
RAD
tr[RAD ]
and
S̄AF := SAF + log tr[RAD ] .
(A16)
This solution satisfies the constraint and the objective value becomes
− tr[S̄AF σAF ] + 1 − tr[R̄AD ] = − tr[SAF σAF ] − log tr[RAD ] ≥ − tr[SAF σAF ] + 1 − tr[RAD ] .
(A17)
This concludes the proof of (96).
To prove (97) first note that it is immediate from the definition of the argument flipped relative
entropy of recovery that
D̄ rec (σAE ⊗ ωA′ E ′ kρAD ⊗ τA′ D′ ) ≤ D̄ rec (σAE kρAD ) + D̄ rec (ωA′ E ′ kτA′ D′ ) ,
(A18)
and in the following we prove inequality in the other direction using the dual representation (96). Given feasible operators RAD , SAF for the quantity D̄ rec (σAE kρAD ) and feasible operators RA′ D′ , SA′ F ′ for the quantity D̄ rec (ωA′ E ′ kτA′ D′ ), we have
1D ⊗ SAF ≥ (log ρAD − log RAD ) ⊗ 1F ∧ 1D′ ⊗ SA′ F ′ ≥ (log τA′ D′ − log RA′ D′ ) ⊗ 1F ′
=⇒ 1DD′ ⊗ (SAF ⊗ 1A′ F ′ + 1AF ⊗ SA′ F ′ )
≥ (log ρAD − log RAD ) ⊗ 1A′ D′ + 1AD ⊗ (log τA′ D′ − log RA′ D′ ) ⊗ 1F F ′
= log(ρAD ⊗ τA′ D′ ) − log(RAD ⊗ RA′ D′ ) ⊗ 1F F ′ ,
(A19)
just by multiplying with identities and adding the resulting operator inequalities. Moreover, we
have tr[RAD ⊗ RA′ D′ ] = 1. Hence, (RAD ⊗ RA′ D′ , SAF ⊗ 1A′ F ′ + 1AF ⊗ SA′ F ′ ) is a feasible pair in
the expression (96) for D̄ rec (σAE ⊗ ωA′ E ′ kρAD ⊗ τA′ D′ ) and we get
D̄ rec (σAE ⊗ ωA′ E ′ kρAD ⊗ τA′ D′ ) ≥ − tr[(SAF ⊗ 1A′ F ′ + 1AF ⊗ SA′ F ′ )(σAF ⊗ ωA′ F ′ )]
= − tr[SAF σAF ] − tr[SA′ F ′ ωA′ F ′ ] .
(A20)
(A21)
Taking the supremum over feasible (RAD , SAF ) and (RA′ D′ , SA′ F ′ ), we find the claimed additivity.
20
[1] P. M. Alberti.
“A Note on the Transition Probability over C*-Algebras”.
Letters in Mathematical Physics 7: 25–32 (1983).
[2] H. Araki. “On an Inequality of Lieb and Thirring”. Letters in Mathematical Physics 19(2): 167–170
(1990).
[3] S. Beigi.
“Sandwiched Rényi Divergence Satisfies Data Processing Inequality”.
Journal of Mathematical Physics 54(12): 122202 (2013).
[4] V. P. Belavkin and P. Staszewski. “C*-algebraic Generalization of Relative Entropy and Entropy”.
Annals Henri Poincaré 37(1): 51–58, (1982).
[5] A. Ben-Tal and A. Nemirovski.
Lectures on Modern Convex Optimization.
Society for Industrial and Applied Mathematics (2001).
[6] M. Berta and M. Tomamichel.
“The Fidelity of Recovery Is Multiplicative”.
IEEE Transactions on Information Theory 62(4): 1758–1763 (2016).
[7] S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University Press (2004).
[8] F. G. S. L. Brandão, A. W. Harrow, J. Oppenheim, and S. Strelchuk. “Quantum Conditional Mutual
Information, Reconstructed States, and State Redistribution”. Physical Review Letters 115(5): 050501
(2015).
[9] V. Chandrasekaran and P. Shah.
“Relative entropy optimization and its applications”.
Mathematical Programming 161(1): 1–32 (2017).
[10] T. Cooney, C. Hirche, C. Morgan, J. P. Olson, K. P. Seshadreesan, J. Watrous, and M. M. Wilde.
“Operational meaning of quantum measures of recovery”. Physical Review A 94(2): 022310 (2016).
[11] D. DiVincenzo, M. Horodecki, D. Leung, J. Smolin, and B. Terhal. “Locking Classical Correlations in
Quantum States”. Physical Review Letters 92(6): 067902 (2004).
[12] M. J. Donald. “On the Relative Entropy”. Communications in Mathematical Physics 105(1): 13–34
(1986).
[13] D. R. Farenick. “Private Communication”, (2015).
[14] D. R. Farenick and F. Zhou.
“Jensen’s Inequality Relative to Matrix-Valued Measures”.
Journal of Mathematical Analysis and Applications 327(2): 919–929 (2007).
[15] H. Fawzi and O. Fawzi. “Relative entropy optimization in quantum information theory via semidefinite
programming approximations”. Preprint, arXiv: 1705.06671 (2017).
[16] O. Fawzi and R. Renner. “Quantum Conditional Mutual Information and Approximate Markov
Chains”. Communications in Mathematical Physics 340(2): 575–611 (2015).
[17] R. L. Frank and E. H. Lieb.
“Monotonicity of a Relative Rényi Entropy”.
Journal of Mathematical Physics 54(12): 122201 (2013).
[18] C. A. Fuchs. Distinguishability and Accessible Information in Quantum Theory. Phd thesis, University
of New Mexico, (1996). Available at arXiv: quant-ph/9601020v1.
[19] S. Golden. “Lower Bounds for the Helmholtz Function”. Physical Review 137(4B): B1127–B1128
(1965).
[20] F. Hiai.
“Equality Cases in Matrix Norm Inequalities of Golden-Thompson Type”.
Linear and Multilinear Algebra 36(4): 239–249 (1994).
[21] F. Hiai and D. Petz. “The Proper Formula for Relative Entropy and its Asymptotics in Quantum
Probability”. Communications in Mathematical Physics 143(1): 99–114 (1991).
[22] B. Ibinson, N. Linden, and A. Winter.
“Robustness of Quantum Markov Chains”.
Communications in Mathematical Physics 277(2): 289–304 (2007).
[23] M. Junge, R. Renner, D. Sutter, M. M. Wilde, and A. Winter. “Universal recovery from a decrease of
quantum relative entropy”. Preprint, arXiv: 1509.07127 (2015).
[24] R. König, R. Renner, and C. Schaffner. “The Operational Meaning of Min- and Max-Entropy”.
IEEE Transactions on Information Theory 55(9): 4337–4347 (2009).
[25] H. Kosaki. “Relative Entropy of States: A Variational Expression”. Journal of Operator Theory
16(2): 335–348, (1986).
[26] S.
Kullback
and
R.
A.
Leibler.
“On
Information
and
Sufficiency”.
Annals of Mathematical Statistics 22(1): 79–86 (1951).
[27] K. Li and A. Winter. “Squashed Entanglement, k-Extendibility, Quantum Markov Chains, and Recovery
Maps”. Preprint, arXiv: 1410.4184 (2014).
21
[28] E. H. Lieb and W. E. Thirring. “Inequalities for the Moments of the Eigenvalues of the Schrödinger
Hamiltonian and Their Relation to Sobolev Inequalities”. In The Stability of Matter: From Atoms to
Stars, chapter III, pages 205–239. Springer (2005).
[29] G.
Lindblad.
“Completely
Positive
Maps
and
Entropy
Inequalities”.
Communications in Mathematical Physics 40(2): 147–151 (1975).
[30] K. Matsumoto. “A New Quantum Version of f-Divergence”. Preprint, arXiv: 1311.4722 (2014).
[31] M. Mosonyi and T. Ogawa. “Quantum Hypothesis Testing and the Operational Interpretation of
the Quantum Rényi Relative Entropies”. Communications in Mathematical Physics 334(3): 1617–1648
(2015).
[32] M. Müller-Lennert. Quantum Relative Rényi Entropies. Master thesis, ETH Zurich,
[33] M. Müller-Lennert, F. Dupuis, O. Szehr, S. Fehr, and M. Tomamichel. “On Quantum Rényi Entropies:
A New Generalization and Some Properties”. Journal of Mathematical Physics 54(12): 122203 (2013).
[34] D. Petz. “Quasi-entropies for Finite Quantum Systems”. Reports on Mathematical Physics 23(1): 57–65
(1986).
[35] D. Petz. “Sufficient Subalgebras and the Relative Entropy of States of a von Neumann Algebra”.
Communications in Mathematical Physics 105(1): 123–131 (1986).
[36] D.
Petz.
“A
Variational
Expression
for
the
Relative
Entropy”.
Communications in Mathematical Physics 114(2): 345–349 (1988).
[37] D. Petz. Quantum Information Theory and Quantum Statistics. Springer (2008).
[38] M. Piani.
“Relative Entropy of Entanglement and Restricted Measurements”.
Physical Review Letters 103(16): 160504 (2009).
[39] E.
Rains.
“A
Semidefinite
Program
for
Distillable
Entanglement”.
IEEE Transactions on Information Theory 47(7): 2921–2933 (2001).
[40] A. Rényi. “On Measures of Information and Entropy”. In Proc. 4th Berkeley Symposium on Mathematical Statistics and Probability, volume 1, pages 547–561, Berkeley, California, USA(1961).
[41] K. P. Seshadreesan and M. M. Wilde. “Fidelity of Recovery, Squashed Entanglement, and Measurement
Recoverability”. Physical Review A 92(4): 042321 (2015).
[42] M. Sion. “On General Minimax Theorems”. Pacific Journal of Mathematics 8: 171–176, (1958).
[43] W.
So.
“Equality
Cases
in
Matrix
Exponential
Inequalities”.
SIAM Journal on Matrix Analysis and Applications 13(4): 1154–1158 (1992).
[44] D. Sutter, O. Fawzi, and R. Renner. “Universal recovery map for approximate Markov chains”.
Proceedings of the Royal Society of London A: Mathematical, Physical and Engineering Sciences 472: 2186
(2016).
[45] D. Sutter, M. Tomamichel, and A. W. Harrow. “Strengthened Monotonicity of Relative Entropy via
Pinched Petz Recovery Map”. IEEE Transactions on Information Theory 62(5): 2907–2913 (2016).
[46] C.
J.
Thompson.
“Inequality
with
Applications
in
Statistical
Mechanics”.
Journal of Mathematical Physics 6(11): 1812 (1965).
[47] M. Tomamichel. Quantum Information Processing with Finite Resources — Mathematical Foundations.
volume 5 of SpringerBriefs in Mathematical Physics, Springer International Publishing (2016).
[48] M. Tomamichel, M. Berta, and M. Hayashi. “Relating Different Quantum Generalizations of the Conditional Rényi Entropy”. Journal of Mathematical Physics 55(8): 082206 (2014).
[49] A. Uhlmann. “Relative entropy and the Wigner-Yanase-Dyson-Lieb concavity in an interpolation theory”. Communications in Mathematical Physics 54(1): 21–32 (1977).
[50] H. Umegaki. “Conditional Expectation in an Operator Algebra”. Kodai Math. Sem. Rep. 14: 59–85,
(1962).
[51] T. van Erven and P. Harremoes.
“Rényi Divergence and Kullback-Leibler Divergence”.
IEEE Transactions on Information Theory 60(7): 3797–3820 (2014).
[52] V. Vedral. “On Bound Entanglement Assisted Distillation”. Physics Letters A 262(2-3): 121–124
(1999).
[53] V. Vedral and M. B. Plenio.
“Entanglement Measures and Purification Procedures”.
Physical Review A 57(3): 1619–1633 (1998).
[54] M.
M.
Wilde.
“Recoverability
in
quantum
information
theory”.
Proceedings of the Royal Society A 471(2182): 20150338 (2015).
[55] M. M. Wilde, A. Winter, and D. Yang.
“Strong Converse for the Classical Capacity
of Entanglement-Breaking and Hadamard Channels via a Sandwiched Rényi Relative Entropy”.
22
Communications in Mathematical Physics 331(2): 593–622 (2014).
| 7cs.IT
|
Compositional Abstractions of Interconnected Discrete-Time
Stochastic Control Systems
arXiv:1709.10312v1 [cs.SY] 29 Sep 2017
Abolfazl Lavaei, Sadegh Esmaeil Zadeh Soudjani, Rupak Majumdar, and Majid Zamani
Abstract— This paper is concerned with a compositional approach for constructing abstractions of interconnected discretetime stochastic control systems. The abstraction framework
is based on new notions of so-called stochastic simulation
functions, using which one can quantify the distance between
original interconnected stochastic control systems and their
abstractions in the probabilistic setting. Accordingly, one can
leverage the proposed results to perform analysis and synthesis
over abstract interconnected systems, and then carry the
results over concrete ones. In the first part of the paper, we
derive sufficient small-gain type conditions for the compositional quantification of the distance in probability between
the interconnection of stochastic control subsystems and that
of their abstractions. In the second part of the paper, we
focus on the class of discrete-time linear stochastic control
systems with independent noises in the abstract and concrete
subsystems. For this class of systems, we propose a computational scheme to construct abstractions together with their
corresponding stochastic simulation functions. We demonstrate
the effectiveness of the proposed results by constructing an
abstraction (totally 4 dimensions) of the interconnection of four
discrete-time linear stochastic control subsystems (together 100
dimensions) in a compositional fashion.
I. I NTRODUCTION
Large-scale interconnected systems have received significant attentions in the last few years due to their presence in
real life systems including power networks, air traffic control,
and so on. Each complex real-world system can be regarded
as an interconnected system composed of several subsystems.
Since these large-scale network of systems are inherently difficult to analyze and control, one can develop compositional
schemes to employ the abstractions of the given systems as a
replacement in the controller design process. In other words,
in order to overcome the computational complexity in largescale interconnected systems, one can abstract the original
concrete system by a simpler one with lower dimension.
Those abstractions allow us to design controllers for them,
and then refine the controllers to the ones for the concrete
complex systems, while provide us with the quantified errors
in this controller synthesis detour.
In the past few years, there have been several results on the
construction of (in)finite abstractions for stochastic systems.
Existing results include infinite approximations for a class of
stochastic hybrid systems [1] and finite approximations for
discrete-time stochastic models with continuous state spaces
[2], [3], [4]. Construction of finite bisimilar abstractions for
stochastic control systems is proposed in [5], [6]. Recent
results address stochastic switched systems [7], [8] and
This work was supported in part by the German Research Foundation
(DFG) through the grant ZA 873/1-1.
A. Lavaei and M. Zamani are with the Department of Electrical and Computer Engineering, Technical University of Munich, D-80290 Munich, Germany. S. Esmaeil Zadeh Soudjani and R. Majumdar are with the Max Planck
Institute for Software Systems, Kaiserslautern 67663, Germany. Email:
{lavaei,zamani}@tum.de, {sadegh,rupak}@mpi-sws.org.
propose compositional construction of infinite abstractions
of continuous-time stochastic control systems [9], [10] using
small-gain type compositional reasoning.
In this paper, we provide a compositional approach for
the construction of infinite abstractions of interconnected
discrete-time stochastic control systems. Our abstraction
framework is based on a new notion of so-called stochastic
simulation functions under which an abstraction, which is
itself a discrete-time stochastic control system with lower
dimension, performs as a substitute in the controller design
process. The stochastic simulation function is leveraged to
quantify the error in probability in this controller synthesis
scheme. As a consequence, one can use the proposed results here to solve particularly safety/reachability problems
over the abstract interconnected systems and then carry the
results over the concrete interconnected ones. It should be
noted that the existing compositional results in [9], [10] are
for continuous-time stochastic systems and assume that the
noises in the concrete and abstract systems are the same,
which means the abstraction has access to the noise of the
concrete system, which is a strong assumption. In this paper,
we do not have such an assumption meaning that the noises
of the abstraction can be completely independent of that of
the concrete system.
II. D ISCRETE -T IME S TOCHASTIC C ONTROL S YSTEMS
A. Preliminaries
We consider a probability space (Ω, FΩ , PΩ ), where Ω is
the sample space, FΩ is a sigma-algebra on Ω comprising
subsets of Ω as events, and PΩ is a probability measure
that assigns probabilities to events. We assume that random
variables introduced in this article are measurable functions
of the form X : (Ω, FΩ ) → (SX , FX ). Any random variable
X induces a probability measure on its space (SX , FX ) as
P rob{A} = PΩ {X −1 (A)} for any A ∈ FX . We often directly discuss the probability measure on (SX , FX ) without
explicitly mentioning the underlying probability space and
the function X itself.
A topological space S is called a Borel space if it is
homeomorphic to a Borel subset of a Polish space (i.e., a
separable and completely metrizable space). Examples of a
Borel space are the Euclidean spaces Rn , its Borel subsets
endowed with a subspace topology, as well as hybrid spaces.
Any Borel space S is assumed to be endowed with a Borel
sigma-algebra, which is denoted by B(S). We say that a map
f : S → Y is measurable whenever it is Borel measurable.
B. Notation
The following notation is used throughout the paper. We
denote the set of nonnegative integers by N := {0, 1, 2, . . .}
and the set of positive integers by Z+ := {1, 2, 3, . . .}. The
symbols R, R>0 , and R≥0 denote the set of real, positive, and
nonnegative real numbers, respectively. Given a vector x ∈
Rn , kxk denotes the Euclidean norm of x. The symbols In
and 1n denote the identity matrix in Rn×n , and the vector in
Rn with all its elements to be one, respectively. We denote by
diag(a1 , . . . , aN ) a diagonal matrix in RN ×N with diagonal
matrix entries a1 , . . . , aN starting from the upper left corner.
Given functions fi : XQ
. . . , N },
i → Yi , for
QNany i ∈ {1,
QN
N
their Cartesian product i=1 fi : i=1 Xi → i=1 Yi is
Q
defined as ( N
i=1 fi )(x1 , . . . , xN ) = [f1 (x1 ); . . . ; fN (xN )].
For any set A we denote by AN the Cartesian
product of a
Q∞
countable number of copies of A, i.e., AN = k=0 A. Given
a measurable function f : N → Rn , the (essential) supremum
of f is denoted by kf k∞ := (ess)sup{kf (k)k, k ≥ 0}. A
+
function γ : R+
0 → R0 , is said to be a class K function if
it is continuous, strictly increasing, and γ(0) = 0. A class
K function γ(·) is said to be a class K∞ if γ(r) → ∞ as
r → ∞.
C. Discrete-Time Stochastic Control Systems
We consider stochastic control systems in discrete time
(dt-SCS) defined over a general state space adopted from
[11] and characterized by the tuple
Σ = (X, W, U, {U (x, ω)|x ∈ X, ω ∈ W }, Y, Tx , h) ,
(1)
where X is a Borel space as the state space of the system.
We denote by (X, B(X)) the measurable space with B(X)
being the Borel sigma-algebra on the state space. Sets W
and U are Borel spaces as the internal and external input
spaces of the system. The set {U (x, ω)|x ∈ X, ω ∈ W }
is a family of non-empty measurable subsets of U with the
property that
K := {(x, ω, ν) : x ∈ X, ω ∈ W, ν ∈ U (x)}
is measurable in X × W × U . Intuitively, U (x, ω) is the set
of inputs that are feasible at state x ∈ X with the internal
input ω ∈ W . Set Y is a Borel space as the output space
of the system. Map Tx : B(X) × X × W × U → [0, 1], is
a conditional stochastic kernel that assigns to any x ∈ X,
ω ∈ W and ν ∈ U (x, ω) a probability measure Tx (·|x, ω, ν)
on the measurable space R(X, B(X)) so that for any set
A ∈ B(X), Px,ω,ν (A) = A Tx (dx̄|x, ω, ν), where Px,ω,ν
denotes the conditional probability P(·|x, ω, ν). Finally, h :
X → Y is a measurable function that maps a state x ∈ X
to its output y = h(x) ∈ Y .
Given the dt-SCS in (1), we are interested in Markov
policies to control the system.
Definition 2.1: A Markov policy for the dt-SCS Σ in (1)
is a sequence ρ = (ρ0 , ρ1 , ρ2 , . . .) of universally measurable
stochastic kernels ρn [12], each defined on the input space
U given X × W and such that for all (xn , ωn ) ∈ X ×
W , ρn (U (xn , ωn )|(xn , ωn )) = 1. The class of all Markov
policies is denoted by ΠM .
For given inputs ω(·), ν(·), the stochastic kernel Tx captures the evolution of the state of the system. This kernel
features an equivalent dynamical representation: there exists
a measurable function fa : X × W × U × Vς → X such that
the evolution of the state of the system can be written as
x(k + 1) = fa (x(k), ω(k), ν(k), ς(k)),
where {ς(k) : Ω → Vς , k ∈ N} is a sequence of independent
and identically distributed (i.i.d.) random variables on the set
Vς . In this paper we assume that the state space X is a subset
of Rn and are interested in the specific form of the function
fa (x, ω, ν, ς) = f (x, ω, ν) + g(x)ς.
Therefore, the dt-SCS Σ in (1) can be described as:
x(k + 1) = f (x(k), ω(k), ν(k)) + g(x(k))ς(k),
Σ:
y(k) = h(x(k)),
(2)
for any x(k) ∈ X, ω(k) ∈ W, and ν(k) ∈ U (x(k), ω(k)).
Note that Tx in (1) contains the information of functions f
and g and the distribution of noise ς(·) in the dynamical
representation (2).
For the sake of simplicity, we also assume that the set of
valid inputs is the whole input space: U (x, ω) = U for all
x ∈ X and ω ∈ W , but the obtained results are generally
applicable. We associate respectively to U and W the sets U
and W to be collections of sequences {ν(k) : Ω → U, k ∈
N} and {ω(k) : Ω → W, k ∈ N}, in which ν(k) and ω(k)
are independent of ς(t) for any k, t ∈ N and t ≥ k.
For any initial state a ∈ X, ν(·) ∈ U, and ω(·) ∈ W,
the random sequences xaων : Ω × N → X and yaων : Ω ×
N → Y that satisfy (2) are called respectively the solution
process and output trajectory of Σ under external input ν,
internal input ω and initial state a. We here call the tuple
(ω, ν, xaων , yaων ) a trajectory of Σ.
III. S TOCHASTIC P SEUDO -S IMULATION AND
S IMULATION F UNCTIONS
In this section we first introduce a notion of so-called
pseudo-simulation functions for the discrete-time stochastic
control systems with both internal and external inputs and
then define the stochastic simulation functions for systems
with only external input. These definitions can be used to
quantify closeness of two dt-SCS with the same internal input
and output spaces.
Definition 3.1: Consider dt-SCS Σ = (X, W, U, Y, Tx , h)
b = (X̂, W, Û , Y, T̂x , ĥ) with the same internal input
and Σ
and output spaces. A function V : X × X̂ → R≥0 is called
b to Σ
a stochastic pseudo-simulation function (SPSF) from Σ
if
• ∃α ∈ K∞ such that
∀x ∈ X, ∀x̂ ∈ X̂,
•
α(kh(x) − ĥ(x̂)k) ≤ V (x, x̂), (3)
∀x ∈ X, x̂ ∈ X̂, ν̂ ∈ Û , and ∀ω̂ ∈ Ŵ , ∃ν ∈ U such
that ∀ω ∈ W
h
E V (x(k + 1), x̂(k + 1)) x(k), x̂(k), ω(k) = ω,
i
, ω̂(k) = ω̂, ν(k) = ν, ν̂(k) = ν̂ − V (x(k), x̂(k)) ≤
−κ(V (x(k), x̂(k)))+ρint (kω − ω̂k)+ρext(kν̂k)+ψ,
(4)
for some κ ∈ K, ρint , ρext ∈ K∞ ∪ {0}, and ψ ∈ R≥0 .
b PS Σ if there exists a pseudoWe utilize notation Σ
b to Σ, in which control system
simulation function V from Σ
b
Σ is considered as an abstraction of concrete (original)
system Σ. The second condition above implies implicitly
existence of a function ν = νν̂ (x, x̂, ν̂, ω̂) for satisfaction
of (4). This function is called the interface function and can
b to a policy ν
be used to refine a synthesized policy ν̂ for Σ
for Σ.
In this paper we study interconnected discrete-time
stochastic control systems without internal inputs, resulting
from the interconnection of discrete-time stochastic control
subsystems having both internal and external signals. In
this case, the interconnected dt-SCS reduces to the tuple
(X, U, Y, Tx , h). Thus we modify the above notion for systems without internal inputs.
Definition 3.2: Consider two dt-SCS Σ = (X, U, Y, Tx , h)
b = (X̂, Û , Y, T̂x , ĥ) with the same output spaces. A
and Σ
function V : X × X̂ → R≥0 is called a stochastic simulation
b to Σ if
function (SSF) from Σ
• ∃α ∈ K∞ such that
∀x ∈ X, ∀x̂ ∈ X̂,
α(kh(x) − ĥ(x̂)k) ≤ V (x, x̂), (5)
∀x ∈ X, x̂ ∈ X̂, ν̂ ∈ Û , ∃ν ∈ U such that
h
i
E V (x(k+1), x̂(k+1)) x(k), x̂(k), ν(k)=ν, ν̂(k)=ν̂
•
−V (x(k), x̂(k)) ≤−κ(V (x(k), x̂(k)))+ρext (kν̂k)+ψ,
(6)
for some κ ∈ K, ρext ∈ K∞ ∪ {0}, and ψ ∈ R≥0 .
The next theorem shows usefulness of SSF in comparing
output trajectories of two dt-SCS in a probabilistic sense.
b be two dt-SCS with the same
Theorem 3.3: Let Σ and Σ
b to Σ, and
output spaces. Suppose V is an SSF from Σ
there exists a constant 0 < κ
b < 1 such that the function
κ ∈ K in (6) satisfies κ(r) ≥ κ
br ∀r ∈ R≥0 . For any
external input trajectory ν̂(·) ∈ Û that preserves Markov
b and for any random variables
property for the closed-loop Σ,
a and â as the initial states of the two dt-SCS, there exists
an input trajectory ν(·) ∈ U of Σ through the interface
function associated with V such that the following inequality
holds provided that there exists a constant ψb ≥ 0 satisfying
ψb ≥ ρext (kν̂k∞ ) + ψ:
P
sup kyaν (k) − ŷâν̂ (k)k ≥ ε | [a; â]
0≤k≤T
(7)
T
b
1 − 1 − V (a,â) 1 − ψb
if α (ε) ≥ ψbκ ,
α(ε)
α(ε)
b
≤
b
ψ
V (a,â) (1−b
κ)T + κbα(ε)
(1−(1−b
κ)T ) if α (ε) < ψbκ .
α(ε)
b to Σ, we have
Proof: Since V is an SSF from Σ
P
sup kyaν (k) − ŷâν̂ (k)k ≥ ε | [a; â]
0≤k≤T
=P
sup α (kyaν (k) − ŷâν̂ (k)k) ≥ α(ε) | [a; â]
0≤k≤T
≤P
sup V (xaν (k), x̂âν̂ (k)) ≥ α(ε) | [a; â] . (8)
0≤k≤T
The equality holds due to α being a K∞ function. The
inequality is also true due to condition (5) on the SSF V .
The results follows by applying Theorem 3 in [13, pp. 81]
to (8) and utilizing inequality (6).
The results shown in Theorem 3.3 provide closeness of
output behaviours of two systems in finite-time horizon. We
can extend the result to infinite-time horizon provided that
constant ψ̂ = 0 as the following.
b be two dt-SCS with the same
Corollary 3.4: Let Σ and Σ
b to Σ such that
output spaces. Suppose V is an SSF from Σ
ρext (·) ≡ 0 and ψ = 0. For any external input trajectory
b
ν̂(·) ∈ Û preserving Markov property for the closed-loop Σ,
and for any random variables a and â as the initial states
of the two dt-SCS, there exists ν(·) ∈ U of Σ through the
interface function associated with V such that the following
inequality holds:
V (a, â)
.
P
sup kyaν (k) − ŷâ0 (k)k ≥ ε | [a; â] ≤
α (ε)
0≤k<∞
b to Σ with ρext (·) ≡ 0
Proof: Since V is an SSF from Σ
and ψ = 0, for any x(k) ∈ X and x̂(k) ∈ X̂ and any
ν̂(k) ∈ Û , there exists ν(k) ∈ U such that
h
i
E V (x(k + 1), x̂(k + 1)) | x(k), x̂(k), ν(k), ν̂(k)
− V ((x(k), x̂(k)) ≤ −κ(V (x(k), x̂(k)),
(9)
showing that V (xaν (k), x̂âν̂ (k)) is a nonnegative supermartingale [14]. Following the same reasoning as in the proof
of Theorem 3.3 we have
P
sup kyaν (k) − ŷâν̂ (k)k ≥ ε | [a; â]
0≤k<∞
=P
sup α kyaν (k) − ŷâν̂ (k)k ≥ α(ε) | [a; â]
0≤k<∞
V (a, â)
≤P
sup V (xaν (k), x̂âν̂ (k)) ≥ α(ε) | [a; â] ≤
,
α(ε)
0≤k<∞
where the last inequality is due to the nonnegative supermartingale property [13].
The stochastic simulation function defined before can be
used to guarantee an upper bound on the probability of the
maximum difference in output trajectories. This idea can
be used in conjunction with stochastic safety/reachability
analysis of the systems, which is discussed next.
Suppose that V is a stochastic simulation function from
b to Σ. Then for any input strategy ν̂ of the system Σ
b
Σ
there exists an input strategy ν of Σ, such that the following
probability is bounded
P sup kyaν (k) − ŷâν̂ (k)k ≥ ε | [a; â] ≤ δ,
0≤k≤T
with δ being defined in Theorem 3.3 based on ε and T . Given
the unsafe set A1 for Σ, we can construct another set A2 ,
which is the ε neighborhood of A1 , i.e.,
A2 = {y ′ |∃y ∈ A1 , ky ′ − yk ≤ ε}.
Now, we can provide the following corollary.
b to Σ. For any
Corollary 3.5: Suppose V is an SSF from Σ
input ν̂(·) there exists ν(·) such that the following inequality
holds:
P{∃k ≤ T, yaν (k) ∈ A1 } ≤ P{∃k ≤ T, ŷâν̂ (k) ∈ A2 }+δ.
Proof: Denote the events E1 := {∃k ≤ T, yaν (k) ∈
A1 } and E2 := {∃k ≤ T, ŷâν̂ (k) ∈ A2 }. Then we have
P{E1 } = P{E1 ∩ E2 } + P{E1 ∩ E¯2 } ≤ P{E2 } + P{E1 ∩ Ē2 },
where Ē2 is the complement of E2 . Notice that the term
P{E1 ∩ Ē2 } is bounded by δ due to the above results, which
concludes the proof.
IV. C OMPOSITIONAL A BSTRACTIONS
I NTERCONNECTED S YSTEMS
κi (s) ≥ λi γi (s)
hji ≡ 0 =⇒ δij = 0 and
FOR
Here, we first provide a formal definition of interconnection between discrete-time stochastic control systems.
A. Interconnected Stochastic Control Systems
Consider a complex stochastic control system Σ composed
of N ∈ N≥1 stochastic control subsystems Σi interconnected
with each other as follows:
Σi = (Xi , Wi , Ui , Yi , Txi , hi ), i ∈ [1; N ],
with partitioned internal inputs and outputs
ωi = ωi1 ; . . . ; ωi(i−1) ; ωi(i+1) ; . . . ; ωiN ,
yi = [yi1 ; . . . ; yiN ],
(10)
and also output space and function
hi (xi ) = [hi1 (xi ); . . . ; hiN (xi )],
Yi =
N
Y
Yij .
Assumption 1: For any i, j ∈ [1; N ], i 6= j, there exist
K∞ functions γi and constants λi ∈ R>0 and δij ∈ R≥0
such that for any s ∈ R≥0
hji 6≡ 0 =⇒ ρiint ((N − 1)α−1
j (s)) ≤ δij γj (s),
µT (−Λ + ∆) < 0
(11)
We interpret the outputs yii as external ones, whereas the
outputs yij with i 6= j are internal ones which are used
to define the interconnected stochastic control systems. In
particular, we assume that the dimension of ωij is equal
to the dimension of yji . If there is no connection from
stochastic control subsystem Σi to Σj , then we assume that
the connecting output function is identically zero for all
arguments, i.e., hij ≡ 0. Now, we define the interconnected
stochastic control systems as the following.
Definition 4.1: Consider N ∈ N≥1 stochastic control
subsystems Σi = (Xi , Wi , Ui , Yi , Txi , hi ), i ∈ [1; N ], with
the input-output configuration as in (10) and (11). The interconnection of Σi for any i ∈ [1, . . . , N ], is the interconnected
stochastic control system Σ = (X,
Y, Tx , h), denoted
QU,
QN by
N
I(Σ1 , . . . , ΣN ), such that X := i=1 Xi , U := i=1 Ui ,
QN
function fa := i=1 fai , characterizing the stochastic kernel
QN
Tx based on those of subsystems (i.e. fai ), Y := i=1 Yii ,
QN
and h = i=1 hii , subjected to the following constraint:
(12)
B. Compositional Abstractions of Interconnected Systems
This subsection contains one of the main contributions of
the paper. We assume that we are given N stochastic control
subsystems
Σi = (Xi , Wi , Ui , Yi , Txi , hi ),
bi =
together with their corresponding abstractions Σ
b i to Σi . For
(X̂i , Wi , Ûi , Yi , T̂xi , ĥi ) with SPSF Vi from Σ
providing the main compositionality result of the paper, we
raise the following assumption.
(15)
where κi , αj , and ρiint represent the corresponding K and
K∞ functions of Vi appearing in Definition 3.1. Prior to presenting the next theorem, we define Λ := diag(λ1 , . . . , λN ),
→
∆ := {δij }, where δii = 0 ∀i ∈ [1; N ], and Γ( s ) :=
→
[γ1 (s1 ); . . . ; γN (sN )], where s = [s1 ; . . . ; sN ]. In the next
theorem, we leverage a small-gain type condition to quantify
the error between the interconnection of stochastic control
subsystems and that of their abstractions in a compositional
way.
Theorem 4.2: Consider the interconnected stochastic control system Σ = I(Σ1 , . . . , ΣN ) induced by N ∈ N≥1
stochastic control subsystems Σi . Suppose that each stochasb i with the
tic control subsystem Σi admits an abstraction Σ
corresponding SPSF Vi . If Assumption 1 holds and there
exists a vector µ ∈ RN
>0 such that the inequality
(16)
is also met, then
j=1
ωij = yji ∀i, j ∈ [1; N ], i 6= j.
(13)
(14)
V (x, x̂) :=
N
X
µi Vi (xi , x̂i )
i=1
b = I(Σ
b 1, . . . , Σ
b N ) to Σ =
is an SSF function from Σ
I(Σ1 , . . . , ΣN ).
The proof is similar to that of Theorem 4.5 in [9], and is
omitted here due to lack of space.
V. D ISCRETE -T IME L INEAR S TOCHASTIC C ONTROL
S YSTEMS
In this section, we focus on a class of discrete-time linear
stochastic control systems, defined as follows:
x(k + 1) = Ax(k) + Bν(k) + Dω(k) + F ς(k),
Σ:
y(k) = Cx(k),
(17)
where the additive noise ς(k) is a sequence of independent
random vectors with multivariate standard normal distributions. We use the tuple Σ = (A, B, C, D, F ) to refer to the
class of systems in (17). Here, we provide conditions under
which a candidate V is an SPSF function facilitating the
construction of an abstraction Σ̂.
Let us assume that there exist matrix K and positive
definite matrix M such that the matrix inequalities
C T C M,
(18)
(1 + π)(A + BK)T M (A + BK) − M −b
κM, (19)
hold for some positive constants π and 0 < κ
b < 1. We
employ the following quadratic SPSF
V (x, x̂) = (x − P x̂)T M (x − P x̂),
(20)
where P ∈ Rn×n̂ is a matrix of appropriate dimension.
Assume that the equalities
AP
D
CP
=
=
=
P Â − BQ
P D̂ − BS
Ĉ,
(21)
(22)
(23)
hold for some matrices Q and S of appropriate dimensions
and possibly with the lowest possible n̂. In the next theorem,
we show that under the aforementioned conditions V in (20)
b to Σ.
is an SPSF from Σ
b =
Theorem 5.1: Let Σ = (A, B, C, D, F ) and Σ
(Â, B̂, Ĉ, D̂, F̂ ) be two discrete-time linear stochastic control subsystems with two independent additive noises. Suppose that there exist matrices M , K, P , Q, and S satisfying
(18), (19), (21), (22), and (23). Then, V defined in (20) is
b to Σ.
an SPSF from Σ
Proof: Here, we show that ∀x, ∀x̂, ∀ν̂, ∀ω̂, ∃ν, ∀ω,
such that V satisfies kCx − Ĉ x̂k2 ≤ V (x, x̂) and
h
E V (x(k + 1), x̂(k + 1)) | x(k), x̂(k), ω(k) = ω, ω̂(k) = ω̂,
i
, ν̂(k) = ν̂ − V (x(k), x̂(k))
π √
2
≤ −b
κ(V (x(k), x̂(k))) + (1 + + )k M Dk2 kω − ω̂k2
π
2
2
2 √
2
e − P B̂)k kν̂k2
+ (1 + + )k M (B R
π
π
+ Tr F T M F + F̂ T P T M P F̂ .
(24)
According to (23), we have kCx − Ĉ x̂k2 = (x −
P x̂)T C T C(x − P x̂). By applying (18), it can be easily
verified that kCx − Ĉ x̂k2 ≤ V (x, x̂) holds ∀x, ∀x̂. Now, we
show inequality (24). Given any x, x̂, ν̂, and ω̂, we choose
ν via the following linear interface function:
e + S ω̂, (25)
ν = νν̂ (x, x̂, ν̂, ŵ) := K(x − P x̂) + Qx̂ + Rν̂
e of appropriate dimension. By Employing
for some matrix R
equations (21), (22), and the definition of the interface
function in (25), we simplify
Ax + Bνν̂ (x, x̂, ν̂, ω̂) + Dω − P (Âx̂ + B̂ ν̂ + D̂ω̂)
+ F ς(k) − P F̂ ςˆ(k)
e
to (A+BK)(x−P x̂)+D(ω− ω̂)+(B R−P
B̂)ν̂ + F ς(k)−
P F̂ ςˆ(k) . One obtains:
h
E V (x(k + 1), x̂(k + 1)) | x(k), x̂(k), ω(k) = ω, ω̂(k) = ω̂,
i
, ν̂(k) = ν̂ − V (x(k), x̂(k))
h
i
= (x − P x̂)T (A + BK)T M (A + BK) − M (x − P x̂)
i h
i
h
+ 2(x − P x̂)T (A + BK)T M D(ω − ω̂)
h
i h
i
e − P B̂)ν̂
+ 2(x − P x̂)T (A + BK)T M (B R
i
i h
h
√
e − P B̂)ν̂ + k M D(ω − ω̂)k2
+ 2(ω − ω̂)T DT M (B R
√
e − P B̂)ν̂k2 + Tr F T M F + F̂ T P T M P F̂ .
+ k M (B R
1 2
b , for any
Using Young’s inequality [15] as ab ≤ π2 a2 + 2π
a, b ≥ 0 and any π > 0, and by employing Cauchy-Schwarz
inequality and (19), one obtains the following upper bound:
h
E V (x(k + 1), x̂(k + 1)) | x(k), x̂(k), ω(k) = ω, ω̂(k) = ω̂,
i
, ν̂(k) = ν̂ − V (x(k), x̂(k))
π √
2
≤ −b
κ(V (x, x̂)) + (1 + + )k M Dk2 kω − ω̂k2
π
2
2 √
2
e − P B̂)k2 kν̂k2
+ (1 + + )k M (B R
π
π
+ Tr F T M F + F̂ T P T M P F̂ .
(26)
b to
Hence, the proposed V in (20) is an SPSF from Σ
Σ, which completes the proof. Note that the K and K∞
functions κ, α, and ρext , in Definition 3.1 associated with
the SPSF in √
(20) are α(s) := s2 , κ(s) := b
κs, and√ρint (s) :=
2
π
e
(1+ π + 2 )k M Dk2 s2 , ρext (s) := (1+ π2 + π2 )k M (B R−
P B̂)k2 s2 , ∀s ∈ R≥0 . Moreover, positive
constant
ψ
in
(4)
is ψ = Tr F T M F + F̂ T P T M P F̂ .
Remark 5.2: One can readily verify from the result of
Theorem 5.1 that choosing F̂ equal to zero results in smaller
constant ψ and, hence, more closeness of linear subsystems
and their abstractions. Observe that this is not the case when
one assumes the noise of the concrete subsystem and its
abstraction are the same as in [9], [10].
Remark 5.3: Note that the results in Theorem 5.1 do not
impose any condition on matrix B̂ and, hence, it can be
chosen arbitrarily. As an example, one can choose B̂ = In̂
b fully actuated and, hence,
which makes the abstract system Σ
the synthesis problem over it much easier.
Remark 5.4: Since Theorem 5.1 does not impose any
e we choose R
e to minimize function
condition on matrix R,
e
ρext for V as suggested in [16]. The following choice for R
minimizes ρext .
e = (B T M B)−1 B T M P B̂.
R
(27)
VI. E XAMPLE
Here, we demonstrate the effectiveness of the proposed
results for an interconnected system consisting of four
discrete-time linear stochastic control subsystems, i.e. Σ =
I(Σ1 , Σ2 , Σ3 , Σ4 ). The interconnection scheme of Σ with
four inputs and two outputs is illustrated in Figure 1.
As seen, the output of Σ1 (resp. Σ2 ) is connected to the
internal input of Σ4 (resp. Σ3 ) and the output of Σ3 (resp.
Σ4 ) connects to the internal input of Σ1 (resp. Σ2 ). The
system matrices are given by
Ai = I25 , Bi = I25 , CiT = 0.1125 , Fi = 0.01125 ,
for i ∈ {1, 2, 3, 4}. The internal input and output matrices
are also given by:
T
T
T
T
C14
= C23
= C31
= C42
= 0.1125 ,
D13 = D24 = D32 = D41 = 0.1125 .
In order to construct an abstraction for I(Σ1 , Σ2 , Σ3 , Σ4 ),
b i of each individual subsystem
we construct an abstraction Σ
Σi , i ∈ {1, 2, 3, 4}. We first fix κ
b and π for each subsystem,
Σ1
y14
ν1
ν3
ν2
ν4
Σ2
Fig. 1.
Σ3
y31
y33
y23
Σ4
y44
y42
The interconnected system Σ = I(Σ1 , Σ2 , Σ3 , Σ4 ).
and then determine the matrices M and K such that (18)
and (19) hold for i ∈ {1, 2, 3, 4}:
Mi = I25 , Ki = −0.95I25 , κ
bi = 0.98, πi = 0.99.
We continue with determining other matrices such that (21),
(22), and (23) hold:
Pi = 125 , Qi = 125 , Si = −0.003125,
for i ∈ {1, 2, 3, 4}. Accordingly, the matrices of abstract
subsystems are computed as:
Âi = 2, Ĉi = 2.5, D̂i = 0.096,
for i ∈ {1, 2, 3, 4}. Note that here F̂i , i ∈ {1, 2, 3, 4}, are
considered zero in order to reduce constants ψi for each
ei , i ∈
Vi . Moreover, B̂i are chosen 1 and we compute R
e
{1, 2, 3, 4}, using (27) as Ri = 125 . The interface function
for i ∈ {1, 2, 3, 4} follows by (25) as:
νi = −0.95I25 (xi − 125 x̂i )+125 x̂i +125 ν̂i −0.003125ω̂i .
Hence, Theorem 5.1 holds and Vi (xi , x̂i ) = (xi −
b i to Σi
125 x̂i )T Mi (xi − 125 x̂i ) is an SPSF function from Σ
2
satisfying conditions (3) and (4) with αi (s) = s , κi (s) =
0.98s, ρiext (s) = 0, ρiint (s) = 0.88s2, and ψi = 0.0025,
for i ∈ {1, 2, 3, 4}. We now proceed with Theorem 4.2
b to Σ.
to construct a stochastic simulation function form Σ
Assumption 1 holds with γi (s) = s and:
0.98
0
0
0
0
0
0.88
0
0.98
0
0
0
0
0.88
0
0
.
, Λ=
∆=
0
0
0.98
0
0
0.88
0
0
0
0
0
0.98
0.88
0
0
0
Additionally, one can readily verify that a vector µ ∈ R4>0
exists here since the spectral radius of Λ−1 ∆ is strictly less
than one [17]. By choosing vector µ as µT = [1 1 1 1],
the function
V (x, x̂) = V1 (x1 , x̂1 )+V2 (x2 , x̂2 )+V3 (x3 , x̂3 )+V4 (x4 , x̂4 )
b 1, Σ
b 2, Σ
b 3, Σ
b 4 ) to I(Σ1 , Σ2 , Σ3 , Σ4 ) satis an SSF from I(Σ
isfying conditions (5) and (6) with α(s) = s2 , κ(s) = 0.1s,
ρext (s) = 0, ∀s ∈ R≥0 , and ψ = 0.01. If the initial states of
b are started from zero,
the interconnected systems Σ and Σ
one can readily verify that the norm of error between outputs
b will not exceed 1 with probability at least
of Σ and of Σ
90% computed by the stochastic simulation function V using
inequality (7) for T = 10.
VII. D ISCUSSION
In this paper, we provided a compositional approach
for abstractions of interconnected discrete-time stochastic
control systems, with independent noises in the abstract and
concrete subsystems. First, we introduced new notions of
stochastic pseudo-simulation and stochastic simulation functions in order to quantify the distance in a probability setting
between original stochastic control subsystems and their abstractions and their interconnections, respectively. Therefore,
one can employ the proposed results here to potentially solve
safety/reachability problems over the abstract interconnected
systems and then refine the results to the concrete interconnected ones. Furthermore, we provided a computational
scheme for the class of discrete-time linear stochastic control
systems to construct abstractions together with their corresponding stochastic pseudo-simulation functions. Finally, we
demonstrated the effectiveness of the results by constructing
an abstraction (totally 4 dimensions) of the interconnection
of four discrete-time linear stochastic control subsystems
(together 100 dimensions) in a compositional fashion.
R EFERENCES
[1] A. A. Julius and G. J. Pappas, “Approximations of stochastic hybrid
systems,” IEEE Transactions on Automatic Control, vol. 54, no. 6, pp.
1193–1203, 2009.
[2] I. Tkachev and A. Abate, “On infinite-horizon probabilistic properties
and stochastic bisimulation functions,” in Proceedings of the 50th
IEEE Conference on Decision and Control and European Control
Conference (CDC-ECC), 2011, pp. 526–531.
[3] S. Esmaeil Zadeh Soudjani and A. Abate, “Adaptive and sequential
gridding procedures for the abstraction and verification of stochastic
processes,” SIAM Journal on Applied Dynamical Systems, vol. 12,
no. 2, pp. 921–956, 2013.
[4] S. Esmaeil Zadeh Soudjani, “Formal abstractions for automated
verification and synthesis of stochastic systems,” Ph.D. dissertation,
Technische Universiteit Delft, The Netherlands, November 2014.
[5] M. Zamani, P. Mohajerin Esfahani, R. Majumdar, A. Abate, and
J. Lygeros, “Symbolic control of stochastic systems via approximately
bisimilar finite abstractions,” IEEE Transactions on Automatic Control,
vol. 59, no. 12, pp. 3135–3150, 2014.
[6] M. Zamani, I. Tkachev, and A. Abate, “Towards scalable synthesis of
stochastic control systems,” Discrete Event Dynamic Systems, vol. 27,
no. 2, pp. 341–369, July 2017.
[7] M. Zamani and A. Abate, “Approximately bisimilar symbolic models
for randomly switched stochastic systems,” Systems & Control Letters,
vol. 69, pp. 38–46, 2014.
[8] M. Zamani, A. Abate, and A. Girard, “Symbolic models for stochastic
switched systems: A discretization and a discretization-free approach,”
Automatica, vol. 55, pp. 183–196, 2015.
[9] M. Zamani, “Compositional approximations of interconnected stochastic hybrid systems,” in Proceedings of the 53rd IEEE Conference on
Decision and Control (CDC), 2014, pp. 3395–3400.
[10] M. Zamani, M. Rungger, and P. Mohajerin Esfahani, “Approximations
of stochastic hybrid systems: a compositional approach,” IEEE Transactions on Automatic Control, 2016.
[11] O. Hernández-Lerma and J. B. Lasserre, Discrete-time Markov control
processes, ser. Applications of Mathematics. Springer-Verlag, 1996,
vol. 30.
[12] D. P. Bertsekas and S. E. Shreve, Stochastic Optimal Control: The
Discrete-Time Case. Athena Scientific, 1996.
[13] H. J. Kushner, Stochastic Stability and Control, ser. Mathematics in
Science and Engineering. Elsevier Science, 1967.
[14] B. Oksendal, Stochastic differential equations: an introduction with
applications. Springer Science & Business Media, 2013.
[15] W. H. Young, “On classes of summable functions and their fourier
series,” Proceedings of the Royal Society of London A: Mathematical,
Physical and Engineering Sciences, vol. 87, no. 594, pp. 225–229,
1912.
[16] A. Girard and G. J. Pappas, “Hierarchical control system design using
approximate simulation,” Automatica, vol. 45, no. 2, pp. 566–571,
2009.
[17] S. Dashkovskiy, H. Ito, and F. Wirth, “On a small gain theorem for iss
networks in dissipative lyapunov form,” European Journal of Control,
vol. 17, no. 4, pp. 357–365, 2011.
| 3cs.SY
|
Ready, Set, Verify!
arXiv:1803.06960v2 [cs.PL] 20 Mar 2018
Applying hs-to-coq to real-world Haskell code
JOACHIM BREITNER, University of Pennsylvania, USA
ANTAL SPECTOR-ZABUSKY, University of Pennsylvania, USA
YAO LI, University of Pennsylvania, USA
CHRISTINE RIZKALLAH, University of New South Wales, Australia
JOHN WIEGLEY, BAE Systems, USA
STEPHANIE WEIRICH, University of Pennsylvania, USA
Good tools can bring mechanical verification to programs written in mainstream functional languages. We use
hs-to-coq to translate significant portions of Haskell’s containers library into Coq, and verify it against
specifications that we derive from a variety of sources including type class laws, the library’s test suite, and
interfaces from Coq’s standard library. Our work shows that it is feasible to verify mature, widely-used, highly
optimized, and unmodified Haskell code. We also learn more about the theory of weight-balanced trees, extend
hs-to-coq to handle partiality, and – since we found no bugs – attest to the superb quality of well-tested
functional code.
CCS Concepts: • Software and its engineering → Software verification;
Additional Key Words and Phrases: Coq, Haskell, verification
1
INTRODUCTION
What would it take to tempt functional programmers to verify their code?
Certainly, better tools would help. We see that functional programmers who use dependentlytyped languages or proof assistants, such as Coq [The Coq development team 2016], Agda [Bove
et al. 2009], Idris [Brady 2017], and Isabelle [Nipkow et al. 2002], do verify their code, since their
tools allow it. However, adopting these platforms means rewriting everything from scratch. What
about the verification of existing code, such as libraries written in mature languages like Haskell?
Haskell programmers can reach for LiquidHaskell [Vazou et al. 2014] which smoothly integrates
the expressive power of refinement types with Haskell, using SMT solvers for fully automatic
verification. But some verification endeavors require the full strength of a mature interactive proof
assistant like Coq. The hs-to-coq tool, developed by Spector-Zabusky, Breitner, Rizkallah, and
Weirich [2018], translates Haskell types, functions and type classes into equivalent Coq code – a
form of shallow embedding – which can be verified just like normal Coq definitions.
But can this approach be used for more than the small, textbook-sized examples it has been
applied to so far? Yes, it can! In this work, we use hs-to-coq to translate and verify the two set
data structures from Haskell’s containers package.1 This codebase is not a toy. It is decades old,
highly tuned for performance, type-class polymorphic, and implemented in terms of low-level
features like bit manipulation operators and raw pointer equality. It is also an integral part of the
Haskell ecosystem. We make the following contributions:
1 Specifically,
we target version 0.5.11.0, which was released on January 22, 2018 and was the most recent release of this
library at the time of publication; it is available at https://github.com/haskell/containers/tree/v0.5.11.0.
Authors’ addresses: Joachim Breitner, [email protected], University of Pennsylvania, 3330 Walnut St, Philadelphia,
PA, 19104, USA; Antal Spector-Zabusky, [email protected], University of Pennsylvania, 3330 Walnut St, Philadelphia,
PA, 19104, USA; Yao Li, [email protected], University of Pennsylvania, 3330 Walnut St, Philadelphia, PA, 19104, USA;
Christine Rizkallah, [email protected], University of New South Wales, Sydney, NSW, Australia; John Wiegley, john.
[email protected], BAE Systems, USA; Stephanie Weirich, [email protected], University of Pennsylvania,
3330 Walnut St, Philadelphia, PA, 19104, USA.
2
Breitner, Spector-Zabusky, Li, Rizkallah, Wiegley and Weirich
• We demonstrate that hs-to-coq is suitable for the verification of unmodified, real-world
Haskell libraries. By “real-world”, we mean code that is up-to-date, in common use, and
optimized for performance. In Section 2 we describe the containers library in more detail
and discuss why it fits this description.
• We present a case study not just of verifying a popular Haskell library, but also of developing
a good specification of that library. This process is worth consideration because it is not at all
obvious what we mean when we say that we have “verified” a library. Section 4 discusses the
techniques that we have used to construct a rich, two-sided specification; one that draws
from diverse, cross-validated sources and yet is suitable for verification.
• We extend hs-to-coq and its associated standard libraries to support our verification goal. In
particular, in Section 5 we describe the challenges that arise when translating the Data.Set
and Data.IntSet modules, and our solutions. Notably, we drop the restriction in previous
work [Spector-Zabusky et al. 2018] that the input of the translation must be intrinsically
total. Instead, we show how to safely defer reasoning about incomplete pattern matching
and potential nontermination to later stages of the verification process.
• We increase confidence in the translation done of hs-to-coq. In one direction, properties
of the Haskell test suite turn into Coq theorems that we prove. In the other direction, the
translated code, when extracted back to Haskell, passes the original test suite.
• We provide new implementation-agnostic insight into the verification of the weight-balanced
tree data structure, as we describe in Section 6. In particular, we find the right precondition
for the central balancing operations needed to verify the particular variant used in Data.Set.
Our work provides a rich specification for Haskell’s finite set libraries that is directly and
mechanically connected to the current implementation. As a result, Haskell programmers can be
assured that these libraries behave as expected. Of course, there is a limit to the assurances that we
can provide through this sort of effort. We discuss the verification gap and other limitations of our
approach in Section 7.
We would like to have been able to claim the contribution of findings bugs in containers, but
there simply were none. Still, our efforts resulted in improvements to the containers library. First,
an insight during the verification process led to an optimization that makes the Data.Set.union
function 4% faster. Second, we discovered an incompleteness in the specification of the validity
checker used in the test suite.
The tangible artifacts of this work have been incorporated into the hs-to-coq distribution and
are available as open source tools and libraries.2
2
THE containers LIBRARY
We select the containers library for our verification efforts because it is a critical component of
the Haskell ecosystem. With over 4000 publicly available Haskell packages using on containers,
it is the third-most dependent-up package on the Haskell package repository Hackage, after base
and bytestring.3
The containers library is both mature and highly optimized. It has existed for over a decade
and has undergone many significant revisions in order to improve its performance. It contains
seven container data structures, covering support for finite sets (Data.Set and Data.IntSet),
finite maps (Data.Map and Data.IntMap), sequences (Data.Sequence), graphs (Data.Graph), and
trees (Data.Tree). However most users of the containers library only use the map and set
2 https://github.com/antalsz/hs-to-coq.
3 http://packdeps.haskellers.com/reverse
Ready, Set, Verify!
3
---------------------------------------------------------------------Sets are size balanced trees
-------------------------------------------------------------------data Set a = Bin {-# UNPACK #-} !Size !a !(Set a) !(Set a)
| Tip
type Size = Int
--| O(log n). Is the element in the set?
member :: Ord a => a -> Set a -> Bool
member = go
where go !_ Tip = False
go x (Bin _ y l r) = case compare x y of
LT -> go x l
GT -> go x r
EQ -> True
Fig. 1. The Set data type and its membership function5
modules;4 moreover, the map modules are essentially analogues of the set modules. Therefore, we
focus on Data.Set and Data.IntSet in this work.
2.1
Weight-balanced trees and big-endian Patricia trees
The Data.Set module implements finite sets using weight-balanced binary search trees. The
definition of the Set datatype in this module, along with its membership function, is given in
Figure 1. These sets and operations are polymorphic over the element type and require only that
this type is linearly ordered, as expressed by the Ord constraint on the member function. The member
function descends the ordered search tree to determine whether it contains a particular element.
The Size component stored with the Bin constructor is used by the operations in the library to
ensure that the tree stays balanced. The implementation maintains the balancing invariant
s 1 + s 2 ≤ 1 ∨ (s 1 ≤ 3s 2 ∧ s 2 ≤ 3s 1 ),
where s 1 and s 2 are the sizes of the left and right subtrees of a Bin constructor. This definition
is based on the description by Adams [1992], who modified the original weight-balanced tree
proposed by Nievergelt and Reingold [1972]. Thanks to this balancing, operations such as insertion,
membership testing, and deletion take time logarithmic in the size of the tree.
This type definition has been tweaked to improve the performance of the library. The annotations
on the Bin data constructor instruct the compiler to unpack the size component, removing a level
of indirection. The ! annotations indicate that all components should be strictly evaluated.
The Data.IntSet module also provides search trees, specialized to values of type Int to provide
more efficient operations, especially union. This implementation is based on big-endian Patricia
trees, as proposed in Morrison’s work on PATRICIA [1968] and described in a pure functional
setting by Okasaki and Gill [1998].
The definition of this data structure is shown in Figure 2. The core idea is to use the bits of the
stored values to decide in which subtree of a node they should be placed. In a node Bin p m s1 s2,
the mask m has exactly one bit set. All bits higher than the mask bit are equal in all elements of
calculated that 78% of the packages on Hackage that depend on containers use only sets and maps.
http://hackage.haskell.org/package/containers-0.5.11.0/docs/src/Data.Set.Internal.html#Set.
All code listings in this paper are manually reformatted and may omit module names from fully qualified names.
4 We
5 From
4
Breitner, Spector-Zabusky, Li, Rizkallah, Wiegley and Weirich
data IntSet = Bin {-# UNPACK #-} !Prefix {-# UNPACK #-} !Mask !IntSet !IntSet
--Invariant: Nil is never found as a child of Bin.
--Invariant: The Mask is a power of 2. It is the largest bit position at
-which two elements of the set differ.
--Invariant: Prefix is the common high-order bits that all elements share to
-the left of the Mask bit.
--Invariant: In Bin prefix mask left right, left consists of the elements
-that don't have the mask bit set; right is all the elements
-that do.
| Tip {-# UNPACK #-} !Prefix {-# UNPACK #-} !BitMap
--Invariant: The Prefix is zero for the last 5 (on 32 bit arches) or 6 bits
-(on 64 bit arches). The values of the set represented by a tip
-are the prefix plus the indices of the set bits in the bit map.
| Nil
--A number stored in a set is stored as
--* Prefix (all but last 5-6 bits) and
--* BitMap (last 5-6 bits stored as a bitmask)
-- Last 5-6 bits are called a Suffix.
type Prefix = Int
type Mask = Int
type BitMap = Word
type Key
= Int
Fig. 2. The IntSet data type6
that node; they form the prefix p. The mask bit is the highest bit that is not shared by all elements.
In particular, all elements in s1 have this bit cleared, while all elements in s2 have it set. When
looking up a value x, the mask bit of x tells us into which branch to descend.
Instead of storing a single value at the leaf of the tree, this implementation improves time and
space performance by storing the membership information of consecutive numbers as the bits of a
machine-word-sized bitmap in the Tip constructor.
The Nil constructor is the only way to represent an empty tree, and will never occur as the child
of a Bin constructor. Every well-formed IntSet is either made of Bins and Tips, or a single Nil.
2.2
A history of performance tuning
The history of the Data.Set module can be traced back to 2004, when a number of competing
search tree implementations were debated in the “Tree Wars” thread on the Haskell libraries mailing
list. Benchmarks showed that Daan Leijen’s implementation had the best performance, and it was
added to containers in 2005 as Data.Set.7
In 2010, Milan Straka thoroughly evaluated the performance of the containers library and
implemented a number of performance tweaks [Straka 2010]. This change8 replaced a fairly readable
balance and several small and descriptive helper functions with a single dense block of code. A
later change9 by Straka created two copies of this scary-looking balance, each specialized and
optimized for different preconditions.
6 From
http://hackage.haskell.org/package/containers-0.5.11.0/docs/src/Data.IntSet.Internal.html#IntSet
7 https://github.com/haskell/containers/commit/bbbba97c
8 https://github.com/haskell/containers/commit/3535fcbe
9 https://github.com/haskell/containers/commit/d17d7182
Ready, Set, Verify!
5
Adams [1992] describes two algorithms for union, intersection, and difference: “hedge
union” and “divide and conquer”. Originally containers used the former, but in 2016 its maintainers
switched to the latter,10 again based on performance measurement.
The module Data.IntSet (and Data.IntMap) has been around even longer. Okasaki and Gill
mention in their 1998 paper [Okasaki and Gill 1998] that GHC had already made use of IntSet and
IntMap for several years. In 2011, the Data.IntSet module was re-written to use machine-words
as bit maps in the leaves of the tree, as discussed at the end of Section 2.1.11 This moved the
containers library further away from the literature on Patricia trees and introduced a fair amount
of low-level bit twiddling operations (e.g., highestBitMask, lowestBitMask, and revNat).
3
OVERVIEW OF OUR VERIFICATION APPROACH
In order to verify Set and IntSet, we use hs-to-coq to translate the unmodified Haskell modules
to Gallina and then use Coq to verify the translated code. For example, consider the excerpt of the
implementation of Set in Figure 1. The hs-to-coq tool translates this input to the following Coq
definitions.12 The strictness and unpacking annotations are ignored, as they do not make sense in
Coq, and the type name Set is renamed to Set_ to avoid clashing with the Coq keyword.
Definition Size := GHC.Num.Int%type.
Inductive Set_ a : Type
:= Bin : Size -> a -> (Set_ a) -> (Set_ a) -> Set_ a
| Tip : Set_ a.
Definition member {a} `{GHC.Base.Ord a} : a -> Set_ a -> bool :=
let fix go arg_0__ arg_1__
:= match arg_0__, arg_1__ with
| _, Tip => false
| x, Bin _ y l r => match GHC.Base.compare x y with
| Lt => go x l
| Gt => go x r
| Eq => true
end
end
in go.
These definitions depend on hs-to-coq’s pre-existing translated version of GHC’s standard library
base. Here, we use the existing translation of Haskell’s Int type, the Ord type class, and Ord’s
compare method.
We carry out this translation for the Set and IntSet along with their attendant functions, and
then verify the resulting Gallina code. In Section 4 we discuss the properties that we prove about
the two data structures.
To further test that the translation from Haskell to Coq, we also used Coq’s extraction mechanism
to translate the generated Gallina code, like that seen above, back to Haskell. This process converts
the implicitly-passed type class dictionaries to ordinary explicitly-passed function arguments, but
otherwise preserves the structure of the code. By providing an interface that restores the typeclass based types, we can run the original containers test suite against this code. This process
10 https://github.com/haskell/containers/commit/c3083cfc
11 https://github.com/haskell/containers/pull/3
12 In
the file examples/containers/lib/Data/Set/Internal.v
Breitner, Spector-Zabusky, Li, Rizkallah, Wiegley and Weirich
Proofs
Gallina
Haskell
6
Set
IntSet
Tests
Set
IntSet
Headers, types and proof preparations
Functions and type class instances (verified)
(unverified)
"
(untranslated)
"
Tests and specifications
Tests
Lemmas about integers and bit-wise operations
Lemmas about lists and sortedness
A theory of dyadic intervals
Automation for reasoning about Ord
Set
IntSet
Instantiating Coq’s FSetInterface
Tests
0
500
1000
1500
2000
2500
3000
Lines of code
3500
4000
4500
5000
Fig. 3. A quantitative overview of the Haskell code, its translation into Coq and our proofs
helps us check that hs-to-coq preserves the semantics of the original Haskell program during the
translation process.
Figure 3 provides an overview of the Haskell code that we target, the Gallina code that we
translate it into, and the Coq proofs that we wrote. The set modules of the containers library
contain 149 functions and 22 type class instances, written in 2207 lines of code (excluding comments
and blank lines). Out of these, 15 functions and 11 type class instances (270 loc) were deemed “out
of scope” and not translated. (We discuss untranslated definitions in more detail in Sections 5.3
and 7.2.) Our translation produces 2709 lines of Gallina code.
The Set and IntSet data structures come with extensive APIs. We specify and verify a representative subset of commonly used functions (listed in Figure 4), covering 66% of the Set API and
49% of the IntSet API. This verified API is complete enough to instantiate Coq’s specification of
finite sets, along with many other specifications at varying levels of abstraction; for more detail,
see Section 4.
As Coq is not an automated theorem prover, verification of these complex data structures requires
significant effort. In total, we verified 1202 lines of Haskell; the verification of Set and IntSet
required 8.5 lines of proof per lines of code. This factor is noticeably higher for IntSet (10.0×) than
for Set (7.1×), as the latter is conceptually simpler to reason about and allowed us to achieve a
higher degree of automation using Coq tactics.
Our proofs also require the formalization of several background theories (not counted in the
proof-to-code ratio above), including: integer arithmetic and bits (1265 loc): lists and sortedness
(1489 loc); dyadic intervals, which are used for verifying IntSet (1169 loc); and support for working
with lawful Ord instances, including a complete decision procedure (453 loc).
Ready, Set, Verify!
7
Set: delete, deleteMax, deleteMin, difference, disjoint, drop, elems, empty, filter,
foldl, foldl’, foldr, foldr’, fromAscList, fromDescList, fromDistinctAscList,
fromDistinctDescList, insert, intersection, isSubsetOf, lookupMax, lookupMin,
mapMonotonic, maxView, member, minView, notMember, null, partition, singleton, size,
split, splitAt, splitMember, take, toAscList, toDescList, toList, union, unions
Instances: Eq, Eq1, Monoid, Ord, Ord1, Semigroup
Internal functions: balanceL, balanceR, combineEq, glue, insertMax, insertMin,
insertR, link, maxViewSure, merge, minViewSure, valid
IntSet: delete, difference, disjoint, elems, empty, filter, foldl, foldr, fromList,
insert, intersection, isProperSubsetOf, isSubsetOf, map, member, notMember, null,
partition, singleton, size, split, splitMember, toAscList, toDescList, toList,
union, unions
Instances: Eq, Monoid, Ord, Semigroup
Internal functions: branchMask, equal, highestBitMask, mask, nequal, nomatch,
revNat, shorter, valid, zero
Fig. 4. The verified subset of functions and type classes in Data.Set and Data.IntSet
4
SPECIFYING Set AND IntSet
The phrase “we have verified this piece of software” on its own is meaningless: the particular
specification that a piece of software is verified against matters. Good specifications are rich, twosided, formal, and live [Appel et al. 2017]. A specification is rich if it “describ[es] complex behaviors
in detail”. It is two-sided if it is “connected to both implementations and clients”. It is formal if it
is “written in a mathematical notation with clear semantics”. And it is live if it is “connected via
machine-checkable proofs to the implementation”.
All specifications of Set and IntSet that we use are formal and live by definition. They are
formal because we express the desired properties using Gallina, the language of the Coq proof
assistant; and they are live because we use hs-to-coq to automatically convert containers to
Coq where we develop and check our proofs. But how can we ensure that our specifications of Set
and IntSet are two-sided and rich? How do we know that the specifications are not just facts that
happen to be true, but are useful for the verification of larger systems? What complex behaviors of
the data structures should we specify?
To ensure that our specifications are two-sided, we use specifications that we did not invent
ourselves. Instead, we draw our specifications from a variety of diverse sources: from several parts
of the containers codebase (Sections 4.1 to 4.4), from Haskell type class laws (Section 4.5), from
pre-existing Coq theories (Section 4.6), and from a mathematical description of sets (Section 4.7).
This way, we also ensure that our specifications are rich because they describe the complex Set and
IntSet data structures at varying levels of abstraction. Finally, by verifying the code against these
disparate specifications, we not only increase the assurance that we captured all the important
behaviors of Set and IntSet, but we also cross-validate the specifications against each other.
4.1
Specifying implementation invariants
Set and IntSet are two examples of abstract types whose correctness depend on invariants.
Therefore, we define well-formedness predicates WF: Set_ e -> Prop and WF : IntSet -> Prop
and show that the operations preserve these properties. The definition of well-formedness differs
between the two types, but specifications of both are already present within containers.
8
Breitner, Spector-Zabusky, Li, Rizkallah, Wiegley and Weirich
Inductive Bounded : Set_ e -> option e -> option e -> Prop :=
| BoundedTip : forall lb ub,
Bounded Tip lb ub
| BoundedBin : forall lb ub s1 s2 x sz,
Bounded s1 lb (Some x) ->
Bounded s2 (Some x) ub ->
isLB lb x = true -> (* If lb is defined, it is less than x *)
isUB ub x = true -> (* If ub is defined, it is greater than x *)
sz = (1 + size s1 + size s2) ->
balance_prop (size s1) (size s2) -> (* weights of tree are balanced *)
Bounded (Bin sz x s1 s2) lb ub.
(** Any set that has bounds is well-formed *)
Definition WF (s : Set_ e) : Prop := Bounded s None None.
Fig. 5. Well-formed weight-balanced sets
Well-formed weight-balanced trees. Our definition of WF for Set is derived from the valid function
defined in the containers library. This function checks whether the input (1) is a balanced tree,
(2) is an ordered tree, and (3) has the correct values in its size fields. It is not part of the normal,
user-facing API of containers (since all exported functions preserve well-formedness), but is
used internally by the developers for debugging and testing. For us, it is valuable as an executable
specification, with less room for ambiguity and interpretation than comments and documentation.
However, rather than using valid directly, we define well-formedness as an inductive predicate,
because we find it more useful from a proof engineering perspective. In particular, our definition
of WF, as shown in Figure 5, relies on the Bounded inductive family. Its indices express lower and
upper bounds of the elements stored in the tree; None means unbounded. At the same time, the
property also checks that the sizes of the two subtrees are balanced in the Bin case.13 Nevertheless,
we can relate the WF predicate to the valid function found in containers:
Lemma Bounded_iff_valid : forall s, WF s <-> valid s = true.
Well-formed Patricia trees. Our well-formedness definition for IntSet is derived from the comments in the IntSet data type, shown in Figure 2, where the type declaration of almost disappears
beneath a large swath of comments describing its invariants.
In this case, the documentation-derived well-formedness predicate is stronger than the corresponding valid function from the implementation – the Haskell function was missing some
necessary conditions. We reported this to the library authors,14 who have since fixed valid.
This fix to valid is an example of how verification allows us to cross-validate specifications
and ensure that the invariants written in the comments are adequately reflected by the code. On
the other hand, sometimes we discover that it is the comments in the code that are incomplete.
For example, the comment describing the precondition for balanceL in the Data.Set module was
misleading and too vague; for more detail, see Section 6.1.
13 In
the file examples/containers/theories/SetProofs.v
14 https://github.com/haskell/containers/issues/522
Ready, Set, Verify!
4.2
9
Property-based testing
At the next level of verification, we would like to show that the implementations of Set and IntSet
are correct according to the implementors of the module. We specify correctness by deriving a
definition directly from the test suite that is distributed with the containers library.
Thanks to the popularity of property-based testing within the Haskell community, this test suite
contains a wealth of precisely specified general properties expressed using QuickCheck [Claessen
and Hughes 2000]. For example, one such property states that the union operation is associative:
prop_UnionAssoc :: IntSet -> IntSet -> IntSet -> Bool
prop_UnionAssoc t1 t2 t3 = union t1 (union t2 t3) == union (union t1 t2) t3
which we can interpret as a theorem about union:15
Theorem thm_UnionAssoc:
forall t1, WF t1 -> forall t2, WF t2 -> forall t3, WF t3 ->
union t1 (union t2 t3) == union (union t1 t2) t3 = true.
We do not have to write these theorems by hand: as we describe in Section 5.8, we use hs-to-coq in
a nonstandard way to automatically turn these executable tests into Gallina propositions (i.e. types).
We have have translated IntSet’s test suite in this manner and have proven that all QuickCheck
properties about verified IntSet functions are theorems (with one exception due to our choice of
integer representation – see Section 5.5).
4.3
Numeric overflow in Set
There is one way in which we have diverged from the specification of correctness given by the
comments of the containers library. The Data.Set module states:16
Warning: The size of the set must not exceed maxBound::Int. Violation of this condition is not detected and if the size limit is exceeded, its behavior is undefined.
In practice, it makes no difference whether Int is bounded or not, as a set with (263 − 1) elements
would require at least 368 exabytes of storage. What does this imply for our specification of
Set? Should we use fixed-width integers to represent Int? This choice would closely match the
implementation, but we would have to carefully add preconditions to all our lemmas to avoid integer
overflow, greatly complicating the proofs, with little verification insight to be gained. Furthermore,
such a specification would be difficult to use by clients, who themselves would need to prove that
they satisfy such preconditions.
Instead, we translate Haskell’s Int type to Coq’s type of unbounded integers (called Z). This
mapping avoids the problem of integer overflow altogether and is arguably consistent with the
comment, as this choice replaces undefined behavior with concrete behavior. (The situation is
slightly different for IntSet; see Section 5.5.)
4.4
Rewrite rules
So far, we have only been concerned with specifying the correctness of the data structures using
the definition of correctness that is present in the original source code; the comments, the valid
functions and the QuickCheck properties. However, to be able to claim that our specifications are
two-sided, we need to show that the properties that we prove are useful to clients of the module.
One source of such properties is rewrite rules [Peyton Jones et al. 2001]. The containers library
includes a small number of such rules. These annotations instruct the compiler to replace any
15 In
the file examples/containers/theories/IntSetPropertyProofs.v
16 http://hackage.haskell.org/package/containers-0.5.11.0/docs/Data-Set.html
10
Breitner, Spector-Zabusky, Li, Rizkallah, Wiegley and Weirich
Class OrdLaws (t : Type) {HEq : Eq_ t} {HOrd : Ord t} {HEqLaw : EqLaws t} := {
(* The axioms *)
Ord_antisym
: forall a b, a <= b = true -> b <= a = true -> a == b = true;
Ord_trans_le : forall a b c, a <= b = true -> b <= c = true -> a <= c = true;
Ord_total
: forall a b, a <= b = true \/ b <= a = true;
(* The other operations, in terms of <= or == *)
Ord_compare_Lt : forall a b, compare a b = Lt <-> b <= a = false;
Ord_compare_Eq : forall a b, compare a b = Eq <-> a == b = true;
Ord_compare_Gt : forall a b, compare a b = Gt <-> a <= b = false;
Ord_lt_le
: forall a b, a < b = negb (b <= a);
Ord_ge_le
: forall a b, a >= b =
(b <= a);
Ord_gt_le
: forall a b, a > b = negb (a <= b)}.
Fig. 6. Our codification of the Ord type class laws
occurrence of the pattern on the left-hand side in the rule by the expression on the right hand side.
The standard example is
{-# RULES "map/map" forall f g xs. map f (map g xs) = map (f . g) xs #-}
which fuses two adjacent calls to map into one, eliminating the intermediate list.
The program transformation list fusion [Peyton Jones et al. 2001] is implemented completely in
terms of rewrite rules, and the rules in the containers library setup its functions for fusion; an
example is
{-# RULES "Set.toAscList" forall s . toAscList s = build (\c n -> foldrFB c n s) #-}
which transforms the toAscList function into an equivalent representation in terms of build.
We can view these rewrite rules as a direct specification of properties that the compiler assumes
are true during compilation. Rewrite rules are used by GHC during optimization; if any of these
properties are actually false, GHC will silently produce incorrect code. Therefore, any proof about
the correctness of GHC’s compilation of these files depends on a proof of these properties. We have
manually translated all the rules into Coq – there are only few of them, so manual translation is
viable – and have proved that the translated operations satisfy this specification.17
4.5 Type classes with laws
Many Haskell type classes come with laws that all instances of the type class should satisfy, which
provides another source of external specification that we can use. For example, an instance of Eq is
expected to implement an equivalence relation, an instance of Ord should describe a linear order,
and an instance of Monoid should be, well, a monoid.
We reflect these laws using type classes whose members are the required properties. For example,
we have defined EqLaws, OrdLaws (shown in Figure 6), and MonoidLaws. These classes can only be
instantiated if the corresponding instance is lawful.
Even though we have defined these laws ourselves, using our understanding of what they mean
for the Haskell standard library, we argue that they form a two-sided specification. In particular, we
have been clients to the Ord laws in our verification of the Set data structure. Almost all theorems
about Set must constrain the element type to one that is an instance of OrdLaws. Therefore, we
know our specification of these laws is sufficiently strong to verify this library.
17 In
the files examples/containers/theories/SetProofs.v and examples/containers/theories/IntSetProofs.v
Ready, Set, Verify!
11
At the same time, we have also shown that multiple type class instances satisfy these laws
including both set modules18 and other types such as Z, unit, tuples, option, and list.19 Because
we successfully instantiated these type classes for many types, we also know that they are not too
strong.
That said, nailing down the precise form of the type class laws can be tricky. Consider the case
of a Monoid instance for a type T. The associativity law can be stated as “for all elements x, y and z
of T, we have that x <> (y <> z) is equal to (x <> y) <> z.” But in order to write this down as part
of MonoidLaws in Coq, we need to make two decisions:
(1) What do we mean by “equal”? The first option is to use Coq’s propositional equality and
require that x <> (y <> z) = (x <> y) <> z. This would, however, prevent us from making
Set_, with (<>) = union, a member of MonoidLaws: two extensionally-equal sets may be
represented by differently structured trees. Therefore, we instead require that the two expressions are equal according to their Eq instance:(x <> (y <> z) == (x <> y) <> z) = true. The
tradeoff with this approach is that it precludes instances like MonoidLaws b -> MonoidLaws
(a -> b), since functions have no instance of Eq (and indeed, cannot have decidable equality).
For many types, however, this distinction is moot, since Haskell’s equality coincides with
structural equality; for example, this is the case for Bool, for Integer, and for IntSet itself
(although not for Set, as mentioned above). To facilitate reasoning about such types, we
provide the type class EqExact, which states that
forall {a} `{EqExact a}, x == y = true <-> x = y
(2) What do we mean by “For all elements of T”? The obvious choice is universal quantification
over all elements of T:
forall (x y z : T), x <> (y <> z) == (x <> y) <> z = true
But again, this collides with common practice in Haskell. Once again, consider Set: union
only works correctly on well-formed sets. Therefore, our approach is to define an instance
of this and other classes not for the type Set_ e, but for the type of well-formed sets, {s :
Set_ e | WF s}, where type class laws hold universally. This instance reflects the “external
view” of the data structure – clients should only have access to well-formed sets.
An alternative could be to instead constrain MonoidLaws’s theorems to hold only on members
of T that are well-formed in some general way (e.g., according to an IsWF type class that could
be instantiated at different types). In this way, we could instantiate MonoidLaws directly with
types that require well-formedness, without the need for subset types.
4.6
Specifications from the Coq standard library
Because we are working in Coq, we have access to a standard library of specifications for finite sets,
which we know are two-sided because they have already appeared in larger Coq developments. The
Coq.FSets.FSetInterface module20 provides module types that cover many common operations
and their properties. The module types come in two varieties: one that specifies sets of elements
that merely have decidable equality (WSfun, WS), and one that specifies sets of elements that can
be linearly ordered (Sfun, S). The WSfun and Sfun modules are presented as module functors that
take an OrderedType module, containing the linearly-ordered element type, as an input; the WS
and S modules are the same, but they inline this information.
18 In
the files examples/containers/theories/SetProofs.v and examples/containers/theories/IntSetProofs.v
the file base-thy/GHC/Base.v
20 https://coq.inria.fr/library/Coq.FSets.FSetInterface.html
19 In
12
Breitner, Spector-Zabusky, Li, Rizkallah, Wiegley and Weirich
For example, the parameterized module type WSfun provides one specification of a finite set type,
called t in the excerpt from this interface below. The element type of this set, E.t, is required to
have decidable equality.
Module Type WSfun (E : DecidableType).
Definition elt := E.t.
Parameter t : Type.
(* Set type *)
Parameter In : elt -> t -> Prop. (* Characteristic function for the Set *)
Parameter mem : elt -> t -> bool. (* Membership function *)
(* Specification of mem *)
Parameter mem_1 : forall x s, In x s -> mem x s = true.
Parameter mem_2 : forall x s, mem x s = true -> In x s.
...
End WSFun.
Every operation in this interface, such as mem above, is accompanied by a small number of properties
that specify the behavior of the operation.
We instantiate all four interfaces for Set and IntSet.21 For example, the instance for Set starts
out:
Module SetFSet (E : OrderedType) <: WSfun(E) <: WS <: Sfun(E) <: S.
Definition t := {s : Set_ elt | WF s}.
Program Definition In (x :elt) (s : t) : Prop := . . . .
Program Definition mem : elt -> t -> bool := member.
Lemma mem_1 : forall (s : t) (x : elt), In x s -> mem x s = true.
Proof. . . . (* Proof may assume that s is well-formed *) . . . Qed.
...
End SetFSet.
Instantiating these interfaces runs into two small hiccups. The first is that they talk about all
sets, not simply all well-formed sets. Therefore, as in the previous section, we instantiate these
interfaces with the subset type {s : Set_ e | WF s}. The second is that Coq’s module system does
not interact with type classes, and Set_ is defined such that its element type must be an instance
of the Ord type class. This impedance mismatch requires us to write a module which can generate
an Ord instance from a Coq OrderedType module.
By successfully instantiating this module interface, we obtain two benefits. First, we must
prove theorems that cover many of the main functions provided by containers; these theorems
are particularly valuable, since as the interface itself is heavily used by Coq users. Second, by
instantiating this interface, we connect our injected Haskell code to the Coq ecosystem, enabling
Coq users to easily use the containers-derived data structures in their developments, should they
so desire.
4.7
Abstract models as specifications
Tests, type classes and the other sources of specifications do not fully describe the intended behavior
of all functions. We therefore also have to also come up with specifications on our own. We do
21 In
the files examples/containers/theories/SetProofs.v and examples/containers/theories/IntSetProofs.v
Ready, Set, Verify!
13
this by relating a concrete search tree to the abstract set that it represents; that is, we provide a
denotational semantics. We denote a set with elements of type e as its indicator function of type
e -> bool; for Set e, we provide a denotation function sem : forall {e} `{Eq_ e}, Set_ e ->
(e -> bool), and for IntSet, we provide a denotation relation Sem : IntSet -> (N -> bool) ->
Prop.
This approach allows us to abstractly describe the meaning of operations like insert. For Set_,
we do this by providing a theorem like
Theorem insert_sem:
forall {a} `{OrdLaws a} (s : Set_ a) (x : a), WF s ->
forall (i : a), sem (insert x s) i = (i == x) || sem s i.
(For IntSet, some techincal details differ.) However, there is more we need to know about insert
than just its denotation. We also need to know that it preserves well-formedness and bounds, and
– to reason about balancing – its size. To avoid having to prove these properties independently,
we define a relation Desc that completely describes a set, by asserting that it is well-founded and
relating it to bounds, its size and its denotation:
Definition Desc (s : Set _e) (lb ub : option e) (sz : Z) (f : e -> bool) :=
Bounded s lb ub /\ size s = sz /\ (forall i, sem s i = f i).
This allows us to state a single theorem about insert, namely
Lemma insert_Desc: forall x s lb ub,
Bounded s lb ub ->
isLB lb x = true -> (* If lb is defined, it is less than x *)
isUB ub x = true -> (* If ub is defined, it is greater than x *)
Desc (insert x s) lb ub
(if sem s y then size s else (1 + size s)) (fun i => (i == x) || sem s i).
and prove everything we need to know about insert in one single inductive proof.22
Since Desc describes insert completely, it introduces a layer of abstraction that we can build
upon. In fact, we specify all functions this way, and use these specifications, rather than the concrete
implementation, to prove the other specifications. (We have an analogous Desc relation for IntSet
that describes the properties of Patricia trees.)
An alternative abstract model for finite sets is the sorted list of their elements, i.e. the result of
toAscList. The meaning of certain operations, like foldr, take or size, can naturally be expressed
in terms of toAscList, but would be very convoluted to state in terms of the indicator function,
and we use this denotation – or both – where appropriate.
5
PRODUCING VERIFIABLE CODE WITH hs-to-coq
Identifying what to prove about the code is only half of the challenge – we also need to get the
Haskell code into Coq. Ideally, the translation of Haskell code into Gallina using hs-to-coq would
be completely automatic and produce code that can be verified as easily as code written directly in
Coq – and for textbook-level examples, that is the case [Spector-Zabusky et al. 2018]. However,
working with real-world code requires adjustments to the translation process to make sure that the
output is both accepted by Coq and amenable to verification.
A core principle of our approach is that the Haskell source code does not need to be modified in
order to be verified. This principle ensures that we verify “the” containers library (not a “verified
fork”) and that the verification can be ported to a newer version of the library.
22 In
the file examples/containers/theories/SetProofs.v
14
Breitner, Spector-Zabusky, Li, Rizkallah, Wiegley and Weirich
The crucial feature of hs-to-coq that enables this approach is the support for edits: instructions
to treat some code differently during translation. Edits are specified in plain text files, which also
serve as a concise summary of our interventions. The hs-to-coq tool already supported many
forms of edits; for example, specifying when names need to be changed, when parts of the module
should be ignored or replaced by some other term, when we want to map Haskell types to existing
Coq types, or when a recursive function definition needs an explicit termination proof. In the
course of this work, we added new features to hs-to-coq – such as the ability to apply rewrite
rules, to handle partiality and to defer termination proofs to the verification stage – and extended
the provided base library.
In this section we demonstrate some of the challenges posed by translating real-world code, and
show how hs-to-coq’s flexibility allowed us to not only to overcome them, but also to facilitate
subsequently proving the input correct.
5.1
Unsafe pointer equality
An example of a Haskell feature that we cannot expect to translate without intervention is unsafe
pointer equality. GHC’s runtime provides the scarily named function reallyUnsafePtrEqualty#,
which the containers library wraps as ptrEq :: a -> a -> Bool. If this function returns True,
then both arguments are represented in memory by the same pointer. If this function returns False,
we know nothing – this function is underspecified and may return False even if the two pointers
are equal. This operation is used, for example, in Set.insert x s when x is already a member of s:
If ptrEq indicates that the subtree is unchanged, the function skips the redundant re-balancing
step – which enhances performance – and returns the original set rather than constructing a
semantically equivalent copy – which increases sharing.
Coq does not provide any way of reasoning about memory, so when we use hs-to-coq, we must
replace ptrEq with something else. But what?
One option is to replace the definition of ptrEq with a definition that does not do any computation
and simply always returns False. This can be done using the following edit:
replace Definition ptrEq : forall {a}, a -> a -> bool := fun _ _ _ => false.
This replacement would behave in a way that is consistent with reallyUnsafePtrEqualty#
and allows us to proceed with translation and verification. However, the code in the True branch
of an unsafe pointer equality test would be dead code in Coq, and our verification would miss bugs
possibly lurking there.
Consequently, we choose a different encoding that captures the semantics of ptrEq more precisely.
We make the definition of ptrEq opaque and partially specify its behavior.23
Definition ptrEq_spec :
{ptrEq : forall a, a -> a -> bool | forall a (x y : a), ptrEq _ x y = true -> x = y}.
Proof. apply (exist _ (fun _ _ _ => false)). intros; congruence. Qed.
Definition ptrEq : forall {a}, a -> a -> bool := proj1_sig ptrEq_spec.
Lemma ptrEq_eq : forall {a} (x:a)(y:a), ptrEq x y = true -> x = y.
Proof. exact (proj2_sig ptrEq_spec). Qed.
Here, we define the ptrEq function together with a specification as ptrEq_spec. Although the
function in ptrEq_spec also always returns false, the Qed at the end of its definition completely
hides this implementation of ptrEq. While the specification ptrEq_eq is vacuously true, making
23 In
the file examples/containers/lib/Utils/Containers/Internal/PtrEquality.v
Ready, Set, Verify!
15
Set: deleteAt, deleteFindMax, deleteFindMin, elemAt, findIndex, findMax, findMin
Instances: Data, IsList, NFData, Read, Show, Show1
Internal functions: showTree
IntSet: deleteFindMax, deleteFindMin, findMax, findMin, fromAscList,
fromDistinctAscList
Instances: Data, IsList, NFData, Read, Show
Internal functions: showTree
Fig. 7. Untranslated functions and type classes in Data.Set and Data.IntSet
this definition opaque forces verification to proceed down both paths. While could achieve the
same using Axiom, our variant protects us from accidentially introducing inconsistencies to Coq.
We do have to trust that GHC’s definition of pointer equality has this specification, but given
this assumption, we can soundly verify that pointer equality is used correctly.
5.2
Evaluation order
A shallow embedding of Haskell into Coq makes the difference between strict and lazy code vanish,
because Gallina is a total language and does not care about evaluation order. Danielsson et al.
[2006] show that such “fast and loose” reasoning does not invalidate our theorems.
Haskell has “magic” functions like seq that allow the programmer to explicitly control strictness,
and the containers library uses it. Its effect is irrelevant in Coq, and we instruct hs-to-coq to
use this simple, magic-free implementation for it:
Definition seq {a} {b} (x : a) (y : b) := y.
5.3
Eliminating unwanted parts of the code
Figure 7 lists the untranslated portions of the Set and IntSet modules. Many of these operations
are functions that we choose to ignore for the sake of verification – for example, the function
showTree in Data.Set prints the internal structure of such a set as an ASCII-art tree. This function
is not used elsewhere in the module. In the interest of a tidier and smaller output, we skip this
function using an edit:
skip showTree
Similarly, we skip functionality related to serialization (the Show and Show1 type classes), deserialization (the Read type class), generic programming (the Data type class), and overloaded list
notation (the IsList type class).
Furthermore, we skip some operations whose public API is partial. For example, evaluating
findMax empty will raise an exception, as the empty set has no maximum element. We cannot
model this exception in Coq, so we skip findMax and similar functions (findMin, deleteFindMax,
deleteFindMin, findIndex, elemAt and deleteAt). This elision is not significant because the
containers API provides total equivalents for many of these functions (e.g., lookupIndex, which
returns Nothing when the index is out of bounds).
Finally, we skip the two functions that use mutual recursion (Data.IntSet.fromAscList and
Data.IntSet.fromDistinctAscList) as this is not yet supported by hs-to-coq.
16
5.4
Breitner, Spector-Zabusky, Li, Rizkallah, Wiegley and Weirich
Partiality in total functions
In contrast to the skipped functions above, some functions use partiality in their implementation
in ways that cannot be triggered by a user of the public API. In particular, they may use calls to
Haskell’s error function when an invariant is violated.
For example, the central balancing functions for Sets, balanceL and balanceR, may call error
when passed an ill-formed Set. Because our proofs only reason about well-formed sets, this code is
actually dead. It does not matter how we translate error – any term that is accepted by Coq is
good enough. However, error in Haskell has the type
error :: String -> a
which means that a call to error can inhabit any type. We cannot define such a function in Coq,
and adding it as an axiom would be glaringly unsound.
Therefore, we extended hs-to-coq to use the following definition for error:
Class Default (a :Type) := { default : a }.
Definition error {a} `{Default a} : String -> a.
Proof. exact (fun _ => default). Qed.
The type class enforces that we use error only at non-empty types, ensuring logical consistency.
Yet we will notice that something is wrong when we have to prove something about it. Just as with
ptrEq (Section 5.1), by making the definition of error opaque using Qed, we are prevented from
accidentally or intentionally using the concrete default value of a given type in a proof about
error. Furthermore, when we extract the Coq code back to Haskell for testing, we translate this
definition back to Haskell’s error function, preserving the original semantics.
This encoding is inspired by Isabelle, where all types are inhabited and there is a polymorphic
term undefined :: a that denotes an unspecified element of any type.
5.5 Translating the Int in IntSet
As discussed in Section 4.1, we map Haskell’s finite-width integer type Int to Coq’s unbounded
integer type Z in the translation of Data.Set in order to match the specification that integer
overflow is outside the scope of the specified behavior.
For IntSet, however, this choice would cause problems. Big-endian Patricia trees require that
two different elements have a highest differing bit. This is not the case for Z, where negative
numbers have an infinite number of bits set to 1; for instance, -1 is effectively an infinite sequence
of set bits. Fortunately, hs-to-coq is flexible enough to allow us to make a different choice when
translating IntSet; we can pick any suitable type where all elements have a finite number of bits
set, such as the natural numbers (N) or a fixed width integer type.
Given that Coq’s standard library provides a fairly comprehensive library of lemmas about N and
decision procedures (omega and lia) that work with it, we chose to use N for now, with the intention
to eventually switch to a 64-bit integer type. This is the appropriate generalization of IntSet to an
infinite domain. In Haskell, the domain is 64-bit words, which happen to be interpretable as negative
numbers. When we generalize to an infinite domain, we generalize to bit strings of unbounded but
finite length, which we can most simply interpret as nonnegative.
The IntSet code uses bit-level operations, like complement and negate, that do not exist for N.
To deal with this we extended hs-to-coq with support for rewrite edits like
rewrite forall x y, (x .&. complement y) = (xor x (x .&. y))
which instruct it to replace any expression that matches the left-hand side by the right hand side.
For signed or bounded integer types, both sides are equivalent. For unbounded unsigned types,
Ready, Set, Verify!
17
like Coq’s type N, the left hand side is undefined (values in N have no complement in N), while the
right hand side is perfectly fine. When we switch to bounded integers in the IntSet code, we can
remove these edits.
5.6
Low-level bit twiddling
The containers library uses highly tuned bit-twiddling algorithms to operate on IntSets. For
example, the function revNat reverses the order of the bits in a 64-bit number:
revNat :: Nat -> Nat
revNat x1 = case ((x1 `shiftRL` 1) .&. 5555555555555555) .|.
((x1 .&. 5555555555555555) `shiftLL` 1) of x2 ->
case ((x2 `shiftRL` 2) .&. 3333333333333333) .|.
((x2 .&. 3333333333333333) `shiftLL` 2) of x3 ->
case ((x3 `shiftRL` 4) .&. 0F0F0F0F0F0F0F0F) .|.
((x3 .&. 0F0F0F0F0F0F0F0F) `shiftLL` 4) of x4 ->
case ((x4 `shiftRL` 8) .&. 00FF00FF00FF00FF) .|.
((x4 .&. 00FF00FF00FF00FF) `shiftLL` 8) of x5 ->
case ((x5 `shiftRL` 16) .&. 0000FFFF0000FFFF) .|.
((x5 .&. 0000FFFF0000FFFF) `shiftLL` 16) of x6 ->
(x6 `shiftRL` 32) .|. (x6 `shiftLL` 32)
Though complicated, this code is within the scope of what hs-to-coq can translate, and we can
verify its correctness.
However, we can’t keep up with all their tricks. For example, indexOfTheOnlyBit, which was
contributed by Edward Kmett,24 takes a number with exactly one bit set and calculates the index of
said bit. It does so by unboxing the input, multiplying it by a magic constant, and using the upper 6
bits of the product as an index into a table stored in an unboxed array literal. This manifests as the
following scary-looking code:
indexOfTheOnlyBit :: Nat -> Int
indexOfTheOnlyBit bitmask = I# (lsbArray `indexInt8OffAddr#` unboxInt
(intFromNat ((bitmask * magic) `shiftRL` offset)))
where
unboxInt (I# i) = i
magic
= 0x07EDD5E59A4E28C2
offset
= 58
!lsbArray
= "\63\0\58\1\5. . . 15\8\23\7\6\5"#
We currently cannot translate this code because hs-to-coq does not yet support unboxed arrays
or unboxed integers. We therefore replace it with a simple definition based on a integer logarithm
function provided by Coq’s standard library:
redefine Definition indexOfTheOnlyBit := fun x => N.log2 x.
Similarly, we provide simpler definitions for the low-level bit-twiddling functions branchMask,
mask, zero and suffixBitMask.
5.7
Non-trivial recursion
In order to prove the correctness of Set and IntSet, we must deal with termination. There are
two reasons for this. First, we intrinsically want to prove that none of the functions provided by
containers go into an infinite loop. Second, Coq requires that all defined functions are terminating,
24 https://github.com/haskell/containers/commit/e076b33f
18
Breitner, Spector-Zabusky, Li, Rizkallah, Wiegley and Weirich
--Less obvious structural recursion
link :: a -> Set a -> Set a -> Set a
link x Tip r = insertMin x r
link x l Tip = insertMax x l
link x l@(Bin sizeL y ly ry) r@(Bin sizeR z lz rz)
| delta*sizeL < sizeR = balanceL z (link x l lz) rz
| delta*sizeR < sizeL = balanceR y ly (link x ry r)
| otherwise
= bin x l r
--Well-founded recursion
foldlBits :: Int -> (a -> Int -> a) -> a -> Nat -> a
foldlBits prefix f z bitmap = go bitmap z
where go 0 acc = acc
go bm acc = go (xor bm bitmask) ((f acc) $! (prefix+bi))
where !bitmask = lowestBitMask bm
!bi = indexOfTheOnlyBit bitmask
--Deferred recursion
fromDistinctAscList :: [a] -> Set a
fromDistinctAscList [] = Tip
fromDistinctAscList (x0 : xs0) = go (1::Int) (Bin 1 x0 Tip Tip) xs0
where go !_ t []
=t
go s l (x : xs) = case create s xs of (r :*: ys) ->
let !t' = link x l r in go (s `shiftL` 1) t' ys
create !_ [] = (Tip :*: [])
create s xs@(x : xs')
| s == 1
= (Bin 1 x Tip Tip :*: xs')
| otherwise = case create (s `shiftR` 1) xs of
res@(_ :*: []) -> res
(l :*: (y:ys)) -> case create (s `shiftR` 1) ys of
(r :*: zs) -> (link y l r :*: zs)
Fig. 8. Recursion styles
as unrestricted recursion would lead to logical inconsistencies. Depending on how involved the
termination argument for a given function is, we use one of the following four approaches.
5.7.1 Obvious structural recursion. By default, hs-to-coq implements recursive functions directly using Coq’s fix keyword. This works smoothly for primitive structural recursion; indeed, a
majority of the recursive functions that we encountered, such as member in Figure 1, were of this
form and required no further attention.
5.7.2 Almost-structural recursion. Another common recursion pattern can be found in binary
operations such as link in Figure 8. Here, every recursive call shrinks either or both of its arguments
to immediate subterms of the originals, leaving the others unchanged. This almost-structural
recursion is already beyond the capabilities of Coq’s termination checker. We therefore instruct
hs-to-coq to use Coq’s Program Fixpoint command to translate these functions in terms of
well-founded recursion by adding the edits
termination link {measure (size_nat arg_0__ + size_nat arg_1__)}
obligations link termination_by_omega
Ready, Set, Verify!
19
This specifies both:
(1) The termination measure, which is the sum of the sizes of the arguments (we defined the
function size_nat : IntSet -> nat).
(2) The termination proof that the measure decreases on every call. This is represented as the
Coq tactic termination_by_omega, which is a thin wrapper we defined around omega, a
Coq tactic to decide linear integer arithmetic.
We can use these edits (with size_nat and termination_by_omega) to get to Coq accepts such
recursive definitions without the need for any further proofs or manual intervention.
5.7.3 Well-founded recursion. A small number of functions recurse in a non-structural way, such
as foldlBits in Figure 8, which recurses on the input after clearing the least-significant set bit (bm
`xor` (lowestBitMask bm)). We can handle this sort of logic using the same machinery as before,
but now we have to write a specialized termination tactic and declare it in the obligations hint.
To do so, we need to prove necessary lemmas before translating the Haskell module in question.
Coq’s Program Fixpoint only supports top-level functions, but we frequently encounter local
recursive functions – the go idiom, as seen here. To support this, we extended hs-to-coq to offer
some of the convenience provided by Program Fixpoint by translating local recursive functions
using the same well-founded-recursion-based fixed-point combinator as Program Fixpoint.
5.7.4 Deferred recursion. Finally, we encounter some functions that require elaborate termination arguments, such as fromDistinctAscList in Figure 8. It has two local recursive functions,
go and create, and to convince ourselves that fromDistinctAscList is indeed terminating, we
have to reason as follows:
The function create bitshifts its first argument to the right upon each recursive call,
until the argument is 1. Ergo, it is terminating, but only for positive input – it clearly
loops if x is 0. The function go recurses on smaller lists as its third argument, but to see
that, we first have to convince ourself that the list in the tuple returned by create is
never larger than the list passed to create. Also, go calls create, so we need to ensure
that the s passed to it is positive. The go function bitshifts s to the left at every call, so
if go is called with a positive s, then s will remain positive in recursive calls. Finally,
we see that fromDistinctAscList calls go with s equal to 1, which is positive, so we
can conclude that fromDistinctAscList terminates.
If we wanted to convince Coq of this termination pattern, we would have to turn create and
go into top-level definitions, change their types to rule out invalid (non-positive) inputs, define
create using Program Fixpoint, and provide an explicit termination argument by well-founded
recursion. Then we could prove that create preserves the list lengths, which we need to define
go, again using Program Fixpoint. This is certainly possible, but it is not simple, especially in an
automatic translation.
For hard cases like these we resort to deferred termination checking, a feature that we added
to hs-to-coq. We can instruct it to use the following axiom as a very permissive fixed-point
combinator and translate the code of fromDistinctAscList essentially unchanged:
Axiom deferredFix: forall {a r} `{Default r}, ((a -> r) -> (a -> r)) -> a -> r.
On its own, deferredFix does not do anything; it merely sits in the translated code applied to
the original function body. It is consistent, since its type could be implemented by a function
that always returns default (see Section 5.4). And it does not prevent the user from running
extracted code – we can extract this axiom to the target language’s unrestricted fixpoint operator
20
Breitner, Spector-Zabusky, Li, Rizkallah, Wiegley and Weirich
(e.g., Data.Function.fix in Haskell), although this costs us the guarantee that the extracted code
is terminating.
When we come to verifying something about a function that is defined using deferredFix, we
need to give it meaning. We do so using a second axiom, deferredFix_eq_on, which states that
for any well-founded relation R (well_founded R), if the recursive calls in f are always at values
that are strictly R-smaller than the input (recurses_on R), then we may unroll the fixpoint of f:
Definition recurses_on {a b}
(P : a -> Prop) (R : a -> a -> Prop) (f : (a -> b) -> (a -> b))
:= forall g h x, P x -> (forall y, P y -> R y x -> g y = h y) -> f g x = f h x.
Axiom deferredFix_eq_on: forall {a b}
`{Default b} (f : (a -> b) -> (a -> b)) (P : a -> Prop) (R : a -> a -> Prop),
well_founded R -> recurses_on P R f ->
forall x, P x -> deferredFix f x = f (deferredFix f) x.
The predicate P : a -> Prop allows us to restrict the domain to inputs for which the function is
actually terminating – crucial for go and create.
The predicate recurses_on P R f characterizes the recursion pattern of f, but does so in a very
extensional way and only considers recursive calls that can actually affect the result of the function.
For instance, it would consider a recursive function defined by f n = if f n then true else true
to be terminating.
We can use this non-trivial axiom without losing too much sleep, because we can actually
implement deferredFix and deferredFix_eq_on in terms of classical logic and the axiom of
choice (as provided by the Coq module Coq.Logic.Epsilon).25 This means that both axioms are
consistent with plain Coq. We do not know if deferredFix_eq_on is strictly weaker than classical
choice, so users of hs-to-coq who want to combine the output of hs-to-coq with developments
known to be inconsistent with classical choice (e.g., homotopy type theory) should be cautious.
Pragmatically, working with deferredFix is quite convenient, as we can prove termination
together with the other specifications about these functions. The actual termination proofs themselves are not fundamentally different from the proof obligations that Program Fixpoint would
generate for us and – although not needed in this example – can be carried out even for nested
recursion through higher-order functions like map.
This approach to defining recursive functions extensionality was inspired by Isabelle’s function
package [Krauss 2006].
5.8
Translating Haskell tests to Coq types
When we translate code, we usually want to preserve the semantics of the code as much as possible.
Things are very different when we translate the QuickCheck tests defined in the containers test
suite, as we discuss in Section 4.2: whereas the semantics of the test suite in Haskell is a program
that creates random input to use as input test executable properties, we want to re-interpret these
properties as logical propositions in Coq. Put differently, we are turning executable code into types.
QuickCheck’s API provides types and type classes for writing property-based tests. In particular,
it defines an opaque type Property that describes properties that can checked using randomized
testing, a type class Testable that converts various testable types into a Property, and a type
constructor Gen that describes how to generate a random value. These are combined, for instance,
in the QuickCheck combinator
forAll :: (Show a, Testable prop) => Gen a -> (a -> prop) -> Property
25 In
the file base/GHC/DeferredFixImpl.v
Ready, Set, Verify!
21
which tests the result of a function on inputs generated by the given generator.
Since we want to prove, not test, these properties, we do not convert the QuickCheck implementation to Coq. Instead, we write a small Coq module that provides the necessary pieces of the interface
of Test.QuickCheck, but interprets these types and functions in terms of Coq propositions. In
particular:
• We use Coq’s non-computational type of propositions, Prop, instead of QuickCheck’s computational Property;
• Gen a is simply a wrapper around a logical predicate on a; and
• forAll quantifies (using Coq’s forall) over the type a, and ensures that the given function
– now a predicate – holds for all members of a that satisfy the given “generator”.
Concretely, this leads to the following adapted Coq code:
Record Gen a := MkGen { unGen : a -> Prop }.
Class Testable (a : Type) := { toProp : a -> Prop }.
Definition forAll {a prop} `{Testable prop} (g : Gen a) (p : a -> prop) : Prop :=
forall (x : a), unGen g x -> toProp (p x).
We provide similar translations for QuickCheck’s operators ===, ==>, .&&. and .||., and we replace
generators such as choose :: Random a => (a, a) -> Gen a with their corresponding predicates.
With this module in place, hs-to-coq translates the test suite into a “proof suite”. As we saw in
Section 4.2, a test like prop_UnionAssoc is now a definition of a Coq proposition, that is to say a
type, and can be used as the type of a theorem:
Theorem thm_UnionAssoc : toProp prop_UnionAssoc.
6
CONTRIBUTIONS TO containers: THEORY AND PRACTICE
We chose the Set and IntSet modules of the containers library as our target because they nicely
represent the kind of Haskell code that we want to see verified in practice. Nevertheless, the deep
understanding required to verify these modules led to new insights into the algorithms themselves
and to improvements to their Haskell implementation.
6.1
New insight into the verification of weight-balanced trees
The data structure underlying Set and Map was originally presented by Nievergelt and Reingold
[1972]. It is a binary search tree with the invariant that if s 1 is the size of the left subtree and s 2 the
size of the right subtree of a node, then
balNR (s 1 , s 2 ) B (s 1 + 1) ≤ δ · (s 2 + 1) ∧ (s 2 + 1) ≤ δ · (s 1 + 1)
holds for a balancing tuning parameter δ . In 1992, Adams suggested a variant of the balancing
condition, namely
bal(s 1 , s 2 ) B s 1 + s 2 ≤ 1 ∨ (s 1 ≤ δ · s 2 ∧ s 2 ≤ δ · s 1 ).
The conditions are very similar, but not equivalent: the former allows, for example, δ = 3, s 1 = 2
and s 2 = 0, which the latter rejects.
Initially, the containers library used Adams’s balancing condition with the parameters δ = 4
(for sets) or δ = 5 (for maps). Campbell [2010] found that these parameters are actually invalid and
exhibited a sequence of insertions and deletions that produced an unbalanced tree. Subsequently,
the containers library switched to δ = 3 in both modules. Inspired by this bug report, Hirai and
Yamamoto [2011] thoroughly analyzed this data structure with the help of a Coq formalization, and
identified the valid range for the balancing parameter, albeit only for Nievergelt and Reingold’s
variant – our proof seems to be the first mechanical verification of Adam’s variant.
22
Breitner, Spector-Zabusky, Li, Rizkallah, Wiegley and Weirich
Given such thorough analysis of the algorithms, we did not expect to learn anything new
about this data structure, and for the most part, this was true. Our proofs are free of any manual
calculations about tree sizes and the balancing condition. We just state the proper preconditions
for each lemma, and Coq’s automation for linear integer arithmetic, lia [Besson 2006], takes care
of the rest.
One exception was the crucial function balanceL which is used, according to the documentation
“when the left subtree might have been inserted to or when the right subtree might have been
deleted from”. This suggests the precondition
(bal(s 1 − 1, s 2 ) ∧ 0 < s 1 ) ∨ bal(s 1 , s 2 ) ∨ bal(s 1 , s 2 + 1)
corresponding to the three cases: left tree inserted to, no change, and right tree deleted from. This
is also what Hirai and Yamamoto used in their formalization of the original variant. And indeed,
this precondition is strong enough to verify that the output of balanceL is balanced – but we
found the precondition too strong. The link operation, shown in Figure 8, is supposed to balance
two arbitrary trees using balanceL. In the verification of link we were unable to satisfy this
precondition.
We found another precondition that is both strong enough for the verification of balanceL and
weak enough to allow the verification of link, namely
(bal∗ (s 1 − 1, s 2 ) ∧ 0 < s 1 ) ∨ bal(s 1 , s 2 ) ∨ bal(s 1 , s 2 + 1)
where
bal∗ (s 1 , s 2 ) B δ · s 1 ≤ δ 2 · s 2 + δ · s 2 + s 2 ∧ s 2 ≤ s 1 .
Unfortunately we cannot give much insight into why this is the right inequality – this is the
price we pay for relying completely on proof automation when verifying the balancing properties.
For us, this inequality fell out of the proof for link, where we call balanceL with trees with sizes
s 1 and s 2 , and what we know about these trees can be expressed as
∃ s sl sr , s 1 = 1 + s + sl ∧ s 2 = sr ∧ bal(sl , sr ) ∧ δ · s < 1 + sl + sr ∧ 1 ≤ s ∧ 0 ≤ sl ∧ 0 ≤ sr
As lia works best when the formulas are quantifier-free, we manually eliminated the existential
quantifiers and simplified this to arrive at bal∗ (s 1 − 1, s 2 ).
6.2
Improvements to containers
Although our verification did not find bugs in the code of the library, we were able to improve
containers: during the verification of Data.Set.union, we noticed that it was using a nested
pattern match to check if an argument is a singleton set, when it could be testing if the size was 1
directly. This change turned out to provide a 4% speedup to union and will be present in the next
version of the containers library.26
Additionally, as we mention in Section 4.1, our well-formedness property for IntSet uncovered
a weakness in the valid function for IntSet and IntMap, which was used in the test suite. The
valid function failed to ensure that some of the invariants hold recursively in the tree structure,
which was necessary for the proof. We notified the containers maintainers,27 who then made the
valid function complete.
We also provide a package, containers-verified,28 which re-exports the types and definitions
we have verified from the precise version of containers we are working with. This way, a developer
26 https://github.com/haskell/containers/commit/b1a05c3a2
27 https://github.com/haskell/containers/issues/522
28 https://hackage.haskell.org/package/containers-verified
Ready, Set, Verify!
23
who wants to use only the verified portion of the implementation can replace their dependency on
containers with a dependency on containers-verified.
7
ASSUMPTIONS AND LIMITATIONS
We have shown a way to make mechanically checked, formal statements about existing Haskell
code, and have applied this technique to verify parts of the containers library. But are the theorems
that we prove actually true? And if they are, how useful is this method?
7.1
The formalization gap
As always, when a theorem is stated about a object that does not purely exist within mathematics,
its validity depends on a number of assumption.
• First and foremost, we have to assume that Coq behaves as documented and does not allow
us to prove false theorems. This is of particular relevance because we rely on fine details of
Coq’s machinery (e.g. the behavior of Qed, Section 5.4) and optionally add consistent axioms
(see Section 5.7.4).
• The biggest assumption is that the semantics of a Gallina program, as defined by the Calculus
of Construction [Coquand and Huet 1988], models the the behavior of a running Haskell
program in a meaningful way. At this time, we cannot even attempt to close this gap, as there
are neither formal semantics nor verified compilers for Haskell.
We know that “Fast and Loose Reasoning is Morally Correct” [Danielsson et al. 2006], which
says that theorems about the total fragment of a non-total language carry over to the full
language. But since we relate two very different languages, this argument alone is not enough
to bridge the gap.
We can gain additional confidence by testing this connection: we extracted our Gallina
versions of Set and IntSet back to Haskell and successfully ran the original test suite of
containers against it. Every successful test indicates that the corresponding theorem (see
Section 4.2) is indeed a theorem about the Haskell program.
• Furthermore, we rely on hs-to-coq translating Haskell code into the correct Gallina code.
The translator itself is a sizable piece of code, unverified (and such verification is currently out
of our reach, not least because there is no formal semantics of Haskell) and therefore surely
not free of bugs. We get some confidence into the tool from manually inspecting its output
and observing that it is indeed what we would consider the “right” translation from Haskell
into Gallina, and additional confidence from the fact that we were actually able to prove the
specifications, which would not be possible if the translated code behaved differently than
intended. Moreover, extracting the translated code back to Haskell and running the test suite
also stress-tests the translation.
• The translation was not completely automatic and required manual edits. With each edit,
we add another assumption to the formalization gap: Does our underspecification of pointer
equality encompass the actual behavior of GHC’s reallyUnsafePtrEqualty#? Given our
choice of using unbounded integers, does the size field in a Set really never overflow in
practice? Are our manually written versions of low-level bit twiddling functions correct? We
list and justify our manual interventions in Section 5.
The formalization gap of our work is relatively large compared to, say, the gap for the verification
of programs written in Gallina in the first place. But for the purpose of ensuring the correctness
of the Haskell code, that is less critical. Even if one of our assumptions is flawed, it is much more
likely that the flaw will get in the way of concluding the proofs, rather than allowing us to conclude
the proofs without noticing a bug. Incomplete proofs can uncover bugs, too.
24
7.2
Breitner, Spector-Zabusky, Li, Rizkallah, Wiegley and Weirich
Limitations of our approach
We proved multiple specifications about a large part of the code, but there are limits to what
theorems we can prove, and what we can prove them about.
Since we work with a shallow embedding of Haskell in Coq, we cannot make statements about
the performance of the Haskell code – which is a pity, given that the containers library contains
highly optimized code and provides clear specifications of the algorithmic complexity of their
operations. Similarly, we cannot verify that the operations are as strict or lazy as documented.
We also cannot translate and verify all code in the containers library, because some functions
use language features not yet supported by hs-to-coq, such as mutual recursion, unboxed values
and generic programming. When this affects a crucial utility function we can provide a manual
translation. This widens the formalization gap, but enables verification of code that depends on
it. When it affects less central code, e.g. the Data instances, we can simply skip the translation
(Section 5.3).
Partiality is a particular interesting limit. Coq only allows total functions, but practical Haskell
code uses partiality, often in a benign way: Calls to error in code paths that are unreachable as long
as invariants are maintained, or recursive functions that terminate on the actual arguments they
are called with, but may diverge on other inputs. Whereas Spector-Zabusky et al. [2018] considered
such code out of scope, we have found ways to deal with “internal partiality” (see Sections 5.4
and 5.7.4).
Taking a step back, it might seem that our approach may see limited adoption in the Haskell
community because it requires expertise in Coq. But though tied to a Haskell artifact, verification is
isolated from the codebase. Haskell programmers can focus on their domain, trying to get the best
performance out of the code and without having to know about verification, while proof engineers
can work solely within Coq and do not have to be fluent in Haskell.
In this paper we verify a specific version of the Haskell code, and do not discuss ongoing
maintenance of such a verification. It remains to be seen how resilient the proofs are against
changes in the Haskell code. Changes of syntactic nature, or changes to a function’s strictness,
might be swallowed by hs-to-coq. Other changes might affect the translated code, but still allow
the proofs to go through, if our proof tactics are flexible enough. In general, though, we expect that
changes in the code require changes in the proofs. Since Coq is an interactive theorem prover, it
will at least clearly point out which parts of our development need to be updated.
8
8.1
RELATED WORK
Verification of purely functional data structures
Purely functional data structures, such as those found in Okasaki’s book [1999] are frequent targets
of mechanical verification. That said, we believe that we are the first to verify the Patricia tree
algorithms that underlie Data.IntSet.
Verifying weight-balanced trees in Haskell. Similar to hs-to-coq, LiquidHaskell [Vazou et al.
2014] can be used to verify existing Haskell code. Users of this tool annotate their Haskell source
files with refinement types and other annotations. LiquidHaskell then uses an SMT solver to automatically discharge proof obligations described by the refinements. This means that LiquidHaskell
provides more automation than hs-to-coq; however, the language of Coq is higher-order and
more expressive than the language of SMT solvers. Vazou et al. [2017] compared the experiences of
using LiquidHaskell vs. plain Coq, and found that both have advantages.
Vazou et al. [2013] also described the use of LiquidHaskell to verify the Data.Map module of
finite maps from containers. Although not the same as Data.Set, this code shares the same
underlying data structure (weight-balanced trees) and the implementations of the two are similar.
Ready, Set, Verify!
25
Their verification has similarities with our work; they also use unbounded integers as the number
representation and left functions like showTree unspecified. However, we develop a richer specification, which includes a semantic description of each operation we verified, constraints about
the tree balance, and the ordering of the elements in the tree. In contrast, Vazou et al. limit their
specifications to ordering only. For example, in addition to showing that the insertion operation
preserves the order of elements of the tree, our work also shows that: (1) insertion preserves the
balancing conditions of the weight-balanced tree, (2) the size field at each node in the tree is
maintained correctly (i.e., the size equals to the number of all its descendants), and (3) the tree
returned by this operation contains the inserted element, and all elements in the original tree, but
nothing more. Although it might be possible to replicate our specifications by using theories of
finite sets and maps in SMT [Kröning et al. 2009] to encode these properties using refinement types
in LiquidHaskell, this approach has not been explored.
Furthermore, our specification also includes type class laws, and we are able to verify that Set
and IntSet have lawful instances of the Eq, Ord, Semigroup, and Monoid type classes. When Vazou
et al.’s original work was developed in 2013, LiquidHaskell did not have the capability to state and
prove these properties. Since then, there have been new developments in LiquidHaskell, particularly
refinement reflection [Vazou et al. 2018], which could make it possible to specify and prove type
class laws.
Both LiquidHaskell and hs-to-coq check for termination of Haskell functions. In LiquidHaskell,
the termination check is an option that can be deactivated, allowing the sound verification of
nonstrict, non-terminating functions [Vazou et al. 2014]. In contrast, a proof of termination is
a requirement for verifying functions using hs-to-coq [Spector-Zabusky et al. 2018]. However,
hs-to-coq is able to take advantage of many options available in Coq for proving termination of
non-trivial recursion, including structural recursion, Program Fixpoint and our own approach
based on deferredFix. This latter approach alowed us to reason about fromDistinctAscList
(see Section 5.7.4) and prove that it is indeed terminating; on the other hand, Vazou et al. deactivate
the termination check for this function.
Verifying weight-balanced trees in other languages. Hirai and Yamamoto [2011] implemented a
weight-balanced tree similar to Haskell’s Data.Set library (albeit using the balancing condition of
Nievergelt and Reingold [1972]) and mechanically verified its balancing properties in Coq. More
recently, Nipkow and Dirix [2018] extended this work to formalize similar weight-balanced trees
in Isabelle and further verified the functional correctness of insertions and deletions.
We verify more functions in the Data.Set library than prior work. Operations that are unique
to our development include foldl, isSubsetOf, and fromDistinctAscList. The code verified
both by Hirai and Yamamoto and by Nipkow and Dirix is also different from the latest containers
library; for example, it does not use of pointer equality (Section 5.1). Another difference is that Hirai
and Yamamoto defined the union, intersection, and difference functions based on the “hedge
union” algorithm, but containers has since changed to use the “divide and conquer” algorithm.
Hirai and Yamamoto specify only the balancing constraints, whereas we develop a richer specification that also includes a semantic description of each operation we verified and the ordering of
the elements in the tree. We also gained new insights about the balancing conditions of the weightbalanced tree through our verification effort (see Section 6.1). Nipkow and Dirix’s specification
contains the same properties as ours, but they only specify the behavior of insert and delete; we
have verified a significantly larger set of functions (see Figure 4).
Other verifications of balanced trees. There are many existing works on mechanically verifying
purely functional balanced trees. We briefly mention a few here. Filliâtre and Letouzey [2004]
implemented AVL trees in Coq, and verified their functional correctness as well as their balancing
26
Breitner, Spector-Zabusky, Li, Rizkallah, Wiegley and Weirich
conditions. Appel [2011] did the same thing for red-black trees. These implementations have now
become parts of Coq standard library. Charguéraud [2010] translated OCaml implementations of
Okasaki [1999]’s functional data structures to characteristic formulae expressed as Coq axioms.
Licata [2012] lectured on verifying red-black trees in Agda at the Oregon Programming Languages
Summer School. McBride [2014] showed how to represent the ordering relationships in Agda for
general data structures, not just binary search trees. Nipkow [2016] showed how to automatically
verify the ordering properties of eight different binary search tree structures, by specifying each
in terms of the sorted list of their elements, a method he used again in his verification of weightbalanced trees [Nipkow and Dirix 2018]. Ralston [2009] verified AVL trees in ACL2.
8.2
Verification tools for Haskell
Previous work using hs-to-coq has only applied it to small examples. Spector-Zabusky et al. [2018]
describe three case studies, two of which require less than 20 lines of Haskell. The longest example
(the Bag module taken directly from the GHC compiler) is 247 lines of code. Furthermore, none of
the reasoning required for these examples is particularly deep. Our work provides experience with
more complex, externally-sourced, industrial-strength examples.
The coq-haskell library [Wiegley 2017] is a handwritten Coq library designed to make it easier
for Haskell programmers to work in Coq. In addition to enabling easier Coq programming, it also
provides support for extracting Coq programs to Haskell.
The prototype contract checker halo [Vytiniotis et al. 2013] takes a Haskell program, uses GHC
to desugar it into the intermediate language Core, and translates the Core program into a first-order
logic formula. It then invokes an SMT solver to prove this formula; a successful proof implies that
the original program is crash-free.
Haskabelle was hs-to-coq’s counterpart in Isabelle. It translated total Haskell code into equivalent Isabelle function definitions. Similar to hs-to-coq, it parsed Haskell, desugared syntactic
constructs, and configurably adapted basic types and functions to their counterparts in Isabelle’s
standard library. It used to be bundled with the Isabelle release, but it has not been updated recently
and was dropped from the distribution.
Haskell has been used as a prototyping language for mechanically verified systems in the past. The
seL4 verified microkernel started with a Haskell prototype that was semi-automatically translated
to Isabelle/HOL [Derrin et al. 2006; Klein et al. 2009]. The authors found that the availability of the
Haskell prototype provided a machine-checkable formal executable specification of the system.
They used this prototype to refine their designs via testing, allowing them to make corrections
before full verification.
The Programmatica project [Hallgren et al. 2004] included a tool that translates Haskell code
into the Alfa proof editor. Their tool only produces valid proofs for total functions over finite data
structures. The logic of the Alfa proof assistant is based on dependent type theory, but without as
many features as Coq. In particular, the Programmatica tool compiles away type classes and nested
pattern matching; both of these features are retained by hs-to-coq.
Dybjer et al. developed a tool for automatically translating Haskell programs to the Agda/Alfa
proof assistant [2004]. They explicitly note the interplay between testing and theorem proving and
show how to verify a tautology checker. Abel et al. [2005] translate Haskell expressions into the
logic of the Agda 2 proof assistant. Their tool works later in the GHC pipeline than hs-to-coq and
translates Core expressions.
8.3
Translating other higher-order functional languages
There are many large and successful verification projects that demonstrate that functional languages
are well suited for verification. In contrast to our work, they require re-implementing the code either
Ready, Set, Verify!
27
in a new functional language, as is the case for Cogent [Amani et al. 2016; O’Connor et al. 2016]
and F* [Swamy et al. 2016], or in a proof assistant, such as HOL4 in the case of CakeML [Kumar
et al. 2014; Myreen and Owens 2014]. The CakeML and Cogent projects have a different focus than
ours, and they provide a higher assurance in their verified code. CakeML [Kumar et al. 2014] has a
verified compiler and Cogent has a certifying compiler [O’Connor et al. 2016; Rizkallah et al. 2016].
Both tools provide mechanically checked proofs that their shallow embeddings correspond to the
functional code being verified.
Cogent is a restricted higher-order functional language [O’Connor et al. 2016] that was used to
verify filesystems [Amani et al. 2016]. Its compiler produces C code, a high-level specification in
Isabelle/HOL, and an Isabelle/HOL refinement proof linking the two [O’Connor et al. 2016; Rizkallah
et al. 2016]. Chen et al. [2017] integrated property-based testing into the Cogent framework. The
authors claim that property-based testing enables an incremental approach to a fully verified system,
as it allows for the replacement of tests of properties stated in the specification by formal proofs.
Our work substantiates this claim, as indeed one of the ways in which we obtain specifications is
through the QuickCheck properties provided by Haskell as discussed in Section 4.2.
CakeML is a large subset of ML with a verified compiler and runtime system [Kumar et al. 2014].
Users of CakeML can write pure functional programs in the HOL4 proof assistant and verify them
in HOL4. They can then extract an equivalent correct by construction CakeML program. This
method has been used to verify several data structures including red black trees, crypto protocols,
and a CakeML version of the HOL light theorem prover [Myreen and Owens 2014].
F* is a general-purpose functional language that allows for a mixture of proving and generalpurpose programming [Swamy et al. 2016]. Programs can be specified using dependent and refinement types and automatically verified using one of F*’s backend SMT solvers. Its subset
Low* [Protzenko et al. 2017] has been used to verify high-assurance optimized cryptographic
libraries.
8.4
Extraction
The semantic proximity of Haskell and Coq, which we rely on, is also used in the other direction by
Coq’s support for code extraction to Haskell [Letouzey 2002]. Several projects use this feature to
verify Haskell code [Chen et al. 2015; Joseph 2014]. However, since extraction starts with Coq code
and generates Haskell, it cannot be used to verify pre-existing Haskell. Although in a certain sense
hs-to-coq and extraction are inverses, round-tripping does not produce syntactically equivalent
output in either direction. In one direction, hs-to-coq extensively annotates the resulting Coq
code; in the other, extraction ignores many Haskell features and inserts unsafe type coercions. In
this work, we use testing to verify that round-tripping produces operationally equivalent output;
this provides assurance about the correctness of both hs-to-coq and extraction.
CertiCoq [Anand et al. 2017] and Œuf [Mullen et al. 2018] are verified compilers from Gallina to
assembly. CertiCoq can compile any Gallina program, while Œuf can only compile a limited subset
of Gallina. However, Œuf provides stronger guarantees about the Gallina code that is in the limited
subset it translates.
9
CONCLUSIONS AND FUTURE WORK
We verified the two finite set modules that are part of the widely used and highly-optimized
containers library. Our efforts provide the deepest specification and verification of this code to
date, covering more of the API and proving stronger, more descriptive properties than prior work.
In future work, there is yet more to verify in containers. For example, we plan to add a version of
IntSet that uses 64-bit ints as the element type in addition to our current version with unbounded
natural numbers. That way users could choose the treatment of overflow that makes the most sense
28
Breitner, Spector-Zabusky, Li, Rizkallah, Wiegley and Weirich
for their application. We have also started to verify Data.Map and Data.IntMap. Because these
modules use the same algorithms as Data.Set and Data.IntSet, we already adapted some of our
existing proofs to this setting.
The fact that we did not find bugs says a lot about the tools that are already available to Haskell
programmers for producing correct code, such as a strong, expressive type system and a mature
property-based testing infrastructure. However, few would dare to extrapolate from these results to
say that all Haskell programs are bug free! Instead, we view verification as a valuable opportunity
for functional programmers and an activity that we hope will become more commonplace.
ACKNOWLEDGMENTS
This material is based upon work supported by the National Science Foundation under Grant
No. 1319880 and Grant No. 1521539.
REFERENCES
Andreas Abel, Marcin Benke, Ana Bove, John Hughes, and Ulf Norell. 2005. Verifying Haskell Programs Using Constructive
Type Theory. In Haskell Workshop. ACM, 62–73. DOI:http://dx.doi.org/10.1145/1088348.1088355
Stephen Adams. 1992. Implementing sets efficiently in a functional language,. Research Report CSTR 92-10. University of
Southampton.
Sidney Amani, Alex Hixon, Zilin Chen, Christine Rizkallah, Peter Chubb, Liam O’Connor, Joel Beeren, Yutaka Nagashima,
Japheth Lim, Thomas Sewell, Joseph Tuong, Gabriele Keller, Toby Murray, Gerwin Klein, and Gernot Heiser. 2016.
Cogent: Verifying High-Assurance File System Implementations. In International Conference on Architectural Support
for Programming Languages and Operating Systems. Atlanta, GA, USA, 175–188. DOI:http://dx.doi.org/10.1145/2872362.
2872404
Abhishek Anand, Andrew Appel, Greg Morrisett, Zoe Paraskevopoulou, Randy Pollack, Olivier Savary Belanger, Matthieu
Sozeau, and Matthew Weaver. 2017. CertiCoq: A verified compiler for Coq. In CoqPL Workshop (CoqPL ’17).
Andrew W. Appel. 2011. Efficient verified red-black trees. (2011). https://www.cs.princeton.edu/~appel/papers/redblack.pdf
Andrew W. Appel, Lennart Beringer, Adam Chlipala, Benjamin C. Pierce, Zhong Shao, Stephanie Weirich, and Steve
Zdancewic. 2017. Position paper: the science of deep specification. Philosophical Transactions of the Royal Society of
London A: Mathematical, Physical and Engineering Sciences 375, 2104 (2017). DOI:http://dx.doi.org/10.1098/rsta.2016.0331
Frédéric Besson. 2006. Fast Reflexive Arithmetic Tactics the Linear Case and Beyond. In TYPES (Lecture Notes in Computer
Science), Vol. 4502. Springer, 48–62.
Ana Bove, Peter Dybjer, and Ulf Norell. 2009. A Brief Overview of Agda - A Functional Language with Dependent Types. In
Theorem Proving in Higher Order Logics, 22nd International Conference, TPHOLs 2009, Munich, Germany, August 17-20,
2009. Proceedings (Lecture Notes in Computer Science), Stefan Berghofer, Tobias Nipkow, Christian Urban, and Makarius
Wenzel (Eds.), Vol. 5674. Springer, 73–78. DOI:http://dx.doi.org/10.1007/978-3-642-03359-9_6
Edwin Brady. 2017. Type-driven Development With Idris. Manning.
Taylor Campbell. 2010. Bug in Data.Map. (2010). http://article.gmane.org/gmane.comp.lang.haskell.libraries/13448 e-mail
to the Haskell libraries mailing list.
Arthur Charguéraud. 2010. Program verification through characteristic formulae. In ICFP. ACM, 321–332. DOI:http:
//dx.doi.org/10.1145/1932681.1863590
Haogang Chen, Daniel Ziegler, Tej Chajed, Adam Chlipala, M. Frans Kaashoek, and Nickolai Zeldovich. 2015. Using Crash
Hoare Logic for Certifying the FSCQ File System. In SOSP. ACM, 18–37. DOI:http://dx.doi.org/10.1145/2815400.2815402
Zilin Chen, Liam O’Connor, Gabriele Keller, Gerwin Klein, and Gernot Heiser. 2017. The Cogent Case for Property-Based
Testing. In Workshop on Programming Languages and Operating Systems (PLOS) (2017-10-28). ACM, Shanghai, China,
1–7.
Koen Claessen and John Hughes. 2000. QuickCheck: a lightweight tool for random testing of Haskell programs. In ICFP.
ACM, 268–279.
Thierry Coquand and Gérard P. Huet. 1988. The Calculus of Constructions. Information and Computation 76, 2/3 (1988),
95–120. DOI:http://dx.doi.org/10.1016/0890-5401(88)90005-3
Nils Anders Danielsson, John Hughes, Patrik Jansson, and Jeremy Gibbons. 2006. Fast and loose reasoning is morally correct.
In POPL. ACM, 206–217. DOI:http://dx.doi.org/10.1145/1111037.1111056
Philip Derrin, Kevin Elphinstone, Gerwin Klein, David Cock, and Manuel M. T. Chakravarty. 2006. Running the Manual: An
Approach to High-assurance Microkernel Development. In Haskell Symposium. ACM, 60–71. DOI:http://dx.doi.org/10.
1145/1159842.1159850
Ready, Set, Verify!
29
Peter Dybjer, Qiao Haiyan, and Makoto Takeyama. 2004. Verifying Haskell programs by combining testing, model
checking and interactive theorem proving. Information & Software Technology 46, 15 (2004), 1011–1025. DOI:http:
//dx.doi.org/10.1016/j.infsof.2004.07.002
Jean-Christophe Filliâtre and Pierre Letouzey. 2004. Functors for Proofs and Programs. In Programming Languages and
Systems, David Schmidt (Ed.). Springer Berlin Heidelberg, Berlin, Heidelberg, 370–384.
Thomas Hallgren, James Hook, Mark P. Jones, and Richard B. Kieburtz. 2004. An overview of the programatica toolset. In
HCSS.
Yoichi Hirai and Kazuhiko Yamamoto. 2011. Balancing weight-balanced trees. Journal of Functional Programming 21, 3
(2011), 287–307. DOI:http://dx.doi.org/10.1017/S0956796811000104
Adam Megacz Joseph. 2014. Generalized Arrows. (May 2014). http://www2.eecs.berkeley.edu/Pubs/TechRpts/2014/
EECS-2014-130.html also see http://www.megacz.com/berkeley/coq-in-ghc/.
Gerwin Klein, Kevin Elphinstone, Gernot Heiser, June Andronick, David Cock, Philip Derrin, Dhammika Elkaduwe, Kai
Engelhardt, Rafal Kolanski, Michael Norrish, Thomas Sewell, Harvey Tuch, and Simon Winwood. 2009. seL4: Formal
Verification of an OS Kernel. In ACM Symposium on Operating Systems Principles. ACM, Big Sky, MT, USA, 207–220.
Alexander Krauss. 2006. Partial Recursive Functions in Higher-Order Logic. In IJCAR (LNCS), Vol. 4130. Springer, 589–603.
Daniel Kröning, Philipp Rümmer, and Georg Weissenbacher. 2009. A proposal for a theory of finite sets, lists, and maps for
the SMT-LIB standard. In Informal proceedings, 7th International Workshop on Satisfiability Modulo Theories at CADE,
Vol. 22.
Ramana Kumar, Magnus O. Myreen, Michael Norrish, and Scott Owens. 2014. CakeML: A Verified Implementation of ML.
In Proceedings of the 41st ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages (POPL ’14). ACM,
New York, NY, USA, 179–191.
Pierre Letouzey. 2002. A New Extraction for Coq. In TYPES (LNCS), Vol. 2646. Springer, 200–219. DOI:http://dx.doi.org/10.
1007/3-540-39185-1_12
Dan Licata. 2012. 15-150 Lecture 21: Red-Black Trees. (2012). https://www.cs.cmu.edu/~15150/previous-semesters/
2012-spring/resources/lectures/21.pdf Lecture at the Oregon Programming Language Summer School.
The Coq development team. 2016. The Coq proof assistant reference manual. LogiCal Project. http://coq.inria.fr Version
8.6.1.
Conor Thomas McBride. 2014. How to Keep Your Neighbours in Order. In Proceedings of the 19th ACM SIGPLAN International
Conference on Functional Programming (ICFP ’14). ACM, New York, NY, USA, 297–309. DOI:http://dx.doi.org/10.1145/
2628136.2628163
Donald R. Morrison. 1968. PATRICIA&Mdash;Practical Algorithm To Retrieve Information Coded in Alphanumeric. J. ACM
15, 4 (Oct. 1968), 514–534. DOI:http://dx.doi.org/10.1145/321479.321481
Eric Mullen, Stuart Pernsteiner, James R. Wilcox, Zachary Tatlock, and Dan Grossman. 2018. ŒUf: Minimizing the Coq
Extraction TCB. In Proceedings of the 7th ACM SIGPLAN International Conference on Certified Programs and Proofs (CPP
2018). ACM, New York, NY, USA, 172–185. DOI:http://dx.doi.org/10.1145/3167089
Magnus O. Myreen and Scott Owens. 2014. Proof-producing translation of higher-order logic into pure and stateful ML.
Journal of Functional Programming 24 (May 2014), 284–315. DOI:http://dx.doi.org/10.1017/S0956796813000282
Jürg Nievergelt and Edward M. Reingold. 1972. Binary Search Trees of Bounded Balance. In STOC. ACM, 137–142. DOI:
http://dx.doi.org/10.1145/800152.804906
Tobias Nipkow. 2016. Automatic Functional Correctness Proofs for Functional Search Trees. In Interactive Theorem Proving
(ITP 2016), J. Blanchette and S. Merz (Eds.), Vol. 9807. 307–322.
Tobias Nipkow and Stefan Dirix. 2018. Weight-Balanced Trees. Archive of Formal Proofs (March 2018). http://isa-afp.org/
entries/Weight_Balanced_Trees.html, Formal proof development.
Tobias Nipkow, Lawrence C. Paulson, and Markus Wenzel. 2002. Isabelle/HOL - A Proof Assistant for Higher-Order Logic.
Lecture Notes in Computer Science, Vol. 2283. Springer. DOI:http://dx.doi.org/10.1007/3-540-45949-9
Liam O’Connor, Zilin Chen, Christine Rizkallah, Sidney Amani, Japheth Lim, Toby Murray, Yutaka Nagashima, Thomas
Sewell, and Gerwin Klein. 2016. Refinement Through Restraint: Bringing Down the Cost of Verification. In International
Conference on Functional Programming. Nara, Japan.
Chris Okasaki. 1999. Purely functional data structures. Cambridge University Press. DOI:http://dx.doi.org/10.1017/
CBO9780511530104
Chris Okasaki and Andrew Gill. 1998. Fast Mergeable Integer Maps. In In Workshop on ML. 77–86.
Simon Peyton Jones, Andrew Tolmach, and Tony Hoare. 2001. Playing by the rules: rewriting as a practical optimisation
technique in GHC. In Haskell Workshop.
Jonathan Protzenko, Jean-Karim Zinzindohoué, Aseem Rastogi, Tahina Ramananandro, Peng Wang, Santiago ZanellaBéguelin, Antoine Delignat-Lavaud, Cătălin Hriţcu, Karthikeyan Bhargavan, Cédric Fournet, and Nikhil Swamy. 2017.
Verified Low-level Programming Embedded in F*. Proc. ACM Program. Lang. 1, ICFP, Article 17 (Aug. 2017), 29 pages.
DOI:http://dx.doi.org/10.1145/3110261
30
Breitner, Spector-Zabusky, Li, Rizkallah, Wiegley and Weirich
Ryan Ralston. 2009. ACL2-certified AVL Trees. In Proceedings of the Eighth International Workshop on the ACL2 Theorem
Prover and Its Applications (ACL2 ’09). ACM, New York, NY, USA, 71–74. DOI:http://dx.doi.org/10.1145/1637837.1637848
Christine Rizkallah, Japheth Lim, Yutaka Nagashima, Thomas Sewell, Zilin Chen, Liam O’Connor, Toby Murray, Gabriele
Keller, and Gerwin Klein. 2016. A Framework for the Automatic Formal Verification of Refinement from Cogent to C. In
International Conference on Interactive Theorem Proving. Nancy, France.
Antal Spector-Zabusky, Joachim Breitner, Christine Rizkallah, and Stephanie Weirich. 2018. Total Haskell is reasonable Coq.
In CPP. ACM, 14–27. DOI:http://dx.doi.org/10.1145/3167092
Milan Straka. 2010. The Performance of the Haskell Containers Package. In Proceedings of the Third ACM Haskell Symposium
on Haskell (Haskell ’10). ACM, New York, NY, USA, 13–24. DOI:http://dx.doi.org/10.1145/1863523.1863526
Nikhil Swamy, Cătălin Hriţcu, Chantal Keller, Aseem Rastogi, Antoine Delignat-Lavaud, Simon Forest, Karthikeyan
Bhargavan, Cédric Fournet, Pierre-Yves Strub, Markulf Kohlweiss, Jean-Karim Zinzindohoue, and Santiago ZanellaBéguelin. 2016. Dependent Types and Multi-monadic Effects in F*. In Proceedings of the 43rd Annual ACM SIGPLANSIGACT Symposium on Principles of Programming Languages (POPL ’16). ACM, New York, NY, USA, 256–270. DOI:
http://dx.doi.org/10.1145/2837614.2837655
Niki Vazou, Leonidas Lampropoulos, and Jeff Polakow. 2017. A Tale of Two Provers: Verifying Monoidal String Matching in
Liquid Haskell and Coq. In Haskell Symposium. ACM, 63–74. DOI:http://dx.doi.org/10.1145/3122955.3122963
Niki Vazou, Patrick M. Rondon, and Ranjit Jhala. 2013. Abstract Refinement Types. In Proceedings of the 22Nd European
Conference on Programming Languages and Systems (ESOP’13). Springer-Verlag, Berlin, Heidelberg, 209–228. DOI:
http://dx.doi.org/10.1007/978-3-642-37036-6_13
Niki Vazou, Eric L. Seidel, Ranjit Jhala, Dimitrios Vytiniotis, and Simon Peyton-Jones. 2014. Refinement Types for Haskell.
In ICFP. ACM, 269–282. DOI:http://dx.doi.org/10.1145/2628136.2628161
Niki Vazou, Anish Tondwalkar, Vikraman Choudhury, Ryan G. Scott, Ryan R. Newton, Philip Wadler, and Ranjit Jhala. 2018.
Refinement reflection: complete verification with SMT. PACMPL 2, POPL (2018), 53:1–53:31. DOI:http://dx.doi.org/10.
1145/3158141
Dimitrios Vytiniotis, Simon Peyton Jones, Koen Claessen, and Dan Rosén. 2013. HALO: Haskell to Logic Through Denotational Semantics. In POPL. ACM, 431–442. DOI:http://dx.doi.org/10.1145/2429069.2429121
John Wiegley. 2017. coq-haskell: A library for formalizing Haskell types and functions in Coq. https://github.com/
jwiegley/coq-haskell. (2017).
| 6cs.PL
|
D EEP CAS: A Deep Reinforcement Learning Algorithm for
Control-Aware Scheduling
arXiv:1803.02998v1 [cs.SY] 8 Mar 2018
Burak Demirel, Arunselvan Ramaswamy, Daniel E. Quevedo and Holger Karl
Abstract— We consider networked control systems consisting
of multiple independent closed-loop control subsystems, operating over a shared communication network. Such systems are
ubiquitous in cyber-physical systems, Internet of Things, and
large-scale industrial systems. In many large-scale settings, the
size of the communication network is smaller than the size of
the system. In consequence, scheduling issues arise. The main
contribution of this paper is to develop a deep reinforcement
learning-based control-aware scheduling (D EEP CAS) algorithm
to tackle these issues. We use the following (optimal) design
strategy: First, we synthesize an optimal controller for each
subsystem; next, we design learning algorithm that adapts to
the chosen subsystem (plant) and controller. As a consequence
of this adaptation, our algorithm finds a schedule that minimizes the control loss. We present empirical results to show
that D EEP CAS finds schedules with better performance than
periodic ones. Finally, we illustrate that our algorithm can be
used for scheduling and resource allocation in more general
networked control settings than the above-mentioned one.
Index Terms— Deep Learning; Reinforcement Learning; Optimal Control; Networked Control Systems
I. I NTRODUCTION
Today’s cyber-physical systems (CPS), Internet of Things
(IoT), large-scale industrial systems, and myriad real-world
systems try to integrate traditional control systems with artificial intelligence to improve the overall system performance
and reliability. Examples of such systems include smart grids,
smart cities, city-wide vehicular traffic control, industrial
process optimization, among others. In recent years, artificial
intelligence (AI) has seen a resurgence as an effective
model-free solution to many problems, including optimal
control problems arising in the aforementioned systems. This
resurgence is partly owing to the advances in computational
capacities and the advent of deep neural networks for function approximation and feature extraction. Oftentimes, the
use of reinforcement learning algorithms or AI in conjunction
with traditional controllers reduces the complexity of system
design while boosting efficiency.
The large size of systems poses tremendous challenges
in resource allocation. In applications involving feedback,
this challenge is exaggerated since resource allocation is
required to be “control-aware”, i.e., it should reduce control
loss. Typical resources including communication channels
and computational resources do not scale with system size.
B. Demirel, A. Ramaswamy, D. E. Quevedo, and H. Karl
are with Paderborn University, Warburger Straße 100, 33098,
Paderborn,
Germany
[email protected]
[email protected] [email protected].
[email protected].
A. Ramaswamy was funded by the German Research Foundation (DFG)
- 315248657.
DeepCAS
Smart Sensor
Kalman
Filter (I)
Plant
Control
Algorithm
Kalman
Filter (II)
Feedback Control Loop (i)
Controller
Control
Algorithm
Net work
with M available
channels
···
Kalman
Filter (II)
Fig. 1.
Networked control system (NCS) that consists of N control
subsystems closed over a shared communication network (with M available
channel).
In large-scale systems, controllers often rely on information
collected from various sensors to make intelligent decisions.
Hence, efficient information dispersion is essential for decision making over communication networks to be effective.
As noted earlier, this is a hard problem since the number of
communication channels available is much smaller than what
is ideally required to transfer data from sensors to controllers.
In this paper, we present a deep reinforcement learningbased scheduling algorithm for the sensor-to-controller communication in large-scale networked control systems. To
illustrate our ideas, we use the system architecture illustrated
in Fig. 1. Later, we will show that our ideas can be used for
scheduling in more general architectures.
Fig. 1 is a simplified representation of (some) CPS and
IoT systems. The system consists of N independent control
subsystems that communicate over a shared communication
network, which contains M channels. We assume that M
N (M is much smaller than N ) and, in particular, that
each channel can be used to serve any communication
request. Each control subsystem consists of a smart sensor, a
controller, and a plant to be controlled. The feedback loops
are closed over this resource-constrained communication
network. For more details on the system model, we refer
the reader to § II.
D EEP CAS, our deep reinforcement learning-based modelfree scheduling algorithm, decides which of the N subsystems are allocated to the M channels in a given time
instant. Further, it adapts to the controller at hand and finds
a schedule that minimizes the control loss. We present an
implementation wherein D EEP CAS obtains feedbacks (i.e.,
rewards) from smart sensors for making decisions. At each
time instant, each smart sensor computes the estimate of the
subsystem’s state, using Kalman Filter (I), and transmits it to
the corresponding controller; see Fig. 1. The controller runs
Kalman Filter (II) to estimate the subsystem’s state in the
absence of transmissions. In addition to Kalman Filter (I), a
smart sensor also implements a copy of Kalman Filter (II)
and the control algorithm. This way the smart sensor knows
the state estimate used by the controller at every time instant.
Several scheduling approaches for control systems have
been proposed in the literature to determine the access order
of different sensors and/or actuators; see [1] and references
therein. A widely considered approach is to employ periodic
schedules [2]–[5]. However, finding such schedules for control applications may not be easy since both specific period
and sequence need to be found. Further, periodic sequences
may not be even optimal in which case searching for them
is futile. With a few exceptions (e.g., [6] and [7]), the
determination of optimal schedules indeed requires solving a
mixed-integer quadratic program, which is computationally
infeasible for all but very small systems.
Our contribution is in the development of a deep reinforcement learning-based control-aware scheduling (D EEP CAS)
algorithm. At its heart lies the Deep Q-Network (DQN), a
modern variant of Q learning, introduced in [8]. In addition to
being readily scalable, D EEP CAS is completely model-free.
To optimize the overall control performance, we propose
the following (optimal) design of control and scheduling:
In the first step, we design an optimal controller for each
independent subsystem. As discussed in [9], under limited
communication, the control loss has two components: (a)
best possible control loss (b) error due to intermittent transmissions. If M = N , then (b) vanishes. Since we are in
setting of M N , the goal of the scheduler is to minimize
(b). To this end, we first construct a Markov decision process
(MDP). The state space of this MDP is the difference in state
estimates of all controllers and sensors (obtainable from the
smart sensors). Its reward is the negative of the previously
mentioned (b). Since we are using DQN (reinforcement
learning) to solve this MDP, we do not need the knowledge
of transition probabilities. Since the goal of DQN is to find a
scheduling strategy that maximizes reward, it naturally aims
at minimizing (b). For more details on the MDP, the reader is
referred to § III. We present empirical results showing that
D EEP CAS finds schedules that have lower control losses
than traditional hand-tuned heuristics.
The organization of this paper is as follows: Section II
introduces models and assumptions for subsystems, sensors,
controllers, and network. This section also presents the
primary objective of the paper and how to reach this objective
via solving control design and control-aware scheduling
problems separately. Section III presents an MDP associated
with control-aware scheduling problems and proposes a
deep reinforcement learning algorithm to solve this MDP
efficiently. In Section IV, numerical examples are used to
illustrate the power of our AI-based scheduling algorithm.
Section V gives concluding remarks and directions for future work.
II. N ETWORKED C ONTROL S YSTEM : M ODEL ,
A SSUMPTIONS , AND G OALS
A. Model for each subsystem
As illustrated in Figure 1, our networked control system
consists of N independent closed-loop subsystems. The
feedback loop within each subsystem (plant) is closed over a
shared communication network. For 1 ≤ i ≤ N , subsystem
i is described by
(i)
(i)
(i)
(i)
(1)
xt+1 = A(i) xt + B (i) ut + wt ,
where A(i) and B (i) are matrices of appropriate dimensions,
(i)
(i)
xt ∈ Rni is the subsystem i’s state, ut ∈ Rmi is the
(i)
ni
control input, and wt ∈ R is zero-mean i.i.d. Gaussian
noise with covariance matrix W (i) . The initial state of
(i)
subsystem i, x0 , is assumed to be a Gaussian random vector
(i)
(i)
with mean x̄0 and covariance matrix X0 .
At a given time t, we assume that only noisy output
measurements are available. We thus have:
(i)
(i)
(i)
(2)
yt = C (i) xt + vt ,
(i)
where vt ∈ Rpi is zero-mean i.i.d. Gaussian noise with
(i)
(i)
covariance matrix V (i) . All noise sources, wt and vt , are
(i)
independent of the initial conditions x0 .
B. Control architecture and loss function
The dynamics of each subsystem is a stochastic linear
time-invariant (LTI) system given by (1). Further, each subsystem is independently controlled although dependencies do
arise from sharing a communication network. Subsystem i
(i)
has a smart sensor which samples the subsystem’s output yt
and computes an estimate of the subsystem’s state. This value
is then sent to the associated controller, provided a channel
is allocated to it by D EEP CAS. If the controller obtains
a new state estimate from the sensor, then it calculates a
control command based on this state estimate. Otherwise, it
calculates a control command based on its own estimate of
the subsystem’s state.
The control actions and scheduling decisions (of D EEP CAS) are taken to minimize the total control loss given by
J=
N
X
(3)
J (i) ,
i=1
where J (i) is the control loss of subsystem i and is given by
(i)> (i) (i)
J (i) = E xT Qf xT
+
T−1
X
(i)>
(i)
xt Q(i) xt
+
(i)>
(i)
ut R(i) ut
, (4)
t=0
(i)
where Q(i) and Qf are positive semi-definite matrices and
R(i) is positive definite.
C. Smart sensors and pre-processing units
Within our setting, the primary role of a smart sensor
is to take measurements of a subsystem’s output. Also, it
plays a vital role in helping D EEP CAS with scheduling
decisions. It is from the smart sensors that D EEP CAS gets
all the necessary feedback information for scheduling. For
these tasks, each smart sensor employs two Kalman filters:
(1) Kalman Filter (I) is used to estimate the subsystem’s
state, (2) a copy of Kalman Filter (II) is used to estimate
the subsystem’s state as perceived by the controller. Note
that the controller employs Kalman Filter (II). Below, we
discuss them in detail.
Kalman filter (I): The sensor employs a standard Kalman
(i)s
(i)s
filter to compute the state estimate x̂t|t and covariance Pt|t
recursively as
(i)s
x̂t|t−1
(i)s
Pt|t−1
=
=
(i)s
(i)
A x̂t−1|t−1 + B (i) ut−1
(i)s
A(i) Pt−1|t−1 A(i)> + W (i)
(i)
−1
(i)s
(i)s
Kt = Pt|t−1 C (i)> C (i) Pt|t−1 C (i)> + V (i)
(i)s
(i)s
(i)
(i)
(i)s
x̂t|t = x̂t|t−1 + Kt yt − C x̂t|t−1
(i)s
(i)
(i)s
Pt|t = I − Kt C (i) Pt|t−1 ,
(i)s
(i)
(i)s
(i)c
(i)
(i)
(i)
with x̂0|−1 = x̄0 .
D. Goal: minimizing the control loss
Under the assumptions presented in § II, the certainty
equivalent controller is still optimal; see [9] for details. The
total control loss in (4) has two components: (a) best possible
control loss (b) error due to intermittent communications.
Hence, the problem of minimizing control loss has two separate components: (i) designing the best (optimal) controller
for each subsystem and (ii) scheduling in a control-aware
manner.
Component I: Controller design. The controller in feedback
(i)
loop i takes the following control action, ut , at time t:
(i)
(i)
(i)c
(7)
ut = −Lt x̂t|t ,
(i)c
where x̂t|t is the state estimate used by the controller,
(i)
(i)
(i)
Lt = (B (i)> St+1 B (i) + R(i) )−1 B (i)> St+1 A(i)
and
(i)
St
(i)
(8)
is recursively computed as
(i)
(i)
St = A(i)> St+1 A(i) + Q(i) − A(i)> St+1 B (i)
(i)
(i)
(i)s
(i)c
J
(i)
=
(i)> (i) (i)
x̄0 S0 x̄0 +Tr
(i) (i)
S0 X0 +
T−1
X
(i)
Tr St+1 W (i)
t=0
+
T−1
X
(i)s (i)
Tr Pt|t Γt
t=0
(i)
where Γt ,
(i)s
(i)c
x̂t|t − x̂t|t .
+
T−1
X
h
i
(i)> (i) (i)
E et|t Γt et|t , (10)
t=0
(i)>
(i)
(i)
(i)
Lt (B (i)> St+1 B (i) + R(i) )Lt and et|t
(i)
Note that et|t is the communication error
,
in
subsystem i. Recall that there are N subsystems and M
(i)
communication channels. If M = N , then et|t = 0 for all
t ≥ 0 and 1 ≤ i ≤ N .
Component II: Control-aware scheduling. The main aim of
the scheduling algorithm is to help minimize J of (3). To
this end, one must minimize
i
h
(i)> (i) (i)
E et|t Γt et|t
(11)
t=0
x̂t|t−1 = A(i) x̂t−1|t−1 + B (i) ut−1 ,
(5)
( (i)s
x̂t|t
if the MMSE estimate received ,
(i)c
(6)
x̂t|t =
(i)c
x̂t|t−1 otherwise ,
(i)c
(i)
loop i have communicated. Otherwise, x̂t|t is the state
estimate obtained from Kalman Filter (II). The minimum
value of the control loss of subsystem i is given by
T−1
X
starting from x̂0|−1 = x̄0 and P0|−1 = X0 .
Kalman filter (II): The controller runs a minimum mean
square error (MMSE) estimator to compute estimates of the
subsystem’s state as follows:
(i)c
(i)
with initial values SN = Qf . Let x̂t|t be the state estimate
of Kalman Filter (I), as employed by the sensor. We have
(i)c
(i)s
x̂t|t = x̂t|t when the sensor and controller of the feedback
× (B (i)> St+1 B (i) + R(i) )−1 B (i)> St+1 A(i) , (9)
of (10) for every 1 ≤ i ≤ N . Note that T in (11) is the
control horizon. At any time t, the scheduler decides which
M among the N subsystems may communicate. Note that
(i)
et|t = 0 when a communication channel is assigned to
subsystem i at time t.
In the following section, we present a deep reinforcement learning algorithm for control-aware scheduling called
D EEP CAS. D EEP CAS communicates only with the smart
sensors. At every time instant, sensors are told if they can
transmit to their associated controllers. Then, the sensors
provide feedback on the scheduling decision for that stage.
Note that we do not consider the overhead involved in
providing feedback.
III. D EEP REINFORCEMENT LEARNING FOR
CONTROL - AWARE SCHEDULING
As stated earlier, at the heart of our D EEP CAS lies the
DQN. The DQN is a modern variant of Q-learning that
effectively counters Bellman’s curse of dimensionality. Essentially, DQN or Q-learning finds a solution to an associated
Markov decision process (MDP) in an iterative model-free
manner. Before proceeding, let us recall the definition of an
MDP. For a more detailed exposition, the reader is referred
to [10]. An MDP, M, is given by the following tuple
(S, A, P, r, γ), where
S is the state-space of M.
A is the set of actions that can be taken.
P is the transition probability, i.e., P (s, s0 ; a) is the probability of transitioning to state s0 when action a is taken
at state s.
r is the one stage reward function, i.e., r(s, a) is the
reward when action a is taken at state s.
γ is the discount factor and ∈ [0, 1].
Below, we state the MDP Md associated with our problem.
S: The state space S consists of all possible augmented
error vectors. Hence, state vector st at time t is given
(1)
(N )
by (et|t , . . . , et|t ).
A: Action space is {S | S ⊂ {1, 2, . . . , N }, |S| = M }.
Hence, the cardinality (size), |A|, of the action space
N
equals M
.
N
P
(i) (i) (i)
r: At time t, the reward r(t) is given by −
et|t Γt et|t .
i=1
γ: Although it would seem natural to use γ = 1, we use
0 < γ < 1 since it hastens the rate of convergence.
Note that the scheduler (D EEP CAS) takes action just before
time t and receives rewards just after time t, based on
transmissions at time t. Also, note that D EEP CAS only gets
non-zero rewards from non-transmitting sensors. D EEP CAS
is model-free. Hence, it does not need to know transition
probabilities.
Let us suppose we use a reinforcement learning algorithm,
such as Q-learning, to solve Md . Since the learning algorithm will find policies that minimize the future expected
cumulative rewards, we expect to find policies that minimize
scheduling effects on the entire system. This is a consequence
of our above definition of reward r. In the following section,
we provide a brief overview of Q-learning and DQN, the
reinforcement learning algorithm at the heart of our D EEP CAS. Simply put, D EEP CAS is a DQN solving the above
defined MDP Md .
D EEP CAS. At any time t0 , the scheduler (D EEP CAS) is
interested in maximizing the following expected discounted
future reward:
"T −1
#
X
t−t0
R(t0 ) := E
γ
r(t) ,
t=t0
where r(t) is the single stage cost given by −
N
P
i=1
(i) (i) (i)
et|t Γt et|t .
Q-learning is a useful methodology to solve such problems.
It is based on finding the following Q-factor for every stateaction pair:
∗
Q (s, a) := max E [Rt | st = s, at = a, π] ,
π
where π is a policy that maps states to actions. The algorithm
itself is based on the Bellman equation:
Q∗ (s, a) = Es0 ∼E [r + γ 0max 0 Q∗ (s0 , a0 ) | s, a] .
a ∈U (s )
Since our state space is continuous, we use a deep neural
network (DNN) for function approximation. Specifically, we
try to find good approximations of the Q-factors iteratively.
In other words, the neural network takes as input state s
and outputs Q(s, a, θ) for every possible action a, such that
Q(s, a, θ) ≈ Q∗ (s, a). This deep function approximator, with
weights θ, is referred to as a Deep Q-Network. The Deep QNetwork is trained by minimizing a time-varying sequence
of loss functions Lt (θt ) given by
Lt (θt ) = 1/2 Es,a∼ρ(s,a) (Q(s, a, θt ) − yt )2 ,
where yt := Es0 ∼E [r + γ maxa0 Q(s0 , a0 , θt−1 ) | s, a] is the
expected cost-to-go based on the latest update of the weights;
ρ is the behavior distribution. Training the neural network
involves finding θ∗ , which minimizes the loss functions.
Since the algorithm is online, training is done in conjunction with scheduling. At time t, after feedback (reward) is
received, one gradient descent step can be performed using
the following gradient term:
∇θt Lt (θt ) = Es,a∼ρ(·);s∼E Q(s, a; θt ) − r
0 0
− γ max
Q(s , a , θt−1 ) ∇θt Q(s, a; θt ) . (12)
0
a
To make the algorithm implementable, we use a sample rather than find the above expectation, when updating
weights. At each time, we pick actions using the -greedy
approach. Specifically, we pick a random action with probability , and we pick a greedy action with probability
1 − . This constitutes the previously mentioned behavior
distribution ρ, which governs how actions are picked. Note
that a greedy action at at time t is one that maximizes
Q(st , a; θ). Initially it is desirable to explore, hence is set
to 1. Once the algorithm has gained some experience, it is
better to exploit this experience. To accomplish this, we use
an attenuating .
Although we train our DNN in an online manner, we
do not perform a gradient descent step using (12), since it
can lead to poor learning. Instead, we store the previous
K experiences (st , at , rt , st+1 ), t0 − K + 1 ≤ t ≤ t0 , in
an experience replay memory D. In other words, at time
t, D EEP CAS stores (st , at , rt , st+1 ) into D whose size is
K. When it comes to training the neural network at time
t, it performs a single mini-batch gradient descent step.
The mini-batch (of gradients) is randomly sampled from
the aforementioned experience replay D. The idea of using
experience replay memory, to overcome biases and to have
a stabilizing effect on algorithms, was introduced in [8].
DQN for control-aware scheduling
1: Initialize the replay memory D to capacity K.
2: Initialize the weights, θ, of the Q-Network.
3: for the entire duration do
4:
With probability select a random action at .
5:
With probability 1 − pick at that maximizes
Q(st , a, θ).
6:
Execute action at to obtain reward rt and observe
st+1 .
7:
Store (st , at , rt , st+1 ) in D.
8:
Sample random mini-batch transitions from D.
9:
Corresponding to (sj , aj , rj , sj+1 ), set
yj := rj + γ max
Q(sj+1 , a0 ; θ).
0
a
10:
Perform a gradient descent step with loss given by
(yj − Q(sj , aj ; θ))2 .
100
We are now ready to present our numerical results. Recall
that a DQN is at the heart of our D EEP CAS, which uses
a deep neural network to approximate Q-factors. Before
proceeding, we specify the algorithm parameters. The input
to the neural network is the appended error vector. The
hidden layer consists of 1024 rectifier units. The output layer
is a fully connected
linear layer with a single output for each
N
of the M
actions. The discount factor γ in our Q-learning
algorithm is 0.95. The size of the experience replay buffer
is fixed at 20000. The exploration parameter is initialized
to 1, then attenuated to 0.001 at the rate of 0.9. For training
the neural network, we use the optimizer ADAM [11] with
a learning rate of e−4 and a decay of 0.001. Note that we
used the same set of parameters for all of the experiments
presented below.
Experiment 1: N=3, M=1, and T=500. For our first experiment, we used D EEP CAS to schedule one channel for
a system with three subsystems, over a control horizon of
length 500. We considered three second-order single-inputsingle-output (SISO) subsystems consisting of one stable
(subsystem 2) and two unstable subsystems (subsystems 1
and 3). If there were three channels, then there would be
no scheduling problem and the total optimal control loss
J would be 13.8487. Since there is a single channel, one
expects a solution to the scheduling problem to allocate it
to subsystems 1 and 3 for a more substantial fraction of the
time, as compared to subsystem 2. This expectation is fair
since subsystems 1 and 3 are unstable while subsystem 2 is
stable. Once trained, D EEP CAS indeed allocates the channel
to subsystem 1 for 52% of the time, to subsystem 2 for 12%
of the time, and to subsystem 3 for 36% of the time.
We train D EEP CAS continuously over many epochs. Each
epoch corresponds to a single run of the control problem with
horizon 500. At the start of each epoch, the initial conditions
for the control problem are chosen as explained in § II. Fig. 2
illustrates the learning progress of D EEP CAS. The abscissa
axis of the graph represents the epoch number while the
ordinate axis represents the average control loss. The plot is
obtained by taking the mean of 30 Monte Carlo runs. Since
DQN is randomly initialized, scheduling decisions are poor
at the beginning, and the average control loss is high. As
learning proceeds, the decisions taken improve. After only
10 epochs, D EEP CAS converges to a scheduling strategy
with an associated control loss of around 21.
Comparison to periodic scheduling. Traditionally, the problem of scheduling for control systems is solved by using
control theoretic heuristics to find periodic schedules. We
do the following to compare the performance of D EEP CAS
to such techniques. For every fixed n ∈ {2, . . . , 11}, we
perform an exhaustive search (among 3n sequences) to find
a scheduling sequence, of length n, which minimizes the
control loss. The results are listed in Table I1 . As illustrated
in Table I, this methodology yields a minimum control loss
1 We stopped at n = 11 since exhaustive search is extremely timeconsuming and impractical beyond that.
Empirical average cost
IV. N UMERICAL RESULTS
80
60
40
20
0
5
10
15
20
25
30
Epochs
Fig. 2.
The convergence of the empirical average control loss for
D EEP CAS governing a networked system having three subsystems and one
communication channel.
TABLE I
O PTIMAL P ERIODIC S CHEDULES FOR THE T HREE S UBSYSTEMS
Length n
2
3
4
5
6
7
8
9
10
11
Sequence
{1, 3}
{1, 3, 2}
{1, 3, 1, 2}
{1, 3, 1, 3, 2}
{1, 3, 1, 3, 1, 2}
{1, 3, 1, 3, 1, 3, 2}
{1, 3, 1, 3, 1, 3, 1, 2}
{1, 3, 1, 3, 1, 3, 1, 3, 2}
{1, 3, 1, 3, 1, 3, 1, 3, 1, 2}
{1, 3, 1, 3, 1, 3, 1, 3, 1, 3, 2}
Total cost J
29.0510
25.7798
25.2363
23.6020
23.7801
22.9636
23.2944
22.8112
23.1738
22.8601
of J = 22.8112 for n = 9. In comparison, D EEP CAS finds a
scheduling strategy with an associated control loss of 21.15.
We conclude that, in addition to being faster, D EEP CAS does
not need any system specification and can schedule efficiently
for very long control horizons.
Another approach to scheduling is to solve an associated mixed-integer quadratic program (MIQP); see [12]
for details. Solving a MIQP accurately is computationally
infeasible for all but small system sizes.
Experiment 2: N=6, M=3, and T=500. For our second
experiment, we train D EEP CAS to schedule three channels
for a system with six second-order SISO subsystems. If
N = M , then the total control loss equals 18.234. As before,
learning is done continuously over many epochs. Fig. 3
illustrates the learning progress of D EEP CAS in scheduling
three channels among six subsystems. The abscissa and
ordinate axes are as before. As evidenced in the figure,
D EEP CAS quickly finds schedules with an associated control
loss of around 20. Due to the difference in scale, the control
loss in Fig. 3 seems to have high variance as compared to
Fig. 2.
We are unable to compare the results of Experiment 2
with any traditional heuristics. This is because traditional
heuristics do not extend to the system size and control
horizon considered here. Further, performing an exhaustive
search for finding periodic schedules is not possible
n since
the number of possibilities are in the order of 63 = 20n ,
where n is the period-length.
Empirical average cost
Empirical average cost
24
23
22
21
20
0
10
20
30
40
50
Epochs
Fig. 3.
The convergence of the empirical average control loss for
D EEP CAS governing a networked system having six subsystems and three
communication channels.
Scheduling in general control settings. The systems considered hitherto have independent subsystems. This facilitates
the splitting of the total control cost into two components;
see (10). The one-stage reward in our algorithm is the
negative of the error due to lack of communication defined
in (11). However, in general multi-agent settings, the previously mentioned splitting may not be possible.
Now, we discuss an extension to our algorithm which
will allow for scheduling in these general scenarios. For
such an extension, we observe that one merely needs to
provide a recipe for obtaining the state-space and reward in
the definition of Md (associated MDP). Since we want our
scheduler to reduce the control loss, it seems that a natural
choice for reward is the negative of the one-stage control
cost. Regarding state-space, one may need to consider the
state (or its estimate) control-action pairs, say (xt , ut ), of the
system. It may also be necessary to append the pairs (xt , ut )
of all the subsystems. The expected control cost associated
with taking action while in some state does not change with
time. This allows for a consistent definition of Md . It may
be noted that there may be other ways to define an associated
MDP.
To show that the above extension is viable, we repeated
Experiments 1 and 2 with negative of the one-stage control
cost as the reward. The results of the modified experiments
are very similar to the original ones. Due to constraints in
space, we only present Fig. 4 associated with the modification of Experiment 1.
V. C ONCLUSIONS
This paper considered the problem of scheduling the
sensor-to-controller communication in a networked control
system, consisting of a multitude of independent control
subsystems, in the presence of scarce communication resources. For this reason, we developed D EEP CAS, i.e., a
reinforcement learning-based control-aware scheduling algorithm. This algorithm is model-free and scalable, and
it outperforms scheduling heuristics tailored for feedback
control applications. Specifically, we compared our algorithm
with periodic schedules. Also, we briefly discussed how
D EEP CAS could be extended to a networked control system,
100
80
60
40
20
0
10
20
30
40
50
Epochs
Fig. 4.
The convergence of the empirical average control loss for
D EEP CAS governing a networked system having three subsystems and one
communication channel (reward as −J).
containing coupled subsystems.
As stated earlier, this paper does not address the problem
associated with overheads. An exciting future direction could
be the development of sophisticated reinforcement learning
algorithm which can reduce overhead. Addressing the aforementioned issues may lead to the associated problem of large
discrete action spaces. The algorithms thus developed would
need to take this into account. Lastly, it is also interesting to
consider imperfect communication scenarios.
R EFERENCES
[1] P. Park, S. C. Ergen, C. Fischione, C. Lu, and K. H. Johansson,
“Wireless network design for control systems: a survey,” IEEE Communications Surveys & Tutorials, 2018.
[2] H. Rehbinder and M. Sanfridson, “Scheduling of a limited communication channel for optimal control,” Automatica, vol. 40, no. 3, pp.
491–500, March 2004.
[3] D. Hristu-Varsakelis and L. Zhang, “LQG control of networked control
systems,” International Journal of Control, vol. 81, no. 8, pp. 1266–
1280, 2008.
[4] L. Shi, P. Cheng, and J. Chen, “Optimal periodic sensor scheduling
with limited resources,” IEEE Transactions on Automatic Control,
vol. 56, no. 9, pp. 2190–2195, 2011.
[5] L. Orihuela, A. Barreiro, F. Gómez-Estern, and F. R. Rubio, “Periodicity of Kalman-based scheduled filters,” IEEE Transactions on
Automatic Control, vol. 50, no. 10, pp. 2672–2676, 2014.
[6] L. Zhao, W. Zhang, J. Hu, A. Abate, and C. J. Tomlin, “On the optimal
solutions of the infinite-horizon linear sensor scheduling problem,”
IEEE Transactions on Automatic Control, vol. 59, no. 10, pp. 2825–
2830, March 2014.
[7] Y. Mo, E. Garone, and B. Sinopoli, “On infinite-horizon sensor
scheduling,” Systems & Control Letters, vol. 67, pp. 65–70, May 2014.
[8] V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou,
D. Wierstra, and M. Riedmiller, “Playing atari with deep reinforcement
learning,” in NIPS Deep Learning Workshop, 2013.
[9] B. Demirel, A. S. Leong, V. Gupta, and D. E. Quevedo, “Trade-offs
in stochastic event-triggered control,” arXiv:1708.02756, 2017.
[10] D. P. Bertsekas and J. N. Tsitsiklis, Neuro-Dynamic Programming.
Athena Scientific, 1996.
[11] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in Proceeding of the 3rd International Conference for Learning
Representations, 2015.
[12] T. Charalambous, A. Ozcelikkale, M. Zanon, P. Falcone, and
H. Wymeersch, “On the resource allocation problem in wireless networked control systems,” in Proceedings of the 57th IEEE Conference
on on Decision and Control, 2017.
| 3cs.SY
|
arXiv:1509.04509v8 [math.GR] 23 Aug 2017
FINITE BANDS ARE FINITELY RELATED
IGOR DOLINKA
Abstract. We prove that every finite idempotent semigroup (band) is finitely
related, which means that the clone of its term operations (i.e. operations
induced by words) is determined by finitely many relations. This solves an
open problem posed by Peter Mayr in 2013.
1. Introduction
Let A = (A, F ) be an algebra, where F is a family of finitary operations on the
set A. As is well known, any term t(x1 , . . . , xn ) (of the same similarity type as A)
induces, by interpretation in A, an operation tA : An → A. Operations obtained
in this way are the term operations of A, and the collection of all term operations
of this algebra is denoted by Clo(A) and called the clone of A.
For an n-ary operation f on a set X (f : X n → X) and a k-ary relation ρ on
the same set (ρ ⊆ X k ) we say that f preserves ρ if for any a1 , . . . , ak ∈ X n (where
ai = (ai,1 , . . . , ai,n ) for 1 ≤ i ≤ k) such that (a1,j , . . . , ak,j ) ∈ ρ for all 1 ≤ j ≤ n
we have (f (a1 ), . . . , f (ak )) ∈ ρ; in other words, either ρ = ∅ or ρ is a subalgebra of
the kth direct power of the algebra (X, f ). For a family R of (finitary) relations on
X we denote by Pol(R) the set of all operations on X preserving all relations from
R – these are the polymorphisms of the relational structure (X, R). Conversely, if
O is a family of operations on X, the set of all relations preserved by all operations
from O is denoted by Inv(O). A foundational result in clone theory tells us that,
for finite A, Clo(A) = Pol(Inv(F )), and in this sense the clone of A is determined
by the collection of relations Inv(F ): an operation on A arises from a term if and
only if it preserves all the relations that are preserved by F .
The set of relations Inv(F ) is always infinite. However, it may happen that there
is in fact a finite subset R ⊆ Inv(F ) such that Clo(A) = Pol(R), so that the clone
of A is determined by a finite set of relations (which then can be reduced to a single
relation). In such a case we say that the algebra A is finitely related.
Finitely related algebras recently received a lot of attention in universal algebra
[8] and its applications in computer science because of their link, exhibited e.g.
by [2, 5], to other key algebraic properties pertinent to the algebraic approach in
studying the computation complexity of constraint satisfaction problems (CSP) [7].
See also [1, 3, 4, 16] for some recent developments and important results regarding
finitely related general algebras. For example, a major result of Aichinger et al. [2]
implies that any finite algebra having a Mal’cev term operation (including all finite
2010 Mathematics Subject Classification. Primary 20M07; Secondary 03C05, 08A40, 20M05.
Key words and phrases. Semigroup; Idempotent semigroup; Band; Finitely related algebra.
This research was supported by the Ministry of Education, Science, and Technological Development of the Republic of Serbia through the grant No.174019.
1
2
IGOR DOLINKA
groups and rings) is finitely related, a consequence of which is that there are only
countably many Mal’cev clones on any finite set.
Concerning the study of finitely related (finite) semigroups, the seminal paper
is the one of Davey et al. [9], where it was shown (among other things) that finite
semigroups that are either commutative or nilpotent enjoy the finitely related property. This paper was followed by an extensive study of Mayr [17], who exhibited
the first example of a non-finitely related finite semigroup (not too surprisingly, this
was the ‘infamous’ six-element Brandt monoid B21 , which seems to behave badly
with respect to almost any conceivable equational property of semigroup varieties).
Also, Mayr proved that any finite regular band (an idempotent semigroup satisfying the identity xyxzx ≈ xyzx) is finitely related. The question of whether all
finite bands are finitely related is left as an open problem (Problem 6.3); the same
question is mentioned following Problem 7.2 in [15]. It is exactly this problem that
we aim to address in the present paper; namely, we prove the following main result.
Theorem 1.1. Let S be a finite idempotent semigroup. Then S is finitely related.
Here is the brief outline of the paper. The next, preliminary section is divided
into three parts. First, we are going to invoke few criteria (from [9, 17]) for a finite
algebra to be finitely related. Along the way, we are going to introduce a handy
concept of an n-scheme of terms, and specialise all these concepts to semigroups
and words, respectively. Then we are going to review the lattice of all varieties of
bands (idempotent semigroups) and proceed to a full effective description of their
equational theories. Finally, we complete the second section by some auxiliary
results that will be used in the proof of the previous theorem, presented in the third
section. This proof uses induction on the ‘height’ of the variety that S generates
in the lattice of all band varieties, depicted in Fig. 1, taking the mentioned result
on regular bands as the induction basis. We first resolve the case when S generates
one of the irreducible varieties in that lattice (Theorem 3.1) and then demonstrate
how to derive Theorem 1.1 in its full generality from the irreducible case (Theorem
3.2).
2. Preliminaries
2.1. Term schemes and criteria for finitely related finite algebras. Let
f : An → A be an n-ary operation on the set A. For i, j ∈ n = {1, . . . , n}, i < j,
we define an operation fij : An → A (sometimes called an identification minor of
f ) by
fij (x1 , . . . , xn ) = f (x1 , . . . , xi−1 , xj , xi+1 , . . . , xj , . . . , xn );
in other words, fij is obtained from f by identifying the variable xi with xj . Similarly, if t = t(x1 , . . . , xn ) is a term of a given similarity type, then by t(ij) we denote
the term obtained from t by replacing each occurrence of xi in it by xj ; furthermore,
by convention we define t(ji) to be t(ij) .
As usual, we say that an operation f : An → A depends on its ith variable xi ,
i ∈ n, if there exist a1 , . . . , ai−1 , ai+1 , . . . , an , b, c ∈ A such that
f (a1 , . . . , ai−1 , b, ai+1 , . . . , an ) 6= f (a1 , . . . , ai−1 , c, ai+1 , . . . , an ).
For example, the identification minor fij does not depend on xi . In turn, as another
example, the ith projection operation fails to depend on any of its variables but xi .
FINITE BANDS ARE FINITELY RELATED
3
Now let V be a variety and assume
S = {tij : 1 ≤ i < j ≤ n}
is a family of terms over Xn = {x1 , . . . , xn } of the similarity type of V satisfying
the following conditions:
(D) For each A ∈ V, the term operation tA
ij (x1 , . . . , xn ) does not depend on the
variable xi .
(C1) For any four distinct 1 ≤ i, j, p, q ≤ m such that i < j and p < q, V satisfies
the identity
(pq)
tij ≈ t(ij)
pq .
(C2) For any three 1 ≤ i < j < k ≤ n, V satisfies the identities
(jk)
tij
(ik)
(jk)
≈ tjk ≈ tik .
The condition (D) is called dependency, while (C1) and (C2) are the consistency
conditions; the family S is called an n-scheme of terms for V. Bearing in mind
Definition 2.3 and Notation 2.7 from [9] it is an easy exercise to see that this is just
equivalent to the notion of an (n, n − 1)-scheme introduced in that paper.
Similarly, it is very easy to see that the following holds.
Lemma 2.1. Let V be an arbitrary variety and t(x1 , . . . , xn ) a term of the same
similarity type as V. The family of all identification minors of t,
{t(ij) : 1 ≤ i < j ≤ n},
is an n-scheme of terms for V.
Again following [9], we say that an n-scheme S = {tij : 1 ≤ i < j ≤ m} for V
comes from the term t if S is V-equivalent to the family of all identification minors
of t in the sense that V satisfies the identities
tij ≈ t(ij)
for all 1 ≤ i < j ≤ n.
The following characterisation is part of Theorem 2.9 from [9], see also [14, 17,
18, 19].
Theorem 2.2. Let A be a finite algebra, and let V be the variety generated by A.
The following are equivalent:
(1) A is finitely related.
(2) There exists n0 ≥ |A| such that for all n > n0 , every n-scheme of terms for
V comes from a term.
(3) There exists n0 ≥ |A| such that for all n > n0 , an operation f : An → A
is a term operation of A whenever all minors fij (i, j ∈ n, i < j) are term
operations of A.
Here we immediately invoke Remark 2.12 from [9] which provides an argument
that in checking if a finite algebra A satisfies condition (3) above, it can be assumed
that the operation f : An → A is essential, i.e. that it depends on all of its
variables. Therefore, we will effectively use the previous theorem in the form where
the condition (3) is replaced by
(3’) There exists n0 ≥ |A| such that for all n > n0 , an operation f : An → A
depending on all of its variables is a term operation of A whenever all
minors fij (i, j ∈ n, i < j) are term operations of A.
4
IGOR DOLINKA
Notice that for semigroups and semigroup varieties all definitions and results
mentioned above remain valid when we replace terms by words (i.e. elements of the
free semigroup Xn+ ), term operations by operations induced by words, etc., even
though, strictly speaking, words are not terms. However, it is true that every term
operation on a semigroup is induced by a word and vice versa. Therefore, it is
meaningful to define the notion of an n-scheme of words for a semigroup variety,
and the ‘semigroup version’ of Theorem 2.2 holds as well: a finite semigroup S is
finitely related if and only if every n-scheme of words for the variety generated by
S comes from a single word (provided n is large enough) if and only if the condition
(3’) holds with respect to operations induced by words. It was proved in [17] that
this is the case when S is a regular band, that is, a finite idempotent semigroup
satisfying xyxzx ≈ xyzx. The aim of this paper is to extend this to all finite bands
in an inductive manner, taking the result of [17] as a basis of the induction. Hence,
in the next subsection we need to gather some basic information about varieties of
bands and few auxiliary results describing their equational theories.
2.2. Varieties of bands and their equational theories. The variety of all
bands will be denoted by B. It is known that any semigroup variety containing B
(including B itself) is not generated by a finite semigroup [20], while every proper
subvariety of B is finitely generated (see [6, 10, 11]).
In what follows, we adopt the notation of [13] for certain operators acting on
words; we briefly recall it for completeness. For a word w over a (finite) alphabet
X, let c(w) denote the content of w, the set of all letters occurring in w. Further,
let s(w) denote the longest prefix of w containing all but one of the letters from
c(w) (so that |c(s(w))| = |c(w)| − 1 and s(w) is maximal with this property), while
σ(w) is the last letter to occur in w from the left (implying that s(w)σ(w) is the
shortest prefix of w with the same content as w). Dually, let e(w) be the longest
suffix of w containing all but one of the letters from c(w), and let ε(w) be the last
letter in w to occur from the right.
Define an operator b on words by induction on the size of their content such that
b(∅) = ∅ (here ∅ stands for the empty word) and
b(w) = bs(w)σ(w)ε(w)be(w)
(note the concatenation of word operators, where bs(w) means b(s(w)), etc.).
Theorem 2.3 (cf. Lemma 2.7 and Theorem 2.9 of [12]). Let X be an alphabet and
u, v ∈ X + . Then u ≈ v holds in B if and only if b(u) = b(v). In particular, for any
word w we have that the identity
w ≈ s(w)σ(w)ε(w)e(w)
holds in any band.
This immediately implies the following well-known fact (recorded, e.g., as Corollary 2.3 in [13]).
Corollary 2.4. Let u, v, w be words such that c(v) ⊆ c(u) = c(w). Then the identity
uvw ≈ uw holds in B.
Proof. Under the given conditions, s(uvw) = s(u) = s(uw) and σ(uvw) = σ(u) =
σ(uw), and, dually, e(uvw) = e(w) = e(uw) and ε(uvw) = ε(w) = ε(uw), so the
result follows by Theorem 2.3.
FINITE BANDS ARE FINITELY RELATED
5
sB
..
.
..
.
❅s A
❅s
A4 s
4
❅
❅
❅
❅
❅s
❅s
❅
❅
❅
❅
❅s
❅s B
B3 s
3
❅
❅
❅
❅
❅s
❅s
❅
❅
❅
❅
❅s
❅s A
A3 s
3
❅
❅
❅
❅
❅s
❅s
❅
❅
❅
❅
❅s
❅s B = RRB
B2 = LRB s
2
❅
❅
❅
❅
❅s
❅s
s
❅
❅
❅
❅
❅s
❅s A = RZ
A2 = LZ s
2
SL
❅
❅
❅s
T
Figure 1. The lattice of all varieties of bands
The description of the lattice L (B) of all subvarieties of B is today still considered as one of the most glaring success stories of the theory of semigroup varieties;
it was achieved independently by Biryukov [6], Fennemore [10], and Gerhard [11],
whereas a more ‘economical’ and coherent treatment of the subject was given later
in [13]. Figure 1 provides a diagram of this lattice: here SL, LZ and RZ are
respectively the varieties of semilattices, left zero and right zero bands.
Of particular importance are the two sequences of join- and meet-irreducible
varieties Am and Bm , m ≥ 2, as well as their duals Am and Bm – any proper
(i.e. finitely generated) band variety not contained in SL ∨ LZ ∨ RZ (the ‘bottom
cube’) is a join of a pair of these. In fact, we shall first prove that any finite band
generating one of the varieties Am , Bm is finitely related; by left-right duality, this
will extend to Am and Bm , and then we shall show how to use this fact to prove
finite relatedness for any finite band generating a proper subvariety of B. The
varieties LRB and RRB are the varieties of left regular and right regular bands
defined by identities xyx ≈ xy and xyx ≈ yx, respectively. Their join is the variety
of regular bands that is the subject of Theorem 6.2 of [17]. As we remarked earlier,
our approach will be inductive with respect to the chains formed by varieties Am ,
6
IGOR DOLINKA
Bm and their duals, with the latter result from [17] serving as a basis of that
induction.
For technical convenience in our main argument it will be useful to select and
fix finite bands Am , Bm , m ≥ 2 (as well as their dual semigroups Am and Bm ),
such that Am generates the variety Am and Bm generates Bm (For example, one
can choose A2 to be the two-element left zero band, B2 to be A2 with an identity
element adjoined, etc. Note that we can choose these bands so that |A2 | < |B2 | <
|A3 | < |B3 | < . . . holds.)
Our main task at this moment is to describe the equational theories of the
considered band varieties, following the approach laid out in [13]. To this end, we
introduce word functions hm and im for m ≥ 2 and their duals, hm , im , in the sense
that if w denotes the reverse of the word w we have tm (w) = tm (w) for t ∈ {h, i}:
• tm (∅) = ∅ for all m ≥ 2 and t ∈ {h, i};
• h2 (w) is the first letter of w from the left (the head of w);
• i2 (w) is the word obtained from w by retaining only the first occurrence
from the left of each letter (the initial part of w, also defined recursively by
i2 (w) = i2 s(w)σ(w));
• for m ≥ 3 and t ∈ {h, i} we set
tm (w) = tm s(w)σ(w)tm−1 (w).
The key feature of these functions is expressed in the following statement.
Theorem 2.5 ([13]). Let m ≥ 2 and let u, v be two words.
• Am satisfies u ≈ v if and only if hm (u) = hm (v).
• Bm satisfies u ≈ v if and only if im (u) = im (v).
Analogous equivalences hold for the dual varieties Am and Bm and functions hm
and im , respectively.
The next few properties will be used in our proofs.
Lemma 2.6.
(1) Let t ∈ {h, i}. If m ≥ 3 or tm = i2 then stm (w) = tm s(w)
and etm (w) = tm e(w) for any word w.
(2) Let t ∈ {h, i}. If m ≥ 4 or tm = i3 then
btm (w) = b(tm s(w)σ(w)ε(w)tm−1 e(w))
and
btm (w) = b(tm−1 s(w)σ(w)ε(w)tm e(w))
for any word w.
(3) Let t ∈ {h, i, h, i} and let u, v be any words. If m ≥ 2 then btm (u) = btm (v)
implies tm (u) = tm (v).
(4) Let t ∈ {h, i}. If m ≥ 3 then
tm (w) = tm−1 (w)ε(w)tm e(w)
for any word w.
Proof. The first part of (1) is just [13, Lemma 3.2(ii)], while the second part is dual
to the first one and follows from it, as etm (w) = etm (w) = stm (w) = tm s(w) =
tm (e(w)) = tm e(w). (2)–(4) are Lemma 3.3(iii),(v), Lemma 4.1 and Lemma 3.3(ii)
of [13], respectively, where (2) and (3) are reformulated in terms of the operator b,
bearing in mind Theorem 2.3.
FINITE BANDS ARE FINITELY RELATED
7
Lemma 2.7. Let X ∈ {A, B}. Assume that the variety Xm satisfies the identity
u ≈ v, where |c(u)|, |c(v)| ≥ 2. Then it also satisfies s(u) ≈ s(v).
Proof. Consider first the case when Xm = A2 . Since both u and v contain at least
two letters, we have h2 s(u) = h2 (u) = h2 (v) = h2 s(v), and the lemma follows.
Otherwise, if either Xm = B2 or m ≥ 3, let tm be the corresponding word function
in the sense of Theorem 2.5. By the given conditions we have tm (u) = tm (v), which
in turn implies stm (u) = stm (v). Then Lemma 2.6(1) tells us that tm s(u) = tm s(v)
(since we assumed that tm 6= h2 ). Another application of Theorem 2.5 yields that
Xm satisfies s(u) ≈ s(v).
2.3. Some auxiliary results. As Theorem 2.2 suggests, we will be concerned with
essential operations f : S n → S on a finite (idempotent) semigroup S such that all
of its minors fij are induced by words. In the next lemma we are looking at some
consequences of such a setting.
Lemma 2.8. Let S be a finite semigroup, and let f : S n → S be an operation
depending on all of its variables such that for any i, j ∈ n, i < j, there is a word
S
wij satisfying fij = wij
.
(1) {wij : 1 ≤ i < j ≤ n} is an n-scheme of words for the semigroup variety
generated by S.
(2) If the variety generated by S contains SL and n ≥ |S| + 4 then c(wij ) =
Xn \ {xi }.
Proof. (1) Easy, and largely analogous to Lemma 2.1, dealing with minors of operations instead of terms.
(2) Since f depends on all of its variables (and thus in particular on xk , k 6= i),
there exist a1 , . . . , ak−1 , ak+1 , . . . , an , b, c ∈ S such that
f (a1 , . . . , ak−1 , b, ak+1 , . . . , an ) 6= f (a1 , . . . , ak−1 , c, ak+1 , . . . , an ).
The pigeonhole principle ensures that there exist p, q ∈ n \ {i, j, k}, p < q, such that
S
depends on xk . However, by
ap = aq . Hence, just as in [17, Lemma 2.6], fpq = wpq
(ij)
(pq)
(1), S satisfies the identity wpq ≈ wij
that wij must contain xk , since k 6= i.
(ij)
(pq)
and so xk ∈ c(wpq ) = c(wij ), ensuring
In the following, n-schemes of words {wij : 1 ≤ i < j ≤ n} over Xn such that
c(wij ) = Xn \ {xi } will be called essential ; so, the above result states that essential
operations on finite semigroups whose varieties contain nontrivial semilattices with
all minors induced by words give rise to essential n-schemes, provided n is large
enough.
Lemma 2.9. Let {wij : 1 ≤ i < j ≤ n} be an essential n-scheme of words for a
semigroup variety V containing B2 (the variety of left regular bands), where n ≥ 5.
Then there exists a unique permutation π of n such that for any i < j,
i2 (wij ) = xα1 · · · xαn−1 ,
where the sequence π (ij) = (α1 , . . . , αn−1 ) is obtained from (1π, . . . , nπ) by replacing
i by j and then deleting the right one of the two occurrences of j. Furthermore, if the
given essential scheme comes from a word w then we must have i2 (w) = x1π · · · xnπ .
Proof. By the very definition of an (essential) n-scheme of words for a variety,
any scheme for a variety V is also a scheme for any of its subvarieties, and so
8
IGOR DOLINKA
is the case for the given scheme with respect to B2 . As already discussed, B2 is
generated by the 3-element band B2 (obtained by adjoining an identity element
to the 2-element left zero band), and by Lemma 6.1 of [17], B2 is finitely related
with degree at most 4. By [9, Lemma 2.6], there exists a unique operation f :
B2
for all 1 ≤ i < j ≤ n. To see that f depends
B2n → B2 such that fij = wij
B2
on all of its variables, fix k ∈ n and p, q ∈ n \ {k}, p < q. As fpq = wpq
and
xk ∈ c(wpq ), fpq depends on xk , so there exist a1 , . . . , ak−1 , ak+1 , . . . , an , b, c ∈
B2 such that fpq (a1 , . . . , ak−1 , b, ak+1 , . . . , an ) 6= fpq (a1 , . . . , ak−1 , c, ak+1 , . . . , an ).
This immediately implies that f depends on xk .
Hence, by [17, Lemma 6.1], f is induced by a word w such that c(w) = Xn . Now
it is routine to see that the permutation π ∈ Sn satisfying
i2 (w) = x1π · · · xnπ
meets all the requirements of the lemma. It is unique because if π ′ would be
another permutation with the required properties, then it would follow that the
given scheme (with respect to B2 ) also comes from the word w′ = x1π′ · · · xnπ′ ,
whence the uniqueness of f implies that (w′ )B2 = f . Therefore, B2 satisfies w′ ≈ w,
implying i2 (w′ ) = i2 (w) and so π ′ = π.
For an essential scheme of words satisfying the assumptions of the above lemma,
we will call π the associated permutation of the scheme.
We proceed with a technical lemma before we learn another important property
of essential n-schemes of words over irreducible band varieties.
Lemma 2.10. Let k ∈ n, and let w ∈ Xn+ be such that c(w) = Xn \ {xk } and
σ(w) = xl . Furthermore, let p, q ∈ n\{l} such that p < q. Then s(w)(pq) = s(w(pq) ).
Proof. We start by writing w in the form
w = s(w)xl u
(pq)
for some word u, implying w
= s(w)(pq) xl u(pq) . If q 6= k then c(w(pq) ) =
(pq)
Xn \{xk , xp } and c(s(w) ) = Xn \{xl , xk , xp }, which immediately gives s(w(pq) ) =
s(w)(pq) . The same conclusion follows if q = k upon noticing that in such a case we
have c(w(pq) ) = Xn \ {xp } and c(s(w)(pq) ) = Xn \ {xl , xp }.
Lemma 2.11. Let n ≥ 5, let S = {wij : 1 ≤ i < j ≤ n} be an essential n-scheme
of words for Xm , where X ∈ {A, B} and either m ≥ 3 or Xm = B2 . Let π be the
associated permutation of S (which exists by Lemma 2.9) and let l = nπ. Then
S ′ = {s(wij ) : 1 ≤ i < j ≤ n, l 6∈ {i, j}}
is an (n − 1)-scheme for Xm over variables Xn \ {xl }.
Proof. Note that if l 6∈ {i, j} then σ(wij ) = xl and so c(s(wij )) = Xn \ {xi , xl }.
Thus (D) holds.
Furthermore, if p < q are such that {p, q} ∩ {i, j, l} = ∅ then by Lemma 2.10 we
(ij)
(pq)
have s(wij )(pq) = s(wij ) and s(wpq )(ij) = s(wpq ). Further the assumption that
(pq)
(ij)
(pq)
(ij)
Xm satisfies wij ≈ wpq implies, by Lemma 2.7, the identity s(wij ) ≈ s(wpq ).
Hence, Xm satisfies s(wij )(pq) ≈ s(wpq )(ij) , and so S ′ satisfies (C1).
Now let i < j < p be such that l 6∈ {i, j, p}. Again, Lemma 2.10 implies
(jp)
(ip)
(jp)
that s(wij )(jp) = s(wij ), s(wjp )(ip) = s(wjp ) and s(wip )(jp) = s(wip ) (the
lemma was applied over the alphabet Xn \ {xi } in the first and the third case,
FINITE BANDS ARE FINITELY RELATED
9
while it was applied over Xn \ {xj } in the second case). Since, by assumption,
(jp)
(ip)
(jp)
Xm satisfies the identities wij ≈ wjp ≈ wip , it also satisfies the identities
(ip)
(jp)
(jp)
s(wij ) ≈ s(wjp ) ≈ s(wip ) by Lemma 2.7. We conclude that Xm satisfies
s(wij )(jp) ≈ s(wjp )(ip) ≈ s(wip )(jp) , hence (C2) holds for S ′ as well.
We are now ready and equipped to present our main arguments.
3. The main proofs
Theorem 3.1. For any m ≥ 2 and T ∈ {A, B}, the bands Tm and Tm are finitely
related.
Proof. We use induction on m. If m = 2, the result already follows from [17,
Theorem 6.2]. Hence fix m ≥ 3 and assume that the statement is true for values of
indices up to m − 1.
We are going to show that Tm is finitely related by verifying condition (3’) in
Theorem 2.2 with n ≥ n0 = max(|Tm | + 4, m + 3). So, under these assumptions,
n
let f : Tm
→ Tm be an operation that depends on all of its variables such that for
all i, j ∈ n, i < j, we have that the operation fij is induced by a word. Let us
Tm
. Since the variety Xm
select words wij ∈ Xn+ , 1 ≤ i < j ≤ n, such that fij = wij
generated by Tm contains SL, Lemma 2.8 tells us that c(wij ) = Xn \ {xi }, so that
S = {wij : 1 ≤ i < j ≤ n} is an essential n-scheme of words for Xm .
Since B2 is contained in Xm and n ≥ m+ 3 > 5, Lemma 2.9 ensures the existence
(and uniqueness) of the associated permutation π of the scheme S. Let us denote
k = (n − 1)π and l = nπ.
Now, since S is an n-scheme of words for Xm , it is also an n-scheme of words for
any of its subvarieties, and thus in particular for Xm−1 . This variety is generated
by the band Tm−1 which is finitely related by the inductive assumption. Hence,
Theorem 2.2 implies that S, considered as a scheme for Xm−1 , comes from a word,
n
say b
u. Equivalently, there is a word operation g : Tm−1 → Tm−1 such that
T
gij = wijm−1 holds for all 1 ≤ i < j ≤ n.
Similarly, by Lemma 2.11, the family of words S ′ = {s(wij ) : 1 ≤ i < j ≤ n, l 6∈
{i, j}} is an (n − 1)-scheme of words for Xm (and thus for Xm−1 ) over the set of
variables Xn \ {xl }. Again, by employing the inductive hypothesis (and bearing
in mind that n − 1 ≥ (m − 1) + 3 and n − 1 ≥ (|Tm | − 1) + 4 ≥ |Tm−1 | + 4), we
find a word e
u such that S ′ comes from e
u; in other words, there is a word operation
n−1
→ Tm−1 with hij = s(wij )Tm−1 for all i < j such that l 6∈ {i, j}.
h : Tm−1
Now define
u,
uxl b
w = s(wk′ l′ )xk e
where k ′ = min(k, l) and l′ = max(k, l). We are going to argue that f = wTm
whence Theorem 2.2 yields that Tm is finitely related.
Claim. If p, q ∈ n, p < q, are such that {p, q} ∩ {k, l} = ∅ then the identity
w(pq) ≈ wpq
holds in Xm (and so in Tm ).
Proof of Claim. We have
(pq)
u(pq) ,
u(pq) xl b
u(pq) = s(wk′ l′ )xk e
u(pq) xl b
w(pq) = (s(wk′ l′ ))(pq) xk e
(3.1)
10
IGOR DOLINKA
where the second equality follows by Lemma 2.10 using c(wk′ l′ ) = Xn \ {xk′ }.
(pq)
(k′ l′ )
Furthermore, by (C1), Xm satisfies wk′ l′ ≈ wpq , so by Lemma 2.7 it also satisfies
(pq)
(k′ l′ )
s(wk′ l′ ) ≈ s(wpq ). Since {p, q}∩{k, l} = ∅, the sequence π (pq) must have the form
(α1 , . . . , αn−3 , k, l) (where the subsequence (α1 , . . . , αn−3 ) is a certain permutation
of n \ {k, l, p}). Therefore, we may highlight the first occurrence from the left of
each letter from c(wpq ) in wpq by writing
wpq = xα1 u1 xα2 · · · xαn−3 un−3 xk un−2 xl v
for some (possibly empty) words u1 , . . . , un−2 , v. This yields
(k′ l′ )
′ ′
′ ′
(k l )
wpq
= xα1 u1 xα2 · · · xαn−3 un−3 xl′ un−2 xl′ v(k l ) ,
implying
′ ′
(k l )
) = xα1 u1 xα2 · · · xαn−3 un−3 = s2 (wpq ).
s(wpq
Bearing in mind (3.1), we deduce that
′ ′
(k l )
u(pq)
u(pq) xl b
u(pq) = s2 (wpq )xk e
u(pq) xl b
)xk e
w(pq) ≈ s(wpq
holds in Xm . By applying Theorem 2.5 and the recursion for tm twice, we get
u(pq)
u(pq) xl b
tm (w(pq) ) = tm s2 (wpq )xk e
u(pq) xl tm−1 s2 (wpq )xk e
u(pq)
u(pq) xl b
= tm s2 (wpq )xk e
= tm s2 (wpq )xk tm−1 s2 (wpq )xk e
u(pq) .
u(pq) xl tm−1 s2 (wpq )xk e
u(pq) xl b
However, by induction assumption on S ′ for Xm−1 we have that Xm−1 satisfies
e
u(pq) ≈ s(wpq ), and by induction assumption on S for Xm−1 we have that Xm−1 satisfies b
u(pq) ≈ wpq . We know that σ(wpq ) = xl , s(wpq ) = xα1 · · · xαn−3 un−3 xk un−2 ,
σs(wpq ) = xk , and s2 (wpq ) = xα1 · · · xαn−3 un−3 . Thus
u(pq) ≈ s2 (wpq )σs(wpq )s(wpq ) ≈ s(wpq )
s2 (wpq )xk e
and consequently
u(pq) ≈ s(wpq )σ(wpq )wpq ≈ wpq
u(pq) xl b
s2 (wpq )xk e
holds in Xm−1 (here we used the idempotent law, upon noticing that s2 (wpq )σs(wpq )
is a prefix of s(wpq ), while s(wpq )σ(wpq ) is a prefix of wpq ). Hence,
tm (w(pq) ) = tm s2 (wpq )xk tm−1 s(wpq )xl tm−1 (wpq )
= tm s(wpq )xl tm−1 (wpq ) = tm (wpq ).
Another application of Theorem 2.5 concludes the proof of the claim that w(pq) ≈
wpq holds in Xm .
n
Let now a = (a1 , . . . , an ) ∈ Tm
be arbitrary. As n ≥ |Tm | + 4 > |Tm | + 3, by the
pigeohole principle there are p, q ∈ n, p < q, such that ap = aq and {p, q} ∩ {k, l} =
∅. Thus
Tm
f (a) = fpq (a) = wpq
(a) = (w(pq) )Tm (a) = (wTm )pq (a) = wTm (a),
where the previous claim was used in the third equality above. Therefore, f = wTm ,
showing that Tm is finitely related.
The result for Tm follows by left-right duality or simply by the observation that
if a finite semigroup S is finitely related, so is its dual semigroup S. (Indeed, if
{uij : 1 ≤ i < j ≤ n} is an n-scheme of words for the variety V generated by S
FINITE BANDS ARE FINITELY RELATED
11
then it is a routine to show that {uij : 1 ≤ i < j ≤ n} is an n-scheme of words
for the variety V generated by S; so, if the latter scheme comes from a word u it
follows immediately that the former one comes from u. Now an appeal to Theorem
2.2 gives the required result.)
Theorem 3.2. For any m ≥ 3, the bands Am × Am , Am × Bm−1 , Bm−1 × Am ,
Bm × Bm , Bm × Am and Am × Bm are finitely related.
Remark 3.3. Note that this theorem does not follow automatically from the previous
one, as it was shown in [9, Example 6.3] that the finitely related property is in
general not preserved by direct products.
Proof. Generally, we are going to argue that a finite band of the form Tm × Ur is
finitely related, where T, U ∈ {A, B} and r ∈ {m − 1, m}, with combinations as in
the formulation; since the third case is dual to the second and sixth case to the fifth,
these two may be safely omitted by our previous left-right duality remarks. Note
that with these restrictions if Tm generates Xm and Ur generates Yr (X , Y ∈ {A, B})
then Yr is always a subvariety of Xm containing Xm−1 . Equivalently, Xm is a
subvariety of Yr+1 containing Yr . Also, B2 can be assumed to be a subvariety of
both Xm and Yr .
So, let n0 = max(|Tm ||Ur | + 4, m + 3) and assume that for some n ≥ n0 , f is an
n-ary term operation of Tm × Ur depending on all of its variables, with the property
that for all 1 ≤ i < j ≤ n, fij is induced by a word wij ∈ Xn+ .
n
n
Claim. There exist operations g : Tm
→ Tm and h : Ur → Ur such that
f ((a1 , b1 ), . . . , (an , bn )) = (g(a1 , . . . , an ), h(b1 , . . . , bn ))
for all a1 , . . . , an ∈ Tm and b1 , . . . , bn ∈ Ur , and, furthermore, for all 1 ≤ i < j ≤ n,
the operations gij and hij are induced by wij on Tm and Ur , respectively.
Proof of Claim. By Lemma 2.8(1), S = {wij : 1 ≤ i < j ≤ n} is an n-scheme
of words for the variety generated by Tm × Ur , and thus for each of the varieties
generated individually by Tm and Ur . Now by [9, Lemma 2.6] there are (unique)
Tm
n-ary operations g and h on Tm and Ur , respectively, such that gij = wij
and
Ur
for all 1 ≤ i < j ≤ n.
hij = wij
As n > |Tm ||Ur |, for arbitrary (a1 , b1 ), . . . , (an , bn ) ∈ Tm × Ur there are p, q ∈ n,
p < q, such that (ap , bp ) = (aq , bq ). Hence,
f ((a1 , b1 ), . . . , (an , bn )) = fpq ((a1 , b1 ), . . . , (an , bn ))
Tm ×Ur
((a1 , b1 ), . . . , (an , bn ))
= wpq
Tm
Ur
(a1 , . . . , an ), wpq
= wpq
(b1 , . . . , bn )
= (gpq (a1 , . . . , an ), hpq (b1 , . . . , bn ))
= (g(a1 , . . . , an ), h(b1 , . . . , bn )).
By the very construction of g and h, the claim follows.
By the choice of n0 and Theorem 3.1 (in fact, by n satisfying the assumptions
in its proof), both g and h are induced by words, say u and v. We claim that f
is induced on Tm × Ur by the word w = uv. We are going to prove that for all
1 ≤ i < j ≤ n, both Tm and Ur satisfy w(ij) ≈ wij . Similarly as in the final part of
12
IGOR DOLINKA
the proof of the previous theorem, this will suffice to establish the theorem, because
then for arbitrary (a1 , b1 ), . . . , (an , bn ) ∈ Tm × Ur the pigeonhole principle provides
p < q with (ap , bp ) = (aq , bq ), so that for c = ((a1 , b1 ), . . . , (an , bn )), a = (a1 , . . . , an )
and b = (b1 , . . . , bn ) we have
Tm
Ur
(a), wpq
(b))
f (c) = fpq (c) = (gpq (a), hpq (b)) = (wpq
= ((w(pq) )Tm (a), (w(pq) )Ur (b))
= (w(pq) )Tm ×Ur (c) = (wTm ×Ur )pq (c) = wTm ×Ur (c),
showing that f = wTm ×Ur .
Let us start by noting that the scheme S is essential (by Lemma 2.8(2)), which
immediately implies c(u) = c(v) = Xn , so s(w) = s(u), σ(w) = σ(u), ε(w) = ε(v)
and e(w) = e(v). This yields s(w(ij) ) = s(u(ij) ) and e(w(ij) ) = e(v(ij) ). Furthermore, i2 (w) = i2 (u), implying i2 (w(ij) ) = i2 (u(ij) ) = i2 (wij ) by Lemma 2.9
because Xm contains B2 . Thus σ(w(ij) ) = σ(u(ij) ) = σ(wij ) and, dually, ε(w(ij) ) =
ε(v(ij) ) = ε(wij ). Also, if tm is the word function corresponding to the variety Xm
(in the sense of Theorem 2.5), then we know, by construction of words u and v,
that Xm satisfies u(ij) ≈ wij — so that tm (u(ij) ) = tm (wij ) — while Yr satisfies
v(ij) ≈ wij . By assumptions made earlier in this proof, the latter identity holds in
Xm−1 , implying tm−1 (v(ij) ) = tm−1 (wij ).
Note that we have h2 (w(ij) ) = h2 (v(ij) ) since e(w) = e(v). So, if Xm = A3 we
deduce, by Lemma 2.6(1),
h3 (w(ij) ) = h3 s(u(ij) )σ(u(ij) )h2 (w(ij) )
= sh3 (u(ij) )σ(u(ij) )h2 (v(ij) )
= sh3 (wij )σ(wij )h2 (wij )
= h3 s(wij )σ(wij )h2 (wij ) = h3 (wij ).
On the other hand, if Xm ∈ {Ak : k ≥ 4} ∪ {Bk : k ≥ 3} then by items (1) and
(2) of Lemma 2.6 we obtain
i
h
btm (w(ij) ) = b tm s(u(ij) )σ(u(ij) )ε(v(ij) )tm−1 e(v(ij) )
h
i
= b stm (u(ij) )σ(u(ij) )ε(v(ij) )etm−1 (v(ij) )
= b stm (wij )σ(wij )ε(wij )etm−1 (wij )
= b tm s(wij )σ(wij )ε(wij )tm−1 e(wij ) = btm (wij ).
But then Lemma 2.6(3) implies that tm (w(ij) ) = tm (wij ), i.e. that Xm (and thus
Tm ) satisfies w(ij) ≈ wij .
The proof that Yr satisfies w(ij) ≈ wij is similar, albeit with slight differences.
Let now tr be the word function corresponding to Yr in the sense of Theorem
2.5. The fact that Xm (and so Yr ) satisfies u(ij) ≈ wij implies tr (u(ij) ) = tr (wij ).
Furthermore, [13, Lemma 3.5] tells us that tr−1 (u(ij) ) = tr−1 (wij ) holds as well,
provided r ≥ 3. In turn, Yr satisfies v(ij) ≈ wij , thus tr (v(ij) ) = tr (wij ).
Now we have three subcases to consider. First, let Yr = B2 (which can happen,
just as the next subcase, only if Xm = A3 ). Since c(u) = c(v) = Xn , we have
i2 (w(ij) ) = i2 (u(ij) v(ij) ) = i2 (v(ij) ). On the other hand, we have already concluded
FINITE BANDS ARE FINITELY RELATED
13
that Yr satisfies v(ij) ≈ wij , so i2 (v(ij) ) = i2 (wij ) yielding i2 (w(ij) ) = i2 (wij ), as
required. Our second subcase is Yr = A3 , when h2 (w(ij) ) = h2 (u(ij) ) because of
s(w) = s(u). By items (1) and (4) of Lemma 2.6 we have:
h3 (w(ij) ) = h2 (w(ij) )ε(v(ij) )h3 e(v(ij) )
= h2 (u(ij) )ε(v(ij) )eh3 (v(ij) )
= h2 (wij )ε(wij )eh3 (wij )
= h2 (wij )ε(wij )h3 e(wij ) = h3 (wij ).
Finally, let Yr ∈ {Ak : k ≥ 4} ∪ {Bk : k ≥ 3}. Then, by invoking parts (1) and
(2) of Lemma 2.6 once again, we deduce
h
i
btr (w(ij) ) = b tr−1 s(u(ij) )σ(u(ij) )ε(v(ij) )tr e(v(ij) )
i
h
= b str−1 (u(ij) )σ(u(ij) )ε(v(ij) )etr (v(ij) )
= b str−1 (wij )σ(wij )ε(wij )etr (wij )
= b tr−1 s(wij )σ(wij )ε(wij )tr e(wij ) = btr (wij ),
whence Lemma 2.6(3) implies that tr (w(ij) ) = tr (wij ).
Hence, both Xm and Yr (with restrictions on m, r as described at the beginning
of the proof) satisfy the identity w(ij) ≈ wij . As remarked earlier, this completes
the proof that f is induced by w and confirms the theorem.
The proof of Theorem 1.1 is now immediate: if S is a finite band then it generates
the same variety as one of the bands covered by [17, Theorem 6.2], Theorem 3.1
and Theorem 3.2. By [9, Theorem 2.11], S is finitely related.
Acknowledgements. The author is grateful to a thorough and perceptive referee
for a handful of comments which significantly improved the paper.
References
[1] E. Aichinger, ‘Constantive Mal’cev clones on finite sets are finitely related’, Proc. Amer.
Math. Soc. 138 (2010), 3501–3507.
[2] E. Aichinger, P. Mayr, and R. McKenzie, ‘On the number of finite algebraic structures’, J.
Eur. Math. Soc. 16 (2014), 1673–1686.
[3] L. Barto, ‘Finitely related algebras in congruence distributive varieties have near unanimity
terms’, Canad. J. Math. 65 (2013), 3–21.
[4] L. Barto, ‘Finitely related algebras in congruence modular varieties have few subpowers’, J.
Eur. Math. Soc. (to appear). Available online at:
http://www.karlin.mff.cuni.cz/∼barto/Articles/ValerioteConjecture.pdf
[5] J. Berman, P. Idziak, P. Marković, R. McKenzie, M. Valeriote, and R. Willard, ‘Varieties
with few subalgebras of powers’ Trans. Amer. Math. Soc. 362 (2010), 1445–1473.
[6] A. P. Biryukov, ‘Varieties of idempotent semigroups’, Algebra i Logika 9 (1970), 255–273 (in
Russian).
[7] A. Bulatov, P. Jeavons, and A. Krokhin, ‘Classifying the complexity of constraints using finite
algebras’, SIAM J. Comput. 34 (2005), 720–742.
[8] S. Burris and H. P. Sankappanavar, A Course in Universal Algebra, Graduate Texts in Mathematics (Springer-Verlag, New York, 1981).
[9] B. A. Davey, M. G. Jackson, J. G. Pitkethly, and Cs. Szabó, ‘Finite degree: algebras in general
and semigroups in particular’, Semigroup Forum 83 (2011), 89–110.
[10] C. F. Fennemore, ‘All varieties of bands. I, II’, Math. Nachr. 48 (1971), 237–252; ibid. 253–
262.
14
IGOR DOLINKA
[11] J. A. Gerhard, ‘The lattice of equational classes of idempotent semigroups’, J. Algebra 15
(1970), 195–224.
[12] J. A. Gerhard and M. Petrich, ‘Free bands and free ∗ -bands’, Glasgow Math. J. 28 (1986),
161–179.
[13] J. A. Gerhard and M. Petrich, ‘Varieties of bands revisited’, Proc. London Math. Soc. (3) 58
(1989), 323–350.
[14] S. V. Jablonskiı̆, ‘The structure of the upper neighbourhood for predicately describable classes
in Pk ’, Dokl. Akad. Nauk SSSR 218 (1974), 304–307 (Russian). English traslation: Sov. Math.
Dokl. 15 (1974), 1353–1356.
[15] M. Jackson (ed.), ‘General Algebra and Its Applications 2013: Problem Session’, Algebra
Univers. 74 (2015), 9–16.
[16] P. Marković, M. Maróti, and R. McKenzie, ‘Finitely related clones and algebras with cube
terms’, Order 29 (2012), 345–359.
[17] P. Mayr, ‘On finitely related semigroups’, Semigroup Forum 86 (2013), 613–633.
[18] B. A. Romov, ‘Local characterizations of Post algebras, I’, Kibernetika 5 (1976), 38–45.
[19] I. G. Rosenberg, Á. Szendrei, ‘Degrees of clones and relations’, Houston J. Math. 9 (1983),
545–580.
[20] O. Sapir, ‘The variety of idempotent semigroups is inherently non-finitely generated’ Semigroup Forum 71 (2005), 140–146.
Department of Mathematics and Informatics, University of Novi Sad, Trg Dositeja
Obradovića 4, 21101 Novi Sad, Serbia
E-mail address: [email protected]
| 4math.GR
|
arXiv:1210.2885v1 [math.CO] 10 Oct 2012
A rank computation problem related to the
HK function of trinomial hypersurfaces
Shyamashree Upadhyay
Department of Mathematics
Indian Institute of Technology, Guwahati
Assam-781039, INDIA
email:[email protected]
Abstract
In this article, I provide a solution to a rank computation problem related to the computation of the Hilbert-Kunz function for any
disjoint-term trinomial hypersurface, over any field of characteristic 2.
This rank computation problem was posed in [3]. The formula for the
Hilbert-Kunz function for disjoint-term trinomial hypersurfaces can
be deduced from the result(s) of this article, modulo a lot of tedious
notation.
Contents
1 Introduction
2
2 Stating the problem in a general way
2.1 Statement of Problem I . . . . . . . . . . . . . . . . . . . . . .
2.2 Statement of Problem II . . . . . . . . . . . . . . . . . . . . .
3
3
3
3 Structure of the string of Binomial coefficients
5
4 Solutions to Problems I and II
6
4.1 Solution to Problem I . . . . . . . . . . . . . . . . . . . . . . . 6
4.2 Solution to Problem II . . . . . . . . . . . . . . . . . . . . . . 11
1
5 Concluding remarks about rationality
1
13
Introduction
A ‘disjoint-term trinomial hypersurface’ is defined in §2 of [3]. In the last
section of [3], I gave an algorithm for computing the Hilbert-Kunz function
for any disjoint-term trinomial hypersurface in general, over any field of arbitrary positive characteristic. But the algorithm was given modulo a rank
computation problem for 3 types of systems of linear equations which were
given in tables 5, 6 and 8 of [3]. In this article, I provide a solution to this
rank computation problem over fields of characteristic 2. It will be nice if one
can provide a similar solution over fields of arbitrary positive characteristic
p.
The formula for the Hilbert-Kunz function for disjoint-term trinomial
hypersurfaces follows as a corollary to the solution of this rank computation
problem, but a lot of tedious notation is needed for deducing the formula.
Hence in this article, I am avoiding the work on the deduction of the HilbertKunz function from this.
In the article [3], I was suspecting that due to the weird combinatorial
pattern of the matrices in tables 5, 6 and 8, there can be examples of disjointterm trinomial hypersurfaces for which the corresponding Hilbert-Kunz multiplicity can be irrational. But my suspicion turns out to be incorrect (see
the papers [2] and [1] for a reasoning). However if we consider trinomial hypersurfaces which are defined by a polynomial not having disjoint terms in it,
then we will encounter a similar-looking (but larger and more complicated)
rank computation problem for those. And I suspect that the Hilbert-Kunz
multiplicity may become irrational for those hypersurfaces because of the
more complicated nature of the system of linear equations involved in it. In
fact, the solution to the rank computation problem given in this article is
also important for the next article in which I will consider trinomial hypersurfaces which are defined by a polynomial not having disjoint terms in it,
because something similar happens there.
2
2
Stating the problem in a general way
The following combinatorial fact is well known (called the Chu-Vandermonde
identity):
For any two natural numbers m and n and for any non-negative integer j,
the standard inner product of the two vectors ( m C0 , m C1 , . . . , m Cj ) and
( n Cj , n Cj−1 , . . . , n C0 ) equals m+n Cj where we declare that for any
positive integer M, M Cl = 0 if l > M.
It follows from the above fact and from the pattern of the matrices present
in tables 5, 6 and 8 of the article [3] that
• for the linear systems corresponding to the matrices in tables 5 and 8,
we need to solve the general combinatorial problem mentioned below
as Problem I and
• for the linear systems corresponding to the matrices in tables 6, we
need to solve the general combinatorial problem mentioned below as
Problem II.
It is worthwhile to mention here that in this article, we will be considering
systems of linear equations over fields of characteristic 2 only.
2.1
Statement of Problem I
Given any positive integer M and any non-negative integer j, what all values
of positive integers k and l solve the following system of (l + 1) many linear
equations in (k + 1) many unknowns given by AX = b where the coefficient
matrix A is the (l + 1) × (k + 1) matrix given below in table 1 and b is
the (l + 1) × (1) column vector (0, 0, . . . , 1)t . The lines along the southwest↔north-east direction (in table 1) indicate the continuation of the same
entry along that direction.
2.2
Statement of Problem II
Given any positive integer M and non-negative integer j, what all values of
positive integers k, l and q solve the following system of (l+q +1) many linear
equations in (k + 1) many unknowns given by AX = b where the coefficient
3
Table 1: The coefficient matrix A of Problem I
M
Cj+l+k
Cj+l+(k−1) · · · · · · · · · M Cj+l+1
M
Cj+l+(k−1)
M
M
M
Cj+l
Cj+(l−1)
..
.
M
Cj+1
Cj
M
matrix A is the (l + q + 1) × (k + 1) matrix given below in table 2 and b is
the (l + q + 1) × (1) column vector (0, 0, . . . , 1)t . In table 2,
Table 2: The coefficient matrix A of Problem II
♥1q
··· ··· ···
♥qk−1
♥kq
..
..
..
..
..
..
.
.
.
.
.
.
k−1
0
1
♥2
♥2
··· ··· ···
♥2
♥k2
♥01
♥11
··· ··· ···
♥1k−1
♥k1
M
M
M
M
Cj+l+k
Cj+l+(k−1) · · · · · · · · ·
Cj+l+1
Cj+l
M
M
Cj+l+(k−1)
Cj+(l−1)
..
.
♥0q
..
.
M
Cj+1
Cj
M
M = α + δ for some natural numbers α and δ (given to us).
For any 1 ≤ i ≤ q and 0 ≤ r ≤ k, ♥ri := the standard inner product
of the two vectors ( α C0 , α C1 , . . . , α Cj+k+l−r )
and ( δ Cj+k+l+i−r , . . . , δ Ci+1 , δ Ci ).
The lines along the south-west↔north-east direction indicate the
continuation of the same entry along that direction.
4
3
Structure of the string of Binomial coefficients
For providing solutions to both problems I and II, we first need to have an
account of the position of all odd entries in the string of binomial coefficients
of any given positive integer M starting from M C0 to M CM . This is given
as follows:
Let M = 2lm + 2lm−1 + · · · + 2l1 be the binary expansion of M where
lm > lm−1 > · · · > l1 are non-negative integers.
Let DM := the set of all possible (m + 1)-tuples (α0 , α1 , . . . , αm ) generated
out of the set {0, 2l1 , 2l2 , . . . , 2lm } which satisfy the following 3 properties
simultaneously:
• α0 = 1.
• α1 ≥ α2 ≥ · · · ≥ αm .
• If αi 6= 0 for some i ∈ {1, . . . , m}, then αi > αj for any j ∈ {i +
1, . . . , m}.
Clearly, the total number of elements in the set DM is 2m . Let us put an
order - on DM by declaring that
(α0 , α1 , . . . , αm ) - (α0 , α1 , . . . , αm )
if and only if
αi ≤ αi ∀i ∈ {0, 1, . . . , m}.
Observe that - is a total order on the set DM . The positions of the odd
entries in the string of binomial coefficients of M are indexed by elements of
the set DM ; and are arranged (in ascending order) according to the order on DM . This indexing is given by:
For any d = (α0 , α1 , . . . , αm ) ∈ DM , the position of the odd entry
corresponding to it in the string of binomial coefficients of M is Σm
i=0 αi .
m
Since the total number of elements in the set DM is 2 , we can denote the
(M )
(M )
(M )
(M ) (M )
(M )
elements of DM by d1 , d2 , . . . , d2m where d1 - d2 - . . . - d2m .
And let us denote the corresponding positions of the odd entries in the
string of binomial coefficients of M by
(M )
(M )
(M )
Σd1 , Σd2 , . . . , Σd2m .
5
Looking at the positions of all odd entries in the string of binomial coefficients
of the given positive integer M, we can deduce some important facts about
the position of all even entries in the string of binomial coefficients of M.
These facts will be very much useful in writing down solutions to both the
Problems I and II. These facts are as given below:
Let the binary expansion of M be as mentioned above. Then
1. the number of even entries between any two consecutive odd entries
in the continuous string of binomial coefficients of M is of the form
2lg − 2lg−1 − . . . − 2l1 − 1 for some g ∈ {1, . . . , m}.
2. If lg < lĝ , then a bunch of 2lg − 2lg−1 − . . . − 2l1 − 1 many even entries
appears on both right and left sides of the bunch of 2lĝ − 2lĝ−1 − . . . −
2l1 − 1 many even entries at equal distance from the central bunch.
3. Between any two consecutive bunches of 2lg − 2lg−1 − . . . − 2l1 − 1 many
even entries, there exists (2g − 1) many bunches of even entries such
that the number of even entries in each of these (2g − 1) many bunches
is different from 2lg − 2lg−1 − . . . − 2l1 − 1.
4
Solutions to Problems I and II
In subsection 4.1 below, the notation and terminology will remain the same
as in subsection 2.1. Similarly, in subsection 4.2 below, the notation and
terminology will remain the same as in subsection 2.2.
4.1
Solution to Problem I
Problem I can be restated in the following way:—
Given any positive integer M and any non-negative integer j, for what all
values of positive integers k and l does the string
M
Cj ,
M
Cj+1 , . . . , . . . ,
M
Cj+l+k
of consecutive binomial coefficients of M satisfy all the following 3
properties simultaneously:
•
M
Cj and
is even).
M
Cj+k+1 have different parity (that is, exactly one of them
6
•
M
Cj+t = M C(j+t)+q(k+1) for every t ∈ {1, . . . , k + 1} and for every
positive integer q such that (j + t) + q(k + 1) ≤ j + l + k. In other
words, the string of length k + 1 beginning from M Cj+1 and ending at
M
Cj+k+1 should be repeated until we reach M Cj+l+k (there can be a
truncated string of length k + 1 at the end).
• The string of length k + 1 beginning from M Cj+1 and ending at
M
Cj+k+1 should contain evenly many odd elements in it.
There are many possible values of positive integers k and l which do the job.
A list of all possible answers is given below in cases I and II.
Case I: When
M
Cj is odd.
(M )
Suppose that M Cj equals Σdi for some i ∈ {1, . . . , 2m }, then we provide
a solution to Problem I by considering different sub cases depending upon
M
M
M
the value ΣdM
i+1 − Σdi − 1 being 0 or > 0. Observe that Σdi+1 − Σdi − 1is
nothing but the total number of zeroes lying in between the positions ΣM
i+1
and ΣM
.
i
(M )
(M )
Sub case I.a: When Σdi+1 − Σdi − 1 = 0.
(M )
(M )
If there exists an even natural number s such that Σdi+s+1 − Σdi+s − 1 > 0,
then take the string M Cj+1, . . . , M Cj+k+1 of length (k + 1) to be the continuous string of binomial coefficients of M starting from position number
(M )
(M )
1 + Σdi
and ending at position number Σdi+s + L (inclusive of the beginning and end points), where L is some natural number such that 0 <
(M )
(M )
L ≤ Σdi+s+1 − Σdi+s − 1. Hence the natural number k is given by k =
(M )
(M )
Σdi+s + L − Σdi − 1 where L and s are as mentioned above.
To determine the possible values of l in this sub case, look at the entry in
(M )
position number Σdi+s +L+1. If it is even, then take l = 1. If it is odd, then
look at the number of odd entries in the string of binomial coefficients of M
(M )
that come in continuum starting from position number 1 + Σdi (including
(M )
the odd entry at this position). Let pj denote this number. Similarly, look
at the number of odd entries in the string of binomial coefficients of M that
(M )
come in continuum starting from position number 1 + L + Σdi+s (including
(M )
the odd entry at this position). Let qj denote this number. Tale l to be
(M ) (M )
any natural number such that l ≤ 1 + min{pj , qj }.
7
Remark 4.1.1. An even natural number s as mentioned above in this sub
(M )
case may not exist for some M and for some position number Σdi . In that
situation, this sub case is invalid.
(M )
(M )
Sub case I.b: When Σdi+1 − Σdi − 1 > 0.
I will now list down the various possibilities under this sub case:
Either
(M )
(M )
Take k to be any natural number such that k + 1 ≤ Σdi+1 − Σdi − 1.
And take the string M Cj+1 , . . . , M Cj+k+1 of length (k + 1) to be the
continuous string of binomial coefficients of M starting from position
(M )
(M )
number 1 + Σdi and ending at position number 1 + k + Σdi . And take
(M )
(M )
l to be any natural number such that l ≤ Σdi+1 − Σdi − (k + 1).
Or
For computing the possible values of the natural number k, proceed
similarly as in sub case (I.a). And to determine the possible values of l,
(M )
look at the entry in position number Σdi+s + L + 1. If it is odd, then take
l = 1. If it is even, then look at the number of even entries in the string of
binomial coefficients of M that come in continuum starting from position
(M )
(M )
number 1 + Σdi (including the even entry at this position). Let pj
denote this number. Similarly, look at the number of even entries in the
string of binomial coefficients of M that come in continuum starting from
(M )
position number 1 + L + Σdi+s (including the even entry at this position).
(M )
Let qj denote this number. Tale l to be any natural number such that
(M ) (M )
l ≤ 1 + min{pj , qj }.
Case II: When M Cj is even.
The position of M Cj in the string of binomial coefficients of M (starting
from M C0 to M CM , that is, from left to right) is (j + 1)-th.
(M )
Sub case II.a: When j + 2 = Σdi for some i ∈ {1, . . . , 2m } and there exists
an odd natural number u such that all the entries at the position numbers
(M )
(M )
(M )
Σdi , 1 + Σdi , . . . , u + Σdi are odd and the entry at position number
(M )
u + 1 + Σdi is even.
I will now list down the various possibilities under this sub case:
Either
8
Take k to be any odd natural number such that k + 1 ≤ u + 1. And take
the string M Cj+1 , . . . , M Cj+k+1 of length (k + 1) to be the continuous
(M )
string of binomial coefficients of M starting from position number Σdi
(M )
and ending at position number k + Σdi . Take l to be any natural number
such that l ≤ (u + 1) − k.
Or
Take the string Cj+1 , . . . , Cj+k+1 of length (k + 1) to be the continuous
string of binomial coefficients of M starting from position number
(M )
(M )
j + 2(= Σdi ) and ending at position number Σdi+s , where s is any odd
natural number such that i + s ≤ 2m . [Remark: Such a odd natural number
s may not exist always. If it does not exist, then this is an invalid
possibility.] And to determine the possible values of l, look at the entry in
(M )
position number Σdi+s + 1. If it is even, then take l = 1. If it is odd, then
look at the number of odd entries in the string of binomial coefficients of M
(M )
that come in continuum starting from position number Σdi (including
(M )
the odd entry at this position). Let pj denote this number. Similarly,
look at the number of odd entries in the string of binomial coefficients of M
(M )
that come in continuum starting from position number 1 + Σdi+s (including
(M )
the odd entry at this position). Let qj denote this number. Tale l to be
(M ) (M )
any natural number such that l ≤ 1 + min{pj , qj }.
M
M
(M )
(M )
Sub case II.b: When j + 2 = Σdi for some i ∈ {1, . . . , 2m } and Σdi+1 −
(M )
Σdi − 1 > 0.
Take the string M Cj+1 , . . . , M Cj+k+1 of length (k + 1) to be the continuous
string of binomial coefficients of M starting from position number j + 2(=
(M )
(M )
Σdi ) and ending at position number Σdi+s , where s is any odd natural
number such that i + s ≤ 2m . [Remark: Such a odd natural number s may
not exist always. If it does not exist, then this is an invalid possibility.] In
this sub case, take l = 1.
(M )
Sub case II.c: When j + 2 6= Σdi for any i ∈ {1, . . . , 2m }.
(M )
Clearly there exists i ∈ {1, . . . , 2m } such that j + 2 < Σdi . Look at the
smallest such i, call it i0 . Take the string M Cj+1, . . . , M Cj+k+1 of length
(k + 1) to be the continuous string of binomial coefficients of M starting
(M )
from position number j + 2 and ending at position number Σdi0 +s , where s
is any odd natural number such that i0 + s ≤ 2m . [Remark: Such a odd
9
natural number s may not exist always. If it does not exist, then this is an
invalid possibility.]
And to determine the possible values of l, look at the entry in position
(M )
number 1 + Σdi0 +s . If it is odd, then take l = 1. If it is even, then look at
the number of even entries in the string of binomial coefficients of M that
come in continuum starting from position number j + 2 (including the even
(M )
entry at this position). Let pj denote this number. Similarly, look at the
number of even entries in the string of binomial coefficients of M that come
(M )
in continuum starting from position number 1 + Σdi0 +s (including the even
(M )
entry at this position). Let qj denote this number. Take l to be any
(M ) (M )
natural number such that l ≤ 1 + min{pj , qj }.
There is another non-trivial possibility under this sub case, which is the following:
Let M = 2Nr + 2Nr−1 + · · · + 2N1 be the binary expansion of M where
Nr > Nr−1 > · · · > N1 are non-negative integers. Let Ns0 be the least
element of the set {Nr , Nr−1 , · · · , N1 } which is ≥ 2. Let
{Nst , Nst−1 , · · · , Ns1 } be the subset of the set {Nr , Nr−1 , · · · , N1 } defined by
{Nb |1 ≤ b ≤ r − 1 and N1+b − Nb > 1}. It may happen in some cases that
Ns0 = Ns1 .
Given any d ∈ {1, . . . , t}, let Ud denote the set of all possible sums
generated from elements of the set {2Nr , 2Nr−1 , . . . , 2N1+sd } (where each
element of this set can appear atmost once in any such sum) which are
strictly less than 2Nr + 2Nr−1 + · · · + 2N1+sd . Given any d ∈ {1, . . . , t} and
any x(d) ∈ Ud , suppose M Cj is at the (1 + x(d) + 21+Nsd )-th position in the
string of binomial coefficients of M counting from the end (that is, from the
right to left). In this situation, the following solutions are possible:
(i) Take k + 1 to be equal to 2Nsd . And to determine the possible values of
l, look at the number of even entries in the string of binomial coefficients of
M that come in continuum after M Cj (i.e., excluding M Cj and to the right
(M )
of it). Let pj denote this number. Similarly, look at the number of even
entries in the string of binomial coefficients of M that come in continuum
after M Cj+2(1+Nsd ) (i.e., excluding M Cj+2(1+Nsd ) and to the right of it). Let
(M )
denote this number. Tale l to be any natural number such that
(M ) (M )
l ≤ 1 + 2Nsd + min{pj , qj }.
(ii) If d > 1 and there exists integer(s) z such that Nsd > z > Nsd−1 and
z ∈ {Nr , Nr−1 , · · · > N1 }, then for any such z, take k + 1 to be equal to 2z .
qj
10
And to determine the possible values of l, look at the number of even
entries in the string of binomial coefficients of M that come in continuum
(M )
after M Cj (i.e., excluding M Cj and to the right of it). Let pj denote this
number. Similarly, look at the number of even entries in the string of
binomial coefficients of M that come in continuum after M Cj+2(1+Nsd ) (i.e.,
(M )
excluding M Cj+2(1+Nsd ) and to the right of it). Let qj denote this
number. Tale l to be any natural number such that
(M ) (M )
l ≤ 1 + 2z (2(Nsd −z+1) − 1) + min{pj , qj }.
(iii) If d = 1 and Ns1 > Ns0 , then there always exists integer(s) z such that
Ns1 > z ≥ Ns0 and z ∈ {Nr , Nr−1 , · · · > N1 }. For any such z, do similarly
as in (ii) above replacing d there by 1.
4.2
Solution to Problem II
Problem II can be restated in the following way:—
Given any two positive integers α and δ and any non-negative integer j, for
what all values of positive integers k, l and q does the string
M
Cj ,
M
Cj+1 , . . . , . . . ,
M
Cj+l+k
(where M = α + δ) of consecutive binomial coefficients of M becomes a
solution to Problem I as mentioned above and the following properties are
also satisfied (simultaneously):
•
•
•
M
Cj+k+l+1 + M Cj+k+l + · · · + M Cj+l+1 and
· · · + α Cj+l+1 should have the same parity.
α
Cj+k+l+1 +
α
Cj+k+l +
M
Cj+k+l+2 + M Cj+k+l+1 +· · ·+ M Cj+l+2 and α Cj+k+l+2 + α Cj+k+l+1 +
· · · + α Cj+l+2 differ in parity if and only if δ C1 ( α Cj+k+l+1 + α Cj+k+l +
· · · + α Cj+l+1) is odd.
M
Cj+k+l+3 + M Cj+k+l+2 +· · ·+ M Cj+l+3 and α Cj+k+l+3 + α Cj+k+l+2 +
· · ·+ α Cj+l+3 differ in parity if and only if δ C1 ( α Cj+k+l+2 + α Cj+k+l+1 +
· · · + α Cj+l+2) + δ C2 ( α Cj+k+l+1 + α Cj+k+l + · · · + α Cj+l+1) is odd.
• and so on till · · · · · ·
•
M
Cj+k+l+q + M Cj+k+l+q−1+· · ·+ M Cj+l+q and α Cj+k+l+q + α Cj+k+l+q−1+
q−1 δ
Cp ( α Cj+k+l+q−p +
· · · + α Cj+l+q differ in parity if and only if Σp=1
α
α
Cj+k+l+q−1−p + · · · + Cj+l+q−p ) is odd.
11
Given any positive integers α, δ and non-negative integer j, we can proceed
similarly as in subsection 4.1 and find the possible values of the positive
integers k and l for which the string
M
Cj ,
M
Cj+1 , . . . , . . . ,
M
Cj+l+k
of consecutive binomial coefficients of M becomes a solution to Problem I.
But these values of positive integers k and l should also satisfy the additional
q-many properties (about parities of sums of strings of length (k + 1) of the
binomial coefficients of M and α) as mentioned above.
For any two natural numbers Z and y, let sy (Z) denote the sum of the
first y-many (starting from Z C0 ) binomial coefficients of Z. It is an easy
exercise to check that for any such Z and y, sy (Z) = Z−1 Cy−1 . Since we
are working over field of characteristic 2, it is easy to see that M Cj+k+l+1 +
M
Cj+k+l + · · · + M Cj+l+1 equals sj+k+l+1(M) + sj+l (M) which in turn equals
M −1
Cj+k+l + M −1 Cj+l−1 . Similarly one can say that α Cj+k+l+1 + α Cj+k+l +
· · · + α Cj+l+1 equals α−1 Cj+k+l + α−1 Cj+l−1. Therefore the condition that
M
Cj+k+l+1 + M Cj+k+l +· · ·+ M Cj+l+1 and α Cj+k+l+1 + α Cj+k+l +· · ·+ α Cj+l+1
should have the same parity translates into the condition that
α+δ−1
Cj+k+l + α+δ−1 Cj+l−1 and α−1 Cj+k+l +
have the same parity. · · · · · · · · · condition (∗)
α−1
Cj+l−1 should
The other (q − 1)-many conditions (about parity of sums of strings of length
(k + 1)) imply that the positive integer q should be such that for every
a ∈ {2, . . . , q}, the sums M Cj+k+l+a + M Cj+k+l+a−1 + · · · + M Cj+l+a and
α
Cj+k+l+a + α Cj+k+l+a−1 + · · · + α Cj+l+a differ in parity if and only if the
following condition holds:
If p1 < · · · < pf is the collection of all elements of the set
{p|p ∈ {1, . . . , a − 1} and α Cj+k+l+a−p + α Cj+k+l+a−1−p + · · · +
α
Cj+l+a−p is odd}, then exactly odd many elements of the set
{ δ Cp1 , . . . , δ Cpf } should be odd. And this should hold true for
every a ∈ {2, . . . , q}.· · · · · · · · · condition (∗∗)
We therefore need to determine when the value of the string sum M Cj+k+l+a+
M
Cj+k+l + · · · + M Cj+l+a becomes odd and when it is even as the integer a
ranges over the set {1, . . . , q}. And similarly for the string sums α Cj+k+l+a +
α
Cj+k+l+a−1 + · · · + α Cj+l+a as the integer a ranges over the set {1, . . . , q}.
But this problem is similar to solving problem I (or a problem equivalent to
12
problem I where the column vector b is replaced by a suitable vector which is
either (1, 0, . . . , 0)t or (0, 1, . . . , 1)t ) for both the integers M and α, assuming
(to begin with) that the string sums M Cj+k+l+1 + M Cj+k+l + · · · + M Cj+l+1
and α Cj+k+l+1 + α Cj+k+l + · · · + α Cj+l+1 have the same parity. Recall that
while solving problem I considering various sub cases, we got solutions for
all possible values of the integer l. Speaking more precisely, we got all possible values of ‘upper bounds’ of the natural number l. These upper bounds
are nothing but the number of maximum possible strings (in continuum) of
length (k + 1) for which the parity of the string sum remains the same. That
means, as soon as the value of l exceeds this ‘upper bound’, the string sum
changes parity. Hence the positive integer q should be such that the various
upper bounds of the values of l under consideration (for α and M both)
should tally with conditions (∗) and (∗∗) mentioned above.
5
Concluding remarks about rationality
Looking at the rhythm of the parity of sums of continuous strings of binomial
coefficients of any given positive integer M (as described in subsections 4.1
and 4.2 above), one can depict that over fields of characteristic 2, the HilbertKunz multiplicity in this case of ‘disjoint-term trinomial hypersurfaces’ will
turn out to be rational. In fact, it will depend upon the positions of the odd
and even entries in the entire string of binomial coefficients of M, a precise
account of which is given in section 3. I hope similar thing will hold true
over fields of arbitrary positive characteristic p.
For any hypersurface defined by polynomials having ‘disjoint terms’ in
it (not just trinomial hypersurfaces), it is known that the corresponding
Hilbert-Kunz multiplicity will be rational [see [2] and [1] for a reasoning].
The ‘phis in the papers of [2] and [1] attached to ‘disjoint-term trinomials
(or more generally to any sum of monomials that are pairwise prime) are ‘pfractals. As a consequence, the Hilbert-Kunz series will be a rational function
and the corresponding HK multiplicity will be in Q.
But for trinomial hypersurfaces NOT having ‘disjoint-terms’ in it, the
situation becomes more interesting. There we need to solve ‘similar’ rank
computation problems related to larger and more complicated systems of
linear equations [in fact, there we can have infinitely many linear equations
inside a single system]. But the solution to the rank computation problem
mentioned in this article is like providing a basement for the work in the
13
more general case. Due to the more complicated nature of the systems in
that case, I suspect that the Hilbert-Kunz multiplicity can become irrational
there.
References
[1] Paul Monsky, Pedro Teixeira, p-Fractals and power series-I. Some 2
variable results, Journal of Algebra, 280 (2004), 505–536.
[2] Paul Monsky, Pedro Teixeira, p-Fractals and power series-II. Some applications to Hilbert-Kunz theory, Journal of Algebra, 304 (2006), 237–
255.
[3] Shyamashree Upadhyay, An algorithm for the HK function for disjointterm trinomial hypersurfaces, arXiv:1204.5417, [math.CO]
14
| 0math.AC
|
Hypotheses tests in boundary regression
models
Holger Drees, Natalie Neumeyer and Leonie Selk∗
arXiv:1408.3979v2 [stat.ME] 11 Oct 2016
University of Hamburg, Department of Mathematics
Bundesstrasse 55, 20146 Hamburg, Germany
October 12, 2016
Abstract
Consider a nonparametric regression model with one-sided errors and regression
function in a general Hölder class. We estimate the regression function via minimization of the local integral of a polynomial approximation. We show uniform rates of
convergence for the simple regression estimator as well as for a smooth version. These
rates carry over to mean regression models with a symmetric and bounded error distribution. In such a setting, one obtains faster rates for irregular error distributions
concentrating sufficient mass near the endpoints than for the usual regular distribu√
tions. The results are applied to prove asymptotic n-equivalence of a residual-based
(sequential) empirical distribution function to the (sequential) empirical distribution
function of unobserved errors in the case of irregular error distributions. This result is
remarkably different from corresponding results in mean regression with regular errors.
It can readily be applied to develop goodness-of-fit tests for the error distribution. We
present some examples and investigate the small sample performance in a simulation
study. We further discuss asymptotically distribution-free hypotheses tests for independence of the error distribution from the points of measurement and for monotonicity
of the boundary function as well.
AMS 2010 Classification: Primary 62G08; Secondary 62G10, 62G30, 62G32
Keywords and Phrases: goodness-of-fit testing, irregular error distribution, one-sided
errors, residual empirical distribution function, uniform rates of convergence
1
Introduction
We consider boundary regression models of the form
Yi = g(xi ) + εi ,
1
i = 1, . . . , n,
with negative errors εi whose survival function 1−F (y) behaves like a multiple of |y|α for some
α > 0 near the origin. Such models naturally arise in image analysis, analysis of auctions
and records, or in extreme value analysis with covariates. For such a boundary regression
model with multivariate random covariates and twice differentiable regression function, Hall
and Van Keilegom (2009) establish a minimax rate for estimation of g(x) (for fixed x) under
quadratic loss and determine pointwise asymptotic distributions of an estimator which is
defined as a solution of a linear optimization problem (cf. Remark 2.6). Relatedly, Müller
and Wefelmeyer (2010) consider a mean regression model with (unknown) symmetric support
of the error distribution and Hölder continuous regression function. They discuss pointwise
MSE rates for estimators of the regression function that are defined as the average of local
maxima and local minima. Meister and Reiß (2013) consider a regression model with known
bounded support of the errors. They show asymptotic equivalence in the strong LeCam sense
to a continuous-time Poisson point process model when the error density has a jump at the
endpoint of its support. For a regression model with error distribution that is one-sided and
regularly varying at 0 with index α > 0, Jirak et al. (2014) suggest an estimator for the
boundary regression function which adapts simultaneously to the unknown smoothness of
the regression function and to the unknown extreme value index α. Reiß and Selk (2016+)
construct efficient and unbiased estimators of linear functionals of the regression function in
the case of exponentially distributed errors as well as in the limiting Poisson point process
experiment by Meister and Reiß (2013).
Closely related to regression estimation in models with one-sided errors is the estimation
of a boundary function g based on a sample from (X, Y ) with support {(x, y) ∈ [0, 1]×[0, ∞] |
y ≤ g(x)}. For such models, Härdle et al. (1995) and Hall et al. (1998) proved minimax
rates both for g(x) and for the L1 -distance between g and its estimator. Moreover, they
showed that an approach using local polynomial approximations of g yields this optimal
rate. Explicit estimators in terms of higher order moments were proposed and analyzed
by Girard and Jacob (2008) and Girard et al. (2013). Daouia et al. (2016) consider spline
estimation of a support frontier curve and obtain uniform rates of convergence.
The aim of the paper is to develop tests for model assumptions in boundary regression
models. In particular we will suggest asymptotically distribution-free tests for
• parametric classes of error distributions (goodness-of-fit)
• independence of the error distribution from the points of measurement
• monotonicity of the boundary function.
The test statistics are based on (sequential) empirical processes of residuals. To investigate
these, we need uniform rates of convergence for the regression estimator, which are of interest
on its own. To our knowledge, uniform rates so far have only been shown by Daouia et al.
(2016) who do not obtain optimal rates. Our results can also be applied to mean regression
2
models with bounded symmetric error distribution. For regression functions g in a Hölder
class of order β, we obtain the rate ((log n)/n)β/(αβ+1) . Thus, for tail index α ∈ (0, 2) of the
error distribution, the rate is faster than the typical rate one has in mean regression models
with regular errors. For pointwise and Lp -rates of convergence, it has been known in the
literature that faster rates are possible for nonparametric regression estimation in models
with irregular error distribution, see e.g. Gijbels and Peng (2000), Hall and Van Keilegom
(2009), or Müller and Wefelmeyer (2010).
The uniform rate of convergence for the regression estimator enables us to derive asymptotic expansions for residual-based empirical distribution functions and to prove weak convergence of the residual-based (sequential) empirical distribution function. We state conditions
under which the influence of the regression estimation is negligible such that the same results
are obtained as in the case of observable errors. We apply the results to derive goodnessof-fit tests for parametric classes of error distributions. Asymptotic properties of residual
empirical distribution functions in mean regression models were investigated by Akritas and
Van Keilegom (2001), among others. As the regression estimation strongly influences the
asymptotic behavior of the empirical distribution function in these regular models, asymptotic distributions of goodness-of-fit test statistics are involved, and typically bootstrap is
applied to obtain critical values, see Neumeyer et al. (2006). In contrast, in the present
situation with an irregular error distribution, standard critical values can be used.
In nonparametric frontier models, Wilson (2003) discusses several possible tests for assumptions of independence, for instance independence between input levels and output inefficiency. Those assumptions are needed to prove validity of bootstrap procedures and are
thus crucial in applications, but they may be violated; see Simar and Wilson (1998). Wilson (2003) points out the analogy to tests for independence between errors and covariates
in regression models, but no asymptotic distributions are derived. Tests for independence
in nonparametric mean and quantile regression models that are similar to the test we will
consider are suggested by Einmahl and Van Keilegom (2008) and Birke et al. (2016+).
There is an extensive literature on regression with one-sided error distributions and similar models (in particular production frontier models) which assume monotonicity of the
boundary function, see Gijbels et al. (1999), the literature cited therein and the monotone
nonparametric maximum likelihood estimator in Reiß and Selk (2016+). Monotonicity of
a production frontier function in each component is given under the strong disposability
assumption, but may often not be fulfilled; see e.g. Färe and Grosskopf (1983). We are
not aware of hypothesis tests for monotonicity or other shape constraints in the context of
boundary regression, but would like to mention Gijbels’ (2005) review on testing for monotonicity in mean regression. Tests similar in spirit to the one we are suggesting here were
considered by Birke and Neumeyer (2013) and Birke et al. (2016+) for mean and quantile
regression models, respectively.
The remainder of the article is organized as follows. In Section 2 the regression model
3
under consideration is presented and model assumptions are formulated. The regression
estimator is defined and uniform rates of convergence are given. A smooth modification of
the estimator is considered and uniform rates of convergence for this estimator as well as its
derivative are shown. In Section 3 residual based empirical distribution functions based on
both regression estimators are investigated. Conditions are stated under which the influence
√
of regression estimation is asymptotically n-negligible. Furthermore, an expansion of the
residual empirical distribution function is shown that is valid under more general conditions.
In Section 4 goodness-of-fit tests for the error distribution are discussed in general and in
some detailed examples. We investigate the finite sample performance of the tests in a small
simulation study. We further discuss hypotheses tests for independence of the error distribution from the design points as well as a test for monotonicity of the boundary function.
All proofs are given in the appendix.
2
The regression function: uniform rates of convergence
We consider a regression model with fixed equidistant design and one-sided errors,
Yi = g( ni ) + εi ,
i = 1, . . . , n,
(2.1)
under the following assumptions:
(F1) The errors ε1 , . . . , εn are independent and identically distributed and supported on
(−∞, 0]. The error distribution function fulfills
F (y) = 1 − c|y|α + r(y),
y < 0,
for some α > 0, with r(y) = o(|y|α ) for y % 0.
(G1) The regression function g belongs to some Hölder class of order β ∈ (0, ∞), i. e. g is
bβc-times differentiable on [0, 1] and the bβc-th derivative satisfies
cg := sup
t,x∈[0,1]
t6=x
|g (bβc) (t) − g (bβc) (x)|
< ∞.
|t − x|β−bβc
In Figure 1 some scatter plots of data according to model (2.1) are shown for different
tail indices α of the error distribution.
Remark 2.1 Strictly speaking, we consider a triangular scheme in (2.1), and the errors εi
depend on n too, as the ith regression point i/n varies with n. For notational simplicity, we
suppress the second index, because the distribution of the errors does not depend on n.
4
0.0
0.2
0.4
0.6
x
0.8
1.0
0.2
0.4
0.6
0.8
1.0
0.2
0.0
g (x)
−0.6
−0.4
g (x)
−0.4
−0.6
0.0
0.0
x
−0.2
0.0
0.2
α=3
−0.2
0.0
−0.6
−0.4
g (x)
−0.2
0.0
−0.2
−0.6
−0.4
g (x)
α=2
0.2
α=1
0.2
α = 0.5
0.2
0.4
0.6
0.8
1.0
0.0
x
0.2
0.4
0.6
0.8
1.0
x
Figure 1: Scatter plots of ( ni , Yi ), i = 1, . . . , n, and the true regression function g(x) =
−3(x − 0.4)3 . The error distribution is Weibull F (y) = exp(−(|y|/θ)α )I(−∞,0) (y) + I[0,∞) (y)
with scale θ = 0.3 and shape parameter α.
We consider an estimator that locally approximates the regression function by a polynomial while lying above the data points. More specifically, for x ∈ [0, 1], let
ĝn (x) := ĝ(x) := p(x)
where p is a polynomial of order dβe − 1 and minimizes the local integral
Z x+hn
p(t) dt
(2.2)
x−hn
under the constraints p( nj ) ≥ Yj for all j ∈ {1, . . . , n} such that | nj − x| ≤ hn . For the
asymptotic analysis of this estimator, we need the following assumption:
(H1) Let (hn )n∈N be a sequence of positive bandwidths that satisfies limn→∞ hn = 0 and
limn→∞ nhn / log n = ∞.
We obtain the following uniform rates of convergence.
Theorem 2.2 In model (2.1), under the assumptions (F1), (G1), and (H1), we have
| log h | 1/α
n
sup |ĝ(x) − g(x)| = O(hβn ) + OP
.
nh
n
x∈[hn ,1−hn ]
Note that the deterministic part O(hβn ) arises from approximating the regression function
by a polynomial, whereas the random part originates from the observational error. Balancing
1
the two sources of error by setting hn ((log n)/n) αβ+1 gives
β
log n αβ+1
sup |ĝ(x) − g(x)| = OP
.
(2.3)
n
x∈[hn ,1−hn ]
(Here an bn means that 0 < lim inf n→∞ |an /bn | ≤ lim supn→∞ |an /bn | < ∞.)
This result is of particular interest in the case of irregular error distributions, i. e. α ∈
β
(0, 2), when the rate improves upon the typical optimal rate OP (((log n)/n) 2β+1 ) for estimating mean regression functions in models with regular errors.
5
Remark 2.3 Jirak et al. (2014) consider a similar boundary regression estimator while
P
replacing the integral in (2.2) by its Riemann approximation ni=1 p( ni )I{| ni − x| ≤ hn }. In
particular, they use the Lepski method to construct a data-driven bandwidth that satisfies
1
hn ((log n)/n) αβ+1 in probability. For this modified estimator, we obtain the same uniform
rate of convergence as in Theorem 2.2 by replacing Proposition A.1 in the proof of Theorem
2.2 by Theorem 3.1 in Jirak et al. (2014).
Remark 2.4 For Hölder continuous regression functions with exponent β ∈ (0, 1] the estimator reduces to a local maximum, i. e. ĝ(x) = max{Yi | i = 1, . . . , n s. t. | ni − x| ≤ hn }.
In this case we obtain the rate of convergence as given in Theorem 2.2 uniformly over the
whole unit interval.
Remark 2.5 Müller and Wefelmeyer (2010) consider a mean regression model Yi = m(Xi )+
ηi , i = 1, . . . , n, with symmetric error distribution supported on [−a, a] (with a unknown);
see the left panel of Figure 2. The error distribution function fulfills F (a − y) ∼ 1 − y α for
y & 0. The local empirical midrange of responses, i. e.
1
min Yi + max Yi
m̂(x) =
i∈{1,...,n}
i∈{1,...,n}
2 |X
|X −x|≤h
−x|≤h
n
i
n
i
is shown to have pointwise rate of convergence O(hβn ) + OP ((nhn )−1/α ) to m(x) if m is
Hölder continuous with exponent β ∈ (0, 1]. Theorem 2.2 enables us to extend Müller’s
and Wefelmeyer’s (2010) results in two ways (in a model with fixed design Xi = ni ): we
consider more general Hölder classes with general index β > 0, and we obtain uniform rates
ˆ
of convergence. To this end, we use the mean regression estimator m̂ = (ĝ − g̃)/2
with ĝ as
i
before and g̃ˆ defined analogously, but based on ( n , −Yi ), i = 1, . . . , n; see the right panel of
Figure 2. The rates obtained for supx∈[hn ,1−hn ] |m̂(x) − m(x)| are the same as in Theorem
2.2.
1.5
1.0
y
0.5
0.0
−1.0
−0.5
0.0
−0.5
−1.0
y
0.5
1.0
1.5
2.0
mean regression: boundary curves
2.0
mean regression with symmetric compactly supported error
0.0
0.2
0.4
0.6
0.8
1.0
0.0
x
0.2
0.4
0.6
x
Figure 2: Example for data as in Remark 2.5.
6
0.8
1.0
Remark 2.6 For β ∈ (1, 2], Hall and Van Keilegom (2009) consider the following local
linear boundary regression estimator:
n
o
2
i
i
ǧ(x) = inf α0 (α0 , α1 ) ∈ R : Yi ≤ α0 + α1 n − x ∀i ∈ {1, . . . , n} s. t. n − x ≤ hn .
(2.4)
R x+hh
Because of x−hn (α0 + α1 (t − x)) dt = 2α0 hn this estimator coincides with ĝ for β ∈ (1, 2].
However, in the case β > 2 replacing the linear function in (2.4) by a polynomial of order
dβe − 1 renders the estimator ǧ useless. One obtains ǧ(x) = −∞ for x 6∈ { nj | j = 1, . . . , n}
while ǧ( nj ) = Yj , j = 1, . . . , n. This was already observed by Jirak et al. (2014).
Note that typically the estimator ĝ is not continuous. One might prefer to consider
a smooth estimator by convoluting ĝ with a kernel. Such a modified estimator will also
be advantageous when deriving an expansion for the residual based empirical distribution
function in the next section. Therefore we define
Z 1−hn
x−z
1
dz
(2.5)
ĝ(z) K
g̃(x) =
bn
bn
hn
and formulate some additional assumptions.
R
(K1) K is a continuous kernel with support [−1, 1] and order bβc + 1, i.e. K(u) du = 1,
R r
u K(u) du = 0 ∀r = 1, . . . , bβc. Furthermore, K is differentiable with Lipschitzcontinuous derivative K 0 on (−1, 1).
(B1) The sequence (bn )n∈N of positive bandwidths satisfies limn→∞ bn = 0.
(
log n 1/α
o b1+2δ
if δ ≤
n
β
(1+2δ)∨(3−(β−1)(1/δ−1))
(B2.δ) hn +
= o bn
=
3−(β−1)(1/δ−1)
nhn
o bn
if δ >
β−1
2
β−1
.
2
Here we assume that the parameter δ, which quantifies the minimal required smoothness of
the estimator of g 0 , lies in (0, 1 ∧ (β − 1)). For example, if β < 3 and the optimal bandwidth
hn ((log n)/n)1/(αβ+1) is chosen, then (B2.δ) is fulfilled with δ = (β − 1)/2 for any bn that
satisfies hn = o(bn ).
The estimator g̃ is differentiable and we obtain the following uniform rates of convergence
for g̃ and its derivative g̃ 0 .
Theorem 2.7 If the model assumptions (2.1), (F1), (G1) with β > 1, (H1), (K1), and
(B1) hold, then for In = [hn + bn , 1 − hn − bn ]
| log h | α1
n
β
β
(i) sup |g̃(x) − g(x)| = O(bn ) + O(hn ) + OP
nhn
x∈In
| log h | α1
n
(ii) sup |g̃ (x) − g (x)| =
+O
+ OP
.
nhn
x∈In
If hβn + (log n/(nhn ))1/α = o(bn ), then supx∈In |g̃ 0 (x) − g 0 (x)| = oP (1); in particular this
holds if (B2.δ) is fulfilled for some δ ∈ (0, 1 ∧ (β − 1)).
0
0
O(bβ−1
n )
β
b−1
n hn
7
b−1
n
(iii) For all δ ∈ (0, 1 ∧ (β − 1)), under the additional assumption (B2.δ),
|g̃ 0 (x) − g 0 (x) − g̃ 0 (y) + g 0 (y)|
= oP (1).
|x − y|δ
x,y∈In ,x6=y
sup
3
3.1
The error distribution
Estimation
In this section we consider estimators of the error distribution in model (2.1). For the
asymptotic analysis we need the following additional assumption.
(F2) The cdf F of the errors is Hölder continuous of order α ∧ 1.
We define residuals ε̂i = Yi − ĝ( ni ), and a resulting modified sequential empirical distribution function by
bnsc
1 X
F̂n (y, s) =
I{ε̂i ≤ y}I{hn <
mn i=1
i
n
≤ 1 − hn },
where mn = ]{i ∈ {1, . . . , n} | hn < ni ≤ 1 − hn } = n − bnhn c − dnhn e. We consider the
sequential process, because it will be useful for testing hypotheses in section 4. With slight
abuse of notation, let F̂n (y) = F̂n (y, 1) denote the corresponding estimator for F (y).
We first treat a simple case where the influence of the regression estimation on the residual
empirical process is negligible. To this end, let Fn denote the standard empirical distribution
function of the unobservable errors ε1 , . . . , εn . Furthermore, define s̄n = bn(s ∧ (1 − hn ))c −
bn(s ∧ hn )c /mn and interpret s̄n /bnsc as 0 for s = 0. Note that s̄n = 1 if s = 1 and sn → s
as n → ∞, for each fixed s.
Theorem 3.1 Assume that the conditions (F1), (G1), and (F2) are fulfilled with β > 1.
Furthermore, assume β1 < α < 2 − β1 and hn ((log n)/n)1/(αβ+1) . Then we have
sup
|F̂n (y, s) − s̄n Fbnsc (y)| = oP (n−1/2 ).
y∈R,s∈[0,1]
√
Thus the process { n(F̂n (y, s) − s̄n F (y)) | s ∈ [0, 1], y ∈ R} converges weakly to a Kiefer
process KF , a centered Gaussian process with covariance function ((s1 , y1 ), (s2 , y2 )) 7→ (s1 ∧
s2 )(F (y1 ∧ y2 ) − F (y1 )F (y2 )).
Remark 3.2 The assertion of Theorem 3.1 holds true under the following weaker conditions
on the (possibly random) bandwidth:
hn = oP n−1/(2(α∧1)β) ,
n(α∨1)/2−1 log n = oP (hn ).
8
(3.1)
In particular, one may use the adaptive bandwidth proposed by Jirak et al. (2014).
Condition (3.1) can be fulfilled if and only if β1 < α < 2 − β1 , which in turn can be satisfied
for all α ∈ (0, 2), provided the regression function g is sufficiently smooth. It ensures that
one can choose a rate an of larger order than the uniform bound on the estimation error
established in Theorem 2.2 such that
−1/2
|F (y + an ) − F (y)| = O(aα∧1
).
n ) = o(n
Remark 3.3 Theorem 3.1 implies that for α ∈ (1/β, 2 − 1/β) the estimation of the regression function has no impact on the estimation of the irregular error distribution. This is
remarkably different from corresponding results on the estimation of the error distribution
in mean regression models with regular error distributions. Here the empirical distribution
√
function of residuals, say F̌n , is not asymptotically n-equivalent to the empirical distribu√
tion function of true errors. The process n(F̌n − F ) converges to a Gaussian process whose
covariance structure depends on the error distribution in a complicated way; cf. Theorem 2
in Akritas and Van Keilegom (2001). In the simple case of a mean regression model with
equidistant design and an error distribution F with bounded density f one has
√
n
f (y) X
n(F̌n (y) − Fn (y)) = √
εi + oP (1)
n i=1
uniformly with respect to y ∈ R when the regression function is estimated by a local polynomial estimator, under appropriate bandwidth conditions (see Proposition 3 in Neumeyer
and Van Keilegom (2009)).
In order to obtain asymptotic results for estimators of the error distribution for α ≥ 2− β1 ,
a finer analysis is needed. In what follows, we will use the smooth regression estimator
g̃ defined in (2.5). Let F̃n denote the empirical distribution function based on residuals
ε̃j = Yj − g̃( nj ), i. e.
n
1 X
F̃n (y) =
I{ε̃j ≤ y}I{ nj ∈ In }
mn j=1
where In = [hn + bn , 1 − hn − bn ] and mn = ]{j ∈ {1, . . . , n} | hn + bn ≤ nj ≤ 1 − hn − bn } =
n − 2dn(hn + bn )e + 1. Then the following asymptotic expansion is valid.
Theorem 3.4 If the conditions (F1), (F2), (G1) with β > 1, (H1), (K1), (B1), and
(B2.δ) for some δ ∈ (1/α − 1, 1 ∧ (β − 1)) are fulfilled, then
F̃n (y) =
n
n
1
1X
1 X
I{εj ≤ y}+
F y + (g̃ − g)( nj ) − F (y) I{ nj ∈ In }+oP √
(3.2)
n j=1
mn j=1
n
uniformly for all y ∈ R.
9
Remark 3.5 One can choose bandwidths hn and bn such that the conditions (H1), (B1)
and (B2.δ) are fulfilled for some δ ∈ (1/α − 1, 1 ∧ (β − 1)) if this interval is not empty, which
in turn is equivalent to α > 1/(β ∧ 2). Thus the expansion given in Theorem 3.4 is also valid
for regular error distributions.
If one assumes (B2.δ) for some δ ∈ (0, 1 ∧ (β − 1)), but drops the condition δ > 1/α − 1 and,
in addition, replaces (F2) with the assumption that F is Lipschitz continuous on (−∞, κ]
for some κ < 0, then expansion (3.2) still holds uniformly on (−∞, κ̃] for all κ̃ < κ. In
particular, this holds if F has a bounded density on (−∞, κ].
Next we examine under which conditions the additional term in (3.2) depending on the
estimation error is asymptotically negligible. We focus on those arguments y which are
bounded away from 0, because in this setting weaker conditions on α and β are needed.
Moreover, for the analysis of the tail behavior of the error distribution at 0, tail empirical
processes are better suited and will be considered in future work.
Note that the estimator ĝ tends to underestimate the true function because it is defined
via a polynomial which is minimal under the constraint that it lies above all observations
(i/n, Yi ), which in turn all lie below the true boundary function. As this systematic underestimation does not vanish from (local or global) averaging, we first have to introduce a bias
correction.
Let Eg≡0 denote the expectation if the true regression function is identical 0. For the
remaining part of this section, we assume that Eg≡0 (ĝ(1/2)) is known or that it can be estimated sufficiently accurately. For example, if the empirical process of residuals shall be used
to test a simple null hypothesis, then one may calculate or simulate this expectation under
the given null distribution. We define a bias corrected version of the smoothed estimator by
g̃n∗ (x) := g̃(x) − Eg≡0 (ĝ(1/2)),
for x ∈ In . The following lemma ensures that the above results for g̃ carry over to this
variant if the following condition on the lower tail of F holds:
(F3) There exists τ > 0 such that F (−t) = o(t−τ ) as t → ∞.
Lemma 3.6 If model (2.1) holds with g identical 0 and the conditions (F1), (F3), (G1),
and (H1) are fulfilled, then for all x ∈ [hn , 1 − hn ]
Eg≡0 (|ĝn (x)|) = Eg≡0 (|ĝn (1/2)|) = O
log n 1/α
.
nhn
We need some additional conditions on the rates at which the bandwidths hn and bn tend
to 0:
(H2) hn = o n−1/(2β) ∧ n−1/(αβ+1) , nα/4−1 log n = o(hn )
10
nh 2/α
n
−1/(2β)
−2β −1
−1
(B3) bn = o n
∧ hn n
∧
n
log n
In particular, these assumptions ensure that the bias terms of order hβn + bβn are of smaller
order than n−1/2 and (nhn )−1/α and hence asymptotically negligible, and that quadratic
terms in the estimation error are uniformly negligible, that is, supx∈In |g̃n∗ (x) − g(x)|2 =
oP (n−1/2 ).
Theorem 3.7 Suppose the model assumptions (2.1) with α ∈ (0, 2), β > 1, (F1), (F3),
(G1), (H1), (H2), (K1), (B1), (B2.δ) for some δ > 0, and (B3) hold and F has a
bounded density on (−∞, κ] for some κ < 0. Then
sup
y∈(−∞,κ]
n
1 X
F y + (g̃n∗ − g)( nj ) − F (y) I{ nj ∈ In } = oP (n−1/2 ).
mn j=1
Remark 3.8 The conditions on hn and bn used in Theorem 3.7 can be fulfilled if and only
if α < 2β − 1. In particular, this theorem is applicable if β ≥ 3/2 and the error distribution
is irregular, i.e., α < 2. A possible choice of bandwidths is
1
β
2β − 1
,
∧
.
hn n−1/(2β) ∧ n−1/(αβ+1) / log n, bn n−λ for some λ ∈
2β αβ + 1
2αβ
We obtain asymptotic equivalence of the empirical process of residuals (restricted to
(−∞, κ]) to the empirical process of the errors. To formulate the result, let F̃n∗ be defined
analogously to F̃n , but with g̃ replaced by g̃ ∗ .
Corollary 3.9 Under the assumptions of Theorems 3.4 and 3.7, we have supy∈(−∞,κ] |F̃n∗ (y)−
√
Fn (y)| = oP (n−1/2 ). Thus the process ( n(F̃n∗ (y) − F (y)))y∈(−∞,κ] converges weakly to a
centered Gaussian process with covariance function (y1 , y2 ) 7→ F (y1 ∧ y2 ) − F (y1 )F (y2 ),
y1 , y2 ∈ (−∞, κ].
Note that for the Corollary one needs the condition 1/(β ∧ 2) < α < (2β − 1) ∧ 2.
4
4.1
Hypotheses testing
Goodness-of-fit testing
Let F = {Fϑ | ϑ ∈ Θ} denote a continuously parametrized class of error distributions such
that for each ϑ ∈ Θ, Fϑ (y) = 1 − cϑ |y|αϑ + rϑ (y) with rϑ (y) = o(|y|αϑ ) for y % 0. Our aim is
to test the null hypothesis H0 : F ∈ F. We assume that αϑ ∈ (1/β, 2 − 1/β) for all ϑ ∈ Θ,
such that Theorem 3.1 can be applied under H0 . Let ϑ̂ denote an estimator for ϑ based
on residuals ε̂i = Yi − ĝ( ni ), i = 1, . . . , n. The goodness-of-fit test is based on the empirical
process
√
Sn (y) = n(F̂n (y) − Fϑ̂ (y)), y ∈ R,
11
where, as before, F̂n (y) = F̂n (y, 1). Under any fixed alternative that fulfills (F1) for some α,
ĝ still uniformly consistently estimates g, and thus F̂n is a consistent estimator of the error
distribution F . If ϑ̂ converges to some ϑ∗ ∈ Θ under the alternative, too, then a consistent
hypothesis test is obtained by rejecting H0 for large values of, e. g., a Kolmogorov-Smirnov
test statistic supy∈R |Sn (y)|. Note that under H0 it follows from Theorem 3.1 that
Sn (y) =
√
n(Fn (y) − Fϑ (y)) −
√
n(Fϑ̂ (y) − Fϑ (y)) + oP (1),
where ϑ denotes the true parameter. We consider two examples.
Example 4.1 Consider the mean regression model Yi = m( ni ) + ηi , i = 1, . . . , n, with
symmetric error cdf F and β > 1, and define m̂ with some bandwidth hn ((log n)/n)1/(αβ+1)
as in Remark 2.5. We want to test the null hypothesis H0 : F ∈ F = {Fϑ | ϑ ∈ Θ} for
some Θ ⊂ (0, ∞), where Fϑ denotes the distribution function of the uniform distribution on
[−ϑ, ϑ] (with αϑ = 1 for all ϑ > 0). Define residuals η̂i = Yi − m̂( ni ), i = 1, . . . , n, and let
max
|η̂i |.
ϑ̂n = max
max
η̂i , − min
η̂i =
nhn ≤i≤n−nhn
nhn ≤i≤n−nhn
nhn ≤i≤n−nhn
Then |ϑ̂n − ϑ| is bounded by | maxnhn ≤i≤n−nhn |ηi | − ϑ| + supx∈[hn ,1−hn ] |m̂(x) − m(x)| =
oP (n−1/2 ). Since Fϑ (y) = y+ϑ
I
(y) + I(ϑ,∞) (y), one may conclude supy∈R |Fϑ̂n (y) −
2ϑ [−ϑ,ϑ]
−1/2
Fϑ (y)| = oP (n
). Thus the process Sn converges weakly to a Brownian bridge B composed
with F . The Kolmogorov-Smirnov test statistic supy∈R |Sn (y)| converges in distribution to
supt∈[0,1] |B(t)|. Thus although our testing problem requires the estimation of a nonparametric function and we have a composite null hypothesis, the same asymptotic distribution
arises as in the Kolmogorov-Smirnov test for the simple hypothesis H0 : F = F0 based on
an iid sample with distribution F .
Example 4.2 Again assume that the Hölder coefficient β is greater than 1. Consider the
α
null hypothesis H0 : F ∈ F = {Fϑ | ϑ ∈ (0, ∞)}, where Fϑ (y) = e−(−ϑy) I(−∞,0) (y)+I[0,∞) (y)
denotes a Weibull distribution with some fixed shape parameter α ∈ (1/β, 2 − 1/β) and
unknown scale parameter ϑ. Note that Fϑ satisfies (F1) with c = ϑ.
P
− α1
n
j
1
α
Define the moment estimator ϑ̂n = mn j=1 (−ε̂j ) I{hn < n ≤ 1 − hn }
which is
α
−α
α
motivated by Eϑ [(−ε1 ) ] = ϑ . A Taylor expansion of x 7→ x at x = −εj yields
ϑ̂αn
n
1 X
j
− ϑ = −(ϑ̂n ϑ)
((−εj )α − ϑ−α )I{hn < ≤ 1 − hn }
mn j=1
n
α
α
n
j
α X
j
j
α−1
+
(−ξj )
ĝ( ) − g( ) I{hn < ≤ 1 − hn }
mn j=1
n
n
n
n
1 X
j
= −ϑ
((−εj )α − ϑ−α )I{hn < ≤ 1 − hn } + oPϑ (n−1/2 )
mn j=1
n
2α
12
= OPϑ (n−1/2 )
for some ξj between ε̂j and εj , where in the last steps we have applied Theorem 2.2, the law
of large numbers and a central limit theorem.
For all z, z̃ ∈ R one has |e−z − e−z̃ − (z − z̃)e−z | = e−z |ez−z̃ − 1 − (z − z̃)| ≤ e−z∧z̃ (z − z̃)2 .
Thus
2
2
α
α
Fϑ̂n (y) − Fϑ (y) − e−(−ϑy) ϑ̂αn − ϑα | ≤ e−(−(ϑ̂n ∧ϑ)y) (ϑ̂αn − ϑα )(−y)α = OPϑ (n−1 )
uniformly for all y ∈ (−∞, 0]. Now analogously to the proof of Theorem 19.23 in van der
Vaart (2000) we can conclude weak convergence of
Sn (y) =
√
n(Fn (y) − F (y)) − e
√ X
n
n
1
j
((−εj )α − α )I{hn < ≤ 1 − hn }
ϑ (−y)
mn j=1
ϑ
n
−(−ϑy)α 2α
α
+ oPϑ (1),
y ∈ R, to a Gaussian process with covariance function (y1 , y2 ) 7→ Fϑ (y1 ∧y2 )−Fϑ (y1 )Fϑ (y2 )−
α α
α
e−(−ϑ) (y1 +y2 ) (y1 y2 )α ϑ2α , where the covariance function follows by simple calculations and
α
the fact that Eϑ [I{ε1 ≤ y}((−ε1 )α − ϑ−α )] = (−y)α e−(−ϑy) .
For the special case of a test for exponentially distributed errors (α = 1), the asymptotic
R
quantiles for the Cramér-von-Mises test statistic Sn (y)2 dFϑ̂n (y) are tabled in Stephens
(1976).
Simulations
To study the finite sample performance of our goodness-of-fit test, we investigate its behaviour on simulated data according to Examples 4.1 and 4.2 for samples of size 50, 100, 200
and 500. In both settings the regression function is given by g(x) = 0.5 sin(2πx) + 4x. We
1
use the local linear estimator (corresponding to β = 2) with bandwidth n− 3 , which is up
to a log term of optimal rate for α = 1 and β = 2. The hypothesis tests are based on the
R
adjusted Cramér-von-Mises test statistic mnn Sn (y)2 dFϑ̂n (y) and have nominal size 5%. The
results reported below are based on 200 Monte Carlo simulations for each model.
In the situation of Example 4.1, the errors are drawn according to the density fε (y) =
0.5(ζ + 1)(1 − |y|)ζ I[−1,1] (y) for different values of ζ ∈ (−1, 0] . Note that the null hypothesis
H0 : ∃ϑ : εi ∼ U [−ϑ, ϑ] holds if and only if ζ = 0. Figure 3 shows the empirical power of
the Cramér-von-Mises type test. The actual size is close to the nominal level for all sample
sizes and the power function is monotone both in ζ and the sample size n. For parameter
values ζ ∈ [−0.2,0), one needs rather large sample sizes to detect the alternative, as the error
distribution is too similar to the uniform distribution.
In the setting of Example 4.2 we simulate Weibull(ϑ, α) distributed errors for ϑ = 1
and different values of α > 0. We test the null hypothesis H0 : ∃ϑ : −εi ∼ Exp(ϑ) of
13
1.0
exponentiality, which is only fulfilled for α = 1. In Figure 4 the empirical power function of
our test is displayed for different sample sizes. Again the actual size is close to the 5% and
the power increases with α departing from one as well as with increasing n.
To examine the influence of the bandwidth choice, in addition we have simulated the
1
same models with hn = c · n− 3 for different values of c ranging from c = 0.2 to c = 1.2. The
results for the test of uniformity in Example 4.1 are similar to those displayed in Figure 3
for all these bandwidths. In the situation of Example 4.2 we obtain similar power functions
as reported above for c between 0.8 and 1.2, whereas for smaller bandwidths the actual size
of the test exceeds its nominal value substantially.
●
●
●
0.6
0.4
Rejection Probability
0.8
●
n=50
n=100
n=200
n=500
0.2
●
level
●
●
0.0
●
−0.8
−0.6
−0.4
−0.2
0.0
ζ
Figure 3: Monte-Carlo simulations for Example 4.1
4.2
Test for independence
In model (2.1) we assume that the distributions of the errors εi (i = 1, . . . , n) do not depend
on the point of measurement xi = i/n. We can test this assumption by comparing the sequential empirical distribution function F̂n (y, s) for the residuals with the estimator s̄n F̂n (y),
which should behave similarly if the errors are iid. The following corollary to Theorem 3.1
describes the asymptotic behavior of the Kolmogorov-Smirnov type test statistic
Tn =
sup
√
n|F̂n (y, s) − s̄n F̂n (y)|
s∈[0,1],y∈R
14
1.0
●
●
●
●
●
●
●
●
●
●
●
0.6
●
0.4
●
0.2
Rejection Probability
0.8
●
●
●
level
●
n=50
n=100
n=200
n=500
0.0
●
0.5
1.0
1.5
2.0
2.5
3.0
α
Figure 4: Monte-Carlo simulations for Example 4.2
under the null hypothesis of iid errors.
Corollary 4.3 Assume model (2.1) with (F1), (F2), (G1), and β1 < α < 2 − β1 . Choose a
bandwidth hn ((log n)/n)1/(αβ+1) .
Then Tn converges in distribution to sups∈[0,1],z∈[0,1] |G(s, z)| where G is a completely
tucked Brownian sheet, i. e. a centered Gaussian process with covariance function
((s1 , z1 ), (s2 , z2 )) 7→ (s1 ∧ s2 − s1 s2 )(z1 ∧ z2 − z1 z2 ).
The proof is given in the appendix. Note that under the assumptions of the corollary the
limit of the test statistic Tn is distribution free. The asymptotic quantiles tabled by Picard
(1985) can be used to determine the critical value for a given asymptotic size of the test.
4.3
Test for monotone boundary functions
We consider model (2.1) and aim at testing the null hypothesis
H0 :
g is increasing,
which is a common assumption in boundary models. Let g̃ denote the smooth local polynomial estimator for g defined in (2.5). Such an unconstrained estimator can be modified
15
to obtain an increasing estimator g̃I . To this end, for any function h : [0, 1] → R define the
increasing rearrangement on [a, b] ⊂ [0, 1] as the function Γ(h) : [a, b] → R with
Z b
o
n
I{h(t) ≤ z} dt ≥ x .
Γ(h)(x) = inf z ∈ R a +
a
Denote by Γn the operator Γ with [a, b] = In . We define the increasing rearrangement of
g̃ as g̃I = Γn (g̃), so that g̃I = g̃ if g̃ is nondecreasing (see Anevski and Fougères, 2007,
and Chernozhukov et al., 2009). We now consider residuals obtained from the monotone
estimator: ε̂I,i = Yi − g̃I ( ni ), i = 1, . . . , n. Under the null hypothesis, these residuals should
be approximately iid, whereas under the alternative they show a varying behavior for ni in
different subintervals of [0, 1]. For illustration see Figure 5 where we have generated a data
set (upper panel) with true non-monotone boundary curve g (dashed curve). The solid curve
is the increasing rearrangement gI . The lower left panel shows the errors εi , i = 1, . . . , n,
with iid-behaviour. The lower right panel shows εI,i = Yi − gI ( ni ), i = 1, . . . , n, with a clear
non-iid pattern.
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ● ●
●
●
●
●
● ●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
s
●
●
●
●
●
●
●
●
0
−1
−2
−3
−4
−4
−3
−2
−1
0
●
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
0.6
0.8
1.0
Figure 5: The upper panel shows data points and the true boundary function (dashed curve)
as well as the increasing rearrangement (solid curve). The lower left panel shows the errors.
The lower right panel shows residuals built with respect to the increasing rearrangement.
16
Similarly as in Subsection 4.2, we compare the sequential empirical distribution function
bnsc
1 X
I{ε̃I,j ≤ y}I{ nj ∈ In }
F̃I,n (y, s) =
mn j=1
based on the increasing estimator g̃I with the product estimator s̄n F̃n (y, 1), where again
In := [hn + bn , 1 − hn − bn ] and mn := n − 2dn(hn + bn )e + 1. Let
√
G̃n (s, y) = n(F̃I,n (y, s) − s̄n F̃n (y)), s ∈ [0, 1], y ∈ R.
To derive its limit distribution under the null hypothesis, we need an additional assumption:
(I1) Let inf x∈[0,1] g 0 (x) > 0.
Theorem 4.4 Assume model (2.1) with (F1), (F2), (G1), (K1), (I1), β > 1 and
α < 2 − β1 . If hn ((log n)/n)1/(αβ+1) and bn ((log n)/n)1/(αβ+1) , then
sup
|F̃I,n (y, s) − s̄n Fbnsc (y)| = oP (n−1/2 ).
1
β
<
(4.1)
y∈R,s∈[0,1]
Thus the Kolmogorov-Smirnov test statistic sups∈[0,1],y∈R |G̃n (s, y)|, converges in distribution
to sups∈[0,1],z∈[0,1] |G(s, z)| where G is the completely tucked Brownian sheet (see Corollary
4.3).
The conditions on the bandwidths can be substantially relaxed; cf. Remark 3.2.
Remark 4.5 A test that rejects H0 for large values of the Kolmogorov-Smirnov test statistic
Tn = sups∈[0,1],y∈R |G̃n (s, y)| is consistent. To see this note that by Theorem 1 of Anevski and
Fougères (2007), supx∈In |g̃I (x) − gI (X)| ≤ supx∈In |g̃(x) − g(x)| = oP (1) with gI denoting
the increasing rearrangement of g. Thus n−1/2 Tn converges to
Z s
T = sup
Fε (y + (gI − g)(x)) dx − sFε (y) .
s∈[0,1],y∈R
0
Since T > 0 under the alternative hypothesis g 6= gI , the test statistic Tn converges to
infinity.
A
A.1
Appendix: Proofs
Auxiliary results
Proposition A.1 Assume that model (2.1) holds and that the regression function g fulfills
condition (G1) for some β ∈ (0, β ∗ ] and some cg ∈ [0, c∗ ]. Then there exist constants
Lβ ∗ ,c∗ , Lβ ∗ > 0 and a natural number jβ ∗ (depending only on the respective subscripts) such
that
|ĝ(x) − g(x)| ≤ Lβ ∗ ,c∗ hβn + Lβ ∗ max
min
|εi | .
1≤j≤2jβ ∗
17
i:−1+(j−1)/jβ ∗ ≤|i/n−x|/hn ≤−1+j/jβ ∗
This proposition can be verified by an obvious modification of the proof of Theorem 3.1
by Jirak et al. (2014).
Lemma A.2 Under assumptions (F1) and (H1) for any fixed set I1 , . . . , Im of disjoint
non-degenerate subintervals of [−1, 1] we have
sup
max
x∈[hn ,1−hn ] 1≤j≤m
min
i∈{1,...,n}
(i/n−x)/hn ∈Ij
|εi | = OP
| log h | 1/α
n
.
nhn
1/α
Proof. Let rn := | log hn |/(nhn )
. Obviously it suffices to prove that for all nondegenerate subintervals I ⊂ [−1, 1] there exists a constant L such that
n
o
lim P
sup
min |εi | > Lrn = 0.
n→∞
x∈[hn ,1−hn ]
i∈{1,...,n}
(i/n−x)/hn ∈I
Denote by d = sup I − inf I > 0 the diameter of I and let dn := dnhn de − 1 and ln := bn/dn c.
Then for all x > 0
n
o
n
o
P
sup
min |εi | > x ≤ P
max
min
|εi | > x
x∈[hn ,1−hn ]
j∈{1,...,n−dn } i∈{j,...,j+dn }
i∈{1,...,n}
(i/n−x)/hn ∈I
≤ P
n
o
n
o
max Mn,l > x + P
max Mn,l > x
l∈{0,...,ln }
l even
l∈{0,...,ln }
l odd
with
Mn,l :=
max
min
j∈{ldn +1,...,(l+1)dn } i∈{j,...,j+dn }
|εi |.
Since the random variables Mn,l for l even are iid, we have
n
o
bln /2c+1
P
max Mn,l > x = 1 − 1 − P {Mn,0 > x}
,
l∈{0,...,ln }
l even
and an analogous equation holds for the maxima over the odd numbered block maxima Mn,l .
Let G be the cdf of |εi |. If Mn,0 exceeds x, then there is a smallest index j ∈ {1, . . . , dn }
for which mini∈{j,...,j+dn } |εi | > x. Hence
P {Mn,0 > x} = P
n
min
i∈{1,...,1+dn }
dn
o X
n
|εi | > x +
P |εj−1 | ≤ x,
j=2
min
i∈{j,...,j+dn }
o
|εi | > x
= (1 − G(x))dn +1 + (dn − 1)G(x)(1 − G(x))dn +1
≤ (1 + dn G(x))(1 − G(x))dn .
To sum up, we have shown that
P
n
sup
min
x∈[hn ,1−hn ]
i∈{1,...,n}
(i/n−x)/hn ∈I
|εi | > Lrn
o
bln /2c+1
dn
≤ 2 1 − 1 − (1 + dn G(Lrn ))(1 − G(Lrn ))
.
18
It remains to be shown that the right hand side tends to 0 for sufficiently large L which is
true if and only if
(1 + dn G(Lrn ))(1 − G(Lrn ))dn = o(1/ln ).
This is an immediate consequence of 1/ln ∼ dhn and
| log hn |
(1 + o(1))
nhn
| log hn |
=⇒ (1 − G(Lrn ))dn = exp − nhn dcLα
(1 + o(1))
nh
n
=⇒ (1 + dn G(Lrn ))(1 − G(Lrn ))dn = O | log hn | exp − cdLα | log hn |(1 + o(1)) = o(hn )
G(Lrn ) = cLα
2
if cdLα > 1.
A.2
Proof of Theorem 2.2
The assertion directly follows from Proposition A.1 and Lemma A.2.
A.3
2
Proof of Theorem 2.7
(i) Using Theorem 2.2, a Taylor expansion of g of order bβc and assumption (K1), one can
show by direct calculations that for some τu ∈ (0, 1)
1−hn
x−z
dz
sup |g̃(x) − g(x)| ≤ sup
bn
x∈In
x∈In
hn
Z 1−hn
1
x−z
(g(z) − g(x)) K
+ sup
dz
bn
bn
x∈In
hn
Z 1
(g(x − ubn ) − g(x))K (u) du
≤
sup |ĝ(z) − g(z)|O(1) + sup
Z
1
(ĝ(z) − g(z)) K
bn
x∈In
z∈[hn ,1−hn ]
−1
| log h | α1
n
β
≤ O(hn ) + OP
nh
Z 1n
1
ubβc (g (bβc) (x − τu ubn ) − g (bβc) (x))K(u)du .
+bbβc
n sup
x∈In bβc! −1
Now the Hölder property of g combined by (K1) yields the desired result.
(ii) Since g is bounded on [hn , 1 − hn ] and supx∈[hn ,1−hn ] |ĝ(x) − g(x)| = oP (1), ĝ is
eventually bounded on [hn , 1 − hn ] too. Note that the partial derivative of ĝ(z)b−1
n K((x −
z)/bn ) with respect to x is continuous and bounded (for fixed n). Thus we can exchange
integration and differentiation and obtain
0
Z
0
1−hn
sup |g̃ (x) − g (x)| = sup
x∈In
x∈In
ĝ(z)
hn
19
1 0 x − z
K
dz − g 0 (x) .
b2n
bn
Integration by parts yields
Z 1−hn
Z 1−hn
1
1 0 x−z
x−z
0
g (z) K
g(z) 2 K
dz =
dz
bn
bn
bn
bn
hn
hn
since K(−1) = K(1) = 0. Therefore
Z 1−hn
1 0 x−z
0
0
(ĝ(z) − g(z)) 2 K
dz
sup |g̃ (x) − g (x)| ≤ sup
bn
bn
x∈In
x∈In
hn
Z 1−hn
1
x−z
0
0
(g (z) − g (x)) K
+ sup
dz
bn
bn
x∈In
hn
Z 1
−1
(g 0 (x − ubn ) − g 0 (x))K (u) du .
≤
sup |ĝ(z) − g(z)|O(bn ) + sup
x∈In
z∈[hn ,1−hn ]
−1
Similarly as in the proof of (i), assertion (ii) follows by Theorem 2.2, a Taylor expansion of
g 0 of order bβc − 1 and the assumptions (K1) and (G1).
(iii) We distinguish the cases |x − y| > an and |x − y| ≤ an for some suitable sequence
(an )n∈N with limn→∞ an = 0 specified later. In the first case, we obtain
|g̃ 0 (x) − g 0 (x) − g̃ 0 (y) + g 0 (y)|
|x − y|δ
sup
x,y∈In ,|x−y|>an
≤ 2 sup |g̃ 0 (x) − g 0 (x)|a−δ
n
x∈In
O(bβ−1
n )+
=
O(hβn ) + OP
| log hn |
nhn
α1 ! ! !
b−1
a−δ
n
n .
(A.1)
In the second case, we use a decomposition like in the proof of (ii):
sup
x,y∈In ,0<|x−y|≤an
≤
sup
|g̃ 0 (x) − g 0 (x) − g̃ 0 (y) + g 0 (y)|
|x − y|δ
R 1−hn
1
0 x−z
0 y−z
(ĝ(z)
−
g(z))
K
−
K
dz
2
b
bn
bn
hn
n
|x − y|δ
R 1−hn
x,y∈In ,0<|x−y|≤an
0
+
sup
x,y∈In
0<|x−y|≤an
0
|g (x) − g (y)|
+
|x − y|δ
hn
sup
x,y∈In
0<|x−y|≤an
y−z
g 0 (z) b1n K x−z
−
K
dz
bn
bn
|x − y|δ
.
By Lipschitz continuity of K 0 and Theorem 2.2, the first term on the right hand side is of
the order
!
| log h | α1 1
n
O(hβn ) + OP
O(an1−δ ).
(A.2)
nhn
b3n
For β ≥ 2, the second term is of the order an1−δ as g 0 is Lipschitz continuous, while for
β ∈ (1, 2) assumption (G1) yields the rate anβ−1−δ . In both cases, condition (B2.δ) ensures
20
that the second term converges to 0.
The last term on the right hand side can be rewritten as
R1 0
(g (x − hn u) − g 0 (y − hn u))K(u) du
−1
sup
|x − y|δ
x,y∈In
0<|x−y|≤an
and is thus of the same order as the second term by assumption (G1).
To conclude the proof, one needs to find a sequence an = o(1) such that (A.1) and (A.2)
tend to 0 in probability, i.e.
b3
ϑn
β−1
δ
1−δ
bn +
= o(an ) and an = o n
bn
ϑn
with ϑn := hβn + (| log hn |/(nhn ))1/α . Obviously, such a sequence an exists if and only if
bβ−1
+
n
δ
b3 1−δ
ϑn
n
,
=o
bn
ϑn
2
which in turn is equivalent to condition (B2.δ).
A.4
Proof of Theorem 3.1
The assumptions about α ensure that β/(αβ + 1) > 1/(2(α ∧ 1)), and so in view of (2.3)
the uniform estimation error of ĝ is stochastically of smaller order than n−1/(2(α∧1)) . Hence
there exists a sequence
1
an = o(n− 2(α∧1) )
(A.3)
such that
P
|ĝ(x) − g(x)| ≤ an −−−→ 1.
sup
n→∞
x∈[hn ,1−hn ]
Let F̄n (y, s) :=
1
mn
Pbnsc
j=1
I{εj ≤ y}I{hn <
j
n
≤ 1 − hn }. Since
bnsc
1 X
F̂n (y, s) =
I{εj ≤ y + (ĝ − g)( nj )}I{hn <
mn j=1
j
n
≤ 1 − hn }
we may conclude
√
√
n(F̄n (y − an , s) − s̄n Fbnsc (y)) ≤ n(F̂n (y, s) − s̄n Fbnsc (y))
√
≤ n(F̄n (y + an , s) − s̄n Fbnsc (y))
for all y ∈ R and s ∈ [0, 1] with probability converging to 1.
We take a closer look at the bounds. The sequential empirical process
−1/2
En (y, s) = n
bnsc
X
(I{εj ≤ y} − F (y)),
j=1
21
y ∈ R, s ∈ [0, 1],
(A.4)
converges weakly to a Kiefer process; see e.g. Theorem 2.12.1 in van der Vaart and Wellner
(1996). Now, n ∼ mn , the asymptotic equicontinuity of the process En , the Hölder continuity
(F2) and (A.3) imply
√
n F̄n (y ± an , s) − s̄n Fbnsc (y)
n
=
En (y ± an , s ∧ (1 − hn )) − En (y, s ∧ (1 − hn )) − En (y ± an , s ∧ hn ) + En (y, s ∧ hn )
mn
√
√
+ ns̄n (F (y ± an ) − F (y)) + n(F̄n (y, s) − s̄n Fbnsc (y))
√
= oP (1) + n(F̄n (y, s) − s̄n Fbnsc (y))
uniformly for all y ∈ R, s ∈ [0, 1].
It remains to be shown that
√
n(F̄n (y, s) − s̄n Fbnsc (y))
√ X
√
bnsc
bnsc
n
ns̄n X
j
=
(I{εj ≤ y} − F (y))I{hn < ≤ 1 − hn } −
(I{εj ≤ y} − F (y))
mn j=1
n
bnsc j=1
n
=
− 1 En (y, s ∧ (1 − hn )) − En (y, s ∧ hn )
mn
ns̄
n
−
− 1 En (y, s)
bnsc
+ En (y, s ∧ (1 − hn )) − En (y, s ∧ hn ) − En (y, s)
(A.5)
tends to 0 in probability uniformly for all y ∈ R, s ∈ [0, 1].
The first term vanishes asymptotically, because En is uniformly stochastically bounded
and n ∼ mn .
Next note that s̄n = 0 for s < hn , while for s ≥ hn
ns̄n
bn(s ∧ (1 − hn )c − bnhn c
O(nhn )
−1=
−1=
,
−1
bnsc
(1 − 2hn + O(n ))bnsc
(1 − 2hn + O(n−1 ))bnsc
(A.6)
which is uniformly bounded for all s ∈ [hn , 1] and tends to 0 uniformly with respect to s ∈
1/2
[hn , 1]. Moreover, En is uniformly stochastically bounded and sup0≤s≤h1/2
|En (y, s)| =
n ,y∈R
oP (1), because En is asymptotically equicontinuous with En (y, 0) = 0. Hence, the second
term in (A.5) converges to 0 in probability, too. Likewise, the convergence of the last term
to 0 follows from the asymptotic equicontinuity of En , which concludes the proof.
2
A.5
Proof of Theorem 3.4 and of Remark 3.5
For any interval I ⊂ R and constant k > 0, define the following class of differentiable
functions:
n
|d0 (x) − d0 (y)| o
1+δ
0
Ck (I) = d : I → R max sup |d(x)|, sup |d (x)|, sup
≤k .
|x − y|δ
x∈I
x∈I
x,y∈I,x6=y
22
1+δ
Then Theorem 2.7 yields P ((g̃ − g) ∈ C1/2
(In )) → 1 as n → ∞. Hence there exist random functions dn : [0, 1] → R such that dn (x) = (g̃ − g)(x) for all x ∈ In and
P dn ∈ C11+δ ([0, 1]) → 1 for n → ∞. (For instance, one may extrapolate g̃ − g linearly on
[0, hn ] and on [1 − hn , 1].)
On the space F := R × C11+δ ([0, 1]) we define the semimetric
o
n
∗
∗
∗ ∗
|F (y + γ(x)) − F (y + γ(x))| , sup |d(x)−d (x)| .
ρ((y, d), (y , d )) = max sup
sup
x∈[0,1] γ∈C 1+δ ([0,1])
x∈[0,1]
1
For ϕ = (y, d) ∈ F let
√
n
1
Znj (ϕ) :=
I{εj ≤ y + d( nj )}I{ nj ∈ In } − √ I{εj ≤ y}
mn
n
and
n
X
Gn (ϕ) :=
(Znj (ϕ) − E[Znj (ϕ)]).
j=1
Note that
Gn (y, dn )
√ X
n
n
I{εj ≤ y − g(j/n) + g̃(j/n)}I{j/n ∈ In }
=
mn j=1
√ X
n
n
√
n
1 X
I{εj ≤ y} + nF (y)
−
F (y + (g̃ − g)(j/n))I{j/n ∈ In } − √
mn j=1
n j=1
=
n
n
√
1X
1 X
n F̃n (y) −
I{εj ≤ y} −
F y + (g̃ − g)( nj ) − F (y) I{ nj ∈ In } .
n j=1
mn j=1
We will apply Theorem 2.11.9 of van der Vaart and Wellner (1996) to show that the process
(Gn (ϕ))ϕ∈F converges to a (Gaussian) limiting process. In particular, Gn is asymptotically equicontinuous, which readily yields the assertion, because supy∈R ρ((y, dn ), (y, 0)) =
supx∈[0,1] |dn (x)| = oP (1) and the variance of
n
1 X n
Gn (y, 0) = √
− 1 I{εj ≤ y}I{j/n ∈ In } − I{εj ≤ y}I{j/n 6∈ In }
n j=1 mn
tends to 0, implying that Gn (y, 0) = oP (1) uniformly in y. Thus Gn (y, dn ) = oP (1) uniformly
in y and the assertion holds.
One may proceed as in the proof of Lemma 3 in Neumeyer and Van Keilegom (2009) (see
the online supporting information to that article) to prove that the conditions of Theorem
2.11.9 of van der Vaart and Wellner (1996) are fulfilled. The proof of the first two displayed
formulas of this theorem are analogous. The only difference is that Neumeyer and Van
23
Keilegom (2009) assume a bounded error density while we use Hölder continuity of F , see
assumption (F2). Next we show that the bracketing entropy condition (i.e., the last displayed
condition in Theorem 2.11.9 of van der Vaart and Wellner, 1996) is fulfilled and that (F, ρ)
is totally bounded.
To this end, let dLm ≤ dUm , m = 1, . . . , M , be brackets for C11+δ ([0, 1]) of length η 2/(α∧1)
w.r.t. the supremum norm. According to van der Vaart and Wellner (1996), Theorem 2.7.1
and Corollary 2.7.2, M = O exp(κη −2/((1+δ)(α∧1)) ) brackets are needed. For each m define
P
L
FmL (y) := n−1 nj=1 F (y + dLm (j/n)) and choose ym,k
, k = 1, . . . , K = O(η −2 ) such that
L
L
L
L
:= −∞ and ym,K+1
:= ∞.
) < η 2 for all k ∈ {1, . . . , K + 1} with ym,0
) − FmL (ym,k−1
FmL (ym,k
U
U
L
L
U
U
Define Fm and ym,k analogously, ỹm,k := ym,k and denote by ỹm,k the smallest ym,l larger
L
than or equal to ym,k+1
. Then F is covered by
L
U
Fmk := {(y, d) ∈ F | ỹm,k
≤ y ≤ ỹm,k
, dLm ≤ d ≤ dUm },
m = 1, . . . , M, k = 1, . . . , K.
Check that by condition (F2)
sup |FmU (y) − FmL (y)| ≤ sup n−1
y∈R
y∈R
n
X
|F (y + dUm (j/n)) − F (y + dLm (j/n))|
j=1
≤ LF sup |dUm (x) − dLm (x)|α∧1 ≤ LF η 2
(A.7)
x∈R
with LF denoting the Hölder constant of F . Thus
n
i2
1X h
E
sup
I{εj ≤ y + d(j/n)} − I{εj ≤ y ∗ + d∗ (j/n)}
n j=1
(y,d),(y ∗ ,d∗ )∈Fmk
n
2
1X
U
L
E I{εj ≤ ỹm,k
≤
+ dUm (j/n)} − I{εj ≤ ỹm,k
+ dLm (j/n)}
n j=1
U
L
≤ FmU (ỹm,k
) − FmL (ỹm,k
)
U
L
L
L
L
L
≤ |FmU (ỹm,k
) − FmU (ỹm,k+1
)| + |FmU (ỹm,k+1
) − FmL (ỹm,k+1
)| + |FmL (ỹm,k+1
) − FmL (ỹm,k
)|
≤ (2 + LF )η 2
L
U
where the last step follows from (A.7) and the definitions of ỹm,k
and ỹm,k
. Hence we obtain
n
for the squared diameter of Fmk w.r.t. L2
n
h
X
E
j=1
sup
i2
|Znj (y, d) − Znj (y , d )|
∗
∗
(y,d),(y ∗ ,d∗ )∈Fmk
n
i2
n X h
∗
∗
E
sup
I{εj ≤ y + d(j/n)} − I{εj ≤ y + d (j/n)} I{j/n ∈ In }
≤ 2 2
mn j=1
(y,d),(y ∗ ,d∗ )∈Fmk
n
i2
2X h
+
E
sup
I{εj ≤ y} − I{εj ≤ y ∗ }
n j=1
(y,d),(y ∗ ,d∗ )∈Fmk
24
≤ 3(2 + LF )η 2
for sufficiently large n. This shows that the bracketing number satisfies log N[ ] (η, F, Ln2 ) =
O(log M + log K) = O η −2/((1+δ)(α∧1)) , and the last displayed condition of Theorem 2.11.9
of van der Vaart and Wellner (1996) follows from δ > 1/α − 1.
It remains to show that (F, ρ) is totally bounded, i.e. that, for all η ∈ (0, 1), the space
F can be covered by finitely many sets with ρ-diameter less than 5η. To this end, choose
dLm and dUm as above. For each m ∈ {1, . . . , M } and j ∈ {0, . . . , J := dη −1 e}, let sj :=
jη 1/(α∧1) ∧ 1 and Fjm (y) := P (ε1 ≤ y + dLm (sj )), and choose an increasing sequence yjm,k ,
k = 1, . . . , K := bη −1 c, such that Fjm (yjm,k ) − Fjm (yjm,k−1 ) < η for all k ∈ {1, . . . , K + 1}
with yjm,0 := −∞ and yjm,K+1 := ∞. Denote by ȳl , 1 ≤ l ≤ L, all points yjm,k , j ∈
{0, . . . , J}, m ∈ {1, . . . , M }, k ∈ {1, . . . , K}, in increasing order. We show that all sets
Fml := {(y, d) | ȳl−1 ≤ y ≤ ȳl , dLm ≤ d ≤ dUm } have ρ-diameter less than 5η. Check that, for
all 1 ≤ l ≤ L, one has
sup
|F (ȳl + γ(x)) − F (ȳl−1 + γ(x))|
sup
x∈[0,1] γ∈C 1+δ ([0,1])
1
≤ max
sup
max
h
|F (ȳl + γ(x)) − F (ȳl + γ(sj ))|
sup
1≤j≤J sj−1 ≤x≤sj 1≤m≤M dL ≤γ≤dU
m
m
+|F (ȳl + γ(sj )) − F (ȳl + dLm (sj ))| + |F (ȳl + dLm (sj )) − F (ȳl−1 + dLm (sj ))|
i
+|F (ȳl−1 + dLm (sj )) − F (ȳl−1 + γ(sj ))| + |F (ȳl−1 + γ(sj )) − F (ȳl−1 + γ(x))|
h
i
< max (sj − sj−1 )α∧1 + η 2 + η + η 2 + (sj − sj−1 )α∧1
1≤j≤J
≤ 5η.
Therefore, for all (y, d), (y ∗ , d∗ ) ∈ Fml
ρ((y, d), (y ∗ , d∗ ))
n
≤ max sup
sup
x∈[0,1] γ∈C 1+δ ([0,1])
|F (ȳl + γ(x)) − F (ȳl−1 + γ(x))|, sup dUm (x) − dLm (x)
o
x∈[0,1]
1
≤ max{5η, η 2/(α∧1) } = 5η,
which concludes the proof of Theorem 3.4.
If we drop the assumption δ > 1/α − 1 but require F to be Lipschitz continuous, then
we use brackets for C11+δ ([0, 1]) of length η 2 (instead of η 2/(α∧1) ) and replace (A.7) with
sup |FmU (y)
y∈R
−
FmL (y)|
−1
≤ sup n
y∈R
n
X
|F (y + dUm (j/n)) − F (y + dLm (j/n))|
j=1
≤ LF sup |dUm (x) − dLm (x)| ≤ LF η 2
x∈R
with LF denoting the Lipschitz constant of F to prove log N[ ] (η, F, Ln2 ) = O η −2/(1+δ) ,
which again yields the third condition of Theorem 2.11.9 of van der Vaart and Wellner
25
(1996). Likewise, in the last part of the proof, one defines sj := jη ∧ 1 and replaces (A.8)
with max1≤j≤J (sj − sj−1 ) + η 2 + η + η 2 + (sj − sj−1 ) ≤ 5η.
2
In the remaining proofs to Section 3, we use the index n for the estimators to emphasis
the dependence on the sample size and to distinguish between estimators and polynomials
corresponding to a given sample on the one hand and corresponding objects in a limiting
setting on the other hand.
A.6
Proof of Lemma 3.6
Proposition A.1 and the proof of Lemma A.2 show that there exist constants d, d˜ > 0
˜
depending only on β and cg such that E(ĝn (x)) ≤ dE(M
n,0 ) and P {Mn,0 > t} ≤ 1 +
dnhn
dnhn (1 − F (−t)) (F (−t))
for all t > 0.
1/α
Let an := a(log n/(nhn ))
for a suitable constant a > 0 and fix some t0 > 0 such that
α
(1 − F (−t))/(ct ) ∈ (1/2, 2) for all t ∈ (0, t0 ]. Then
E(Mn,0 )
Z ∞
P {Mn,0 > t} dt
=
0
Z t0
Z
dnhn
≤ an +
1 + dnhn (1 − F (−t)) (F (−t))
dt + (1 + dnhn )
∞
((F (−t))dnhn dt.
t0
an
Now, for sufficiently large n,
Z t0
an
≤
≤
≤
≤
=
1 + dnhn (1 − F (−t)) (F (−t))dnhn dt
Z t0
c dnhn
dt
1 + 2cdnhn tα 1 − tα
2
an
Z t0
c
(1 + 2cd)nhn
tα exp − dnhn tα dt
2
an
Z tα0
c
t0
exp − dnhn u du
(1 + 2cd)nhn
α aαn
2
2(1 + 2cd)nhn t0
c
exp − daα log n
αcdnhn
2
−ξ
o(n )
for all ξ > 0 if a is chosen sufficiently large. Hence the assertion follows from (H1) and (F3)
which imply
Z ∞
Z ∞
dnhn
dnhn
+
t−dτ nhn dt = o(n−ξ )
(F (−t))
dt ≤ nhn ((F (−t0 ))
nhn
t0
2
for all ξ > 0.
26
A.7
Proof of Theorem 3.7
As the density f is bounded and Lipschitz continuous, one has
Z (g̃n∗ −g)(j/n)
∗
∗
f (y + t) − f (y) dt
F y + (g̃n − g)(j/n)) − F (y) − f (y)(g̃n − g)(j/n) =
0
= O (g̃n∗ − g)2 (j/n)
uniformly for y ≤ y0 and j/n ∈ In . Hence the remainder term can be approximated by a
sum of estimation errors as follows:
n
n
1 X
f (y) X ∗
∗
F y + (g̃n − g)(j/n)) − F (y) I{j/n ∈ In } −
(g̃ − g)(j/n)I{j/n ∈ In }
mn j=1
mn j=1 n
n
log n 2/α
1 X ∗
2β
2
2β
= O
(g̃n − g) (j/n)I{j/n ∈ In } = OP hn + bn +
= oP (n−1/2 )
mn j=1
nhn
where for the last conclusions we have used Theorem 2.2, Lemma 3.6 and the assumptions
(H2) and (B3). Thus the assertion follows if we show that
n
1 X ∗
(g̃ − g)(j/n)I{j/n ∈ In } = oP (n−1/2 ).
mn j=1 n
To this end, note that g̃n∗ (x) and g̃n∗ (y) are independent for |x − y| > 2(hn + bn ). For
simplicity, we assume that 2n(hn + bn ) =: kn is a natural number. If we split the whole sum
into blocks with kn consecutive summands, then all blocks with odd numbers are independent
and all blocks with even numbers are independent. It suffices to show that
1
mn
bn/(2kn )c
X
∆n,2`−1 = oP (n−1/2 )
`=1
1
mn
bn/(2kn )c
X
∆n,2` = oP (n−1/2 )
`=1
P(l+2)kn −1 ∗
where ∆n,l = j=(l+1)k
(g̃n − g)(j/n), 1 ≤ ` ≤ bn/kn c. We only consider the second sum,
n
because the first convergence obviously follows by the same arguments.
It suffices to verify
(A.8)
E ∆2n,2` = o(kn )
E ∆n,2` = o n−1/2 kn = o n1/2 (hn + bn )
(A.9)
uniformly for all 1 ≤ ` ≤ bn/(2kn )c, since then
E
bn/(2k
Xn )c
`=1
∆n,2`
2
bn/(2kn )c
=
X
V ar(∆n,2` ) +
`=1
bn/(2k
Xn )c
`=1
27
E(∆n,2` )
2
= o(n),
which implies the assertion.
To prove (A.8), note that according to Lemma 3.6, Proposition A.1 and the proofs of
Lemma A.2 and of Theorem 2.7(i), there exist constants c1 , c2 , c3 > 0 (depending only on β,
cg and the kernel K) such that
log n 1/α
+ max(M1∗ , M2∗ )
sup |g̃n∗ (x) − g(x)| ≤ c1 hβn + bβn +
nhn
x∈In
where M1∗ , M2∗ are independent random variables such that P {Mi∗ > t} ≤ 1 − (1 − P {Mn,0 >
t})c2 (hn +bn )/hn with
P {Mn,0 > t} ≤ 1 + c3 nhn (1 − F (−t)) (F (−t))c3 nhn .
1/2
Because kn hβn + bβn + (log n/(nhn ))1/α = o(kn ) by (H2) and (B3), it suffices to show that
Z ∞
∗ 2
E (Mi ) =
P {Mi∗ > t1/2 } dt = o(1/kn ).
(A.10)
0
Fix some t0 ∈ (0, (2c)−2/α ) such that (1 − F (−t))/(ctα ) ∈ (1/2, 2) for all t ∈ (0, t0 ]. In
what follows, d denotes a generic constant (depending only on β, cg , c and K) which may
vary from line to line. Applying the inequalities exp(−2ρu) ≤ (1 − u)ρ ≤ exp(−ρu), which
holds for all ρ > 0 and u ∈ (0, 1/2), we obtain for (nhn / log n)−2/α < t ≤ t0 and sufficiently
large n
c (hn +bn )/hn
P {Mi∗ > t1/2 } ≤ 1 − 1 − (1 + c3 nhn 2ctα/2 )(1 − ctα/2 /2)c3 nhn 2
≤ 1 − 1 − 3c3 cnhn tα/2 exp − c3 cnhn tα/2 /2
≤ 1 − exp − dn(hn + bn )tα/2 exp − c3 cnhn tα/2 /2
≤ dn(hn + bn )tα/2 exp − c3 cnhn tα/2 /2 .
Therefore, for sufficiently large a > 0,
Z
t20
P {Mi∗ > t1/2 } dt
0
≤ a
nh −2/α
n
log n
Z
t0
tα/2−1 exp − c3 cnhn tα/2 /2 dt
+ dt0 n(hn + bn )
a(nhn / log n)−2/α
≤ o(1/(n(hn + bn ))) + dt0 n(hn + bn ) exp − c3 caα/2 log n/2
= o(1/(n(hn + bn )))
(A.11)
where in the last but one step we apply the conditions (B3) and (H2). Now, assertion
(A.10) (and hence (A.8)) follows from
Z ∞
Z ∞
h
ic2 (hn +bn )/hn
∗
1/2
1/2 c3 nhn
1 − 1 − c3 nhn (F (−t ))
dt
P {Mi > t } dt ≤
t20
t20
28
Z
∞
1/2
1 − exp − dn(hn + bn )(F (−t
Z
c3 nhn
≤ dn(hn + bn ) nhn (F (−t0 ))
+
≤
c3 nhn
))
dt
t20
∞
t−τ c3 nhn /2 dt
nhn
−ξ
= o(n )
for all ξ > 0 and sufficiently large n, where we have used (H2) and (F3).
To establish (A.9), first note that for a kernel K of order d + 1 with d := bβc
E(g̃n (x) − g(x)) = E
Z
1
ĝn (x + bn u) −
−1
Z
d
X
g (j) (x)
j=0
j!
(bn u)j K(u) du
1
=
E(ĝn (x + bn u) − g(x + bn u))K(u) du + O(bβn )
−1
uniformly for all x ∈ [hn + bn , 1 − hn − bn ]. In view of (K1), (H2) and (B3), it thus suffices
to show that
E(ĝn (x) − g(x)) − Eg≡0 (ĝn (1/2)) = E(ĝn (x) − g(x)) − Eg≡0 (ĝn (x)) = o(n−1/2 ) (A.12)
uniformly for Lebesgue almost all x ∈ [hn , 1 − hn ]. Note that the distribution of ĝn (x) does
not depend on x if g equals 0.
Recall that ĝn (x) = p̃n (0) where p̃n is a polynomial on [−1, 1] of degree d that solves the
linear optimization problem
Z 1
p̃n (t) dt → min!
−1
under the constraints
p̃n
i/n − x
hn
≥ Yi ,
∀ i ∈ [n(x − hn ), n(x + hn )].
Define polynomials
d
X
1 (k)
qx (t) :=
g (x)(hn t)k ,
k!
k=0
pn (t) := (nhn )1/α (p̃n (t) − qx (t)),
t ∈ [−1, 1].
Then qx ((u − x)/hn ) is the Taylor expansion of order d of g(u) at x and the estimation error
can be written as
ĝn (x) − g(x) = (nhn )−1/α pn (0).
(A.13)
Note that pn is a polynomial of degree d that solves the linear optimization problem
Z
1
pn (t) dt → min!
−1
29
subject to
pn
i/n − x
hn
≥ (nhn )1/α ε̄i ,
∀ i ∈ [n(x − hn ), n(x + hn )],
(A.14)
with
ε̄i := εi + g(i/n) − qx
i/n − x
hn
.
We now use point process techniques to analyze the asymptotic behavior of this linear
program.
Denote by
X
Nn :=
δ
((i/n − x)/hn , (nhn )1/α ε̄i )
i∈[n(x−hn ),n(x+hn )]
a point process of standardized error random variables. Then the constraints (A.14) can be
reformulated as Nn (Apn ) = 0 where Af := {(t, u) ∈ [−1, 1] × R | u > f (t)} denotes the open
epigraph of a function f .
Since by (H2) |ε̄i − εi | = g(in ) − qx ((i/n − x)/hn )) = O(hβn ) = o((nhn )−1/α ) uniformly
for all i ∈ [n(x − hn ), n(x + hn )], one has
E Nn ([−1, 1] × (−1, ∞))) ∼ 2nhn P ε̄1 > −(nhn )−1/α → 2c.
Therefore, Nn converges weakly to a Poisson process N on [−1, 1]×R with intensity measure
2cU[−1,1] ⊗να where να has Lebesgue density x 7→ α|x|α−1 I(−∞, 0) (see, e.g., Resnick (2007),
Theorem 6.3). By Skorohod’s representation theorem, we may assume that the convergence
holds a.s.
Next we analyze the corresponding linear program in the limiting model to minimize
R1
p(t) dt over polynomials of degree d subject to N (Ap ) = 0. In what follows we use a
−1
P
representation of the Poisson process as N = ∞
i=1 δ(Ti ,Zi ) where Ti are independent random
variables which are uniformly distributed on [−1, 1].
First we prove by contradiction that the optimal solution is almost surely unique. Suppose
that there exist more than one solution. From the theory of linear programs it is known
that then there exists a solution p such that J := {j ∈ N | p(Tj ) = Zj } has at most d
elements. Because p is bounded and N has a.s. finitely many points in any bounded set,
η := inf{|p(Ti ) − Zi | | i ∈ N \ J} > 0 a.s. Since p is an optimal solution, all polynomials ∆
R1
of degree d such that ∆(Tj ) = 0, j ∈ J, and k∆k∞ < η must satisfy −1 ∆(t) dt = 0, because
both p + ∆ and p − ∆ satisfy the constraints N (Ap±∆ ) = 0. In particular, for all polynomials
Q
q of degree d − |J|, ∆(t) = τ i∈J (t − Ti )q(t) is of that type if τ > 0 is sufficiently small.
Q
P|J|−1
Write i∈J (t − Ti ) in the form t|J| + l=0 al tl . Then necessarily
Z
1
Y
(t − Ti )tj dt =
−1 i∈J
|J|−1
X
2
2al
I{|J| + j even} +
I{l + j even} = 0,
|J| + j + 1
l
+
j
+
1
l=0
30
for all j ∈ {0, . . . d − |J|}. This implies that (Ti )i∈J lies on a manifold M|J|,d of dimension
|J| − (d − |J| + 1) = 2|J| − d − 1 which only depends on |J| and d. However, by Proposition
A.1, kpk∞ ≤ Kd Zmax where
Zmax := max min{|Zi | | Ti ∈ [−1 + (j − 1)/jd , −1 + j/jd ]}.
1≤i≤jd
The above conclusion contradicts P {Zmax > K} → 0 as K → ∞, since
P ∃J ⊂ N : |J| ≤ d, (Tj )j∈J ∈ M|J|,d , max |Zj | ≤ Kd K = 0
j∈J
for all K > 0 (i.e., the fact that among finitely many values Ti a.s. there does not exist a
subset which lies on a given manifold of lower dimension).
Therefore the solution p must be a.s. unique which in turn implies that it is a basic
feasible solution, i.e., |J| ≥ d + 1. On the other hand, because the intensity measure of N is
absolutely continuous, |J| ≤ d + 1 a.s. and thus |J| = d + 1. Because of Nn → N a.s., one
has Nn ([−1, 1] × [−Kd Zmax , ∞)) = N ([−1, 1] × [−Kd Zmax , ∞)) =: M for sufficiently large n.
Moreover, one can find a numeration of the points (Tn,i , Zn,i ), 1 ≤ i ≤ M , of Nn and (Ti , Zi ),
1 ≤ i ≤ M , of N in [−1, 1] × [−Kd Zmax , ∞) such that (Tn,i , Zn,i ) → (Ti , Zi ).
R1
Next we prove that the solution to the linear program to minimize −1 pn (t) dt subject
to Nn (Apn ) = 0 is eventually unique with pn → p a.s. Since any optimal solution can be
written as a convex combination of basic feasible solutions, w.l.o.g. we may assume that
Jn := {1 ≤ i ≤ M | pn (Tn,i ) = Zn,i } has at least d + 1 elements. The polynomial pn is
uniquely determined by this set Jn . Suppose that along a subsequence n0 the set Jn0 is
constant, but not equal to J. Then p0n converges uniformly to the polynomial p̄ of degree
d that is uniquely determined by the conditions p̄(Ti ) = Zi for all i ∈ Jn0 . In particular,
p̄ is different from the unique optimal polynomial p for the limit Poisson process, but it
R1
R1
satisfies the constraints N (Ap ) = 0. Thus −1 p̄(t) dt > −1 p(t) dt. On the other hand, for
all η > 0 the polynomial p + η eventually satisfies the constraints Nn (Ap+η ) = 0 and thus
R1
R1
p(t)
+
η
dt
≥
p̄ (t) dt, which leads to a contradiction.
−1
−1 n
Hence, Jn = J for all sufficiently large n and the optimal solution pn for Nn is unique and
it converges uniformly to the optimal solution p for the Poisson process N . Moreover, using
the relation (pn (Tn,j ))j∈J = (Zn,j )j∈J (which is a system of linear equation in the coefficients
of pn ), pn (0) can be calculated as wnt (Zn,j )j∈J for some vector wn which converges to a limit
vector w (corresponding to the analogous relation for p).
Exactly the same arguments apply if we replace ε̄i with εi , which corresponds to the
case that g is identical 0. Since the points (T̃n,i , Z̃n,i ) of the pertaining point process equal
Tn,i , Zn,i − (nhn )1/α (g(i/n) − qx ((i/n) − x)/hn ) and thus |Z̃n,i − Zn,i | ≤ cg (nhn )1/α hβn , the
difference of the resulting values for optimal polynomial at 0 is bounded by a multiple of
(nhn )1/α hβn . In view of (A.13) and (H2), we may conclude that the difference between the
estimation errors can be bounded by a multiple of hβn = o(n−1/2 ), which finally yields (A.12)
and thus the assertion.
2
31
A.8
Proof of Corollary 4.3
bnsc
√ (Fbnsc (y)−Fn (y)) = En (y, s)−
En (y, 1) with En defined in (A.4). A similar
Note that bnsc
n
n
reasoning as in the proof of Theorem 3.1 (see (A.6)) shows that
ns̄
bnsc
n
sup
− 1 √ (Fbnsc (y) − Fn (y)) = oP (1).
bnsc
n
y∈R,s∈[0,1]
Hence, by Theorem 3.1, uniformly for all y ∈ R, s ∈ [0, 1],
√
n(F̂n (y, s) − s̄n F̂n (y))
√
√
√
= n(F̂n (y, s) − s̄n Fbnsc (y)) − s̄n n((F̂n (y) − Fn (y))) + s̄n n(Fbnsc (y) − Fn (y))
=
=
ns̄n bnsc
√ (Fbnsc (y) − Fn (y)) + oP (1)
bnsc n
En (y, 1) + oP (1)
En (y, s) − bnsc
n
= En (y, s) − sEn (y, 1) + oP (1)
which converges weakly to KF (y, s)−sKF (y, 1) for the Kiefer process KF defined in Theorem
3.1. Check that this Gaussian process has the same law as G(s, F (y)), because they have
the same covariance function. Thus the Kolmogorov-Smirnov statistic Tn converges weakly
to sups∈[0,1],y∈R |G(s, F (y)| = sups∈[0,1],z∈[0,1] |G(s, z)|, where the last equality holds by the
continuity of F .
2
A.9
Proof of Theorem 4.4
Note that under the given assumptions, the statements of Theorem 2.7 (i) and (ii) are valid
with rate oP (1). Let Ωn := {inf x∈In g̃ 0 (x) > 0}. From assumption (I1) and Theorem 2.7 (ii)
it follows that P (Ωn ) → 1 for n → ∞. But on Ωn the estimators g̃I and g̃ are identical,
and thus F̃I,bnsc = F̃bnsc . Now (4.1) can be concluded as in the proof of Theorem 3.1,
because Theorem 2.7 (i) yields supx∈In |g̃(x) − g(x)| = oP (n−1/(2(α∧1)) ). The convergence of
the Kolmogorov-Smirnov test statistic then follows exactly as in the proof of Corollary 4.3.
2
Acknowledgement
Financial support by the DFG (Research Unit FOR 1735 Structural Inference in Statistics:
Adaptation and Effciency) is gratefully acknowledged.
References
Akritas, M. and Van Keilegom, I. (2001). Nonparametric estimation of the residual distribution. Scand. J. Statist. 28, 549–567.
32
Anevski, D. and Fougères, A.-L. (2007). Limit properties of the monotone rearrangement
for density and regression function estimation. arXiv:0710.4617v1
Birke, M. and Neumeyer, N. (2013). Testing Monotonicity of Regression Functions - An
Empirical Process Approach. Scand. J. Statist. 40, 438–454.
Birke, M., Neumeyer, N. and Volgushev, S. (2016+). The independence process in conditional quantile location-scale models and an application to testing for monotonicity.
Statistica Sinica, to appear.
Chernozhukov, V., Fernández-Val, I. and Galichon, A. (2009). Improving point and interval
estimators of monotone functions by rearrangement. Biometrika 96, 559–575.
Daouia, A., Noh, H. and Park, B. U. (2016). Data envelope fitting with constrained polynomial splines. J. R. Stat. Soc. B. 78, 3–30.
Einmahl, J. H. J. and Van Keilegom, I. (2008). Specification tests in nonparametric regression. Journal of Econometrics 143, 88–102.
Färe, R. and Grosskopf, S. (1983). Measuring output efficiency. European Journal of
Operational Research 13, 173–179.
Gijbels, I. (2005). Monotone regression. In: N. Balakrishnan, S. Kotz, C.B. Read and B.
Vadakovic (eds), The Encyclopedia of Statistical Sciences, 2nd edition. Hoboken, NJ:
Wiley.
Gijbels, I., Mammen, E., Park, B. and Simar, L. (2000). On estimation of monotone and
concave frontier functions. J. Amer. Statist. Assoc. 94, 220–228.
Gijbels, I. and Peng, L. (2000). Estimation of a support curve via order statistics. Extremes
3, 251–277.
Girard, S. and Jacob, P. (2008). Frontier estimation via kernel regression on high powertransformed data. J. Multivariate Anal. 99, 403–420.
Girard, S., Guillou, A. and Stupfler, G. (2013). Frontier estimation with kernel regression
on high order moments. J. Multivariate Anal. 116, 172–189.
Hall, P., Park, B.U. and Stern, S.E. (1998). On polynomial estimators of frontiers and
boundaries. J. Multivariate Anal. 66, 71–98.
Hall, P. and Van Keilegom, I. (2009). Nonparametric “regression” when errors are positioned at end-points. Bernoulli 15, 614–633.
33
Härdle, W., Park, B.U. and Tsybakov, A.B. (1995). Estimation of non-sharp support
boundaries. J. Multivariate Anal. 55, 205–218.
Jirak, M., Meister, A. and Reiß, M. (2014). Adaptive estimation in nonparametric regression with one-sided errors. Ann. Statist. 42, 1970–2002.
Meister, A. and Reiß, M. (2013). Asymptotic equivalence for nonparametric regression with
non-regular errors. Probab. Th. Rel. Fields 155, 201–229.
Müller, U.U. and Wefelmeyer W. (2010). Estimation in Nonparametric Regression with
Non-Regular Errors. Comm. Statist. Theory Methods 39, 1619–1629.
Neumeyer, N. and Van Keilegom, I. (2009). Change-Point Tests for the Error Distribution
in Nonparametric Regression. Scand. J. Statist. 36, 518–541.
online supporting information available at
http://onlinelibrary.wiley.com/doi/10.1111/j.1467-9469.2009.00639.x/suppinfo
Picard, D. (1985). Testing and estimating change-points in time series. Adv. Appl. Probab.
17, 841–867.
Reiß, M. and Selk, L. (2016+). Efficient nonparametric functional estimation for one-sided
regression. Bernoulli, to appear.
Resnick, S.I. (2007). Heavy-Tail Phenomena. Springer.
Simar, L. and Wilson, P.W. (1998). Sensitivity analysis of efficiency scores: how to bootstrap in nonparametric frontier models. Management Science 44, 49–61.
Stephens, M.A. (1976). Asymptotic Results for Goodness-of-Fit Statistics with Unknown
Parameters. Ann. Statist. 4, 357–369.
van der Vaart, A.W. (2000). Asymptotic Statistics. Cambridge University Press.
van der Vaart, A. W. and Wellner, J. A. (1996). Weak Convergence and Empirical Processes. Springer, New York.
Wilson, P.W. (2003). Testing independence in models of productive efficiency. Journal of
Productivity Analysis 20, 361–390.
34
| 10math.ST
|
Published as a conference paper at ICLR 2018
arXiv:1703.02000v7 [cs.LG] 30 Jan 2018
ACTIVATION M AXIMIZATION G ENERATIVE A DVER SARIAL N ETS
Yuxuan Song, Kan Ren
Shanghai Jiao Tong University
songyuxuan,[email protected]
Zhiming Zhou, Han Cai
Shanghai Jiao Tong University
heyohai,[email protected]
Shu Rong
Yitu Tech
[email protected]
Jun Wang
University College London
[email protected]
Weinan Zhang, Yu Yong
Shanghai Jiao Tong University
[email protected], [email protected]
A BSTRACT
Class labels have been empirically shown useful in improving the sample quality
of generative adversarial nets (GANs). In this paper, we mathematically study the
properties of the current variants of GANs that make use of class label information.
With class aware gradient and cross-entropy decomposition, we reveal how class
labels and associated losses influence GAN’s training. Based on that, we propose
Activation Maximization Generative Adversarial Networks (AM-GAN) as an advanced solution. Comprehensive experiments have been conducted to validate our
analysis and evaluate the effectiveness of our solution, where AM-GAN outperforms other strong baselines and achieves state-of-the-art Inception Score (8.91)
on CIFAR-10. In addition, we demonstrate that, with the Inception ImageNet
classifier, Inception Score mainly tracks the diversity of the generator, and there is,
however, no reliable evidence that it can reflect the true sample quality. We thus
propose a new metric, called AM Score, to provide more accurate estimation on
the sample quality. Our proposed model also outperforms the baseline methods in
the new metric.
1
I NTRODUCTION
Generative adversarial nets (GANs) (Goodfellow et al., 2014) as a new way for learning generative
models, has recently shown promising results in various challenging tasks, such as realistic image
generation (Nguyen et al., 2016b; Zhang et al., 2016; Gulrajani et al., 2017), conditional image
generation (Huang et al., 2016b; Cao et al., 2017; Isola et al., 2016), image manipulation (Zhu et al.,
2016) and text generation (Yu et al., 2016).
Despite the great success, it is still challenging for the current GAN models to produce convincing
samples when trained on datasets with high variability, even for image generation with low resolution,
e.g., CIFAR-10. Meanwhile, people have empirically found taking advantages of class labels can
significantly improve the sample quality.
There are three typical GAN models that make use of the label information: CatGAN (Springenberg,
2015) builds the discriminator as a multi-class classifier; LabelGAN (Salimans et al., 2016) extends
the discriminator with one extra class for the generated samples; AC-GAN (Odena et al., 2016) jointly
trains the real-fake discriminator and an auxiliary classifier for the specific real classes. By taking
the class labels into account, these GAN models show improved generation quality and stability.
However, the mechanisms behind them have not been fully explored (Goodfellow, 2016).
In this paper, we mathematically study GAN models with the consideration of class labels. We derive
the gradient of the generator’s loss w.r.t. class logits in the discriminator, named as class-aware
gradient, for LabelGAN (Salimans et al., 2016) and further show its gradient tends to guide each
generated sample towards being one of the specific real classes. Moreover, we show that AC-GAN
(Odena et al., 2016) can be viewed as a GAN model with hierarchical class discriminator. Based on
1
Published as a conference paper at ICLR 2018
the analysis, we reveal some potential issues in the previous methods and accordingly propose a new
method to resolve these issues.
Specifically, we argue that a model with explicit target class would provide clearer gradient guidance
to the generator than an implicit target class model like that in (Salimans et al., 2016). Comparing
with (Odena et al., 2016), we show that introducing the specific real class logits by replacing the
overall real class logit in the discriminator usually works better than simply training an auxiliary
classifier. We argue that, in (Odena et al., 2016), adversarial training is missing in the auxiliary
classifier, which would make the model more likely to suffer mode collapse and produce low quality
samples. We also experimentally find that predefined label tends to result in intra-class mode collapse
and correspondingly propose dynamic labeling as a solution. The proposed model is named as
Activation Maximization Generative Adversarial Networks (AM-GAN). We empirically study the
effectiveness of AM-GAN with a set of controlled experiments and the results are consistent with our
analysis and, note that, AM-GAN achieves the state-of-the-art Inception Score (8.91) on CIFAR-10.
In addition, through the experiments, we find the commonly used metric needs further investigation.
In our paper, we conduct a further study on the widely-used evaluation metric Inception Score
(Salimans et al., 2016) and its extended metrics. We show that, with the Inception Model, Inception
Score mainly tracks the diversity of generator, while there is no reliable evidence that it can measure
the true sample quality. We thus propose a new metric, called AM Score, to provide more accurate
estimation on the sample quality as its compensation. In terms of AM Score, our proposed method
also outperforms other strong baseline methods.
The rest of this paper is organized as follows. In Section 2, we introduce the notations and formulate
the LabelGAN (Salimans et al., 2016) and AC-GAN∗ (Odena et al., 2016) as our baselines. We
then derive the class-aware gradient for LabelGAN, in Section 3, to reveal how class labels help its
training. In Section 4, we reveal the overlaid-gradient problem of LabelGAN and propose AM-GAN
as a new solution, where we also analyze the properties of AM-GAN and build its connections to
related work. In Section 5, we introduce several important extensions, including the dynamic labeling
as an alternative of predefined labeling (i.e., class condition), the activation maximization view and
a technique for enhancing the AC-GAN∗ . We study Inception Score in Section 6 and accordingly
propose a new metric AM Score. In Section 7, we empirically study AM-GAN and compare it to the
baseline models with different metrics. Finally we conclude the paper and discuss the future work in
Section 8.
2
P RELIMINARIES
In the original GAN formulation (Goodfellow et al., 2014), the loss functions of the generator G and
the discriminator D are given as:
Lori
G = −Ez∼pz (z) [log Dr (G(z))] , −Ex∼G [log Dr (x)],
Lori
D = −Ex∼pdata [log Dr (x)] − Ex∼G [log(1 − Dr (x))],
(1)
where D performs binary classification between the real and the generated samples and Dr (x)
represents the probability of the sample x coming from the real data.
2.1
L ABEL GAN
The framework (see Eq. (1)) has been generalized to multi-class case where each sample x has
its associated class label y ∈ {1, . . ., K, K+1}, and the K+1th label corresponds to the generated
samples (Salimans et al., 2016). Its loss functions are defined as:
Llab
G = − Ex∼G [log
Llab
D
PK
, −Ex∼G [log Dr (x)],
(2)
= − E(x,y)∼pdata [log Dy (x)] − Ex∼G [log DK+1 (x)],
(3)
i=1 Di (x)]
where Di (x) denotes the probability of the sample x being class i. The loss can be written in the
form of cross-entropy, which will simplify our later analysis:
Llab
G = Ex∼G [H([1, 0], [Dr (x), DK+1 (x)])],
(4)
Llab
D
(5)
= E(x,y)∼pdata [H(v(y), D(x))] + Ex∼G [H(v(K+1), D(x))],
where D(x) = [D1 (x), D2 (x), ..., DK+1 (x)] and v(y) = [v1 (y), . . . , vK+1
P (y)] with vi (y) = 0 if i 6=
y and vi (y) = 1 if i = y. H is the cross-entropy, defined as H(p, q)=− i pi log qi . We would refer
the above model as LabelGAN (using class labels) throughout this paper.
2
Published as a conference paper at ICLR 2018
AC-GAN∗
2.2
Besides extending the original two-class discriminator as discussed in the above section, Odena et al.
(2016) proposed an alternative approach, i.e., AC-GAN, to incorporate class label information, which
introduces an auxiliary classifier C for real classes in the original GAN framework. With the core
idea unchanged, we define a variant of AC-GAN as the following, and refer it as AC-GAN∗ :
Lac
G (x, y) = E(x,y)∼G H [1, 0], [Dr (x), Df (x)]
+ E(x,y)∼G H u(y), C(x) ,
Lac
+ E(x,y)∼G H [0, 1], [Dr (x), Df (x)]
D (x, y) = E(x,y)∼pdata H [1, 0], [Dr (x), Df (x)]
+ E(x,y)∼pdata H u(y), C(x) ,
(6)
(7)
(8)
(9)
where Dr (x) and Df (x) = 1 − Dr (x) are outputs of the binary discriminator which are the same as
vanilla GAN, u(·) is the vectorizing operator that is similar to v(·) but defined with K classes, and
C(x) is the probability distribution over K real classes given by the auxiliary classifier.
In AC-GAN, each sample has a coupled target class y, and a loss on the auxiliary classifier w.r.t. y is
added to the generator to leverage the class label information. We refer the losses on the auxiliary
classifier, i.e., Eq. (7) and (9), as the auxiliary classifier losses.
The above formulation is a modified version of the original AC-GAN. Specifically, we omit the
auxiliary classifier loss E(x,y)∼G [H(u(y), C(x))] which encourages the auxiliary classifier C to
classify the fake sample x to its target class y. Further discussions are provided in Section 5.3. Note
that we also adopt the − log(Dr (x)) loss in generator.
3
C LASS -AWARE G RADIENT
In this section, we introduce the class-aware gradient, i.e., the gradient of the generator’s loss w.r.t.
class logits in the discriminator. By analyzing the class-aware gradient of LabelGAN, we find that the
gradient tends to refine each sample towards being one of the classes, which sheds some light on how
the class label information helps the generator to improve the generation quality. Before delving into
the details, we first introduce the following lemma on the gradient properties of the cross-entropy
loss to make our analysis clearer.
Lemma 1. With l being the logits vector and σ being the softmax function, let σ(l) be the current
softmax probability distribution and p̂ denote the target probability distribution, then
−
∂H p̂, σ(l)
= p̂ − σ(l).
∂l
(10)
For a generated sample x, the loss in LabelGAN is Llab
G (x) = H([1, 0], [Dr (x), DK+1 (x)]), as
defined in Eq. (4). With Lemma 1, the gradient of Llab
(x)
w.r.t. the logits vector l(x) is given as:
G
Dk (x)
∂H [1, 0], [Dr (x), DK+1 (x)] ∂lr (x)
∂Llab
G (x)
−
=−
= 1 − Dr (x)
, k ∈ {1, . . . , K},
∂lk (x)
∂lr (x)
∂lk (x)
Dr (x)
∂H [1, 0], [Dr (x), DK+1 (x)]
∂Llab
G (x)
−
=−
= 0 − DK+1 (x) = − 1 − Dr (x) .
(11)
∂lK+1 (x)
∂lK+1 (x)
With the above equations, the gradient of Llab
G (x) w.r.t. x is:
K
−
X ∂Llab (x) ∂lk (x)
∂Llab
∂Llab
G (x)
G (x) ∂lK+1 (x)
=
− G
−
∂x
∂lk (x) ∂x
∂lK+1 (x)
∂x
k=1
K
X lab
X
K+1
Dk (x) ∂lk (x)
∂lK+1 (x)
∂lk (x)
= 1 − Dr (x)
−
= 1 − Dr (x)
αk (x)
,
Dr (x) ∂x
∂x
∂x
k=1
(12)
k=1
where
(D
αklab (x)
=
k (x)
Dr (x)
k ∈ {1, . . . , K}
−1
k = K+1
3
.
(13)
Published as a conference paper at ICLR 2018
Class 1
Class 2
Final Gradient
Gradient 1
Gradient 2
Generated Sample
Figure 1: An illustration of the overlaid-gradient problem. When two or more classes are encouraged
at the same time, the combined gradient may direct to none of these classes. It could be addressed by
assigning each generated sample a specific target class instead of the overall real class.
From the formulation, we find that the overall gradient w.r.t. a generated example x is 1−Dr (x),
which is the same as that in vanilla GAN (Goodfellow et al., 2014). And the gradient on real classes
is further distributed to each specific real class logit lk (x) according to its current probability ratio
Dk (x)
Dr (x) .
As such, the gradient naturally takes the label information into consideration: for a generated sample,
higher probability of a certain class will lead to a larger step towards the direction of increasing the
corresponding confidence for the class. Hence, individually, the gradient from the discriminator for
each sample tends to refine it towards being one of the classes in a probabilistic sense.
That is, each sample in LabelGAN is optimized to be one of the real classes, rather than simply to be
real as in the vanilla GAN. We thus regard LabelGAN as an implicit target class model. Refining
each generated sample towards one of the specific classes would help improve the sample quality.
Recall that there are similar inspirations in related work. Denton et al. (2015) showed that the result
could be significantly better if GAN is trained with separated classes. And AC-GAN (Odena et al.,
2016) introduces an extra loss that forces each sample to fit one class and achieves a better result.
4
T HE P ROPOSED M ETHOD
In LabelGAN, the generator gets its gradients from the K specific real class logits in discriminator
and tends to refine each sample towards being one of the classes. However, LabelGAN actually
suffers from the overlaid-gradient problem: all real class logits are encouraged at the same time.
Though it tends to make each sample be one of these classes during the training, the gradient of each
sample is a weighted averaging over multiple label predictors. As illustrated in Figure 1, the averaged
gradient may be towards none of these classes.
In multi-exclusive classes setting, each valid sample should only be classified to one of classes by
the discriminator with high confidence. One way to resolve the above problem is to explicitly assign
each generated sample a single specific class as its target.
4.1
AM-GAN
Assigning each sample a specific target class y, the loss functions of the revised-version LabelGAN
can be formulated as:
Lam
G = E(x,y)∼G [H(v(y), D(x))],
(14)
Lam
D
(15)
= E(x,y)∼pdata [H(v(y), D(x))] + Ex∼G [H(v(K+1), D(x))],
where v(y) is with the same definition as in Section 2.1. The model with aforementioned formulation
is named as Activation Maximization Generative Adversarial Networks (AM-GAN) in our paper.
And the further interpretation towards naming will be in Section 5.2. The only difference between
AM-GAN and LabelGAN lies in the generator’s loss function. Each sample in AM-GAN has a
specific target class, which resolves the overlaid-gradient problem.
AC-GAN (Odena et al., 2016) also assigns each sample a specific target class, but we will show that
the AM-GAN and AC-GAN are substantially different in the following part of this section.
4
Published as a conference paper at ICLR 2018
Figure 2: AM-GAN (left) v.s. AC-GAN∗ (right). AM-GAN can be viewed as a combination of
LabelGAN and auxiliary classifier, while AC-GAN∗ is a combination of vanilla GAN and auxiliary
classifier. AM-GAN can naturally conduct adversarial training among all the classes, while in
AC-GAN∗ , adversarial training is only conducted at the real-fake level and missing in the auxiliary
classifier.
4.2
L ABEL GAN + AUXILIARY C LASSIFIER
Both LabelGAN and AM-GAN are GAN models with K+1 classes. We introduce the following
cross-entropy decomposition lemma to build their connections to GAN models with two classes and
the K-classes models (i.e., the auxiliary classifiers).
PK
Lemma 2. Given v = [v1 , . . . , vK+1 ], v1:K , [v1 , . . . , vK ], vr ,
k=1 vk , R(v) , v1:K /vr and
F (v) , [vr , vK+1 ], let p̂ = [p̂1 , . . . , p̂K+1 ], p = [p1 , . . . , pK+1 ], then we have
H p̂, p = p̂r H R(p̂), R(p) + H F (p̂), F (p) .
(16)
With Lemma 2, the loss function of the generator in AM-GAN can be decomposed as follows:
Lam
+ H F v(x) , F D(x) .
G (x) = H v(x), D(x) = vr (x) · H R v(x) , R D(x)
|
{z
}
{z
}
|
Auxiliary Classifier G Loss
(17)
LabelGAN G Loss
The second term of Eq. (17) actually equals to the loss function of the generator in LabelGAN:
H F v(x) , F D(x) = H 1, 0 , Dr (x), DK+1 (x) = Llab
G (x).
(18)
Similar analysis can be adapted to the first term and the discriminator. Note that vr (x) equals to
one. Interestingly, we find by decomposing the AM-GAN losses, AM-GAN can be viewed as a
combination of LabelGAN and auxiliary classifier (defined in Section 2.2). From the decomposition
perspective, disparate to AM-GAN, AC-GAN is a combination of vanilla GAN and the auxiliary
classifier.
The auxiliary classifier loss in Eq. (17) can also be viewed as the cross-entropy version of generator
loss in CatGAN: the generator of CatGAN directly optimizes entropy H(R(D(x))) to make each
sample have a high confidence of being one of the classes, while AM-GAN achieves this by the
first term of its decomposed loss H(R(v(x)), R(D(x))) in terms of cross-entropy with given target
distribution. That is, the AM-GAN is the combination of the cross-entropy version of CatGAN and
LabelGAN. We extend the discussion between AM-GAN and CatGAN in the Appendix B.
4.3
N ON -H IERARCHICAL M ODEL
With the Lemma 2, we can also reformulate the AC-GAN∗ as a K+1 classes model. Take the
generator’s loss function as an example:
Lac
G (x, y) = E(x,y)∼G H [1, 0], [Dr (x), Df (x)] + H u(y), C(x)
= E(x,y)∼G H v(y), [Dr (x) · C(x), Df (x)] .
(19)
In the K+1 classes model, the K+1 classes distribution is formulated as [Dr (x) · C(x), Df (x)].
AC-GAN introduces the auxiliary classifier in the consideration of leveraging the side information
5
Published as a conference paper at ICLR 2018
of class label, it turns out that the formulation of AC-GAN∗ can be viewed as a hierarchical K+1
classes model consists of a two-class discriminator and a K-class auxiliary classifier, as illustrated in
Figure 2. Conversely, AM-GAN is a non-hierarchical model. All K+1 classes stay in the same level
of the discriminator in AM-GAN.
In the hierarchical model AC-GAN∗ , adversarial training is only conducted at the real-fake twoclass level, while misses in the auxiliary classifier. Adversarial training is the key to the theoretical
guarantee of global convergence pG = pdata . Taking the original GAN formulation as an instance, if
generated samples collapse to a certain point x, i.e., pG (x) > pdata (x), then there must exit another
pdata (x)
point x0 with pG (x0 ) < pdata (x0 ). Given the optimal D(x) = pG (x)+p
, the collapsed point x will
data (x)
get a relatively lower score. And with the existence of higher score points (e.g. x0 ), maximizing the
generator’s expected score, in theory, has the strength to recover from the mode-collapsed state. In
practice, the pG and pdata are usually disjoint (Arjovsky & Bottou, 2017), nevertheless, the general
behaviors stay the same: when samples collapse to a certain point, they are more likely to get a
relatively lower score from the adversarial network.
Without adversarial training in the auxiliary classifier, a mode-collapsed generator would not get any
penalties from the auxiliary classifier loss. In our experiments, we find AC-GAN is more likely to
get mode-collapsed, and it was empirically found reducing the weight (such as 0.1 used in Gulrajani
et al. (2017)) of the auxiliary classifier losses would help. In Section 5.3, we introduce an extra
adversarial training in the auxiliary classifier with which we improve AC-GAN∗ ’s training stability
and sample-quality in experiments. On the contrary, AM-GAN, as a non-hierarchical model, can
naturally conduct adversarial training among all the class logits.
5
5.1
E XTENSIONS
DYNAMIC L ABELING
In the above section, we simply assume each generated sample has a target class. One possible solution
is like AC-GAN (Odena et al., 2016), predefining each sample a class label, which substantially
results in a conditional GAN. Actually, we could assign each sample a target class according to its
current probability estimated by the discriminator. A natural choice could be the class which is of the
maximal probability currently: y(x) , argmaxi∈{1,...,K} Di (x) for each generated sample x. We
name this dynamic labeling.
According to our experiments, dynamic labeling brings important improvements to AM-GAN, and is
applicable to other models that require target class for each generated sample, e.g. AC-GAN, as an
alternative to predefined labeling.
We experimentally find GAN models with pre-assigned class label tend to encounter intra-class mode
collapse. In addition, with dynamic labeling, the GAN model remains generating from pure random
noises, which has potential benefits, e.g. making smooth interpolation across classes in the latent
space practicable.
5.2
T HE ACTIVATION M AXIMIZATION V IEW
Activation maximization is a technique which is traditionally applied to visualize the neuron(s) of
pretrained neural networks (Nguyen et al., 2016a;b; Erhan et al., 2009).
The GAN training can be viewed as an Adversarial Activation Maximization Process. To be more
specific, the generator is trained to perform activation maximization for each generated sample on
the neuron that represents the log probability of its target class, while the discriminator is trained to
distinguish generated samples and prevents them from getting their desired high activation.
It is worth mentioning that the sample that maximizes the activation of one neuron is not necessarily
of high quality. Traditionally people introduce various priors to counter the phenomenon (Nguyen
et al., 2016a;b). In GAN, the adversarial process of GAN training can detect unrealistic samples
and thus ensures the high-activation is achieved by high-quality samples that strongly confuse the
discriminator.
We thus name our model the Activation Maximization Generative Adversarial Network (AM-GAN).
6
Published as a conference paper at ICLR 2018
5.3
AC-GAN∗+
Experimentally we find AC-GAN easily get mode collapsed and a relatively low weight for the
auxiliary classifier term in the generator’s loss function would help. In the Section 4.3, we attribute
mode collapse to the miss of adversarial training in the auxiliary classifier. From the adversarial
activation maximization view: without adversarial training, the auxiliary classifier loss that requires
high activation on a certain class, cannot ensure the sample quality.
That is, in AC-GAN, the vanilla GAN loss plays the role for ensuring sample quality and avoiding
mode collapse. Here we introduce an extra loss to the auxiliary classifier in AC-GAN∗ to enforce
adversarial training and experimentally find it consistently improve the performance:
Lac+
D (x, y) = E(x,y)∼G H u(·), C(x) ,
(20)
where u(·) represents the uniform distribution, which in spirit is the same as CatGAN (Springenberg,
2015).
Recall that we omit the auxiliary classifier loss E(x,y)∼G H u(y)] in AC-GAN∗ . According to our
experiments, E(x,y)∼G [H(u(y)] does improve AC-GAN∗ ’s stability and make it less likely to get
mode collapse, but it also leads to a worse Inception Score. We will report the detailed results in
Section 7. Our understanding on this phenomenon is that: by encouraging the auxiliary classifier also
to classify fake samples to their target classes, it actually reduces the auxiliary classifier’s ability on
providing gradient guidance towards the real classes, and thus also alleviates the conflict between the
GAN loss and the auxiliary classifier loss.
6
E VALUATION M ETRICS
One of the difficulties in generative models is the evaluation methodology (Theis et al., 2015). In this
section, we conduct both the mathematical and the empirical analysis on the widely-used evaluation
metric Inception Score (Salimans et al., 2016) and other relevant metrics. We will show that Inception
Score mainly works as a diversity measurement and we propose the AM Score as a compensation to
Inception Score for estimating the generated sample quality.
6.1
I NCEPTION S CORE
As a recently proposed metric for evaluating the performance of generative models, Inception Score
has been found well correlated with human evaluation (Salimans et al., 2016), where a publiclyavailable Inception model C pre-trained on ImageNet is introduced. By applying the Inception
model to each generated sample x and getting the corresponding class probability distribution C(x),
Inception Score is calculated via
Inception Score = exp Ex KL C(x) k C̄ G
,
(21)
where Ex is short of Ex∼G and C̄ G = Ex [C(x)] is the overall probability distribution of the generated
samples over classes, which
is judged by C,and KL denotes the Kullback-Leibler divergence. As
proved in Appendix D, Ex KL C(x) k C̄ G can be decomposed into two terms in entropy:
Ex KL(C(x) k C̄ G ) = H(C̄ G ) + (−Ex H C(x) ).
6.2
(22)
T HE P ROPERTIES OF I NCEPTION M ODEL
A common understanding of how Inception Score works lies in that a high score in the first term
H(C̄ G ) indicates the generated samples have high diversity (the overall class probability distribution
evenly distributed), and a high score in the second term −Ex [H(C(x))] indicates that each individual
sample has high quality (each generated sample’s class probability distribution is sharp, i.e., it can be
classified into one of the real classes with high confidence) (Salimans et al., 2016).
However, taking CIFAR-10 as an illustration, the data are not evenly distributed over the classes
under the Inception model trained on ImageNet, which is presented in Figure 4a. It makes Inception
Score problematic in the view of the decomposed scores, i.e., H(C̄ G ) and −Ex [H(C(x))]. Such as
that one would ask whether a higher H(C̄ G ) indicates a better mode coverage and whether a smaller
H(C(x)) indicates a better sample quality.
7
5.80
5.75
E[H(C(x))]
9.0
8.5
8.0
7.5
7.0
6.5
6.0
5.5
5.00
H(C̄ G )
Inception Score
Published as a conference paper at ICLR 2018
5.70
5.65
5.60
0 10 20 30 40 50 60 70 80
Iterations
10 20 30 40 50 60 70 80
Iterations
(a)
4.1
4.0
3.9
3.8
3.7
3.6
3.5
3.40 10 20 30 40 50 60 70 80
Iterations
(b)
(c)
200
400 600 800
ImageNet Classes
(a)
1000
2000
5000
1500
4000
Number of Samples
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0.000
Number of Samples
Avg Probability Density
Figure 3: Training curves of Inception Score and its decomposed terms. a) Inception Score, i.e.
exp(H(C̄ G ) − Ex [H(C(x))]); b) H(C̄ G ); c) Ex [H(C(x))]. A common understanding of Inception
Score is that: the value of H(C̄ G ) measures the diversity of generated samples and is expected to
increase in the training process. However, it usually tends to decrease in practice as illustrated in (c).
1000
500
00
1
2
3
4
5
6
H(C(x)) with ImageNet Classifier
7
(b)
airplane
automobile
bird
cat
deer
dog
frog
horse
ship
truck
3000
2000
1000
0
0.00
0.01 0.02 0.03 0.04 0.05
H(C(x)) with CIFAR-10 Classifier
(c)
Figure 4: Statistics of the CIFAR-10 training images. a) C̄ G over ImageNet classes; b) H(C(x))
distribution with ImageNet classifier of each class; c) H(C(x)) distribution with CIFAR-10 classifier
of each class. With the Inception model, the value of H(C(x)) score of CIFAR-10 training data is
variant, which means, even in real data, it would still strongly prefer some samples than some others.
H(C(x)) on a classifier that pre-trained on CIFAR-10 has low values for all CIFAR-10 training data
and thus can be used as an indicator of sample quality.
We experimentally find that, as in Figure 3b, the value of H(C̄ G ) is usually going down during
the training process, however, which is expected to increase. And when we delve into the detail of
H(C(x)) for each specific sample in the training data, we find the value of H(C(x)) score is also
variant, as illustrated in Figure 4b, which means, even in real data, it would still strongly prefer some
samples than some others. The exp operator in Inception Score and the large variance of the value of
H(C(x)) aggravate the phenomenon. We also observe the preference on the class level in Figure 4b,
e.g., Ex [H(C(x))]=2.14 for trucks, while Ex [H(C(x))]=3.80 for birds.
It seems, for an ImageNet Classifier, both the two indicators of Inception Score cannot work correctly.
Next we will show that Inception Score actually works as a diversity measurement.
6.3
I NCEPTION S CORE AS A D IVERSITY M EASUREMENT
Since the two individual indicators are strongly correlated, here we go back to Inception Score’s
original formulation Ex [KL(C(x) k C̄ G )]. In this form, we could interpret Inception Score as that it
requires each sample’s distribution C(x) highly different from the overall distribution of the generator
C̄ G , which indicates a good diversity over the generated samples.
As is empirically observed, a mode-collapsed generator usually gets a low Inception Score. In an
extreme case, assuming all the generated samples collapse to a single point, then C(x)=C G and we
would get the minimal Inception Score 1.0, which is the exp result of zero. To simulate mode collapse
in a more complicated case, we design synthetic experiments as following: given a set of N points
{x0 , x1 , x2 , ..., xN −1 }, with each point xi adopting the distribution C(xi ) = v(i) and representing
class i, where v(i) is the vectorization operator of length N , as defined in Section 2.1, we randomly
drop m points, evaluate Ex [KL(C(x) k C̄ G )] and draw the curve. As is showed in Figure 5, when
N − m increases, the value of Ex [KL(C(x) k C̄ G )] monotonically increases in general, which means
that it can well capture the mode dropping and the diversity of the generated distributions.
8
5
5
4
4
log (Inception Score)
log (Inception Score)
Published as a conference paper at ICLR 2018
3
2
1
00
20
40
60
80
Number of Kept Classes
3
2
1
00
100
20
40
60
80
Number of Kept Classes
(a)
100
(b)
0.35
0.10
0.25
0.30
0.08
0.20
0.20
0.15
0.06
E[H(C(x))]
0.25
KL(C̄ T , C̄ G )
AM Score
Figure 5: Mode dropping analysis of Inception Score. a) Uniform density over classes; b) Gaussian
density over classes. The value of Ex [KL(C(x) k C̄ G )] monotonically increases in general as the
number of kept classes increases, which illustrates Inception Score is able to capture the mode
dropping and the diversity of the generated distributions. The error bar indicates the min and max
values in 1000 random dropping.
0.04
0.15
0.10
0.02
0.10
0.050 10 20 30 40 50 60 70 80
Iterations
0.000 10 20 30 40 50 60 70 80
Iterations
0.050 10 20 30 40 50 60 70 80
Iterations
(a)
(b)
(c)
Figure 6: Training curves of AM Score and its decomposed terms. a) AM Score, i.e. KL(C̄ train , C̄ G )+
Ex [H(C(x))]; b) KL(C̄ train , C̄ G ); c) Ex [H(C(x))]. All of them works properly (going down) in the
training process.
One remaining question is that whether good mode coverage and sample diversity mean high quality
of the generated samples. From the above analysis, we do not find any evidence. A possible
explanation is that, in practice, sample diversity is usually well correlated with the sample quality.
6.4
AM S CORE WITH ACCORDINGLY P RETRAINED C LASSIFIER
Note that if each point xi has multiple variants such as x1i , x2i , x3i , one of the situations, where x2i and
x3i are missing and only x1 is generated, cannot be detected by Ex [KL(C(x) k C̄ G )] score. It means
that with an accordingly pretrained classifier, Ex [KL(C(x) k C̄ G )] score cannot detect intra-class
level mode collapse. This also explains why the Inception Network on ImageNet could be a good
candidate C for CIFAR-10. Exploring the optimal C is a challenge problem and we shall leave it as a
future work.
However, there is no evidence that using an Inception Network trained on ImageNet can accurately
measure the sample quality, as shown in Section 6.2. To compensate Inception Score, we propose to
introduce an extra assessment using an accordingly pretrained classifier. In the accordingly pretrained
classifier, most real samples share similar H(C(x)) and 99.6% samples hold scores less than 0.05 as
showed in Figure 4c, which demonstrates that H(C(x)) of the classifier can be used as an indicator
of sample quality.
The entropy term on C̄ G is actually problematic when training data is not evenly distributed over
classes, for that argmin H(C̄ G ) is a uniform distribution. To take the C̄ train into account, we replace
H(C̄ G ) with a KL divergence between C̄ train and C̄ G . So that
AM Score , KL(C̄ train , C̄ G ) + Ex H C(x) ,
(23)
which requires C̄ G close to C̄ train and each sample x has a low entropy C(x). The minimal value of
AM Score is zero, and the smaller value, the better. A sample training curve of AM Score is showed
in Figure 6, where all indicators in AM Score work as expected. 1
1
Inception Score and AM Score measure the diversity and quality of generated samples, while FID (Heusel
et al., 2017) measures the distance between the generated distribution and the real distribution.
9
Published as a conference paper at ICLR 2018
Model
GAN
GAN∗
AC-GAN∗
AC-GAN∗+
LabelGAN
AM-GAN
Inception Score
CIFAR-10
Tiny ImageNet
dynamic
predefined
dynamic
predefined
7.04 ± 0.06 7.27 ± 0.07
7.25 ± 0.07 7.31 ± 0.10
7.41 ± 0.09 7.79 ± 0.08
7.28 ± 0.07 7.89 ± 0.11
8.56 ± 0.11 8.01 ± 0.09 10.25 ± 0.14 8.23 ± 0.10
8.63 ± 0.08 7.88 ± 0.07 10.82 ± 0.16 8.62 ± 0.11
8.83 ± 0.09 8.35 ± 0.12 11.45 ± 0.15 9.55 ± 0.11
AM Score
CIFAR-10
Tiny ImageNet
dynamic
predefined
dynamic
predefined
0.45 ± 0.00 0.43 ± 0.00
0.40 ± 0.00 0.41 ± 0.00
0.17 ± 0.00 0.16 ± 0.00 1.64 ± 0.02 1.01 ± 0.01
0.10 ± 0.00 0.14 ± 0.00 1.04 ± 0.01 1.20 ± 0.01
0.13 ± 0.00 0.25 ± 0.00 1.11 ± 0.01 1.37 ± 0.01
0.08 ± 0.00 0.05 ± 0.00 0.88 ± 0.01 0.61 ± 0.01
Table 1: Inception Score and AM Score Results. Models in the same column share the same network
structures & hyper-parameters. We applied dynamic / predefined labeling for models that require
target classes.
dynamic
predefined
AC-GAN∗
0.61
0.35
AC-GAN∗+
0.39
0.36
LabelGAN
0.35
0.32
AM-GAN
0.36
0.36
Table 2: The maximum value of mean MS-SSIM of various models over the ten classes on CIFAR-10.
High-value indicates obvious intra-class mode collapse. Please refer to the Figure 11 in the Appendix
for the visual results.
7
E XPERIMENTS
To empirically validate our analysis and the effectiveness of the proposed method, we conduct
experiments on the image benchmark datasets including CIFAR-10 and Tiny-ImageNet2 which
comprises 200 classes with 500 training images per class. For evaluation, several metrics are used
throughout our experiments, including Inception Score with the ImageNet classifier, AM Score with a
corresponding pretrained classifier for each dataset, which is a DenseNet (Huang et al., 2016a) model.
We also follow Odena et al. (2016) and use the mean MS-SSIM (Wang et al., 2004) of randomly
chosen pairs of images within a given class, as a coarse detector of intra-class mode collapse.
A modified DCGAN structure, as listed in the Appendix F, is used in experiments. Visual results of
various models are provided in the Appendix considering the page limit, such as Figure 9, etc. The
repeatable experiment code is published for further research3 .
7.1
E XPERIMENTS ON CIFAR-10
7.1.1
GAN WITH AUXILIARY C LASSIFIER
The first question is whether training an auxiliary classifier without introducing correlated losses to
the generator would help improve the sample quality. In other words, with the generator only with
the GAN loss in the AC-GAN∗ setting. (referring as GAN∗ )
As is shown in Table 1, it improves GAN’s sample quality, but the improvement is limited comparing
to the other methods. It indicates that introduction of correlated loss plays an essential role in the
remarkable improvement of GAN training.
7.1.2
C OMPARISON A MONG D IFFERENT M ODELS
The usage of the predefined label would make the GAN model transform to its conditional version,
which is substantially disparate with generating samples from pure random noises. In this experiment,
we use dynamic labeling for AC-GAN∗ , AC-GAN∗+ and AM-GAN to seek for a fair comparison
among different discriminator models, including LabelGAN and GAN. We keep the network structure
and hyper-parameters the same for different models, only difference lies in the output layer of the
discriminator, i.e., the number of class logits, which is necessarily different across models.
As is shown in Table 1, AC-GAN∗ achieves improved sample quality over vanilla GAN, but sustains
mode collapse indicated by the value 0.61 in MS-SSIM as in Table 2. By introducing adversarial
2
3
https://tiny-imagenet.herokuapp.com/
Link for anonymous experiment code: https://github.com/ZhimingZhou/AM-GAN
10
Published as a conference paper at ICLR 2018
Model
Score ± Std.
DFM (Warde-Farley & Bengio, 2017)
7.72 ± 0.13
Improved GAN (Salimans et al., 2016)
8.09 ± 0.07
AC-GAN (Odena et al., 2016)
8.25 ± 0.07
WGAN-GP + AC (Gulrajani et al., 2017) 8.42 ± 0.10
SGAN (Huang et al., 2016b)
8.59 ± 0.12
AM-GAN (our work)
8.91 ± 0.11
Splitting GAN (Guillermo et al., 2017)
8.87 ± 0.09
Real data
11.24 ± 0.12
Table 3: Inception Score comparison on CIFAR-10. Splitting GAN uses the class splitting technique
to enhance the class label information, which is orthogonal to AM-GAN.
training in the auxiliary classifier, AC-GAN∗+ outperforms AC-GAN∗ . As an implicit target class
model, LabelGAN suffers from the overlaid-gradient problem and achieves a relatively higher per
sample entropy (0.124) in the AM Score, comparing to explicit target class model AM-GAN (0.079)
and AC-GAN∗+ (0.102). In the table, our proposed AM-GAN model reaches the best scores against
these baselines.
1
We also test AC-GAN∗ with decreased weight on auxiliary classifier losses in the generator ( 10
relative to the GAN loss). It achieves 7.19 in Inception Score, 0.23 in AM Score and 0.35 in MS-SSIM.
The 0.35 in MS-SSIM indicates there is no obvious mode collapse, which also conform with our
above analysis.
7.1.3
I NCEPTION S CORE C OMPARING WITH R ELATED W ORK
AM-GAN achieves Inception Score 8.83 in the previous experiments, which significantly outperforms
the baseline models in both our implementation and their reported scores as in Table 3. By further
enhancing the discriminator with more filters in each layer, AM-GAN also outperforms the orthogonal
work (Guillermo et al., 2017) that enhances the class label information via class splitting. As the
result, AM-GAN achieves the state-of-the-art Inception Score 8.91 on CIFAR-10.
7.1.4
DYNAMIC L ABELING AND C LASS C ONDITION
It’s found in our experiments that GAN models with class condition (predefined labeling) tend to
encounter intra-class mode collapse (ignoring the noise), which is obvious at the very beginning of
GAN training and gets exasperated during the process.
In the training process of GAN, it is important to ensure a balance between the generator and the
discriminator. With the same generator’s network structures and switching from dynamic labeling to
class condition, we find it hard to hold a good balance between the generator and the discriminator:
to avoid the initial intra-class mode collapse, the discriminator need to be very powerful; however, it
usually turns out the discriminator is too powerful to provide suitable gradients for the generator and
results in poor sample quality.
Nevertheless, we find a suitable discriminator and conduct a set of comparisons with it. The results
can be found in Table 1. The general conclusion is similar to the above, AC-GAN∗+ still outperforms
AC-GAN∗ and our AM-GAN reaches the best performance. It’s worth noticing that the AC-GAN∗
does not suffer from mode collapse in this setting.
In the class conditional version, although with fine-tuned parameters, Inception Score is still relatively
low. The explanation could be that, in the class conditional version, the sample diversity still tends to
decrease, even with a relatively powerful discriminator. With slight intra-class mode collapse, the
per-sample-quality tends to improve, which results in a lower AM Score. A supplementary evidence,
P
not very strict, of partial mode collapse in the experiments is that: the | ∂G(z)
∂z | is around 45.0 in
dynamic labeling setting, while it is 25.0 in the conditional version.
The LabelGAN does not need explicit labels and the model is the same in the two experiment settings.
But please note that both Inception Score and the AM Score get worse in the conditional version. The
only difference is that the discriminator becomes more powerful with an extended layer, which attests
11
9
1.0
8
0.8
7
6
50
10
20
30 40
Iterations
AM Score
Inception Score
Published as a conference paper at ICLR 2018
AM-GAN
AC-GAN*+
AC-GAN*
LAB-GAN
GAN
GAN*
50
AM-GAN
AC-GAN*+
AC-GAN*
LAB-GAN
GAN
GAN*
0.6
0.4
0.2
0.00
60
(a) Inception Score training curves
10
20
30 40
Iterations
50
60
(b) AM Score training curves
Figure 7: The training curves of different models in the dynamic labeling setting.
that the balance between the generator and discriminator is crucial. We find that, without the concern
of intra-class mode collapse, using the dynamic labeling makes the balance between generator and
discriminator much easier.
7.1.5
T HE E(x,y)∼G [H(u(y), C(x))] LOSS
Note that we report results of the modified version of AC-GAN, i.e., AC-GAN∗ in Table 1. If we take
the omitted loss E(x,y)∼G [H(u(y), C(x))] back to AC-GAN∗ , which leads to the original AC-GAN
(see Section 2.2), it turns out to achieve worse results on both Inception Score and AM Score on
CIFAR-10, though dismisses mode collapse. Specifically, in dynamic labeling setting, Inception
Score decreases from 7.41 to 6.48 and the AM Score increases from 0.17 to 0.43, while in predefined
class setting, Inception Score decreases from 7.79 to 7.66 and the AM Score increases from 0.16 to
0.20.
This performance drop might be because we use different network architectures and hyper-parameters
from AC-GAN (Odena et al., 2016). But we still fail to achieve its report Inception Score, i.e., 8.25,
on CIFAR-10 when using the reported hyper-parameters in the original paper. Since they do not
publicize the code, we suppose there might be some unreported details that result in the performance
gap. We would leave further studies in future work.
7.1.6
T HE L EARNING P ROPERTY
We plot the training curve in terms of Inception Score and AM Score in Figure 7. Inception Score and
AM Score are evaluated with the same number of samples 50k, which is the same as Salimans et al.
(2016). Comparing with Inception Score, AM Score is more stable in general. With more samples,
Inception Score would be more stable, however the evaluation of Inception Score is relatively costly.
A better alternative of the Inception Model could help solve this problem.
The AC-GAN∗ ’s curves appear stronger jitter relative to the others. It might relate to the counteract
between the auxiliary classifier loss and the GAN loss in the generator. Another observation is that
the AM-GAN in terms of Inception Score is comparable with LabelGAN and AC-GAN∗+ at the
beginning, while in terms of AM Score, they are quite distinguishable from each other.
7.2
E XPERIMENTS ON T INY-I MAGE N ET
In the CIFAR-10 experiments, the results are consistent with our analysis and the proposed method
outperforms these strong baselines. We demonstrate that the conclusions can be generalized with
experiments in another dataset Tiny-ImageNet.
The Tiny-ImageNet consists with more classes and fewer samples for each class than CIFAR-10,
which should be more challenging. We downsize Tiny-ImageNet samples from 64×64 to 32×32
and simply leverage the same network structure that used in CIFAR-10, and the experiment result is
showed also in Table 1. From the comparison, AM-GAN still outperforms other methods remarkably.
And the AC-GAN∗+ gains better performance than AC-GAN∗ .
12
Published as a conference paper at ICLR 2018
8
C ONCLUSION
In this paper, we analyze current GAN models that incorporate class label information. Our analysis
shows that: LabelGAN works as an implicit target class model, however it suffers from the overlaidgradient problem at the meantime, and explicit target class would solve this problem. We demonstrate
that introducing the class logits in a non-hierarchical way, i.e., replacing the overall real class logit in
the discriminator with the specific real class logits, usually works better than simply supplementing an
auxiliary classifier, where we provide an activation maximization view for GAN training and highlight
the importance of adversarial training. In addition, according to our experiments, predefined labeling
tends to lead to intra-class mode collapsed, and we propose dynamic labeling as an alternative. Our
extensive experiments on benchmarking datasets validate our analysis and demonstrate our proposed
AM-GAN’s superior performance against strong baselines. Moreover, we delve deep into the widelyused evaluation metric Inception Score, reveal that it mainly works as a diversity measurement. And
we also propose AM Score as a compensation to more accurately estimate the sample quality.
In this paper, we focus on the generator and its sample quality, while some related work focuses on
the discriminator and semi-supervised learning. For future work, we would like to conduct empirical
studies on discriminator learning and semi-supervised learning. We extend AM-GAN to unlabeled
data in the Appendix C, where unsupervised and semi-supervised is accessible in the framework of
AM-GAN. The classifier-based evaluation metric might encounter the problem related to adversarial
samples, which requires further study. Combining AM-GAN with Integral Probability Metric based
GAN models such as Wasserstein GAN (Arjovsky et al., 2017) could also be a promising direction
since it is orthogonal to our work.
R EFERENCES
Arjovsky, Martin and Bottou, Léon. Towards principled methods for training generative adversarial
networks. In ICLR, 2017.
Arjovsky, Martin, Chintala, Soumith, and Bottou, Léon.
arXiv:1701.07875, 2017.
Wasserstein gan.
arXiv preprint
Cao, Yun, Zhou, Zhiming, Zhang, Weinan, and Yu, Yong. Unsupervised diverse colorization via
generative adversarial networks. arXiv preprint, 2017.
Che, Tong, Li, Yanran, Jacob, Athul Paul, Bengio, Yoshua, and Li, Wenjie. Mode regularized
generative adversarial networks. arXiv preprint arXiv:1612.02136, 2016.
Denton, Emily L, Chintala, Soumith, Fergus, Rob, et al. Deep generative image models using a
laplacian pyramid of adversarial networks. In Advances in neural information processing systems,
pp. 1486–1494, 2015.
Erhan, Dumitru, Bengio, Yoshua, Courville, Aaron, and Vincent, Pascal. Visualizing higher-layer
features of a deep network. University of Montreal, 1341:3, 2009.
Goodfellow, Ian. Nips 2016 tutorial:
arXiv:1701.00160, 2016.
Generative adversarial networks.
arXiv preprint
Goodfellow, Ian, Pouget-Abadie, Jean, Mirza, Mehdi, Xu, Bing, Warde-Farley, David, Ozair, Sherjil,
Courville, Aaron, and Bengio, Yoshua. Generative adversarial nets. In Advances in neural
information processing systems, pp. 2672–2680, 2014.
Guillermo, L. Grinblat, Lucas, C. Uzal, and Pablo, M. Granitto. Class-splitting generative adversarial
networks. arXiv preprint arXiv:1709.07359, 2017.
Gulrajani, Ishaan, Ahmed, Faruk, Arjovsky, Martin, Dumoulin, Vincent, and Courville, Aaron.
Improved training of wasserstein gans. arXiv preprint arXiv:1704.00028, 2017.
Heusel, Martin, Ramsauer, Hubert, Unterthiner, Thomas, Nessler, Bernhard, Klambauer, Günter, and
Hochreiter, Sepp. Gans trained by a two time-scale update rule converge to a nash equilibrium.
arXiv preprint arXiv:1706.08500, 2017.
13
Published as a conference paper at ICLR 2018
Huang, Gao, Liu, Zhuang, Weinberger, Kilian Q, and van der Maaten, Laurens. Densely connected
convolutional networks. arXiv preprint arXiv:1608.06993, 2016a.
Huang, Xun, Li, Yixuan, Poursaeed, Omid, Hopcroft, John, and Belongie, Serge. Stacked generative
adversarial networks. arXiv preprint arXiv:1612.04357, 2016b.
Isola, Phillip, Zhu, Jun-Yan, Zhou, Tinghui, and Efros, Alexei A. Image-to-image translation with
conditional adversarial networks. arXiv preprint arXiv:1611.07004, 2016.
Karras, Tero, Aila, Timo, Laine, Samuli, and Lehtinen, Jaakko. Progressive growing of gans for
improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017.
Mao, Xudong, Li, Qing, Xie, Haoran, Lau, Raymond YK, Wang, Zhen, and Smolley, Stephen Paul.
Least squares generative adversarial networks. arXiv preprint ArXiv:1611.04076, 2016.
Nguyen, Anh, Dosovitskiy, Alexey, Yosinski, Jason, Brox, Thomas, and Clune, Jeff. Synthesizing
the preferred inputs for neurons in neural networks via deep generator networks. In Advances in
Neural Information Processing Systems, pp. 3387–3395, 2016a.
Nguyen, Anh, Yosinski, Jason, Bengio, Yoshua, Dosovitskiy, Alexey, and Clune, Jeff. Plug & play
generative networks: Conditional iterative generation of images in latent space. arXiv preprint
arXiv:1612.00005, 2016b.
Odena, Augustus, Olah, Christopher, and Shlens, Jonathon. Conditional image synthesis with
auxiliary classifier gans. arXiv preprint arXiv:1610.09585, 2016.
Salimans, Tim, Goodfellow, Ian, Zaremba, Wojciech, Cheung, Vicki, Radford, Alec, and Chen, Xi.
Improved techniques for training gans. In Advances in Neural Information Processing Systems,
pp. 2226–2234, 2016.
Springenberg, Jost Tobias. Unsupervised and semi-supervised learning with categorical generative
adversarial networks. arXiv preprint arXiv:1511.06390, 2015.
Szegedy, Christian, Vanhoucke, Vincent, Ioffe, Sergey, Shlens, Jon, and Wojna, Zbigniew. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, pp. 2818–2826, 2016.
Theis, Lucas, Oord, Aäron van den, and Bethge, Matthias. A note on the evaluation of generative
models. arXiv preprint arXiv:1511.01844, 2015.
Wang, Zhou, Simoncelli, Eero P, and Bovik, Alan C. Multiscale structural similarity for image quality
assessment. In Signals, Systems and Computers, 2004. Conference Record of the Thirty-Seventh
Asilomar Conference on, volume 2, pp. 1398–1402. IEEE, 2004.
Warde-Farley, D. and Bengio, Y. Improving generative adversarial networks with denoising feature
matching. In ICLR, 2017.
Yu, Lantao, Zhang, Weinan, Wang, Jun, and Yu, Yong. Seqgan: sequence generative adversarial nets
with policy gradient. arXiv preprint arXiv:1609.05473, 2016.
Zhang, Han, Xu, Tao, Li, Hongsheng, Zhang, Shaoting, Huang, Xiaolei, Wang, Xiaogang, and
Metaxas, Dimitris. Stackgan: Text to photo-realistic image synthesis with stacked generative
adversarial networks. arXiv preprint arXiv:1612.03242, 2016.
Zhu, Jun-Yan, Krähenbühl, Philipp, Shechtman, Eli, and Efros, Alexei A. Generative visual manipulation on the natural image manifold. In European Conference on Computer Vision, pp. 597–613.
Springer, 2016.
14
Published as a conference paper at ICLR 2018
A
A.1
G RADIENT VANISHING & − log(Dr (x)) & L ABEL S MOOTHING
L ABEL S MOOTHING
Label smoothing that avoiding extreme logits value was showed to be a good regularization (Szegedy
et al., 2016). A general version of label smoothing could be: modifying the target probability of
discriminator)
[λ1 , 1 − λ1 ] x ∼ G
D̂r (x), D̂f (x) =
.
(24)
[1 − λ2 , λ2 ] x ∼ pdata
Salimans et al. (2016) proposed to use only one-side label smoothing. That is, to only apply label
smoothing for real samples: λ1 = 0 and λ2 > 0. The reasoning of one-side label smoothing is
applying label smoothing on fake samples will lead to fake mode on data distribution, which is too
obscure.
We will next show the exact problems when applying label smoothing to fake samples along with the
log(1−Dr (x)) generator loss, in the view of gradient w.r.t. class logit, i.e., the class-aware gradient,
and we will also show that the problem does not exist when using the − log(Dr (x)) generator loss.
A.2
T HE log(1−Dr (x)) GENERATOR LOSS
The log(1−Dr (x)) generator loss with label smoothing in terms of cross-entropy is
Llog(1-D)
= −Ex∼G H [λ1 , 1 − λ1 ], [Dr (x), DK+1 (x)] ,
G
(25)
with lemma 1, its negative gradient is
−
∂Llog(1-D)
(x)
G
= Dr (x) − λ1 ,
∂lr (x)
Dr (x) = λ1
D (x) < λ1
r
Dr (x) > λ1
gradient vanishing
Dr (x) is optimized towards 0 .
Dr (x) is optimized towards 1
(26)
(27)
Gradient vanishing is a well know training problem of GAN. Optimizing Dr (x) towards 0 or 1 is
also not what desired, because the discriminator is mapping real samples to the distribution with
Dr (x) = 1 − λ2 .
A.3
T HE − log(Dr (x)) GENERATOR LOSS
The − log(Dr (x)) generator loss with target [1−λ, λ] in terms of cross-entropy is
L-log(D)
= Ex∼G H [1 − λ, λ], [Dr (x), DK+1 (x)] ,
G
(28)
the negative gradient of which is
−
∂L-log(D)
(x)
G
= (1 − λ) − Dr (x),
∂lr (x)
Dr (x) = 1 − λ stationary point
D (x) < 1 − λ Dr (x) towards 1 − λ .
r
Dr (x) > 1 − λ Dr (x) towards 1 − λ
(29)
(30)
Without label smooth λ, the − log(Dr (x)) always∗ preserves the same gradient direction as
log(1−Dr (x)) though giving a difference gradient scale. We must note that non-zero gradient
does not mean that the gradient is efficient or valid.
The both-side label smoothed version has a strong connection to Least-Square GAN (Mao et al.,
2016): with the fake logit fixed to zero, the discriminator maps real to α on the real logit and maps
fake to β on the real logit, the generator in contrast tries to map fake sample to α. Their gradient on
the logit are also similar.
15
Published as a conference paper at ICLR 2018
B
C AT GAN
The auxiliary classifier loss of AM-GAN can also be viewed as the cross-entropy version of CatGAN:
generator of CatGAN directly optimizes entropy H(R(D(x))) to make each sample be one class,
while AM-GAN achieves this by the first term of its decomposed loss H(R(v(x)), R(D(x))) in
terms of cross-entropy with given target distribution. That is, the AM-GAN is the cross-entropy
version of CatGAN that is combined with LabelGAN by introducing an additional fake class.
B.1
D ISCRIMINATOR LOSS ON FAKE SAMPLE
The discriminator of CatGAN maximizes the prediction entropy of each fake sample:
LCat’
D = Ex∼G − H D(x) .
(31)
In AM-GAN, as we have an extra class on fake, we can achieve this in a simpler manner by minimizing
the probability on real logits.
LAM’
(32)
D = Ex∼G H F (v(K+1)), F (D(x)) .
If vr (K+1) is not zero, that is, when we did negative label smoothing Salimans et al. (2016), we
could define R(v(K+1)) to be a uniform distribution.
LAM”
= Ex∼G H R(v(K+1)), R(D(x)) × vr (K+1).
(33)
D
As a result, the label smoothing part probability will be required to be uniformly distributed, similar
to CatGAN.
C
U NLABELED DATA
In this section, we extend AM-GAN to unlabeled data. Our solution is analogous to CatGAN
Springenberg (2015).
C.1
S EMI - SUPERVISED SETTING
Under semi-supervised setting, we can add the following loss to the original solution to integrate the
unlabeled data (with the distribution denoted as punl (x)):
Lunl’
(34)
D = Ex∼punl H v(x), D(x) .
C.2
U NSUPERVISED SETTING
Under unsupervised setting, we need to introduce one extra loss, analogy to categorical GAN
Springenberg (2015):
Lunl”
(35)
D = H pref , R(Ex∼punl [D(x)]) ,
where the pref is a reference label distribution for the prediction on unsupervised data. For example,
pref could be set as a uniform distribution, which requires the unlabeled data to make use of all the
candidate class logits.
This loss can be optionally added to semi-supervised setting, where the pref could be defined as the
predicted label distribution on the labeled training data Ex∼pdata [D(x)].
16
Published as a conference paper at ICLR 2018
D
I NCEPTION S CORE
As a recently proposed metric for evaluating the performance of the generative models, the InceptionScore has been found well correlated with human evaluation (Salimans et al., 2016), where a
pre-trained publicly-available Inception model C is introduced. By applying the Inception model to
each generated sample x and getting the corresponding class probability distribution C(x), Inception
Score is calculated via
Inception Score = exp Ex KL C(x) k C̄ G ,
(36)
where Ex is short of Ex∼G and C̄ G = Ex [C(x)] is the overall probability distribution of the generated
samples over classes, which is judged by C, and KL denotes the Kullback-Leibler divergence which
is defined as
P
P
P
KL(p k q) = i pi log pqii = i pi log pi − i pi log qi = −H(p) + H(p, q).
(37)
An extended metric, the Mode Score, is proposed in Che et al. (2016) to take the prior distribution of
the labels into account, which is calculated via
Mode Score = exp Ex KL C(x) k C̄ train − KL(C̄ G k C̄ train ) ,
(38)
where the overall class distribution from the training data C̄ train has been added as a reference. We
show in the following that, in fact, Mode Score and Inception Score are equivalent.
Lemma 3. Let p(x) be the class probability distribution of the sample x, and p̄ denote another
probability distribution, then
Ex H p(x), p̄
= H Ex p(x) , p̄ .
(39)
With Lemma 3, we have
log(Inception Score)
= Ex KL(C(x) k C̄ G )
= Ex H C(x), C̄ G − Ex H C(x)
= H Ex C(x) , C̄ G − Ex H C(x)
= H(C̄ G ) + (−Ex H C(x) ),
(40)
log(Mode Score)
= Ex KL C(x) k C̄ train − KL(C̄ G k C̄ train )
= Ex H C(x), C̄ train − Ex H C(x) − H(C̄ G , C̄ train ) + H(C̄ G )
= H(C̄ G ) + (−Ex H C(x) ).
(41)
17
Published as a conference paper at ICLR 2018
E
T HE L EMMA AND P ROOFS
Lemma 1. With l being the logits vector and σ being the softmax function, let σ(l) be the current
softmax probability distribution and p̂ denote any target probability distribution, then:
∂H p̂, σ(l)
−
= p̂ − σ(l).
∂l
(42)
Proof.
P
P
i)
∂ i p̂i log Pexp(l
∂H p̂, σ(l)
∂H p̂, σ(l)
∂ i p̂i log σ(l)i
j exp(lj )
=
=
−
=−
∂l
∂lk
∂lk
∂lk
k
P
P
P
P
∂ i p̂i li − log j exp(lj )
∂ log
∂ i p̂i li
exp(lk )
j exp(lj )
=
=
−
= p̂k − P
∂lk
∂lk
∂lk
j exp(lj )
∂H p̂, σ(l)
= p̂ − σ(l).
⇒−
∂l
PK
Lemma 2. Given v = [v1 , . . . , vK+1 ], v1:K , [v1 , . . . , vK ], vr ,
k=1 vk , R(v) , v1:K /vr and
F (v) , [vr , vK+1 ], let p̂ = [p̂1 , . . . , p̂K+1 ], p = [p1 , . . . , pK+1 ], then we have:
H p̂, p = p̂r H R(p̂), R(p) + H F (p̂), F (p) .
(43)
Proof.
H(p̂, p) = −
PK
k=1
p̂k log pk − p̂K+1 log pK+1 = −p̂r
p̂k
k=1 p̂r
= −p̂r
PK
= −p̂r
PK
p̂k
k=1 p̂r
(log
pk
pr
p̂k
k=1 p̂r
PK
p
log( pk pr ) − p̂K+1 log pK+1
r
+ log pr ) − p̂K+1 log pK+1
p
log pk − p̂r log pr − p̂K+1 log pK+1
r
= p̂r H R(p̂), R(p) + H F (p̂), F (p) .
Lemma 3. Let p(x) be the class probability distribution of the sample x that from a certain data
distribution, and p̄ denote the reference probability distribution, then
Ex H p(x), p̄
= H Ex p(x) , p̄ .
Proof.
P
P
Ex H p(x), p̄ = Ex − i pi (x) log p̄i = − i Ex [pi (x)] log p̄i
X
=−
Ex [p(x)] i log p̄i = H Ex p(x) , p̄ .
i
18
(44)
Published as a conference paper at ICLR 2018
(a)
(b)
(c)
Figure 8: H(C(x) of Inception Score in Real Images. a) 0<H(C(x)<1 ; b) 3<H(C(x)< 4; c)
6<H(C(x)<7.
F
N ETWORK S TRUCTURE & H YPER -PARAMETERS
Generator:
Operation
Kernel
Strides
Output Dims
Output Drop
Activation
Noise
N/A
N/A
100 | 110
0.0
N(0.0, 1.0)
Linear
N/A
N/A
4×4×768
0.0
Leaky ReLU
True
Deconvolution
3×3
2×2
8×8×384
0.0
Leaky ReLU
True
Deconvolution
3×3
2×2
16×16×192
0.0
Leaky ReLU
True
Deconvolution
3×3
2×2
32×32×96
0.0
Leaky ReLU
True
Deconvolution
3×3
1×1
32×32×3
0.0
Tanh
Kernel
Strides
Output Dims
Output Drop
Activation
Add Gaussian Noise
N/A
N/A
32×32×3
0.0
N(0.0, 0.1)
Convolution
3×3
1×1
32×32×64
0.3
Leaky ReLU
True
Convolution
3×3
2×2
16×16×128
0.3
Leaky ReLU
True
Convolution
3×3
2×2
8×8×256
0.3
Leaky ReLU
True
Convolution
3×3
2×2
4×4×512
0.3
Leaky ReLU
True
Convolution*
3×3
1×1
4×4×512
0.3
Leaky ReLU
True
AvgPool
N/A
N/A
1×1×512
0.3
N/A
Linear
N/A
N/A
10 | 11 | 12
0.0
Softmax
Discriminator:
Operation
The *layer was only used for class condition experiments
Optimizer: Adam with beta1=0.5, beta2=0.999; Batch size=100.
Learning rate: Exponential decay with stair, initial learning rate 0.0004.
We use weight normalization for each weight
19
BN
Published as a conference paper at ICLR 2018
Figure 9: Random Samples of AM-GAN: Dynamic Labeling
Figure 10: Random Samples of AM-GAN: Class Condition
20
Published as a conference paper at ICLR 2018
Figure 11: Random Samples of AC-GAN∗ : Dynamic Labeling
Figure 12: Random Samples of AC-GAN∗ : Class Condition
21
Published as a conference paper at ICLR 2018
Figure 13: Random Samples of AC-GAN∗+ : Dynamic Labeling
Figure 14: Random Samples of AC-GAN∗+ : Class Condition
22
Published as a conference paper at ICLR 2018
Figure 15: Random Samples of LabelGAN: Under Dynamic Labeling Setting
Figure 16: Random Samples of LabelGAN: Under Class Condition Setting
23
Published as a conference paper at ICLR 2018
Figure 17: Random Samples of GAN: Under Dynamic Labeling Setting
Figure 18: Random Samples of GAN: Under Class Condition Setting
24
| 2cs.AI
|
Graph sketching-based Space-efficient Data Clustering
Anne Morvan∗†‡
[email protected]
Krzysztof Choromanski§
[email protected]
Cédric Gouy-Pailler∗
[email protected]
arXiv:1703.02375v4 [cs.LG] 20 Dec 2017
Jamal Atif†
[email protected]
Abstract
In this paper, we address the problem of recovering
arbitrary-shaped data clusters from datasets while facing
high space constraints, as this is for instance the case in many
real-world applications when analysis algorithms are directly
deployed on resources-limited mobile devices collecting the
data. We present DBMSTClu a new space-efficient densitybased non-parametric method working on a Minimum Spanning Tree (MST) recovered from a limited number of linear
measurements i.e. a sketched version of the dissimilarity
graph G between the N objects to cluster. Unlike k-means,
k-medians or k-medoids algorithms, it does not fail at distinguishing clusters with particular forms thanks to the property of the MST for expressing the underlying structure of
a graph. No input parameter is needed contrarily to DBSCAN or the Spectral Clustering method. An approximate
MST is retrieved by following the dynamic semi-streaming
model in handling the dissimilarity graph G as a stream of
edge weight updates which is sketched in one pass over the
data into a compact structure requiring O(N polylog(N ))
space, far better than the theoretical memory cost O(N 2 ) of
G. The recovered approximate MST T as input, DBMSTClu then successfully detects the right number of nonconvex
clusters by performing relevant cuts on T in a time linear in
N . We provide theoretical guarantees on the quality of the
clustering partition and also demonstrate its advantage over
the existing state-of-the-art on several datasets.
1
Introduction
Clustering is one of the principal data mining tasks consisting in grouping related objects in an unsupervised
manner. It is expected that objects belonging to the
same cluster are more similar to each other than to objects belonging to different clusters. There exists a variety of algorithms performing this task. Methods like k∗ CEA,
LIST, 91191 Gif-sur-Yvette, France.
Paris-Dauphine, PSL Research University, CNRS,
LAMSADE, 75016 Paris, France.
‡ Partly supported by the DGA (French Ministry of Defense).
§ Google Brain Robotics, New York, USA.
† Université
means [1], k-medians [2] or k-medoids [3] are useful unless the number and the shape of clusters are unknown
which is unfortunately often the case in real-world applications. They are typically unable to find clusters with
a nonconvex shape. Although DBSCAN [4] does not
have these disadvantages, its resulting clustering still
depends on the chosen parameter values.
One of the successful approaches relies on a graph
representation of the data. Given a set of N data points
{x1 , . . . , xN }, a graph can be built based on the dissimilarity of data where points of the dataset are the vertices
and weighted edges express distances between these objects. Besides, the dataset can be already a graph G
modeling a network in many fields, such as bioinformatics - where gene-activation dependencies are described
through a network - or social, computer, information,
transportation network analysis. The clustering task
consequently aims at detecting clusters as groups of
nodes that are densely connected with each other and
sparsely connected to vertices of other groups. In this
context, Spectral Clustering (SC) [9] is a popular tool
to recover clusters with particular structures for which
classical k-means algorithm fails. When dealing with
large scale datasets, a main bottleneck of the technique
is to perform the partial eigendecomposition of the associated graph Laplacian matrix, though. Another inherent difficulty is to handle the huge number of nodes
and edges of the induced dissimilarity graph: storing all
edges can cost up to O(N 2 ) where N is the number of
nodes. Over the last decade, it has been established that
the dynamic streaming model [5] associated with linear
sketching techniques [6] - also suitable for distributed
processing -, is a good way for tackling this last issue. In
this paper, the addressed problem falls within a framework of storage limits allowing O(N polylog(N )) space
complexity but not O(N 2 ).
Contributions. The new clustering algorithm
DBMSTClu presented in this paper brings a solution
to these issues: 1) detecting arbitrary-shaped data clusters, 2) with no parameter, 3) in a time linear to the
number of points, 4) in a space-efficient manner by
Copyright c 2018 by SIAM
Unauthorized reproduction of this article is prohibited
working on a limited number of linear measurements, a
sketched version of the streamed dissimilarity graph G.
DBMSTClu returns indeed a partition of the N points
to cluster relying only on a Minimum Spanning Tree
(MST) of the dissimilarity graph G taking O(N ) space.
This MST can be space-efficiently approximatively retrieved in the dynamic semi-streaming model by handling G as a stream of edge weight updates sketched
in only one pass over the data into a compact structure taking O(N polylog(N )) space. DBMSTClu then
automatically identifies the right number of nonconvex
clusters by cutting suitable edges of the resulting approximate MST in O(N ) time.
The remaining of this paper is organized as follows.
In §2 the related work about graph clustering and other
space-efficient clustering algorithms is described. §3
gives fundamentals of the sketching technique that can
be used to obtain an approximate MST of the dissimilarity graph. DBMSTClu (DB for Density-Based), the
proposed MST-based algorithm for clustering is then
explained in §4.1, its theoretical guarantees discussed
in §4.2 and the implementation enabling its scalability
detailed in §4.3. §5 presents the experimental results
comparing the proposed clustering algorithm to other
existing methods. Finally, §6 concludes the work and
discusses future directions.
2
pected clusters, consists in deleting the heaviest edges
from the Euclidean MST of the considered graph but
this completely fails when the intra-cluster distance is
lower than the inter-clusters one. For decades since
MST-based clustering methods [20, 21] have been developed and can be classified into the group of densitybased methods. MSDR [21] relies on the mean and the
standard deviation of edge weights within clusters but
will encourage clusters with points far from each other
as soon as they are equally “far”. Moreover, it does not
handle clusters with less than three points.
2.3 Space-efficient
clustering
algorithms.
Streaming k-means [23] is a one-pass streaming method
for the k-means problem but still fails to detect clusters
with nonconvex shapes since only the centroid point
of each cluster is stored. This is not the case of
CURE algorithm [24] which represents each cluster as
a random sample of data points contained in it but
this offline method has a prohibitive time complexity
of O(N 2 log(N )) not suitable for large datasets. More
time-efficient, CluStream [25] and DenStream [26]
create microclusters based on local densities in an
online fashion and aggregate them later to build bigger
clusters in offline steps. Though, only DenStream can
capture non-spherical clusters but needs parameters
like DBSCAN from which it is inspired.
Related work
2.1 General graph clustering. The approach of
graph representation of the data has led to an extensive
literature over graph clustering related to graph partitioning [7]. From the clustering methods point of view,
DenGraph [10] proposes a graph version of DBSCAN
which is able to deal with noise while work from [11]
focuses on the problem of recovering clusters with considerably dissimilar sizes. Recent works include also approaches from convex optimization using low-rank decomposition of the adjacency matrix [12, 13, 14, 15].
These methods bring theoretical guarantees about the
exact recovery of the ground truth clustering for the
Stochastic Block Model [16, 17, 18] but demand to
compute the eigendecomposition of a N × N matrix
(resp. O(N 3 ) and O(N 2 ) for time and space complexity). Moreover they are restricted to unweighted graphs
(weights in the work from [15] refer to likeliness of existence of an edge, not a distance between points).
2.2 MST-based graph clustering. The MST is
known to help recognizing clusters with arbitrary
shapes. Clustering algorithms from this family identify
clusters by performing suitable cuts among the MST
edges. The first algorithm, called Standard Euclidean
MST (SEMST) is from [19] and given a number of ex-
3
Context, notations and graph sketching
preprocessing step
Consider a dataset with N points. Either the underlying network already exists, or it is assumed that a
dissimilarity graph G between points can be built where
points of the dataset are the vertices and weighted edges
express distances between these objects. For instance,
this can be the Euclidean distance. In both cases, the
graphs considered here should follow this definition:
Definition 3.1. (graph G = (V, E)) A graph G =
(V, E) consists in a set of nodes V and a set of edges
E ⊆ V × V . The graph is undirected but weighted.
The weight w on an edge between node i and j - if this
edge exists - corresponds to the normalized predefined
distance between i and j, s.t. 0 < w ≤ 1. |V | = N
and |E| = M stand resp. for the cardinality of sets V
and E. E = {e1 , ..., eM } and for all edges ei a weight
wi represents a distance between two vertices. In the
sequel, E(G) describes the set of edges of a graph G.
Freely of any parameter, DBMSTClu performs the
nodes clustering from its unique input: an MST of
G. So an independent preprocessing is required for
its recovery by any existing method. To respect our
space restrictions though, the use of a graph sketching
Copyright c 2018 by SIAM
Unauthorized reproduction of this article is prohibited
technique is motivated here: from the stream of its edge
weights, a sketch of G is built and then an approximate
MST is retrieved exclusively from it.
Streaming graph sketching and approximate MST recovery. Processing data in the dynamic
streaming model [5] for graph sketching implies: 1) The
graph should be handled as a stream s of edge weight
updates: s = (a1 , ... aj , ...) where aj is the j-th update in the stream corresponding to the tuple aj =
(i, wold,i , ∆wi ) with i denoting the index of the edge to
update, wold,i its previous weight and ∆wi the update to
perform. Thus, after reading aj in the stream, the i-th
edge is assigned the new weight wi = wold,i + ∆wi ≥ 0.
2) The method should make only one pass over this
stream . 3) Edges can be both inserted or deleted (turnstile model), i.e. weights can be increased or decreased
(but have always to be nonnegative). So weights change
regularly, as in social networks where individuals can be
friends for some time then not anymore.
The algorithm in [6] satisfies these conditions and is
used here to produce in an online fashion a limited number of linear measurements summarizing edge weights
of G, as new data aj are read from the stream s of
edge weight updates. Its general principle is briefly described here. For a given small 1 , G is seen as a set of
unweighted subgraphs Gk containing all the edges with
weight lower than (1 + 1 )k , hence Gk ⊂ Gk+1 . The Gk
are embodied as N virtual vectors v (i) ∈ {−1, 0, 1}M for
i ∈ [N ]1 expressing for each node the belonging to an
(i)
existing edge: for j ∈ [M ], vj equals to 0 if node i is not
in ej , 1 (resp. −1) if ej exists and i is its left (resp. right)
node. All v (i) are described at L different “levels”, i.e. L
virtual copies of the true vectors are made with some entries randomly set to zero s.t. the v (i),l get sparser as the
corresponding level l ∈ [L] increases. The v (i),l for each
level are explicitly coded in memory by three counters:
PM (i),l
PM
PM (i),l j
(i),l
φ = j=1 vj ; ι = j=1 j vj ; τ = j=1 vj
z
mod p, with p a suitably large prime and z ∈ Zp . The
resulting compact data structure further named a sketch
enables to draw almost uniformly at random a nonzero
weighted edge among Gk at any time among the levels vectors v (i),l which are 1-sparse (with exactly one
nonzero coefficient) thanks to `0 -sampling [27]:
practice only one pass over the data is needed and the
space cost is significantly lower than the theoretical
O(N 2 ) bound. The time cost for each update of
the sketch is polylog(N ). The authors from [6] also
proposed an algorithm to compute in a single-pass the
approximate weight W̃ of an MST T - the sum of all
its edge weights - by appropriate samplings from the
sketch in O(N polylog(N )) time. They show that W ≤
W̃ ≤ (1+1 ) W where W stands
Pr for the true weight and
W̃ = N −(1+1 )r+1 cc(Gr )+ k=0 λk cc(Gk ) with λk =
(1 + 1 )k+1 − (1 + 1 )i , r = dlog1+1 (wmax )e s.t. wmax is
the maximal weight of G and cc denotes the number
of connected components of the graph in parameter.
Here an extended method is applied for obtaining rather
an approximate MST - and not simply its weight - by
registering edges as they are sampled. Referring to
the proof of Lemma 3.4 in [6], the approach is simply
justified by applying Kruskal’s algorithm where edges
with lower weights are first sampled2 .
Note that the term MST is kept in the whole paper
for the sake of simplicity, but the sketching technique
and so does our algorithm enables to recover a Minimum
Spanning Forest if the initial graph is disconnected.
4
The proposed MST-based graph clustering
method: DBMSTClu
4.1 Principle. Let us consider a dataset with N
points. After the sketching phase, an approximate MST
further named T has been obtained with N − 1 edges
s.t. ∀i ∈ [N − 1], 0 < wi ≤ 1. Our density-based
clustering method DBMSTClu exclusively relies on this
object by performing some cuts among the edges of the
tree s.t. K − 1 cuts result in K clusters. Note that
independently of the technique used to obtain an MST
(the sketching method is just a possible one), the space
complexity of the algorithm is O(N ) which is better
than the O(N 2 ) of Spectral Clustering (SC). The time
complexity of DBMSTClu is O(N K) which is clearly
less than the O(N 3 ) one implied by SC. After a cut,
obtained clusters can be seen as subtrees of the initial
T and the analysis of their qualities is only based on
edges contained in those subtrees. In the sequel, all
clusters Ci , i ∈ [K], are assimilated to their associated
subtree of T and for instance the maximal edge of a
Definition 3.2. (`0 -sampling) An (, δ) `0 -sampler cluster will refer to the edge with the maximum weight
for a nonzero vector x ∈ Rn fails with a probability from the subtree associated to this cluster.
at most δ or returns some i ∈ [n] with probability
Our algorithm is a parameter-free divisive top1
where
supp
x
=
{i
∈
[n]
|
x
=
6
0}.
(1 ± ) | supp
down
procedure: it starts from one cluster containing
i
x|
the whole dataset and at each iteration, a cut which
The sketch requires O(N log3 (N )) space. It follows
that the sketching is technically semi-streamed but in
2
1 In
the sequel, for a given integer a, [a] = {1, . . . , a}.
We would like to thank Mario Lucic for the fruitful private
conversation and for coauthoring with Krzysztof Choromanski the
MSE sketching extension during his internship at Google.
Copyright c 2018 by SIAM
Unauthorized reproduction of this article is prohibited
maximizes some criterion is performed. The criterion
for identifying the best cut to do (if any should be
made) at a given stage is a measure of the validity of
the resulting clustering partition. This is a function of
two positive quantities defined below: Dispersion and
Separation of one cluster. The quality of a given cluster
is then measured from Dispersion and Separation while
the quality of the clustering partition results from the
weighted average of all cluster validity indices. Finally,
all those measures are based on the value of edge weights
and the two latter ones lie between −1 and 1.
Definition 4.1. (Cluster Dispersion) The
Dispersion of a cluster Ci (DISP) is defined as the
maximum edge weight of Ci . If the cluster is a
singleton (i.e. contains only one node), the associated
Dispersion is set to 0. More formally:
(4.1)
(
6 0
max wj if |E(Ci )| =
j, ej ∈Ci
∀i ∈ [K], DISP(Ci ) =
0
otherwise.
Definition 4.2. (Cluster Separation) The Separation of a cluster Ci (SEP) is defined as the minimum
distance between the nodes of Ci and the ones of all
other clusters Cj , j 6= i, 1 ≤ i, j ≤ K, K 6= 1 where K
is the total number of clusters. In practice, it corresponds to the minimum weight among all already cut
edges from T comprising a node from Ci . If K = 1, the
Separation is set to 1. More formally, with Cuts(Ci )
denoting cut edges incident to Ci ,
(4.2)
(
min
wj if K 6= 1
j, ej ∈Cuts(Ci )
∀i ∈ [K], SEP(Ci ) =
1
otherwise.
b
b
b
DISP(C1 )
b
b
b
b
b
b
b
b
b
SEP(C1 )
C1
Figure 1: SEP and DISP definitions with N = 12,
K = 3 for dashed cluster C1 in the middle.
Fig. 1 sums up the introduced definitions. The
higher the Separation, the farther is the cluster separated from the other clusters, while low values suggest
that the cluster is close to the nearest one.
Definition 4.3. (Validity Index of a Cluster)
The Validity Index of a cluster Ci is defined as:
(4.3)
VC (Ci ) =
SEP(Ci ) − DISP(Ci )
max(SEP(Ci ), DISP(Ci ))
The Validity Index of a Cluster (illustration in Fig. 2)
is defined s.t. −1 ≤ VC (Ci ) ≤ 1 where 1 stands for the
best validity index and −1 for the worst one. No division
by zero (i.e. max(DISP(Ci ), SEP(Ci )) = 0) happens
because Separation is always strictly positive. When
Dispersion is higher than Separation, −1 < VC (Ci ) < 0.
Conversely, when Separation is higher than Dispersion,
0 < VC (Ci ) < 1. So our clustering algorithm will
naturally encourage clusters with a higher Separation
over those with a higher Dispersion.
ϵ
b
1
b
ϵ
VC (Clef
t)
b
ϵ
1
VC (Clef
t)
b
b
ϵ
)
VC (Cright
1
b
1
VC (Cright
)
Figure 2: Validity Index of a Cluster’s example with
N = 3. For a small , cutting edge with weight or 1
gives resp. left and right partitions. (up) VC (Clef
t ) = 1;
1
VC (Cright
) = −1 < 0. (bottom) VC (Clef
)
=
1−
> 0;
t
1
VC (Cright ) = 1. The bottom partition, for which
validity indices of each cluster are positive, is preferred.
Definition 4.4. (Validity Index of a Clustering
Partition) The Density-Based Validity Index of a
Clustering partition Π = {Ci }, 1 ≤ i ≤ K, DBCVI(Π)
is defined as the weighted average of the Validity Indices
of all clusters in the partition where N is the number of
points in the dataset.
K
X
|Ci |
(4.4)
DBCVI(Π) =
VC (Ci )
N
i=1
The Validity Index of Clustering lies also between
−1 and 1 where 1 stands for an optimal density-based
clustering partition while −1 stands for the worst one.
Our defined quantities are significantly distinct
from the separation and sparseness defined in [28]. Indeed, firstly, their quantities are not well defined for
special cases when clusters have less than four nodes or
a partition containing a lonely cluster. Secondly, the
way they differentiate internal and external nodes or
edges does not properly recover easy clusters like convex blobs. Moreover, our DBCVI differs from the Silhouette Coefficient [29]. It does not perform well with
nonconvex-shaped clusters and although this is based
on close concepts like tightness and also separation, the
global coefficient is based on the average values of Silhouette coefficients of each point, while our computation
of DBCVI begins at the cluster level.
DBMSTClu is summarized in Algorithm 1. It starts
from a partition with one cluster containing the whole
dataset whereas the associated initial DBCVI is set to
the worst possible value: −1. As long as there exists
a cut which makes the DBCVI greater from (or equal
Copyright c 2018 by SIAM
Unauthorized reproduction of this article is prohibited
to) the one of the current partition, a cut is greedily
chosen by maximizing the obtained DBCVI among all
the possible cuts. When no direct improvement is
possible, the algorithm stops. It is guaranteed that
the cut edge locally maximizes the DBCVI at each
iteration since by construction, the algorithm will try
each possible cut. In practice, the algorithm stops
after a reasonable number of cuts, getting trapped in a
local maximum corresponding to a meaningful cluster
partition. This prevents from obtaining a partition
where all points are in singleton clusters. Indeed, such
a result (K = N ) is not desirable, although it is
optimal in the sense of the DBCVI, since in this case,
∀i ∈ [K], DISP(Ci ) = 0 and VC (Ci ) = 1. Moreover,
the non-parametric characteristic helps achieving stable
partitions. In Algorithm 1, evaluateCut(.) computes the
DBCVI when the cut in parameter is applied to T .
Algorithm 1 Clustering algorithm DBMSTClu
1: Input: T , the MST
2: dbcvi← − 1.0; clusters = [ ]; cut list←[E(T )]
3: while dbcvi < 1.0 do
4:
cut tp←N one; dbcvi tp←dbcvi
5:
for each cut in cut list do
6:
newDbcvi← evaluateCut(T , cut)
7:
if newDbcvi ≥ dbcvi tp then
8:
cut tp←cut; dbcvi tp←newDbcvi
9:
if cut tp 6= N one then
10:
clusters← cut(clusters, cut tp)
11:
dbcvi←dbcvi tp; remove(cut list, cut tp)
12:
else
13:
break
14: return clusters, dbcvi
Proof. For the first cut, both separations of obtained
clusters C1 and C2 are equal to the weight of the
considered edge for cut. Here, this is the one with
the highest weight. Thus, for i = 1, 2, DISP(Ci ) ≤
SEP(Ci ) =⇒ VC (Ci ) ≥ 0. Finally, the DBCVI of
the partition, as a convex sum of two nonnegative
quantities, is clearly nonnegative.
Lemma 4.2. Lowest weighted edge case Let T be
an MST of the dissimilarity data graph. If the first
cut from E(T ) done by DBMSTClu is the one with the
lowest weight, then resulting DBCVI is negative or zero.
Proof. Same reasoning in the opposite case s.t.
SEP(Ci ) − DISP(Ci ) ≤ 0 for i ∈ {1, 2}.
Proposition 4.1. When the first cut is not the
heaviest Let T be an MST of the dissimilarity data
graph with N nodes. Let us consider this specific case:
all edges have a weight equal to w except two edges e1
and e2 resp. with weight w1 and w2 s.t. w1 > w2 >
w > 0. DBMSTClu does not cut any edge with weight
w and cuts e2 instead of e1 as a first cut iff:
p
2n2 w1 − n1 + n21 + 4w1 (n22 w1 + N 2 − N n1 − n22 )
w2 >
2(N − n1 + n2 )
where n1 (resp. n2 ) is the number of nodes in the
first cluster resulting from the cut of e1 (resp. e2 ).
Otherwise, e1 gets cut.
Proof. See supplementary material.
This proposition emphasizes that the algorithm is
cleverer than simply cutting the heaviest edge first.
Indeed, although w2 < w1 , cutting e2 could be preferred
4.2 Quality of clusters. An analysis of the algo- over e . Moreover, no edge with weight w can get
1
rithm and the quality of the recovered clusters is now cut at the first iteration as they have the minimal
given. The main results are: 1) DBMSTClu differs sig- weight in the tree. Indeed it really happens since an
nificantly from the naive approach of SEMST by pre- approximate MST with discrete rounded weights is used
ferring cuts which do not necessarily correspond to the when sketching is applied.
heaviest edge (Prop. 4.1 and 4.2). 2) As long as the
current partition contains at least one cluster with a Proposition 4.2. First cut on the heaviest edge
negative validity index, DBMSTClu will find a cut im- in the middle Let T be an MST of the dissimilarity
proving the global index (Prop. 4.3). 3) Conditions are data graph with N nodes. Let us consider this specific
given to determine in advance if and which cut will be case: all edges have a weight equal to w except two
performed in a cluster with a positive validity index edges e1 and e2 resp. with weight w1 and w2 s.t.
(Prop. 4.4 and 4.5 ). All are completely independent of w1 > w2 > w > 0. Denote n1 (resp. n2 ) the number
the sketching phase. Prop. 4.1 and 4.2 rely on the two of nodes in the first cluster resulting from the cut of e1
basic lemmas regarding the first cut in the MST:
(resp. e2 ). In the particular case where edge e1 with
maximal weight w1 stands between two subtrees with the
Lemma 4.1. Highest weighted edge case Let T be same number of points, i.e. n1 = N/2, e1 is always
an MST of the dissimilarity data graph. If the first cut preferred over e2 as the first optimal cut.
from E(T ) made by DBMSTClu is the heaviest edge,
Proof. See supplementary material.
then resulting DBCVI is nonnegative.
Copyright c 2018 by SIAM
Unauthorized reproduction of this article is prohibited
b
S1
S2
b
b
b
Remark 4.1. Let us consider the MST in Fig. 3 with
S
N = 8, w1 = 1, w2 = w3 = 1 − , and other weights
e
set to . Clearly, it is not always preferred to cut e1
in the middle since for = 0.1, DBCV I2 ≈ 0.27 >
DBCV I1 = = 0.1. So, it is a counter-example to a Figure 5: Illustration of the recursive relationship for
possible generalization of Prop. 4.2 where there would be left and right Dispersions resulting from the cut of
edge e: DISPlef t (e) = max(w(S1 )), DISPright (e) =
more than three possible distinct weights in T .
max(w(S2 ), w(S3 )) where w(.) returns the edge weights.
Separation
works analogically.
w2
w1
w3
ϵ
ϵ
ϵ
ϵ
3
b
b
b
b
b
b
b
b
b
b
b
b
b
Figure 3: Counter-example for Remark 4.1
These last propositions hold for every iteration in the
algorithm.
Proposition 4.3. Fate of negative VC cluster
Let K = t + 1 be the number of clusters in the clustering
partition at iteration t. If for some i ∈ [K], VC (Ci ) < 0,
then DBMSTClu will cut an edge at this stage.
Proof. See supplementary material.
Ci
l
wsep
Cil
wmax
Cir
r
wsep
Figure 4: Generic example of Prop. 4.3 and 4.5’s proofs.
Thus, at the end of the clustering algorithm, each
cluster will have a nonnegative VC so at each step of
the algorithm, this bound holds for the final DBCVI:
PK
DBCV I ≥ i=1 |CNi | max(VC (Ci ), 0).
Proposition 4.4. Fate of positive VC cluster I
Let T be an MST of the dissimilarity data graph and C
a cluster s.t. VC (C) > 0 and SEP(C) = s. DBMSTClu
does not cut an edge e of C with weight w < s if
both resulting clusters have at least one edge with weight
greater than w.
Proof. See supplementary material.
Proposition 4.5. Fate of positive VC cluster II
Consider a partition with K clusters s.t. some cluster
Ci , i ∈ [K] with VC (Ci ) > 0 is in the setting of Fig. 10
i.e. cutting the heaviest edge e with weight wmax results
in two clusters: the left (resp. right) cluster Cil (resp.
Cir ) with n1 points (resp. n2 ) s.t. DISP(Cil ) = d1 ,
l
r
SEP(Cil ) = wsep
, DISP(Cir ) = d2 and SEP(Cir ) = wsep
.
l
r
Assuming w.l.o.g. wsep > wsep , e gets cut iff:
n1 d1 +n2 d2
n1 +n2
wmax
wmax
≤ r .
wsep
Proof. See supplementary material.
4.3 Implementation for linear time and space
complexities. Algorithm 1 previously described could
lead to a naive implementation. We now briefly explain how to make the implementation efficient in order to achieve the linear time and space complexities in N (code on https://github.com/annemorvan/
DBMSTClu/). The principle is based on two tricks. 1)
As observed in §4.2: for a performed cut in cluster
Ci , VC (Cj ) for any j 6= i remain unchanged. Hence,
if VC (Cj ) are stored for each untouched cluster after a
given cut, only the edges of Cl and Cr resp. the left
and right clusters induced by e’s cut need to be evaluated again to determine the DBCVI in case of cut.
Thus the number of operations to find the optimal cut
decreases drastically over time as the clusters become
smaller through the cuts. 2) However finding the first
cut already costs O(N ) time hence paying this price for
each cut evaluation would lead to O(N 2 ) operations.
Fortunately, this can be avoided as SEP and DISP exhibit some recurrence relationship in T : when knowing
these values for a given cut, we can deduce the value
for a neighboring cut (cf. Fig. 5). To determine the
first cut, T should be hence completely crossed following the iterative version of the Depth-First Search. The
difficulty though is that the recursive relationship between the quantities to update is directional: left and
right w.r.t. the edge to cut. So we develop here double
Depth-First search (see principle in Algorithm 2): from
any given edge of T , edges left and right are all visited consecutively with a Depth-First search, and SEP
and DISP are updated recursively thanks to a carefully
defined order in the priority queue of edges to handle.
5
Experiments
Tight lower and upper bounds are given in §3 for the
weight of the approximate MST retrieved by sketching.
First experiments in §5.1 show that the clustering
results do not suffer from the use of an approximate
MST instead of a real one. Experiments from §5.2 prove
then the scalability of DBMSTClu for large values of N .
5.1 Safety of the sketching. The results of DBMSTClu are first compared with DBSCAN [4] because it
can compete with DBMSTClu as 1) nonconvex-shaped
Copyright c 2018 by SIAM
Unauthorized reproduction of this article is prohibited
Algorithm 2 Generic Double Depth-First Search
1: Input: T , the MST; e, the edge of T where the
search starts; n src, source node of e
2: Q = deque() //Empty priority double-ended queue
3: for incident edges to n src do
4:
pushBack(Q, (incident e, n src, F ALSE))
5: pushBack(Q, (e, n src, T RU E))
6: for incident edges to n trgt do
7:
pushBack(Q, (incident e, n trgt, F ALSE))
8: pushBack(Q, (e, n trgt, T RU E))
9: while Q is not empty do
10:
e, node, marked = popFront(Q)
11:
opposite node = getOtherNodeEdge(e, node)
12:
if not marked then
13:
for incident edges to node do
14:
pushBack(Q, (incident e, node, F ))
15:
pushBack(Q, (e, node, T ))
16:
pushFront(Q, (e, opposite node, T ))
17:
else
18:
doTheJob(e) //Perform here the task
19: return
clusters are recognized, 2) it does not require explicitly
the number of expected clusters, 3) it is both time and
space-efficient (resp. O(N log N ) and O(N )). Then,
the results of another MST-based algorithm are shown
for comparison. The latter called Standard Euclidean
MST (SEMST) [19] cuts the K − 1 heaviest edges of
a standard Euclidean MST given a targeted number of
clusters K. For DBMSTClu, the dissimilarity graph
is built from computing the Euclidean distance between
data points and passed into the sketch phase to produce
an approximate version of the exact MST.
Synthetic datasets. Experiments were performed
on two classic synthetic datasets from the Euclidean
space: noisy circles and noisy moons. Each dataset
contains 1000 data points in 20 dimensions: the first
two dimensions are randomly drawn from predefined
2D-clusters, as shown in Fig. 6 and 7, while the other
18 dimensions are random Gaussian noise.
Real dataset.
DBMSTClu performances are
also measured on the mushroom dataset (https://
archive.ics.uci.edu/ml/datasets/mushroom).
It
contains 8124 records of 22 categorical attributes corresponding to 23 species of gilled mushrooms in the
Agaricus and Lepiota family. 118 binary attributes are
created from the 22 categorical ones, then the complete
graph (about 33 millions of edges) is built by computing
the normalized Hamming distance (i.e. the number of
distinct bits) between points.
Results. Fig. 6, 7 and 8 show the results for all
previously defined datasets and aforementioned meth-
ods. The synthetic datasets were projected onto 2D
spaces for visualization purposes. They were produced
with a noise level such that SEMST fails and DBSCAN
does not perform well without parameters optimization.
In particular, for DBSCAN all the cross points correspond to noise. With the concentric circles, SEMST
does not cut on the consistent edges, hence leads to an
isolated singleton cluster. DBSCAN classifies the same
point plus a near one as noise while recovering the two
circles well. Finally, DBMSTClu finds the two main
clusters and also creates five singleton clusters which
can be legitimately considered as noise as well. With
noisy moons, while DBSCAN considers three outliers,
DBMSTClu detects the same as singleton clusters. As
theoretically proved above, experiments emphasize the
fact that our algorithm is more subtle than simply cutting the heaviest edges as the failure of SEMST shows.
Moreover our algorithm exhibits an ability to detect
outliers, which could be labeled as noise in a postprocessing phase. Another decisive advantage of our algorithm is the absence of any required parameters, contrarily to DBSCAN. For the mushroom dataset, if suitable parameters are given to SEMST and DBSCAN,
the right 23 clusters get found while DBMSTClu retrieves them without tuning any parameter. Quantitative results for the synthetic datasets are shown in Table 3: the achieved silhouette coefficient (between −1
and 1), Adjusted Rand Index (ARI) (between 0 and 1)
and DBCVI. For all the indices, the higher, the better.
Further analysis can be read in supplementary material.
For the mushroom dataset, the corresponding DBCVI
and silhouette coefficient are resp. 0.75 and 0.47.
SEMST
DBSCAN
DBMSTClu
Silhouette coeff.
0.16 -0.12
0.02
0.26
-0.26 0.26
ARI
0
0.99
0.99
0
0.99
0.99
DBCVI
0.001 0.06
-0.26 0.15
0.18 0.15
Table 1: Silhouette coefficients, ARI and DBCVI for the
noisy circles (left) and noisy moons (right) datasets.
5.2 Scalability of the clustering. For mushroom
dataset, DBMSTClu’s execution time (avg. on 5 runs)
is 3.36s while DBSCAN requires 9.00s. This gives a
first overview of its ability to deal with high number
of clusters. Further experiments on execution time
were conducted on large-scale random weighted graphs
generated from the Stochastic Block Model with varying
equal-sized number of clusters K and N . The scalability
of DBMSTClu is shown in Fig. 9 and Table 2 by
exhibiting the linear time complexity in N . Graphs with
1M of nodes and 100 clusters were easily clustered. In
Table 2, the row with the execution time ratio between
K = 100 and K = 5 illustrates the first trick from
§4.3 as the observed time ratio is around 2/3 of the
Copyright c 2018 by SIAM
Unauthorized reproduction of this article is prohibited
1.5
1.5
1.5
1.0
1.0
1.0
0.5
0.5
0.5
0.0
0.0
0.0
0.5
0.5
0.5
1.0
1.0
1.0
1.5
1.5
2
1
0
1
2
2
1
0
1
1.5
2
2
1
0
1
2
Figure 6: Noisy circles: SEMST, DBSCAN ( = 0.15, minP ts = 5), DBMSTClu with an approximate MST.
1.5
1.5
1.5
1.0
1.0
1.0
0.5
0.5
0.5
0.0
0.0
0.0
0.5
0.5
0.5
1.0
1.0
0.5
0.0
0.5
1.0
1.5
2.0
1.0
1.0
0.5
0.0
0.5
1.0
1.5
1.0
2.0
1.0
0.5
0.0
0.5
1.0
1.5
2.0
Figure 7: Noisy moons: SEMST, DBSCAN ( = 0.16, minP ts = 5), DBMSTClu with an approximate MST.
4 3 2
1 0 1
2 3 4
5
3
2
1
0
1
2
3
3
2
1
0
1
2
3
4 3 2
1 0 1
2 3 4
5
3
2
1
0
1
2
3
3
2
1
0
1
2
3
4 3 2
1 0 1
2 3 4
5
3
2
1
0
1
2
3
3
2
1
0
1
2
3
Figure 8: Mushroom dataset: SEMST, DBSCAN ( = 1.5, minP ts = 2), DBMSTClu with an approximate MST
(projection on the first three principal components).
Clu is non-parametric, unlike most existing clustering
methods: it automatically determines the right number
4000
of nonconvex clusters. Although the approach is funK=5
K=20
3500
damentally independent from the sketching phase, its
K=100
3000
robustness has been assessed by using as input an ap2500
proximate MST of the sketched G rather than an exact
2000
one. The graph sketch is computed dynamically on the
1500
1000
fly as new edge weight updates are read in only one
500
pass over the data. This brings a space-efficient solu0
0
200000
400000
600000
800000
1000000
tion for finding an MST of G when the O(N 2 ) edges
Number of points
cannot fit in memory. Hence, our algorithm adapts to
Figure 9: DBMSTClu’s execution time with values of
the semi-streaming setting with O(N polylog N ) space.
N ∈ {1K, 10K, 50K, 100K, 250K, 500K, 750K, 1M}.
Our approach shows promising results, as evidenced by
the experimental part regarding time and space scala6 Conclusion
bility and clustering performance even with sketching.
In this paper we introduced DBMSTClu a novel space- Further work would consist in using this algorithm in
efficient Density-Based Clustering algorithm which only privacy issues, as the lost information when sketching
relies on a Minimum Spanning Tree (MST) of the dis- might ensure data privacy. Moreover, as it is already the
similarity graph G: the spatial and time costs are resp. case for the graph sketching, we could look for adaptO(N ) and O(N K) with N the number of data points ing both the MST recovery and DBMSTClu to the fully
and K the number of clusters. This enables to deal eas- online setting, i.e. to be able to modify dynamically curily with graphs of million of nodes. Moreover, DBMSTTime (s)
theoretical one 100/5 = 20.
Copyright c 2018 by SIAM
Unauthorized reproduction of this article is prohibited
K\N
5
20
100
“100/5”
1000
0.34
0.95
4.36
12.82
10000
2.96
8.73
40.25
13.60
50000
14.37
43.71
201.76
14.04
100000
28.91
88.51
398.41
13.78
250000
73.04
223.18
995.42
13.63
500000
148.85
449.37
2011.79
13.52
750000
218.11
669.29
3015.61
13.83
1000000
292.25
889.88
4016.13
13.74
Table 2: Numerical values for DBMSTClu’s execution time (in s) varying N and K (avg. on 5 runs). The last
row shows the execution time ratio between K = 100 and K = 5.
rent MST and clustering partition as a new edge weight
update from the stream is seen.
[16]
References
[1] S. Lloyd, Least Squares Quantization in PCM, IEEE
Trans. Inf. Theor., 28 (1982), pp. 129–137.
[2] A. Jain and R. Dubes, Algorithms for Clustering Data,
Prentice Hall, Upper Saddle River, NJ, USA, 1988.
[3] L. Kaufman and P. Rousseeuw, Clustering by means
of medoids, Statistical Data Analysis Based on the
L1 Norm and Related Methods, pp. 405–416, 1987.
[4] M. Ester, H. Kriegel, J. Sander, X. Xu, A densitybased algorithm for discovering clusters in large spatial
databases with noise, AAAI Press (1996), pp. 226–231.
[5] S. Muthukrishnan, Data Streams: Algorithms and
Applications, Found. Trends Theor. Comput. Sci., 1
(2005), pp. 117–236.
[6] K. Ahn, S. Guha and A. McGregor, Analyzing Graph
Structure via Linear Measurements, ACM-SIAM Symposium on Discrete Algorithms, 2012, Kyoto, Japan,
pp. 459–467.
[7] S. Schaeffer, Survey: Graph Clustering, Comput. Sci.
Rev., 1 (2007), pp. 27–64.
[8] M. Girvan and M. Newman, Community Structure
in Social and Biological Networks, Proceedings of the
National Academy of Sciences of the United States of
America, 99 (2012), pp.7821-7826.
[9] M. Nascimento and A. de Carvalho, Spectral methods
for graph clustering - A survey, European Journal of
Operational Research, 211 (2011), pp.221-231.
[10] T. Falkowski, A. Barth and M. Spiliopoulou, DENGRAPH: A Density-based Community Detection Algorithm, IEEE/WIC/ACM International Conference on
Web Intelligence, pp. 112–115, 2007.
[11] N. Ailon, Y. Chen and H. Xu, Breaking the Small Cluster Barrier of Graph Clustering, CoRR, abs/1302.4549,
2013.
[12] S. Oymak and B. Hassibi, Finding Dense Clusters
via ”Low Rank + Sparse” Decomposition, CoRR,
abs/1104.5186, 2011.
[13] Y. Chen, S. Sanghavi and H. Xu, Clustering Sparse
Graphs, Advances in Neural Information Processing
Systems, 25 (2012), pp. 2204–2212.
[14] Y. Chen, A. Jalali, S. Sanghavi and H. Xu, Clustering
Partially Observed Graphs via Convex Optimization, J.
Mach. Learn. Res., 15 (2014), pp. 2213–2238.
[15] Y. Chen, and S. Lim and H. Xu, Weighted Graph Clustering with Non-uniform Uncertainties, International
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
Conference on Machine Learning, 32 (2014), Beijing,
China, pp. II-1566–II-1574.
P. Holland, K. Laskey and S. Leinhardt, Stochastic
blockmodels: First steps, Social Networks, 5 (1983),
pp.109–137.
A. Condon and R. Karp, Algorithms for Graph Partitioning on the Planted Partition Model, Random
Struct. Algorithms, 18 (2001), pp.116–140.
K. Rohe, S. Chatterjee and B. Yu, Spectral clustering
and the high-dimensional stochastic blockmodel, Ann.
Statist., 39 (2011), pp. 1878–1915.
C. Zahn, Graph-Theoretical Methods for Detecting and
Describing Gestalt Clusters, IEEE Trans. Comput., 20,
(1971), pp. 68–86.
T. Asano, B. Bhattacharya, M. Keil, F. Yao, Clustering
Algorithms Based on Minimum and Maximum Spanning Trees, Symposium on Computational Geometry,
Urbana-Champaign, IL, USA, pp. 252–257, 1988.
O. Grygorash, Y. Zhou and Z. Jorgensen, Minimum
Spanning Tree Based Clustering Algorithms, IEEE
Int.Conference on Tools with Artificial Intelligence,
pp. 73-81, 2006.
K. Ahn, S.Guha and A. McGregor, Graph Sketches:
Sparsification, Spanners, and Subgraphs, ACM
SIGMOD-SIGACT-SIGAI Symposium on Principles
of Database Systems, Scottsdale, AZ, USA, pp. 5–14,
2012.
N. Ailon, R. Jaiswal, C. Monteleoni, Streaming kmeans approximation, Advances in Neural Information
Processing Systems, 22 (2009), pp. 10–18.
S. Guha, R. Rastogi and K. Shim, Cure: An Efficient
Clustering Algorithm for Large Databases, Inf. Syst.,
26 (2001), pp. 35–58.
C. Aggarwal, J. Han, J. Wang, and P. Yu, A Framework
for Clustering Evolving Data Streams, International
Conference on Very Large Data Bases, 29 (2003),
Berlin, Germany, pp.81–92.
F. Cao, M. Ester, W. Qian and A. Zhou, Density-based
clustering over an evolving data stream with noise,
SIAM Conference on Data Mining (2006), pp. 328–339.
G. Cormode and D. Firmani, A unifying framework
for `0 -sampling algorithms, Distributed and Parallel
Databases, 32 (2014), pp. 315–335.
D. Moulavi, P. Jaskowiak, R. Campello, A. Zimek and
J. Sander, Density-Based Clustering Validation, SIAM
ICDM, Philadelphia, PA, USA, pp. 839–847, 2014.
P. Rousseeuw, Silhouettes: A graphical aid to the
interpretation and validation of cluster analysis, J.
Comput. and Appl. Math., 20 (1987), pp. 53–65.
Copyright c 2018 by SIAM
Unauthorized reproduction of this article is prohibited
7
Proofs.
7.1
maximal weight w1 stands between two subtrees with the
same number of points, i.e. n1 = N/2, e1 is always
preferred over e2 as the first optimal cut.
Proof of Prop. 4.1.
Proposition 4.1. When the first cut is not the
heaviest Let T be an MST of the dissimilarity data
graph with N nodes. Let us consider this specific case:
all edges have a weight equal to w except two edges e1
and e2 resp. with weight w1 and w2 s.t. w1 > w2 >
w > 0. DBMSTClu does not cut any edge with weight
w and cuts e2 instead of e1 as a first cut iff:
p
2n2 w1 − n1 + n21 + 4w1 (n22 w1 + N 2 − N n1 − n22 )
w2 >
2(N − n1 + n2 )
where n1 (resp. n2 ) is the number of nodes in the
first cluster resulting from the cut of e1 (resp. e2 ).
Otherwise, e1 gets cut.
Proof. Let DBCV I1 (resp. DBCV I2 ) be the DBCVI
after cut of e1 (resp. e2 ). As w (resp. w1 ) is the
minimum (resp. maximal) weight, the algorithm does
not cut e since the resulting DBCVI would be negative
(cf. Lemma 4.2) while DBCV I1 is guaranteed to be
positive (cf. Lemma 4.1). So, the choice will be
between e1 and e2 but e2 gets cut iff DBCV I2 >
DBCV I1 . DBCV I1 and DBCV I2 expressions are
simplified w.l.o.g. by scaling the weights by w s.t.
w ← 1, w1 ← w1 /w, w2 ← w2 /w, hence w1 > w2 > 1.
Then,
DBCV I2 > DBCV I1 > 0
1
n2 w 2
n2
(
)(1 −
⇐⇒
− 1) + (1 −
)
N w1
N
w2
n1
1
w2
n1
−
(1 −
)(1 −
) + (1 −
)>0
N
w1
N
w1
⇐⇒ w22 (N + n2 − n1 ) +w2 (n1 − 2n2 w1 )
{z
}
|
{z
}
|
a
+ (n2 − N )w1 > 0.
|
{z
}
b
Proof. A reductio ad absurdum is made by showing that
cutting edge e2 i.e. DBCV I2 > DBCV I1 leads to the
contradiction w1 /w < 1. With the scaling process from
Prop. 4.1’proof:
1
1
w2
1
w2
1
(1 −
) + (1 −
)=1−
−
2
w1
2
w1
2w1
2w1
n2 w2
n2
1
DBCV I2 =
(
− 1) + (1 −
)(1 −
)
N w1
N
w2
1
n2 w2
1
=1−
+
(
+
− 2)
w2
N w1
w
|
{z 2 }
DBCV I1 =
=A
There is w2 > w = 1, so w12 < 1. Besides w2 < w1 so
w2
w1 < 1 thus, A < 0. Let now consider w.l.o.g. that
edge e2 is on the ”right side” (right cluster/subtree) of
e1 (similar proof if e2 is on the left side of e1 ). Hence,
it is clear that for maximizing DBCV I2 as a function
of n2 , we need n2 = n1 + 1. Then,
DBCV I2 > DBCV I1
1
1
1 w2
1
1
w2
⇐⇒ −
+ ( + )(
−2+
)>−
−
w2
2 N w1
w2
w1
w1
1
1
1
1
2
⇐⇒ (
+
+
+
)w2 − 1 −
2w1
N w1
2w1
N
2w1
1
1 1
+ (−1 + + )
>0
2 N w2
2
1
1
⇐⇒ (1 + ) w22 + w2 ( − w1 (1 + ))
N
2
N
| {z }
|
{z
}
a>0
b<0
1
+ w1 ( −
| N {z
c<0
1
)>0
2}
As c/a < 0 and w2 > 0, w2 > 2(NN+1) [ w1 (1 + N2 ) − 12 +
√
∆ ] with ∆ = (w1 (1 + N2 ) − 12 )2 + 4(1 + N1 )( 12 − N1 )w1 .
Clearly, ∆ = b2 − 4ac is positive
and
c/a
is
negative.
This inequality is incompatible with w1 > w2 since:
√
b2 −4ac
But w2 > 0, then w2 > −b+ 2a
which gives the
N
2
1 √
final result after some simplifications.
w1 > w2 ⇐⇒ w1 >
[ w1 (1 + ) − + ∆ ]
2(N + 1)
N
2
√
1
7.2 Proof of Prop. 4.2.
⇐⇒ w1 + > ∆
2
Proposition 4.2. First cut on the heaviest edge
4 2
1
4
1
⇐⇒
w (1 + ) + w1 (−1 − ) < 0
in the middle Let T be an MST of the dissimilarity
N 1
N
N
N
data graph with N nodes. Let us consider this specific
⇐⇒ w1 < 1 : ILLICIT
case: all edges have a weight equal to w except two
edges e1 and e2 resp. with weight w1 and w2 s.t. Indeed, after the scaling process, w1 < 1 = w is not
w1 > w2 > w > 0. Denote n1 (resp. n2 ) the number possible since by hypothesis, w1 > w. Finally, it is not
of nodes in the first cluster resulting from the cut of e1 allowed to cut e2 , the only remaining possible edge to
(resp. e2 ). In the particular case where edge e1 with cut is e1 .
c<0
Copyright c 2018 by SIAM
Unauthorized reproduction of this article is prohibited
7.4
Ci
l
wsep
Cil
wmax
Cir
r
wsep
Proof of Prop. 4.4.
Proposition 4.4. Fate of positive VC cluster I
Let T be an MST of the dissimilarity data graph and C
a cluster s.t. VC (C) > 0 and SEP(C) = s. DBMSTClu
does not cut an edge e of C with weight w < s if
both resulting clusters have at least one edge with weight
greater than w.
Proof. Let us consider clusters C1 and C2 resulting
from the cut of edge e. Assume that in the associated
subtree of C1 (resp. C2 ), there is an edge e1 (resp.
Figure 10: Generic example for proof of Prop. 4.3 and e2 ) with a weight w1 (resp. w2 ) higher than w s.t.
without loss of generality, w1 > w2 . Since VC (C) > 0,
4.5.
s > w1 > w2 > w. But cutting edge e implies that
for i ∈ {1, 2}, DISP(Ci ) > SEP(Ci ) = w, and thus
7.3 Proof of Prop. 4.3.
VC (Ci ) < 0. Cutting edge e would therefore mean to
Proposition 4.3. Fate of negative VC cluster replace a cluster C s.t. VC (C) > 0 by two clusters s.t.
Let K = t + 1 be the number of clusters in the clustering for i ∈ {1, 2}, VC (Ci ) < 0 which obviously decreases the
partition at iteration t. If for some i ∈ [K], VC (Ci ) < 0, current DBCVI. Thus, e does not get cut at this step of
the algorithm.
then DBMSTClu will cut an edge at this stage.
Proof. Let i ∈ [K] s.t. VC (Ci ) < 0 i.e. SEP(Ci ) <
l
DISP(Ci ). We denote wsep
the minimal weight outing
cluster Ci and wmax the maximal weight in subtree Si
def
def
l
of Ci i.e. SEP(Ci ) = wsep
and DISP(Ci ) = wmax .
l
< wmax . By cutting the cluster Ci on the
Hence, wsep
edge with weight wmax , we define Cil and Cir resp. the
left and right resulting clusters.
Let us look at VC (Cil ). If SEP(Cil ) ≥ DISP(Cil )
SEP(C l )
then VC (Cil ) ≥ 0 ≥ VC (Ci ) else VC (Cil ) = DISP(Cil ) − 1.
i
The definition of the Separation as a minimum and our
cut imply that
SEP(Cil ) ≥ min(SEP(Ci ), wmax ) ≥ SEP(Ci ).
7.5
Proof of Prop. 4.5.
Proposition 4.5. Fate of positive VC cluster II
Consider a partition with K clusters s.t. some cluster
Ci , i ∈ [K] with VC (Ci ) > 0 is in the setting of Fig. 10
i.e. cutting the heaviest edge e with weight wmax results
in two clusters: the left (resp. right) cluster Cil (resp.
Cir ) with n1 points (resp. n2 ) s.t. DISP(Cil ) = d1 ,
r
l
.
, DISP(Cir ) = d2 and SEP(Cir ) = wsep
SEP(Cil ) = wsep
l
r
Assuming w.l.o.g. wsep > wsep , e gets cut iff:
n1 d1 +n2 d2
n1 +n2
wmax
≤
wmax
.
r
wsep
Also the definition of the Dispersion as a maximum
implies that DISP(Cil ) ≤ DISP(Ci ). Hence we get that
r
Proof. As VC (Ci ) > 0, there is SEP(Ci ) = wsep
>
w
.
Then,
the
DBCVI
before
(K
clusters)
and
after
max
− 1 i.e. VC (Cil ) ≥ VC (Ci ) in
−1 ≥
cut of wmax (K + 1 clusters) are:
r
this case too. The same reasoning holds for Ci showing
that VC (Cir ) ≥ VC (Ci ). Finally,
K
X
n1 + n2
wmax
DBCV
I
=
V
(C
)
+
1
−
K
C
j
r
X nj
N
wsep
nl
nr
j6=i
DBCV Iaf tercut =
VC (Cj ) + i VC (Cil ) + i VC (Cir )
N
N
N
K
X
j6=i
n1
d1
DBCV
I
=
V
(C
)
+
1
−
K+1
C
j
X nj
nl
nr
N
wmax
j6=i
≥
VC (Cj ) + i VC (Ci ) + i VC (Ci )
N
N
N
j6=i
n2
d2
+
1−
= DBCV Ibef orecut .
N
wmax
SEP(Cil )
DISP(Cil )
SEP(Ci )
DISP(Ci )
Hence cutting the edge with maximal weight in Ci
improves the resulting DBCVI.
DBMSTClu cuts wmax iff DBCV IK+1 ≥ DBCV IK .
So the result after simplification.
Copyright c 2018 by SIAM
Unauthorized reproduction of this article is prohibited
8
Complements on experiments.
Experiments were conducted using Python and scikitlearn library [1] on a single-thread process on an intel
processor based node.
8.1 Safety of the sketching. Fig. 11 shows another
result on a synthetic dataset: three blobs generated
from three Gaussian distributions. With the three
blobs, each method SEMST, DBSCAN and DBMSTClu
performs well: they all manage to retrieve three clusters.
Quantitative results for the three synthetic datasets
are shown in Table 3: the achieved silhouette coefficient, Adjusted Rand Index (ARI) and DBCVI. For all
the indices, the higher, the better. Silhouette coefficient (between −1 and 1) is used to measure a clustering partition without any external information. For
DBSCAN it is computed by considering noise points as
singletons. We see that this measure is not very suitable
for nonconvex clusters like noisy circles or moons. The
ARI (between 0 and 1) measures the similarity between
the experimental clustering partition and the known
groundtruth. DBSCAN and DBMSTClu give similar
almost optimal results. Finally, the obtained DBCVIs
are consistent, since the best ones are reached for DBMSTClu.
References
[1] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel,
B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer,
R. Weiss, V. Dubourg, J. Vanderplas, A. Passos,
D. Cournapeau, M. Brucher, M. Perrot and E. Duchesnay, Scikit-learn: Machine Learning in Python, Journal
of Machine Learning Research, 12 (2011), pp. 2825–
2830.
Copyright c 2018 by SIAM
Unauthorized reproduction of this article is prohibited
15
15
15
10
10
10
5
5
5
0
0
0
5
5
5
10
10
10
15
15
20
15
10
5
0
5
10
15
20
25
20
15
10
5
0
5
10
15
20
25
15
20
15
10
5
0
5
10
15
20
25
Figure 11: Three blobs: SEMST, DBSCAN ( = 1.4, minP ts = 5), DBMSTClu with an approximate MST.
SEMST
DBSCAN
DBMSTClu
Silhouette coeff.
0.84 0.16 -0.12
0.84 0.02
0.26
0.84 -0.26 0.26
Adjusted Rand Index
1 0
0
1 0.99 0.99
1 0.99 0.99
DBCVI
0.84 0.001
0.84 -0.26
0.84 0.18
0.06
0.15
0.15
Table 3: Silhouette coefficients, Adjusted Rand Index and DBCVI for the blobs, noisy circles and noisy moons
datasets with SEMST, DBSCAN and DBMSTClu.
Copyright c 2018 by SIAM
Unauthorized reproduction of this article is prohibited
| 8cs.DS
|
Learning Multi-item Auctions with (or without) Samples
arXiv:1709.00228v1 [cs.GT] 1 Sep 2017
Yang Cai
McGill University, Canada
[email protected]
Constantinos Daskalakis
EECS and CSAIL, MIT, USA
[email protected]
September 4, 2017
Abstract
We provide algorithms that learn simple auctions whose revenue is approximately optimal in multi-item
multi-bidder settings, for a wide range of bidder valuations including unit-demand, additive, constrained
additive, XOS, and subadditive. We obtain our learning results in two settings. The first is the commonly
studied setting where sample access to the bidders’ distributions over valuations is given, for both regular
distributions and arbitrary distributions with bounded support. Here, our algorithms require polynomially
many samples in the number of items and bidders. The second is a more general max-min learning setting
that we introduce, where we are given “approximate distributions,” and we seek to compute a mechanism
whose revenue is approximately optimal simultaneously for all “true distributions” that are close to the ones
we were given. These results are more general in that they imply the sample-based results, and are also
applicable in settings where we have no sample access to the underlying distributions but have estimated
them indirectly via market research or by observation of bidder behavior in previously run, potentially
non-truthful auctions.
All our results hold for valuation distributions satisfying the standard (and necessary) independenceacross-items property. They also generalize and improve upon recent works of Goldner and Karlin [28]
and Morgenstern and Roughgarden [35], which have provided algorithms that learn approximately optimal
multi-item mechanisms in more restricted settings with additive, subadditive and unit-demand valuations
using sample access to distributions. We generalize these results to the complete unit-demand, additive,
and XOS setting, to i.i.d. subadditive bidders, and to the max-min setting.
Our results are enabled by new uniform convergence bounds for hypotheses classes under product
measures. Our bounds result in exponential savings in sample complexity compared to bounds derived by
bounding the VC dimension, and are of independent interest.
1 Introduction
The design of revenue-optimal auctions is a central problem in Economics and Computer Science, which has
found myriad applications in online and offline settings, ranging from sponsored search and online advertising
to selling artwork by auction houses, and public goods such as drilling rights and radio spectrum by governments. The problem involves a seller who wants to sell one or several items to one or multiple strategic bidders
with private valuation functions, mapping each bundle of items they may receive to how much value they
derive from the bundle. As no meaningful revenue guarantee can possibly be achieved without any information about the valuations of the bidders, the problem has been classically studied under Bayesian assumptions,
where a joint distribution from which all bidders’ valuations are drawn is common knowledge, and the goal is
to maximize revenue in expectation with respect to this distribution.
In the single-item setting, Bayesian assumptions have enabled beautiful and influential developments in
auction theory. Already 36 years ago, a breakthrough result by Myerson identified the optimal single-item
auction when bidder values are independent [36], and the ensuing decades saw a great deal of further understanding and practical applications of single-item auctions, importantly in online settings.
However, the quest for optimal multi-item auctions has been quite more challenging. It has been recognized
that revenue-optimal multi-item auctions can be really complex, may exhibit counter-intuitive properties, and
be fragile to changes in the underlying distributions; for a discussion and examples see survey [18]. As such,
it is doubtful that there is a crisp characterization of the structure of optimal multi-item auctions, at least not
beyond single-bidder settings [19]. On the other hand, there has been significant recent progress in efficient
computation of revenue-optimal auctions [14, 15, 1, 7, 3, 8, 9, 12, 10, 2, 6, 20]. Importantly, this progress has
enabled identifying simple auctions (mostly variations of sequential posted pricing mechanisms) that achieve
constant factor approximations to the revenue of the optimum [5, 42, 11, 16, 13], under the item-independence
assumption of Definition 1 and Example 1. These auctions are way simpler than the optimum, and have strong
incentive properties: they are dominant strategy truthful, while still competing against the optimal Bayesian
truthful mechanism. The current state-of-the-art is given as Theorem 8, which applies to bidders with valuation
functions from the broad class of fractionally subbaditive (a.k.a. XOS) valuations, which contains submodular.
As our discussion illustrates, studying auctions assuming Bayesian priors has been quite fruitful, enabling
us to identify guiding principles for how to structure auctions to achieve optimal (in single-item settings)
or approximately optimal (in multi-item settings) revenue. To apply this theory to practice, however, one
needs knowledge of the underlying distributions. Typically, one would estimate these distributions via market
research or by observations of bidder behavior in prior auctions, then use the estimated distributions to design
a good auction. However, estimation involves approximation, and the performance of mechanisms can be
quite fragile to errors in the distributions. This motivates studying whether optimal or approximately optimal
auctions can be identified when one has imperfect knowledge of the true distributions.
With this motivation, recent work in Computer Science has studied whether approximately optimal mechanisms can be “learned” given sample access to the underlying distributions. This work has lead to an almost
complete picture for the single-item (and the more general single-parameter) setting where Myerson’s theory
applies, showing how near-optimal mechanisms can be learned from polynomially many (in the approximation
and the number of bidders) samples [26, 17, 33, 31, 34, 22, 38, 29].
On the multi-item front, however, where the analogue of Myerson’s theory is elusive, and unlikely, our
understanding is much sparser. Recent work of Morgenstern and Roughgarden [35] has taken a computational
learning theory approach to identify the sample complexity required to optimize over classes of simple auctions. Combined with the afore-described results on the revenue guarantees of simple auctions, their work
leads to algorithms that learn approximately optimal auctions in multi-item settings with multiple unit-demand
bidders, or a single subadditive bidder, from polynomially many samples in the number of items and bidders. These results apply to distributions satisfying the item-independence assumption of Definition 1 and
1
Example 1, under which the approximate optimality of simple auctions has been established.
While well-suited for identifying the sample complexity required to optimize over a class of simple mechanisms, which is a perfectly reasonable goal to have but not the one in this paper, the approach taken in [35] is
arguably imperfect towards proving polynomial sample bounds for learning approximately optimal auctions in
the settings where simple mechanisms are known to perform well in the first place. This is due to the following
discordance: (i) On the one hand, simple and approximately optimal mechanisms in multi-item settings are
mostly only known under item-independence. (ii) On the other hand, the computational learning techniques
employed in [35], and in particular bounding the pseudo-dimension of a class of auctions, are not fine enough
to discern the difference in sample complexity required to optimize under item-independence and without itemindependence. As such, this technique can only obtain polynomial sample bounds for approximate revenue
optimization if it so happens that a class of mechanisms is both learnable from polynomially-many samples
under arbitrary distributions, and it guarantees approximately optimal revenue under item-independence, or
for some other interesting class of distributions.1
In particular, bounding the pseudo-dimension of classes of auctions as a means to prove polynomialsample bounds for approximate revenue optimization hits a barrier even for multiple additive bidders with
independent values for items. In this setting, the approximately optimal auctions that are known are the best of
selling the items separately or running a VCG mechanism with entry fees [42, 11], as described in Section 5.2.
Unfortunately, the latter can easily be seen to have pseudo-dimension that is exponential in the number of
bidders, thus only implying a sufficient exponentially large sample size to optimize over these mechanisms.
Is this exponential sample size really necessary or an artifact of the approach? Recent work of Goldner and
Karlin [28] gives us hope that it is the latter. They show how to learn approximately optimal auctions in the
multi-item multi-bidder setting with additive bidders using only one sample from each bidder’s distribution,
assuming that it is regular and independent across items.
Our results. We show that simple and approximately optimal mechanisms are learnable from polynomiallymany samples for multi-item multi-bidder settings, whenever:
• the bidder valuations are fractionally subadditive (XOS), i.e. we can accommodate additive, unitdemand, constrained additive, and submodular valuations;
• the distributions over valuations satisfy the standard item-independence assumption of Definition 1 and
Example 1, and their single-item marginals are arbitrary and bounded, or (have arbitrary supports but
are) regular.2
In particular, our results constitute vast extensions of known results on the polynomial learnability of approximately optimal auctions in multi-item settings [35, 28]. Additionally we show that:
• whenever the valuations are additive and unit-demand, or whenever the bidders are symmetric and have
XOS valuations, our approximately optimal mechanisms can be identified from polynomially many
samples and in polynomial time;
• whenever the bidders are symmetric (i.e. their valuations are independent and identically distributed) and
have subadditive valuations, we can computefrom polynomially
many samples and in polynomial-time
n
a simple mechanism whose revenue is a Ω max{m,n} -fraction of the optimum, where m and n are
respectively the number of items and bidders. In particular, if the number of bidders is at least a constant
fraction of the number of items, the mechanism is a constant factor approximation; and
1
It is known that some restriction needs to be made on the distribution to gain polynomial sample complexity, as otherwise exponential lower bounds are known for learning approximately optimal auctions even for a single unit-demand bidder [24].
2
We note again that without the standard item-independence (or some other) restriction on the distributions, we cannot hope to
learn approximately optimal auctions from sub-exponentially many samples, even for a single unit-demand bidder [24].
2
• in the setting of the previous bullet, if the item marginals are regular, our mechanism is prior-independent,
i.e. there is a single mechanism, identifiable without any samples from the distributions, providing the
afore-described revenue guarantee.
Finally, the mechanisms learned by our algorithms for XOS bidders are either rationed sequential posted price
mechanisms (RSPMs) or anonymous sequential posted price mechanisms with entry fees (ASPEs) as defined
in Section 6. The mechanisms learned for symmetric subadditive bidders are RSPMs. RSPMs maintain a price
pij for every bidder and item pair and, in some order over bidders i = 1, . . . , n, they give one opportunity to
bidder i to purchase one item j that has not been purchased yet at price pij . ASPEs maintain one price pj for
every item and, in some order over bidders i = 1, . . . , n, they give one opportunity to bidder i to purchase any
subset S ′ of the items S that have not been purchased yet as long as he also pays an “entry fee” that depends
on S and the identity of the bidder. See Algorithm 3.
Learning without Samples. Thus far, our algorithms used samples from the valuation distributions to identify an approximately optimal and simple mechanism under item-independence. However, having sample access to the distributions may be impractical. Often we can observe the actions used by bidders in non-truthful
auctions that were previously run, and use these observations to estimate the distributions over valuations using
econometric methods [30, 37, 4]. In fact, it may likely be the case we have never sold all the items together in
the past, and only have observations of bidder behavior in non-truthful auctions selling each item separately.
Econometric methods would achieve better approximations in this case, but only for the item marginals. Finally, we may want to combine multiple sources of information about the distributions, combining past bidder
behavior in several different auctions and with market research data.
With this motivation in mind, we would like to extend our learnability results beyond the setting where
sample access to the valuation distributions is provided. We propose “learning” approximately optimal multiitem auctions given distributions that are close to the true distributions under some distribution distance d(·, ·).
In particular, given approximate distributions D̂1 , . . . , D̂n over bidder valuations, we are looking to identify a
mechanism M satisfying the following max-min style objective:
∀D1 , . . . , Dn s.t. d(Di , D̂i ) ≤ ǫ, ∀i : RevM (D1 , . . . , Dn ) ≥ Ω(OPT(D1 , . . . , Dn )) − poly(ǫ, m, n). (1)
That is, we want to find a mechanism M whose revenue is within a constant multiplicative and a poly(ǫ, m, n)
additive error from optimum, simultaneously in all possible worlds D1 , . . . , Dn , where d(Di , D̂i ) ≤ ǫ, ∀i. It
is not a priori clear that such a “one-fits-all” mechanism actually exists.
There are several notions of distance d(·, ·) between distributions that we could study in the formulation
of Goal (1), but we opt for an easy one to satisfy. We only require that we know every bidder’s marginal
distributions over single-item values to within ǫ in Kolmogorov distance;3 see Definition 2. All that this
requires is that the cumulative density functions of the approximating distributions over single-item values
is within ǫ in infinity norm from the corresponding cumulative density functions of the corresponding true
distributions. As such, it is an easy property to satisfy. For example, given sample access to any single-item
marginal, the DKW inequality [25] implies that O(log(1/δ)/ǫ2 ) samples suffice to learn it to within ǫ in
Kolmogorov distance, with probability at least 1 − δ. So achieving Goal (1) directly also implies polynomial
sample learnability of approximately optimal auctions. But a Kolmogorov approximation can also be arrived
at by combining different sources of information about the single-item marginals such as the ones described
above. Regardless of how the approximations were obtained, the max-min goal outlined above guarantees
3
Indeed, Goal (1) is achievable only for bounded distributions even in the single-item single-bidder setting. Given any bounded
distribution D̂, create D by moving ǫ probability mass in D̂ to +∞. It is not hard to see that D and D̂ are within ǫ in Kolmogorov
distance, but no single mechanism can satisfy the approximation guarantee for both D and D̂ simultaneously. Using a similar argument,
we can argue that the additive error has to depend on H which is the upper bound on any bidder’s value for a single item. See Section 2
for our formal model.
3
robustness of the revenue of the identified mechanism M with respect to all sources of error that came into the
estimation of the single-item marginal distributions.
While Goal (1) is not a priori feasible, we show how to achieve it in multi-item multi-bidder settings with
constrained additive bidders, or symmetric bidders with subadditive valuations, under the standard assumption
of item-independence. Our results are polynomial-time in the same cases as our sample-based results discussed
above.
Roadmap and Technical Ideas. In Section 4, we present a new approach for obtaining uniform convergence
bounds for hypotheses classes under product distributions; see Theorem 2 and Corollary 1. We show that our
approach can significantly improve the sample complexity bound obtained via traditional methods such as
VC theory. In particular, Table 3 compares the sample complexity bounds obtained via our approach to those
obtained by VC theory for different classes of hypotheses.
Our results for mechanisms make use of recent work on the revenue guarantees of simple mechanisms,
which are mainly variants of sequential posted pricing mechanisms [11, 13]. Using our results from Section 4, in Section 5, we derive uniform convergence bounds for the revenue of a class of mechanisms shown to
achieve a constant fraction of optimal revenue when all bidders have valuations that are constrained additive
over independent items. These mechanisms are called Sequential Posted Price with Entry Fee Mechanisms,
a.k.a. SPEMs,4 . As a corollary of the uniform convergence of SPEMs, we obtain our sample based results
for constrained additive bidders. In fact, we obtain a slightly stronger statement than uniform convergence of
the revenue of SPEMs, which also implies our max-min results for constrained-additive bidders; see Theorems 3 and 4. In particular, Theorem 4 and the DKW inequality imply the polynomial-sample learnability of
approximately revenue-optimal auctions for constrained additive bidders.
Technically speaking, our sample based and max-min approximation results for constrained additive bidders provide a crisp illustration of how we leverage item-independence and our new uniform convergence
bounds for product measures to sidestep the exponential pseudo-dimension of the class of mechanisms that
we are optimizing over. Let us discuss our max-min results which are stronger. Suppose Di = ×j Dij is the
true distribution over bidder i’s valuation and D̂i = ×j D̂ij is the approximating distribution, where Dij and
D̂ij are respectively the item j marginals. To argue that the revenue of some anonymous sequential posted
price with entry fees (ASPE) mechanism is similar under D = ×i Di and D̂ = ×i D̂i , we need to couple in
total variation distance the decisions of what sets all bidders buy in the execution of the mechanism under
D and D̂. The issue that we encounter is that there are exponentially many subsets each bidder may buy,
hence the naive use of the Kolmogorov bound ||Dij − D̂ij ||K ≤ ǫ, on each single-item marginal results in
an exponential blow-up in the total variation distance of what subset of items bidder i buys, invalidating our
desired coupling. To circumvent this challenge, we argue in Lemma 4 that the events corresponding to which
subset of items each buyer will buy are single-intersecting, according to Definition 4, when seen as events on
the buyer’s single-item values. Single-intersecting events may be non-convex and have infinite VC dimension.
Nevertheless, because single-item values are independent, our new uniform convergence bounds for product
measures (Lemma 3) imply that the difference in probabilities of any such event under D and D̂ is only a
factor of m, the number of items, larger than the bound ǫ on the Kolmogorov distance between single-item
marginals.
We specialize our results to unit-demand bidders in Section 5.1 to obtain computationally efficient solutions for both max-min and sample-based models. Similarly, Section 5.2 contains our results for additive
bidders. We also generalize our sample-based results for constrained additive bidders to XOS bidders in Section 6. Finally, we provide computationally efficient solutions for symmetric XOS and even symmetric subadditive bidders in Section 7. These results are based on showing that (i) the right parameters of SPEMs can
be efficiently and approximately identified with sample or max-min access to the distributions; and (ii) that
the revenue guarantees of simple mechanisms can be robustified to accommodate error in the setting of the
4
Note that any RSPM or ASPE is an SPEM.
4
parameters. In particular, our sample-based result for unit-demand bidders robustifies the ex-ante relaxation
of the revenue maximization problem from [1] and its conversion to a sequential posted pricing mechanism
from [15], and makes use of the extreme-value theorem for regular distributions from [7]. Our sample-based
result for additive bidders shows how to use samples to design mechanisms that approximate the revenue of
Yao’s VCG with entry fees mechanism [42]. Our sample-based results for XOS bidders show how to use
samples to approximate the parameters of the RSPM and ASPE mechanisms of [13], and argue, by re-doing
their duality proofs, that their revenue guarantees are robust to errors in the approximation. Finally, our sample based result for symmetric subadditive bidders is based on a new, duality based, approximation, showing
how to eliminate the use of ASPEs from the result of [13]. This even allows us to obtain prior-independent
mechanisms when the item marginals are regular.
2 Preliminaries
We focus on revenue maximization in the combinatorial auction with n independent bidders and m heterogenous items. Each bidder has a valuation that is subadditive over independent items (see Definition 1). We
denote bidder i’s type ti as htij im
j=1 , where tij is bidder i’s private information about item j. For each i, j,
we assume tij is drawn independently from the distribution Dij . Let Di = ×m
j=1 Dij be the distribution of
bidder i’s type and D = ×ni=1 Di be the distribution of the type profile. We use Tij (or Ti , T ) and fij (or fi , f )
to denote the support and density function of Dij (or Di , D). For notational convenience, we let t−i to be
the types of all bidders except i. Similarly, we define D−i , T−i and f−i for the corresponding distributions,
support sets and density functions. When bidder i’s type is ti , her valuation for a set of items S is denoted by
vi (ti , S). Throughout the paper we use OPT to denote the optimal revenue obtainable by any randomized and
Bayesian truthful mechanism.
Definition 1. [40] For every bidder i, whose type ti is drawn from a product distribution Fi = ×j Fij , her
distribution, Vi , over valuation functions vi (ti , ·) is subadditive over independent items if:
- vi (·, ·) has no externalities, i.e., for each ti ∈ Ti and S ⊆ [m], vi (ti , S) only depends on htij ij∈S , formally,
for any t′i ∈ Ti such that t′ij = tij for all j ∈ S, vi (t′i , S) = vi (ti , S).
- vi (·, ·) is monotone, i.e., for all ti ∈ Ti and U ⊆ V ⊆ [m], vi (ti , U ) ≤ vi (ti , V ).
- vi (·, ·) is subadditive, i.e., for all ti ∈ Ti and U , V ⊆ [m], vi (ti , U ∪ V ) ≤ vi (ti , U ) + vi (ti , V ).
We use Vi (tij ) to denote vi (ti , {j}), as it only depends on tij . When vi (ti , ·) is XOS (or constrained
additive) for all i and ti ∈ Ti , we say Vi is XOS (or constrained additive) over independent items.
Example 1. [40] We may instantiate Definition 1 to define restricted families of subadditive valuations as
follows. In all cases, suppose t = {tj }j∈[m] is drawn from ×j Dj . To define a valuation function that is:
- unit-demand, we can take tj to be the value of item j, and set v(t, S) = maxj∈S tj .
P
- additive, we can take tj to be the value of item j, and set v(t, S) = j∈S tj .
- constrained additive, we can take tj to be the value of item j, and set v(t, S) = maxR⊆S,R∈I
some downward closed set system I ⊆ 2[m] .
(k)
P
j∈R tj ,
for
- XOS (a.k.a. fractionally subadditive), we can take tj = {tj }k∈[K] to encode all possible values associated
P
(k)
with item j, and take v(t, S) = maxk∈[K] j∈S tj .
Note that constrained additive valuations contain additive and unit-demand valuations as special cases,
and are contained in XOS valuations.
Distribution Access Models
We consider the following three different models to access the distributions.
5
• Sample access to bounded distributions. We assume that for any buyer i and any type ti ∈ Ti , her
value Vi (tij ) for any single item j lies in [0, H].
• Sample access to regular distributions. We assume that for any buyer i and any type ti ∈ Ti , the
distribution of her value Vi (tij ) for any item j is regular.
• Direct access to approximate distributions. We assume that we have direct access to a distribution
D̂ = ×i∈[n],j∈[m]D̂ij , for example we can query the pdf, cdf of D̂ and take samples from D̂. Moreover,
for any buyer i and any type ti ∈ Ti , the distributions of the random variable Vi (tij ) when tij is sampled
from D̂ij or Dij are within ǫ in Kolmogorov distance, and both distributions are supported on [0, H].
Definition 2. The Kolmogorov distance between two distributions P and Q over R, denoted ||P − Q||K , is
defined as supx∈R | PrX∼P [X ≤ x] − PrX∼Q [X ≤ x]|. The total variation distance between two probability
measures P and Q on a sigma-algebra F of subsets of some sample space Ω, denoted ||P − Q||T V , is defined
as supE∈F |P (E) − Q(E)|.
3 Summary of Our Results
We summarize our results in the following two tables. Table 1 contains all sample-based results and Table 2
contains all results under the max-min learning model.
Valuations
# bidders
Distributions
Approximation
Sample Complexity
additive [28]
additive
n
n
regular
arbitrary [0, H]
Ω(OPT)
Ω(OPT) − ǫ · H
1
poly(n, m, 1/ǫ)
unit-demand [35]
unit-demand
n
n
arbitrary [0, H]
regular
Ω(OPT) − ǫ · H
Ω(OPT)
poly(n, m, 1/ǫ)
poly(n, m)
constrained additive
constrained additive
n
n
arbitrary [0, H]
regular
Ω(OPT) − ǫ · H
Ω(OPT)
poly(n, m, 1/ǫ)
poly(n, m)
XOS
XOS
n
n
arbitrary [0, H]
regular
Ω(OPT) − ǫ · H
Ω(OPT)
poly(n, m, 1/ǫ)
poly(n, m)
subadditive [35]
1
arbitrary [0, H]
poly(m, 1/ǫ)
subadditive
n i.i.d.
arbitrary [0, H]
subadditive
n i.i.d.
regular
Ω(OPT)
−ǫ·H
n
Ω max{n,m} · OPT − ǫ · H
n
· OPT
Ω max{n,m}
poly(n, m, 1/ǫ)
prior-independent
Table 1: Summary of Our Sample-based Results.
4 Uniform Convergence under Product Measures
In this section, we develop machinery for obtaining uniform convergence bounds for hypotheses over product
measures. Our goal is to save on the sample complexity implied by VC dimension bounds, as summarized
in Table 3. Indeed, we obtain low sample complexity bounds for indicators over single-intersecting sets (see
6
Valuations
# bidders
Distributions
Approximation
additive
n
arbitrary [0, H]
Ω(OPT) − O(ǫ · n · m · H)
unit-demand
n
arbitrary [0, H]
Ω(OPT) − O(ǫ · n · m · H)
constrained additive
n
arbitrary [0, H]
Ω(OPT) − O(ǫ · n · m2 · H)
subadditive
n i.i.d.
arbitrary [0, H]
Ω
n
max{n,m}
· OPT − O(ǫ · n · m · H)
Table 2: Summary of Our Max-min Learning Results.
Definition 4), which play a key role in proving our results for learning approximately revenue-optimal auctions.
Our main results of this section are Theorem 2 for general functions, and Corollary 1 for sets.
We first define what type of uniform convergence bounds we seek to prove.
Definition 3 ((ǫ, δ)-uniform convergence with respect to proxy measure). A hypothesis class H of functions
mapping domain set X to R has (ǫ, δ)-uniform convergence with sample complexity s(ǫ, δ) iff, for all ǫ, δ > 0,
there exists a processing P : X s(ǫ,δ) → ∆(X ) such that for any distribution D ∈ ∆(X ) when k = s(ǫ, δ):
"
#
Pr
z1 ,··· ,zk ∼D
sup Ez∼P(z1 ,··· ,zk ) [g(z)] − Ez∼D [g(z)] ≤ ǫ ≥ 1 − δ.
g∈H
When X is the Cartesian product of a collection of sets X1 , . . . , Xk , i.e. X = ×i Xi , we say that a hypothesis
class H as above has (ǫ, δ)-p.m. uniform convergence with sample complexity s(ǫ, δ) if the above holds for
all D that are product measures over X .
Next we provide a simple lemma, which leads to a simple version of our main result stated as Theorem 1.
Our main result, stated as Theorem 2, follows.
Lemma 1. Let X1 , . . . , Xd be d domain sets and H be a hypothesis class with functions mapping from the
product space ×di=1 Xi to R. For all i ∈ [d], let Hi be the projected hypothesis class of H on Xi , that is,
Hi = {g | ∃f ∈ H ∃ a−i ∈ ×j6=i Xj ∀ xi ∈ Xi , g(xi ) = f (xi , a−i )}. For every i ∈ [d], let Di and D̂i be two
distributions supported on Xi . Suppose for all i ∈ [d],
sup Ex∼Di [g(x)] − Ex∼D̂i [g(x)] ≤ ǫ,
g∈Hi
then
sup Ex∼×d
f ∈H
i=1 Di
[f (x)] − Ex∼×d
i=1 D̂i
[f (x)] ≤ d · ǫ.
Proof. Let Fi and F̂i be the probability measure function for Di and D̂i respectively. We will prove the
statement using a hybrid argument. We create a sequence of product distributions {D (j) }j≤d , where D (j) =
D̂1 × · · · × D̂j × Dj+1 × · · · × Dd , and D (0) = D, D (d) = D̂. To prove our claim, it suffices to show that for
any integer j ∈ [d],
|Ex∼D(j−1) [f (x)] − Ex∼D(j) [f (x)]| ≤ ǫ.
7
Next, we show how to derive this inequality.
|Ex∼D(j−1) [f (x)] − Ex∼D(j) [f (x)]|
!
Z
Z
f (xj , x−j )dFj (xj ) dF̂1 (x1 ) · · · dF̂j−1 (xj−1 )dFj+1 (xj+1 ) · · · dFd (xd )
=
Xj
×i6=j Xi
−
=
Z
≤ǫ ·
×i6=j Xi
Z
Z
Z
Xj
×i6=j Xi
!
f (xj , x−j )dF̂j (xj ) dF̂1 (x1 ) · · · dF̂j−1 (xj−1 )dFj+1 (xj+1 ) · · · dFd (xd )
Exj ∼Dj [f (xj , x−j )] − Exj ∼D̂j [f (xj , x−j )] dF̂1 (x1 ) · · · dF̂j−1 (xj−1 )dFj+1 (xj+1 ) · · · dFd (xd )
dF̂1 (x1 ) · · · dF̂j−1 (xj−1 )dFj+1 (xj+1 ) · · · dFd (xd )
×i6=j Xi
=ǫ
Theorem 1. Let X1 , . . . , Xd be d domain sets and H a hypothesis class of functions mapping from the product
space ×di=1 Xi to R. For all i ∈ [d], let Hi be the projected hypothesis class of H on Xi , that is Hi =
{g | ∃f ∈ H ∃ a−i ∈ ×j6=i Xj ∀ xi ∈ Xi , g(xi ) = f (xi , a−i )}.
Suppose that, for all i ∈ [d], Hi has (ǫ, δ)-uniform convergence with sample complexity si (ǫ, δ). Then H
has (ǫ, δ)-p.m. uniform convergence with sample complexity s(ǫ, δ) = maxi∈[d] si (ǫ/d, δ/d).
In particular, let z (1) , . . . , z (ℓ) be a sample of size ℓ = s(ǫ, δ) from a product measure ×i∈[d] Di . Define
(1)
(ℓ)
(j)
D̂i = Pi (zi , . . . , zi ), for all i ∈ [d], where zi
corresponding to Hi ’s uniform convergence. Then
"
Pr
z (1) ,...,z (ℓ)
is the i-th entry of sample z (j) and Pi is the processing
#
sup Ez∼×i∈[d] D̂i [f (z)] − Ez∼×i∈[d]Di [f (z)] ≤ ǫ ≥ 1 − δ.
f ∈H
i
h
Proof of Theorem 1: Since ℓ ≥ si (ǫ/d, δ/d), Pr supg∈Hi Ez∼D̂i [g(z)] − Ez∼Di [g(z)] ≤ ǫ/d ≥ 1−δ/d for
all i ∈ [d]. By the union bound, with probability at least 1 − δ, supg∈Hi Ez∼D̂i [g(z)] − Ez∼Di [g(z)] ≤ ǫ/d
for all i ∈ [d]. According to Lemma 1, supf ∈H Ez∼× D̂i [f (z)] − Ez∼×i∈[d] Di [f (z)] ≤ ǫ with probability
i∈[d]
at least 1 − δ. ✷
Theorem 2. Let X1 , . . . , Xd be d domain sets and H a hypothesis class of functions mapping from the product
d
space ×
i=1 Xi to R. For all T ⊆ [d], let HT be the projected hypothesis class of H on XT ≡ ×i∈T Xi , that is,
HT = g | ∃f ∈ H ∃ a−T ∈ ×j ∈/ T Xj ∀ xT ∈ XT , g(xT ) = f (xT , a−T ) . Suppose that, for all T ⊆ [d], HT
has (ǫ, δ)-p.m. uniform convergence with sample complexity sT (ǫ, δ), and define
s(ǫ, δ) =
min
max sTi (ǫ/k, δ/k).
i=1,...,k
k, partitions
T1 ⊔T2 ⊔. . .⊔Tk = [d]
(2)
Then H has (ǫ, δ)-p.m. uniform convergence with sample complexity s(ǫ, δ).
In particular, let z (1) , . . . , z (ℓ) be a sample of size ℓ = s(ǫ, δ) from a product measure ×i∈[d] Di . Suppose that the optimum of (2) is attained at k = k̃ for partition T̃1 ⊔ T̃2 ⊔ . . . ⊔ T̃k̃ = [d]. Define D̂T̃i =
8
(1)
(j)
(ℓ)
PT̃i (z T̃ , . . . , z T̃ ), for all i ∈ [d], where z T̃ contains the entries of sample z (j) in coordinates T̃i and PT̃i is
i
i
i
the processing corresponding to HT̃i ’s uniform convergence. Then
Pr
z (1) ,...,z (ℓ)
"
sup Ez∼×
f ∈H
i∈[k̃] D̂T̃i
#
[f (z)] − Ez∼×i∈[d] Di [f (z)] ≤ ǫ ≥ 1 − δ.
Proof of Theorem 2: For every possible partition use Theorem 1. ✷
Nest, we specialize Theorem 2 to indicator functions over sets.
Corollary 1. We use the same notation as in Theorem 2. Suppose that all functions in H map ×di=1 Xi to
{0, 1}, i.e. they are indicators over sets. Suppose also that the VC dimension of HT (viewed as a collection of
sets) is VT . Define
Vmax =
min
k2 · max VTi .
(3)
i=1,...,k
k, partitions
T1 ⊔T2 ⊔. . .⊔Tk = [d]
Assume that the
optimum of (3) is attained
at k = k̃ for partition T̃1 ⊔ T̃2 ⊔ . . . ⊔ T̃k̃ = [d].
· ln k̃ǫ +
Then ℓ = O Vmax
ǫ2
gence for H. Formally,
"
Pr
z (1) ,...,z (ℓ)
k̃ 2
ǫ2
· ln k̃δ samples from ×i∈[d] Di suffice to obtain (ǫ, δ)-p.m. uniform conver-
sup Ez∼×
f ∈H
i∈[k̃] D̂T̃i
#
[f (z)] − Ez∼×i∈[d] Di [f (z)] ≤ ǫ ≥ 1 − δ,
where for a given sample z (1) , . . . , z (ℓ) from a product distribution ×i∈[d] Di the distributions D̂T̃i are defined
(1)
(ℓ)
(j)
i
i
i
to be uniform over z T̃ , . . . , z T̃ , where z T̃ contains the entries of sample z (j) in coordinates T̃i .
Table 3 compares the sample complexity for uniform convergence implied by Theorem 2 and Corollary 1
to that implied by VC theory, when the underlying measures are product. Suppose H contains the indicator
functions of all convex sets in Rd . VC theory does not provide any finite sample bound for uniform convergence, as the VC dimension of H is ∞. Do our results provide a finite bound? Notice that, for all i, H
i
2
simply contains all intervals in R. Hence, Vi = 2 and Corollary 1 implies that ℓ = O( dǫ2 · log dδ + log dǫ )
samples suffice to obtain (ǫ, δ)-p.m.
uniform
convergence for H . In fact, our sample complexity bound can
2
log
1
be improved to O dǫ2 · log dδ , as O ǫ2 δ samples suffice to guarantee (ǫ, δ)-uniform convergence for all
intervals in R due to the DKW inequality [25].
In the next a few sections, we apply our uniform convergence results to learn a mechanism with approximately optimal revenue. A type of events called single-intersecting (see Definition 4) plays a key role in our
analysis. These events are defined based on the geometric shape of the corresponding sets. For example, balls,
rectangles and all convex sets are single-intersecting, but this definition includes some non-convex sets as well,
for example, “cross-shaped” sets. It turns out that being able to handle these non-convex sets is crucial for our
results, as many events we care about are not convex but nonetheless are single-intersecting.
Definition 4 (Single-intersecting Events). For any event E in Rℓ , E is single-intersecting if the intersection of E and any line
an interval. More formally, for any i ∈ [ℓ]
thatℓ is parallel to one of the axes is
ℓ−1 , the intersection of L and E is of the form
,
where
a
∈
R
and
any
line
L
=
x
∈
R
|
x
=
a
−i
i
i
−i
−i
x ∈ Rℓ | x−i = a−i , xi ∈ [a, ā] where a ≤ ā. In particular, we allow a to be −∞ and ā to be +∞.
¯
¯
¯
9
We establish a uniform convergence bound for single-intersecting events by combing the DKW inequality and
Theorem 1.
Lemma 2. For any integer ℓ, let H be the hypothesis class that contains all indicator functions
single for
ℓ2
ℓ
intersecting events in R . Then H has (ǫ, δ)-p.m. uniform convergence with sample complexity O ǫ2 · log δℓ .
Proof. As the projected hypothesis class for the i-th coordinate simply contains all intervals in R, the sample
complexity for (ǫ, δ)-uniform convergence is O( ǫ12 · log 1δ ) due to the DKW inequality. The claim follows from
Theorem 1.
Next, we show a slightly stronger statement, which is a type of uniform convergence bound when access
to approximate distributions is given. More specifically, we argue that for any single-intersecting event, the
difference in the probability of this event under two product distributions D = ×i∈[ℓ] Di and D̂ = ×i∈[ℓ] D̂i is
at most 2ξ · ℓ, if ||Di − D̂i ||K ≤ ξ for all i. It is not hard to see that Lemma 3 and the DKW inequality imply
Lemma 2.
Lemma 3. For any integer ℓ, let D = ×ℓi=1 Di and D̂ = ×ℓi=1 D̂i , where Di and D̂i are both supported on R
for any i ∈ [ℓ]. If ||Di − D̂i ||K ≤ ξ, PrD [E] − PrD̂ [E] ≤ 2ξ · ℓ for any single-intersecting event E.
Proof. Let H = {1x∈E : E is single-intersecting}. By the definition of single-intersecting events, Hi is the set
of the indicator functions of all intervals in R for any i ∈ [ℓ]. Since ||Di − D̂i ||K ≤ ξ,
sup Ex∼Di [g(x)] − Ex∼D̂i [g(x)] ≤ 2ξ.
g∈Hi
By Lemma 1,
sup Ex∼D [f (x)] − Ex∼D̂ [f (x)] ≤ 2ξ · ℓ.
f ∈H
The following table (Table 3) summarizes some uniform convergence bounds implied by our results in this
section.
Hypotheses Class
VC Bound
Bounds from Theorem 2 and Corollary 1
axis-aligned rectangles in Rd
polytopes with k facets in Rd
arbitrary convex sets in Rd
single-intersecting sets in Rd
Õ(d/ǫ2 )
Õ(dk/ǫ2 )
∞
∞
Õ(d/ǫ2 )
Õ(d · min{d, k}/ǫ2 )
Õ(d2 /ǫ2 )
Õ(d2 /ǫ2 )
Table 3: Number of samples required for (ǫ, Θ(1))-p.m. uniform convergence for different H’s.
5 Constrained Additive Bidders: Uniform Convergence of the Revenue of Sequential Posted Price with Entry Fee Mechanisms
We consider a specific class of mechanisms, namely Sequential Posted Price with Entry fee Mechanisms, a.k.a.
SPEMs; see Algorithm 1 for details. Cai and Zhao [13] recently showed that if the bidders’ valuations are XOS
10
over independent items, the best SPEM achieves a constant fraction of the optimal revenue. 5 This section
has two goals. The first is to show that, when bidders have constrained additive valuations over independent
items, polynomially many samples suffice to guarantee uniform convergence for the revenue of all SPEMs,
and hence our ability to select a near-optimal SPEM from polynomially many samples. This can be proven by
applying our uniform convergence result for single-intersecting events (Lemma 2). The second (and stronger
goal) is to show that we can learn a near-optimal SPEM under the max-min learning model (Theorem 4). We
show that the revenue of any SPEM changes no more than O(ǫ · m2 · n · H) under the true and approximate
valuation distributions (Theorem 3), where ǫ is an upper bound of the Kolmogorov distance between the true
and approximate distributions for every item marginal of every bidder. It is, of course, not hard to see that
Theorem 3 and the DKW inequality imply uniform convergence of the revenue of all SPEMS. To establish
Theorem 3, we need to apply Lemma 3 instead of Lemma 2.
Algorithm 1 Sequential Posted Price with Entry Fee Mechanism (SPEM)
Require: A collection of prices {pij }i∈[n],j∈[m] and a collection of entry fee functions {δi (·)}i∈[n] where
δi : 2[m] 7→ R is bidder i’s entry fee function.
1: S ← [m]
2: for i ∈ [n] do
3:
Show bidder i the set of available items S and set the entry fee for bidder i to be δi (S).
4:
if Bidder i pays the entry fee δi (S) then
P
5:
i receives her favorite bundle Si∗ and pays j∈S ∗ pij .
i
6:
S ← S\Si∗ .
7:
else
8:
i gets nothing and pays 0.
9:
end if
10: end for
We first establish a technical lemma, which states that, for any set of items S, any set of prices {pj }j∈[m]
and entry fee δ, the distribution over the set of items purchased by a constrained additive bidder whose valuation
is drawn from D = ×j∈[m] Dj and D̂ = ×j∈[m] D̂j has total variation distance at most 2mξ, if ||Dj − D̂j ||K ≤ ξ
for every item j ∈ [m]. This is quite surprising. Given that, for each set of items S ′ ⊆ S, the difference in
the probability that the buyer will purchase this particular set S ′ under D and D̂ could already be as large as
Θ(mξ), and the distribution has an exponentially large support size, a trivial argument would give a bound of
2m · Θ(mξ). To overcome this analytical difficulty, we argue instead that for any collection of sets of items,
the event that the buyer’s favorite set lies in this collection is single-intersecting. Then our result follows from
Lemma 3. Notice that it is crucial that Lemma 3 holds for all events that are single-intersecting, as the event
we consider here is clearly non-convex in general.
Lemma 4. For any set S ⊆ [m], any prices {pj }j∈[m] and entry fee δ(S), let L and L̂ be the distributions
over the set of items purchased from S by a constrained additive bidder under prices {pj }j∈[m] and entry fee
δ when her type is drawn from D = ×j∈[m] Dj and D̂ = ×j∈[m] D̂j respectively. If ||Dj − D̂j ||K ≤ ξ for all
item j, ||L − L̂||T V ≤ 2mξ.
Proof. For any set R ⊆ S, let ER be the event that the bidder purchases set R. Proving that the total vari|S|
ation distance
between
h
i L andh L̂ isS no morei than 2m · ξ is the same as proving that for any K ≤ 2 ,
SK
≤ 2m · ξ where R1 , · · · RK are arbitrary distinct subsets of
PrD t ∈ ℓ=1 ERℓ − PrD̂ t ∈ K
ℓ=1 ERℓ
5
Cai and Zhao [13] showed that the best ASPE or RSPM achieves a constant fraction of the optimal revenue. Clearly, any ASPE is
also a SPEM, and any RSPM is simply a SPEM if we force the bidders to be unit-demand by only allowing each of them to purchase
at most one item.
11
S
S. Since the dimension of the bidder’s type space is m, if we can prove that K
ℓ=1 ERℓ is always singleintersecting, our claim follows from Lemma 3.
For any j ∈ [m] and a−j ∈ Rm−1
≥0 , let Lj (a−j ) = {(tj , a−j ) |tj ∈ R≥0 }. We claim that Lj (a−j ) intersects
with at most two different EU and EV where U and V are subsets of S. WLOG, we assume that (0, a−j ) ∈ EU .
• If U = ∅, that means the utility of the favorite set for type (0, a−j ) is smaller than the entry fee δ(S). If
we increase the value of tj , two cases could happen: (1) the utility of the favorite set is still lower than
the entry fee; (2) the utility of the favorite set is higher than the entry fee. In case (1), (tj , a−j ) ∈ E∅ .
In case (2), the bidder pays the entry fee and purchases her favorite set V . Then item j must be in
V , because otherwise the utility for set V does not change from type (0, a−j ) to type (tj , a−j ). If we
keep increasing tj , bidder i’s favorite set remains to be V and she keeps accepting the entry fee and
purchasing V . Hence, Lj (a−j ) can intersect with at most one event ER where R is non-empty.
• If U 6= ∅, that means U is the favorite set of type (0, a−j ) and the utility for winning set U is higher than
the entry fee. If we increase the value of tj , two cases could happen: (1) U remains the favorite set; (2)
a different set V becomes the new favorite set. In case (1), (tj , a−j ) ∈ EU . In case (2), item j must lie in
V but not in U , otherwise how could U be better than V for type (0, a−j ) but worse for type (tj , a−j ).
If we keep increasing tj , the bidder’s favorite set remains to be V and she keeps accepting the entry fee
and purchasing V . Hence, Lj (a−j ) can intersect at most two different events.
It is not hard to see that any event ER is an intersection of halfspaces, so the intersection of Lj (a−j ) with
any event ER is an interval. Also, notice that any type t ∈ Rm
≥0 must lie in an event ER for some set R ⊆ S.
If Lj (a−j ) intersects with two different events EU and EV , the two intersected intervals must lie back to back
on Lj (a−j ). Otherwise, Lj (a−j ) intersects with at least three different events. Contradiction. Since Lj (a−j )
intersects with at most two different events, no matter which of these events are in {ERℓ }ℓ∈[K] , the intersection
SK
S
of Lj (a−j ) and K
ℓ=1 ERℓ is single-intersecting. Now our
ℓ=1 ERℓ is either empty or an interval meaning
claim simply follows from Lemma 3.
Theorem 3. Suppose all bidders’ valuations are constrained additive over independent items. For any SPEM,
let R EV and Rd
EV be its expected revenue under D and D̂ respectively. If Dij and D̂ij are both supported on
[0, H], and ||Dij − D̂ij ||K ≤ ξ for all i ∈ [n] and j ∈ [m],
R EV − Rd
EV ≤ 2nmξ · (mH + OPT) .
Proof. We use a hybrid argument. Consider a sequence of distributions {D (i) }i≤n , where D (i) = D̂1 × · · · ×
D̂i × Di+1 × · · · × Dn , and D (0) = D, D (n) = D̂. We use R EV (i) to denote the expected revenue of the SPEM
under D (i) . To prove our claim, it suffices to argue that R EV (i−1) − R EV (i) ≤ 2ξm · (m · H + OPT) . We
denote by Sk and Sk′ the random set of items that remain available after visiting the first k bidders under D (i−1)
and D (i) . Clearly, for k ≤ i − 1, ||Sk − Sk′ ||T V = 0, so the expected revenue collected from the first i − 1
bidders under D (i−1) and D (i) is the same. According to Lemma 4, ||Si − Si′ ||T V ≤ 2m · ξ. The total amount
of money bidder i spends can never be higher than her value for receiving all the items which is at most m · H.
So the difference in the expected revenue collected from bidder i under D (i−1) and D (i) is at most 2ξ · m2 H.
Suppose R is the set of remaining items after visiting the first i bidders, then the expected revenue collected
from the last n − i bidders is the same under D (i−1) and D (i) , as these bidders have the same distributions.
Moreover, this expected revenue is no more than OPT, since the optimal mechanism can simply just sell R
to the last n − i bidders using the same prices and entry fee as in the SPEM we consider. Of course, for any
fixed R, the probabilities that Si = R and Si′ = R are different, but since for any R the expected revenue from
the last n − i bidders is at most OPT, the difference in the expected revenue from the last n − i bidders under
12
D (i−1) and D (i) is at most ||Si − Si′ ||T V · OPT ≤ 2ξ · mOPT. Hence, the total difference between R EV (i−1)
and R EV (i) is at most 2ξm · (mH + OPT).
Theorem 4. (Max-min Learning for Constrained Additive Bidders) When all bidders’ valuations are constrained additive over independent items and for any bidder i and any item j, Dij and D̂ij are supported on
1
), then with only access to D̂ = ×i,j D̂ij , our algorithm can
[0, H] and ||Dij − D̂ij ||K ≤ ǫ for some ǫ = O( nm
2
learn an RSPM or ASPE whose revenue is at least OPT
c − ǫ · O(m nH), where OPT is the optimal revenue by
any BIC mechanism under D = ×i,j Dij . c > 1 is an absolute constant.
Clearly, Theorem 4 also implies a polynomial sample complexity bound for learning an approximately
revenue-optimal mechanism. A better sample complexity bound can be obtained directly, i.e. without invoking
the uniform convergence of the revenue of SPEMs, and is stated as Theorem 9 for the broader class of XOS
valuations. Similarly, when bidders have simpler valuations, i.e., additive or unit-demand valuations, we can
sharpen our results and achieve polynomial-time learnability of the approximately optimal mechanism using
more specialized techniques. See Sections 5.1 and 5.2 for details.
5.1 Unit-demand Valuations: Polynomial-Time Learning
In this section, we consider bidders with unit-demand valuations, sharpening our results to show how to learn
approximately revenue-optimal mechanisms in polynomial time. It is shown in a sequence of works [15, 32,
11] that there exists a sequential posted price mechanism (SPM see Algorithm 2 for details) that achieves at
1
of the optimal revenue when bidders are unit-demand. We show that under all three distribution access
least 24
models of Section 2 there exists a polynomial-time algorithm that learns a sequential posted price mechanism
whose revenue approximates the optimal revenue. We only sketch the proof here and postpone the details to
Appendix B.
Theorem 5. When all bidders have unit-demand valuations and
• Dij is supported on [0, H] for all bidder i and item j, there exists a polynomial
algorithm that learns
time
1 2
OPT
m2 n log nǫ + log 1δ
an SPM whose revenue is at least 144 − ǫH with probability 1 − δ given O ǫ
samples from D; or
• Dij is a regular distribution for all bidder i and item j, there exists a polynomial time algorithm that
2 2 2
learns a randomized SPM whose revenue is at least OPT
33 with probability 1−δ given O(max{m, n} m n ·
nm
log δ ) samples from D; or
• we are only given access to D̂ij where ||D̂ij −Dij ||K ≤ ǫ for all bidder i and item j, there is a polynomial
time algorithm that constructs a randomized SPM whose revenue under D is at least 14 − (n + m) · ǫ ·
OPT
6
8 − 2ǫ · mnH .
Sample Access to Bounded Distributions: the result is due to Morgenstern and Roughgarden [35].
Direct Access to Approximate Distributions: we first consider a convex program based on D (see Figure 1)
which is usually referred to as the ex-ante relaxation of the revenue maximization problem [1], and use its
optimum as a proxy for OPT. Next, we consider a similar convex program based on D̂ (see Figure 2) and
show that the optima of the two convex programs are close to each other. Finally, we use techniques developed
by Chawla et al. [15] to convert the optimal solution of the second convex program into a randomized SPM.
We can show that the constructed randomized SPM achieves a revenue that approximates the optimum of the
second convex program under D, which implies that the mechanism’s revenue also approximates the OPT. As
6
1
), this is the max-min guarantee we want to achieve.
If we set ǫ to be O( m+n
13
we are given D̂, we can solve the second convex program and convert its optimal solution into a randomized
SPM in polynomial time. See Theorem 10 in Appendix B.1 for further details.
Sample Access to Regular Distributions: we use a similar convex program relaxation based approach as
in the previous case. The main difference is that regular distributions could be unbounded and thus ruin the
approximation guarantee. We show how to use the Extreme Value theorem in [7] to truncate the distributions
without hurting the revenue by much. See Theorem 12 in Appendix B.3 for further details.
5.2 Additive Valuations: Polynomial-Time Learning
In this section, we consider bidders with additive valuations, again sharpening our results to show polynomialtime learnability. It is known that the better of the following two mechanisms achieves at least 81 of the optimal
revenue when all bidders have additive valuations [42, 11]:
Selling Separately: the mechanism sells each item separately using Myerson’s optimal auction.
VCG with Entry Fee: the mechanism solicits bids b = (b1 , · · · , bn ) from the bidders, then offers each
bidder i the option to participate for an entry fee ei (b−i , Di ), which is the median of the random variable
P
7
+
j∈[m] (tij − maxk6=i bkj ) , where ti ∼ Di . This random variable is exactly bidder i’s utility when her type
is ti and the other bidders’ are b−i . If bidder i chooses to participate, she pays the entry fee and can take any
item j at price maxk6=i bkj . Notice that the mechanism never over allocate any item, as only the highest bidder
for an item can afford it.
Indeed, only counting the revenue from the entry fee in the second mechanism and the optimal revenue
from selling the items separately already suffices to provide an 8-approximation [42, 11].
Theorem 6 ([11]). Let SR EV be the optimal revenue for selling the items separately and BR EV be the expected
entry fee collected from the VCG with entry fee mechanism. Then OPT ≤ 6 · SR EV + 2 · BR EV .
Goldner and Karlin [28] showed that one sample suffices to learn a mechanism that achieves a constant
fraction of the optimal revenue when Dij is regular for all i ∈ [n] and j ∈ [m]. We show how to learn an
approximately optimal mechanism in the other two models.
Theorem 7. When the bidders have additive valuations and
• Dij is supported on [0, H] for all bidder i and item j, we can learn in polynomial
time a mechanism
m 2
OPT
· n log n log 1ǫ + log 1δ
whose expected revenue is at least 32 −ǫ·H with probability 1−δ given O
ǫ
samples from D; or
• we are only given access to distributions D̂ij where ||D̂ij − Dij ||K ≤ ǫ for all bidder i and item j, there
is a polynomial time algorithm that constructs a mechanism whose expected revenue under D is at least
1
OPT
266 − 96ǫ · mnH when ǫ ≤ 16 max{m,n} .
Sample Access to Bounded Distributions: Goldner and Karlin’s proof [28] can be directly applied to the
bounded distributions to show a single sample suffices to learn a mechanism whose expected revenue approximates the BR EV . Then as SR EV is the revenue of m separate single-item auctions, we can use the result
in [35] to approximate it. See Theorem 13 in Appendix C.1 for further details.
Direct Access to Approximate Distributions: for each single item, we apply Theorem 5 to learn an individual
auction, then run these learned auctions in parallel. Clearly, the combined auction’s revenue approximates
7
The entry fee function defined in [42, 11] is slightly different. They showed that there exists an entry fee Xi , such that bidder
i accepts the entry fee with probability at least 1/2. Then they argued that extracting Xi /2 as the revenue in the VCG with entry
fee mechanism is enough to obtain a factor 8 approximation. It is not hard to observe that our entry fee is accepted with probability
exactly 1/2, thus our entry fee is at least as large as Xi . So our mechanism also suffices to provide a factor 8 approximation.
14
SR EV . For BR EV , we show that for every bidder i and every bid profile b−i of the other bidders, the event that
corresponds to bidder i accepting any entry fee is single-intersecting (see Definition 4). This implies that the
probability forP
a bidder to accept an entry fee under D̂ and D is close (Lemma 3). So we can essentially use
the median of j∈[m] (tij − maxk6=i bkj )+ with ti ∼ D̂i as the entry fee. See Theorem 14 in Appendix C.2 for
further details.
6 XOS Valuations
In this section we go beyond constained additive valuations to show learnability of approximately revenueoptimal auctions from polynomially many samples. The better of the following two mechanisms is known to
achieve a constant fraction of the optimal revenue, when bidders have valuations that are XOS over independent
items [13].
Rationed Sequential Posted Price Mechanism (RSPM): the mechanism is almost the same as SPM in Algorithm 2, except there is an extra constraint that every bidder can purchase at most one item.
Anonymous Sequential Posted Price with Entry Fee Mechanism (ASPE): every buyer faces the same
collection of item prices {pj }j∈[m] . The seller visits the bidders sequentially. For every bidder, the seller
shows her all the available items (i.e. items that have not yet been purchased) and the associated price for
each item, then asks her to pay a personalized entry fee which depends on her type distribution and the set of
available items. If the bidder accepts the entry fee, she can proceed to purchase any available item at the given
price; if she rejects the entry fee, she neither receives nor pays anything. See Algorithm 3 for details.
Theorem 8. [13] There exists a collection of prices {p∗j }j∈[m] , such that if we set the entry fee function
δi∗ (S) to be the median of bidder i’s utility for set S, either the ASPE(p∗ , δ∗ ) or the best RSPM achieves at
least a constant fraction of the optimal revenue when bidders’
valuations are XOS over independent items.
P
More formally, let u∗i (ti , S) = maxS ∗ ⊆S vi (ti , S ∗ ) − j∈S ∗ p∗j be bidder i’s utility for the set of items S
when her type is ti . We define δi∗ (S) to be the median of the random variable u∗i (ti , S) (with ti ∼ Di ) for
any set S ⊆ [m]. Moreover, the price p∗j for any item j is no larger than 2G, where G = maxi,j Gij and
o
n
1
.
Gij := supx Prtij ∼Dij [Vi (tij ) ≥ x] ≥ 5 max{m,n}
Our goal next is to bound the sample complexity for learning a near-optimal RSPM and the ASPE described
in Theorem 8 under XOS valuations.
We consider first the task of learning a near-optimal RSPM. In a RSPM, all bidders are restricted to be
unit-demand, so the revenue of the best RSPM is upper bounded by the optimal revenue in the corresponding
unit-demand setting. In Section 5.1, we have shown how to learn an approximately optimal mechanism for
unit-demand bidders, and those algorithms can be used to approximate the best RSPM.
So, for the rest of this section, it suffices to focus on learning an ASPE whose revenue approximates the
revenue of the ASPE described in Theorem 8. We will do this in Section 6.1. Before that, we need a robust
version of Theorem 8. In the next Lemma, we argue that if we use a collection of prices {p′j }j∈[m] sufficiently
close to {p∗j }j∈[m] and entry fee δi′ (S) sufficiently close to the median of the utility for every bidder i and
subset S, the better of the corresponding ASPE and the best RSPM still approximates the optimal revenue. We
postpone the proof to Appendix D.
Lemma 5. For any ǫ > 0 and µ ∈ [0, 14 ], let {p′j }j∈[m] be a collection of prices such that |p′j − p∗j | ≤ ǫ
for all j ∈ [m], where {p∗j }j∈[m] is the collection of prices in Theorem 8. Let δi′ (S) be bidder i’s entry fee
′
′
function such that Prti ∼D
+ µ] for any set S ⊆ [m], where u′i (ti , S) =
P i [ui (t′i , S) ≥ δi (S)] ∈ [1/2 − µ, 1/2
′
′
∗
maxS∗⊆S vi (ti , S ) − j∈S ∗ pj . Then, either the ASPE(p , δ ) or the best RSPM achieves revenue at least
OPT
C1 (µ) − C2 (µ) · (m + n) · ǫ when bidders’ valuations are XOS over independent items. Both C1 (·) and C2 (·)
are monotonically increasing functions that only depend on µ.
15
Definition 5. We say a collection of prices {pj }j∈[m] is in the B-bounded ǫ-net if pj is a multiple of ǫ and
no larger than B for any item j. For any collection of prices {pj }j∈[m] , we say the entry fee functions are µbalanced if for every bidder i and every set S ⊆ [m], her entryP
fee δi (S) satisfies Prti ∼Di [ui (ti , S) ≥ δi (S)] ∈
[1/2 − µ, 1/2 + µ], where ui (ti , S) = maxS∗⊆S vi (ti , S ∗ ) − j∈S ∗ pj .
Corollary 2. For bidders with valuations that are XOS over independent items and any ǫ > 0, there exists
a collection of prices {pj }j∈[m] in the 2G-bounded ǫ-net such that for any µ-balanced entry fee functions
− C2 (µ) ·
{δi (·)}i∈[n] with µ ∈ [0, 14 ], either the ASPE(p, δ) or the best RSPM achieves revenue at least COPT
1 (µ)
(m + n) · ǫ.
6.1 XOS Valuations: sample access to bounded and regular distributions
In this section, we consider how to learn an ASPE with high revenue given sample access to D. Our learning
algorithm is a two-step procedure. In the first step, we take a few samples from D and use these samples to
set the entry fee for every collection of prices {pj }j∈[m] in the ǫ-net. More specifically, to decide δi (S) we
compute the utility of bidder i for set S under {pj }j∈[m] over all the samples and take the empirical median
among all these utilities to be δi (S). With a polynomial number of samples, we can guarantee that for any
{pj }j∈[m] in the ǫ-net the computed entry fee functions {δi (·)}i∈[n] are µ-balanced. Now, we have created
an ASPE for every {pj }j∈[m] in the ǫ-net. In the second step, we take some fresh samples from D and use
them to estimate the revenue for each of the ASPEs we created in the first step, then pick the one that has the
highest empirical revenue. It is not hard to argue that with a polynomial number of samples the mechanism
we pick has high revenue with probability almost 1. Combining our algorithm with Theorem 5, we obtain the
following theorem.
Theorem 9. When all bidders’ valuations are XOS over independent items and
• the random variable Vi (tij ) is supported on [0, H] for each bidder i and item j, we can learn an RSPM
and an ASPE such that with probability at least 1 − δ the better
at
of the two mechanisms has revenue
mn 2
m+n
1
OPT
least c1 − ξ · H for some absolute constant c1 > 1 given O ( ξ ) · (m · log ξ + log δ ) samples
from D;
• the random variable Vi (tij ) is regular for each bidder i and item j, we can learn an RSPM and an ASPE
such that with probability at least 1 − δ the better of the two mechanisms has revenue at least OPT
c2 for
1
2
2
2
some absolute constant c2 > 1 given O max{m, n} m n m log(m + n) + log δ samples from D.
The bounded case is proved as Theorem 15 in Appendix D.1. The regular case is proved as Theorem 16 in
Appendix D.1.
7 Symmetric Bidders
In this section, we consider symmetric bidders (Di = Di′ for all i and i′ ∈ [n]) with XOS and subadditive
valuations. For XOS valuations, our goal is to improve our algorithms from Section 6 to be computationally efficient under bidder symmetry. For subadditive valuations, our goal is to establish the learnability of
approximately optimal mechanisms whose revenue improves as the number of bidders becomes comparable
to the number of items. We only describe the results here and postpone the formal statements and proofs to
Appendix E.
• XOS valuations: we can learn in polynomial time an approximately optimal mechanism with a polynomial number of samples when the valuations are XOS over independent items. Our algorithm essentially
16
estimates all the parameters needed to run the RSPM and ASPE used in [13]. In general, it is not clear
how to estimate these parameters efficiently. But when the bidders are symmetric, one only needs to
consider “symmetric parameters” which greatly simplifies the search space and allows us to estimate all
the parameters in polynomial time. See Appendix E.2 for details.
• subadditive valuations:
when
the valuations are subadditive over independent items, the optimal revn
enue is at most O max{m,n} times larger than the highest revenue obtainable by an RSPM. In other
words, if the number of items is within a constant times the number of bidders, an RSPM suffices to
extract a constant fraction of the optimal revenue. Applying our results for unit-demand bidders in Section 5.1, we can learn a nearly-optimal RSPM, which is also a good approximation to OPT. In fact, when
the distribution for random variable Vi (tij ) is regular for every bidder i and item j, we can design a priorindependent mechanism that achieves a constant fraction of the optimal revenue. See Appendix E.3 for
details.
Appendix
A
Our Mechanisms
Here are the detailed description of the two major mechanisms we use: Sequential Posted Price Mechanism
(SPM) and Anonymous Sequential Posted Price with Entry Fee Mechanism (ASPE). We also use the Rationed
Sequential Posted Price Mechanism (RSPM) when bidders are not unit-demand. RSPM is almost identical to
SPM except that there is an extra constraint saying that no bidder can purchase more than one item.
Algorithm 2 Sequential Posted Price Mechanism (SPM)
Require: Pij is the price for bidder i to purchase item j.
1: S ← [m]
2: for i ∈ [n] do
3:
Show bidder i the set of available items S.
P
P
4:
i purchases her favorite bundle Si∗ ∈ maxS ′ ⊆S vi (ti , S ′ ) − j∈S ′ Pij and pays j∈S ∗ Pij .
i
5:
S ← S\Si∗ .
6: end for
Algorithm 3 Anonymous Sequential Posted Price with Entry Fee Mechanism (ASPE)
Require: A collection of prices {pj }j∈[m] and a collection of entry fee functions {δi (·)}i∈[n] where δi :
2[m] 7→ R is bidder i’s entry fee function.
1: S ← [m]
2: for i ∈ [n] do
3:
Show bidder i the set of available items S and set the entry fee for bidder i to be δi (S).
4:
if Bidder i pays the entry fee δi (S) then
P
5:
i receives her favorite bundle Si∗ and pays j∈S ∗ pj .
i
6:
S ← S\Si∗ .
7:
else
8:
i gets nothing and pays 0.
9:
end if
10: end for
17
B Missing Details from Section 5.1
B.1
Unit-demand Valuations: direct access to approximate distributions
We first consider the model where we only have access to an approximate distribution D̂. The following
definition is crucial for proving our result.
Definition 6. For any single dimensional distribution D with cdf F , we define its revenue curve RD : [0, 1] 7→
R≥0 as
RD (q) = max x · q · F −1 (1 − q) + (1 − x) · q̄ · F −1 (1 − q̄)
¯
¯
s.t. x · q + (1 − x) · q̄ = q
¯
x, q, q̄ ∈ [0, 1]
¯
where F −1 (1 − p) = sup{x ∈ R : Prv∼D [v ≥ x] ≥ p}.
Lemma 6 (Folklore). Let ϕij (·) and ϕ̂ij (·) be the ironed virtual value function for distribution Dij and D̂ij
RH
RH
respectively, then for any q ∈ [0, 1], RDij (q) = F −1 (1−q) ϕ(x)dF (x) and RD̂ij (q) = F̂ −1 (1−q) ϕ̂(x)dF (x).
ij
ij
Since the ironed virtual value function is monotonically non-decreasing, RDij (·) and RD̂ij (·) are concave
functions.
We provide an upper bound of the optimal revenue using RDij in the next Lemma. To do that, we first
need the definition of the Single-Dimensional Copies Setting.
Single-Dimensional Copies Setting: In the analysis for unit-demand bidders in [15, 11], the optimal revenue
is upper bounded by the optimal revenue in the single-dimensional copies setting defined in [15]. We use the
same technique. We construct nm agents, where agent (i, j) has value Vi (tij ) of being served with tij ∼ Dij ,
and we are only allow to use matchings, that is, for each i at most one agent (i, k) is served and for each j
at most one agent (k, j) is served8 . Notice that this is a single-dimensional setting, as each agent’s type is
specified by a single number. Let OPT C OPIES -UD be the optimal BIC revenue in this copies setting.
Lemma
there exists a collection of non-negative numbers {qij }i∈[n],j∈[m] satisP 7. For unit-demand bidders,P
fying i qij ≤ 1 for all j ∈ [m] and j qij ≤ 1 for all i ∈ [n], such that the optimal revenue
OPT ≤ 4 ·
X
RDij (qij ).
i,j
Proof. As shown in [11], OPT ≤ 4OPT C OPIES -UD. Let qij be the ex-ante probability that agent (i, j) is
served in the optimal mechanism for the copies setting. Chawla et al. [15] showed that OPT C OPIES -UD ≤
P
i,j RDij (qij ). Our statement follows from the two inequalities above.
Next, we consider a convex program (Figure 1) and argue that the value of the optimal solution of this
program is at least 18 of the optimal revenue.
Lemma 8. The optimal solution of convex program in Figure 1 is at least
8
OPT
8 .
This is exactly the copies setting used in [15], if every bidder i is unit-demand and has value Vi (tij ) with type ti . Notice that this
unit-demand multi-dimensional setting is equivalent as adding an extra constraint, each buyer can purchase at most one item, to the
original setting with subadditive bidders.
18
max
X
RDij (qij )
i,j
s.t.
X
qij ≤
1
2
for all j ∈ [m]
qij ≤
1
2
for all i ∈ [n]
i
X
j
qij ≥ 0
for all i ∈ [n] and j ∈ [m]
Figure 1: A Convex Program for Unit-demand Bidders with Exact Distributions.
n q′ o
ij
′ } be the collection of nonnegative numbers in Lemma 7. Clearly,
is a set of feasible
Proof. Let {qij
2
q′
′ )
′ )
R
RDij (qij
R
(q
Dij (0)
Dij ij
solution for the convex program. Since RDij (·) is concave, RDij 2ij ≥
+
=
.
2
2
2
Therefore,
′
X
qij
OPT
1 X
′
RDij
RDij (qij
)≥
.
≥ ·
2
2
8
i,j
i,j
If we know all Fij exactly, we can solve the convex program (Figure 1) and use the optimal solution to
construct an SPM via an approach provided in [15, 11]. The constructed sequential posted mechanism has
revenue at least 41 of the optimal value of the convex program, which is at least OPT
32 . Next, we show that
with only access to F̂ij , we can essentially carry out the same approach. Consider a different convex program
(Figure 2).
max
X
RD̂ij (qij )
i,j
s.t.
X
qij ≤
1
+n·ǫ
2
for all j ∈ [m]
qij ≤
1
+m·ǫ
2
for all i ∈ [n]
i
X
j
qij ≥ 0
for all i ∈ [n] and j ∈ [m]
Figure 2: A Convex Program for Unit-demand Bidders with Approximate Distributions.
Not that if the support size for all D̂ij is upper bounded by some finite number s, the convex program
above can be rewritten as a linear program with size poly(n, m, s). In the following Lemma, we prove that the
optimal values of the two convex programs above are close.
∗}
Lemma 9. Let {qij
i∈[n],j∈[m] and {q̂ij }i∈[n],j∈[m] be the optimal solution of the convex program in Figure 1
and 2 respectively.
X
X
∗
RD̂ij (q̂ij ) ≥
RDij (qij
) − ǫ · mnH.
i,j
i,j
∗ , q̄ ∗ and x ∈ [0, 1] be the numbers
Proof. We first fix some notations. For any bidder i and item j, let qij
ij
ij
¯
−1
−1
∗
∗
∗
∗
∗
∗ +(1−x )· q̄ ∗ = q ∗ .
satisfy that xij · qij ·Fij (1− qij )+(1−xij )· q̄ij ·Fij (1− q̄ij ) = RDij (qij ) and xij · qij
ij
ij
ij
¯
¯
¯
19
∗ ), p̄ = F −1 (1 − q̄ ∗ ), and q ′ = x · 1 − F̂ (p ) + (1 − x ) · 1 − F̂ (p̄ ) .
Let pij = Fij −1 (1 − qij
ij
ij
ij
ij ij
ij
ij ij
ij
ij
¯
¯
¯
By the definition of RD̂ij (·),
′
RD̂ij (qij
) ≥ xij · 1 − F̂ij (pij ) · pij + (1 − xij ) · 1 − F̂ij (p̄ij ) · p̄ij
¯
¯
(4)
∗ − ǫ, 1 − q ∗ + ǫ] and F̂ (p̄ ) ∈ [1 − q̄ ∗ − ǫ, 1 − q̄ ∗ + ǫ]. Hence,
Since ||D̂ij − Dij ||K ≤ ǫ, F̂ij (pij ) ∈ [1 − qij
ij ij
ij
ij
ij
∗ ) − ǫ¯· H. Therefore, R
′ )≥R
∗ ) − ǫ · H.
¯
¯
(q
(q
the RHS of inequality (4) is greater than RDij (qij
Dij ij
D̂ij ij
′ }
Next, we argue that {qij
is
a
feasible
solution
for
the
convex
program
in Figure 2. Since 1 −
i∈[n],j∈[m]
P ′
P ∗
∗
∗
′
∗
F̂ij (pij ) ≤ qij + ǫ and 1 − F̂ij (p̄ij ) ≤ q̄ij + ǫ, qij ≤ qij + ǫ. Thus, i qij ≤ i qij + n · ǫ ≤ 21 + n · ǫ for all
P ′
¯
¯
≤ 21 + m P
j ∈ [m]. Similarly, we can proveP j qij
· ǫ for all i ∈ [n]. As
is the optimal solution
P {q̂ij }i∈[n],j∈[m]
′ )≥
∗ ) − ǫ · mnH.
for the second convex program, i,j RD̂ij (q̂ij ) ≥ i,j RD̂ij (qij
R
(q
i,j Dij ij
Finally, we show how to use the optimal solution of the convex program in Figure 2 to construct an
SPM that approximates the optimal revenue well. We first provide a general transformation that turns any
approximately feasible solution of convex program in Figure 1 to an SPM mechanism.
Lemma 10. For any distribution D = ×i∈[n],j∈[m]Dij , given a collection of independent random variables
{pij }i∈[n],j∈[m] such that
X
i∈[n]
and
X
Pr
[tij ≥ pij ] ≤ 1 − η1 ,
Pr
[tij ≥ pij ] ≤ 1 − η2 , for all i ∈ [n],
pij ,tij ∼Dij
j∈[m]
pij ,tij ∼Dij
for all j ∈ [m]
we can construct in polynomial time a randomized SPM such that the revenue under D is at least
X
η1 η2 ·
Epij pij · Pr [tij ≥ pij ] .
i,j
tij ∼Dij
Proof. Consider a randomized SPM that sells item j to bidder i at price pij . Notice that bidder i purchases
exactly item j if all of the following three conditions hold: (i) for all bidders ℓ 6= i, tℓj is smaller than the
corresponding price pℓj , (ii) for all items k 6= j, tik is smaller than the corresponding price pik , and (iii) tij
is greater than the corresponding price pij . These
P three conditions are independent from each other. The first
condition holds with probability
at
least
1
−
ℓ6=i Prpℓj ,tℓj ∼Dℓj [tℓj ≥ pℓj ] ≥ η1 . The second condition holds
P
with probability at least 1 − k6=j Prpik ,tik ∼Dik [tik ≥ pik ] ≥ η2 . When the first two conditions hold, bidder i
purchases item j whenever she can afford it. Her expected payment is E
pij pij · Prtij ∼Dij [tij ≥pij ] . Hence,
the expected revenue for selling item
i is at least η1 η2 · Epij pij · Prtij ∼Dij [tij ≥ pij ] and the total
P j to bidder
expected revenue is at least η1 η2 · i,j Epij pij · Prtij ∼Dij [tij ≥ pij ] .
Lemma 11. Given any feasible solution {qij }i∈[n],j∈[m] of the convex program in Figure 2, we can con
1
−
(n
+
m)
·
ǫ
·
struct
a
(randomized)
SPM
in
polynomial
time
such
that
its
revenue
under
D
is
at
least
4
P
i,j RD̂ij (qij ) − ǫ · nmH .
Proof. We first fix some notations. For any bidder i and item j, let qij , q̄ij and xij ∈ [0, 1] be the numbers
¯
satisfying xij · qij · F̂ij−1 (1 − qij ) + (1 − xij ) · q̄ij · F̂ij−1 (1 − q̄ij ) = RD̂ij (qij ) and xij · qij + (1 − xij ) · q̄ij = qij .
¯
¯
¯
20
We use pij to denote a random variable that is pij = F̂ij−1 (1 − qij ) with probability xij and p̄ij = F̂ij−1 (1 − q̄ij )
¯
¯
with probability 1 − xij .
Next, we construct a randomized SPM based on {pij }i∈[n],j∈[m] according to Lemma 10. Note that
X
i∈[n]
Pr
pij ,tij ∼Dij
[tij ≥ pij ] ≤
X
i∈[n]
Pr
pij ,tij ∼D̂ij
!
=
!
=
[tij ≥ pij ] + ǫ
X
qij + nǫ ≤
X
qij + mǫ ≤
i∈[n]
1
+ 2nǫ
2
for all item j, and
X
j∈[m]
Pr
pij ,tij ∼Dij
[tij ≥ pij ] ≤
X
j∈[m]
Pr
pij ,tij ∼D̂ij
[tij ≥ pij ] + ǫ
i∈[m]
1
+ 2mǫ
2
for all bidder i. Hence, we can construct in polynomial time a randomized SPM with revenue at least
X
1
1
− 2nǫ
− 2mǫ ·
Epij pij · Pr [tij ≥ pij ]
tij ∼Dij
2
2
i,j
!#
"
X
1
Pr [tij ≥ pij ] − ǫ
Epij pij ·
− (n + m)ǫ
≥
4
t
∼D̂ij
ij
i,j
X
1
≥
RD̂ij (qij ) − ǫ · nmH
− (n + m)ǫ
4
i,j
≤ ǫ, and the second inequality is because pij is upper bounded
The first inequality is because Dij − D̂ij
h
i K
by H and Epij pij · Prtij ∼D̂ij [tij ≥ pij ] = RD̂ij (qij ) by the definition of pij .
Theorem 10. For unit-demand bidders, given distributions D̂ij where D̂ij − Dij
K
≤ ǫ for all i ∈ [n] and
j ∈ [m], there is a polynomial
OPT time algorithm
that constructs a randomized SPM whose revenue under D is at
1
least 4 − (n + m) · ǫ · 8 − 2ǫ · mnH .
Proof. Our algorithm first computes the optimal solution {q̂ij }i∈[n],j∈[m] for the convex program in Figure 2,
then constructs a randomized SPM based on {q̂ij }i∈[n],j∈[m] using Lemma 11. It is not hard to see that our
algorithm runs in polynomial time. By chaining the inequalities
OPT in Lemma 8, 9 and 11, we can argue that the
1
revenue of our mechanism is at least 4 − (n + m) · ǫ · 8 − 2ǫ · mnH .
B.2
Unit-demand Valuations: sample access to bounded distributions
When the distributions Dij are all bounded, the following theorem provides the sample complexity.
Theorem 11. [35] When Dij is supported on [0, H] for all bidder i and item j, the sample complexity for(ǫ, δ)2
uniformly learning the revenue of SPMs for unit-demand bidders is O 1ǫ
m2 n log n log 1ǫ + log 1δ . That
is, with probability 1 − δ, the empirical revenue based on the samples for any SPM is within ǫ · H of its true
expected revenue. Moreover, with the same number of samples, there is a polynomial time algorithm that
learns an SPM whose revenue is at least OPT
144 − ǫH with probability 1 − δ.
21
B.3
Unit-demand Valuations: sample access to regular distributions
In this section, we show there exists a polynomial time algorithm that learns an SPM whose revenue is at least a
constant fraction of the optimal revenue with polynomial in n and m samples. Note that unlike in the previous
two models, the error of our learning algorithm is only multiplicative when the distributions are regular. First,
we present a Lemma regarding the revenue curve function for regular distributions.
Lemma 12. [7] For any regular distribution F , let RF (·) be the corresponding revenue curve. For any
0 < q ′ ≤ q ≤ p < 1,
(1 − p) · RF (q ′ ) ≤ RF (q).
Throughout this section, we use Z to denote max{m, n} and C to be a constant that will be specified later.
1
does not affect the objective
Using Lemma 12, we show in the next Lemma that restricting qij to be at least CZ
value of the convex program in Figure 1 by too much.
′
∗}
Lemma 13. Suppose {qij
i∈[n],j∈[m] is the optimal solution of the convex program in Figure 1. Let qij =
P
OPT
P
1
1
1
∗
′
∗
max{ CZ , qij }, then i,j RDij (qij ) ≥ 1 − CZ · i,j RDij (qij ) ≥ 1 − CZ · 8 .
P
∗ ) ≥ OPT . So to prove the statement, it suffices to argue that
Proof. According to Lemma 8, i,j RDij (qij
8
′ ) ≥ 1− 1
∗ ). If q ∗ = q ′ , this inequality clearly holds. If q ∗ 6= q ′ ,
for any i and j, RDij (qij
·
R
(q
Dij ij
ij
ij
ij
ij
CZ
′ )≥
∗ ≤ q ′ = 1 . Since F is regular, we can apply Lemma 12 to q ′ and q ∗ and obtain inequality R
(q
qij
ij
D
ij
ij
ij
ij
ij CZ
1
∗ ).
· RDij (qij
1 − CZ
′
Dij
Using Lemma 13, we argue how to compute in polynomial time an approximately optimal SPM. Suppose
is the distribution that we obtain after truncating Dij at a threshold Hij 9 , and we have direct access to a
′ such that D̂ ′ − D ′
discrete distribution D̂ij
ij
ij
K
≤ ǫ for all i and j. We show in the following Lemma that
′ }
the optimal solution of a convex program similar to the one in Figure 2 but for {D̂ij
i∈[n],j∈[m] can guide us to
design an approximately optimal SPM under D in polynomial time. As we have sample access to D, we will
′ }
argue later that a polynomial number of samples suffices to generate {D̂ij
i∈[n],j∈[m] .
1
1
, 1 − 3C·Z
]
Lemma 14. Let {Hij }i∈[n],j∈[m] be a collection of positive numbers satisfying Fij (Hij ) ∈ [1 − C·Z
′
for all i ∈ [n] and j ∈ [m]. Let Dij be the distribution of the random variable min{tij , Hij } where tij ∼ Dij ,
′ − D′
D̂ij
ij
′ be a discrete distribution such that
and D̂ij
K
≤ ǫ for all i ∈ [n] and j ∈ [m]. Suppose
′ , then given direct access to D̂ ′ , we can
s is an upper bound of the support size for any distribution D̂ij
ij
1
1
compute in time polynomial in n, m and s a randomized
SPM
that
achieves
revenue
at
least
−
2
C − 2nǫ ·
1
1
1
1 − CZ
· OPT
2 − C − 2mǫ ·
8 − 2ǫ · nmH under D, where H = maxi,j Hij .
Proof. Consider the following convex program:
max
X
RD̂′ (qij )
i,j
s.t.
X
qij ≤
1
1
+ +n·ǫ
2 C
for all j ∈ [m]
qij ≤
1
1
+ +m·ǫ
2 C
for all i ∈ [n]
i
X
j
qij ≥ 0
9
ij
for all i ∈ [n] and j ∈ [m]
′
Let tij ∼ Dij , then min{tij , Hij } is the corresponding truncated random variable drawn from Dij
.
22
∗}
′ = max{ 1 , q ∗ }.
Let {qij
Figure 1 and qij
i∈[n],j∈[m] be the optimal solution of the convex hprogram in
C·Z ij
i
−1
′
′
′
For every i and j, let pij = Fij (1 − qij ) and q̃ij = Prtij ∼D̂′ tij ≥ pij . By the definition of Hij , p′ij ≤ Hij ,
ij
so
!
′
′
Pr ′ tij ≥ p′ij − ǫ ≥ p′ij qij
p′ij q̃ij ≥ p′ij
− ǫ · Hij = RDij (qij
(5)
) − ǫ · Hij .
tij ∼Dij
′ equals to R
′
p′ij qij
Dij (qij ) because Dij is a regular distribution.
Next, we argue that {q̃ij }i∈[n],j∈[m] is a feasible solution of the convex program above. Observe that
X
X
X
1
1
1
′
∗
q̃ij ≤
qij
+ nǫ ≤
qij
+
+ nǫ ≤ + + nǫ
CZ
2 C
i
i
i
for all item j ∈ [m] and
X
q̃ij ≤
j
X
′
qij
+ nǫ ≤
j
X
∗
qij
+
j
1
CZ
1
1
+ + mǫ
2 C
+ mǫ ≤
for all bidder i ∈ [n].
d be the optimal solution of the convex program above. As {q̃ij }i∈[n],j∈[m] is a feasible solution,
Let OPT
X
X
X
1
OPT
′
′
d
pij q̃ij ≥
RDij (qij ) − ǫ · nmH ≥ 1 −
OPT ≥
RD̂′ (q̃ij ) ≥
− ǫ · nmH.
·
ij
CZ
8
i,j
i,j
i,j
The second last inequality is due to inequality (5) and the last inequality is due to Lemma 13.
So far, we have argued that the optimal solution of our convex program has value close to the OPT. We will
show in the second part of the proof that using the optimal solution of our convex program, we can construct
d Let q̂ij be the optimal solution of the convex program
an SPM whose revenue under D is close to OPT.
above and p̂ij be the corresponding random price, that is, Prp̂ij ,tij ∼D̂′ [tij ≥ p̂ij ] = q̂ij and RD̂′ (q̂ij ) =
ij
ij
i
h
Ep̂ij p̂ij · Prtij ∼D̂′ [tij ≥ p̂ij ] . As p̂ij ≤ Hij ,
ij
Pr
p̂ij ,tij ∼Dij
[tij ≥ p̂ij ] =
Pr
′
p̂ij ,tij ∼Dij
[tij ≥ p̂ij ] ∈ [q̂ij − ǫ, q̂ij + ǫ].
Therefore, for all item j
X
i
Pr
p̂ij ,tij ∼Dij
and for all bidder i
X
Pr
j
p̂ij ,tij ∼Dij
[tij ≥ p̂ij ] ≤
[tij ≥ p̂ij ] ≤
X
i
X
j
Pr
′
p̂ij ,tij ∼D̂ij
Pr
′
p̂ij ,tij ∼D̂ij
[tij ≥ p̂ij ] + nǫ =
X
q̂ij + nǫ ≤
1
1
+ + 2nǫ
2 C
q̂ij + mǫ ≤
1
1
+ + 2mǫ.
2 C
i
[tij ≥ p̂ij ] + mǫ =
X
j
According to Lemma 10, we can construct a randomized SPM with {p̂ij }i∈[n],j∈[m] whose revenue is at least
1
P
1
1
1
i,j Ep̂ij p̂ij · Prtij ∼Dij [tij ≥ p̂ij ] under D. Clearly,
2 − C − 2nǫ · 2 − C − 2mǫ ·
Ep̂ij p̂ij ·
"
Pr [tij ≥ p̂ij ] ≥ Ep̂ij p̂ij ·
tij ∼Dij
Pr [tij ≥ p̂ij ] − ǫ
′
tij ∼D̂ij
23
!#
≥ RD̂′ (q̂ij ) − ǫ · Hij .
ij
Therefore, the revenue of the constructed randomized SPM under D is at least
1
1
1
1
d − ǫ · nmH
− − 2nǫ ·
− − 2mǫ · OPT
2 C
2 C
1
1
1
1
1
OPT
≥
− − 2nǫ ·
− − 2mǫ ·
1−
− 2ǫ · nmH .
·
2 C
2 C
CZ
8
It is not hard to see that both {q̂ij }i∈[n],j∈[m] and {p̂ij }i∈[n],j∈[m] can be computed in time polynomial in n, m
and s.
When ǫ is small enough, the additive error in Lemma 14 can be converted into a multiplicative error. Next,
′ }
we argue that with a polynomial number of samples, we can learn {Hij }i∈[n],j∈[m] and {D̂ij
i∈[n],j∈[m] with
enough accuracy.
Theorem 12. If for all bidder i and item j, Dij is a regular distribution, we can learn in polynomial
time
nm
OPT
2
2
2
with probability 1 − δ a randomized SPM whose revenue is at least 33 with O Z m n · log δ (Z =
max{m, n}) samples.
samples from each Dij , we can find an Hij such that Fij (Hij )
Proof. First, if we take O C 2 · Z 2 · log nm
δ
1
1
δ
lies in[1 − CZ
, 1 − 3CZ
] with probability 1 − 2nm
. By the union bound, the probability that all Hij satisfy
1
1
δ
, 1 − 3C·Z
] for all i and j.
the requirement is at least 1 − 2 . From now on, we assume Fij (Hij ) ∈ [1 − C·Z
1
Observe that OPT ≥ maxi,j Hij · 3C·Z , as the expected revenue for selling item j to bidder i at price Hij
Hij
1
. Therefore, there exists sufficiently large constant d and C, if ǫ = d·Znm
the randomized
is at least 3C·Z
OPT
SPM learned in Lemma 14 has revenue at least 33
. According to′ the Dvoretzky-Kiefer-Wolfowitz (DKW)
samples from Dij (we can take samples from Dij then cap
inequality [25], if we take O d2 Z 2 n2 m2 · log nm
δ
′ be the uniform distribution over the samples, D ′ − D̂ ′
the samples at Hij ) and let D̂ij
ij
ij
δ
′ − D̂ ′
. By the union bound, Dij
probability 1− 2nm
ij
≤
K
≤
1
d·Znm
with
1
d·Znm
for all i ∈ [n] and j ∈ [m] with probability
′ we learned from O Z 2 n2 m2 · log nm
at least 1 − δ/2. Finally, by another union bound, the Hij and D̂ij
δ
samples satisfy Fij (Hij ) ∈ [1 −
1
C·Z , 1
−
1
3C·Z ]
K
′ − D̂ ′
and Dij
ij
K
≤
1
d·Znm
for all i and j with probability
at least 1 − δ. In other words, we can learn a randomized SPM whose revenue is at least OPT
33 with probability
′ is at most
at least 1 − δ using O Z 2 n2 m2 · log nm
samples.
Furthermore,
the
support
size
of
any D̂ij
δ
samples, so our learning algorithm runs in time polynomial in n and m.
O Z 2 n2 m2 · log nm
δ
C
Missing Details from Section 5.2
C.1 Additive Valuations: sample access to bounded distributions
As shown by Goldner and Karlin [28], one sample suffices to design a mechanism that approximates BR EV .
The idea is toP
use the VCG with entry fee mechanism but replace the entry fee ei (b−i , Di ) for bidder i with
ei (b−i , si ) = j∈[m] (sij − maxk6=i bkj )+ , where si is a sample drawn from Di . It is easy to argue that for any
b−i , over the randomness of the sample si and bidder i’s real type ti , the event that eP
i (b−i , si ) ≥ ei (b−i , Di )
and bidder i accepts the entry fee ei (b−i , si ) happens with probability at least 81 . As 12 · i∈[n] Et [ei (t−i , Di )] =
BR EV , the
P expected revenue (over the randomness of the types and the samples) from their mechanism is at
least 81 · i∈[n] Et [ei (t−i , Di )] = BR4EV . Next, we show how to learn a mechanism that approximates SR EV .
24
Lemma 15. When Dij is supported on [0, H] for all bidder i and item j, the sample complexity for
(ǫ, δ)
1 2
1
1
2
uniformly learning the revenue of SPMs for additive bidders is O ǫ
m n log n log ǫ + log δ . More-
over, we can learn in polynomial time an SPM whose revenue is at least
given the same number of samples.
SR EV
4
−
3ǫ
2
· H with probability 1 − δ
Proof. The first half of the Lemma was proved by Morgenstern and Roughgarden [35]. We show how to
prove the second half of the claim. Let OPTj be the optimal revenue for selling item j. By the prophet
inequality [41], there exists an SPM for selling item j with a collection of prices {pij }i∈[n] that achieves
revenue at least OPTj /2. As the bidders are additive, if we run the SPMs for selling each item simultaneously,
the expected revenue is exactly the sum of the revenue of the SPM mechanisms for auctioning a single item.
Note that the simultaneous SPM is indeed a SPM for selling all items. Hence, there exists an SPM that
achieves revenue
sample complexity for (ǫ, δ)-uniformly learning the revenue of
at least OPT/2. Since the
1
1
m 2
n log n log ǫ + log δ , the empirical revenue induced by the samples is within ǫ · H of
SPMs is O
ǫ
the true expected revenue with probability 1 − δ for any SPM.
We use ERopt to denote the optimal empirical revenue obtained by any SPM. If we apply the prophet
inequality to the empirical distribution, we can construct an SPM whose empirical revenue ER is at least
ERopt /2. Notice that ERopt is at most ǫ · H less than the optimal true expected revenue obtained by any
SPM, which is at least OPT/2. Combining the two inequalities above, we have ER ≥ OPT/4 − ǫ/2 · H with
probability 1 − δ. Also, the true expected revenue of our SPM is at least ER − ǫ · H, so our SPM achieves
3ǫ
expected revenue at least OPT
4 − 2 · H with probability 1 − δ.
Now we are ready to prove our Theorem for additive bidders when their valuations are bounded.
Theorem 13. When the bidders have additive valuations and Dij is supported on [0, H] for all bidder i and
item j, we can learn in polynomial time a mechanism whose expected revenue is at least OPT
32 − ǫ · H with
probability 1 − δ given
1
1
m 2
· n log n log + log
O
ǫ
ǫ
δ
samples from D.
SR EV
ǫ
Proof. According to Lemma
we can learn a mechanism
whose revenue is at least 4 − 24 · H with
15,
2
probability 1 − δ given O mǫ · n log n log 1ǫ + log 1δ samples. As we explained in the beginning of this
section, with one sample from the distribution we can construct a randomized mechanism whose expected
revenue is at least BR4EV . Therefore, the better of our two mechanisms has expected revenue at least OPT
32 − ǫ · H
with probability 1 − δ.
C.2 Additive Valuations: direct access to approximate distributions
In this section, we discuss how to learn an approximately optimal mechanism for additive bidders when we are
given direct access to approximate value distributions. Again, we first show how to learn a mechanism whose
revenue approximates SR EV then we provide another mechanism whose revenue approximates BR EV .
Lemma 16. For additive bidders, given distributions D̂ij where
D̂ij − Dij
K
≤ ǫ for all i ∈ [n] and
j ∈ [m], there is a polynomial time algorithm
that constructs a randomized SPM whose revenue under D is at
1
SR EV
least 4 − ǫ · n ·
8 − 2ǫ · mnH .
Proof. Let OPTj be the optimal revenue for selling item j. As the bidders are additive, if we can construct
a randomized SPM Mj for every item j such that its expected revenue under D is at least 14 − ǫ · n ·
25
OPTj
8
− 2ǫ · nH , running these m randomized SPMs in parallel generates expected revenue at least
X 1
OPTj
1
SR EV
−ǫ·n ·
− 2ǫ · nH =
−ǫ·n ·
− 2ǫ · mnH
4
8
4
8
j∈[m]
under D. Due to Theorem 10, we can construct in polynomial time such a randomized SPM Mj for each item
j based on ×i∈[n] D̂ij .
Next, we show how to choose the entry fee based on D̂ = ×i,j D̂ij , so that the VCG with entry fee
mechanism has revenue that approximates BR EV under the true distribution D. More specifically, we use
the median of i’s utility under D̂i = ×j∈[m] D̂ij as bidder i’s entry fee. We prove the result in two steps.
We first show that if we can use an entry fee function such that every bidder i accepts her entry fee with
probability between [1/2 − η, 1/2] for any possible bid profiles b−i of the other bidders, the expected revenue
is at least (1/2 − η) · BR EV . Second, we show how to compute in polynomial time such entry fee functions
with η = O(mǫ) based on D̂.
Lemma 17. Suppose for every bidder i, di (·) : T−i 7→ R is a randomized entry fee function such that for any
bid profile b−i ∈ T−i of the other bidders
+
X
1
1
− η,
≥ di (b−i ) ∈
tij − max bkj
Pr
k6=i
ti ∼Di
2
2
j∈[m]
with probability at least 1 − δ. Then if we use di (·) as the entry fee function in the VCG with entry fee
mechanism, the expected revenue is at least (1 − δ − 2η) · BR EV .
hP
i
+
Proof. When Prti ∼Di
(t
−
max
b
)
≥
d
(b
)
∈ 21 − η, 21 , di (b−i ) is no less than the origij
i
−i
k6
=
i
kj
j∈[m]
inal entry fee ei (b−i , Di ) for any bid profile
−i of the other bidders. The expected revenue under the new entry
bP
fee functions is at least ((1 − δ) · 12 − η · i∈[n] Eb−i ∼D−i [ei (b−i , Di )] ≥ (1 − δ − 2η) · BR EV .
Lemma 18. For any bidder i and any bid profile b−i from the other bidders, let Fi,b−i and F̂i,b−i be the distriP
butions for the random variable j∈[m] (tij − maxk6=i bkj )+ when ti is drawn from Di and D̂i respectively.
If Dij − D̂ij
K
≤ ǫ for all bidder i and item j, Fi,b−i − F̂i,b−i
K
≤ 2mǫ for all i and b−i . Moreover,
when mǫ ≤ 1/16, we can compute a randomized mechanism whose expected revenue is at least BR5EV .
o
n
P
+
(t
−
max
b
)
≥
x
. It is easy
Proof. For any real number x, consider event Ei,b−i ,x = ti
k6=i kj
j∈[m] ij
to see that Ei,b−i ,x is single-intersecting for any any i, b−i and x. According to Lemma 3,
Pr
ti ∼Di
Ei,b−i ,x − Pr Ei,b−i ,x ≤ 2mǫ
ti ∼D̂i
for any i, b−i and x. Hence, Fi,b−i − F̂i,b−i
K
≤ 2mǫ.
Next, we argue how to construct a randomized entry fee di (b−i ) in polynomial time with only sample access of F̂i,b−i . Suppose we take k samples from F̂i,b−i and sort them in descending order s1 ≥ s2 ≥ · · · ≥ sk .
Let the entry fee di (b−i ) to be s⌈ 5k ⌉ . By the Chernoff bound, with probability at least 1 − exp(−k/128) (over
16
hP
i
+
(t
−
max
b
)
≥
d
(b
)
= Prti ∼D̂i Ei,b−i ,di (b−i )
the randomness of the samples) Prti ∼D̂i
i −i
k6=i kj
j∈[m] ij
lies in 14 , 38 . Since Prti ∼Di Ei,b−i ,di (b−i ) = Prti ∼D̂i Ei,b−i ,di (b−i ) ± 2mǫ,
Pr
ti ∼Di
1 1
Ei,b−i ,di (b−i ) ∈ [ , ],
8 2
26
if mǫ ≤ 1/16. According
to Lemma 17, the expected revenue under our entry fee di (b−i ) is at least
1
BR EV
if we choose k to be larger than some absolute constant. Clearly,
4 − exp(−k/128) · BR EV ≥
5
the procedure above can be completed in polynomial time with access to D̂.
Combining Lemma 16 and 18, we are ready to prove our main result of this section.
Theorem 14. If all bidders have additive valuations, given distributions D̂ij where D̂ij − Dij
K
≤ ǫ for
all i ∈ [n] and j ∈ [m], there is a polynomial time algorithm that constructs a mechanism whose expected
1
revenue under D is at least OPT
266 − 96ǫ · mnH when ǫ ≤ 16 max{m,n} .
Proof. Since ǫ ≤
3
16
SR EV
8
1
,
16 max{m,n}
we can learn in polynomial time a randomized SPM whose revenue is at
least
·
− 2ǫ · mnH and a VCG with entry fee mechanism whose revenue is at least BR EV /5. As
OPT ≤ 6 · SR EV + 2BR EV (Theorem 6), the better of the two mechanisms we can learn in polynomial time
has revenue at least OPT
266 − 96ǫ · mnH.
D
Missing Details from Section 6
Proof of Lemma 5: We only sketch the proof here. Let P OST R EV denote the highest revenue obtainable
by any RSPM. In [13], Cai and Zhao constructed an upper bound of the optimal revenue using duality and
separated the upper bound into three components: S INGLE , TAIL and C ORE. Both S INGLE and TAIL are
within constant times the P OST R EV , and the ASPE(p∗ , δ∗ ) is used to bound the C ORE. It turns out one
can use essentially the same proof as in [13] to prove that the mechanism ASPE(p′, δ′ ) has revenue at least
a1 (µ) · C ORE − a2 (µ) · P OST R EV − a3 (µ) · (n + m) · ǫ where a1 (µ), a2 (µ) and a3 (µ) are functions that
map µ to positive numbers. In other words, we can replace ASPE(p∗ , δ∗ ) with ASPE(p′, δ′ ) and still obtain a
constant factor approximation. ✷
D.1 Missing Proofs from Section 6.1
We formalize the first step of our algorithm in the following lemma.
Lemma 19. For any B > 0, ǫ > 0, η ∈ [0, 1] and µ ∈ [0,
1
4 ],
suppose we take K = O
log
1
+log n+m log B
η
ǫ
µ2
samples t(1) , · · · , t(K) from D. For any collection of prices {pj }j∈[m] in the B-bounded ǫ-net, define the
(p)
(1)
(K)
entry fee δi (S) of bidder i for setPS under {pj }j∈[m] to be the median of ui (ti , S), · · · , ui (ti , S), where
ui (ti , S) = maxS∗⊆S vi (ti , S ∗ )− j∈S ∗ pj . Then with probability 1−η, for any collection of prices {pj }j∈[m]
n
o
(p)
in the B-bounded ǫ-net, δi (·)
is a collection of µ-balanced entry fee functions.
i∈[n]
Proof. For any fixed {pj }j∈[m] , fixed bidder i and fixed set S, it is easy to argue that the probability for
(p)
Prti ∼Di [ui (ti , S) ≥ δi (S)] to be larger than 21 + µ or smaller than 12 − µ is at most 2exp(−2Kµ2 ) due to
the Chernoff bound. Next, we take a union bound over all {pj }j∈[m] in the ǫ-net, all bidders and all possible
(p)
subsets of [m], so the probability that for any collection of prices {pj }j∈[m] in the ǫ-net {δi (·)}i∈[n] is a
m m
collection of µ-balanced entry fee functions is at least 1 − 2 exp(−2Kµ2 ) · Bǫ
· 2 · n. If we take K to be
at least
log η1 +log n+m log
µ2
B
ǫ
, the success probability is at least 1 − η.
Next, we formalize the second step of our learning algorithm.
27
Lemma 20. For any B ≥ 2G, ǫ, ǫ′ > 0, η ∈ [0, 1] and µ ∈ [0, 14 ], suppose for every collection of
(p)
prices {pj }j∈[m] in the B-bounded ǫ-net, {δi (·)}i∈[n] is a collection of µ-balanced entry fee functions.
We use S to denote the set that contains ASPE(p, δ(p) ) for every p in the B-bounded ǫ-net. If we take
K = O
log
1
+m log B
η
ǫ
ǫ′2
′
samples t(1) , · · · , t(K) from D and let ASPE(p′ , δ(p ) ) be the mechanism that has
′
the highest revenue in S. Then with probability at least 1 − η, the better of ASPE(p′ , δ(p ) ) and the best RSPM
achieves revenue at least COPT
− C2 (µ) · (m + n) · ǫ − 2mnB · ǫ′ .
1 (µ)
Proof. For any {pj }j∈[m] in the ǫ-net, define R EV (p) to be the expected revenue of ASPE(p, δ(p) ) and Rd
EV (p)
(p)
d
be the average revenue of ASPE(p, δ ) among the K samples. First, we argue that R EV (p) is a random
variable that lies between [0, mnB]. The revenue from selling the items can be at most mB as there are only
m items and pj ≤ B for all j ∈ [m]. How about the entry fee? For any bidder i,
X
1
m
≤ .
Pr [vi (ti , [m]) ≥ mG] ≤
Pr [Vi (tij ) ≥ G] ≤
ti ∼Di
tij ∼Dij
5 max{m, n}
5
j∈[m]
The first inequality is because vi (ti , ·) is a subadditive function for every type ti ∈ Ti , so for vi (ti , [m]) to
be greater than mG, there must exist a item j such that Vi (tij ) ≥ G. The second inequality follows from the
definition of G in Theorem 8.
(p)
If there exists a set S ⊆ [m] such that δi (S) > mG, we have
h
i 1
1
(p)
Pr [vi (ti , [m]) ≥ mG] ≥ Pr vi (ti , [m]) ≥ δi (S) ≥ − µ ≥ .
ti ∼Di
ti ∼Di
2
4
(p)
Contradiction. Note that the second inequality is because δi (·) is µ-balanced. Hence, the entry fee is always
upper bounded by mG and Rd
EV (p) is at most mnG + mB ≤ mnB. Also, notice that the expectation of
Rd
EV (p) is exactly R EV (p). By the Chernoff bound,
i
h
′2
′
Pr R EV (p) − Rd
EV (p) ≤ mnB · ǫ ≥ 1 − 2 exp(−2K · ǫ )
for any fixed {pj }j∈[m] . By the union bound, the probability that for all {pj }j∈[m] in the ǫ-net
′
R EV (p) − Rd
EV (p) ≤ mnB · ǫ
m
is at least 1 − 2 exp(−2K · ǫ′2 ) · Bǫ , which is lower bounded by 1 − η due to our choice of K. When this
′
happens, the expected revenue of ASPE(p′, δ(p ) ) is at most 2mnB · ǫ′ less than the highest expected revenue
achievable by any of these mechanisms, because
′
′
′
′
d
R EV (p′ ) ≥ Rd
EV (p ) − mnB · ǫ ≥ R
EV (p) − mnB · ǫ ≥ R EV (p) − 2mnB · ǫ
for any p in the ǫ-net. Combining this inequality with Corollary 2 completes our proof.
Note that Lemma 19 and 20 hold for all distributions D. The reason we require D to be bounded or regular
is because without these restrictions, we do not know how to approximate the best RSPM. In the following
Theorem, we combine Lemma 19, 20 and Theorem 11 to obtain the sample complexity of our learning
algorithm for bounded distributions.
Theorem 15. When all bidders’ valuations are XOS over independent
and the random variable
Vi (tij )
items
2
mn
1
is supported on [0, H] for any bidder i and any item j, with O
· m · log m+n
samples
ξ
ξ + log δ
from D, we can learn an RSPM and an ASPE such that with probability at least 1 − δ the better of the two
mechanisms has revenue at least OPT
c − ξ · H for some absolute constant c > 1.
28
Proof. With O
2
1
2 n log n log 1 + log 1
m
samples, we can obtain an RSPM whose revenue is at
ξ
ξ
δ
1
least 24
of the revenue of the best RSPM minus 2ξ · H with probability 1 − δ/2 according to Theorem 11.
ξ
ξ·H
and ǫ′ = 12mn
Let µ be some fixed constant in [0, 41 ], B = 2H, ǫ = 6C2 (µ)(m+n)
. According to Lemma 19,
samples, we can construct an entry fee function for each price vector
given O log 1δ + log n + m log m+n
ξ
in the B-bounded ǫ-net, such that all these entry fee functions
with probabilityat least 1 − δ/4.
are µ-balanced
2
mn
1
According to Lemma 20, we can learn an ASPE with O
· m · log m+n
fresh samples
ξ
ξ + log δ
− 2ξ · H
from D, such that the better of the ASPE we learned and the best RSPM has revenue of at least COPT
1 (µ)
with probability 1 − δ/4. Combining the statements
we can learn with probability
1 − δ a mechanism
above,
2
mn
1
whose revenue is at least OPT
· m · log m+n
samples.
c − ξ · H with O
ξ
ξ + log δ
In the next Theorem, we combine Lemma 19, 20 and Theorem 12 to obtain the sample complexity of our
learning algorithm for regular distributions.
Theorem 16. When all bidders’ valuations are XOS over independent items and the random variable
Vi (tij )
is regular for each item j ∈ [m] and bidder i ∈ [n], with O Z 2 m2 n2 · m · log(m + n) + log 1δ (Z =
max{m, n}) samples from D, we can learn an RSPM and an ASPE such that with probability at least 1 − δ
the better of the two mechanisms has revenue at least OPT
c for some absolute constant c > 1.
Proof. According to Theorem 12, we can learn with probability
1 − δ/2 a randomized RSPM whose revenue
nm
1
2
2
2
with O Z m n · log δ samples. Next, we learn an ASPE with high
is at least 33 of the optimal RSPM
revenue. With O Z 2 · log nm
samples
from each Dij , we can estimate Wij such that
δ
1 1
,
Pr [Vi (tij ) ≥ Wij ] ∈
tij ∼Dij
6Z 5Z
δ
. By the union bound, the probability that all Wij satisfy the requirement is at least
with probability 1 − 4nm
δ
1 − 4 . So with probability at least 1 − 4δ , Wij ≥ Gij for all i ∈ [n] and j ∈ [m].
ξ·B
ξ
Let B = 2·maxi,j Wij , µ be some fixed constant in [0, 41 ], ǫ = C2 (µ)Z(m+n)
for some small
and ǫ′ = 2mnZ
1
constant ξ, which will be specified later. We know that given O log δ + log n + m log(m + n) samples, we
can construct µ-balanced entry fee functions for all price vectors in the B-bounded ǫ-net with probability
1 − δ/8 due to Lemma 19. According to Lemma 20, we can learn an ASPE with
1
2 2 2
O Z m n · m · log(m + n) + log
δ
fresh samples from D, such that the better of the ASPE we learned and the best RSPM has revenue of at least
2ξ·B
OPT
C1 (µ) − Z with probability 1 − δ/8. Note that there exists a bidder i and an item j such that Wij = B/2,
so OPT ≥
B
2
·
1
6Z
and for sufficiently small ξ,
OPT
C1 (µ)
−
2ξ·B
Z
≥
OPT
2C1 (µ) .
can learn with probability 1 − δ a mechanism
whose revenue is at least
1
2
2
2
O Z m n · m · log(m + n) + log δ samples.
29
Combining the statements above, we
OPT
c
for some absolute constant c with
E Learning Algorithms for Symmetric Bidders
E.1
An Upper Bound of the Optimal Revenue for Symmetric Bidders
In this section, we introduce an upper bound to OPT based on duality [13], which is crucial for us to prove
the approximation ratios of our learning algorithms. We first fix some notation. Let P OST R EV be the highest
revenue obtainable by any RSPM. As the bidders are symmetric, we drop the subscript i when there is no
confusion. In particular, we use V (tij ) to denote bidder i’s value for winning item j if her private information
for item j is tij , and v(ti , S) to denote bidder i’s value for set S when her type is ti . We use Dj to denote the
distribution of the private information about item j. Let σiS (t) be the interim probability for bidder i to receive
exactly set S ⊆ [m] when her type is t.
In [13], an upper bound of the optimal revenue is derived using duality theory. Their upper bound applies
to asymmetric bidders with valuations that are subadditive over independent items. When the bidders are
symmetric, we can simplify their upper bound. First, we need the definition of b-balanced thresholds.
Definition 7 (b-balanced Thresholds). For any constant b ∈ (0, 1), a collection of positive real numbers
b
].
{βj }j∈[m] is b-balanced if for all i ∈ [n] and j ∈ [m], Prtij ∼Dj [V (tij ) ≥ βj ] ∈ [ nb , n−1
Note that when bidders are asymmetric, b-balanced thresholds are not guaranteed to exist, as there may not
b
exist any βj that satisfies Prtij ∼Dj [V (tij ) ≥ βj ] ∈ [ nb , n−1
] for all bidder i simultaneously. Next, we define
the C OREη (β) which will be crucial for upper bounding the optimal revenue.10
Definition 8 (C ORE). Given any collection of thresholds {βj }j∈[n] and a nonnegative constant η ≤ 41 ,
P
• if j∈[m] Prtj ∼Dj [V (tj ) ≥ βj ] ≤ 12 − η, let cη (β) be 0;
P
• otherwise, let cη (β) be a nonnegative number such that j∈[m] Prtj ∼Dj [V (tj ) ≥ βj + cη (β)] ∈ 21 − η, 12 .
For every type t, let Cη (t) = {j | V (tj ) < βj + cη (β)}. Then,
X X
X
C OREη (β) = max
f (ti ) ·
σiS (ti ) · v (ti , S ∩ Cη (ti )) ,
σ∈P (D)
i∈[n] ti ∈Ti
S⊆[m]
where P (D) is the set of all feasible interim allocation rules. That is, C OREη (β) is the maximum welfare a
mechanism can extract out of the allocation of items whose individual value for the bidder they are allocated
to is lower than the adjusted thresholds.
It was shown in [13] that every collection of thresholds induces an upper bound to the optimal revenue. In
particular, for any choice of thresholds {βj }j∈[m] and η 11 , the revenue R EV (M ) of any BIC mechanism M is
upper bounded by
2 · S INGLE (M, β) + 4 · TAIL η (M, β) + 4 · C OREη (M, β) (Adapted from Theorem 2 in [13]).
These terms depend on the choice of {βj }j∈[m] , η as well as the mechanism M . We refer interested readers
to [13] for the definitions of these terms. To obtain a benchmark/upper bound of the optimal revenue, one can
simply replace the above expression with
2 · max S INGLE (M, β) + 4 · max TAIL η (M, β) + 4 · max C OREη (M, β).
M
M
M
10
For readers that are familiar with the definition of the C ORE in [13], C OREη (β) is essentially the same term but adapted for
symmetric bidders.
11
In [13], the thresholds are allowed to depend on the identity of the bidder. More specifically, for any i ∈ [n] and j ∈ [m], there
is an associated threshold βij . Their upper bound applies to asymmetric thresholds as well. Indeed, when the bidders are asymmetric,
their upper bound is induced by a set of asymmetric thresholds. As we only discuss symmetric bidders in this section, we focus on
symmetric thresholds for simplicity. Regarding η, Cai and Zhao only considered the case when η = 0, but their analysis can be easily
modified to accommodate any η ≤ 1/4. See Theorem 17 for the modified upper bound.
30
It is not hard to see that this benchmark may be impossible to approximate for certain choices of the thresholds.
Just imagine the case when the thresholds are extremely high, then maxM C OREη (M, β) becomes the optimal
social welfare which can be arbitrarily large comparing to the optimal revenue. What Cai and Zhao [13]
showed was that when the thresholds are b-balanced, this upper bound can indeed be approximated by the
revenue of an RSPM and an ASPE. From now on, we only consider b-balanced thresholds.
Using results in [13], we can further simplify the benchmark. In particular, maxM S INGLE (M, β) is less
2
· P OST R EV for any
than 6 · P OST R EV for all choices of {βj }j∈[n] and maxM TAIL η (M, β) is less than 1−b
choice of η and b-balanced thresholds {βj }j∈[n] . Moreover, maxM C OREη (M, β) ≤ C OREη (β). Combining
the inequalities above, we obtain the following Theorem.
Theorem 17 (Adapted from [13]). When the bidders are symmetric and have valuations that are subadditive
over independent items, for any constant b ∈ (0, 1), η ≤ 14 and a collection of b-balanced thresholds {βj }j∈[m] ,
OPT ≤
E.2
12 +
8
1−b
· P OST R EV + 4 · C OREη (β).
Symmetric Bidders with XOS Valuations
In this section, we show how to learn in polynomial time an approximately optimal mechanism for symmetric
bidders with XOS valuations given sample access to the distributions. According to Theorem 17, we only need
to learn a mechanism that approximates P OST R EV and C OREη (β). From Section 5.1, we know how to approximated P OST R EV in polynomial time, so we focus on learning a mechanism whose revenue approximates
C OREη (β).
First, we need a crucial property about XOS valuations.
Lemma 21 (Supporting Pricesn[23]). oIf v(t, ·) is an XOS function, for any subset S ⊆ [m] there exists a
collection of supporting prices θjS (t)
for v(t, S) such that
j∈S
P
1. v(t, S ′ ) ≥ j∈S ′ θjS (t) for all S ′ ⊆ S and
P
S
2.
j∈S θj (t) = v(t, S).
Let v ′ (ti , S) = v (ti , S ∩ Cη (ti )) and Fi be the distribution of the valuation v ′ (ti , S). As the bidders are
symmetric, Fi = Fi′ for any i and i′ . The C OREη (β) is exactly the maximum expected social welfare if
every bidder i’s valuation is drawn independently from Fi . Cai and Zhao [13] showed how to use an ASPE to
approximate this term. In the next Lemma, we construct the prices used in their ASPE and show its relation to
C OREη (β).
Lemma 22. (Adapted from [13]) Let every bidder i’s valuation be v ′ (ti , S) = v (ti , S ∩ Cη (ti )) when her type
is ti and σ ∗ be a symmetric allocation that achieves α-fraction of the optimal social welfare with respect to
v ′ (·, ·). For every item j ∈ [m], let
Qη,j =
X
1 X X
S∩C (t )
∗
·
f (ti ) ·
σiS
(ti ) · θj η i (ti ),
2
i∈[n] ti ∈Ti
n
o
S∩C (t )
where θj η i (ti )
j∈S∩Cη (ti )
S:j∈S
is the supporting prices for v (ti , S ∩ Cη (ti )). Let
u∗ (t, S) = max
v(t, S ∗ ) −
∗
S ⊆S
X
j∈S ∗
31
Qη,j
be a bidder’s utility for the set of items S when her type is t. We define δ∗ (S) to be the median of the
random
∗
∗
variable u (t, S) (with t ∼ ×j∈[m]Dj ) for any set S ⊆ [m]. The revenue of ASPE {Qη,j }j∈[m] , δ is at least
α · C OREη (β)
− C(b, η) · P OST R EV ,
2
where C(b, η) is a function that only depends on b and η.
Proof. We can essentially use the same proof in [13] to prove that the expected revenue of the ASPE is at least
X
Qη,j − C(b, η) · P OST R EV .
j∈[m]
For readers that are familiar with that proof, the only thing we need to make sure is that our choice of σ ∗ and
{βj }j∈[m] satisfy Lemma 5 in [13]. Since σ ∗ is symmetric and {βj }j∈[m] is b-balanced, for all bidder i and
item j
X
b
· (n − 1) = b,
Pr [V (tkj ) ≥ βj ] ≤
tkj ∼Dj
n−1
k6=i
and
Pr [V (tij ) ≥ βj ] /b ≥ 1/n ≥ ·
tij ∼Dj
Next, we argue
P
ti ∈Ti
j∈[m] Qη,j
X
j∈[m]
Qη,j =
X
≥
α·C OREη (β)
.
2
fi (ti ) ·
X
∗
σiS
(ti ).
S:j∈S
Observe that
X
1 X X
∗
·
f (ti ) ·
σiS
(ti ) · v ′ (ti , S) ≥ α · C OREη (β).
2
i∈[n] ti ∈Ti
S
The last inequality is because C OREη (β) is the maximum social welfare under v ′ (·, ·) and σ ∗ achieves α
fraction of that.
Lemma 23. For any ǫ > 0 and µ ∈ [0, 41 ], let {Qj }j∈[m] be a collection of prices such that |Qj − Qη,j | ≤ ǫ
for all j ∈ [m]. Let δ(S) be the entry fee function such that Pr
Pt∼×j∈[m] Dj [u(t, S) ≥ δ(S)] ∈ [1/2−µ, 1/2+µ]
∗
for any set S ⊆ [m], where u(t, S) = maxS∗⊆S v(t, S ) − j∈S ∗ Qj . Then, the ASPE(Q, δ) achieves at least
α·C OREη (β)
B1 (µ)
− B2 (b, η, µ) · P OST R EV − B3 (µ) · (m + n) · ǫ revenue when bidders’ valuations are XOS over
independent item. Both B1 (µ) and B3 (µ) are functions that only depend on µ and B2 (b, η, µ) is a function that
only depends on µ, b and η.
Proof. It turns out the proof in [13] is robust enough to accommodate the error ǫ and µ. We can prove the
claim by following essentially the same analysis as in [13]. We do not include the details here.
E.2.1 Leaning the ASPE in Polynomial Time
We first show how to learn a collection of b-balanced thresholds and the corresponding cη (β).
a
Lemma 24. For any positive constant b < 1 and η ≤ 14 , there is a polynomial time algorithm that computes
collection of b-balanced thresholds {βj }j∈[m] and cη (β) with probability 1−δ using O m2 n4 log m
samples
δ
from distribution ×j∈[m] Dj .
32
(1)
(K)
Proof. Given K = O m2 n4 log m + log 1δ samples tj , . . . , tj from distribution Dj , we construct Fj as
(K)
(1)
. According to the DKW Theorem [25], with probability
the uniform distribution over V tj , . . . , V tj
at least 1 − δ/m,
1
for all x
(6)
Pr [V (tj ) ≥ x] − Pr [vj ≥ x] ≤
tj ∼Dj
vj ∼Fj
c · mn2
where c is a constant that will be specified later. From now on, we assume that Inequality (6) holds for every
j, which happens with probability 1 − δ.
h
i
(ℓ)
(ℓ)
b
As 1/K ≤ 3nb 2 ≤ n−1
− 3nb 2 − nb − 3nb 2 , there must exist a sample tj such that Prvj ∼Fj vj ≥ V tj
∈
i
h
(ℓ)
b
b
b
b
. Note that
n + 3n2 , n−1 − 3n2 . Let βj = V tj
1
1
Pr [V (tj ) ≥ βj ] ∈ Pr [vj ≥ βj ] −
, Pr [vj ≥ βj ] +
.
tj ∼Dj
vj ∼Fj
c · mn2 vj ∼Fj
c · mn2
i
h
b
. Thus, βj is b-balanced for all item j.
If c is less than 3b , Prtj ∼Dj [V (tj ) ≥ βj ] ∈ nb , n−1
P
Next, we argue how to learn cη (β). If j∈[m] Prvj ∼Fj [vj ≥ βj ] ≤ 21 − η2 , let cη (β) = 0. This is a
P
valid choice, as j Prtj ∼Dj [V (tj ) ≥ βj ] is at most 12 − η2 + cn1 2 ≤ 21 according to inequality (6). SupP
pose j∈[m] Prvj ∼Fj [vj ≥ βj ] > 21 − η2 , as m/K < η4 , there must exist some item k ∈ [m] and a sami
h
P
(ℓ)
(ℓ)
≥ βk such that j∈[m] Prvj ∼Fj vj ≥ βj + V tk − βk ∈ 21 − η4 , 21 − η2 . Let cη (β) =
ple V tk
(ℓ)
V tk − βk . According to inequality (6),
X
j∈[m]
1 1 η
1
1 η
− −
, − +
.
Pr [V (tj ) ≥ βj + cη (β)] ∈
tj ∼Dj
2 4 cn2 2 2 cn2
P
For sufficiently large c, j Prtj ∼Dj [V (tj ) ≥ βj + cη (β)] ∈ 12 − η, 12 .
Finding each βj takes O(K log K) time and finding the cη (β) takes O(mK) time. So we can learn
in polynomial time
a collection of b-balanced thresholds {βj }j∈[m] and cη (β) with probability 1 − δ using
m
2
4
O m n log δ samples.
Next, we show how to learn the prices of the ASPE. As showed by Feige [27], there exists a polynomial
time algorithm that achieves 1− 1e fraction of the optimal social welfare when bidders have XOS valuations. We
let σ ∗ be the interim allocation rule induced by Feige’s algorithm and estimate the prices by running Feige’s
algorithm on sampled valuation profiles. To run Feige’s algorithm, we need a demand oracle for bidder’s
valuations. In the following Lemma, we argue that v ′ (t, ·) is an XOS function for any type t, and given a value
(or demand, XOS) oracle for v(t, ·), we can construct in polynomial time the corresponding oracle for v ′ (t, ·).
First, we define these oracles formally.
Definition 9. We consider the following three oracles for a bidder’s valuation function v(t, ·):
• Value oracle: takes a set S ⊆ [m] as the input and returns v(t, S).
• Demand oracle: takes a collection of prices {pj }P
j∈[m] as an input and returns the favorite set under
these prices, that is, S ∗ ∈ argmaxS∈[m] v(t, S) − j∈S pj .
• XOS oracle (only when v(t, ·) is XOS): takes a set S ⊆ [m] as the input and returns the supporting
prices {θjS (t)}j∈S for v(t, S).
33
Lemma 25. Given a collection of thresholds {βj }j∈[m] and cη (β). For any set S ⊆ [m], let v ′ (t, S) =
v(t, S ∩ Cη (t)). If v(t, ·) is an XOS function, v ′ (t, ·) is also an XOS function. Given a value (or demand, XOS)
oracle for v(t, ·), we can construct in polynomial time a value (or demand, XOS) oracle for v ′ (t, ·).
Proof. If v(t, ·) is an XOS function, v(t, ·) can be represented as the max of a collection of additive functions.
Observe that if we change the values for items in Cη (t) to 0 in each of these additive functions, v ′ (t, ·) equals
to the max of this new collection of additive functions. Hence, v ′ (t, ·) is also an XOS function.
If we are given a value oracle for v(t, ·), it is straightforward to construct a value oracle for v ′ (t, ·). If
we are given a demand oracle for v(t, ·), here is how to construct a demand oracle for v ′ (t, ·). For every
queried price vector {pj }j∈[m] , we change the price for each item outside Cη (t) to 2v(t, [m]) and keep the
prices for the items in Cη (t). Let this new price vector be p′ . We query the demand oracle of v(t, ·) on p′ .
The output set should also be the demand set for v ′ (t, ·) under prices p, as the bidder can only afford items
in Cnη (t) and v ′ (t,
o S) = v(t, S) for any set S ⊆ Cη (t). Finally, we consider the XOS oracle. For any set S,
S∩Cη (t)
let θj
(t)
S∩Cη (t)
j∈S∩Cη (t)
be the supporting prices for v(t, S ∩ Cη (t)). Let γjS (t) = θj
for all item j in
Cη (t) ∩ S and γjS (t) = 0 for all item j in S − Cη (t). According to the definition of v ′ (t, ·), {γjS (t)}j∈S is the
supporting price for v ′ (t, S). So given an XOS oracle for v(t, ·), we can compute the supporting price of any
set S for v ′ (t, ·) in polynomial time.
Lemma 25 shows that v ′ (t, ·) is also an XOS function for any type t and with access to a demand oracle
for v(t, ·) we can construct a demand oracle for v ′ (t, ·) in polynomial time. So we can indeed run Feige’s
algorithm on v ′ . In the next Lemma, we show how to learn a collection of prices {Qj }j∈[m] and entry fee
function δ(·, ·) such that the corresponding ASPE has high revenue.
Lemma 26. Given a collection of b-balanced thresholds {βj }j∈[m] and cη (β), and access to value, demand
and XOS oracles for valuation v(t, ·) for every type t, there is a polynomial time algorithm that learns an
C OREη (β)
ASPE({Qj }j∈[m] , δ) whose revenue is at least
− g(b, η) · P OST R EV − K2 · ξ · OPT with probability
K1
samples from ×j∈[m] Dj , where K1 and K2 are positive absolute
at least 1 − ζ using O n3 (m + n)2 log m
ζ
constants, and g(b, η) is a function that only depends on b and η.
Proof. According to Lemma 25, we can construct value, demand and XOS oracles for valuation v ′ (t, ·) given
access to the corresponding oracles for v(t, ·). We use {γjS (t)}j∈S to denote the output of the XOS oracle for
S∩Cη (t)
v ′ (t, ·) on set S. In particular, γjS (t) = 0 for all j ∈ S − Cη (t) and γjS (t) = θj
(t) for all j ∈ S ∩ Cη (t),
S∩C (t)
{θj η (t)}j∈S∩Cη (t)
where
is the supporting prices for v(t, S ∩ Cη (t)). Let A(t) be the allocation computed
by Feige’s algorithm on the valuation profile (v ′ (t1 , ·), . . . , v ′ (tn , ·)), where Ai (t) denotes the set of items that
bidder i receives. Let σ ∗ be the interim allocation rule induced by A(·) when bidders types are all drawn from
∗ (t ) = Pr
×j∈[m] Dj independently. That is, σiS
i
t−i [Ai (t) = S]. We use the same definition for Qη,j as in
Lemma 22. In other words, Qη,j is the contribution of item j to the social welfare under allocation rule σ ∗ , so
we can rewrite it as
X
1
A (t)
· Et
1 [j ∈ Ai (t)] · γj i (ti ) .
2
i∈[n]
A (t(ℓ) ) (ℓ)
P
(ti ). We
Let t(1), . . . , t(K) be K sampled type profiles, and q (ℓ) = 21 i∈[n] 1 j ∈ Ai (t(ℓ) ) · γj i
1 P
(ℓ)
S
set Qj to be K · ℓ∈[K] q . Since γj (t) ≤ βj + cη (β) for any j, S and t, Qj ≤ βj + cη (β). By the Chernoff
bound,
Pr [|Qj − Qη,j | ≤ ǫ · (βj + cη (β))] ≥ 1 − 2 exp(−2K · ǫ2 ).
34
As {βj }j∈[m] is a collection of b-balanced thresholds, we can obtain revenue βj · nb by only selling
item j to one bidder at price βj . Hence, βj ≤ n·OPT
b . Now, consider a posted price mechanism that
sells
item
j
at
price
β
+
c
(β).
A
single
bidder
will
purchase at least one item with probability at least
j
η
P
1
j Prtj ∼Dj [V (tj ) ≥ βj + cη (β)] which is no less than 2 − η if cη (β) > 0. Hence, the revenue of this mech
ξ
anism is at least cη (β) · 12 − η . As η ≤ 14 , cη (β) ≤ 4OPT. If we let ǫ = (m+n)·(n/b+4)
for some small
i
h
log 4m
ζ
ξ
constant ξ which will be specified later and K = 2ǫ2ζ , we have Pr |Qj − Qη,j | ≤ m+n
· OPT ≥ 1 − 2m
.
In other words, with O n3 (m + n)2 log m
samples from ×j∈[m] Dj (as each tℓ costs n samples), we can
ζ
ξ
· OPT for all item j
learn in polynomial time a collection of prices {Qj }j∈[m] such that |Qj − Qη,j | ≤ m+n
with probability 1 − ζ/2.
Next, we consider the entry fee function. We use essentially the same argument as in Lemma 19. Suppose
we take L samples t(1) , · · · , t(L) from ×j∈[m]Dj . Define the entry fee δ(S) for set S under {Qj }j∈[m] to be
P
the median of u(t(1) , S), · · · , u(t(L) , S), where u(t, S) = maxS∗⊆S v(t, S ∗ ) − j∈S ∗ pj . Given any constant
µ ∈ [0, 1/4], for any fixed set S, it is easy to argue that the probability for Prt∼×j∈[m] Dj [u(t, S) ≥ δ(S)] to
be larger than 21 + µ or less than 21 − µ is at most 2 exp(−2Lµ2 ) due to the Chernoff bound. If we let L to be
1/ζ
a · m+log
for a sufficiently large constant a, the probability that δ(·) is a µ-balanced entry fee function is at
µ2
least 1 − ζ/2 by the union bound.
samples from ×j∈[m] Dj , we can compute in polynomial time a colHence, with O n3 (m + n)2 log m
ζ
lection of prices {Qj }j∈[m] and a entry fee function δ(·) such that the revenue of the ASPE {Qj }j∈[m] , δ(·)
(1−1/e)·C ORE (β)
η
− B2 (b, η, µ) · P OST R EV − ξ · B3 (µ) · OPT with probability 1 − ζ due to Lemma 23.
is at least
B1 (µ)
Our claim follows by fixing the value of µ to be some constant.
Theorem 18. For symmetric bidders with valuations that are XOS over independent items,
1. when V (tj ) is upper bounded by H for any j ∈ [m] and any tj , with
O
n 5 + m2 n
4
m
+
· log
δ
2
!
1
1
1
m2 n log n log + log
ǫ
ǫ
δ
samples from ×j∈[m] Dj , we can learn in polynomial time with probability 1 − δ a mechanism whose
revenue is at least c1 · OPT − ǫ · H for some absolute constant c1 ;
2. when the distribution of random variable V (tj ) with tj ∼ Dj is regular for all item j ∈ [m], with
m
nm
O n5 · log
+ max{m, n}2 m2 n2 · log
δ
δ
samples from ×j∈[m] Dj , we can learn in polynomial time with probability 1 − δ a mechanism whose
revenue is at least c2 · OPT for some absolute constant c2 .
Proof of Theorem 18: Combining Lemma 24, Lemma 26 and Theorem 17, we know how to compute in
polynomial time an ASPE whose revenue is at least a1 · OPT −a2 · P OST
R EV with probability 1 − δ/2 for
some absolute constant a1 , a2 , and we only need O n5 + m2 n4 · log m
δ samples from ×j∈[m] Dj . When the
P OST R EV
distributions are bounded, we canlearn in polynomial time an RSPM
whose revenue is at least 144 − ξH
2
1
with probability 1 − δ/2 using O
m2 n log n log 1ǫ + log 1δ
samples (Theorem 11). By choosing the
ξ
ratio between ξ and ǫ to be the right constant, we can show the first part of our claim. When V (tj ) is a regular
R EV
random variable for every item j, we can learn in polynomial time an RSPM whose revenue is at least P OST
33
35
samples (Theorem 12). Therefore, we can
with probability 1 − δ/2 using O max{m, n}2 m2 n2 · log nm
δ
learn a mechanism in polynomial time such that with probability 1 − δ whose revenue is at least a constant
fraction of the OPT. This proves the second part of our claim.✷
E.3
Symmetric Bidders with Subadditive Valuations
In this section, we argue that if the bidders are symmetric and m = O(n), there exists a collection of bbalanced thresholds {βj }j∈[m] for a fixed constant b, such that P OST R EV is within a constant fraction of the
benchmark. Note that this argument only applies to symmetric bidders, as b-balanced thresholds may not even
exist for asymmetric bidders.
We set η = 0 for this section and drop the subscript η when there is no confusion. We show how to
upper bound C ORE(β) with P OST R EV by choosing a particular collection of b-balanced thresholds. Let Z =
n
n
. It is not hard to see that 3Z
-balanced thresholds exist, as we can choose βj such that
max{m, n} and b = 3Z
1
Prtj ∼Dj [V (tj ) ≥ βj ] = 3Z .
P
n
-balanced thresholds, then C ORE(β) ≤ j∈[m] βj .
Lemma 27. Let {βj }j∈[m] be a collection of 3Z
Proof. As {βj }j∈[m] are
n
3Z -balanced,
Prtij ∼Dj [V (tij ) ≥ βj ] ≤
X
j∈[m]
σ∈P (D)
≤ max
σ∈P (D)
= max
σ∈P (D)
= max
σ∈P (D)
≤
X
βj
P
X X
1
2Z .
Therefore,
j∈[m] βj .
X
f (ti ) ·
i∈[n] ti ∈Ti
X X
i∈[n] ti ∈Ti
X X
X
σiS (ti ) · v(ti , S ∩ C(ti ))
S⊆[m]
f (ti ) ·
i∈[n] j∈[m]
j
≤
1
Pr [V (tj ) ≥ βj ] ≤ ,
tj ∼Dj
2
so c(β) = 0. Next, we upper bound C ORE(β) by
C ORE(β) = max
n
(n−1)·3Z
βj ·
βj ·
X
X
σiS (ti ) ·
j∈S
S⊆[m]
X
f (ti ) ·
ti ∈Ti
X X
i∈[n] ti ∈Ti
X
S:j∈S
fi (ti ) ·
X
S:j∈S
βj
σiS (ti )
σiS (ti )
j
The first inequality is because v(ti , ·) is a subadditive function, so
X
X
v(ti , S ∩ Ci (ti )) ≤
V (tij ) ≤
j∈S∩Ci (ti )
j∈S∩Ci (ti )
βj ≤
X
βj .
j∈S
P
P
P
The last inequality is because i∈[n] ti ∈Ti fi (ti ) · S:j∈S σiS (ti ) ≤ 1 is the ex-ante probability for bidder
i to receive item j, and for any feasible interim allocation σ, the sum of all bidders’ ex-ante probabilities for
receiving item j should not exceed 1.
P
In the following Lemma, we demonstrate that j∈[m] βj is upper bounded by 9Z
n · P OST R EV .
36
Lemma 28. Let {βj }j∈[m] be a collection of
n
3Z -balanced
thresholds, P OST R EV ≥
n
9Z
·
P
j∈[m] βj .
Proof. Let us consider an RSPM where the price for selling item j to bidder i is βj . Bidder i purchases item
n
j if that is the only item she can afford and no one else can afford item j. As {βj }j∈[m] are 3Z
-balanced, the
probability that no one else can afford item j is at least
X
2
n
1 −
)≥ .
Pr [V (tkj ) ≥ βj ] ≥ (1 −
tkj ∼Dj
3Z
3
k6=i
Also, the probability that i cannot afford any item other than j is at least
X
n(m − 1)
1
1 −
Pr [V (tiℓ ) ≥ βℓ ] ≥ 1 −
≥ .
tiℓ ∼Dℓ
3Z(n − 1)
2
ℓ6=j
1
Therefore, bidder i purchases item j with probability at least 13 Prtij ∼Dj [V (tij ≥ βj )] ≥ 9Z
. Whenever this
P P βj
n P
event happens, it contributes βj to the revenue. So the total revenue is at least j i 9Z = 9Z · j βj .
Combining Theorem 17, Lemma 27 and 28, we obtain the following Theorem.
Theorem 19. For symmetric bidders with valuations that are subadditive over independent items,
36 max{n, m}
· P OST R EV .
OPT ≤ 24 +
n
n
9 max{n,m} · C ORE (β) if {βj }j∈[m] is a collection
9 max{n,m}
n
·
n
3 max{n,m} and replacing C ORE (β) with
Proof. Combining Lemma 27 and 28, we have P OST R EV ≥
n
of 3 max{n,m}
-balanced thresholds. By setting b to be
P OST R EV in Theorem 17, we have
OPT ≤
As
n
3 max{n,m}
12 +
8
1−
≤ 1/3,
OPT ≤
n
3 max{n,m}
36 max{n, m}
+
n
36 max{n, m}
24 +
n
!
· P OST R EV .
· P OST R EV .
E.3.1 Learning an Approximately Optimal Mechanism for Symmetric Subadditive Bidders
With Theorem 19, we only need to learn a mechanism that approximates the optimal revenue obtainable by
any RSPM. The next Lemma connects RSPMs with SPMs in an induced unit-demand setting.
Lemma 29. Consider n symmetric bidders whose types are drawn independently from ×m
j=1 Dj . Let Fj be
the distribution for random variable V (tj ) where tj ∼ Dj . We define an induced unit-demand setting with n
symmetric unit-demand bidders whose values for item j are drawn independently from Fj . For any collection
of prices {pij }i∈[n],j∈[m], the revenue of the RSPM with these prices in the original setting is exactly the same
as the revenue of the SPM with these prices in the induced unit-demand setting.
Proof. As in an RSPM bidders can purchase at most one item, bidders behave exactly the same as in the
induced unit-demand setting. Since the prices in the SPM and RSPM are the same, bidders purchase exactly
the same items. Hence, the revenue is the same.
37
Corollary 3. For symmetric bidders with valuations that are subadditive over independent items,
n
· OPT,
OPTU D ≥ Ω
Z
where Z = max{m, n} and OPTU D is the optimal revenue for the induced unit-demand setting.
Proof. Combine Theorem 19 and Lemma 29.
Lemma 29 implies that learning an approximately optimal RSPM is equivalent as learning an approximately optimal SPM in the induced unit-demand setting. Next, we apply our results in Section 5.1 to the
induced unit-demand setting to learn an RSMP that approximates the optimal revenue in the original setting.
In the next Theorem, we show that even though the bidders’ valuations could be complex set functions, e.g.,
submodular, XOS and subadditive, as long as m = O(n), the approximate distributions for the bidders’ values
for winning any single item provides sufficient information to learn an approximately optimal mechanism.
Theorem 20. For symmetric bidders with valuations that are subadditive over independent items, let Fj be
the distribution of V (tj ) where tj ∼ Dj . If Fj is supported on [0, H] for all j ∈ [m], given distributions F̂j
≤ ǫ for all j ∈ [m], there is a polynomial time algorithm that constructs a randomized
where F̂j − Fj
K
RSPM whose revenue under the true distribution D is at least
1
n
− (n + m) · ǫ · Ω
· OPT − 2ǫ · mnH .
4
max{m, n}
Proof. Let Z = max{m, n}. According to Corollary 3, OPTU D = Ω Zn · OPT. Since ||F̂j − Fj ||K ≤ ǫ
for all j ∈ [m], we can learn a randomized SPM in the induced unit-demand
setting whose revenue under the
OPTU D
1
true distribution is at least 4 − (n + m) · ǫ ·
− 2ǫ · mnH based on Theorem 10. By Lemma 29,
8
we can construct an RSPM with the same collection of (randomized) prices and achieve revenue
n
1
· OPT − 2ǫ · mnH
− (n + m) · ǫ · Ω
4
Z
in the original setting.
If we are given sample access to bounded distributions, we show in the following Theorem that a polynomial number of samples suffices to learn an approximately optimal mechanism, when m = O(n).
Theorem 21. For symmetric bidders with valuations that are subadditive over independent items, let Fj be
the distribution of V (tj ) where tj ∼ Dj . If Fj is supported
on [0,H] for all j ∈ [m], there is a polynomial
n
max{m,n}
· OPT − ǫH with probability 1 − δ using
!
2
1
1
1
· m2 n log n log + log
O
ǫ
ǫ
δ
time algorithm that learns an RSPM whose revenue is Ω
samples.
Proof. According to Corollary 3, OPTU D = Ω
n
max{m,n}
· OPT. Due to Theorem 11,
2
!
1
1
1
O
· m2 n log n log + log
ǫ
ǫ
δ
samples suffices to learn in polynomial time with probability 1 − δ an SPM with revenue at least Ω(OPTU D ) −
ǫ · H for the induced unit-demand
Lemma 29, we can construct an RSPM with the same collection
setting. By
n
of prices and achieve revenue Ω max{m,n} · OPT − ǫH in the original setting.
38
Finally, if the distribution of the random variable V (tj ) with tj ∼ Dj is regular for all item j ∈ [m], we
prove in the next theorem that there exists a prior-independent mechanism that achieves a constant fraction
of the optimal revenue if m = O(n). Note that approximately optimal prior-independent mechanisms for
symmetric unit-demand bidders are known due to the work by Devanur et al. [21] and Roughgarden et al. [39].
Our result is obtained by combining Theorem 19 and the afore-mentioned prior independent mechanisms.
Theorem 22. For symmetric bidders with valuations that are subadditive over independent items, let Fj be the
distribution of V (tj ) where
j . If Fj is regular for all j ∈ [m], there is a prior-independent mechanism
tj ∼ D
with revenue at least Ω
n
max{m,n}
· OPT. Moreover, the mechanism can be implemented efficiently.
Proof. The mechanism in [21] or [39] provides an approximately optimal prior-independent mechanism in
the induced unit-demand setting. Let us use M to denote this mechanism. Suppose we restrict every bidder
to purchase at most one item in the original setting and then run mechanism M . The expected revenue is
UD
the same as M
revenue in the induced setting. Since M ’s expected
revenue
is Ω(OPT ) and
’s expected
n
n
·OPT, the mechanism we constructed has revenue Ω max{m,n}
·OPT. Since M can
OPTU D = Ω max{m,n}
be implemented efficiently for unit-demand bidders, our mechanism can also be implemented efficiently.
39
References
[1] Saeed Alaei. Bayesian Combinatorial Auctions: Expanding Single Buyer Mechanisms to Many Buyers.
In the 52nd Annual IEEE Symposium on Foundations of Computer Science (FOCS), 2011. 1, 1, 5.1
[2] Saeed Alaei, Hu Fu, Nima Haghpanah, and Jason Hartline. The Simple Economics of Approximately
Optimal Auctions. In the 54th Annual IEEE Symposium on Foundations of Computer Science (FOCS),
2013. 1
[3] Saeed Alaei, Hu Fu, Nima Haghpanah, Jason Hartline, and Azarakhsh Malekian. Bayesian Optimal
Auctions via Multi- to Single-agent Reduction. In the 13th ACM Conference on Electronic Commerce
(EC), 2012. 1
[4] Susan Athey and Philip Haile. Nonparametric approaches to auctions. Handbook of econometrics, 6(Part
A):3847–3965, 2007. 1
[5] Moshe Babaioff, Nicole Immorlica, Brendan Lucier, and S. Matthew Weinberg. A Simple and Approximately Optimal Mechanism for an Additive Buyer. In the 55th Annual IEEE Symposium on Foundations
of Computer Science (FOCS), 2014. 1
[6] Anand Bhalgat, Sreenivas Gollapudi, and Kamesh Munagala. Optimal auctions via the multiplicative
weight method. In ACM Conference on Electronic Commerce, EC ’13, Philadelphia, PA, USA, June
16-20, 2013, pages 73–90, 2013. 1
[7] Yang Cai and Constantinos Daskalakis. Extreme-Value Theorems for Optimal Multidimensional Pricing.
In the 52nd Annual IEEE Symposium on Foundations of Computer Science (FOCS), 2011. 1, 1, 5.1, 12
[8] Yang Cai, Constantinos Daskalakis, and S. Matthew Weinberg. An Algorithmic Characterization of
Multi-Dimensional Mechanisms. In the 44th Annual ACM Symposium on Theory of Computing (STOC),
2012. 1
[9] Yang Cai, Constantinos Daskalakis, and S. Matthew Weinberg. Optimal Multi-Dimensional Mechanism
Design: Reducing Revenue to Welfare Maximization. In the 53rd Annual IEEE Symposium on Foundations of Computer Science (FOCS), 2012. 1
[10] Yang Cai, Constantinos Daskalakis, and S. Matthew Weinberg. Understanding Incentives: Mechanism
Design becomes Algorithm Design. In the 54th Annual IEEE Symposium on Foundations of Computer
Science (FOCS), 2013. 1
[11] Yang Cai, Nikhil R. Devanur, and S. Matthew Weinberg. A duality based unified approach to bayesian
mechanism design. In the 48th Annual ACM Symposium on Theory of Computing (STOC), 2016. 1, 1,
5.1, 5.2, 6, 7, B.1, B.1, B.1
[12] Yang Cai and Zhiyi Huang. Simple and Nearly Optimal Multi-Item Auctions. In the 24th Annual ACMSIAM Symposium on Discrete Algorithms (SODA), 2013. 1
[13] Yang Cai and Mingfei Zhao. Simple mechanisms for subadditive buyers via duality. In the 49th Annual
ACM Symposium on Theory of Computing (STOC), 2017. 1, 1, 5, 5, 6, 8, 7, D, E.1, E.1, 10, 11, 17, E.2,
22, E.2, E.2
[14] Shuchi Chawla, Jason D. Hartline, and Robert D. Kleinberg. Algorithmic Pricing via Virtual Valuations.
In the 8th ACM Conference on Electronic Commerce (EC), 2007. 1
40
[15] Shuchi Chawla, Jason D. Hartline, David L. Malec, and Balasubramanian Sivan. Multi-Parameter Mechanism Design and Sequential Posted Pricing. In the 42nd ACM Symposium on Theory of Computing
(STOC), 2010. 1, 1, 5.1, 5.1, B.1, B.1, 8, B.1
[16] Shuchi Chawla and J. Benjamin Miller. Mechanism design for subadditive agents via an ex-ante relaxation. In Proceedings of the 2016 ACM Conference on Economics and Computation, EC ’16, Maastricht,
The Netherlands, July 24-28, 2016, pages 579–596, 2016. 1
[17] Richard Cole and Tim Roughgarden. The sample complexity of revenue maximization. In Proceedings
of the 46th Annual ACM Symposium on Theory of Computing, 2014. 1
[18] Constantinos Daskalakis. Multi-item auctions defying intuition? ACM SIGecom Exchanges, 14(1):41–
75, 2015. 1
[19] Constantinos Daskalakis, Alan Deckelbaum, and Christos Tzamos. Strong duality for a multiple-good
monopolist. Econometrica, 2017. 1
[20] Constantinos Daskalakis, Nikhil R. Devanur, and S. Matthew Weinberg. Revenue maximization and
ex-post budget constraints. In Proceedings of the Sixteenth ACM Conference on Economics and Computation, EC ’15, Portland, OR, USA, June 15-19, 2015, pages 433–447, 2015. 1
[21] Nikhil R. Devanur, Jason D. Hartline, Anna R. Karlin, and C. Thach Nguyen. Prior-independent multiparameter mechanism design. In Internet and Network Economics - 7th International Workshop, WINE
2011, Singapore, December 11-14, 2011. Proceedings, pages 122–133, 2011. E.3.1, E.3.1
[22] NR Devanur, Z Huang, and CA Psomas. The sample complexity of auctions with side information. In
48th Annual ACM SIGACT Symposium on Theory of Computing, 2016. 1
[23] Shahar Dobzinski, Noam Nisan, and Michael Schapira. Approximation algorithms for combinatorial
auctions with complement-free bidders. In STOC, pages 610–618, 2005. 21
[24] Shaddin Dughmi, Li Han, and Noam Nisan. Sampling and representation complexity of revenue maximization. In International Conference on Web and Internet Economics (WINE), 2014. 1, 2
[25] Aryeh Dvoretzky, Jack Kiefer, and Jacob Wolfowitz. Asymptotic minimax character of the sample distribution function and of the classical multinomial estimator. The Annals of Mathematical Statistics, pages
642–669, 1956. 1, 4, B.3, E.2.1
[26] Edith Elkind. Designing and learning optimal finite support auctions. In Proceedings of the eighteenth
annual ACM-SIAM symposium on Discrete algorithms, pages 736–745. Society for Industrial and Applied Mathematics, 2007. 1
[27] Uriel Feige. On maximizing welfare when utility functions are subadditive. SIAM Journal on Computing,
39(1):122–142, 2009. E.2.1
[28] Kira Goldner and Anna R. Karlin. A prior-independent revenue-maximizing auction for multiple additive
bidders. In Web and Internet Economics - 12th International Conference, WINE 2016, Montreal, Canada,
December 11-14, 2016, Proceedings, pages 160–173, 2016. (document), 1, 1, 3, 5.2, 5.2, C.1
[29] Yannai A. Gonczarowski and Noam Nisan. Efficient empirical revenue maximization in single-parameter
auction environments. CoRR, abs/1610.09976, 2016. 1
[30] Emmanuel Guerre, Isabelle Perrigne, and Quang Vuong. Optimal nonparametric estimation of first-price
auctions. Econometrica, 68(3):525–574, 2000. 1
41
[31] Zhiyi Huang, Yishay Mansour, and Tim Roughgarden. Making the most of your samples. In Proceedings
of the Sixteenth ACM Conference on Economics and Computation, 2015. 1
[32] Robert Kleinberg and S. Matthew Weinberg. Matroid Prophet Inequalities. In the 44th Annual ACM
Symposium on Theory of Computing (STOC), 2012. 5.1
[33] Mehryar Mohri and Andres Munoz Medina. Learning theory and algorithms for revenue optimization in
second price auctions with reserve. In ICML, pages 262–270, 2014. 1
[34] Jamie Morgenstern and Tim Roughgarden. The pseudo-dimension of near-optimal auctions. arXiv
preprint arXiv:1506.03684, 2015. 1
[35] Jamie Morgenstern and Tim Roughgarden. Learning simple auctions. In Proceedings of the 29th Conference on Learning Theory, COLT 2016, New York, USA, June 23-26, 2016, pages 1298–1318, 2016.
(document), 1, 1, 3, 5.1, 5.2, 11, C.1
[36] Roger B. Myerson. Optimal Auction Design. Mathematics of Operations Research, 6(1):58–73, 1981. 1
[37] Harry J Paarsch and Han Hong. An introduction to the structural econometrics of auction data. MIT
Press Books, 1, 2006. 1
[38] Tim Roughgarden and Okke Schrijvers. Ironing in the dark. In Proceedings of the 2016 ACM Conference
on Economics and Computation, pages 1–18. ACM, 2016. 1
[39] Tim Roughgarden, Inbal Talgam-Cohen, and Qiqi Yan. Supply-limiting mechanisms. In 13th ACM
Conference on Electronic Commerce (EC), 2012. E.3.1, E.3.1
[40] Aviad Rubinstein and S. Matthew Weinberg. Simple mechanisms for a subadditive buyer and applications to revenue monotonicity. In Proceedings of the Sixteenth ACM Conference on Economics and
Computation, EC ’15, Portland, OR, USA, June 15-19, 2015, pages 377–394, 2015. 1, 1
[41] Ester Samuel-Cahn. Comparison of threshold stop rules and maximum for independent nonnegative
random variables. Ann. Probab., 12(4):1213–1216, 11 1984. C.1
[42] Andrew Chi-Chih Yao. An n-to-1 bidder reduction for multi-item auctions and its applications. In SODA,
2015. 1, 1, 5.2, 7
42
| 8cs.DS
|
Subsets and Splits
No saved queries yet
Save your SQL queries to embed, download, and access them later. Queries will appear here once saved.